sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "ipoib.h"
  36
  37 static  int ipoib_resolvemulti(struct ifnet *, struct sockaddr **,
  38                 struct sockaddr *);
  39
  40
  41 #include <linux/module.h>
  42
  43 #include <linux/slab.h>
  44 #include <linux/kernel.h>
  45 #include <linux/vmalloc.h>
  46
  47 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  48 #include <linux/if_vlan.h>
  49 #include <net/ip.h>
  50 #include <net/ipv6.h>
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
  57 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
  58
  59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  63
  64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  65 int ipoib_debug_level = 1;
  66
  67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  69 #endif
  70
  71 struct ipoib_path_iter {
  72         struct ipoib_dev_priv *priv;
  73         struct ipoib_path  path;
  74 };
  75
  76 static const u8 ipv4_bcast_addr[] = {
  77         0x00, 0xff, 0xff, 0xff,
  78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  80 };
  81
  82 struct workqueue_struct *ipoib_workqueue;
  83
  84 struct ib_sa_client ipoib_sa_client;
  85
  86 static void ipoib_add_one(struct ib_device *device);
  87 static void ipoib_remove_one(struct ib_device *device);
  88 static void ipoib_start(struct ifnet *dev);
  89 static int ipoib_output(struct ifnet *ifp, struct mbuf *m,
  90             const struct sockaddr *dst, struct route *ro);
  91 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
  92 static void ipoib_input(struct ifnet *ifp, struct mbuf *m);
  93
  94 #define IPOIB_MTAP(_ifp, _m)                                    \
  95 do {                                                            \
  96         if (bpf_peers_present((_ifp)->if_bpf)) {                \
  97                 M_ASSERTVALID(_m);                              \
  98                 ipoib_mtap_mb((_ifp), (_m));                    \
  99         }                                                       \
 100 } while (0)
 101
 102 /*
 103  * This is for clients that have an ipoib_header in the mbuf.
 104  */
 105 static void
 106 ipoib_mtap_mb(struct ifnet *ifp, struct mbuf *mb)
 107 {
 108         struct ipoib_header *ih;
 109         struct ether_header eh;
 110
 111         ih = mtod(mb, struct ipoib_header *);
 112         eh.ether_type = ih->proto;
 113         bcopy(ih->hwaddr, &eh.ether_dhost, ETHER_ADDR_LEN);
 114         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 115         mb->m_data += sizeof(struct ipoib_header);
 116         mb->m_len -= sizeof(struct ipoib_header);
 117         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 118         mb->m_data -= sizeof(struct ipoib_header);
 119         mb->m_len += sizeof(struct ipoib_header);
 120 }
 121
 122 void
 123 ipoib_mtap_proto(struct ifnet *ifp, struct mbuf *mb, uint16_t proto)
 124 {
 125         struct ether_header eh;
 126
 127         eh.ether_type = proto;
 128         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 129         bzero(&eh.ether_dhost, ETHER_ADDR_LEN);
 130         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 131 }
 132
 133 static struct ib_client ipoib_client = {
 134         .name   = "ipoib",
 135         .add    = ipoib_add_one,
 136         .remove = ipoib_remove_one
 137 };
 138
 139 int
 140 ipoib_open(struct ipoib_dev_priv *priv)
 141 {
 142         struct ifnet *dev = priv->dev;
 143
 144         ipoib_dbg(priv, "bringing up interface\n");
 145
 146         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 147
 148         if (ipoib_pkey_dev_delay_open(priv))
 149                 return 0;
 150
 151         if (ipoib_ib_dev_open(priv))
 152                 goto err_disable;
 153
 154         if (ipoib_ib_dev_up(priv))
 155                 goto err_stop;
 156
 157         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 158                 struct ipoib_dev_priv *cpriv;
 159
 160                 /* Bring up any child interfaces too */
 161                 mutex_lock(&priv->vlan_mutex);
 162                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 163                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 164                                 ipoib_open(cpriv);
 165                 mutex_unlock(&priv->vlan_mutex);
 166         }
 167         dev->if_drv_flags |= IFF_DRV_RUNNING;
 168         dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
 169
 170         return 0;
 171
 172 err_stop:
 173         ipoib_ib_dev_stop(priv, 1);
 174
 175 err_disable:
 176         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 177
 178         return -EINVAL;
 179 }
 180
 181 static void
 182 ipoib_init(void *arg)
 183 {
 184         struct ifnet *dev;
 185         struct ipoib_dev_priv *priv;
 186
 187         priv = arg;
 188         dev = priv->dev;
 189         if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 190                 ipoib_open(priv);
 191         queue_work(ipoib_workqueue, &priv->flush_light);
 192 }
 193
 194
 195 static int
 196 ipoib_stop(struct ipoib_dev_priv *priv)
 197 {
 198         struct ifnet *dev = priv->dev;
 199
 200         ipoib_dbg(priv, "stopping interface\n");
 201
 202         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 203
 204         dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
 205
 206         ipoib_ib_dev_down(priv, 0);
 207         ipoib_ib_dev_stop(priv, 0);
 208
 209         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 210                 struct ipoib_dev_priv *cpriv;
 211
 212                 /* Bring down any child interfaces too */
 213                 mutex_lock(&priv->vlan_mutex);
 214                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 215                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
 216                                 ipoib_stop(cpriv);
 217                 mutex_unlock(&priv->vlan_mutex);
 218         }
 219
 220         return 0;
 221 }
 222
 223 int
 224 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu)
 225 {
 226         struct ifnet *dev = priv->dev;
 227
 228         /* dev->if_mtu > 2K ==> connected mode */
 229         if (ipoib_cm_admin_enabled(priv)) {
 230                 if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
 231                         return -EINVAL;
 232
