net/smc/af_smc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   6  *  applies to SOCK_STREAM sockets only
   7  *  offers an alternative communication option for TCP-protocol sockets
   8  *  applicable with RoCE-cards only
   9  *
  10  *  Initial restrictions:
  11  *    - support for alternate links postponed
  12  *
  13  *  Copyright IBM Corp. 2016, 2018
  14  *
  15  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  16  *              based on prototype from Frank Blaschka
  17  */
  18
  19 #define KMSG_COMPONENT "smc"
  20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  21
  22 #include <linux/module.h>
  23 #include <linux/socket.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/in.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/rcupdate_wait.h>
  29 #include <linux/ctype.h>
  30
  31 #include <net/sock.h>
  32 #include <net/tcp.h>
  33 #include <net/smc.h>
  34 #include <asm/ioctls.h>
  35
  36 #include <net/net_namespace.h>
  37 #include <net/netns/generic.h>
  38 #include "smc_netns.h"
  39
  40 #include "smc.h"
  41 #include "smc_clc.h"
  42 #include "smc_llc.h"
  43 #include "smc_cdc.h"
  44 #include "smc_core.h"
  45 #include "smc_ib.h"
  46 #include "smc_ism.h"
  47 #include "smc_pnet.h"
  48 #include "smc_netlink.h"
  49 #include "smc_tx.h"
  50 #include "smc_rx.h"
  51 #include "smc_close.h"
  52 #include "smc_stats.h"
  53
  54 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  55                                                  * creation on server
  56                                                  */
  57 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  58                                                  * creation on client
  59                                                  */
  60
  61 struct workqueue_struct *smc_hs_wq;     /* wq for handshake work */
  62 struct workqueue_struct *smc_close_wq;  /* wq for close work */
  63
  64 static void smc_tcp_listen_work(struct work_struct *);
  65 static void smc_connect_work(struct work_struct *);
  66
  67 static void smc_set_keepalive(struct sock *sk, int val)
  68 {
  69         struct smc_sock *smc = smc_sk(sk);
  70
  71         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  72 }
  73
  74 static struct smc_hashinfo smc_v4_hashinfo = {
  75         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  76 };
  77
  78 static struct smc_hashinfo smc_v6_hashinfo = {
  79         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  80 };
  81
  82 int smc_hash_sk(struct sock *sk)
  83 {
  84         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  85         struct hlist_head *head;
  86
  87         head = &h->ht;
  88
  89         write_lock_bh(&h->lock);
  90         sk_add_node(sk, head);
  91         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  92         write_unlock_bh(&h->lock);
  93
  94         return 0;
  95 }
  96 EXPORT_SYMBOL_GPL(smc_hash_sk);
  97
  98 void smc_unhash_sk(struct sock *sk)
  99 {
 100         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 101
 102         write_lock_bh(&h->lock);
 103         if (sk_del_node_init(sk))
 104                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 105         write_unlock_bh(&h->lock);
 106 }
 107 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 108
 109 struct proto smc_proto = {
 110         .name           = "SMC",
 111         .owner          = THIS_MODULE,
 112         .keepalive      = smc_set_keepalive,
 113         .hash           = smc_hash_sk,
 114         .unhash         = smc_unhash_sk,
 115         .obj_size       = sizeof(struct smc_sock),
 116         .h.smc_hash     = &smc_v4_hashinfo,
 117         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 118 };
 119 EXPORT_SYMBOL_GPL(smc_proto);
 120
 121 struct proto smc_proto6 = {
 122         .name           = "SMC6",
 123         .owner          = THIS_MODULE,
 124         .keepalive      = smc_set_keepalive,
 125         .hash           = smc_hash_sk,
 126         .unhash         = smc_unhash_sk,
 127         .obj_size       = sizeof(struct smc_sock),
 128         .h.smc_hash     = &smc_v6_hashinfo,
 129         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 130 };
 131 EXPORT_SYMBOL_GPL(smc_proto6);
 132
 133 static void smc_restore_fallback_changes(struct smc_sock *smc)
 134 {
 135         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 136                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 137                 smc->clcsock->file = NULL;
 138         }
 139 }
 140
 141 static int __smc_release(struct smc_sock *smc)
 142 {
 143         struct sock *sk = &smc->sk;
 144         int rc = 0;
 145
 146         if (!smc->use_fallback) {
 147                 rc = smc_close_active(smc);
 148                 sock_set_flag(sk, SOCK_DEAD);
 149                 sk->sk_shutdown |= SHUTDOWN_MASK;
 150         } else {
 151                 if (sk->sk_state != SMC_CLOSED) {
 152                         if (sk->sk_state != SMC_LISTEN &&
 153                             sk->sk_state != SMC_INIT)
 154                                 sock_put(sk); /* passive closing */
 155                         if (sk->sk_state == SMC_LISTEN) {
 156                                 /* wake up clcsock accept */
 157                                 rc = kernel_sock_shutdown(smc->clcsock,
 158                                                           SHUT_RDWR);
 159                         }
 160                         sk->sk_state = SMC_CLOSED;
 161                         sk->sk_state_change(sk);
 162                 }
 163                 smc_restore_fallback_changes(smc);
 164         }
 165
 166         sk->sk_prot->unhash(sk);
 167
 168         if (sk->sk_state == SMC_CLOSED) {
 169                 if (smc->clcsock) {
 170                         release_sock(sk);
 171                         smc_clcsock_release(smc);
 172                         lock_sock(sk);
 173                 }
 174                 if (!smc->use_fallback)
 175                         smc_conn_free(&smc->conn);
 176         }
 177
 178         return rc;
 179 }
 180
 181 static int smc_release(struct socket *sock)
 182 {
 183         struct sock *sk = sock->sk;
 184         struct smc_sock *smc;
 185         int rc = 0;
 186
 187         if (!sk)
 188                 goto out;
 189
 190         sock_hold(sk); /* sock_put below */
 191         smc = smc_sk(sk);
 192
 193         /* cleanup for a dangling non-blocking connect */
 194         if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
 195                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 196
 197         if (cancel_work_sync(&smc->connect_work))
 198                 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
 199
 200         if (sk->sk_state == SMC_LISTEN)
 201                 /* smc_close_non_accepted() is called and acquires
 202                  * sock lock for child sockets again
 203                  */
 204                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 205         else
 206                 lock_sock(sk);
 207
 208         rc = __smc_release(smc);
 209
 210         /* detach socket */
 211         sock_orphan(sk);
 212         sock->sk = NULL;
 213         release_sock(sk);
 214
 215         sock_put(sk); /* sock_hold above */
 216         sock_put(sk); /* final sock_put */
 217 out:
 218         return rc;
 219 }
 220
 221 static void smc_destruct(struct sock *sk)
 222 {
 223         if (sk->sk_state != SMC_CLOSED)
 224                 return;
 225         if (!sock_flag(sk, SOCK_DEAD))
 226                 return;
 227
 228         sk_refcnt_debug_dec(sk);
 229 }
 230
 231 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 232                                    int protocol)
 233 {
 234         struct smc_sock *smc;
 235         struct proto *prot;
 236         struct sock *sk;
 237
 238         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 239         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 240         if (!sk)
 241                 return NULL;
 242
 243         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 244         sk->sk_state = SMC_INIT;
 245         sk->sk_destruct = smc_destruct;
 246         sk->sk_protocol = protocol;
 247         smc = smc_sk(sk);
 248         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 249         INIT_WORK(&smc->connect_work, smc_connect_work);
 250         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 251         INIT_LIST_HEAD(&smc->accept_q);
 252         spin_lock_init(&smc->accept_q_lock);
 253         spin_lock_init(&smc->conn.send_lock);
 254         sk->sk_prot->hash(sk);
 255         sk_refcnt_debug_inc(sk);
 256         mutex_init(&smc->clcsock_release_lock);
 257
 258         return sk;
 259 }
 260
 261 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 262                     int addr_len)
 263 {
 264         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 265         struct sock *sk = sock->sk;
 266         struct smc_sock *smc;
 267         int rc;
 268
 269         smc = smc_sk(sk);
 270
 271         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 272         rc = -EINVAL;
 273         if (addr_len < sizeof(struct sockaddr_in))
 274                 goto out;
 275
 276         rc = -EAFNOSUPPORT;
 277         if (addr->sin_family != AF_INET &&
 278             addr->sin_family != AF_INET6 &&
 279             addr->sin_family != AF_UNSPEC)
 280                 goto out;
 281         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 282         if (addr->sin_family == AF_UNSPEC &&
 283             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 284                 goto out;
 285
 286         lock_sock(sk);
 287
 288         /* Check if socket is already active */
 289         rc = -EINVAL;
 290         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 291                 goto out_rel;
 292
 293         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 294         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 295
 296 out_rel:
 297         release_sock(sk);
 298 out:
 299         return rc;
 300 }
 301
 302 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 303                                    unsigned long mask)
 304 {
 305         /* options we don't get control via setsockopt for */
 306         nsk->sk_type = osk->sk_type;
 307         nsk->sk_sndbuf = osk->sk_sndbuf;
 308         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 309         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 310         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 311         nsk->sk_mark = osk->sk_mark;
 312         nsk->sk_priority = osk->sk_priority;
 313         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 314         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 315         nsk->sk_err = osk->sk_err;
 316
 317         nsk->sk_flags &= ~mask;
 318         nsk->sk_flags |= osk->sk_flags & mask;
 319 }
 320
 321 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 322                              (1UL << SOCK_KEEPOPEN) | \
 323                              (1UL << SOCK_LINGER) | \
 324                              (1UL << SOCK_BROADCAST) | \
 325                              (1UL << SOCK_TIMESTAMP) | \
 326                              (1UL << SOCK_DBG) | \
 327                              (1UL << SOCK_RCVTSTAMP) | \
 328                              (1UL << SOCK_RCVTSTAMPNS) | \
 329                              (1UL << SOCK_LOCALROUTE) | \
 330                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 331                              (1UL << SOCK_RXQ_OVFL) | \
 332                              (1UL << SOCK_WIFI_STATUS) | \
 333                              (1UL << SOCK_NOFCS) | \
 334                              (1UL << SOCK_FILTER_LOCKED) | \
 335                              (1UL << SOCK_TSTAMP_NEW))
 336 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 337  * clc socket (since smc is not called for these options from net/core)
 338  */
 339 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 340 {
 341         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 342 }
 343
 344 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 345                              (1UL << SOCK_KEEPOPEN) | \
 346                              (1UL << SOCK_LINGER) | \
 347                              (1UL << SOCK_DBG))
 348 /* copy only settings and flags relevant for smc from clc to smc socket */
 349 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 350 {
 351         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 352 }
