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

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
   3  *
   4  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include "ipoib.h"
  38
  39 static  int ipoib_resolvemulti(struct ifnet *, struct sockaddr **,
  40                 struct sockaddr *);
  41
  42
  43 #include <linux/module.h>
  44
  45 #include <linux/slab.h>
  46 #include <linux/kernel.h>
  47 #include <linux/vmalloc.h>
  48
  49 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  50 #include <linux/if_vlan.h>
  51 #include <net/ip.h>
  52 #include <net/ipv6.h>
  53
  54 MODULE_AUTHOR("Roland Dreier");
  55 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
  59 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
  60
  61 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  62 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  63 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  64 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  65
  66 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  67 int ipoib_debug_level = 1;
  68
  69 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  70 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  71 #endif
  72
  73 struct ipoib_path_iter {
  74         struct ipoib_dev_priv *priv;
  75         struct ipoib_path  path;
  76 };
  77
  78 static const u8 ipv4_bcast_addr[] = {
  79         0x00, 0xff, 0xff, 0xff,
  80         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  81         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  82 };
  83
  84 struct workqueue_struct *ipoib_workqueue;
  85
  86 struct ib_sa_client ipoib_sa_client;
  87
  88 static void ipoib_add_one(struct ib_device *device);
  89 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  90 static void ipoib_start(struct ifnet *dev);
  91 static int ipoib_output(struct ifnet *ifp, struct mbuf *m,
  92             const struct sockaddr *dst, struct route *ro);
  93 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
  94 static void ipoib_input(struct ifnet *ifp, struct mbuf *m);
  95
  96 #define IPOIB_MTAP(_ifp, _m)                                    \
  97 do {                                                            \
  98         if (bpf_peers_present((_ifp)->if_bpf)) {                \
  99                 M_ASSERTVALID(_m);                              \
 100                 ipoib_mtap_mb((_ifp), (_m));                    \
 101         }                                                       \
 102 } while (0)
 103
 104 static struct unrhdr *ipoib_unrhdr;
 105
 106 static void
 107 ipoib_unrhdr_init(void *arg)
 108 {
 109
 110         ipoib_unrhdr = new_unrhdr(0, 65535, NULL);
 111 }
 112 SYSINIT(ipoib_unrhdr_init, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_init, NULL);
 113
 114 static void
 115 ipoib_unrhdr_uninit(void *arg)
 116 {
 117
 118         if (ipoib_unrhdr != NULL) {
 119                 struct unrhdr *hdr;
 120
 121                 hdr = ipoib_unrhdr;
 122                 ipoib_unrhdr = NULL;
 123
 124                 delete_unrhdr(hdr);
 125         }
 126 }
 127 SYSUNINIT(ipoib_unrhdr_uninit, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_uninit, NULL);
 128
 129 /*
 130  * This is for clients that have an ipoib_header in the mbuf.
 131  */
 132 static void
 133 ipoib_mtap_mb(struct ifnet *ifp, struct mbuf *mb)
 134 {
 135         struct ipoib_header *ih;
 136         struct ether_header eh;
 137
 138         ih = mtod(mb, struct ipoib_header *);
 139         eh.ether_type = ih->proto;
 140         bcopy(ih->hwaddr, &eh.ether_dhost, ETHER_ADDR_LEN);
 141         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 142         mb->m_data += sizeof(struct ipoib_header);
 143         mb->m_len -= sizeof(struct ipoib_header);
 144         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 145         mb->m_data -= sizeof(struct ipoib_header);
 146         mb->m_len += sizeof(struct ipoib_header);
 147 }
 148
 149 void
 150 ipoib_mtap_proto(struct ifnet *ifp, struct mbuf *mb, uint16_t proto)
 151 {
 152         struct ether_header eh;
 153
 154         eh.ether_type = proto;
 155         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 156         bzero(&eh.ether_dhost, ETHER_ADDR_LEN);
 157         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 158 }
 159
 160 static struct ib_client ipoib_client = {
 161         .name   = "ipoib",
 162         .add    = ipoib_add_one,
 163         .remove = ipoib_remove_one
 164 };
 165
 166 int
 167 ipoib_open(struct ipoib_dev_priv *priv)
 168 {
 169         struct ifnet *dev = priv->dev;
 170
 171         ipoib_dbg(priv, "bringing up interface\n");
 172
 173         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 174
 175         if (ipoib_pkey_dev_delay_open(priv))
 176                 return 0;
 177
 178         if (ipoib_ib_dev_open(priv))
 179                 goto err_disable;
 180
 181         if (ipoib_ib_dev_up(priv))
 182                 goto err_stop;
 183
