net/smc/af_smc.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   5  *  applies to SOCK_STREAM sockets only
   6  *  offers an alternative communication option for TCP-protocol sockets
   7  *  applicable with RoCE-cards only
   8  *
   9  *  Initial restrictions:
  10  *    - non-blocking connect postponed
  11  *    - IPv6 support postponed
  12  *    - support for alternate links postponed
  13  *    - partial support for non-blocking sockets only
  14  *    - support for urgent data postponed
  15  *
  16  *  Copyright IBM Corp. 2016
  17  *
  18  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  19  *              based on prototype from Frank Blaschka
  20  */
  21
  22 #define KMSG_COMPONENT "smc"
  23 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  24
  25 #include <linux/module.h>
  26 #include <linux/socket.h>
  27 #include <linux/inetdevice.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/in.h>
  30 #include <net/sock.h>
  31 #include <net/tcp.h>
  32
  33 #include "smc.h"
  34 #include "smc_clc.h"
  35 #include "smc_llc.h"
  36 #include "smc_cdc.h"
  37 #include "smc_core.h"
  38 #include "smc_ib.h"
  39 #include "smc_pnet.h"
  40 #include "smc_tx.h"
  41
  42 static DEFINE_MUTEX(smc_create_lgr_pending);    /* serialize link group
  43                                                  * creation
  44                                                  */
  45
  46 struct smc_lgr_list smc_lgr_list = {            /* established link groups */
  47         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
  48         .list = LIST_HEAD_INIT(smc_lgr_list.list),
  49 };
  50
  51 static void smc_tcp_listen_work(struct work_struct *);
  52
  53 static void smc_set_keepalive(struct sock *sk, int val)
  54 {
  55         struct smc_sock *smc = smc_sk(sk);
  56
  57         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  58 }
  59
  60 static struct proto smc_proto = {
  61         .name           = "SMC",
  62         .owner          = THIS_MODULE,
  63         .keepalive      = smc_set_keepalive,
  64         .obj_size       = sizeof(struct smc_sock),
  65         .slab_flags     = SLAB_DESTROY_BY_RCU,
  66 };
  67
  68 static int smc_release(struct socket *sock)
  69 {
  70         struct sock *sk = sock->sk;
  71         struct smc_sock *smc;
  72
  73         if (!sk)
  74                 goto out;
  75
  76         smc = smc_sk(sk);
  77         lock_sock(sk);
  78
  79         sk->sk_state = SMC_CLOSED;
  80         if (smc->clcsock) {
  81                 sock_release(smc->clcsock);
  82                 smc->clcsock = NULL;
  83         }
  84
  85         /* detach socket */
  86         sock_orphan(sk);
  87         sock->sk = NULL;
  88         release_sock(sk);
  89
  90         sock_put(sk);
  91 out:
  92         return 0;
  93 }
  94
  95 static void smc_destruct(struct sock *sk)
  96 {
  97         if (sk->sk_state != SMC_CLOSED)
  98                 return;
  99         if (!sock_flag(sk, SOCK_DEAD))
 100                 return;
 101
 102         sk_refcnt_debug_dec(sk);
 103 }
 104
 105 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
 106 {
 107         struct smc_sock *smc;
 108         struct sock *sk;
 109
 110         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
 111         if (!sk)
 112                 return NULL;
 113
 114         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 115         sk->sk_state = SMC_INIT;
 116         sk->sk_destruct = smc_destruct;
 117         sk->sk_protocol = SMCPROTO_SMC;
 118         smc = smc_sk(sk);
 119         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 120         INIT_LIST_HEAD(&smc->accept_q);
 121         spin_lock_init(&smc->accept_q_lock);
 122         sk_refcnt_debug_inc(sk);
 123
 124         return sk;
 125 }
 126
 127 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 128                     int addr_len)
 129 {
 130         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 131         struct sock *sk = sock->sk;
 132         struct smc_sock *smc;
 133         int rc;
 134
 135         smc = smc_sk(sk);
 136
 137         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 138         rc = -EINVAL;
 139         if (addr_len < sizeof(struct sockaddr_in))
 140                 goto out;
 141
 142         rc = -EAFNOSUPPORT;
 143         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 144         if ((addr->sin_family != AF_INET) &&
 145             ((addr->sin_family != AF_UNSPEC) ||
 146              (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
 147                 goto out;
 148
 149         lock_sock(sk);
 150
 151         /* Check if socket is already active */
 152         rc = -EINVAL;
 153         if (sk->sk_state != SMC_INIT)
 154                 goto out_rel;
 155
 156         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 157         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 158
 159 out_rel:
 160         release_sock(sk);
 161 out:
 162         return rc;
 163 }
 164
 165 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 166                                    unsigned long mask)
 167 {
 168         /* options we don't get control via setsockopt for */
 169         nsk->sk_type = osk->sk_type;
 170         nsk->sk_sndbuf = osk->sk_sndbuf;
 171         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 172         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 173         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 174         nsk->sk_mark = osk->sk_mark;
 175         nsk->sk_priority = osk->sk_priority;
 176         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 177         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 178         nsk->sk_err = osk->sk_err;