 353
 354 /* register the new rmb on all links */
 355 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 356                              struct smc_buf_desc *rmb_desc)
 357 {
 358         struct smc_link_group *lgr = link->lgr;
 359         int i, rc = 0;
 360
 361         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 362         if (rc)
 363                 return rc;
 364         /* protect against parallel smc_llc_cli_rkey_exchange() and
 365          * parallel smcr_link_reg_rmb()
 366          */
 367         mutex_lock(&lgr->llc_conf_mutex);
 368         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 369                 if (!smc_link_active(&lgr->lnk[i]))
 370                         continue;
 371                 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
 372                 if (rc)
 373                         goto out;
 374         }
 375
 376         /* exchange confirm_rkey msg with peer */
 377         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 378         if (rc) {
 379                 rc = -EFAULT;
 380                 goto out;
 381         }
 382         rmb_desc->is_conf_rkey = true;
 383 out:
 384         mutex_unlock(&lgr->llc_conf_mutex);
 385         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 386         return rc;
 387 }
 388
 389 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 390 {
 391         struct smc_link *link = smc->conn.lnk;
 392         struct smc_llc_qentry *qentry;
 393         int rc;
 394
 395         /* receive CONFIRM LINK request from server over RoCE fabric */
 396         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 397                               SMC_LLC_CONFIRM_LINK);
 398         if (!qentry) {
 399                 struct smc_clc_msg_decline dclc;
 400
 401                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 402                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 403                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 404         }
 405         smc_llc_save_peer_uid(qentry);
 406         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 407         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 408         if (rc)
 409                 return SMC_CLC_DECL_RMBE_EC;
 410
 411         rc = smc_ib_modify_qp_rts(link);
 412         if (rc)
 413                 return SMC_CLC_DECL_ERR_RDYLNK;
 414
 415         smc_wr_remember_qp_attr(link);
 416
 417         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
 418                 return SMC_CLC_DECL_ERR_REGRMB;
 419
 420         /* confirm_rkey is implicit on 1st contact */
 421         smc->conn.rmb_desc->is_conf_rkey = true;
 422
 423         /* send CONFIRM LINK response over RoCE fabric */
 424         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 425         if (rc < 0)
 426                 return SMC_CLC_DECL_TIMEOUT_CL;
 427
 428         smc_llc_link_active(link);
 429         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 430
 431         /* optional 2nd link, receive ADD LINK request from server */
 432         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 433                               SMC_LLC_ADD_LINK);
 434         if (!qentry) {
 435                 struct smc_clc_msg_decline dclc;
 436
 437                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 438                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 439                 if (rc == -EAGAIN)
 440                         rc = 0; /* no DECLINE received, go with one link */
 441                 return rc;
 442         }
 443         smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 444         smc_llc_cli_add_link(link, qentry);
 445         return 0;
 446 }
 447
 448 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 449                                      struct smc_clc_msg_accept_confirm *clc)
 450 {
 451         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 452
 453         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 454         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 455         smc->conn.peer_rmbe_size = bufsize;
 456         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 457         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 458 }
 459
 460 static bool smc_isascii(char *hostname)
 461 {
 462         int i;
 463
 464         for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
 465                 if (!isascii(hostname[i]))
 466                         return false;
 467         return true;
 468 }
 469
 470 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 471                                      struct smc_clc_msg_accept_confirm *clc)
 472 {
 473         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 474
 475         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 476         smc->conn.peer_token = clc->d0.token;
 477         /* msg header takes up space in the buffer */
 478         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 479         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 480         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 481         if (clc->hdr.version > SMC_V1 &&
 482             (clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK)) {
 483                 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
 484                         (struct smc_clc_msg_accept_confirm_v2 *)clc;
 485                 struct smc_clc_first_contact_ext *fce =
 486                         (struct smc_clc_first_contact_ext *)
 487                                 (((u8 *)clc_v2) + sizeof(*clc_v2));
 488
 489                 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->eid,
 490                        SMC_MAX_EID_LEN);
 491                 smc->conn.lgr->peer_os = fce->os_type;
 492                 smc->conn.lgr->peer_smc_release = fce->release;
 493                 if (smc_isascii(fce->hostname))
 494                         memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
 495                                SMC_MAX_HOSTNAME_LEN);
 496         }
 497 }
 498
 499 static void smc_conn_save_peer_info(struct smc_sock *smc,
 500                                     struct smc_clc_msg_accept_confirm *clc)
 501 {
 502         if (smc->conn.lgr->is_smcd)
 503                 smcd_conn_save_peer_info(smc, clc);
 504         else
 505                 smcr_conn_save_peer_info(smc, clc);
 506 }
 507
 508 static void smc_link_save_peer_info(struct smc_link *link,
 509                                     struct smc_clc_msg_accept_confirm *clc)
 510 {
 511         link->peer_qpn = ntoh24(clc->r0.qpn);
 512         memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
 513         memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
 514         link->peer_psn = ntoh24(clc->r0.psn);
 515         link->peer_mtu = clc->r0.qp_mtu;
 516 }
 517
 518 static void smc_stat_inc_fback_rsn_cnt(struct smc_sock *smc,
 519                                        struct smc_stats_fback *fback_arr)
 520 {
 521         int cnt;
 522
 523         for (cnt = 0; cnt < SMC_MAX_FBACK_RSN_CNT; cnt++) {
 524                 if (fback_arr[cnt].fback_code == smc->fallback_rsn) {
 525                         fback_arr[cnt].count++;
 526                         break;
 527                 }
 528                 if (!fback_arr[cnt].fback_code) {
 529                         fback_arr[cnt].fback_code = smc->fallback_rsn;
 530                         fback_arr[cnt].count++;
 531                         break;
 532                 }
 533         }
 534 }
 535
 536 static void smc_stat_fallback(struct smc_sock *smc)
 537 {
 538         struct net *net = sock_net(&smc->sk);
 539
 540         mutex_lock(&net->smc.mutex_fback_rsn);
 541         if (smc->listen_smc) {
 542                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->srv);
 543                 net->smc.fback_rsn->srv_fback_cnt++;
 544         } else {
 545                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->clnt);
 546                 net->smc.fback_rsn->clnt_fback_cnt++;
 547         }
 548         mutex_unlock(&net->smc.mutex_fback_rsn);
 549 }
 550
 551 /* must be called under rcu read lock */
 552 static void smc_fback_wakeup_waitqueue(struct smc_sock *smc, void *key)
 553 {
 554         struct socket_wq *wq;
 555         __poll_t flags;
 556
 557         wq = rcu_dereference(smc->sk.sk_wq);
 558         if (!skwq_has_sleeper(wq))
 559                 return;
 560
 561         /* wake up smc sk->sk_wq */
 562         if (!key) {
 563                 /* sk_state_change */
 564                 wake_up_interruptible_all(&wq->wait);
 565         } else {
 566                 flags = key_to_poll(key);
 567                 if (flags & (EPOLLIN | EPOLLOUT))
 568                         /* sk_data_ready or sk_write_space */
 569                         wake_up_interruptible_sync_poll(&wq->wait, flags);
 570                 else if (flags & EPOLLERR)
 571                         /* sk_error_report */
 572                         wake_up_interruptible_poll(&wq->wait, flags);
 573         }
 574 }
 575
 576 static int smc_fback_mark_woken(wait_queue_entry_t *wait,
 577                                 unsigned int mode, int sync, void *key)
 578 {
 579         struct smc_mark_woken *mark =
 580                 container_of(wait, struct smc_mark_woken, wait_entry);
 581
 582         mark->woken = true;
 583         mark->key = key;
 584         return 0;
 585 }
 586
 587 static void smc_fback_forward_wakeup(struct smc_sock *smc, struct sock *clcsk,
 588                                      void (*clcsock_callback)(struct sock *sk))
 589 {
 590         struct smc_mark_woken mark = { .woken = false };
 591         struct socket_wq *wq;
 592
 593         init_waitqueue_func_entry(&mark.wait_entry,
 594                                   smc_fback_mark_woken);
 595         rcu_read_lock();
 596         wq = rcu_dereference(clcsk->sk_wq);
 597         if (!wq)
 598                 goto out;
 599         add_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 600         clcsock_callback(clcsk);
 601         remove_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 602
 603         if (mark.woken)
 604                 smc_fback_wakeup_waitqueue(smc, mark.key);
 605 out:
 606         rcu_read_unlock();
 607 }
 608
 609 static void smc_fback_state_change(struct sock *clcsk)
 610 {
 611         struct smc_sock *smc =
 612                 smc_clcsock_user_data(clcsk);
 613
 614         if (!smc)
 615                 return;
 616         smc_fback_forward_wakeup(smc, clcsk, smc->clcsk_state_change);
 617 }
 618
 619 static void smc_fback_data_ready(struct sock *clcsk)
 620 {
 621         struct smc_sock *smc =
 622                 smc_clcsock_user_data(clcsk);
 623
 624         if (!smc)
 625                 return;
 626         smc_fback_forward_wakeup(smc, clcsk, smc->clcsk_data_ready);
 627 }
 628
 629 static void smc_fback_write_space(struct sock *clcsk)
 630 {
 631         struct smc_sock *smc =
 632                 smc_clcsock_user_data(clcsk);
 633
 634         if (!smc)
 635                 return;
 636         smc_fback_forward_wakeup(smc, clcsk, smc->clcsk_write_space);
 637 }
 638
 639 static void smc_fback_error_report(struct sock *clcsk)
 640 {
 641         struct smc_sock *smc =
 642                 smc_clcsock_user_data(clcsk);
 643
 644         if (!smc)
 645                 return;
 646         smc_fback_forward_wakeup(smc, clcsk, smc->clcsk_error_report);
 647 }
 648
 649 static int smc_switch_to_fallback(struct smc_sock *smc, int reason_code)
 650 {
 651         struct sock *clcsk;
 652         int rc = 0;
 653
 654         mutex_lock(&smc->clcsock_release_lock);
 655         if (!smc->clcsock) {
 656                 rc = -EBADF;
 657                 goto out;
 658         }
 659         clcsk = smc->clcsock->sk;
 660
 661         if (smc->use_fallback)
 662                 goto out;
 663         smc->use_fallback = true;
 664         smc->fallback_rsn = reason_code;
 665         smc_stat_fallback(smc);
 666         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 667                 smc->clcsock->file = smc->sk.sk_socket->file;
 668                 smc->clcsock->file->private_data = smc->clcsock;
 669                 smc->clcsock->wq.fasync_list =
 670                         smc->sk.sk_socket->wq.fasync_list;
 671
 672                 /* There might be some wait entries remaining
 673                  * in smc sk->sk_wq and they should be woken up
 674                  * as clcsock's wait queue is woken up.
