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 <net/sock.h>
  30 #include <net/tcp.h>
  31
  32 #include "smc.h"
  33 #include "smc_clc.h"
  34 #include "smc_ib.h"
  35 #include "smc_pnet.h"
  36
  37 static void smc_tcp_listen_work(struct work_struct *);
  38
  39 static void smc_set_keepalive(struct sock *sk, int val)
  40 {
  41         struct smc_sock *smc = smc_sk(sk);
  42
  43         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  44 }
  45
  46 static struct proto smc_proto = {
  47         .name           = "SMC",
  48         .owner          = THIS_MODULE,
  49         .keepalive      = smc_set_keepalive,
  50         .obj_size       = sizeof(struct smc_sock),
  51         .slab_flags     = SLAB_DESTROY_BY_RCU,
  52 };
  53
  54 static int smc_release(struct socket *sock)
  55 {
  56         struct sock *sk = sock->sk;
  57         struct smc_sock *smc;
  58
  59         if (!sk)
  60                 goto out;
  61
  62         smc = smc_sk(sk);
  63         lock_sock(sk);
  64
  65         sk->sk_state = SMC_CLOSED;
  66         if (smc->clcsock) {
  67                 sock_release(smc->clcsock);
  68                 smc->clcsock = NULL;
  69         }
  70
  71         /* detach socket */
  72         sock_orphan(sk);
  73         sock->sk = NULL;
  74         release_sock(sk);
  75
  76         sock_put(sk);
  77 out:
  78         return 0;
  79 }
  80
  81 static void smc_destruct(struct sock *sk)
  82 {
  83         if (sk->sk_state != SMC_CLOSED)
  84                 return;
  85         if (!sock_flag(sk, SOCK_DEAD))
  86                 return;
  87
  88         sk_refcnt_debug_dec(sk);
  89 }
  90
  91 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
  92 {
  93         struct smc_sock *smc;
  94         struct sock *sk;
  95
  96         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
  97         if (!sk)
  98                 return NULL;
  99
 100         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 101         sk->sk_state = SMC_INIT;
 102         sk->sk_destruct = smc_destruct;
 103         sk->sk_protocol = SMCPROTO_SMC;
 104         smc = smc_sk(sk);
 105         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 106         INIT_LIST_HEAD(&smc->accept_q);
 107         spin_lock_init(&smc->accept_q_lock);
 108         sk_refcnt_debug_inc(sk);
 109
 110         return sk;
 111 }
 112
 113 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 114                     int addr_len)
 115 {
 116         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 117         struct sock *sk = sock->sk;
 118         struct smc_sock *smc;
 119         int rc;
 120
 121         smc = smc_sk(sk);
 122
 123         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 124         rc = -EINVAL;
 125         if (addr_len < sizeof(struct sockaddr_in))
 126                 goto out;
 127
 128         rc = -EAFNOSUPPORT;
 129         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 130         if ((addr->sin_family != AF_INET) &&
 131             ((addr->sin_family != AF_UNSPEC) ||
 132              (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
 133                 goto out;
 134
 135         lock_sock(sk);
 136
 137         /* Check if socket is already active */
 138         rc = -EINVAL;
 139         if (sk->sk_state != SMC_INIT)
 140                 goto out_rel;
 141
 142         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 143         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 144
 145 out_rel:
 146         release_sock(sk);
 147 out:
 148         return rc;
 149 }
 150
 151 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 152                                    unsigned long mask)
 153 {
 154         /* options we don't get control via setsockopt for */
 155         nsk->sk_type = osk->sk_type;
 156         nsk->sk_sndbuf = osk->sk_sndbuf;
 157         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 158         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 159         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 160         nsk->sk_mark = osk->sk_mark;
 161         nsk->sk_priority = osk->sk_priority;
 162         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 163         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 164         nsk->sk_err = osk->sk_err;
 165
 166         nsk->sk_flags &= ~mask;