 179
 180         nsk->sk_flags &= ~mask;
 181         nsk->sk_flags |= osk->sk_flags & mask;
 182 }
 183
 184 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 185                              (1UL << SOCK_KEEPOPEN) | \
 186                              (1UL << SOCK_LINGER) | \
 187                              (1UL << SOCK_BROADCAST) | \
 188                              (1UL << SOCK_TIMESTAMP) | \
 189                              (1UL << SOCK_DBG) | \
 190                              (1UL << SOCK_RCVTSTAMP) | \
 191                              (1UL << SOCK_RCVTSTAMPNS) | \
 192                              (1UL << SOCK_LOCALROUTE) | \
 193                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 194                              (1UL << SOCK_RXQ_OVFL) | \
 195                              (1UL << SOCK_WIFI_STATUS) | \
 196                              (1UL << SOCK_NOFCS) | \
 197                              (1UL << SOCK_FILTER_LOCKED))
 198 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 199  * clc socket (since smc is not called for these options from net/core)
 200  */
 201 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 202 {
 203         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 204 }
 205
 206 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 207                              (1UL << SOCK_KEEPOPEN) | \
 208                              (1UL << SOCK_LINGER) | \
 209                              (1UL << SOCK_DBG))
 210 /* copy only settings and flags relevant for smc from clc to smc socket */
 211 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 212 {
 213         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 214 }
 215
 216 /* determine subnet and mask of internal TCP socket */
 217 int smc_netinfo_by_tcpsk(struct socket *clcsock,
 218                          __be32 *subnet, u8 *prefix_len)
 219 {
 220         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 221         struct sockaddr_in addr;
 222         int rc = -ENOENT;
 223         int len;
 224
 225         if (!dst) {
 226                 rc = -ENOTCONN;
 227                 goto out;
 228         }
 229         if (!dst->dev) {
 230                 rc = -ENODEV;
 231                 goto out_rel;
 232         }
 233
 234         /* get address to which the internal TCP socket is bound */
 235         kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
 236         /* analyze IPv4 specific data of net_device belonging to TCP socket */
 237         for_ifa(dst->dev->ip_ptr) {
 238                 if (ifa->ifa_address != addr.sin_addr.s_addr)
 239                         continue;
 240                 *prefix_len = inet_mask_len(ifa->ifa_mask);
 241                 *subnet = ifa->ifa_address & ifa->ifa_mask;
 242                 rc = 0;
 243                 break;
 244         } endfor_ifa(dst->dev->ip_ptr);
 245
 246 out_rel:
 247         dst_release(dst);
 248 out:
 249         return rc;
 250 }
 251
 252 static int smc_clnt_conf_first_link(struct smc_sock *smc, union ib_gid *gid)
 253 {
 254         struct smc_link_group *lgr = smc->conn.lgr;
 255         struct smc_link *link;
 256         int rest;
 257         int rc;
 258
 259         link = &lgr->lnk[SMC_SINGLE_LINK];
 260         /* receive CONFIRM LINK request from server over RoCE fabric */
 261         rest = wait_for_completion_interruptible_timeout(
 262                 &link->llc_confirm,
 263                 SMC_LLC_WAIT_FIRST_TIME);
 264         if (rest <= 0) {
 265                 struct smc_clc_msg_decline dclc;
 266
 267                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 268                                       SMC_CLC_DECLINE);
 269                 return rc;
 270         }
 271
 272         rc = smc_ib_modify_qp_rts(link);
 273         if (rc)
 274                 return SMC_CLC_DECL_INTERR;
 275
 276         smc_wr_remember_qp_attr(link);
 277         /* send CONFIRM LINK response over RoCE fabric */
 278         rc = smc_llc_send_confirm_link(link,
 279                                        link->smcibdev->mac[link->ibport - 1],
 280                                        gid, SMC_LLC_RESP);
 281         if (rc < 0)
 282                 return SMC_CLC_DECL_TCL;
 283
 284         return rc;
 285 }
 286
 287 static void smc_conn_save_peer_info(struct smc_sock *smc,
 288                                     struct smc_clc_msg_accept_confirm *clc)
 289 {
 290         smc->conn.peer_conn_idx = clc->conn_idx;
 291         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 292         smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
 293         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 294 }
 295
 296 static void smc_link_save_peer_info(struct smc_link *link,
 297                                     struct smc_clc_msg_accept_confirm *clc)
 298 {
 299         link->peer_qpn = ntoh24(clc->qpn);
 300         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 301         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 302         link->peer_psn = ntoh24(clc->psn);
 303         link->peer_mtu = clc->qp_mtu;
 304 }
 305
 306 /* setup for RDMA connection of client */
 307 static int smc_connect_rdma(struct smc_sock *smc)
 308 {
 309         struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
 310         struct smc_clc_msg_accept_confirm aclc;
 311         int local_contact = SMC_FIRST_CONTACT;
 312         struct smc_ib_device *smcibdev;
 313         struct smc_link *link;
 314         u8 srv_first_contact;
 315         int reason_code = 0;
 316         int rc = 0;
 317         u8 ibport;
 318
 319         /* IPSec connections opt out of SMC-R optimizations */
 320         if (using_ipsec(smc)) {