 184         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 185                 struct ipoib_dev_priv *cpriv;
 186
 187                 /* Bring up any child interfaces too */
 188                 mutex_lock(&priv->vlan_mutex);
 189                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 190                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 191                                 ipoib_open(cpriv);
 192                 mutex_unlock(&priv->vlan_mutex);
 193         }
 194         dev->if_drv_flags |= IFF_DRV_RUNNING;
 195         dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
 196
 197         return 0;
 198
 199 err_stop:
 200         ipoib_ib_dev_stop(priv, 1);
 201
 202 err_disable:
 203         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 204
 205         return -EINVAL;
 206 }
 207
 208 static void
 209 ipoib_init(void *arg)
 210 {
 211         struct ifnet *dev;
 212         struct ipoib_dev_priv *priv;
 213
 214         priv = arg;
 215         dev = priv->dev;
 216         if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 217                 ipoib_open(priv);
 218         queue_work(ipoib_workqueue, &priv->flush_light);
 219 }
 220
 221
 222 static int
 223 ipoib_stop(struct ipoib_dev_priv *priv)
 224 {
 225         struct ifnet *dev = priv->dev;
 226
 227         ipoib_dbg(priv, "stopping interface\n");
 228
 229         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 230
 231         dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
 232
 233         ipoib_ib_dev_down(priv, 0);
 234         ipoib_ib_dev_stop(priv, 0);
 235
 236         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 237                 struct ipoib_dev_priv *cpriv;
 238
 239                 /* Bring down any child interfaces too */
 240                 mutex_lock(&priv->vlan_mutex);
 241                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 242                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
 243                                 ipoib_stop(cpriv);
 244                 mutex_unlock(&priv->vlan_mutex);
 245         }
 246
 247         return 0;
 248 }
 249
 250 int
 251 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu)
 252 {
 253         struct ifnet *dev = priv->dev;
 254
 255         /* dev->if_mtu > 2K ==> connected mode */
 256         if (ipoib_cm_admin_enabled(priv)) {
 257                 if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
 258                         return -EINVAL;
 259
 260                 if (new_mtu > priv->mcast_mtu)
 261                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 262                                    priv->mcast_mtu);
 263
 264                 dev->if_mtu = new_mtu;
 265                 return 0;
 266         }
 267
 268         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 269                 return -EINVAL;
 270
 271         priv->admin_mtu = new_mtu;
 272
 273         dev->if_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 274
 275         queue_work(ipoib_workqueue, &priv->flush_light);
 276
 277         return 0;
 278 }
 279
 280 static int
 281 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
 282 {
 283         struct ipoib_dev_priv *priv = ifp->if_softc;
 284         struct ifaddr *ifa = (struct ifaddr *) data;
 285         struct ifreq *ifr = (struct ifreq *) data;
 286         int error = 0;
 287
 288         /* check if detaching */
 289         if (priv == NULL || priv->gone != 0)
 290                 return (ENXIO);
 291
 292         switch (command) {
 293         case SIOCSIFFLAGS:
 294                 if (ifp->if_flags & IFF_UP) {
 295                         if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
 296                                 error = -ipoib_open(priv);
 297                 } else
 298                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 299                                 ipoib_stop(priv);
 300                 break;
 301         case SIOCADDMULTI:
 302         case SIOCDELMULTI:
 303                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 304                         queue_work(ipoib_workqueue, &priv->restart_task);
 305                 break;
 306         case SIOCSIFADDR:
 307                 ifp->if_flags |= IFF_UP;
 308
 309                 switch (ifa->ifa_addr->sa_family) {
 310 #ifdef INET
 311                 case AF_INET:
 312                         ifp->if_init(ifp->if_softc);    /* before arpwhohas */
 313                         arp_ifinit(ifp, ifa);
 314                         break;
 315 #endif
 316                 default:
 317                         ifp->if_init(ifp->if_softc);
 318                         break;
 319                 }
 320                 break;
 321
 322         case SIOCGIFADDR:
 323                 bcopy(IF_LLADDR(ifp), &ifr->ifr_addr.sa_data[0],
 324                     INFINIBAND_ALEN);
 325                 break;
 326
 327         case SIOCSIFMTU:
 328                 /*
 329                  * Set the interface MTU.