 167         nsk->sk_flags |= osk->sk_flags & mask;
 168 }
 169
 170 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 171                              (1UL << SOCK_KEEPOPEN) | \
 172                              (1UL << SOCK_LINGER) | \
 173                              (1UL << SOCK_BROADCAST) | \
 174                              (1UL << SOCK_TIMESTAMP) | \
 175                              (1UL << SOCK_DBG) | \
 176                              (1UL << SOCK_RCVTSTAMP) | \
 177                              (1UL << SOCK_RCVTSTAMPNS) | \
 178                              (1UL << SOCK_LOCALROUTE) | \
 179                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 180                              (1UL << SOCK_RXQ_OVFL) | \
 181                              (1UL << SOCK_WIFI_STATUS) | \
 182                              (1UL << SOCK_NOFCS) | \
 183                              (1UL << SOCK_FILTER_LOCKED))
 184 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 185  * clc socket (since smc is not called for these options from net/core)
 186  */
 187 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 188 {
 189         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 190 }
 191
 192 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 193                              (1UL << SOCK_KEEPOPEN) | \
 194                              (1UL << SOCK_LINGER) | \
 195                              (1UL << SOCK_DBG))
 196 /* copy only settings and flags relevant for smc from clc to smc socket */
 197 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 198 {
 199         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 200 }
 201
 202 /* determine subnet and mask of internal TCP socket */
 203 int smc_netinfo_by_tcpsk(struct socket *clcsock,
 204                          __be32 *subnet, u8 *prefix_len)
 205 {
 206         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 207         struct sockaddr_in addr;
 208         int rc = -ENOENT;
 209         int len;
 210
 211         if (!dst) {
 212                 rc = -ENOTCONN;
 213                 goto out;
 214         }
 215         if (!dst->dev) {
 216                 rc = -ENODEV;
 217                 goto out_rel;
 218         }
 219
 220         /* get address to which the internal TCP socket is bound */
 221         kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
 222         /* analyze IPv4 specific data of net_device belonging to TCP socket */
 223         for_ifa(dst->dev->ip_ptr) {
 224                 if (ifa->ifa_address != addr.sin_addr.s_addr)
 225                         continue;
 226                 *prefix_len = inet_mask_len(ifa->ifa_mask);
 227                 *subnet = ifa->ifa_address & ifa->ifa_mask;
 228                 rc = 0;
 229                 break;
 230         } endfor_ifa(dst->dev->ip_ptr);
 231
 232 out_rel:
 233         dst_release(dst);
 234 out:
 235         return rc;
 236 }
 237
 238 /* setup for RDMA connection of client */
 239 static int smc_connect_rdma(struct smc_sock *smc)
 240 {
 241         struct smc_clc_msg_accept_confirm aclc;
 242         struct smc_ib_device *smcibdev;
 243         int reason_code = 0;
 244         int rc = 0;
 245         u8 ibport;
 246
 247         /* IPSec connections opt out of SMC-R optimizations */
 248         if (using_ipsec(smc)) {
 249                 reason_code = SMC_CLC_DECL_IPSEC;
 250                 goto decline_rdma;
 251         }
 252
 253         /* PNET table look up: search active ib_device and port
 254          * within same PNETID that also contains the ethernet device
 255          * used for the internal TCP socket
 256          */
 257         smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
 258         if (!smcibdev) {
 259                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 260                 goto decline_rdma;
 261         }
 262
 263         /* do inband token exchange */
 264         reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
 265         if (reason_code < 0) {
 266                 rc = reason_code;
 267                 goto out_err;
 268         }
 269         if (reason_code > 0) /* configuration error */
 270                 goto decline_rdma;
 271         /* receive SMC Accept CLC message */
 272         reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
 273                                        SMC_CLC_ACCEPT);
 274         if (reason_code < 0) {