 321                 reason_code = SMC_CLC_DECL_IPSEC;
 322                 goto decline_rdma;
 323         }
 324
 325         /* PNET table look up: search active ib_device and port
 326          * within same PNETID that also contains the ethernet device
 327          * used for the internal TCP socket
 328          */
 329         smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
 330         if (!smcibdev) {
 331                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 332                 goto decline_rdma;
 333         }
 334
 335         /* do inband token exchange */
 336         reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
 337         if (reason_code < 0) {
 338                 rc = reason_code;
 339                 goto out_err;
 340         }
 341         if (reason_code > 0) /* configuration error */
 342                 goto decline_rdma;
 343         /* receive SMC Accept CLC message */
 344         reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
 345                                        SMC_CLC_ACCEPT);
 346         if (reason_code < 0) {
 347                 rc = reason_code;
 348                 goto out_err;
 349         }
 350         if (reason_code > 0)
 351                 goto decline_rdma;
 352
 353         srv_first_contact = aclc.hdr.flag;
 354         mutex_lock(&smc_create_lgr_pending);
 355         local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
 356                                         ibport, &aclc.lcl, srv_first_contact);
 357         if (local_contact < 0) {
 358                 rc = local_contact;
 359                 if (rc == -ENOMEM)
 360                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 361                 else if (rc == -ENOLINK)
 362                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 363                 goto decline_rdma_unlock;
 364         }
 365         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 366
 367         smc_conn_save_peer_info(smc, &aclc);
 368
 369         rc = smc_sndbuf_create(smc);
 370         if (rc) {
 371                 reason_code = SMC_CLC_DECL_MEM;
 372                 goto decline_rdma_unlock;
 373         }
 374         rc = smc_rmb_create(smc);
 375         if (rc) {
 376                 reason_code = SMC_CLC_DECL_MEM;
 377                 goto decline_rdma_unlock;
 378         }
 379
 380         if (local_contact == SMC_FIRST_CONTACT)
 381                 smc_link_save_peer_info(link, &aclc);
 382
 383         rc = smc_rmb_rtoken_handling(&smc->conn, &aclc);
 384         if (rc) {
 385                 reason_code = SMC_CLC_DECL_INTERR;
 386                 goto decline_rdma_unlock;
 387         }
 388
 389         if (local_contact == SMC_FIRST_CONTACT) {
 390                 rc = smc_ib_ready_link(link);
 391                 if (rc) {
 392                         reason_code = SMC_CLC_DECL_INTERR;
 393                         goto decline_rdma_unlock;
 394                 }
 395         }
 396
 397         rc = smc_clc_send_confirm(smc);
 398         if (rc)
 399                 goto out_err_unlock;
 400
 401         if (local_contact == SMC_FIRST_CONTACT) {
 402                 /* QP confirmation over RoCE fabric */
 403                 reason_code = smc_clnt_conf_first_link(
 404                         smc, &smcibdev->gid[ibport - 1]);
 405                 if (reason_code < 0) {
 406                         rc = reason_code;
 407                         goto out_err_unlock;
 408                 }
 409                 if (reason_code > 0)
 410                         goto decline_rdma_unlock;
 411         }
 412
 413         mutex_unlock(&smc_create_lgr_pending);
 414         smc_tx_init(smc);
 415
 416 out_connected:
 417         smc_copy_sock_settings_to_clc(smc);
 418         smc->sk.sk_state = SMC_ACTIVE;
 419
 420         return rc ? rc : local_contact;
 421
 422 decline_rdma_unlock:
 423         mutex_unlock(&smc_create_lgr_pending);
 424         smc_conn_free(&smc->conn);
 425 decline_rdma:
 426         /* RDMA setup failed, switch back to TCP */
 427         smc->use_fallback = true;
 428         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 429                 rc = smc_clc_send_decline(smc, reason_code, 0);
 430                 if (rc < sizeof(struct smc_clc_msg_decline))
 431                         goto out_err;
 432         }
 433         goto out_connected;
 434
 435 out_err_unlock:
 436         mutex_unlock(&smc_create_lgr_pending);
 437         smc_conn_free(&smc->conn);
 438 out_err:
 439         return rc;
 440 }
 441
 442 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 443                        int alen, int flags)
 444 {
 445         struct sock *sk = sock->sk;
 446         struct smc_sock *smc;
 447         int rc = -EINVAL;
 448
 449         smc = smc_sk(sk);
 450
 451         /* separate smc parameter checking to be safe */
 452         if (alen < sizeof(addr->sa_family))
 453                 goto out_err;
 454         if (addr->sa_family != AF_INET)
 455                 goto out_err;
 456         smc->addr = addr;       /* needed for nonblocking connect */
 457
 458         lock_sock(sk);
 459         switch (sk->sk_state) {
 460         default:
 461                 goto out;
 462         case SMC_ACTIVE:
 463                 rc = -EISCONN;
 464                 goto out;
 465         case SMC_INIT:
 466                 rc = 0;
 467                 break;
 468         }
 469
 470         smc_copy_sock_settings_to_clc(smc);
 471         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 472         if (rc)
 473                 goto out;
 474
 475         /* setup RDMA connection */
 476         rc = smc_connect_rdma(smc);
 477         if (rc < 0)
 478                 goto out;