 330                  */
 331                 error = -ipoib_change_mtu(priv, ifr->ifr_mtu);
 332                 break;
 333         default:
 334                 error = EINVAL;
 335                 break;
 336         }
 337         return (error);
 338 }
 339
 340
 341 static struct ipoib_path *
 342 __path_find(struct ipoib_dev_priv *priv, void *gid)
 343 {
 344         struct rb_node *n = priv->path_tree.rb_node;
 345         struct ipoib_path *path;
 346         int ret;
 347
 348         while (n) {
 349                 path = rb_entry(n, struct ipoib_path, rb_node);
 350
 351                 ret = memcmp(gid, path->pathrec.dgid.raw,
 352                              sizeof (union ib_gid));
 353
 354                 if (ret < 0)
 355                         n = n->rb_left;
 356                 else if (ret > 0)
 357                         n = n->rb_right;
 358                 else
 359                         return path;
 360         }
 361
 362         return NULL;
 363 }
 364
 365 static int
 366 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 367 {
 368         struct rb_node **n = &priv->path_tree.rb_node;
 369         struct rb_node *pn = NULL;
 370         struct ipoib_path *tpath;
 371         int ret;
 372
 373         while (*n) {
 374                 pn = *n;
 375                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 376
 377                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 378                              sizeof (union ib_gid));
 379                 if (ret < 0)
 380                         n = &pn->rb_left;
 381                 else if (ret > 0)
 382                         n = &pn->rb_right;
 383                 else
 384                         return -EEXIST;
 385         }
 386
 387         rb_link_node(&path->rb_node, pn, n);
 388         rb_insert_color(&path->rb_node, &priv->path_tree);
 389
 390         list_add_tail(&path->list, &priv->path_list);
 391
 392         return 0;
 393 }
 394
 395 void
 396 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 397 {
 398
 399         _IF_DRAIN(&path->queue);
 400
 401         if (path->ah)
 402                 ipoib_put_ah(path->ah);
 403         if (ipoib_cm_get(path))
 404                 ipoib_cm_destroy_tx(ipoib_cm_get(path));
 405
 406         kfree(path);
 407 }
 408
 409 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 410
 411 struct ipoib_path_iter *
 412 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
 413 {
 414         struct ipoib_path_iter *iter;
 415
 416         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 417         if (!iter)
 418                 return NULL;
 419
 420         iter->priv = priv;
 421         memset(iter->path.pathrec.dgid.raw, 0, 16);
 422
 423         if (ipoib_path_iter_next(iter)) {
 424                 kfree(iter);
 425                 return NULL;
 426         }
 427
 428         return iter;
 429 }
 430
 431 int
 432 ipoib_path_iter_next(struct ipoib_path_iter *iter)
 433 {
 434         struct ipoib_dev_priv *priv = iter->priv;
 435         struct rb_node *n;
 436         struct ipoib_path *path;
 437         int ret = 1;
 438
 439         spin_lock_irq(&priv->lock);
 440
 441         n = rb_first(&priv->path_tree);
 442
 443         while (n) {
 444                 path = rb_entry(n, struct ipoib_path, rb_node);
 445
 446                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 447                            sizeof (union ib_gid)) < 0) {
 448                         iter->path = *path;
 449                         ret = 0;
 450                         break;
 451                 }
 452
 453                 n = rb_next(n);
 454         }
 455
 456         spin_unlock_irq(&priv->lock);
 457
 458         return ret;
 459 }
 460
 461 void
 462 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
 463 {
 464         *path = iter->path;
 465 }
 466
 467 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 468
 469 void
 470 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
 471 {
 472         struct ipoib_path *path, *tp;
 473
 474         spin_lock_irq(&priv->lock);
 475
 476         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 477                 ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
 478                         be16_to_cpu(path->pathrec.dlid),
 479                         path->pathrec.dgid.raw, ":");
 480                 path->valid =  0;
 481         }
 482
 483         spin_unlock_irq(&priv->lock);
 484 }
 485
 486 void
 487 ipoib_flush_paths(struct ipoib_dev_priv *priv)
 488 {
 489         struct ipoib_path *path, *tp;
 490         LIST_HEAD(remove_list);
 491         unsigned long flags;
 492
 493         spin_lock_irqsave(&priv->lock, flags);
 494
 495         list_splice_init(&priv->path_list, &remove_list);
 496
 497         list_for_each_entry(path, &remove_list, list)
 498                 rb_erase(&path->rb_node, &priv->path_tree);
 499
 500         list_for_each_entry_safe(path, tp, &remove_list, list) {
 501                 if (path->query)
 502                         ib_sa_cancel_query(path->query_id, path->query);
 503                 spin_unlock_irqrestore(&priv->lock, flags);
 504                 wait_for_completion(&path->done);
 505                 ipoib_path_free(priv, path);
 506                 spin_lock_irqsave(&priv->lock, flags);
 507         }
 508
 509         spin_unlock_irqrestore(&priv->lock, flags);
 510 }
 511
 512 static void
 513 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
 514 {
 515         struct ipoib_path *path = path_ptr;
 516         struct ipoib_dev_priv *priv = path->priv;
 517         struct ifnet *dev = priv->dev;
 518         struct ipoib_ah *ah = NULL;
 519         struct ipoib_ah *old_ah = NULL;
 520         struct ifqueue mbqueue;
 521         struct mbuf *mb;
 522         unsigned long flags;
 523
 524         if (!status)
 525                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
 526                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
 527         else
 528                 ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
 529                           status, path->pathrec.dgid.raw, ":");
 530
 531         bzero(&mbqueue, sizeof(mbqueue));
 532
 533         if (!status) {
 534                 struct ib_ah_attr av;
 535
 536                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 537                         ah = ipoib_create_ah(priv, priv->pd, &av);
 538         }
 539
 540         spin_lock_irqsave(&priv->lock, flags);
 541
 542         if (ah) {