 275                 rc = reason_code;
 276                 goto out_err;
 277         }
 278         if (reason_code > 0)
 279                 goto decline_rdma;
 280
 281         /* tbd in follow-on patch: more steps to setup RDMA communcication,
 282          * create connection, link group, link
 283          */
 284
 285         /* tbd in follow-on patch: more steps to setup RDMA communcication,
 286          * create rmbs, map rmbs, rtoken_handling, modify_qp
 287          */
 288
 289         rc = smc_clc_send_confirm(smc);
 290         if (rc)
 291                 goto out_err;
 292
 293         /* tbd in follow-on patch: llc_confirm */
 294
 295 out_connected:
 296         smc_copy_sock_settings_to_clc(smc);
 297         smc->sk.sk_state = SMC_ACTIVE;
 298
 299         return rc;
 300
 301 decline_rdma:
 302         /* RDMA setup failed, switch back to TCP */
 303         smc->use_fallback = true;
 304         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 305                 rc = smc_clc_send_decline(smc, reason_code, 0);
 306                 if (rc < sizeof(struct smc_clc_msg_decline))
 307                         goto out_err;
 308         }
 309         goto out_connected;
 310
 311 out_err:
 312         return rc;
 313 }
 314
 315 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 316                        int alen, int flags)
 317 {
 318         struct sock *sk = sock->sk;
 319         struct smc_sock *smc;
 320         int rc = -EINVAL;
 321
 322         smc = smc_sk(sk);
 323
 324         /* separate smc parameter checking to be safe */
 325         if (alen < sizeof(addr->sa_family))
 326                 goto out_err;
 327         if (addr->sa_family != AF_INET)
 328                 goto out_err;
 329         smc->addr = addr;       /* needed for nonblocking connect */
 330
 331         lock_sock(sk);
 332         switch (sk->sk_state) {
 333         default:
 334                 goto out;
 335         case SMC_ACTIVE:
 336                 rc = -EISCONN;
 337                 goto out;
 338         case SMC_INIT:
 339                 rc = 0;
 340                 break;
 341         }
 342
 343         smc_copy_sock_settings_to_clc(smc);
 344         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 345         if (rc)
 346                 goto out;
 347
 348         /* setup RDMA connection */
 349         rc = smc_connect_rdma(smc);
 350         if (rc < 0)
 351                 goto out;
 352         else
 353                 rc = 0; /* success cases including fallback */
 354
 355 out:
 356         release_sock(sk);
 357 out_err:
 358         return rc;
 359 }
 360
 361 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 362 {
 363         struct sock *sk = &lsmc->sk;
 364         struct socket *new_clcsock;
 365         struct sock *new_sk;
 366         int rc;
 367
 368         release_sock(&lsmc->sk);
 369         new_sk = smc_sock_alloc(sock_net(sk), NULL);
 370         if (!new_sk) {
 371                 rc = -ENOMEM;
 372                 lsmc->sk.sk_err = ENOMEM;
 373                 *new_smc = NULL;
 374                 lock_sock(&lsmc->sk);
 375                 goto out;
 376         }
 377         *new_smc = smc_sk(new_sk);
 378
 379         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 380         lock_sock(&lsmc->sk);
 381         if  (rc < 0) {
 382                 lsmc->sk.sk_err = -rc;
 383                 new_sk->sk_state = SMC_CLOSED;
 384                 sock_set_flag(new_sk, SOCK_DEAD);
 385                 sock_put(new_sk);
 386                 *new_smc = NULL;
 387                 goto out;
 388         }
 389         if (lsmc->sk.sk_state == SMC_CLOSED) {
 390                 if (new_clcsock)
 391                         sock_release(new_clcsock);
 392                 new_sk->sk_state = SMC_CLOSED;
 393                 sock_set_flag(new_sk, SOCK_DEAD);
 394                 sock_put(new_sk);
 395                 *new_smc = NULL;
 396                 goto out;
 397         }
 398
 399         (*new_smc)->clcsock = new_clcsock;
 400 out:
 401         return rc;
 402 }
 403
 404 /* add a just created sock to the accept queue of the listen sock as
 405  * candidate for a following socket accept call from user space
 406  */
 407 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 408 {
 409         struct smc_sock *par = smc_sk(parent);
 410
 411         sock_hold(sk);