 479         else
 480                 rc = 0; /* success cases including fallback */
 481
 482 out:
 483         release_sock(sk);
 484 out_err:
 485         return rc;
 486 }
 487
 488 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 489 {
 490         struct sock *sk = &lsmc->sk;
 491         struct socket *new_clcsock;
 492         struct sock *new_sk;
 493         int rc;
 494
 495         release_sock(&lsmc->sk);
 496         new_sk = smc_sock_alloc(sock_net(sk), NULL);
 497         if (!new_sk) {
 498                 rc = -ENOMEM;
 499                 lsmc->sk.sk_err = ENOMEM;
 500                 *new_smc = NULL;
 501                 lock_sock(&lsmc->sk);
 502                 goto out;
 503         }
 504         *new_smc = smc_sk(new_sk);
 505
 506         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 507         lock_sock(&lsmc->sk);
 508         if  (rc < 0) {
 509                 lsmc->sk.sk_err = -rc;
 510                 new_sk->sk_state = SMC_CLOSED;
 511                 sock_set_flag(new_sk, SOCK_DEAD);
 512                 sock_put(new_sk);
 513                 *new_smc = NULL;
 514                 goto out;
 515         }
 516         if (lsmc->sk.sk_state == SMC_CLOSED) {
 517                 if (new_clcsock)
 518                         sock_release(new_clcsock);
 519                 new_sk->sk_state = SMC_CLOSED;
 520                 sock_set_flag(new_sk, SOCK_DEAD);
 521                 sock_put(new_sk);
 522                 *new_smc = NULL;
 523                 goto out;
 524         }
 525
 526         (*new_smc)->clcsock = new_clcsock;
 527 out:
 528         return rc;
 529 }
 530
 531 /* add a just created sock to the accept queue of the listen sock as
 532  * candidate for a following socket accept call from user space
 533  */
 534 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 535 {
 536         struct smc_sock *par = smc_sk(parent);
 537
 538         sock_hold(sk);
 539         spin_lock(&par->accept_q_lock);
 540         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 541         spin_unlock(&par->accept_q_lock);
 542         sk_acceptq_added(parent);
 543 }
 544
 545 /* remove a socket from the accept queue of its parental listening socket */
 546 static void smc_accept_unlink(struct sock *sk)
 547 {
 548         struct smc_sock *par = smc_sk(sk)->listen_smc;
 549
 550         spin_lock(&par->accept_q_lock);
 551         list_del_init(&smc_sk(sk)->accept_q);
 552         spin_unlock(&par->accept_q_lock);
 553         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 554         sock_put(sk);
 555 }
 556
 557 /* remove a sock from the accept queue to bind it to a new socket created
 558  * for a socket accept call from user space
 559  */
 560 static struct sock *smc_accept_dequeue(struct sock *parent,
 561                                        struct socket *new_sock)
 562 {
 563         struct smc_sock *isk, *n;
 564         struct sock *new_sk;
 565
 566         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 567                 new_sk = (struct sock *)isk;
 568
 569                 smc_accept_unlink(new_sk);
 570                 if (new_sk->sk_state == SMC_CLOSED) {
 571                         /* tbd in follow-on patch: close this sock */
 572                         continue;
 573                 }
 574                 if (new_sock)
 575                         sock_graft(new_sk, new_sock);
 576                 return new_sk;
 577         }
 578         return NULL;
 579 }
 580
 581 /* clean up for a created but never accepted sock */
 582 static void smc_close_non_accepted(struct sock *sk)
 583 {
 584         struct smc_sock *smc = smc_sk(sk);
 585
 586         sock_hold(sk);
 587         if (smc->clcsock) {
 588                 struct socket *tcp;
 589
 590                 tcp = smc->clcsock;
 591                 smc->clcsock = NULL;
 592                 sock_release(tcp);
 593         }
 594         /* more closing stuff to be added with socket closing patch */
 595         sock_put(sk);
 596 }
 597
 598 static int smc_serv_conf_first_link(struct smc_sock *smc)
 599 {
 600         struct smc_link_group *lgr = smc->conn.lgr;
 601         struct smc_link *link;
 602         int rest;
 603         int rc;
 604
 605         link = &lgr->lnk[SMC_SINGLE_LINK];
 606         /* send CONFIRM LINK request to client over the RoCE fabric */
 607         rc = smc_llc_send_confirm_link(link,
 608                                        link->smcibdev->mac[link->ibport - 1],
 609                                        &link->smcibdev->gid[link->ibport - 1],
 610                                        SMC_LLC_REQ);
 611         if (rc < 0)
 612                 return SMC_CLC_DECL_TCL;
 613
 614         /* receive CONFIRM LINK response from client over the RoCE fabric */
 615         rest = wait_for_completion_interruptible_timeout(
 616                 &link->llc_confirm_resp,
 617                 SMC_LLC_WAIT_FIRST_TIME);
 618         if (rest <= 0) {
 619                 struct smc_clc_msg_decline dclc;
 620
 621                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 622                                       SMC_CLC_DECLINE);
 623         }
 624
 625         return rc;
 626 }
 627
 628 /* setup for RDMA connection of server */
 629 static void smc_listen_work(struct work_struct *work)
 630 {
 631         struct smc_sock *new_smc = container_of(work, struct smc_sock,
 632                                                 smc_listen_work);
 633         struct socket *newclcsock = new_smc->clcsock;
 634         struct smc_sock *lsmc = new_smc->listen_smc;
 635         struct smc_clc_msg_accept_confirm cclc;
 636         int local_contact = SMC_REUSE_CONTACT;
 637         struct sock *newsmcsk = &new_smc->sk;