 543                 path->pathrec = *pathrec;
 544
 545                 old_ah   = path->ah;
 546                 path->ah = ah;
 547
 548                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 549                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 550
 551                 for (;;) {
 552                         _IF_DEQUEUE(&path->queue, mb);
 553                         if (mb == NULL)
 554                                 break;
 555                         _IF_ENQUEUE(&mbqueue, mb);
 556                 }
 557
 558 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 559                 if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
 560                         ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
 561 #endif
 562
 563                 path->valid = 1;
 564         }
 565
 566         path->query = NULL;
 567         complete(&path->done);
 568
 569         spin_unlock_irqrestore(&priv->lock, flags);
 570
 571         if (old_ah)
 572                 ipoib_put_ah(old_ah);
 573
 574         for (;;) {
 575                 _IF_DEQUEUE(&mbqueue, mb);
 576                 if (mb == NULL)
 577                         break;
 578                 mb->m_pkthdr.rcvif = dev;
 579                 if (dev->if_transmit(dev, mb))
 580                         ipoib_warn(priv, "dev_queue_xmit failed "
 581                                    "to requeue packet\n");
 582         }
 583 }
 584
 585 static struct ipoib_path *
 586 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
 587 {
 588         struct ipoib_path *path;
 589
 590         if (!priv->broadcast)
 591                 return NULL;
 592
 593         path = kzalloc(sizeof *path, GFP_ATOMIC);
 594         if (!path)
 595                 return NULL;
 596
 597         path->priv = priv;
 598
 599         bzero(&path->queue, sizeof(path->queue));
 600
 601 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 602         memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
 603 #endif
 604         memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
 605         path->pathrec.sgid          = priv->local_gid;
 606         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 607         path->pathrec.numb_path     = 1;
 608         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 609
 610         return path;
 611 }
 612
 613 static int
 614 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 615 {
 616         struct ifnet *dev = priv->dev;
 617
 618         ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
 619         struct ib_sa_path_rec p_rec;
 620
 621         p_rec = path->pathrec;
 622         p_rec.mtu_selector = IB_SA_GT;
 623
 624         switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
 625         case 512:
 626                 p_rec.mtu = IB_MTU_256;
 627                 break;
 628         case 1024:
 629                 p_rec.mtu = IB_MTU_512;
 630                 break;
 631         case 2048:
 632                 p_rec.mtu = IB_MTU_1024;
 633                 break;
 634         case 4096:
 635                 p_rec.mtu = IB_MTU_2048;
 636                 break;
 637         default:
 638                 /* Wildcard everything */
 639                 comp_mask = 0;
 640                 p_rec.mtu = 0;
 641                 p_rec.mtu_selector = 0;
 642         }
 643
 644         ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
 645                   p_rec.dgid.raw, ":",
 646                   comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
 647
 648         init_completion(&path->done);
 649
 650         path->query_id =
 651                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 652                                    &p_rec, comp_mask            |
 653                                    IB_SA_PATH_REC_DGID          |
 654                                    IB_SA_PATH_REC_SGID          |
 655                                    IB_SA_PATH_REC_NUMB_PATH     |
 656                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 657                                    IB_SA_PATH_REC_PKEY,
 658                                    1000, GFP_ATOMIC,
 659                                    path_rec_completion,
 660                                    path, &path->query);
 661         if (path->query_id < 0) {
 662                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 663                 path->query = NULL;
 664                 complete(&path->done);
 665                 return path->query_id;
 666         }
 667
 668         return 0;
 669 }
 670
 671 static void
 672 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
 673 {
 674         struct ipoib_path *path;
 675
 676         path = __path_find(priv, eh->hwaddr + 4);
 677         if (!path || !path->valid) {
 678                 int new_path = 0;
 679
 680                 if (!path) {
 681                         path = path_rec_create(priv, eh->hwaddr);
 682                         new_path = 1;
 683                 }
 684                 if (path) {
 685                         if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 686                                 _IF_ENQUEUE(&path->queue, mb);
 687                         else {
 688                                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 689                                 m_freem(mb);
 690                         }
 691
 692                         if (!path->query && path_rec_start(priv, path)) {
 693                                 spin_unlock_irqrestore(&priv->lock, flags);
 694                                 if (new_path)
 695                                         ipoib_path_free(priv, path);
 696                                 return;
 697                         } else
 698                                 __path_add(priv, path);
 699                 } else {
 700                         if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 701                         m_freem(mb);
 702                 }
 703
 704                 return;
 705         }
 706
 707         if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
 708                 ipoib_cm_send(priv, mb, ipoib_cm_get(path));
 709         } else if (path->ah) {
 710                 ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
 711         } else if ((path->query || !path_rec_start(priv, path)) &&
 712                     path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
 713                 _IF_ENQUEUE(&path->queue, mb);
 714         } else {