 675                  */
 676                 smc->clcsk_state_change = clcsk->sk_state_change;
 677                 smc->clcsk_data_ready = clcsk->sk_data_ready;
 678                 smc->clcsk_write_space = clcsk->sk_write_space;
 679                 smc->clcsk_error_report = clcsk->sk_error_report;
 680
 681                 clcsk->sk_state_change = smc_fback_state_change;
 682                 clcsk->sk_data_ready = smc_fback_data_ready;
 683                 clcsk->sk_write_space = smc_fback_write_space;
 684                 clcsk->sk_error_report = smc_fback_error_report;
 685
 686                 smc->clcsock->sk->sk_user_data =
 687                         (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
 688         }
 689 out:
 690         mutex_unlock(&smc->clcsock_release_lock);
 691         return rc;
 692 }
 693
 694 /* fall back during connect */
 695 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 696 {
 697         struct net *net = sock_net(&smc->sk);
 698         int rc = 0;
 699
 700         rc = smc_switch_to_fallback(smc, reason_code);
 701         if (rc) { /* fallback fails */
 702                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 703                 if (smc->sk.sk_state == SMC_INIT)
 704                         sock_put(&smc->sk); /* passive closing */
 705                 return rc;
 706         }
 707         smc_copy_sock_settings_to_clc(smc);
 708         smc->connect_nonblock = 0;
 709         if (smc->sk.sk_state == SMC_INIT)
 710                 smc->sk.sk_state = SMC_ACTIVE;
 711         return 0;
 712 }
 713
 714 /* decline and fall back during connect */
 715 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
 716                                         u8 version)
 717 {
 718         struct net *net = sock_net(&smc->sk);
 719         int rc;
 720
 721         if (reason_code < 0) { /* error, fallback is not possible */
 722                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 723                 if (smc->sk.sk_state == SMC_INIT)
 724                         sock_put(&smc->sk); /* passive closing */
 725                 return reason_code;
 726         }
 727         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 728                 rc = smc_clc_send_decline(smc, reason_code, version);
 729                 if (rc < 0) {
 730                         this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 731                         if (smc->sk.sk_state == SMC_INIT)
 732                                 sock_put(&smc->sk); /* passive closing */
 733                         return rc;
 734                 }
 735         }
 736         return smc_connect_fallback(smc, reason_code);
 737 }
 738
 739 static void smc_conn_abort(struct smc_sock *smc, int local_first)
 740 {
 741         struct smc_connection *conn = &smc->conn;
 742         struct smc_link_group *lgr = conn->lgr;
 743
 744         smc_conn_free(conn);
 745         if (local_first)
 746                 smc_lgr_cleanup_early(lgr);
 747 }
 748
 749 /* check if there is a rdma device available for this connection. */
 750 /* called for connect and listen */
 751 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 752 {
 753         /* PNET table look up: search active ib_device and port
 754          * within same PNETID that also contains the ethernet device
 755          * used for the internal TCP socket
 756          */
 757         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 758         if (!ini->ib_dev)
 759                 return SMC_CLC_DECL_NOSMCRDEV;
 760         return 0;
 761 }
 762
 763 /* check if there is an ISM device available for this connection. */
 764 /* called for connect and listen */
 765 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
 766 {
 767         /* Find ISM device with same PNETID as connecting interface  */
 768         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
 769         if (!ini->ism_dev[0])
 770                 return SMC_CLC_DECL_NOSMCDDEV;
 771         else
 772                 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
 773         return 0;
 774 }
 775
 776 /* is chid unique for the ism devices that are already determined? */
 777 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
 778                                            int cnt)
 779 {
 780         int i = (!ini->ism_dev[0]) ? 1 : 0;
 781
 782         for (; i < cnt; i++)
 783                 if (ini->ism_chid[i] == chid)
 784                         return false;
 785         return true;
 786 }
 787
 788 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
 789  * PNETID matching net_device)
 790  */
 791 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
 792                                        struct smc_init_info *ini)
 793 {
 794         int rc = SMC_CLC_DECL_NOSMCDDEV;
 795         struct smcd_dev *smcd;
 796         int i = 1;
 797         u16 chid;
 798
 799         if (smcd_indicated(ini->smc_type_v1))
 800                 rc = 0;         /* already initialized for V1 */
 801         mutex_lock(&smcd_dev_list.mutex);
 802         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
 803                 if (smcd->going_away || smcd == ini->ism_dev[0])
 804                         continue;
 805                 chid = smc_ism_get_chid(smcd);
 806                 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
 807                         continue;
 808                 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
 809                     smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
 810                         ini->ism_dev[i] = smcd;
 811                         ini->ism_chid[i] = chid;
 812                         ini->is_smcd = true;
 813                         rc = 0;
 814                         i++;
 815                         if (i > SMC_MAX_ISM_DEVS)
 816                                 break;
 817                 }
 818         }
 819         mutex_unlock(&smcd_dev_list.mutex);
 820         ini->ism_offered_cnt = i - 1;
 821         if (!ini->ism_dev[0] && !ini->ism_dev[1])
 822                 ini->smcd_version = 0;
 823
 824         return rc;
 825 }
 826
 827 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 828 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 829                                       struct smc_init_info *ini)
 830 {
 831         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
 832                 return SMC_CLC_DECL_ISMVLANERR;
 833         return 0;
 834 }
 835
 836 static int smc_find_proposal_devices(struct smc_sock *smc,
 837                                      struct smc_init_info *ini)
 838 {
 839         int rc = 0;
 840
 841         /* check if there is an ism device available */
 842         if (ini->smcd_version & SMC_V1) {
 843                 if (smc_find_ism_device(smc, ini) ||
 844                     smc_connect_ism_vlan_setup(smc, ini)) {
 845                         if (ini->smc_type_v1 == SMC_TYPE_B)
 846                                 ini->smc_type_v1 = SMC_TYPE_R;
 847                         else
 848                                 ini->smc_type_v1 = SMC_TYPE_N;
 849                 } /* else ISM V1 is supported for this connection */
 850                 if (smc_find_rdma_device(smc, ini)) {
 851                         if (ini->smc_type_v1 == SMC_TYPE_B)
 852                                 ini->smc_type_v1 = SMC_TYPE_D;
 853                         else
 854                                 ini->smc_type_v1 = SMC_TYPE_N;
 855                 } /* else RDMA is supported for this connection */
 856         }
 857         if (smc_ism_is_v2_capable() && smc_find_ism_v2_device_clnt(smc, ini))
 858                 ini->smc_type_v2 = SMC_TYPE_N;
 859
 860         /* if neither ISM nor RDMA are supported, fallback */
 861         if (!smcr_indicated(ini->smc_type_v1) &&
 862             ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
 863                 rc = SMC_CLC_DECL_NOSMCDEV;
 864
 865         return rc;
 866 }
 867
 868 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 869  * used, the VLAN ID will be registered again during the connection setup.
 870  */
 871 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
 872                                         struct smc_init_info *ini)
 873 {
 874         if (!smcd_indicated(ini->smc_type_v1))
 875                 return 0;
 876         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
 877                 return SMC_CLC_DECL_CNFERR;
 878         return 0;
 879 }
 880
 881 #define SMC_CLC_MAX_ACCEPT_LEN \
 882         (sizeof(struct smc_clc_msg_accept_confirm_v2) + \
 883          sizeof(struct smc_clc_first_contact_ext) + \
 884          sizeof(struct smc_clc_msg_trail))
 885
 886 /* CLC handshake during connect */
 887 static int smc_connect_clc(struct smc_sock *smc,
 888                            struct smc_clc_msg_accept_confirm_v2 *aclc2,
 889                            struct smc_init_info *ini)
 890 {
 891         int rc = 0;
 892
 893         /* do inband token exchange */
 894         rc = smc_clc_send_proposal(smc, ini);
 895         if (rc)
 896                 return rc;
 897         /* receive SMC Accept CLC message */
 898         return smc_clc_wait_msg(smc, aclc2, SMC_CLC_MAX_ACCEPT_LEN,
 899                                 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
 900 }
 901
 902 /* setup for RDMA connection of client */
 903 static int smc_connect_rdma(struct smc_sock *smc,
 904                             struct smc_clc_msg_accept_confirm *aclc,
 905                             struct smc_init_info *ini)
 906 {
 907         int i, reason_code = 0;
 908         struct smc_link *link;
 909
 910         ini->is_smcd = false;
 911         ini->ib_lcl = &aclc->r0.lcl;
 912         ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
 913         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
 914
 915         mutex_lock(&smc_client_lgr_pending);
 916         reason_code = smc_conn_create(smc, ini);
 917         if (reason_code) {
 918                 mutex_unlock(&smc_client_lgr_pending);
 919                 return reason_code;
 920         }
 921
 922         smc_conn_save_peer_info(smc, aclc);
 923
 924         if (ini->first_contact_local) {
 925                 link = smc->conn.lnk;
 926         } else {
 927                 /* set link that was assigned by server */
 928                 link = NULL;
 929                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 930                         struct smc_link *l = &smc->conn.lgr->lnk[i];
 931
 932                         if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
 933                             !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
 934                                     SMC_GID_SIZE) &&
 935                             !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
 936                                     sizeof(l->peer_mac))) {
 937                                 link = l;
 938                                 break;
 939                         }
 940                 }
 941                 if (!link) {
 942                         reason_code = SMC_CLC_DECL_NOSRVLINK;
 943                         goto connect_abort;
 944                 }
 945                 smc_switch_link_and_count(&smc->conn, link);
 946         }
 947
 948         /* create send buffer and rmb */
 949         if (smc_buf_create(smc, false)) {
 950                 reason_code = SMC_CLC_DECL_MEM;
 951                 goto connect_abort;
 952         }
 953
 954         if (ini->first_contact_local)
 955                 smc_link_save_peer_info(link, aclc);
 956
 957         if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
 958                 reason_code = SMC_CLC_DECL_ERR_RTOK;
 959                 goto connect_abort;
 960         }
 961
 962         smc_close_init(smc);
 963         smc_rx_init(smc);
 964
 965         if (ini->first_contact_local) {
 966                 if (smc_ib_ready_link(link)) {
 967                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
 968                         goto connect_abort;
 969                 }
 970         } else {
 971                 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
 972                         reason_code = SMC_CLC_DECL_ERR_REGRMB;
 973                         goto connect_abort;
 974                 }
 975         }
 976         smc_rmb_sync_sg_for_device(&smc->conn);
 977
 978         reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
 979                                            SMC_V1);
 980         if (reason_code)
 981                 goto connect_abort;
 982
 983         smc_tx_init(smc);
 984
 985         if (ini->first_contact_local) {
 986                 /* QP confirmation over RoCE fabric */
 987                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
 988                 reason_code = smcr_clnt_conf_first_link(smc);
 989                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
 990                 if (reason_code)
 991                         goto connect_abort;
 992         }
 993         mutex_unlock(&smc_client_lgr_pending);
 994
 995         smc_copy_sock_settings_to_clc(smc);
 996         smc->connect_nonblock = 0;
 997         if (smc->sk.sk_state == SMC_INIT)
 998                 smc->sk.sk_state = SMC_ACTIVE;
 999
1000         return 0;
1001 connect_abort:
1002         smc_conn_abort(smc, ini->first_contact_local);
1003         mutex_unlock(&smc_client_lgr_pending);
1004         smc->connect_nonblock = 0;
1005
1006         return reason_code;
1007 }
1008
1009 /* The server has chosen one of the proposed ISM devices for the communication.