 412         spin_lock(&par->accept_q_lock);
 413         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 414         spin_unlock(&par->accept_q_lock);
 415         sk_acceptq_added(parent);
 416 }
 417
 418 /* remove a socket from the accept queue of its parental listening socket */
 419 static void smc_accept_unlink(struct sock *sk)
 420 {
 421         struct smc_sock *par = smc_sk(sk)->listen_smc;
 422
 423         spin_lock(&par->accept_q_lock);
 424         list_del_init(&smc_sk(sk)->accept_q);
 425         spin_unlock(&par->accept_q_lock);
 426         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 427         sock_put(sk);
 428 }
 429
 430 /* remove a sock from the accept queue to bind it to a new socket created
 431  * for a socket accept call from user space
 432  */
 433 static struct sock *smc_accept_dequeue(struct sock *parent,
 434                                        struct socket *new_sock)
 435 {
 436         struct smc_sock *isk, *n;
 437         struct sock *new_sk;
 438
 439         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 440                 new_sk = (struct sock *)isk;
 441
 442                 smc_accept_unlink(new_sk);
 443                 if (new_sk->sk_state == SMC_CLOSED) {
 444                         /* tbd in follow-on patch: close this sock */
 445                         continue;
 446                 }
 447                 if (new_sock)
 448                         sock_graft(new_sk, new_sock);
 449                 return new_sk;
 450         }
 451         return NULL;
 452 }
 453
 454 /* clean up for a created but never accepted sock */
 455 static void smc_close_non_accepted(struct sock *sk)
 456 {
 457         struct smc_sock *smc = smc_sk(sk);
 458
 459         sock_hold(sk);
 460         if (smc->clcsock) {
 461                 struct socket *tcp;
 462
 463                 tcp = smc->clcsock;
 464                 smc->clcsock = NULL;
 465                 sock_release(tcp);
 466         }
 467         /* more closing stuff to be added with socket closing patch */
 468         sock_put(sk);
 469 }
 470
 471 /* setup for RDMA connection of server */
 472 static void smc_listen_work(struct work_struct *work)
 473 {
 474         struct smc_sock *new_smc = container_of(work, struct smc_sock,
 475                                                 smc_listen_work);
 476         struct socket *newclcsock = new_smc->clcsock;
 477         struct smc_sock *lsmc = new_smc->listen_smc;
 478         struct smc_clc_msg_accept_confirm cclc;
 479         struct sock *newsmcsk = &new_smc->sk;
 480         struct smc_clc_msg_proposal pclc;
 481         struct smc_ib_device *smcibdev;
 482         struct sockaddr_in peeraddr;
 483         int reason_code = 0;
 484         int rc = 0, len;
 485         __be32 subnet;
 486         u8 prefix_len;
 487         u8 ibport;
 488
 489         /* do inband token exchange -
 490          *wait for and receive SMC Proposal CLC message
 491          */
 492         reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
 493                                        SMC_CLC_PROPOSAL);
 494         if (reason_code < 0)
 495                 goto out_err;
 496         if (reason_code > 0)
 497                 goto decline_rdma;
 498
 499         /* IPSec connections opt out of SMC-R optimizations */
 500         if (using_ipsec(new_smc)) {
 501                 reason_code = SMC_CLC_DECL_IPSEC;
 502                 goto decline_rdma;
 503         }
 504
 505         /* PNET table look up: search active ib_device and port
 506          * within same PNETID that also contains the ethernet device
 507          * used for the internal TCP socket
 508          */
 509         smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
 510         if (!smcibdev) {
 511                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 512                 goto decline_rdma;
 513         }
 514
 515         /* determine subnet and mask from internal TCP socket */
 516         rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
 517         if (rc) {
 518                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 519                 goto decline_rdma;
 520         }
 521         if ((pclc.outgoing_subnet != subnet) ||
 522             (pclc.prefix_len != prefix_len)) {
 523                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 524                 goto decline_rdma;