 715                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 716                 m_freem(mb);
 717         }
 718 }
 719
 720 static int
 721 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
 722 {
 723         struct ipoib_header *eh;
 724
 725         eh = mtod(mb, struct ipoib_header *);
 726         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
 727                 /* Add in the P_Key for multicast*/
 728                 eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 729                 eh->hwaddr[9] = priv->pkey & 0xff;
 730
 731                 ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
 732         } else
 733                 ipoib_unicast_send(mb, priv, eh);
 734
 735         return 0;
 736 }
 737
 738
 739 static void
 740 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
 741 {
 742         struct mbuf *mb;
 743
 744         if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
 745             IFF_DRV_RUNNING)
 746                 return;
 747
 748         spin_lock(&priv->lock);
 749         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
 750             (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
 751                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 752                 if (mb == NULL)
 753                         break;
 754                 IPOIB_MTAP(dev, mb);
 755                 ipoib_send_one(priv, mb);
 756         }
 757         spin_unlock(&priv->lock);
 758 }
 759
 760 static void
 761 ipoib_start(struct ifnet *dev)
 762 {
 763         _ipoib_start(dev, dev->if_softc);
 764 }
 765
 766 static void
 767 ipoib_vlan_start(struct ifnet *dev)
 768 {
 769         struct ipoib_dev_priv *priv;
 770         struct mbuf *mb;
 771
 772         priv = VLAN_COOKIE(dev);
 773         if (priv != NULL)
 774                 return _ipoib_start(dev, priv);
 775         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
 776                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 777                 if (mb == NULL)
 778                         break;
 779                 m_freem(mb);
 780                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 781         }
 782 }
 783
 784 int
 785 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
 786 {
 787
 788         /* Allocate RX/TX "rings" to hold queued mbs */
 789         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 790                                 GFP_KERNEL);
 791         if (!priv->rx_ring) {
 792                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 793                        ca->name, ipoib_recvq_size);
 794                 goto out;
 795         }
 796
 797         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
 798         if (!priv->tx_ring) {
 799                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 800                        ca->name, ipoib_sendq_size);
 801                 goto out_rx_ring_cleanup;
 802         }
 803         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 804
 805         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 806
 807         if (ipoib_ib_dev_init(priv, ca, port))
 808                 goto out_tx_ring_cleanup;
 809
 810         return 0;
 811
 812 out_tx_ring_cleanup:
 813         kfree(priv->tx_ring);
 814
 815 out_rx_ring_cleanup:
 816         kfree(priv->rx_ring);
 817
 818 out:
 819         return -ENOMEM;
 820 }
 821
 822 static void
 823 ipoib_detach(struct ipoib_dev_priv *priv)
 824 {
 825         struct ifnet *dev;
 826
 827         dev = priv->dev;
 828         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 829                 priv->gone = 1;
 830                 bpfdetach(dev);
 831                 if_detach(dev);
 832                 if_free(dev);
 833                 free_unr(ipoib_unrhdr, priv->unit);
 834         } else
 835                 VLAN_SETCOOKIE(priv->dev, NULL);
 836
 837         free(priv, M_TEMP);
 838 }
 839
 840 void
 841 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
 842 {
 843         struct ipoib_dev_priv *cpriv, *tcpriv;
 844
 845         /* Delete any child interfaces first */
 846         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 847                 ipoib_dev_cleanup(cpriv);
 848                 ipoib_detach(cpriv);
 849         }
 850
 851         ipoib_ib_dev_cleanup(priv);
 852
 853         kfree(priv->rx_ring);
 854         kfree(priv->tx_ring);
 855
 856         priv->rx_ring = NULL;
 857         priv->tx_ring = NULL;
 858 }
 859
 860 static struct ipoib_dev_priv *
 861 ipoib_priv_alloc(void)
 862 {
 863         struct ipoib_dev_priv *priv;
 864
 865         priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
 866         spin_lock_init(&priv->lock);
 867         spin_lock_init(&priv->drain_lock);
 868         mutex_init(&priv->vlan_mutex);
 869         INIT_LIST_HEAD(&priv->path_list);
 870         INIT_LIST_HEAD(&priv->child_intfs);
 871         INIT_LIST_HEAD(&priv->dead_ahs);
 872         INIT_LIST_HEAD(&priv->multicast_list);
 873         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 874         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 875         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
 876         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
 877         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
 878         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
 879         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 880         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 881         memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
 882
 883         return (priv);
 884 }
 885
 886 struct ipoib_dev_priv *
 887 ipoib_intf_alloc(const char *name)
 888 {
 889         struct ipoib_dev_priv *priv;
 890         struct sockaddr_dl *sdl;
 891         struct ifnet *dev;
 892
 893         priv = ipoib_priv_alloc();
 894         dev = priv->dev = if_alloc(IFT_INFINIBAND);
 895         if (!dev) {
 896                 free(priv, M_TEMP);
 897                 return NULL;
 898         }
 899         dev->if_softc = priv;
 900         priv->unit = alloc_unr(ipoib_unrhdr);
 901         if (priv->unit == -1) {
 902                 if_free(dev);
 903                 free(priv, M_TEMP);
 904                 return NULL;
 905         }
 906         if_initname(dev, name, priv->unit);
 907         dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
 908         dev->if_addrlen = INFINIBAND_ALEN;
 909         dev->if_hdrlen = IPOIB_HEADER_LEN;
 910         if_attach(dev);
 911         dev->if_init = ipoib_init;