1010  * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
1011  */
1012 static int
1013 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 *aclc,
1014                                struct smc_init_info *ini)
1015 {
1016         int i;
1017
1018         for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
1019                 if (ini->ism_chid[i] == ntohs(aclc->chid)) {
1020                         ini->ism_selected = i;
1021                         return 0;
1022                 }
1023         }
1024
1025         return -EPROTO;
1026 }
1027
1028 /* setup for ISM connection of client */
1029 static int smc_connect_ism(struct smc_sock *smc,
1030                            struct smc_clc_msg_accept_confirm *aclc,
1031                            struct smc_init_info *ini)
1032 {
1033         int rc = 0;
1034
1035         ini->is_smcd = true;
1036         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1037
1038         if (aclc->hdr.version == SMC_V2) {
1039                 struct smc_clc_msg_accept_confirm_v2 *aclc_v2 =
1040                         (struct smc_clc_msg_accept_confirm_v2 *)aclc;
1041
1042                 rc = smc_v2_determine_accepted_chid(aclc_v2, ini);
1043                 if (rc)
1044                         return rc;
1045         }
1046         ini->ism_peer_gid[ini->ism_selected] = aclc->d0.gid;
1047
1048         /* there is only one lgr role for SMC-D; use server lock */
1049         mutex_lock(&smc_server_lgr_pending);
1050         rc = smc_conn_create(smc, ini);
1051         if (rc) {
1052                 mutex_unlock(&smc_server_lgr_pending);
1053                 return rc;
1054         }
1055
1056         /* Create send and receive buffers */
1057         rc = smc_buf_create(smc, true);
1058         if (rc) {
1059                 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
1060                 goto connect_abort;
1061         }
1062
1063         smc_conn_save_peer_info(smc, aclc);
1064         smc_close_init(smc);
1065         smc_rx_init(smc);
1066         smc_tx_init(smc);
1067
1068         rc = smc_clc_send_confirm(smc, ini->first_contact_local,
1069                                   aclc->hdr.version);
1070         if (rc)
1071                 goto connect_abort;
1072         mutex_unlock(&smc_server_lgr_pending);
1073
1074         smc_copy_sock_settings_to_clc(smc);
1075         smc->connect_nonblock = 0;
1076         if (smc->sk.sk_state == SMC_INIT)
1077                 smc->sk.sk_state = SMC_ACTIVE;
1078
1079         return 0;
1080 connect_abort:
1081         smc_conn_abort(smc, ini->first_contact_local);
1082         mutex_unlock(&smc_server_lgr_pending);
1083         smc->connect_nonblock = 0;
1084
1085         return rc;
1086 }
1087
1088 /* check if received accept type and version matches a proposed one */
1089 static int smc_connect_check_aclc(struct smc_init_info *ini,
1090                                   struct smc_clc_msg_accept_confirm *aclc)
1091 {
1092         if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1093              !smcr_indicated(ini->smc_type_v1)) ||
1094             (aclc->hdr.typev1 == SMC_TYPE_D &&
1095              ((!smcd_indicated(ini->smc_type_v1) &&
1096                !smcd_indicated(ini->smc_type_v2)) ||
1097               (aclc->hdr.version == SMC_V1 &&
1098                !smcd_indicated(ini->smc_type_v1)) ||
1099               (aclc->hdr.version == SMC_V2 &&
1100                !smcd_indicated(ini->smc_type_v2)))))
1101                 return SMC_CLC_DECL_MODEUNSUPP;
1102
1103         return 0;
1104 }
1105
1106 /* perform steps before actually connecting */
1107 static int __smc_connect(struct smc_sock *smc)
1108 {
1109         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1110         struct smc_clc_msg_accept_confirm_v2 *aclc2;
1111         struct smc_clc_msg_accept_confirm *aclc;
1112         struct smc_init_info *ini = NULL;
1113         u8 *buf = NULL;
1114         int rc = 0;
1115
1116         if (smc->use_fallback)
1117                 return smc_connect_fallback(smc, smc->fallback_rsn);
1118
1119         /* if peer has not signalled SMC-capability, fall back */
1120         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
1121                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
1122
1123         /* IPSec connections opt out of SMC optimizations */
1124         if (using_ipsec(smc))
1125                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
1126                                                     version);
1127
1128         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1129         if (!ini)
1130                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
1131                                                     version);
1132
1133         ini->smcd_version = SMC_V1;
1134         ini->smcd_version |= smc_ism_is_v2_capable() ? SMC_V2 : 0;
1135         ini->smc_type_v1 = SMC_TYPE_B;
1136         ini->smc_type_v2 = smc_ism_is_v2_capable() ? SMC_TYPE_D : SMC_TYPE_N;
1137
1138         /* get vlan id from IP device */
1139         if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
1140                 ini->smcd_version &= ~SMC_V1;
1141                 ini->smc_type_v1 = SMC_TYPE_N;
1142                 if (!ini->smcd_version) {
1143                         rc = SMC_CLC_DECL_GETVLANERR;
1144                         goto fallback;
1145                 }
1146         }
1147
1148         rc = smc_find_proposal_devices(smc, ini);
1149         if (rc)
1150                 goto fallback;
1151
1152         buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
1153         if (!buf) {
1154                 rc = SMC_CLC_DECL_MEM;
1155                 goto fallback;
1156         }
1157         aclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
1158         aclc = (struct smc_clc_msg_accept_confirm *)aclc2;
1159
1160         /* perform CLC handshake */
1161         rc = smc_connect_clc(smc, aclc2, ini);
1162         if (rc)
1163                 goto vlan_cleanup;
1164
1165         /* check if smc modes and versions of CLC proposal and accept match */
1166         rc = smc_connect_check_aclc(ini, aclc);
1167         version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
1168         ini->smcd_version = version;
1169         if (rc)
1170                 goto vlan_cleanup;
1171
1172         /* depending on previous steps, connect using rdma or ism */
1173         if (aclc->hdr.typev1 == SMC_TYPE_R)
1174                 rc = smc_connect_rdma(smc, aclc, ini);
1175         else if (aclc->hdr.typev1 == SMC_TYPE_D)
1176                 rc = smc_connect_ism(smc, aclc, ini);
1177         if (rc)
1178                 goto vlan_cleanup;
1179
1180         SMC_STAT_CLNT_SUCC_INC(sock_net(smc->clcsock->sk), aclc);
1181         smc_connect_ism_vlan_cleanup(smc, ini);
1182         kfree(buf);
1183         kfree(ini);
1184         return 0;
1185
1186 vlan_cleanup:
1187         smc_connect_ism_vlan_cleanup(smc, ini);
1188         kfree(buf);
1189 fallback:
1190         kfree(ini);
1191         return smc_connect_decline_fallback(smc, rc, version);
1192 }
1193
1194 static void smc_connect_work(struct work_struct *work)
1195 {
1196         struct smc_sock *smc = container_of(work, struct smc_sock,
1197                                             connect_work);
1198         long timeo = smc->sk.sk_sndtimeo;
1199         int rc = 0;
1200
1201         if (!timeo)
1202                 timeo = MAX_SCHEDULE_TIMEOUT;
1203         lock_sock(smc->clcsock->sk);
1204         if (smc->clcsock->sk->sk_err) {
1205                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1206         } else if ((1 << smc->clcsock->sk->sk_state) &
1207                                         (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
1208                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1209                 if ((rc == -EPIPE) &&
1210                     ((1 << smc->clcsock->sk->sk_state) &
1211                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1212                         rc = 0;
1213         }
1214         release_sock(smc->clcsock->sk);
1215         lock_sock(&smc->sk);
1216         if (rc != 0 || smc->sk.sk_err) {
1217                 smc->sk.sk_state = SMC_CLOSED;
1218                 if (rc == -EPIPE || rc == -EAGAIN)
1219                         smc->sk.sk_err = EPIPE;
1220                 else if (signal_pending(current))
1221                         smc->sk.sk_err = -sock_intr_errno(timeo);
1222                 sock_put(&smc->sk); /* passive closing */
1223                 goto out;
1224         }
1225
1226         rc = __smc_connect(smc);
1227         if (rc < 0)
1228                 smc->sk.sk_err = -rc;
1229
1230 out:
1231         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1232                 if (smc->sk.sk_err) {
1233                         smc->sk.sk_state_change(&smc->sk);
1234                 } else { /* allow polling before and after fallback decision */
1235                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1236                         smc->sk.sk_write_space(&smc->sk);
1237                 }
1238         }
1239         release_sock(&smc->sk);
1240 }
1241
1242 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1243                        int alen, int flags)
1244 {
1245         struct sock *sk = sock->sk;
1246         struct smc_sock *smc;
1247         int rc = -EINVAL;
1248
1249         smc = smc_sk(sk);
1250
1251         /* separate smc parameter checking to be safe */
1252         if (alen < sizeof(addr->sa_family))
1253                 goto out_err;
1254         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1255                 goto out_err;
1256
1257         lock_sock(sk);
1258         switch (sk->sk_state) {
1259         default:
1260                 goto out;
1261         case SMC_ACTIVE:
1262                 rc = -EISCONN;
1263                 goto out;
1264         case SMC_INIT:
1265                 break;
1266         }
1267
1268         smc_copy_sock_settings_to_clc(smc);
1269         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1270         if (smc->connect_nonblock) {
1271                 rc = -EALREADY;
1272                 goto out;
1273         }
1274         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1275         if (rc && rc != -EINPROGRESS)
1276                 goto out;
1277
1278         sock_hold(&smc->sk); /* sock put in passive closing */
1279         if (smc->use_fallback)
1280                 goto out;
1281         if (flags & O_NONBLOCK) {
1282                 if (queue_work(smc_hs_wq, &smc->connect_work))
1283                         smc->connect_nonblock = 1;
1284                 rc = -EINPROGRESS;
1285         } else {
1286                 rc = __smc_connect(smc);
1287                 if (rc < 0)
1288                         goto out;
1289                 else
1290                         rc = 0; /* success cases including fallback */
1291         }
1292
1293 out:
1294         release_sock(sk);
1295 out_err:
1296         return rc;
1297 }
1298
1299 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1300 {
1301         struct socket *new_clcsock = NULL;
1302         struct sock *lsk = &lsmc->sk;
1303         struct sock *new_sk;
1304         int rc = -EINVAL;
1305
1306         release_sock(lsk);
1307         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1308         if (!new_sk) {
1309                 rc = -ENOMEM;
1310                 lsk->sk_err = ENOMEM;
1311                 *new_smc = NULL;
1312                 lock_sock(lsk);
1313                 goto out;
1314         }
1315         *new_smc = smc_sk(new_sk);
1316
1317         mutex_lock(&lsmc->clcsock_release_lock);
1318         if (lsmc->clcsock)
1319                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1320         mutex_unlock(&lsmc->clcsock_release_lock);
1321         lock_sock(lsk);
1322         if  (rc < 0 && rc != -EAGAIN)
1323                 lsk->sk_err = -rc;
1324         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1325                 new_sk->sk_prot->unhash(new_sk);
1326                 if (new_clcsock)
1327                         sock_release(new_clcsock);
1328                 new_sk->sk_state = SMC_CLOSED;
1329                 sock_set_flag(new_sk, SOCK_DEAD);
1330                 sock_put(new_sk); /* final */
1331                 *new_smc = NULL;
1332                 goto out;
1333         }
1334
1335         /* new clcsock has inherited the smc listen-specific sk_data_ready
1336          * function; switch it back to the original sk_data_ready function
1337          */
1338         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1339         (*new_smc)->clcsock = new_clcsock;
1340 out:
1341         return rc;
1342 }
1343
1344 /* add a just created sock to the accept queue of the listen sock as
1345  * candidate for a following socket accept call from user space
1346  */
1347 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1348 {
1349         struct smc_sock *par = smc_sk(parent);
1350
1351         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1352         spin_lock(&par->accept_q_lock);
1353         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1354         spin_unlock(&par->accept_q_lock);
1355         sk_acceptq_added(parent);
1356 }
1357
1358 /* remove a socket from the accept queue of its parental listening socket */
1359 static void smc_accept_unlink(struct sock *sk)
1360 {
1361         struct smc_sock *par = smc_sk(sk)->listen_smc;
1362
1363         spin_lock(&par->accept_q_lock);
1364         list_del_init(&smc_sk(sk)->accept_q);
1365         spin_unlock(&par->accept_q_lock);
1366         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1367         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1368 }
1369
1370 /* remove a sock from the accept queue to bind it to a new socket created
1371  * for a socket accept call from user space
1372  */