 638         struct smc_clc_msg_proposal pclc;
 639         struct smc_ib_device *smcibdev;
 640         struct sockaddr_in peeraddr;
 641         struct smc_link *link;
 642         int reason_code = 0;
 643         int rc = 0, len;
 644         __be32 subnet;
 645         u8 prefix_len;
 646         u8 ibport;
 647
 648         /* do inband token exchange -
 649          *wait for and receive SMC Proposal CLC message
 650          */
 651         reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
 652                                        SMC_CLC_PROPOSAL);
 653         if (reason_code < 0)
 654                 goto out_err;
 655         if (reason_code > 0)
 656                 goto decline_rdma;
 657
 658         /* IPSec connections opt out of SMC-R optimizations */
 659         if (using_ipsec(new_smc)) {
 660                 reason_code = SMC_CLC_DECL_IPSEC;
 661                 goto decline_rdma;
 662         }
 663
 664         /* PNET table look up: search active ib_device and port
 665          * within same PNETID that also contains the ethernet device
 666          * used for the internal TCP socket
 667          */
 668         smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
 669         if (!smcibdev) {
 670                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 671                 goto decline_rdma;
 672         }
 673
 674         /* determine subnet and mask from internal TCP socket */
 675         rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
 676         if (rc) {
 677                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 678                 goto decline_rdma;
 679         }
 680         if ((pclc.outgoing_subnet != subnet) ||
 681             (pclc.prefix_len != prefix_len)) {
 682                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 683                 goto decline_rdma;
 684         }
 685
 686         /* get address of the peer connected to the internal TCP socket */
 687         kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
 688
 689         /* allocate connection / link group */
 690         mutex_lock(&smc_create_lgr_pending);
 691         local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
 692                                         smcibdev, ibport, &pclc.lcl, 0);
 693         if (local_contact == SMC_REUSE_CONTACT)
 694                 /* lock no longer needed, free it due to following
 695                  * smc_clc_wait_msg() call
 696                  */
 697                 mutex_unlock(&smc_create_lgr_pending);
 698         if (local_contact < 0) {
 699                 rc = local_contact;
 700                 if (rc == -ENOMEM)
 701                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 702                 else if (rc == -ENOLINK)
 703                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 704                 goto decline_rdma;
 705         }
 706         link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 707
 708         rc = smc_sndbuf_create(new_smc);
 709         if (rc) {
 710                 reason_code = SMC_CLC_DECL_MEM;
 711                 goto decline_rdma;
 712         }
 713         rc = smc_rmb_create(new_smc);
 714         if (rc) {
 715                 reason_code = SMC_CLC_DECL_MEM;
 716                 goto decline_rdma;
 717         }
 718
 719         rc = smc_clc_send_accept(new_smc, local_contact);
 720         if (rc)
 721                 goto out_err;
 722
 723         /* receive SMC Confirm CLC message */
 724         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
 725                                        SMC_CLC_CONFIRM);
 726         if (reason_code < 0)
 727                 goto out_err;
 728         if (reason_code > 0)
 729                 goto decline_rdma;
 730         smc_conn_save_peer_info(new_smc, &cclc);
 731         if (local_contact == SMC_FIRST_CONTACT)
 732                 smc_link_save_peer_info(link, &cclc);
 733
 734         rc = smc_rmb_rtoken_handling(&new_smc->conn, &cclc);
 735         if (rc) {
 736                 reason_code = SMC_CLC_DECL_INTERR;
 737                 goto decline_rdma;
 738         }
 739
 740         if (local_contact == SMC_FIRST_CONTACT) {
 741                 rc = smc_ib_ready_link(link);
 742                 if (rc) {
 743                         reason_code = SMC_CLC_DECL_INTERR;
 744                         goto decline_rdma;
 745                 }
 746                 /* QP confirmation over RoCE fabric */
 747                 reason_code = smc_serv_conf_first_link(new_smc);
 748                 if (reason_code < 0) {
 749                         /* peer is not aware of a problem */
 750                         rc = reason_code;
 751                         goto out_err;
 752                 }
 753                 if (reason_code > 0)
 754                         goto decline_rdma;
 755         }
 756
 757         smc_tx_init(new_smc);
 758
 759 out_connected:
 760         sk_refcnt_debug_inc(newsmcsk);
 761         newsmcsk->sk_state = SMC_ACTIVE;
 762 enqueue:
 763         if (local_contact == SMC_FIRST_CONTACT)
 764                 mutex_unlock(&smc_create_lgr_pending);
 765         lock_sock(&lsmc->sk);
 766         if (lsmc->sk.sk_state == SMC_LISTEN) {
 767                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
 768         } else { /* no longer listening */
 769                 smc_close_non_accepted(newsmcsk);
 770         }
 771         release_sock(&lsmc->sk);
 772
 773         /* Wake up accept */
 774         lsmc->sk.sk_data_ready(&lsmc->sk);
 775         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
 776         return;
 777
 778 decline_rdma:
 779         /* RDMA setup failed, switch back to TCP */