 525         }
 526
 527         /* get address of the peer connected to the internal TCP socket */
 528         kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
 529
 530         /* tbd in follow-on patch: more steps to setup RDMA communcication,
 531          * create connection, link_group, link
 532          */
 533
 534         /* tbd in follow-on patch: more steps to setup RDMA communcication,
 535          * create rmbs, map rmbs
 536          */
 537
 538         rc = smc_clc_send_accept(new_smc);
 539         if (rc)
 540                 goto out_err;
 541
 542         /* receive SMC Confirm CLC message */
 543         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
 544                                        SMC_CLC_CONFIRM);
 545         if (reason_code < 0)
 546                 goto out_err;
 547         if (reason_code > 0)
 548                 goto decline_rdma;
 549
 550         /* tbd in follow-on patch: more steps to setup RDMA communcication,
 551          * rtoken_handling, modify_qp
 552          */
 553
 554 out_connected:
 555         sk_refcnt_debug_inc(newsmcsk);
 556         newsmcsk->sk_state = SMC_ACTIVE;
 557 enqueue:
 558         lock_sock(&lsmc->sk);
 559         if (lsmc->sk.sk_state == SMC_LISTEN) {
 560                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
 561         } else { /* no longer listening */
 562                 smc_close_non_accepted(newsmcsk);
 563         }
 564         release_sock(&lsmc->sk);
 565
 566         /* Wake up accept */
 567         lsmc->sk.sk_data_ready(&lsmc->sk);
 568         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
 569         return;
 570
 571 decline_rdma:
 572         /* RDMA setup failed, switch back to TCP */
 573         new_smc->use_fallback = true;
 574         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 575                 rc = smc_clc_send_decline(new_smc, reason_code, 0);
 576                 if (rc < sizeof(struct smc_clc_msg_decline))
 577                         goto out_err;
 578         }
 579         goto out_connected;
 580
 581 out_err:
 582         newsmcsk->sk_state = SMC_CLOSED;
 583         goto enqueue; /* queue new sock with sk_err set */
 584 }
 585
 586 static void smc_tcp_listen_work(struct work_struct *work)
 587 {
 588         struct smc_sock *lsmc = container_of(work, struct smc_sock,
 589                                              tcp_listen_work);
 590         struct smc_sock *new_smc;
 591         int rc = 0;
 592
 593         lock_sock(&lsmc->sk);
 594         while (lsmc->sk.sk_state == SMC_LISTEN) {
 595                 rc = smc_clcsock_accept(lsmc, &new_smc);
 596                 if (rc)
 597                         goto out;
 598                 if (!new_smc)
 599                         continue;
 600
 601                 new_smc->listen_smc = lsmc;
 602                 new_smc->use_fallback = false; /* assume rdma capability first*/
 603                 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
 604                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
 605                 smc_copy_sock_settings_to_smc(new_smc);
 606                 schedule_work(&new_smc->smc_listen_work);
 607         }
 608
 609 out:
 610         release_sock(&lsmc->sk);
 611         lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
 612 }
 613
 614 static int smc_listen(struct socket *sock, int backlog)
 615 {
 616         struct sock *sk = sock->sk;
 617         struct smc_sock *smc;
 618         int rc;
 619
 620         smc = smc_sk(sk);
 621         lock_sock(sk);
 622
 623         rc = -EINVAL;
 624         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
 625                 goto out;
 626
 627         rc = 0;
 628         if (sk->sk_state == SMC_LISTEN) {
 629                 sk->sk_max_ack_backlog = backlog;
 630                 goto out;
 631         }
 632         /* some socket options are handled in core, so we could not apply
 633          * them to the clc socket -- copy smc socket options to clc socket
 634          */
 635         smc_copy_sock_settings_to_clc(smc);
 636
 637         rc = kernel_listen(smc->clcsock, backlog);
 638         if (rc)
 639                 goto out;
 640         sk->sk_max_ack_backlog = backlog;
 641         sk->sk_ack_backlog = 0;
 642         sk->sk_state = SMC_LISTEN;
 643         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 644         schedule_work(&smc->tcp_listen_work);