 233                 if (new_mtu > priv->mcast_mtu)
 234                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 235                                    priv->mcast_mtu);
 236
 237                 dev->if_mtu = new_mtu;
 238                 return 0;
 239         }
 240
 241         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 242                 return -EINVAL;
 243
 244         priv->admin_mtu = new_mtu;
 245
 246         dev->if_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 247
 248         queue_work(ipoib_workqueue, &priv->flush_light);
 249
 250         return 0;
 251 }
 252
 253 static int
 254 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
 255 {
 256         struct ipoib_dev_priv *priv = ifp->if_softc;
 257         struct ifaddr *ifa = (struct ifaddr *) data;
 258         struct ifreq *ifr = (struct ifreq *) data;
 259         int error = 0;
 260
 261         /* check if detaching */
 262         if (priv == NULL || priv->gone != 0)
 263                 return (ENXIO);
 264
 265         switch (command) {
 266         case SIOCSIFFLAGS:
 267                 if (ifp->if_flags & IFF_UP) {
 268                         if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
 269                                 error = -ipoib_open(priv);
 270                 } else
 271                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 272                                 ipoib_stop(priv);
 273                 break;
 274         case SIOCADDMULTI:
 275         case SIOCDELMULTI:
 276                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 277                         queue_work(ipoib_workqueue, &priv->restart_task);
 278                 break;
 279         case SIOCSIFADDR:
 280                 ifp->if_flags |= IFF_UP;
 281
 282                 switch (ifa->ifa_addr->sa_family) {
 283 #ifdef INET
 284                 case AF_INET:
 285                         ifp->if_init(ifp->if_softc);    /* before arpwhohas */
 286                         arp_ifinit(ifp, ifa);
 287                         break;
 288 #endif
 289                 default:
 290                         ifp->if_init(ifp->if_softc);
 291                         break;
 292                 }
 293                 break;
 294
 295         case SIOCGIFADDR:
 296                 bcopy(IF_LLADDR(ifp), &ifr->ifr_addr.sa_data[0],
 297                     INFINIBAND_ALEN);
 298                 break;
 299
 300         case SIOCSIFMTU:
 301                 /*
 302                  * Set the interface MTU.
 303                  */
 304                 error = -ipoib_change_mtu(priv, ifr->ifr_mtu);
 305                 break;
 306         default:
 307                 error = EINVAL;
 308                 break;
 309         }
 310         return (error);
 311 }
 312
 313
 314 static struct ipoib_path *
 315 __path_find(struct ipoib_dev_priv *priv, void *gid)
 316 {
 317         struct rb_node *n = priv->path_tree.rb_node;
 318         struct ipoib_path *path;
 319         int ret;
 320
 321         while (n) {
 322                 path = rb_entry(n, struct ipoib_path, rb_node);
 323
 324                 ret = memcmp(gid, path->pathrec.dgid.raw,
 325                              sizeof (union ib_gid));
 326
 327                 if (ret < 0)
 328                         n = n->rb_left;
 329                 else if (ret > 0)
 330                         n = n->rb_right;
 331                 else
 332                         return path;
 333         }
 334
 335         return NULL;
 336 }
 337
 338 static int
 339 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 340 {
 341         struct rb_node **n = &priv->path_tree.rb_node;
 342         struct rb_node *pn = NULL;
 343         struct ipoib_path *tpath;
 344         int ret;
 345
 346         while (*n) {
 347                 pn = *n;
 348                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 349
 350                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 351                              sizeof (union ib_gid));
 352                 if (ret < 0)
 353                         n = &pn->rb_left;
 354                 else if (ret > 0)
 355                         n = &pn->rb_right;
 356                 else
 357                         return -EEXIST;
 358         }
 359
 360         rb_link_node(&path->rb_node, pn, n);
 361         rb_insert_color(&path->rb_node, &priv->path_tree);
 362
 363         list_add_tail(&path->list, &priv->path_list);
 364
 365         return 0;
 366 }
 367
 368 void
 369 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 370 {
 371
 372         _IF_DRAIN(&path->queue);
 373
 374         if (path->ah)
 375                 ipoib_put_ah(path->ah);
 376         if (ipoib_cm_get(path))
 377                 ipoib_cm_destroy_tx(ipoib_cm_get(path));
 378
 379         kfree(path);
 380 }
 381
 382 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 383
 384 struct ipoib_path_iter *
 385 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
 386 {
 387         struct ipoib_path_iter *iter;
 388
 389         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 390         if (!iter)
 391                 return NULL;
 392
 393         iter->priv = priv;
 394         memset(iter->path.pathrec.dgid.raw, 0, 16);
 395
 396         if (ipoib_path_iter_next(iter)) {
 397                 kfree(iter);
 398                 return NULL;
 399         }
 400
 401         return iter;
 402 }
 403
 404 int
 405 ipoib_path_iter_next(struct ipoib_path_iter *iter)
 406 {
 407         struct ipoib_dev_priv *priv = iter->priv;
 408         struct rb_node *n;
 409         struct ipoib_path *path;
 410         int ret = 1;
 411
 412         spin_lock_irq(&priv->lock);
 413
 414         n = rb_first(&priv->path_tree);
 415
 416         while (n) {
 417                 path = rb_entry(n, struct ipoib_path, rb_node);
 418
 419                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 420                            sizeof (union ib_gid)) < 0) {
 421                         iter->path = *path;
 422                         ret = 0;
 423                         break;
 424                 }
 425
 426                 n = rb_next(n);
 427         }
 428
 429         spin_unlock_irq(&priv->lock);
 430
 431         return ret;
 432 }
 433
 434 void
 435 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
 436 {