 780         smc_conn_free(&new_smc->conn);
 781         new_smc->use_fallback = true;
 782         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 783                 rc = smc_clc_send_decline(new_smc, reason_code, 0);
 784                 if (rc < sizeof(struct smc_clc_msg_decline))
 785                         goto out_err;
 786         }
 787         goto out_connected;
 788
 789 out_err:
 790         newsmcsk->sk_state = SMC_CLOSED;
 791         goto enqueue; /* queue new sock with sk_err set */
 792 }
 793
 794 static void smc_tcp_listen_work(struct work_struct *work)
 795 {
 796         struct smc_sock *lsmc = container_of(work, struct smc_sock,
 797                                              tcp_listen_work);
 798         struct smc_sock *new_smc;
 799         int rc = 0;
 800
 801         lock_sock(&lsmc->sk);
 802         while (lsmc->sk.sk_state == SMC_LISTEN) {
 803                 rc = smc_clcsock_accept(lsmc, &new_smc);
 804                 if (rc)
 805                         goto out;
 806                 if (!new_smc)
 807                         continue;
 808
 809                 new_smc->listen_smc = lsmc;
 810                 new_smc->use_fallback = false; /* assume rdma capability first*/
 811                 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
 812                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
 813                 smc_copy_sock_settings_to_smc(new_smc);
 814                 schedule_work(&new_smc->smc_listen_work);
 815         }
 816
 817 out:
 818         release_sock(&lsmc->sk);
 819         lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
 820 }
 821
 822 static int smc_listen(struct socket *sock, int backlog)
 823 {
 824         struct sock *sk = sock->sk;
 825         struct smc_sock *smc;
 826         int rc;
 827
 828         smc = smc_sk(sk);
 829         lock_sock(sk);
 830
 831         rc = -EINVAL;
 832         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
 833                 goto out;
 834
 835         rc = 0;
 836         if (sk->sk_state == SMC_LISTEN) {
 837                 sk->sk_max_ack_backlog = backlog;
 838                 goto out;
 839         }
 840         /* some socket options are handled in core, so we could not apply
 841          * them to the clc socket -- copy smc socket options to clc socket
 842          */
 843         smc_copy_sock_settings_to_clc(smc);
 844
 845         rc = kernel_listen(smc->clcsock, backlog);
 846         if (rc)
 847                 goto out;
 848         sk->sk_max_ack_backlog = backlog;
 849         sk->sk_ack_backlog = 0;
 850         sk->sk_state = SMC_LISTEN;
 851         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 852         schedule_work(&smc->tcp_listen_work);
 853
 854 out:
 855         release_sock(sk);
 856         return rc;
 857 }
 858
 859 static int smc_accept(struct socket *sock, struct socket *new_sock,
 860                       int flags)
 861 {
 862         struct sock *sk = sock->sk, *nsk;
 863         DECLARE_WAITQUEUE(wait, current);
 864         struct smc_sock *lsmc;
 865         long timeo;
 866         int rc = 0;
 867
 868         lsmc = smc_sk(sk);
 869         lock_sock(sk);
 870
 871         if (lsmc->sk.sk_state != SMC_LISTEN) {
 872                 rc = -EINVAL;
 873                 goto out;
 874         }
 875
 876         /* Wait for an incoming connection */
 877         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 878         add_wait_queue_exclusive(sk_sleep(sk), &wait);
 879         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
 880                 set_current_state(TASK_INTERRUPTIBLE);
 881                 if (!timeo) {
 882                         rc = -EAGAIN;
 883                         break;
 884                 }
 885                 release_sock(sk);
 886                 timeo = schedule_timeout(timeo);
 887                 /* wakeup by sk_data_ready in smc_listen_work() */
 888                 sched_annotate_sleep();
 889                 lock_sock(sk);
 890                 if (signal_pending(current)) {
 891                         rc = sock_intr_errno(timeo);
 892                         break;
 893                 }
 894         }
 895         set_current_state(TASK_RUNNING);
 896         remove_wait_queue(sk_sleep(sk), &wait);
 897
 898         if (!rc)
 899                 rc = sock_error(nsk);
 900
 901 out:
 902         release_sock(sk);
 903         return rc;
 904 }
 905
 906 static int smc_getname(struct socket *sock, struct sockaddr *addr,
 907                        int *len, int peer)
 908 {
 909         struct smc_sock *smc;
 910
 911         if (peer && (sock->sk->sk_state != SMC_ACTIVE))
 912                 return -ENOTCONN;
 913
 914         smc = smc_sk(sock->sk);
 915
 916         return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
 917 }
 918
 919 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 920 {
 921         struct sock *sk = sock->sk;
 922         struct smc_sock *smc;
 923         int rc = -EPIPE;
 924
 925         smc = smc_sk(sk);
 926         lock_sock(sk);
 927         if (sk->sk_state != SMC_ACTIVE)
 928                 goto out;
 929         if (smc->use_fallback)
 930                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
 931         else
 932                 rc = smc_tx_sendmsg(smc, msg, len);
 933 out:
 934         release_sock(sk);
 935         return rc;
 936 }
 937
 938 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
 939                        int flags)
 940 {
 941         struct sock *sk = sock->sk;
 942         struct smc_sock *smc;
 943         int rc = -ENOTCONN;
 944
 945         smc = smc_sk(sk);
 946         lock_sock(sk);
 947         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