 645
 646 out:
 647         release_sock(sk);
 648         return rc;
 649 }
 650
 651 static int smc_accept(struct socket *sock, struct socket *new_sock,
 652                       int flags)
 653 {
 654         struct sock *sk = sock->sk, *nsk;
 655         DECLARE_WAITQUEUE(wait, current);
 656         struct smc_sock *lsmc;
 657         long timeo;
 658         int rc = 0;
 659
 660         lsmc = smc_sk(sk);
 661         lock_sock(sk);
 662
 663         if (lsmc->sk.sk_state != SMC_LISTEN) {
 664                 rc = -EINVAL;
 665                 goto out;
 666         }
 667
 668         /* Wait for an incoming connection */
 669         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 670         add_wait_queue_exclusive(sk_sleep(sk), &wait);
 671         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
 672                 set_current_state(TASK_INTERRUPTIBLE);
 673                 if (!timeo) {
 674                         rc = -EAGAIN;
 675                         break;
 676                 }
 677                 release_sock(sk);
 678                 timeo = schedule_timeout(timeo);
 679                 /* wakeup by sk_data_ready in smc_listen_work() */
 680                 sched_annotate_sleep();
 681                 lock_sock(sk);
 682                 if (signal_pending(current)) {
 683                         rc = sock_intr_errno(timeo);
 684                         break;
 685                 }
 686         }
 687         set_current_state(TASK_RUNNING);
 688         remove_wait_queue(sk_sleep(sk), &wait);
 689
 690         if (!rc)
 691                 rc = sock_error(nsk);
 692
 693 out:
 694         release_sock(sk);
 695         return rc;
 696 }
 697
 698 static int smc_getname(struct socket *sock, struct sockaddr *addr,
 699                        int *len, int peer)
 700 {
 701         struct smc_sock *smc;
 702
 703         if (peer && (sock->sk->sk_state != SMC_ACTIVE))
 704                 return -ENOTCONN;
 705
 706         smc = smc_sk(sock->sk);
 707
 708         return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
 709 }
 710
 711 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
 712 {
 713         struct sock *sk = sock->sk;
 714         struct smc_sock *smc;
 715         int rc = -EPIPE;
 716
 717         smc = smc_sk(sk);
 718         lock_sock(sk);
 719         if (sk->sk_state != SMC_ACTIVE)
 720                 goto out;
 721         if (smc->use_fallback)
 722                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
 723         else
 724                 rc = sock_no_sendmsg(sock, msg, len);
 725 out:
 726         release_sock(sk);
 727         return rc;
 728 }
 729
 730 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
 731                        int flags)
 732 {
 733         struct sock *sk = sock->sk;
 734         struct smc_sock *smc;
 735         int rc = -ENOTCONN;
 736
 737         smc = smc_sk(sk);
 738         lock_sock(sk);
 739         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
 740                 goto out;
 741
 742         if (smc->use_fallback)
 743                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
 744         else
 745                 rc = sock_no_recvmsg(sock, msg, len, flags);
 746 out:
 747         release_sock(sk);
 748         return rc;
 749 }
 750
 751 static unsigned int smc_accept_poll(struct sock *parent)
 752 {
 753         struct smc_sock *isk;
 754         struct sock *sk;
 755
 756         lock_sock(parent);
 757         list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
 758                 sk = (struct sock *)isk;
 759
 760                 if (sk->sk_state == SMC_ACTIVE) {
 761                         release_sock(parent);
 762                         return POLLIN | POLLRDNORM;
 763                 }
 764         }
 765         release_sock(parent);
 766
 767         return 0;
 768 }
 769
 770 static unsigned int smc_poll(struct file *file, struct socket *sock,
 771                              poll_table *wait)
 772 {
 773         struct sock *sk = sock->sk;
 774         unsigned int mask = 0;
 775         struct smc_sock *smc;
 776         int rc;
 777
 778         smc = smc_sk(sock->sk);
 779         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
 780                 /* delegate to CLC child sock */
 781                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