 912         dev->if_ioctl = ipoib_ioctl;
 913         dev->if_start = ipoib_start;
 914         dev->if_output = ipoib_output;
 915         dev->if_input = ipoib_input;
 916         dev->if_resolvemulti = ipoib_resolvemulti;
 917         dev->if_baudrate = IF_Gbps(10);
 918         dev->if_broadcastaddr = priv->broadcastaddr;
 919         dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
 920         sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
 921         sdl->sdl_type = IFT_INFINIBAND;
 922         sdl->sdl_alen = dev->if_addrlen;
 923         priv->dev = dev;
 924         if_link_state_change(dev, LINK_STATE_DOWN);
 925         bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
 926
 927         return dev->if_softc;
 928 }
 929
 930 int
 931 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
 932 {
 933         struct ib_device_attr *device_attr = &hca->attrs;
 934
 935         priv->hca_caps = device_attr->device_cap_flags;
 936
 937         priv->dev->if_hwassist = 0;
 938         priv->dev->if_capabilities = 0;
 939
 940 #ifndef CONFIG_INFINIBAND_IPOIB_CM
 941         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
 942                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
 943                 priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
 944                 priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
 945         }
 946
 947 #if 0
 948         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
 949                 priv->dev->if_capabilities |= IFCAP_TSO4;
 950                 priv->dev->if_hwassist |= CSUM_TSO;
 951         }
 952 #endif
 953 #endif
 954         priv->dev->if_capabilities |=
 955             IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
 956         priv->dev->if_capenable = priv->dev->if_capabilities;
 957
 958         return 0;
 959 }
 960
 961
 962 static struct ifnet *
 963 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
 964 {
 965         struct ipoib_dev_priv *priv;
 966         struct ib_port_attr attr;
 967         int result = -ENOMEM;
 968
 969         priv = ipoib_intf_alloc(format);
 970         if (!priv)
 971                 goto alloc_mem_failed;
 972
 973         if (!ib_query_port(hca, port, &attr))
 974                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
 975         else {
 976                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
 977                        hca->name, port);
 978                 goto device_init_failed;
 979         }
 980
 981         /* MTU will be reset when mcast join happens */
 982         priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
 983         priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
 984
 985         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 986         if (result) {
 987                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 988                        hca->name, port, result);
 989                 goto device_init_failed;
 990         }
 991
 992         if (ipoib_set_dev_features(priv, hca))
 993                 goto device_init_failed;
 994
 995         /*
 996          * Set the full membership bit, so that we join the right
 997          * broadcast group, etc.
 998          */
 999         priv->pkey |= 0x8000;
1000
1001         priv->broadcastaddr[8] = priv->pkey >> 8;
1002         priv->broadcastaddr[9] = priv->pkey & 0xff;
1003
1004         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1005         if (result) {
1006                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1007                        hca->name, port, result);
1008                 goto device_init_failed;
1009         }
1010         memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
1011
1012         result = ipoib_dev_init(priv, hca, port);
1013         if (result < 0) {
1014                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1015                        hca->name, port, result);
1016                 goto device_init_failed;
1017         }
1018         if (ipoib_cm_admin_enabled(priv))
1019                 priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1020
1021         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1022                               priv->ca, ipoib_event);
1023         result = ib_register_event_handler(&priv->event_handler);
1024         if (result < 0) {
1025                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1026                        "port %d (ret = %d)\n",
1027                        hca->name, port, result);
1028                 goto event_failed;
1029         }
1030         if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1031
1032         return priv->dev;
1033
1034 event_failed:
1035         ipoib_dev_cleanup(priv);
1036
1037 device_init_failed:
1038         ipoib_detach(priv);
1039
1040 alloc_mem_failed:
1041         return ERR_PTR(result);
1042 }
1043
1044 static void
1045 ipoib_add_one(struct ib_device *device)
1046 {
1047         struct list_head *dev_list;
1048         struct ifnet *dev;
1049         struct ipoib_dev_priv *priv;
1050         int s, e, p;
1051
1052         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1053                 return;
1054
1055         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1056         if (!dev_list)
1057                 return;
1058
1059         INIT_LIST_HEAD(dev_list);
1060
1061         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1062                 s = 0;
1063                 e = 0;
1064         } else {
1065                 s = 1;
1066                 e = device->phys_port_cnt;
1067         }
1068
1069         for (p = s; p <= e; ++p) {
1070                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1071                         continue;
1072                 dev = ipoib_add_port("ib", device, p);
1073                 if (!IS_ERR(dev)) {
1074                         priv = dev->if_softc;
1075                         list_add_tail(&priv->list, dev_list);
1076                 }
1077         }
1078
1079         ib_set_client_data(device, &ipoib_client, dev_list);
1080 }
1081
1082 static void
1083 ipoib_remove_one(struct ib_device *device, void *client_data)
1084 {
1085         struct ipoib_dev_priv *priv, *tmp;
1086         struct list_head *dev_list = client_data;
1087
1088         if (!dev_list)
1089                 return;
1090
1091         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1092                 return;
1093