 948                 goto out;
 949
 950         if (smc->use_fallback)
 951                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
 952         else
 953                 rc = sock_no_recvmsg(sock, msg, len, flags);
 954 out:
 955         release_sock(sk);
 956         return rc;
 957 }
 958
 959 static unsigned int smc_accept_poll(struct sock *parent)
 960 {
 961         struct smc_sock *isk;
 962         struct sock *sk;
 963
 964         lock_sock(parent);
 965         list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
 966                 sk = (struct sock *)isk;
 967
 968                 if (sk->sk_state == SMC_ACTIVE) {
 969                         release_sock(parent);
 970                         return POLLIN | POLLRDNORM;
 971                 }
 972         }
 973         release_sock(parent);
 974
 975         return 0;
 976 }
 977
 978 static unsigned int smc_poll(struct file *file, struct socket *sock,
 979                              poll_table *wait)
 980 {
 981         struct sock *sk = sock->sk;
 982         unsigned int mask = 0;
 983         struct smc_sock *smc;
 984         int rc;
 985
 986         smc = smc_sk(sock->sk);
 987         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
 988                 /* delegate to CLC child sock */
 989                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
 990                 /* if non-blocking connect finished ... */
 991                 lock_sock(sk);
 992                 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
 993                         sk->sk_err = smc->clcsock->sk->sk_err;
 994                         if (sk->sk_err) {
 995                                 mask |= POLLERR;
 996                         } else {
 997                                 rc = smc_connect_rdma(smc);
 998                                 if (rc < 0)
 999                                         mask |= POLLERR;
1000                                 else
1001                                         /* success cases including fallback */
1002                                         mask |= POLLOUT | POLLWRNORM;
1003                         }
1004                 }
1005                 release_sock(sk);
1006         } else {
1007                 sock_poll_wait(file, sk_sleep(sk), wait);
1008                 if (sk->sk_state == SMC_LISTEN)
1009                         /* woken up by sk_data_ready in smc_listen_work() */
1010                         mask |= smc_accept_poll(sk);
1011                 if (sk->sk_err)
1012                         mask |= POLLERR;
1013                 if (atomic_read(&smc->conn.sndbuf_space)) {
1014                         mask |= POLLOUT | POLLWRNORM;
1015                 } else {
1016                         sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1017                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1018                 }
1019                 /* for now - to be enhanced in follow-on patch */
1020         }
1021
1022         return mask;
1023 }
1024
1025 static int smc_shutdown(struct socket *sock, int how)
1026 {
1027         struct sock *sk = sock->sk;
1028         struct smc_sock *smc;
1029         int rc = -EINVAL;
1030
1031         smc = smc_sk(sk);
1032
1033         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1034                 goto out_err;
1035
1036         lock_sock(sk);
1037
1038         rc = -ENOTCONN;
1039         if (sk->sk_state == SMC_CLOSED)
1040                 goto out;
1041         if (smc->use_fallback) {
1042                 rc = kernel_sock_shutdown(smc->clcsock, how);
1043                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1044                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1045                         sk->sk_state = SMC_CLOSED;
1046         } else {
1047                 rc = sock_no_shutdown(sock, how);
1048         }
1049
1050 out:
1051         release_sock(sk);
1052
1053 out_err:
1054         return rc;
1055 }
1056
1057 static int smc_setsockopt(struct socket *sock, int level, int optname,
1058                           char __user *optval, unsigned int optlen)
1059 {
1060         struct sock *sk = sock->sk;
1061         struct smc_sock *smc;
1062
1063         smc = smc_sk(sk);
1064
1065         /* generic setsockopts reaching us here always apply to the
1066          * CLC socket
1067          */
1068         return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1069                                              optval, optlen);
1070 }
1071
1072 static int smc_getsockopt(struct socket *sock, int level, int optname,
1073                           char __user *optval, int __user *optlen)
1074 {
1075         struct smc_sock *smc;
1076
1077         smc = smc_sk(sock->sk);
1078         /* socket options apply to the CLC socket */
1079         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1080                                              optval, optlen);
1081 }
1082
1083 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1084                      unsigned long arg)
1085 {
1086         struct smc_sock *smc;
1087
1088         smc = smc_sk(sock->sk);
1089         if (smc->use_fallback)
1090                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1091         else
1092                 return sock_no_ioctl(sock, cmd, arg);
1093 }
1094
1095 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1096                             int offset, size_t size, int flags)
1097 {
1098         struct sock *sk = sock->sk;
1099         struct smc_sock *smc;
1100         int rc = -EPIPE;
1101
1102         smc = smc_sk(sk);
1103         lock_sock(sk);
1104         if (sk->sk_state != SMC_ACTIVE)
1105                 goto out;