 437         *path = iter->path;
 438 }
 439
 440 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 441
 442 void
 443 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
 444 {
 445         struct ipoib_path *path, *tp;
 446
 447         spin_lock_irq(&priv->lock);
 448
 449         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 450                 ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
 451                         be16_to_cpu(path->pathrec.dlid),
 452                         path->pathrec.dgid.raw, ":");
 453                 path->valid =  0;
 454         }
 455
 456         spin_unlock_irq(&priv->lock);
 457 }
 458
 459 void
 460 ipoib_flush_paths(struct ipoib_dev_priv *priv)
 461 {
 462         struct ipoib_path *path, *tp;
 463         LIST_HEAD(remove_list);
 464         unsigned long flags;
 465
 466         spin_lock_irqsave(&priv->lock, flags);
 467
 468         list_splice_init(&priv->path_list, &remove_list);
 469
 470         list_for_each_entry(path, &remove_list, list)
 471                 rb_erase(&path->rb_node, &priv->path_tree);
 472
 473         list_for_each_entry_safe(path, tp, &remove_list, list) {
 474                 if (path->query)
 475                         ib_sa_cancel_query(path->query_id, path->query);
 476                 spin_unlock_irqrestore(&priv->lock, flags);
 477                 wait_for_completion(&path->done);
 478                 ipoib_path_free(priv, path);
 479                 spin_lock_irqsave(&priv->lock, flags);
 480         }
 481
 482         spin_unlock_irqrestore(&priv->lock, flags);
 483 }
 484
 485 static void
 486 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
 487 {
 488         struct ipoib_path *path = path_ptr;
 489         struct ipoib_dev_priv *priv = path->priv;
 490         struct ifnet *dev = priv->dev;
 491         struct ipoib_ah *ah = NULL;
 492         struct ipoib_ah *old_ah = NULL;
 493         struct ifqueue mbqueue;
 494         struct mbuf *mb;
 495         unsigned long flags;
 496
 497         if (!status)
 498                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
 499                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
 500         else
 501                 ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
 502                           status, path->pathrec.dgid.raw, ":");
 503
 504         bzero(&mbqueue, sizeof(mbqueue));
 505
 506         if (!status) {
 507                 struct ib_ah_attr av;
 508
 509                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 510                         ah = ipoib_create_ah(priv, priv->pd, &av);
 511         }
 512
 513         spin_lock_irqsave(&priv->lock, flags);
 514
 515         if (ah) {
 516                 path->pathrec = *pathrec;
 517
 518                 old_ah   = path->ah;
 519                 path->ah = ah;
 520
 521                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 522                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 523
 524                 for (;;) {
 525                         _IF_DEQUEUE(&path->queue, mb);
 526                         if (mb == NULL)
 527                                 break;
 528                         _IF_ENQUEUE(&mbqueue, mb);
 529                 }
 530
 531 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 532                 if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
 533                         ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
 534 #endif
 535
 536                 path->valid = 1;
 537         }
 538
 539         path->query = NULL;
 540         complete(&path->done);
 541
 542         spin_unlock_irqrestore(&priv->lock, flags);
 543
 544         if (old_ah)
 545                 ipoib_put_ah(old_ah);
 546
 547         for (;;) {
 548                 _IF_DEQUEUE(&mbqueue, mb);
 549                 if (mb == NULL)
 550                         break;
 551                 mb->m_pkthdr.rcvif = dev;
 552                 if (dev->if_transmit(dev, mb))
 553                         ipoib_warn(priv, "dev_queue_xmit failed "
 554                                    "to requeue packet\n");
 555         }
 556 }
 557
 558 static struct ipoib_path *
 559 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
 560 {
 561         struct ipoib_path *path;
 562
 563         if (!priv->broadcast)
 564                 return NULL;
 565
 566         path = kzalloc(sizeof *path, GFP_ATOMIC);
 567         if (!path)
 568                 return NULL;
 569
 570         path->priv = priv;
 571
 572         bzero(&path->queue, sizeof(path->queue));
 573
 574 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 575         memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
 576 #endif
 577         memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
 578         path->pathrec.sgid          = priv->local_gid;
 579         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 580         path->pathrec.numb_path     = 1;
 581         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 582
 583         return path;
 584 }
 585
 586 static int
 587 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 588 {
 589         struct ifnet *dev = priv->dev;
 590
 591         ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
 592         struct ib_sa_path_rec p_rec;
 593
 594         p_rec = path->pathrec;
 595         p_rec.mtu_selector = IB_SA_GT;
 596
 597         switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
 598         case 512:
 599                 p_rec.mtu = IB_MTU_256;
 600                 break;
 601         case 1024:
 602                 p_rec.mtu = IB_MTU_512;
 603                 break;
 604         case 2048:
 605                 p_rec.mtu = IB_MTU_1024;
 606                 break;
 607         case 4096:
 608                 p_rec.mtu = IB_MTU_2048;
 609                 break;
 610         default:
 611                 /* Wildcard everything */
 612                 comp_mask = 0;
 613                 p_rec.mtu = 0;
 614                 p_rec.mtu_selector = 0;
 615         }
 616
 617         ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
 618                   p_rec.dgid.raw, ":",
 619                   comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
 620
 621         init_completion(&path->done);
 622
 623         path->query_id =