 782                 /* if non-blocking connect finished ... */
 783                 lock_sock(sk);
 784                 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
 785                         sk->sk_err = smc->clcsock->sk->sk_err;
 786                         if (sk->sk_err) {
 787                                 mask |= POLLERR;
 788                         } else {
 789                                 rc = smc_connect_rdma(smc);
 790                                 if (rc < 0)
 791                                         mask |= POLLERR;
 792                                 else
 793                                         /* success cases including fallback */
 794                                         mask |= POLLOUT | POLLWRNORM;
 795                         }
 796                 }
 797                 release_sock(sk);
 798         } else {
 799                 sock_poll_wait(file, sk_sleep(sk), wait);
 800                 if (sk->sk_state == SMC_LISTEN)
 801                         /* woken up by sk_data_ready in smc_listen_work() */
 802                         mask |= smc_accept_poll(sk);
 803                 if (sk->sk_err)
 804                         mask |= POLLERR;
 805                 /* for now - to be enhanced in follow-on patch */
 806         }
 807
 808         return mask;
 809 }
 810
 811 static int smc_shutdown(struct socket *sock, int how)
 812 {
 813         struct sock *sk = sock->sk;
 814         struct smc_sock *smc;
 815         int rc = -EINVAL;
 816
 817         smc = smc_sk(sk);
 818
 819         if ((how < SHUT_RD) || (how > SHUT_RDWR))
 820                 goto out_err;
 821
 822         lock_sock(sk);
 823
 824         rc = -ENOTCONN;
 825         if (sk->sk_state == SMC_CLOSED)
 826                 goto out;
 827         if (smc->use_fallback) {
 828                 rc = kernel_sock_shutdown(smc->clcsock, how);
 829                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
 830                 if (sk->sk_shutdown == SHUTDOWN_MASK)
 831                         sk->sk_state = SMC_CLOSED;
 832         } else {
 833                 rc = sock_no_shutdown(sock, how);
 834         }
 835
 836 out:
 837         release_sock(sk);
 838
 839 out_err:
 840         return rc;
 841 }
 842
 843 static int smc_setsockopt(struct socket *sock, int level, int optname,
 844                           char __user *optval, unsigned int optlen)
 845 {
 846         struct sock *sk = sock->sk;
 847         struct smc_sock *smc;
 848
 849         smc = smc_sk(sk);
 850
 851         /* generic setsockopts reaching us here always apply to the
 852          * CLC socket
 853          */
 854         return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
 855                                              optval, optlen);
 856 }
 857
 858 static int smc_getsockopt(struct socket *sock, int level, int optname,
 859                           char __user *optval, int __user *optlen)
 860 {
 861         struct smc_sock *smc;
 862
 863         smc = smc_sk(sock->sk);
 864         /* socket options apply to the CLC socket */
 865         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
 866                                              optval, optlen);
 867 }
 868
 869 static int smc_ioctl(struct socket *sock, unsigned int cmd,
 870                      unsigned long arg)
 871 {
 872         struct smc_sock *smc;
 873
 874         smc = smc_sk(sock->sk);
 875         if (smc->use_fallback)
 876                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
 877         else
 878                 return sock_no_ioctl(sock, cmd, arg);
 879 }
 880
 881 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
 882                             int offset, size_t size, int flags)
 883 {
 884         struct sock *sk = sock->sk;
 885         struct smc_sock *smc;
 886         int rc = -EPIPE;
 887
 888         smc = smc_sk(sk);
 889         lock_sock(sk);
 890         if (sk->sk_state != SMC_ACTIVE)
 891                 goto out;
 892         if (smc->use_fallback)
 893                 rc = kernel_sendpage(smc->clcsock, page, offset,
 894                                      size, flags);
 895         else
 896                 rc = sock_no_sendpage(sock, page, offset, size, flags);
 897
 898 out:
 899         release_sock(sk);
 900         return rc;
 901 }
 902