1094         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1095                 if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1096                         continue;
1097
1098                 ipoib_stop(priv);
1099
1100                 ib_unregister_event_handler(&priv->event_handler);
1101
1102                 /* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1103
1104                 flush_workqueue(ipoib_workqueue);
1105
1106                 ipoib_dev_cleanup(priv);
1107                 ipoib_detach(priv);
1108         }
1109
1110         kfree(dev_list);
1111 }
1112
1113 static void
1114 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1115 {
1116         struct ipoib_dev_priv *parent;
1117         struct ipoib_dev_priv *priv;
1118         struct ifnet *dev;
1119         uint16_t pkey;
1120         int error;
1121
1122         if (ifp->if_type != IFT_INFINIBAND)
1123                 return;
1124         dev = VLAN_DEVAT(ifp, vtag);
1125         if (dev == NULL)
1126                 return;
1127         priv = NULL;
1128         error = 0;
1129         parent = ifp->if_softc;
1130         /* We only support 15 bits of pkey. */
1131         if (vtag & 0x8000)
1132                 return;
1133         pkey = vtag | 0x8000;   /* Set full membership bit. */
1134         if (pkey == parent->pkey)
1135                 return;
1136         /* Check for dups */
1137         mutex_lock(&parent->vlan_mutex);
1138         list_for_each_entry(priv, &parent->child_intfs, list) {
1139                 if (priv->pkey == pkey) {
1140                         priv = NULL;
1141                         error = EBUSY;
1142                         goto out;
1143                 }
1144         }
1145         priv = ipoib_priv_alloc();
1146         priv->dev = dev;
1147         priv->max_ib_mtu = parent->max_ib_mtu;
1148         priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1149         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1150         error = ipoib_set_dev_features(priv, parent->ca);
1151         if (error)
1152                 goto out;
1153         priv->pkey = pkey;
1154         priv->broadcastaddr[8] = pkey >> 8;
1155         priv->broadcastaddr[9] = pkey & 0xff;
1156         dev->if_broadcastaddr = priv->broadcastaddr;
1157         error = ipoib_dev_init(priv, parent->ca, parent->port);
1158         if (error)
1159                 goto out;
1160         priv->parent = parent->dev;
1161         list_add_tail(&priv->list, &parent->child_intfs);
1162         VLAN_SETCOOKIE(dev, priv);
1163         dev->if_start = ipoib_vlan_start;
1164         dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1165         dev->if_hdrlen = IPOIB_HEADER_LEN;
1166         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1167                 ipoib_open(priv);
1168         mutex_unlock(&parent->vlan_mutex);
1169         return;
1170 out:
1171         mutex_unlock(&parent->vlan_mutex);
1172         if (priv)
1173                 free(priv, M_TEMP);
1174         if (error)
1175                 ipoib_warn(parent,
1176                     "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1177                     parent->ca->name, parent->port, vtag);
1178         return;
1179 }
1180
1181 static void
1182 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1183 {
1184         struct ipoib_dev_priv *parent;
1185         struct ipoib_dev_priv *priv;
1186         struct ifnet *dev;
1187         uint16_t pkey;
1188
1189         if (ifp->if_type != IFT_INFINIBAND)
1190                 return;
1191
1192         dev = VLAN_DEVAT(ifp, vtag);
1193         if (dev)
1194                 VLAN_SETCOOKIE(dev, NULL);
1195         pkey = vtag | 0x8000;
1196         parent = ifp->if_softc;
1197         mutex_lock(&parent->vlan_mutex);
1198         list_for_each_entry(priv, &parent->child_intfs, list) {
1199                 if (priv->pkey == pkey) {
1200                         ipoib_dev_cleanup(priv);
1201                         list_del(&priv->list);
1202                         break;
1203                 }
1204         }
1205         mutex_unlock(&parent->vlan_mutex);
1206 }
1207
1208 eventhandler_tag ipoib_vlan_attach;
1209 eventhandler_tag ipoib_vlan_detach;
1210
1211 static int __init
1212 ipoib_init_module(void)
1213 {
1214         int ret;
1215
1216         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1217         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1218         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1219
1220         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1221         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1222         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1223                                                      IPOIB_MIN_QUEUE_SIZE));
1224 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1225         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1226 #endif
1227
1228         ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1229                 ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1230         ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1231                 ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1232
1233         /*
1234          * We create our own workqueue mainly because we want to be
1235          * able to flush it when devices are being removed.  We can't
1236          * use schedule_work()/flush_scheduled_work() because both
1237          * unregister_netdev() and linkwatch_event take the rtnl lock,
1238          * so flush_scheduled_work() can deadlock during device
1239          * removal.
1240          */
1241         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1242         if (!ipoib_workqueue) {
1243                 ret = -ENOMEM;
1244                 goto err_fs;
1245         }
1246
1247         ib_sa_register_client(&ipoib_sa_client);
1248
1249         ret = ib_register_client(&ipoib_client);
1250         if (ret)
1251                 goto err_sa;
1252
1253         return 0;
1254
1255 err_sa:
1256         ib_sa_unregister_client(&ipoib_sa_client);
1257         destroy_workqueue(ipoib_workqueue);
1258
1259 err_fs:
1260         return ret;
1261 }
1262
1263 static void __exit
1264 ipoib_cleanup_module(void)
1265 {
1266
1267         EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1268         EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1269         ib_unregister_client(&ipoib_client);
1270         ib_sa_unregister_client(&ipoib_sa_client);
1271         destroy_workqueue(ipoib_workqueue);
1272 }
1273
1274 /*
1275  * Infiniband output routine.
