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