 624                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 625                                    &p_rec, comp_mask            |
 626                                    IB_SA_PATH_REC_DGID          |
 627                                    IB_SA_PATH_REC_SGID          |
 628                                    IB_SA_PATH_REC_NUMB_PATH     |
 629                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 630                                    IB_SA_PATH_REC_PKEY,
 631                                    1000, GFP_ATOMIC,
 632                                    path_rec_completion,
 633                                    path, &path->query);
 634         if (path->query_id < 0) {
 635                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 636                 path->query = NULL;
 637                 complete(&path->done);
 638                 return path->query_id;
 639         }
 640
 641         return 0;
 642 }
 643
 644 static void
 645 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
 646 {
 647         struct ipoib_path *path;
 648
 649         path = __path_find(priv, eh->hwaddr + 4);
 650         if (!path || !path->valid) {
 651                 int new_path = 0;
 652
 653                 if (!path) {
 654                         path = path_rec_create(priv, eh->hwaddr);
 655                         new_path = 1;
 656                 }
 657                 if (path) {
 658                         _IF_ENQUEUE(&path->queue, mb);
 659                         if (!path->query && path_rec_start(priv, path)) {
 660                                 spin_unlock_irqrestore(&priv->lock, flags);
 661                                 if (new_path)
 662                                         ipoib_path_free(priv, path);
 663                                 return;
 664                         } else
 665                                 __path_add(priv, path);
 666                 } else {
 667                         ++priv->dev->if_oerrors;
 668                         m_freem(mb);
 669                 }
 670
 671                 return;
 672         }
 673
 674         if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
 675                 ipoib_cm_send(priv, mb, ipoib_cm_get(path));
 676         } else if (path->ah) {
 677                 ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
 678         } else if ((path->query || !path_rec_start(priv, path)) &&
 679                     path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
 680                 _IF_ENQUEUE(&path->queue, mb);
 681         } else {
 682                 ++priv->dev->if_oerrors;
 683                 m_freem(mb);
 684         }
 685 }
 686
 687 static int
 688 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
 689 {
 690         struct ipoib_header *eh;
 691
 692         eh = mtod(mb, struct ipoib_header *);
 693         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
 694                 /* Add in the P_Key for multicast*/
 695                 eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 696                 eh->hwaddr[9] = priv->pkey & 0xff;
 697
 698                 ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
 699         } else
 700                 ipoib_unicast_send(mb, priv, eh);
 701
 702         return 0;
 703 }
 704
 705
 706 static void
 707 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
 708 {
 709         struct mbuf *mb;
 710
 711         if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
 712             IFF_DRV_RUNNING)
 713                 return;
 714
 715         spin_lock(&priv->lock);
 716         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
 717             (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
 718                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 719                 if (mb == NULL)
 720                         break;
 721                 IPOIB_MTAP(dev, mb);
 722                 ipoib_send_one(priv, mb);
 723         }
 724         spin_unlock(&priv->lock);
 725 }
 726
 727 static void
 728 ipoib_start(struct ifnet *dev)
 729 {
 730         _ipoib_start(dev, dev->if_softc);
 731 }
 732
 733 static void
 734 ipoib_vlan_start(struct ifnet *dev)
 735 {
 736         struct ipoib_dev_priv *priv;
 737         struct mbuf *mb;
 738
 739         priv = VLAN_COOKIE(dev);
 740         if (priv != NULL)
 741                 return _ipoib_start(dev, priv);
 742         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
 743                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 744                 if (mb == NULL)
 745                         break;
 746                 m_freem(mb);
 747                 dev->if_oerrors++;
 748         }
 749 }
 750
 751 int
 752 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
 753 {
 754
 755         /* Allocate RX/TX "rings" to hold queued mbs */
 756         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 757                                 GFP_KERNEL);
 758         if (!priv->rx_ring) {
 759                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 760                        ca->name, ipoib_recvq_size);
 761                 goto out;
 762         }
 763
 764         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
 765         if (!priv->tx_ring) {
 766                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 767                        ca->name, ipoib_sendq_size);
 768                 goto out_rx_ring_cleanup;
 769         }
 770         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 771
 772         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 773
 774         if (ipoib_ib_dev_init(priv, ca, port))
 775                 goto out_tx_ring_cleanup;
 776
 777         return 0;
 778
 779 out_tx_ring_cleanup:
 780         kfree(priv->tx_ring);
 781
 782 out_rx_ring_cleanup:
 783         kfree(priv->rx_ring);
 784
 785 out:
 786         return -ENOMEM;
 787 }
 788
 789 static void
 790 ipoib_detach(struct ipoib_dev_priv *priv)
 791 {
 792         struct ifnet *dev;
 793
 794         dev = priv->dev;
 795         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 796                 priv->gone = 1;
 797                 bpfdetach(dev);
 798                 if_detach(dev);
 799                 if_free(dev);
 800         } else
 801                 VLAN_SETCOOKIE(priv->dev, NULL);
 802
 803         free(priv, M_TEMP);