1373 struct sock *smc_accept_dequeue(struct sock *parent,
1374                                 struct socket *new_sock)
1375 {
1376         struct smc_sock *isk, *n;
1377         struct sock *new_sk;
1378
1379         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1380                 new_sk = (struct sock *)isk;
1381
1382                 smc_accept_unlink(new_sk);
1383                 if (new_sk->sk_state == SMC_CLOSED) {
1384                         new_sk->sk_prot->unhash(new_sk);
1385                         if (isk->clcsock) {
1386                                 sock_release(isk->clcsock);
1387                                 isk->clcsock = NULL;
1388                         }
1389                         sock_put(new_sk); /* final */
1390                         continue;
1391                 }
1392                 if (new_sock) {
1393                         sock_graft(new_sk, new_sock);
1394                         if (isk->use_fallback) {
1395                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1396                                 isk->clcsock->file->private_data = isk->clcsock;
1397                         }
1398                 }
1399                 return new_sk;
1400         }
1401         return NULL;
1402 }
1403
1404 /* clean up for a created but never accepted sock */
1405 void smc_close_non_accepted(struct sock *sk)
1406 {
1407         struct smc_sock *smc = smc_sk(sk);
1408
1409         sock_hold(sk); /* sock_put below */
1410         lock_sock(sk);
1411         if (!sk->sk_lingertime)
1412                 /* wait for peer closing */
1413                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1414         __smc_release(smc);
1415         release_sock(sk);
1416         sock_put(sk); /* sock_hold above */
1417         sock_put(sk); /* final sock_put */
1418 }
1419
1420 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1421 {
1422         struct smc_link *link = smc->conn.lnk;
1423         struct smc_llc_qentry *qentry;
1424         int rc;
1425
1426         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1427                 return SMC_CLC_DECL_ERR_REGRMB;
1428
1429         /* send CONFIRM LINK request to client over the RoCE fabric */
1430         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1431         if (rc < 0)
1432                 return SMC_CLC_DECL_TIMEOUT_CL;
1433
1434         /* receive CONFIRM LINK response from client over the RoCE fabric */
1435         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1436                               SMC_LLC_CONFIRM_LINK);
1437         if (!qentry) {
1438                 struct smc_clc_msg_decline dclc;
1439
1440                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1441                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1442                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1443         }
1444         smc_llc_save_peer_uid(qentry);
1445         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1446         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1447         if (rc)
1448                 return SMC_CLC_DECL_RMBE_EC;
1449
1450         /* confirm_rkey is implicit on 1st contact */
1451         smc->conn.rmb_desc->is_conf_rkey = true;
1452
1453         smc_llc_link_active(link);
1454         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1455
1456         /* initial contact - try to establish second link */
1457         smc_llc_srv_add_link(link);
1458         return 0;
1459 }
1460
1461 /* listen worker: finish */
1462 static void smc_listen_out(struct smc_sock *new_smc)
1463 {
1464         struct smc_sock *lsmc = new_smc->listen_smc;
1465         struct sock *newsmcsk = &new_smc->sk;
1466
1467         if (lsmc->sk.sk_state == SMC_LISTEN) {
1468                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1469                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1470                 release_sock(&lsmc->sk);
1471         } else { /* no longer listening */
1472                 smc_close_non_accepted(newsmcsk);
1473         }
1474
1475         /* Wake up accept */
1476         lsmc->sk.sk_data_ready(&lsmc->sk);
1477         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1478 }
1479
1480 /* listen worker: finish in state connected */
1481 static void smc_listen_out_connected(struct smc_sock *new_smc)
1482 {
1483         struct sock *newsmcsk = &new_smc->sk;
1484
1485         sk_refcnt_debug_inc(newsmcsk);
1486         if (newsmcsk->sk_state == SMC_INIT)
1487                 newsmcsk->sk_state = SMC_ACTIVE;
1488
1489         smc_listen_out(new_smc);
1490 }
1491
1492 /* listen worker: finish in error state */
1493 static void smc_listen_out_err(struct smc_sock *new_smc)
1494 {
1495         struct sock *newsmcsk = &new_smc->sk;
1496         struct net *net = sock_net(newsmcsk);
1497
1498         this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt);
1499         if (newsmcsk->sk_state == SMC_INIT)
1500                 sock_put(&new_smc->sk); /* passive closing */
1501         newsmcsk->sk_state = SMC_CLOSED;
1502
1503         smc_listen_out(new_smc);
1504 }
1505
1506 /* listen worker: decline and fall back if possible */
1507 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1508                                int local_first, u8 version)
1509 {
1510         /* RDMA setup failed, switch back to TCP */
1511         smc_conn_abort(new_smc, local_first);
1512         if (reason_code < 0 ||
1513             smc_switch_to_fallback(new_smc, reason_code)) {
1514                 /* error, no fallback possible */
1515                 smc_listen_out_err(new_smc);
1516                 return;
1517         }
1518         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1519                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1520                         smc_listen_out_err(new_smc);
1521                         return;
1522                 }
1523         }
1524         smc_listen_out_connected(new_smc);
1525 }
1526
1527 /* listen worker: version checking */
1528 static int smc_listen_v2_check(struct smc_sock *new_smc,
1529                                struct smc_clc_msg_proposal *pclc,
1530                                struct smc_init_info *ini)
1531 {
1532         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1533         struct smc_clc_v2_extension *pclc_v2_ext;
1534         int rc = SMC_CLC_DECL_PEERNOSMC;
1535
1536         ini->smc_type_v1 = pclc->hdr.typev1;
1537         ini->smc_type_v2 = pclc->hdr.typev2;
1538         ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1539         if (pclc->hdr.version > SMC_V1)
1540                 ini->smcd_version |=
1541                                 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1542         if (!(ini->smcd_version & SMC_V2)) {
1543                 rc = SMC_CLC_DECL_PEERNOSMC;
1544                 goto out;
1545         }
1546         if (!smc_ism_is_v2_capable()) {
1547                 ini->smcd_version &= ~SMC_V2;
1548                 rc = SMC_CLC_DECL_NOISM2SUPP;
1549                 goto out;
1550         }
1551         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1552         if (!pclc_v2_ext) {
1553                 ini->smcd_version &= ~SMC_V2;
1554                 rc = SMC_CLC_DECL_NOV2EXT;
1555                 goto out;
1556         }
1557         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1558         if (!pclc_smcd_v2_ext) {
1559                 ini->smcd_version &= ~SMC_V2;
1560                 rc = SMC_CLC_DECL_NOV2DEXT;
1561         }
1562
1563 out:
1564         if (!ini->smcd_version)
1565                 return rc;
1566
1567         return 0;
1568 }
1569
1570 /* listen worker: check prefixes */
1571 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1572                                  struct smc_clc_msg_proposal *pclc)
1573 {
1574         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1575         struct socket *newclcsock = new_smc->clcsock;
1576
1577         if (pclc->hdr.typev1 == SMC_TYPE_N)
1578                 return 0;
1579         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1580         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1581                 return SMC_CLC_DECL_DIFFPREFIX;
1582
1583         return 0;
1584 }
1585
1586 /* listen worker: initialize connection and buffers */
1587 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1588                                 struct smc_init_info *ini)
1589 {
1590         int rc;
1591
1592         /* allocate connection / link group */
1593         rc = smc_conn_create(new_smc, ini);
1594         if (rc)
1595                 return rc;
1596
1597         /* create send buffer and rmb */
1598         if (smc_buf_create(new_smc, false))
1599                 return SMC_CLC_DECL_MEM;
1600
1601         return 0;
1602 }
1603
1604 /* listen worker: initialize connection and buffers for SMC-D */
1605 static int smc_listen_ism_init(struct smc_sock *new_smc,
1606                                struct smc_init_info *ini)
1607 {
1608         int rc;
1609
1610         rc = smc_conn_create(new_smc, ini);
1611         if (rc)
1612                 return rc;
1613
1614         /* Create send and receive buffers */
1615         rc = smc_buf_create(new_smc, true);
1616         if (rc) {
1617                 smc_conn_abort(new_smc, ini->first_contact_local);
1618                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1619                                          SMC_CLC_DECL_MEM;
1620         }
1621
1622         return 0;
1623 }
1624
1625 static bool smc_is_already_selected(struct smcd_dev *smcd,
1626                                     struct smc_init_info *ini,
1627                                     int matches)
1628 {
1629         int i;
1630
1631         for (i = 0; i < matches; i++)
1632                 if (smcd == ini->ism_dev[i])
1633                         return true;
1634
1635         return false;
1636 }
1637
1638 /* check for ISM devices matching proposed ISM devices */
1639 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1640                                    u16 proposed_chid, u64 proposed_gid,
1641                                    unsigned int *matches)
1642 {
1643         struct smcd_dev *smcd;
1644
1645         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1646                 if (smcd->going_away)
1647                         continue;
1648                 if (smc_is_already_selected(smcd, ini, *matches))
1649                         continue;
1650                 if (smc_ism_get_chid(smcd) == proposed_chid &&
1651                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1652                         ini->ism_peer_gid[*matches] = proposed_gid;
1653                         ini->ism_dev[*matches] = smcd;
1654                         (*matches)++;
1655                         break;
1656                 }
1657         }
1658 }
1659
1660 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
1661 {
1662         if (!ini->rc)
1663                 ini->rc = rc;
1664 }
1665
1666 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1667                                         struct smc_clc_msg_proposal *pclc,
1668                                         struct smc_init_info *ini)
1669 {
1670         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1671         struct smc_clc_v2_extension *smc_v2_ext;
1672         struct smc_clc_msg_smcd *pclc_smcd;
1673         unsigned int matches = 0;
1674         u8 smcd_version;
1675         u8 *eid = NULL;
1676         int i, rc;
1677
1678         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1679                 goto not_found;
1680
1681         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1682         smc_v2_ext = smc_get_clc_v2_ext(pclc);
1683         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1684         if (!smcd_v2_ext ||
1685             !smc_v2_ext->hdr.flag.seid) { /* no system EID support for SMCD */
1686                 smc_find_ism_store_rc(SMC_CLC_DECL_NOSEID, ini);
1687                 goto not_found;
1688         }
1689
1690         mutex_lock(&smcd_dev_list.mutex);
1691         if (pclc_smcd->ism.chid)
1692                 /* check for ISM device matching proposed native ISM device */
1693                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1694                                        ntohll(pclc_smcd->ism.gid), &matches);
1695         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1696                 /* check for ISM devices matching proposed non-native ISM
1697                  * devices
1698                  */
1699                 smc_check_ism_v2_match(ini,
1700                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1701                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1702                                        &matches);
1703         }
1704         mutex_unlock(&smcd_dev_list.mutex);
1705
1706         if (ini->ism_dev[0]) {
1707                 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1708                 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1709                         goto not_found;
1710         } else {
1711                 goto not_found;
1712         }
1713
1714         /* separate - outside the smcd_dev_list.lock */
1715         smcd_version = ini->smcd_version;
1716         for (i = 0; i < matches; i++) {
1717                 ini->smcd_version = SMC_V2;
1718                 ini->is_smcd = true;
1719                 ini->ism_selected = i;
1720                 rc = smc_listen_ism_init(new_smc, ini);
1721                 if (rc) {
1722                         smc_find_ism_store_rc(rc, ini);
1723                         /* try next active ISM device */
1724                         continue;
1725                 }
1726                 return; /* matching and usable V2 ISM device found */
1727         }
1728         /* no V2 ISM device could be initialized */