 903 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
 904                                struct pipe_inode_info *pipe, size_t len,
 905                                     unsigned int flags)
 906 {
 907         struct sock *sk = sock->sk;
 908         struct smc_sock *smc;
 909         int rc = -ENOTCONN;
 910
 911         smc = smc_sk(sk);
 912         lock_sock(sk);
 913         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
 914                 goto out;
 915         if (smc->use_fallback) {
 916                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
 917                                                     pipe, len, flags);
 918         } else {
 919                 rc = -EOPNOTSUPP;
 920         }
 921 out:
 922         release_sock(sk);
 923         return rc;
 924 }
 925
 926 /* must look like tcp */
 927 static const struct proto_ops smc_sock_ops = {
 928         .family         = PF_SMC,
 929         .owner          = THIS_MODULE,
 930         .release        = smc_release,
 931         .bind           = smc_bind,
 932         .connect        = smc_connect,
 933         .socketpair     = sock_no_socketpair,
 934         .accept         = smc_accept,
 935         .getname        = smc_getname,
 936         .poll           = smc_poll,
 937         .ioctl          = smc_ioctl,
 938         .listen         = smc_listen,
 939         .shutdown       = smc_shutdown,
 940         .setsockopt     = smc_setsockopt,
 941         .getsockopt     = smc_getsockopt,
 942         .sendmsg        = smc_sendmsg,
 943         .recvmsg        = smc_recvmsg,
 944         .mmap           = sock_no_mmap,
 945         .sendpage       = smc_sendpage,
 946         .splice_read    = smc_splice_read,
 947 };
 948
 949 static int smc_create(struct net *net, struct socket *sock, int protocol,
 950                       int kern)
 951 {
 952         struct smc_sock *smc;
 953         struct sock *sk;
 954         int rc;
 955
 956         rc = -ESOCKTNOSUPPORT;
 957         if (sock->type != SOCK_STREAM)
 958                 goto out;
 959
 960         rc = -EPROTONOSUPPORT;
 961         if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
 962                 goto out;
 963
 964         rc = -ENOBUFS;
 965         sock->ops = &smc_sock_ops;
 966         sk = smc_sock_alloc(net, sock);
 967         if (!sk)
 968                 goto out;
 969
 970         /* create internal TCP socket for CLC handshake and fallback */
 971         smc = smc_sk(sk);
 972         smc->use_fallback = false; /* assume rdma capability first */
 973         rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
 974                               IPPROTO_TCP, &smc->clcsock);
 975         if (rc)
 976                 sk_common_release(sk);
 977
 978 out:
 979         return rc;
 980 }
 981
 982 static const struct net_proto_family smc_sock_family_ops = {
 983         .family = PF_SMC,
 984         .owner  = THIS_MODULE,
 985         .create = smc_create,
 986 };
 987
 988 static int __init smc_init(void)
 989 {
 990         int rc;
 991
 992         rc = smc_pnet_init();
 993         if (rc)
 994                 return rc;
 995
 996         rc = proto_register(&smc_proto, 1);
 997         if (rc) {
 998                 pr_err("%s: proto_register fails with %d\n", __func__, rc);
 999                 goto out_pnet;
1000         }
1001
1002         rc = sock_register(&smc_sock_family_ops);
1003         if (rc) {
1004                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1005                 goto out_proto;
1006         }
1007
1008         rc = smc_ib_register_client();
1009         if (rc) {
1010                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1011                 goto out_sock;
1012         }
1013
1014         return 0;
1015
1016 out_sock:
1017         sock_unregister(PF_SMC);
1018 out_proto:
1019         proto_unregister(&smc_proto);
1020 out_pnet:
1021         smc_pnet_exit();
1022         return rc;
1023 }
1024
1025 static void __exit smc_exit(void)
1026 {
1027         smc_ib_unregister_client();
1028         sock_unregister(PF_SMC);
1029         proto_unregister(&smc_proto);
1030         smc_pnet_exit();
1031 }
1032
1033 module_init(smc_init);
1034 module_exit(smc_exit);
1035
1036 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1037 MODULE_DESCRIPTION("smc socket address family");
1038 MODULE_LICENSE("GPL");
1039 MODULE_ALIAS_NETPROTO(PF_SMC);