 804 }
 805
 806 void
 807 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
 808 {
 809         struct ipoib_dev_priv *cpriv, *tcpriv;
 810
 811         /* Delete any child interfaces first */
 812         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 813                 ipoib_dev_cleanup(cpriv);
 814                 ipoib_detach(cpriv);
 815         }
 816
 817         ipoib_ib_dev_cleanup(priv);
 818
 819         kfree(priv->rx_ring);
 820         kfree(priv->tx_ring);
 821
 822         priv->rx_ring = NULL;
 823         priv->tx_ring = NULL;
 824 }
 825
 826 static volatile int ipoib_unit;
 827
 828 static struct ipoib_dev_priv *
 829 ipoib_priv_alloc(void)
 830 {
 831         struct ipoib_dev_priv *priv;
 832
 833         priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
 834         spin_lock_init(&priv->lock);
 835         mutex_init(&priv->vlan_mutex);
 836         INIT_LIST_HEAD(&priv->path_list);
 837         INIT_LIST_HEAD(&priv->child_intfs);
 838         INIT_LIST_HEAD(&priv->dead_ahs);
 839         INIT_LIST_HEAD(&priv->multicast_list);
 840         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 841         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 842         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
 843         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
 844         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
 845         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
 846         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 847         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 848         memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
 849
 850         return (priv);
 851 }
 852
 853 struct ipoib_dev_priv *
 854 ipoib_intf_alloc(const char *name)
 855 {
 856         struct ipoib_dev_priv *priv;
 857         struct sockaddr_dl *sdl;
 858         struct ifnet *dev;
 859
 860         priv = ipoib_priv_alloc();
 861         dev = priv->dev = if_alloc(IFT_INFINIBAND);
 862         if (!dev) {
 863                 free(priv, M_TEMP);
 864                 return NULL;
 865         }
 866         dev->if_softc = priv;
 867         if_initname(dev, name, atomic_fetchadd_int(&ipoib_unit, 1));
 868         dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
 869         dev->if_addrlen = INFINIBAND_ALEN;
 870         dev->if_hdrlen = IPOIB_HEADER_LEN;
 871         if_attach(dev);
 872         dev->if_init = ipoib_init;
 873         dev->if_ioctl = ipoib_ioctl;
 874         dev->if_start = ipoib_start;
 875         dev->if_output = ipoib_output;
 876         dev->if_input = ipoib_input;
 877         dev->if_resolvemulti = ipoib_resolvemulti;
 878         if_initbaudrate(dev, IF_Gbps(10));
 879         dev->if_broadcastaddr = priv->broadcastaddr;
 880         dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
 881         sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
 882         sdl->sdl_type = IFT_INFINIBAND;
 883         sdl->sdl_alen = dev->if_addrlen;
 884         priv->dev = dev;
 885         if_link_state_change(dev, LINK_STATE_DOWN);
 886         bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
 887
 888         return dev->if_softc;
 889 }
 890
 891 int
 892 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
 893 {
 894         struct ib_device_attr *device_attr;
 895         int result = -ENOMEM;
 896
 897         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
 898         if (!device_attr) {
 899                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
 900                        hca->name, sizeof *device_attr);
 901                 return result;
 902         }
 903
 904         result = ib_query_device(hca, device_attr);
 905         if (result) {
 906                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
 907                        hca->name, result);
 908                 kfree(device_attr);
 909                 return result;
 910         }
 911         priv->hca_caps = device_attr->device_cap_flags;
 912
 913         kfree(device_attr);
 914
 915         priv->dev->if_hwassist = 0;
 916         priv->dev->if_capabilities = 0;
 917
 918 #ifndef CONFIG_INFINIBAND_IPOIB_CM
 919         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
 920                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
 921                 priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
 922                 priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
 923         }
 924
 925 #if 0
 926         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
 927                 priv->dev->if_capabilities |= IFCAP_TSO4;
 928                 priv->dev->if_hwassist |= CSUM_TSO;
 929         }
 930 #endif
 931 #endif
 932         priv->dev->if_capabilities |=
 933             IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
 934         priv->dev->if_capenable = priv->dev->if_capabilities;
 935
 936         return 0;
 937 }
 938
 939
 940 static struct ifnet *
 941 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
 942 {
 943         struct ipoib_dev_priv *priv;
 944         struct ib_port_attr attr;
 945         int result = -ENOMEM;
 946
 947         priv = ipoib_intf_alloc(format);
 948         if (!priv)
 949                 goto alloc_mem_failed;
 950
 951         if (!ib_query_port(hca, port, &attr))
 952                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
 953         else {
 954                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
 955                        hca->name, port);
 956                 goto device_init_failed;
 957         }
 958
 959         /* MTU will be reset when mcast join happens */
 960         priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
 961         priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
 962
 963         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 964         if (result) {
 965                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 966                        hca->name, port, result);
 967                 goto device_init_failed;
 968         }
 969
 970         if (ipoib_set_dev_features(priv, hca))
 971                 goto device_init_failed;
 972
 973         /*
 974          * Set the full membership bit, so that we join the right
 975          * broadcast group, etc.