1729         ini->smcd_version = smcd_version;       /* restore original value */
1730
1731 not_found:
1732         ini->smcd_version &= ~SMC_V2;
1733         ini->ism_dev[0] = NULL;
1734         ini->is_smcd = false;
1735 }
1736
1737 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1738                                         struct smc_clc_msg_proposal *pclc,
1739                                         struct smc_init_info *ini)
1740 {
1741         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1742         int rc = 0;
1743
1744         /* check if ISM V1 is available */
1745         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1746                 goto not_found;
1747         ini->is_smcd = true; /* prepare ISM check */
1748         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1749         rc = smc_find_ism_device(new_smc, ini);
1750         if (rc)
1751                 goto not_found;
1752         ini->ism_selected = 0;
1753         rc = smc_listen_ism_init(new_smc, ini);
1754         if (!rc)
1755                 return;         /* V1 ISM device found */
1756
1757 not_found:
1758         smc_find_ism_store_rc(rc, ini);
1759         ini->ism_dev[0] = NULL;
1760         ini->is_smcd = false;
1761 }
1762
1763 /* listen worker: register buffers */
1764 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1765 {
1766         struct smc_connection *conn = &new_smc->conn;
1767
1768         if (!local_first) {
1769                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1770                         return SMC_CLC_DECL_ERR_REGRMB;
1771         }
1772         smc_rmb_sync_sg_for_device(&new_smc->conn);
1773
1774         return 0;
1775 }
1776
1777 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1778                                         struct smc_clc_msg_proposal *pclc,
1779                                         struct smc_init_info *ini)
1780 {
1781         int rc;
1782
1783         if (!smcr_indicated(ini->smc_type_v1))
1784                 return SMC_CLC_DECL_NOSMCDEV;
1785
1786         /* prepare RDMA check */
1787         ini->ib_lcl = &pclc->lcl;
1788         rc = smc_find_rdma_device(new_smc, ini);
1789         if (rc) {
1790                 /* no RDMA device found */
1791                 if (ini->smc_type_v1 == SMC_TYPE_B)
1792                         /* neither ISM nor RDMA device found */
1793                         rc = SMC_CLC_DECL_NOSMCDEV;
1794                 return rc;
1795         }
1796         rc = smc_listen_rdma_init(new_smc, ini);
1797         if (rc)
1798                 return rc;
1799         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1800 }
1801
1802 /* determine the local device matching to proposal */
1803 static int smc_listen_find_device(struct smc_sock *new_smc,
1804                                   struct smc_clc_msg_proposal *pclc,
1805                                   struct smc_init_info *ini)
1806 {
1807         int rc;
1808
1809         /* check for ISM device matching V2 proposed device */
1810         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1811         if (ini->ism_dev[0])
1812                 return 0;
1813
1814         if (!(ini->smcd_version & SMC_V1))
1815                 return ini->rc ?: SMC_CLC_DECL_NOSMCD2DEV;
1816
1817         /* check for matching IP prefix and subnet length */
1818         rc = smc_listen_prfx_check(new_smc, pclc);
1819         if (rc)
1820                 return ini->rc ?: rc;
1821
1822         /* get vlan id from IP device */
1823         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1824                 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
1825
1826         /* check for ISM device matching V1 proposed device */
1827         smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1828         if (ini->ism_dev[0])
1829                 return 0;
1830
1831         if (pclc->hdr.typev1 == SMC_TYPE_D)
1832                 /* skip RDMA and decline */
1833                 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
1834
1835         /* check if RDMA is available */
1836         rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1837         smc_find_ism_store_rc(rc, ini);
1838
1839         return (!rc) ? 0 : ini->rc;
1840 }
1841
1842 /* listen worker: finish RDMA setup */
1843 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1844                                   struct smc_clc_msg_accept_confirm *cclc,
1845                                   bool local_first)
1846 {
1847         struct smc_link *link = new_smc->conn.lnk;
1848         int reason_code = 0;
1849
1850         if (local_first)
1851                 smc_link_save_peer_info(link, cclc);
1852
1853         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1854                 return SMC_CLC_DECL_ERR_RTOK;
1855
1856         if (local_first) {
1857                 if (smc_ib_ready_link(link))
1858                         return SMC_CLC_DECL_ERR_RDYLNK;
1859                 /* QP confirmation over RoCE fabric */
1860                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1861                 reason_code = smcr_serv_conf_first_link(new_smc);
1862                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1863         }
1864         return reason_code;
1865 }
1866
1867 /* setup for connection of server */
1868 static void smc_listen_work(struct work_struct *work)
1869 {
1870         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1871                                                 smc_listen_work);
1872         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1873         struct socket *newclcsock = new_smc->clcsock;
1874         struct smc_clc_msg_accept_confirm *cclc;
1875         struct smc_clc_msg_proposal_area *buf;
1876         struct smc_clc_msg_proposal *pclc;
1877         struct smc_init_info *ini = NULL;
1878         int rc = 0;
1879
1880         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1881                 return smc_listen_out_err(new_smc);
1882
1883         if (new_smc->use_fallback) {
1884                 smc_listen_out_connected(new_smc);
1885                 return;
1886         }
1887
1888         /* check if peer is smc capable */
1889         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1890                 rc = smc_switch_to_fallback(new_smc, SMC_CLC_DECL_PEERNOSMC);
1891                 if (rc)
1892                         smc_listen_out_err(new_smc);
1893                 else
1894                         smc_listen_out_connected(new_smc);
1895                 return;
1896         }
1897
1898         /* do inband token exchange -
1899          * wait for and receive SMC Proposal CLC message
1900          */
1901         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1902         if (!buf) {
1903                 rc = SMC_CLC_DECL_MEM;
1904                 goto out_decl;
1905         }
1906         pclc = (struct smc_clc_msg_proposal *)buf;
1907         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1908                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1909         if (rc)
1910                 goto out_decl;
1911         version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1912
1913         /* IPSec connections opt out of SMC optimizations */
1914         if (using_ipsec(new_smc)) {
1915                 rc = SMC_CLC_DECL_IPSEC;
1916                 goto out_decl;
1917         }
1918
1919         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1920         if (!ini) {
1921                 rc = SMC_CLC_DECL_MEM;
1922                 goto out_decl;
1923         }
1924
1925         /* initial version checking */
1926         rc = smc_listen_v2_check(new_smc, pclc, ini);
1927         if (rc)
1928                 goto out_decl;
1929
1930         mutex_lock(&smc_server_lgr_pending);
1931         smc_close_init(new_smc);
1932         smc_rx_init(new_smc);
1933         smc_tx_init(new_smc);
1934
1935         /* determine ISM or RoCE device used for connection */
1936         rc = smc_listen_find_device(new_smc, pclc, ini);
1937         if (rc)
1938                 goto out_unlock;
1939
1940         /* send SMC Accept CLC message */
1941         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1942                                  ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1943         if (rc)
1944                 goto out_unlock;
1945
1946         /* SMC-D does not need this lock any more */
1947         if (ini->is_smcd)
1948                 mutex_unlock(&smc_server_lgr_pending);
1949
1950         /* receive SMC Confirm CLC message */
1951         memset(buf, 0, sizeof(*buf));
1952         cclc = (struct smc_clc_msg_accept_confirm *)buf;
1953         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1954                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1955         if (rc) {
1956                 if (!ini->is_smcd)
1957                         goto out_unlock;
1958                 goto out_decl;
1959         }
1960
1961         /* finish worker */
1962         if (!ini->is_smcd) {
1963                 rc = smc_listen_rdma_finish(new_smc, cclc,
1964                                             ini->first_contact_local);
1965                 if (rc)
1966                         goto out_unlock;
1967                 mutex_unlock(&smc_server_lgr_pending);
1968         }
1969         smc_conn_save_peer_info(new_smc, cclc);
1970         smc_listen_out_connected(new_smc);
1971         SMC_STAT_SERV_SUCC_INC(sock_net(newclcsock->sk), ini);
1972         goto out_free;
1973
1974 out_unlock:
1975         mutex_unlock(&smc_server_lgr_pending);
1976 out_decl:
1977         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1978                            version);
1979 out_free:
1980         kfree(ini);
1981         kfree(buf);
1982 }
1983
1984 static void smc_tcp_listen_work(struct work_struct *work)
1985 {
1986         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1987                                              tcp_listen_work);
1988         struct sock *lsk = &lsmc->sk;
1989         struct smc_sock *new_smc;
1990         int rc = 0;
1991
1992         lock_sock(lsk);
1993         while (lsk->sk_state == SMC_LISTEN) {
1994                 rc = smc_clcsock_accept(lsmc, &new_smc);
1995                 if (rc) /* clcsock accept queue empty or error */
1996                         goto out;
1997                 if (!new_smc)
1998                         continue;
1999
2000                 new_smc->listen_smc = lsmc;
2001                 new_smc->use_fallback = lsmc->use_fallback;
2002                 new_smc->fallback_rsn = lsmc->fallback_rsn;
2003                 sock_hold(lsk); /* sock_put in smc_listen_work */
2004                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
2005                 smc_copy_sock_settings_to_smc(new_smc);
2006                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
2007                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
2008                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
2009                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
2010                         sock_put(&new_smc->sk);
2011         }
2012
2013 out:
2014         release_sock(lsk);
2015         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
2016 }
2017
2018 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
2019 {
2020         struct smc_sock *lsmc =
2021                 smc_clcsock_user_data(listen_clcsock);
2022
2023         if (!lsmc)
2024                 return;
2025         lsmc->clcsk_data_ready(listen_clcsock);
2026         if (lsmc->sk.sk_state == SMC_LISTEN) {
2027                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
2028                 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
2029                         sock_put(&lsmc->sk);
2030         }
2031 }
2032
2033 static int smc_listen(struct socket *sock, int backlog)
2034 {
2035         struct sock *sk = sock->sk;
2036         struct smc_sock *smc;
2037         int rc;
2038
2039         smc = smc_sk(sk);
2040         lock_sock(sk);
2041
2042         rc = -EINVAL;
2043         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
2044             smc->connect_nonblock)
2045                 goto out;
2046
2047         rc = 0;
2048         if (sk->sk_state == SMC_LISTEN) {
2049                 sk->sk_max_ack_backlog = backlog;
2050                 goto out;
2051         }
2052         /* some socket options are handled in core, so we could not apply
2053          * them to the clc socket -- copy smc socket options to clc socket
2054          */
2055         smc_copy_sock_settings_to_clc(smc);
2056         if (!smc->use_fallback)
2057                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
2058
2059         /* save original sk_data_ready function and establish
2060          * smc-specific sk_data_ready function
2061          */
2062         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
2063         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
2064         smc->clcsock->sk->sk_user_data =
2065                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
2066         rc = kernel_listen(smc->clcsock, backlog);
2067         if (rc) {
2068                 smc->clcsock->sk->sk_data_ready = smc->clcsk_data_ready;
2069                 goto out;
2070         }
2071         sk->sk_max_ack_backlog = backlog;
2072         sk->sk_ack_backlog = 0;
2073         sk->sk_state = SMC_LISTEN;
2074
2075 out:
2076         release_sock(sk);
2077         return rc;
2078 }
2079
2080 static int smc_accept(struct socket *sock, struct socket *new_sock,
2081                       int flags, bool kern)
2082 {
2083         struct sock *sk = sock->sk, *nsk;
2084         DECLARE_WAITQUEUE(wait, current);
2085         struct smc_sock *lsmc;
2086         long timeo;
2087         int rc = 0;
2088
2089         lsmc = smc_sk(sk);
2090         sock_hold(sk); /* sock_put below */