1106         if (smc->use_fallback)
1107                 rc = kernel_sendpage(smc->clcsock, page, offset,
1108                                      size, flags);
1109         else
1110                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1111
1112 out:
1113         release_sock(sk);
1114         return rc;
1115 }
1116
1117 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1118                                struct pipe_inode_info *pipe, size_t len,
1119                                     unsigned int flags)
1120 {
1121         struct sock *sk = sock->sk;
1122         struct smc_sock *smc;
1123         int rc = -ENOTCONN;
1124
1125         smc = smc_sk(sk);
1126         lock_sock(sk);
1127         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1128                 goto out;
1129         if (smc->use_fallback) {
1130                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1131                                                     pipe, len, flags);
1132         } else {
1133                 rc = -EOPNOTSUPP;
1134         }
1135 out:
1136         release_sock(sk);
1137         return rc;
1138 }
1139
1140 /* must look like tcp */
1141 static const struct proto_ops smc_sock_ops = {
1142         .family         = PF_SMC,
1143         .owner          = THIS_MODULE,
1144         .release        = smc_release,
1145         .bind           = smc_bind,
1146         .connect        = smc_connect,
1147         .socketpair     = sock_no_socketpair,
1148         .accept         = smc_accept,
1149         .getname        = smc_getname,
1150         .poll           = smc_poll,
1151         .ioctl          = smc_ioctl,
1152         .listen         = smc_listen,
1153         .shutdown       = smc_shutdown,
1154         .setsockopt     = smc_setsockopt,
1155         .getsockopt     = smc_getsockopt,
1156         .sendmsg        = smc_sendmsg,
1157         .recvmsg        = smc_recvmsg,
1158         .mmap           = sock_no_mmap,
1159         .sendpage       = smc_sendpage,
1160         .splice_read    = smc_splice_read,
1161 };
1162
1163 static int smc_create(struct net *net, struct socket *sock, int protocol,
1164                       int kern)
1165 {
1166         struct smc_sock *smc;
1167         struct sock *sk;
1168         int rc;
1169
1170         rc = -ESOCKTNOSUPPORT;
1171         if (sock->type != SOCK_STREAM)
1172                 goto out;
1173
1174         rc = -EPROTONOSUPPORT;
1175         if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1176                 goto out;
1177
1178         rc = -ENOBUFS;
1179         sock->ops = &smc_sock_ops;
1180         sk = smc_sock_alloc(net, sock);
1181         if (!sk)
1182                 goto out;
1183
1184         /* create internal TCP socket for CLC handshake and fallback */
1185         smc = smc_sk(sk);
1186         smc->use_fallback = false; /* assume rdma capability first */
1187         rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1188                               IPPROTO_TCP, &smc->clcsock);
1189         if (rc)
1190                 sk_common_release(sk);
1191         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1192         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1193
1194 out:
1195         return rc;
1196 }
1197
1198 static const struct net_proto_family smc_sock_family_ops = {
1199         .family = PF_SMC,
1200         .owner  = THIS_MODULE,
1201         .create = smc_create,
1202 };
1203
1204 static int __init smc_init(void)
1205 {
1206         int rc;
1207
1208         rc = smc_pnet_init();
1209         if (rc)
1210                 return rc;
1211
1212         rc = smc_llc_init();
1213         if (rc) {
1214                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1215                 goto out_pnet;
1216         }
1217
1218         rc = smc_cdc_init();
1219         if (rc) {
1220                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1221                 goto out_pnet;
1222         }
1223
1224         rc = proto_register(&smc_proto, 1);
1225         if (rc) {
1226                 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1227                 goto out_pnet;
1228         }
1229
1230         rc = sock_register(&smc_sock_family_ops);
1231         if (rc) {
1232                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1233                 goto out_proto;
1234         }
1235
1236         rc = smc_ib_register_client();
1237         if (rc) {
1238                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1239                 goto out_sock;
1240         }
1241
1242         return 0;
1243
1244 out_sock:
1245         sock_unregister(PF_SMC);
1246 out_proto:
1247         proto_unregister(&smc_proto);
1248 out_pnet:
1249         smc_pnet_exit();
1250         return rc;
1251 }
1252
1253 static void __exit smc_exit(void)
1254 {
1255         struct smc_link_group *lgr, *lg;
1256         LIST_HEAD(lgr_freeing_list);
1257
1258         spin_lock_bh(&smc_lgr_list.lock);
1259         if (!list_empty(&smc_lgr_list.list))
1260                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1261         spin_unlock_bh(&smc_lgr_list.lock);
1262         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1263                 list_del_init(&lgr->list);
1264                 smc_lgr_free(lgr); /* free link group */
1265         }
1266         smc_ib_unregister_client();
1267         sock_unregister(PF_SMC);
1268         proto_unregister(&smc_proto);
1269         smc_pnet_exit();
1270 }
1271
1272 module_init(smc_init);
1273 module_exit(smc_exit);
1274
1275 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1276 MODULE_DESCRIPTION("smc socket address family");
1277 MODULE_LICENSE("GPL");
1278 MODULE_ALIAS_NETPROTO(PF_SMC);