 976          */
 977         priv->pkey |= 0x8000;
 978
 979         priv->broadcastaddr[8] = priv->pkey >> 8;
 980         priv->broadcastaddr[9] = priv->pkey & 0xff;
 981
 982         result = ib_query_gid(hca, port, 0, &priv->local_gid);
 983         if (result) {
 984                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
 985                        hca->name, port, result);
 986                 goto device_init_failed;
 987         }
 988         memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
 989
 990         result = ipoib_dev_init(priv, hca, port);
 991         if (result < 0) {
 992                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
 993                        hca->name, port, result);
 994                 goto device_init_failed;
 995         }
 996         if (ipoib_cm_admin_enabled(priv))
 997                 priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
 998
 999         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1000                               priv->ca, ipoib_event);
1001         result = ib_register_event_handler(&priv->event_handler);
1002         if (result < 0) {
1003                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1004                        "port %d (ret = %d)\n",
1005                        hca->name, port, result);
1006                 goto event_failed;
1007         }
1008         if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1009
1010         return priv->dev;
1011
1012 event_failed:
1013         ipoib_dev_cleanup(priv);
1014
1015 device_init_failed:
1016         ipoib_detach(priv);
1017
1018 alloc_mem_failed:
1019         return ERR_PTR(result);
1020 }
1021
1022 static void
1023 ipoib_add_one(struct ib_device *device)
1024 {
1025         struct list_head *dev_list;
1026         struct ifnet *dev;
1027         struct ipoib_dev_priv *priv;
1028         int s, e, p;
1029
1030         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1031                 return;
1032
1033         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1034         if (!dev_list)
1035                 return;
1036
1037         INIT_LIST_HEAD(dev_list);
1038
1039         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1040                 s = 0;
1041                 e = 0;
1042         } else {
1043                 s = 1;
1044                 e = device->phys_port_cnt;
1045         }
1046
1047         for (p = s; p <= e; ++p) {
1048                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1049                         continue;
1050                 dev = ipoib_add_port("ib", device, p);
1051                 if (!IS_ERR(dev)) {
1052                         priv = dev->if_softc;
1053                         list_add_tail(&priv->list, dev_list);
1054                 }
1055         }
1056
1057         ib_set_client_data(device, &ipoib_client, dev_list);
1058 }
1059
1060 static void
1061 ipoib_remove_one(struct ib_device *device)
1062 {
1063         struct ipoib_dev_priv *priv, *tmp;
1064         struct list_head *dev_list;
1065
1066         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1067                 return;
1068
1069         dev_list = ib_get_client_data(device, &ipoib_client);
1070
1071         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1072                 if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1073                         continue;
1074
1075                 ipoib_stop(priv);
1076
1077                 ib_unregister_event_handler(&priv->event_handler);
1078
1079                 /* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1080
1081                 flush_workqueue(ipoib_workqueue);
1082
1083                 ipoib_dev_cleanup(priv);
1084                 ipoib_detach(priv);
1085         }
1086
1087         kfree(dev_list);
1088 }
1089
1090 static void
1091 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1092 {
1093         struct ipoib_dev_priv *parent;
1094         struct ipoib_dev_priv *priv;
1095         struct ifnet *dev;
1096         uint16_t pkey;
1097         int error;
1098
1099         if (ifp->if_type != IFT_INFINIBAND)
1100                 return;
1101         dev = VLAN_DEVAT(ifp, vtag);
1102         if (dev == NULL)
1103                 return;
1104         priv = NULL;
1105         error = 0;
1106         parent = ifp->if_softc;
1107         /* We only support 15 bits of pkey. */
1108         if (vtag & 0x8000)
1109                 return;
1110         pkey = vtag | 0x8000;   /* Set full membership bit. */
1111         if (pkey == parent->pkey)
1112                 return;
1113         /* Check for dups */
1114         mutex_lock(&parent->vlan_mutex);
1115         list_for_each_entry(priv, &parent->child_intfs, list) {
1116                 if (priv->pkey == pkey) {
1117                         priv = NULL;
1118                         error = EBUSY;
1119                         goto out;
1120                 }
1121         }
1122         priv = ipoib_priv_alloc();
1123         priv->dev = dev;
1124         priv->max_ib_mtu = parent->max_ib_mtu;
1125         priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1126         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1127         error = ipoib_set_dev_features(priv, parent->ca);
1128         if (error)
1129                 goto out;
1130         priv->pkey = pkey;
1131         priv->broadcastaddr[8] = pkey >> 8;
1132         priv->broadcastaddr[9] = pkey & 0xff;
1133         dev->if_broadcastaddr = priv->broadcastaddr;
1134         error = ipoib_dev_init(priv, parent->ca, parent->port);
1135         if (error)
1136                 goto out;
1137         priv->parent = parent->dev;
1138         list_add_tail(&priv->list, &parent->child_intfs);
1139         VLAN_SETCOOKIE(dev, priv);
1140         dev->if_start = ipoib_vlan_start;
1141         dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1142         dev->if_hdrlen = IPOIB_HEADER_LEN;
1143         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1144                 ipoib_open(priv);
1145         mutex_unlock(&parent->vlan_mutex);
1146         return;
1147 out:
1148         mutex_unlock(&parent->vlan_mutex);
1149         if (priv)
1150                 free(priv, M_TEMP);
1151         if (error)
1152                 ipoib_warn(parent,
1153                     "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1154                     parent->ca->name, parent->port, vtag);
1155         return;
1156 }
1157
1158 static void
1159 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1160 {
1161         struct ipoib_dev_priv *parent;
1162         struct ipoib_dev_priv *priv;
1163         struct ifnet *dev;
1164         uint16_t pkey;
1165
1166         if (ifp->if_type != IFT_INFINIBAND)
1167                 return;
1168
1169         dev = VLAN_DEVAT(ifp, vtag);
1170         if (dev)
1171                 VLAN_SETCOOKIE(dev, NULL);
1172         pkey = vtag | 0x8000;
1173         parent = ifp->if_softc;
1174         mutex_lock(&parent->vlan_mutex);
1175         list_for_each_entry(priv, &parent->child_intfs, list) {
1176                 if (priv->pkey == pkey) {
1177                         ipoib_dev_cleanup(priv);
1178                         list_del(&priv->list);
1179                         break;
1180                 }
1181         }
1182         mutex_unlock(&parent->vlan_mutex);
1183 }
1184
1185 eventhandler_tag ipoib_vlan_attach;
1186 eventhandler_tag ipoib_vlan_detach;
1187
1188 static int __init
1189 ipoib_init_module(void)
1190 {
1191         int ret;
1192
1193         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1194         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1195         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1196
1197         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1198         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1199         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1200                                                      IPOIB_MIN_QUEUE_SIZE));
1201 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1202         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1203 #endif
1204
1205         ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1206                 ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1207         ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1208                 ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1209
1210         /*
1211          * We create our own workqueue mainly because we want to be
1212          * able to flush it when devices are being removed.  We can't
1213          * use schedule_work()/flush_scheduled_work() because both
1214          * unregister_netdev() and linkwatch_event take the rtnl lock,
1215          * so flush_scheduled_work() can deadlock during device
1216          * removal.