2091         lock_sock(sk);
2092
2093         if (lsmc->sk.sk_state != SMC_LISTEN) {
2094                 rc = -EINVAL;
2095                 release_sock(sk);
2096                 goto out;
2097         }
2098
2099         /* Wait for an incoming connection */
2100         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
2101         add_wait_queue_exclusive(sk_sleep(sk), &wait);
2102         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
2103                 set_current_state(TASK_INTERRUPTIBLE);
2104                 if (!timeo) {
2105                         rc = -EAGAIN;
2106                         break;
2107                 }
2108                 release_sock(sk);
2109                 timeo = schedule_timeout(timeo);
2110                 /* wakeup by sk_data_ready in smc_listen_work() */
2111                 sched_annotate_sleep();
2112                 lock_sock(sk);
2113                 if (signal_pending(current)) {
2114                         rc = sock_intr_errno(timeo);
2115                         break;
2116                 }
2117         }
2118         set_current_state(TASK_RUNNING);
2119         remove_wait_queue(sk_sleep(sk), &wait);
2120
2121         if (!rc)
2122                 rc = sock_error(nsk);
2123         release_sock(sk);
2124         if (rc)
2125                 goto out;
2126
2127         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
2128                 /* wait till data arrives on the socket */
2129                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
2130                                                                 MSEC_PER_SEC);
2131                 if (smc_sk(nsk)->use_fallback) {
2132                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
2133
2134                         lock_sock(clcsk);
2135                         if (skb_queue_empty(&clcsk->sk_receive_queue))
2136                                 sk_wait_data(clcsk, &timeo, NULL);
2137                         release_sock(clcsk);
2138                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
2139                         lock_sock(nsk);
2140                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
2141                         release_sock(nsk);
2142                 }
2143         }
2144
2145 out:
2146         sock_put(sk); /* sock_hold above */
2147         return rc;
2148 }
2149
2150 static int smc_getname(struct socket *sock, struct sockaddr *addr,
2151                        int peer)
2152 {
2153         struct smc_sock *smc;
2154
2155         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
2156             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
2157                 return -ENOTCONN;
2158
2159         smc = smc_sk(sock->sk);
2160
2161         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
2162 }
2163
2164 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
2165 {
2166         struct sock *sk = sock->sk;
2167         struct smc_sock *smc;
2168         int rc = -EPIPE;
2169
2170         smc = smc_sk(sk);
2171         lock_sock(sk);
2172         if ((sk->sk_state != SMC_ACTIVE) &&
2173             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2174             (sk->sk_state != SMC_INIT))
2175                 goto out;
2176
2177         if (msg->msg_flags & MSG_FASTOPEN) {
2178                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2179                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2180                         if (rc)
2181                                 goto out;
2182                 } else {
2183                         rc = -EINVAL;
2184                         goto out;
2185                 }
2186         }
2187
2188         if (smc->use_fallback) {
2189                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2190         } else {
2191                 rc = smc_tx_sendmsg(smc, msg, len);
2192                 SMC_STAT_TX_PAYLOAD(smc, len, rc);
2193         }
2194 out:
2195         release_sock(sk);
2196         return rc;
2197 }
2198
2199 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2200                        int flags)
2201 {
2202         struct sock *sk = sock->sk;
2203         struct smc_sock *smc;
2204         int rc = -ENOTCONN;
2205
2206         smc = smc_sk(sk);
2207         lock_sock(sk);
2208         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2209                 /* socket was connected before, no more data to read */
2210                 rc = 0;
2211                 goto out;
2212         }
2213         if ((sk->sk_state == SMC_INIT) ||
2214             (sk->sk_state == SMC_LISTEN) ||
2215             (sk->sk_state == SMC_CLOSED))
2216                 goto out;
2217
2218         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2219                 rc = 0;
2220                 goto out;
2221         }
2222
2223         if (smc->use_fallback) {
2224                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2225         } else {
2226                 msg->msg_namelen = 0;
2227                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2228                 SMC_STAT_RX_PAYLOAD(smc, rc, rc);
2229         }
2230
2231 out:
2232         release_sock(sk);
2233         return rc;
2234 }
2235
2236 static __poll_t smc_accept_poll(struct sock *parent)
2237 {
2238         struct smc_sock *isk = smc_sk(parent);
2239         __poll_t mask = 0;
2240
2241         spin_lock(&isk->accept_q_lock);
2242         if (!list_empty(&isk->accept_q))
2243                 mask = EPOLLIN | EPOLLRDNORM;
2244         spin_unlock(&isk->accept_q_lock);
2245
2246         return mask;
2247 }
2248
2249 static __poll_t smc_poll(struct file *file, struct socket *sock,
2250                              poll_table *wait)
2251 {
2252         struct sock *sk = sock->sk;
2253         struct smc_sock *smc;
2254         __poll_t mask = 0;
2255
2256         if (!sk)
2257                 return EPOLLNVAL;
2258
2259         smc = smc_sk(sock->sk);
2260         if (smc->use_fallback) {
2261                 /* delegate to CLC child sock */
2262                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2263                 sk->sk_err = smc->clcsock->sk->sk_err;
2264         } else {
2265                 if (sk->sk_state != SMC_CLOSED)
2266                         sock_poll_wait(file, sock, wait);
2267                 if (sk->sk_err)
2268                         mask |= EPOLLERR;
2269                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2270                     (sk->sk_state == SMC_CLOSED))
2271                         mask |= EPOLLHUP;
2272                 if (sk->sk_state == SMC_LISTEN) {
2273                         /* woken up by sk_data_ready in smc_listen_work() */
2274                         mask |= smc_accept_poll(sk);
2275                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2276                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2277                                                            wait);
2278                         sk->sk_err = smc->clcsock->sk->sk_err;
2279                 } else {
2280                         if ((sk->sk_state != SMC_INIT &&
2281                              atomic_read(&smc->conn.sndbuf_space)) ||
2282                             sk->sk_shutdown & SEND_SHUTDOWN) {
2283                                 mask |= EPOLLOUT | EPOLLWRNORM;
2284                         } else {
2285                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2286                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2287                         }
2288                         if (atomic_read(&smc->conn.bytes_to_rcv))
2289                                 mask |= EPOLLIN | EPOLLRDNORM;
2290                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2291                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2292                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2293                                 mask |= EPOLLIN;
2294                         if (smc->conn.urg_state == SMC_URG_VALID)
2295                                 mask |= EPOLLPRI;
2296                 }
2297         }
2298
2299         return mask;
2300 }
2301
2302 static int smc_shutdown(struct socket *sock, int how)
2303 {
2304         struct sock *sk = sock->sk;
2305         bool do_shutdown = true;
2306         struct smc_sock *smc;
2307         int rc = -EINVAL;
2308         int old_state;
2309         int rc1 = 0;
2310
2311         smc = smc_sk(sk);
2312
2313         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2314                 return rc;
2315
2316         lock_sock(sk);
2317
2318         rc = -ENOTCONN;
2319         if ((sk->sk_state != SMC_ACTIVE) &&
2320             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2321             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2322             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2323             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2324             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2325                 goto out;
2326         if (smc->use_fallback) {
2327                 rc = kernel_sock_shutdown(smc->clcsock, how);
2328                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2329                 if (sk->sk_shutdown == SHUTDOWN_MASK)
2330                         sk->sk_state = SMC_CLOSED;
2331                 goto out;
2332         }
2333         switch (how) {
2334         case SHUT_RDWR:         /* shutdown in both directions */
2335                 old_state = sk->sk_state;
2336                 rc = smc_close_active(smc);
2337                 if (old_state == SMC_ACTIVE &&
2338                     sk->sk_state == SMC_PEERCLOSEWAIT1)
2339                         do_shutdown = false;
2340                 break;
2341         case SHUT_WR:
2342                 rc = smc_close_shutdown_write(smc);
2343                 break;
2344         case SHUT_RD:
2345                 rc = 0;
2346                 /* nothing more to do because peer is not involved */
2347                 break;
2348         }
2349         if (do_shutdown && smc->clcsock)
2350                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2351         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2352         sk->sk_shutdown |= how + 1;
2353
2354 out:
2355         release_sock(sk);
2356         return rc ? rc : rc1;
2357 }
2358
2359 static int smc_setsockopt(struct socket *sock, int level, int optname,
2360                           sockptr_t optval, unsigned int optlen)
2361 {
2362         struct sock *sk = sock->sk;
2363         struct smc_sock *smc;
2364         int val, rc;
2365
2366         if (level == SOL_TCP && optname == TCP_ULP)
2367                 return -EOPNOTSUPP;
2368
2369         smc = smc_sk(sk);
2370
2371         /* generic setsockopts reaching us here always apply to the
2372          * CLC socket
2373          */
2374         mutex_lock(&smc->clcsock_release_lock);
2375         if (!smc->clcsock) {
2376                 mutex_unlock(&smc->clcsock_release_lock);
2377                 return -EBADF;
2378         }
2379         if (unlikely(!smc->clcsock->ops->setsockopt))
2380                 rc = -EOPNOTSUPP;
2381         else
2382                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2383                                                    optval, optlen);
2384         if (smc->clcsock->sk->sk_err) {
2385                 sk->sk_err = smc->clcsock->sk->sk_err;
2386                 sk_error_report(sk);
2387         }
2388         mutex_unlock(&smc->clcsock_release_lock);
2389
2390         if (optlen < sizeof(int))
2391                 return -EINVAL;
2392         if (copy_from_sockptr(&val, optval, sizeof(int)))
2393                 return -EFAULT;
2394
2395         lock_sock(sk);
2396         if (rc || smc->use_fallback)
2397                 goto out;
2398         switch (optname) {
2399         case TCP_FASTOPEN:
2400         case TCP_FASTOPEN_CONNECT:
2401         case TCP_FASTOPEN_KEY:
2402         case TCP_FASTOPEN_NO_COOKIE:
2403                 /* option not supported by SMC */
2404                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2405                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2406                 } else {
2407                         rc = -EINVAL;
2408                 }
2409                 break;
2410         case TCP_NODELAY:
2411                 if (sk->sk_state != SMC_INIT &&
2412                     sk->sk_state != SMC_LISTEN &&
2413                     sk->sk_state != SMC_CLOSED) {
2414                         if (val) {
2415                                 SMC_STAT_INC(smc, ndly_cnt);
2416                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2417                                                  &smc->conn.tx_work, 0);
2418                         }
2419                 }
2420                 break;
2421         case TCP_CORK:
2422                 if (sk->sk_state != SMC_INIT &&
2423                     sk->sk_state != SMC_LISTEN &&
2424                     sk->sk_state != SMC_CLOSED) {
2425                         if (!val) {
2426                                 SMC_STAT_INC(smc, cork_cnt);
2427                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2428                                                  &smc->conn.tx_work, 0);
2429                         }
2430                 }
2431                 break;
2432         case TCP_DEFER_ACCEPT:
2433                 smc->sockopt_defer_accept = val;
2434                 break;
2435         default:
2436                 break;
2437         }
2438 out:
2439         release_sock(sk);
2440
2441         return rc;
2442 }
2443
2444 static int smc_getsockopt(struct socket *sock, int level, int optname,
2445                           char __user *optval, int __user *optlen)
2446 {
2447         struct smc_sock *smc;
2448         int rc;
2449
2450         smc = smc_sk(sock->sk);