1217          */
1218         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1219         if (!ipoib_workqueue) {
1220                 ret = -ENOMEM;
1221                 goto err_fs;
1222         }
1223
1224         ib_sa_register_client(&ipoib_sa_client);
1225
1226         ret = ib_register_client(&ipoib_client);
1227         if (ret)
1228                 goto err_sa;
1229
1230         return 0;
1231
1232 err_sa:
1233         ib_sa_unregister_client(&ipoib_sa_client);
1234         destroy_workqueue(ipoib_workqueue);
1235
1236 err_fs:
1237         return ret;
1238 }
1239
1240 static void __exit
1241 ipoib_cleanup_module(void)
1242 {
1243
1244         EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1245         EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1246         ib_unregister_client(&ipoib_client);
1247         ib_sa_unregister_client(&ipoib_sa_client);
1248         destroy_workqueue(ipoib_workqueue);
1249 }
1250
1251 /*
1252  * Infiniband output routine.
1253  */
1254 static int
1255 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1256         const struct sockaddr *dst, struct route *ro)
1257 {
1258         u_char edst[INFINIBAND_ALEN];
1259         struct llentry *lle = NULL;
1260         struct rtentry *rt0 = NULL;
1261         struct ipoib_header *eh;
1262         int error = 0;
1263         short type;
1264
1265         if (ro != NULL) {
1266                 if (!(m->m_flags & (M_BCAST | M_MCAST)))
1267                         lle = ro->ro_lle;
1268                 rt0 = ro->ro_rt;
1269         }
1270 #ifdef MAC
1271         error = mac_ifnet_check_transmit(ifp, m);
1272         if (error)
1273                 goto bad;
1274 #endif
1275
1276         M_PROFILE(m);
1277         if (ifp->if_flags & IFF_MONITOR) {
1278                 error = ENETDOWN;
1279                 goto bad;
1280         }
1281         if (!((ifp->if_flags & IFF_UP) &&
1282             (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1283                 error = ENETDOWN;
1284                 goto bad;
1285         }
1286
1287         switch (dst->sa_family) {
1288 #ifdef INET
1289         case AF_INET:
1290                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1291                         memcpy(edst, &lle->ll_addr.mac8, sizeof(edst));
1292                 else if (m->m_flags & M_MCAST)
1293                         ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1294                 else
1295                         error = arpresolve(ifp, rt0, m, dst, edst, &lle);
1296                 if (error)
1297                         return (error == EWOULDBLOCK ? 0 : error);
1298                 type = htons(ETHERTYPE_IP);
1299                 break;
1300         case AF_ARP:
1301         {
1302                 struct arphdr *ah;
1303                 ah = mtod(m, struct arphdr *);
1304                 ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1305
1306                 switch(ntohs(ah->ar_op)) {
1307                 case ARPOP_REVREQUEST:
1308                 case ARPOP_REVREPLY:
1309                         type = htons(ETHERTYPE_REVARP);
1310                         break;
1311                 case ARPOP_REQUEST:
1312                 case ARPOP_REPLY:
1313                 default:
1314                         type = htons(ETHERTYPE_ARP);
1315                         break;
1316                 }
1317
1318                 if (m->m_flags & M_BCAST)
1319                         bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1320                 else
1321                         bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1322
1323         }
1324         break;
1325 #endif
1326 #ifdef INET6
1327         case AF_INET6:
1328                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1329                         memcpy(edst, &lle->ll_addr.mac8, sizeof(edst));
1330                 else if (m->m_flags & M_MCAST)
1331                         ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1332                 else
1333                         error = nd6_storelladdr(ifp, m, dst, (u_char *)edst, &lle);
1334                 if (error)
1335                         return error;
1336                 type = htons(ETHERTYPE_IPV6);
1337                 break;
1338 #endif
1339
1340         default:
1341                 if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1342                 error = EAFNOSUPPORT;
1343                 goto bad;
1344         }
1345
1346         /*
1347          * Add local net header.  If no space in first mbuf,
1348          * allocate another.
1349          */
1350         M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1351         if (m == NULL) {
1352                 error = ENOBUFS;
1353                 goto bad;
1354         }
1355         eh = mtod(m, struct ipoib_header *);
1356         (void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1357         (void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1358
1359         /*
1360          * Queue message on interface, update output statistics if
1361          * successful, and start output if interface not yet active.
1362          */
1363         return ((ifp->if_transmit)(ifp, m));
1364 bad:
1365         if (m != NULL)
1366                 m_freem(m);
1367         return (error);
1368 }
1369
1370 /*
1371  * Upper layer processing for a received Infiniband packet.
1372  */
1373 void
1374 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1375 {
1376         int isr;
1377
1378 #ifdef MAC
1379         /*
1380          * Tag the mbuf with an appropriate MAC label before any other
1381          * consumers can get to it.
1382          */
1383         mac_ifnet_create_mbuf(ifp, m);
1384 #endif
1385         /* Allow monitor mode to claim this frame, after stats are updated. */
1386         if (ifp->if_flags & IFF_MONITOR) {
1387                 if_printf(ifp, "discard frame at IFF_MONITOR\n");
1388                 m_freem(m);
1389                 return;
1390         }
1391         /*
1392          * Dispatch frame to upper layer.
1393          */
1394         switch (proto) {
1395 #ifdef INET
1396         case ETHERTYPE_IP:
1397                 isr = NETISR_IP;
1398                 break;
1399
1400         case ETHERTYPE_ARP:
1401                 if (ifp->if_flags & IFF_NOARP) {
1402                         /* Discard packet if ARP is disabled on interface */
1403                         m_freem(m);
1404                         return;
1405                 }
1406                 isr = NETISR_ARP;
1407                 break;
1408 #endif
1409 #ifdef INET6
1410         case ETHERTYPE_IPV6:
1411                 isr = NETISR_IPV6;
1412                 break;
1413 #endif
1414         default:
1415                 goto discard;
1416         }
1417         netisr_dispatch(isr, m);
1418         return;
1419
1420 discard:
1421         m_freem(m);
1422 }
1423
1424 /*
1425  * Process a received Infiniband packet.
1426  */
1427 static void
1428 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1429 {
1430         struct ipoib_header *eh;
1431
1432         if ((ifp->if_flags & IFF_UP) == 0) {
1433                 m_freem(m);
1434                 return;
1435         }
1436         CURVNET_SET_QUIET(ifp->if_vnet);
1437
1438         /* Let BPF have it before we strip the header. */
1439         IPOIB_MTAP(ifp, m);
1440         eh = mtod(m, struct ipoib_header *);
1441         /*
1442          * Reset layer specific mbuf flags to avoid confusing upper layers.
1443          * Strip off Infiniband header.
1444          */
1445         m->m_flags &= ~M_VLANTAG;
1446         m_clrprotoflags(m);
1447         m_adj(m, IPOIB_HEADER_LEN);
1448
1449         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1450                 if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1451                     ifp->if_addrlen) == 0)
1452                         m->m_flags |= M_BCAST;
1453                 else
1454                         m->m_flags |= M_MCAST;
1455                 ifp->if_imcasts++;
1456         }
1457
1458         ipoib_demux(ifp, m, ntohs(eh->proto));
1459         CURVNET_RESTORE();
1460 }
1461
1462 static int
1463 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1464         struct sockaddr *sa)
1465 {
1466         struct sockaddr_dl *sdl;
1467 #ifdef INET
1468         struct sockaddr_in *sin;
1469 #endif
1470 #ifdef INET6
1471         struct sockaddr_in6 *sin6;
1472 #endif
1473         u_char *e_addr;
1474
1475         switch(sa->sa_family) {
1476         case AF_LINK:
1477                 /*
1478                  * No mapping needed. Just check that it's a valid MC address.
1479                  */
1480                 sdl = (struct sockaddr_dl *)sa;
1481                 e_addr = LLADDR(sdl);
1482                 if (!IPOIB_IS_MULTICAST(e_addr))
1483                         return EADDRNOTAVAIL;
1484                 *llsa = 0;
1485                 return 0;
1486
1487 #ifdef INET
1488         case AF_INET:
1489                 sin = (struct sockaddr_in *)sa;
1490                 if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1491                         return EADDRNOTAVAIL;
1492                 sdl = malloc(sizeof *sdl, M_IFMADDR,
1493                        M_NOWAIT|M_ZERO);
1494                 if (sdl == NULL)
1495                         return ENOMEM;
1496                 sdl->sdl_len = sizeof *sdl;
1497                 sdl->sdl_family = AF_LINK;
1498                 sdl->sdl_index = ifp->if_index;
1499                 sdl->sdl_type = IFT_INFINIBAND;
1500                 sdl->sdl_alen = INFINIBAND_ALEN;
1501                 e_addr = LLADDR(sdl);
1502                 ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1503                     e_addr);
1504                 *llsa = (struct sockaddr *)sdl;
1505                 return 0;
1506 #endif
1507 #ifdef INET6
1508         case AF_INET6:
1509                 sin6 = (struct sockaddr_in6 *)sa;
1510                 /*
1511                  * An IP6 address of 0 means listen to all
1512                  * of the multicast address used for IP6.
1513                  * This has no meaning in ipoib.
1514                  */
1515                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1516                         return EADDRNOTAVAIL;
1517                 if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1518                         return EADDRNOTAVAIL;
1519                 sdl = malloc(sizeof *sdl, M_IFMADDR,
1520                        M_NOWAIT|M_ZERO);
1521                 if (sdl == NULL)
1522                         return (ENOMEM);
1523                 sdl->sdl_len = sizeof *sdl;
1524                 sdl->sdl_family = AF_LINK;
1525                 sdl->sdl_index = ifp->if_index;
1526                 sdl->sdl_type = IFT_INFINIBAND;
1527                 sdl->sdl_alen = INFINIBAND_ALEN;
1528                 e_addr = LLADDR(sdl);
1529                 ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1530                 *llsa = (struct sockaddr *)sdl;
1531                 return 0;
1532 #endif
1533
1534         default:
1535                 return EAFNOSUPPORT;
1536         }
1537 }
1538
1539 module_init(ipoib_init_module);
1540 module_exit(ipoib_cleanup_module);
1541
1542 #undef MODULE_VERSION
1543 #include <sys/module.h>
1544 static int
1545 ipoib_evhand(module_t mod, int event, void *arg)
1546 {
1547         return (0);
1548 }
1549
1550 static moduledata_t ipoib_mod = {
1551         .name = "ipoib",
1552         .evhand = ipoib_evhand,
1553 };
1554
1555 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_SMP, SI_ORDER_ANY);
1556 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1557