2451         mutex_lock(&smc->clcsock_release_lock);
2452         if (!smc->clcsock) {
2453                 mutex_unlock(&smc->clcsock_release_lock);
2454                 return -EBADF;
2455         }
2456         /* socket options apply to the CLC socket */
2457         if (unlikely(!smc->clcsock->ops->getsockopt)) {
2458                 mutex_unlock(&smc->clcsock_release_lock);
2459                 return -EOPNOTSUPP;
2460         }
2461         rc = smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2462                                            optval, optlen);
2463         mutex_unlock(&smc->clcsock_release_lock);
2464         return rc;
2465 }
2466
2467 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2468                      unsigned long arg)
2469 {
2470         union smc_host_cursor cons, urg;
2471         struct smc_connection *conn;
2472         struct smc_sock *smc;
2473         int answ;
2474
2475         smc = smc_sk(sock->sk);
2476         conn = &smc->conn;
2477         lock_sock(&smc->sk);
2478         if (smc->use_fallback) {
2479                 if (!smc->clcsock) {
2480                         release_sock(&smc->sk);
2481                         return -EBADF;
2482                 }
2483                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2484                 release_sock(&smc->sk);
2485                 return answ;
2486         }
2487         switch (cmd) {
2488         case SIOCINQ: /* same as FIONREAD */
2489                 if (smc->sk.sk_state == SMC_LISTEN) {
2490                         release_sock(&smc->sk);
2491                         return -EINVAL;
2492                 }
2493                 if (smc->sk.sk_state == SMC_INIT ||
2494                     smc->sk.sk_state == SMC_CLOSED)
2495                         answ = 0;
2496                 else
2497                         answ = atomic_read(&smc->conn.bytes_to_rcv);
2498                 break;
2499         case SIOCOUTQ:
2500                 /* output queue size (not send + not acked) */
2501                 if (smc->sk.sk_state == SMC_LISTEN) {
2502                         release_sock(&smc->sk);
2503                         return -EINVAL;
2504                 }
2505                 if (smc->sk.sk_state == SMC_INIT ||
2506                     smc->sk.sk_state == SMC_CLOSED)
2507                         answ = 0;
2508                 else
2509                         answ = smc->conn.sndbuf_desc->len -
2510                                         atomic_read(&smc->conn.sndbuf_space);
2511                 break;
2512         case SIOCOUTQNSD:
2513                 /* output queue size (not send only) */
2514                 if (smc->sk.sk_state == SMC_LISTEN) {
2515                         release_sock(&smc->sk);
2516                         return -EINVAL;
2517                 }
2518                 if (smc->sk.sk_state == SMC_INIT ||
2519                     smc->sk.sk_state == SMC_CLOSED)
2520                         answ = 0;
2521                 else
2522                         answ = smc_tx_prepared_sends(&smc->conn);
2523                 break;
2524         case SIOCATMARK:
2525                 if (smc->sk.sk_state == SMC_LISTEN) {
2526                         release_sock(&smc->sk);
2527                         return -EINVAL;
2528                 }
2529                 if (smc->sk.sk_state == SMC_INIT ||
2530                     smc->sk.sk_state == SMC_CLOSED) {
2531                         answ = 0;
2532                 } else {
2533                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2534                         smc_curs_copy(&urg, &conn->urg_curs, conn);
2535                         answ = smc_curs_diff(conn->rmb_desc->len,
2536                                              &cons, &urg) == 1;
2537                 }
2538                 break;
2539         default:
2540                 release_sock(&smc->sk);
2541                 return -ENOIOCTLCMD;
2542         }
2543         release_sock(&smc->sk);
2544
2545         return put_user(answ, (int __user *)arg);
2546 }
2547
2548 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2549                             int offset, size_t size, int flags)
2550 {
2551         struct sock *sk = sock->sk;
2552         struct smc_sock *smc;
2553         int rc = -EPIPE;
2554
2555         smc = smc_sk(sk);
2556         lock_sock(sk);
2557         if (sk->sk_state != SMC_ACTIVE) {
2558                 release_sock(sk);
2559                 goto out;
2560         }
2561         release_sock(sk);
2562         if (smc->use_fallback) {
2563                 rc = kernel_sendpage(smc->clcsock, page, offset,
2564                                      size, flags);
2565         } else {
2566                 SMC_STAT_INC(smc, sendpage_cnt);
2567                 rc = sock_no_sendpage(sock, page, offset, size, flags);
2568         }
2569
2570 out:
2571         return rc;
2572 }
2573
2574 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2575  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2576  * updates till whenever a respective page has been fully processed.
2577  * Note that subsequent recv() calls have to wait till all splice() processing
2578  * completed.
2579  */
2580 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2581                                struct pipe_inode_info *pipe, size_t len,
2582                                unsigned int flags)
2583 {
2584         struct sock *sk = sock->sk;
2585         struct smc_sock *smc;
2586         int rc = -ENOTCONN;
2587
2588         smc = smc_sk(sk);
2589         lock_sock(sk);
2590         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2591                 /* socket was connected before, no more data to read */
2592                 rc = 0;
2593                 goto out;
2594         }
2595         if (sk->sk_state == SMC_INIT ||
2596             sk->sk_state == SMC_LISTEN ||
2597             sk->sk_state == SMC_CLOSED)
2598                 goto out;
2599
2600         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2601                 rc = 0;
2602                 goto out;
2603         }
2604
2605         if (smc->use_fallback) {
2606                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2607                                                     pipe, len, flags);
2608         } else {
2609                 if (*ppos) {
2610                         rc = -ESPIPE;
2611                         goto out;
2612                 }
2613                 if (flags & SPLICE_F_NONBLOCK)
2614                         flags = MSG_DONTWAIT;
2615                 else
2616                         flags = 0;
2617                 SMC_STAT_INC(smc, splice_cnt);
2618                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2619         }
2620 out:
2621         release_sock(sk);
2622
2623         return rc;
2624 }
2625
2626 /* must look like tcp */
2627 static const struct proto_ops smc_sock_ops = {
2628         .family         = PF_SMC,
2629         .owner          = THIS_MODULE,
2630         .release        = smc_release,
2631         .bind           = smc_bind,
2632         .connect        = smc_connect,
2633         .socketpair     = sock_no_socketpair,
2634         .accept         = smc_accept,
2635         .getname        = smc_getname,
2636         .poll           = smc_poll,
2637         .ioctl          = smc_ioctl,
2638         .listen         = smc_listen,
2639         .shutdown       = smc_shutdown,
2640         .setsockopt     = smc_setsockopt,
2641         .getsockopt     = smc_getsockopt,
2642         .sendmsg        = smc_sendmsg,
2643         .recvmsg        = smc_recvmsg,
2644         .mmap           = sock_no_mmap,
2645         .sendpage       = smc_sendpage,
2646         .splice_read    = smc_splice_read,
2647 };
2648
2649 static int smc_create(struct net *net, struct socket *sock, int protocol,
2650                       int kern)
2651 {
2652         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2653         struct smc_sock *smc;
2654         struct sock *sk;
2655         int rc;
2656
2657         rc = -ESOCKTNOSUPPORT;
2658         if (sock->type != SOCK_STREAM)
2659                 goto out;
2660
2661         rc = -EPROTONOSUPPORT;
2662         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2663                 goto out;
2664
2665         rc = -ENOBUFS;
2666         sock->ops = &smc_sock_ops;
2667         sk = smc_sock_alloc(net, sock, protocol);
2668         if (!sk)
2669                 goto out;
2670
2671         /* create internal TCP socket for CLC handshake and fallback */
2672         smc = smc_sk(sk);
2673         smc->use_fallback = false; /* assume rdma capability first */
2674         smc->fallback_rsn = 0;
2675         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2676                               &smc->clcsock);
2677         if (rc) {
2678                 sk_common_release(sk);
2679                 goto out;
2680         }
2681         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2682         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2683
2684 out:
2685         return rc;
2686 }
2687
2688 static const struct net_proto_family smc_sock_family_ops = {
2689         .family = PF_SMC,
2690         .owner  = THIS_MODULE,
2691         .create = smc_create,
2692 };
2693
2694 unsigned int smc_net_id;
2695
2696 static __net_init int smc_net_init(struct net *net)
2697 {
2698         return smc_pnet_net_init(net);
2699 }
2700
2701 static void __net_exit smc_net_exit(struct net *net)
2702 {
2703         smc_pnet_net_exit(net);
2704 }
2705
2706 static __net_init int smc_net_stat_init(struct net *net)
2707 {
2708         return smc_stats_init(net);
2709 }
2710
2711 static void __net_exit smc_net_stat_exit(struct net *net)
2712 {
2713         smc_stats_exit(net);
2714 }
2715
2716 static struct pernet_operations smc_net_ops = {
2717         .init = smc_net_init,
2718         .exit = smc_net_exit,
2719         .id   = &smc_net_id,
2720         .size = sizeof(struct smc_net),
2721 };
2722
2723 static struct pernet_operations smc_net_stat_ops = {
2724         .init = smc_net_stat_init,
2725         .exit = smc_net_stat_exit,
2726 };
2727
2728 static int __init smc_init(void)
2729 {
2730         int rc;
2731
2732         rc = register_pernet_subsys(&smc_net_ops);
2733         if (rc)
2734                 return rc;
2735
2736         rc = register_pernet_subsys(&smc_net_stat_ops);
2737         if (rc)
2738                 return rc;
2739
2740         smc_ism_init();
2741         smc_clc_init();
2742
2743         rc = smc_nl_init();
2744         if (rc)
2745                 goto out_pernet_subsys;
2746
2747         rc = smc_pnet_init();
2748         if (rc)
2749                 goto out_nl;
2750
2751         rc = -ENOMEM;
2752         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2753         if (!smc_hs_wq)
2754                 goto out_pnet;
2755
2756         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2757         if (!smc_close_wq)
2758                 goto out_alloc_hs_wq;
2759
2760         rc = smc_core_init();
2761         if (rc) {
2762                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2763                 goto out_alloc_wqs;
2764         }
2765
2766         rc = smc_llc_init();
2767         if (rc) {
2768                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2769                 goto out_core;
2770         }
2771
2772         rc = smc_cdc_init();
2773         if (rc) {
2774                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2775                 goto out_core;
2776         }
2777
2778         rc = proto_register(&smc_proto, 1);
2779         if (rc) {
2780                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2781                 goto out_core;
2782         }
2783
2784         rc = proto_register(&smc_proto6, 1);
2785         if (rc) {
2786                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2787                 goto out_proto;
2788         }
2789
2790         rc = sock_register(&smc_sock_family_ops);
2791         if (rc) {
2792                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2793                 goto out_proto6;
2794         }
2795         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2796         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2797
2798         rc = smc_ib_register_client();
2799         if (rc) {
2800                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2801                 goto out_sock;
2802         }
2803
2804         static_branch_enable(&tcp_have_smc);
2805         return 0;
2806
2807 out_sock:
2808         sock_unregister(PF_SMC);
2809 out_proto6:
2810         proto_unregister(&smc_proto6);
2811 out_proto:
2812         proto_unregister(&smc_proto);
2813 out_core:
2814         smc_core_exit();
2815 out_alloc_wqs:
2816         destroy_workqueue(smc_close_wq);
2817 out_alloc_hs_wq:
2818         destroy_workqueue(smc_hs_wq);
2819 out_pnet:
2820         smc_pnet_exit();
2821 out_nl:
2822         smc_nl_exit();
2823 out_pernet_subsys:
2824         unregister_pernet_subsys(&smc_net_ops);
2825
2826         return rc;
2827 }
2828
2829 static void __exit smc_exit(void)
2830 {
2831         static_branch_disable(&tcp_have_smc);
2832         sock_unregister(PF_SMC);
2833         smc_core_exit();
2834         smc_ib_unregister_client();
2835         destroy_workqueue(smc_close_wq);
2836         destroy_workqueue(smc_hs_wq);
2837         proto_unregister(&smc_proto6);
2838         proto_unregister(&smc_proto);
2839         smc_pnet_exit();
2840         smc_nl_exit();
2841         unregister_pernet_subsys(&smc_net_stat_ops);
2842         unregister_pernet_subsys(&smc_net_ops);
2843         rcu_barrier();
2844 }
2845
2846 module_init(smc_init);
2847 module_exit(smc_exit);
2848
2849 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2850 MODULE_DESCRIPTION("smc socket address family");
2851 MODULE_LICENSE("GPL");
2852 MODULE_ALIAS_NETPROTO(PF_SMC);