1276  */
1277 static int
1278 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1279         const struct sockaddr *dst, struct route *ro)
1280 {
1281         u_char edst[INFINIBAND_ALEN];
1282 #if defined(INET) || defined(INET6)
1283         struct llentry *lle = NULL;
1284 #endif
1285         struct ipoib_header *eh;
1286         int error = 0, is_gw = 0;
1287         short type;
1288
1289         if (ro != NULL)
1290                 is_gw = (ro->ro_flags & RT_HAS_GW) != 0;
1291 #ifdef MAC
1292         error = mac_ifnet_check_transmit(ifp, m);
1293         if (error)
1294                 goto bad;
1295 #endif
1296
1297         M_PROFILE(m);
1298         if (ifp->if_flags & IFF_MONITOR) {
1299                 error = ENETDOWN;
1300                 goto bad;
1301         }
1302         if (!((ifp->if_flags & IFF_UP) &&
1303             (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1304                 error = ENETDOWN;
1305                 goto bad;
1306         }
1307
1308         switch (dst->sa_family) {
1309 #ifdef INET
1310         case AF_INET:
1311                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1312                         memcpy(edst, lle->ll_addr, sizeof(edst));
1313                 else if (m->m_flags & M_MCAST)
1314                         ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1315                 else
1316                         error = arpresolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1317                 if (error)
1318                         return (error == EWOULDBLOCK ? 0 : error);
1319                 type = htons(ETHERTYPE_IP);
1320                 break;
1321         case AF_ARP:
1322         {
1323                 struct arphdr *ah;
1324                 ah = mtod(m, struct arphdr *);
1325                 ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1326
1327                 switch(ntohs(ah->ar_op)) {
1328                 case ARPOP_REVREQUEST:
1329                 case ARPOP_REVREPLY:
1330                         type = htons(ETHERTYPE_REVARP);
1331                         break;
1332                 case ARPOP_REQUEST:
1333                 case ARPOP_REPLY:
1334                 default:
1335                         type = htons(ETHERTYPE_ARP);
1336                         break;
1337                 }
1338
1339                 if (m->m_flags & M_BCAST)
1340                         bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1341                 else
1342                         bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1343
1344         }
1345         break;
1346 #endif
1347 #ifdef INET6
1348         case AF_INET6:
1349                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1350                         memcpy(edst, lle->ll_addr, sizeof(edst));
1351                 else if (m->m_flags & M_MCAST)
1352                         ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1353                 else
1354                         error = nd6_resolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1355                 if (error)
1356                         return error;
1357                 type = htons(ETHERTYPE_IPV6);
1358                 break;
1359 #endif
1360
1361         default:
1362                 if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1363                 error = EAFNOSUPPORT;
1364                 goto bad;
1365         }
1366
1367         /*
1368          * Add local net header.  If no space in first mbuf,
1369          * allocate another.
1370          */
1371         M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1372         if (m == NULL) {
1373                 error = ENOBUFS;
1374                 goto bad;
1375         }
1376         eh = mtod(m, struct ipoib_header *);
1377         (void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1378         (void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1379
1380         /*
1381          * Queue message on interface, update output statistics if
1382          * successful, and start output if interface not yet active.
1383          */
1384         return ((ifp->if_transmit)(ifp, m));
1385 bad:
1386         if (m != NULL)
1387                 m_freem(m);
1388         return (error);
1389 }
1390
1391 /*
1392  * Upper layer processing for a received Infiniband packet.
1393  */
1394 void
1395 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1396 {
1397         int isr;
1398
1399 #ifdef MAC
1400         /*
1401          * Tag the mbuf with an appropriate MAC label before any other
1402          * consumers can get to it.
1403          */
1404         mac_ifnet_create_mbuf(ifp, m);
1405 #endif
1406         /* Allow monitor mode to claim this frame, after stats are updated. */
1407         if (ifp->if_flags & IFF_MONITOR) {
1408                 if_printf(ifp, "discard frame at IFF_MONITOR\n");
1409                 m_freem(m);
1410                 return;
1411         }
1412         /*
1413          * Dispatch frame to upper layer.
1414          */
1415         switch (proto) {
1416 #ifdef INET
1417         case ETHERTYPE_IP:
1418                 isr = NETISR_IP;
1419                 break;
1420
1421         case ETHERTYPE_ARP:
1422                 if (ifp->if_flags & IFF_NOARP) {
1423                         /* Discard packet if ARP is disabled on interface */
1424                         m_freem(m);
1425                         return;
1426                 }
1427                 isr = NETISR_ARP;
1428                 break;
1429 #endif
1430 #ifdef INET6
1431         case ETHERTYPE_IPV6:
1432                 isr = NETISR_IPV6;
1433                 break;
1434 #endif
1435         default:
1436                 goto discard;
1437         }
1438         netisr_dispatch(isr, m);
1439         return;
1440
1441 discard:
1442         m_freem(m);
1443 }
1444
1445 /*
1446  * Process a received Infiniband packet.
1447  */
1448 static void
1449 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1450 {
1451         struct ipoib_header *eh;
1452
1453         if ((ifp->if_flags & IFF_UP) == 0) {
1454                 m_freem(m);
1455                 return;
1456         }
1457         CURVNET_SET_QUIET(ifp->if_vnet);
1458
1459         /* Let BPF have it before we strip the header. */
1460         IPOIB_MTAP(ifp, m);
1461         eh = mtod(m, struct ipoib_header *);
1462         /*
1463          * Reset layer specific mbuf flags to avoid confusing upper layers.
1464          * Strip off Infiniband header.
1465          */
1466         m->m_flags &= ~M_VLANTAG;
1467         m_clrprotoflags(m);
1468         m_adj(m, IPOIB_HEADER_LEN);
1469
1470         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1471                 if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1472                     ifp->if_addrlen) == 0)
1473                         m->m_flags |= M_BCAST;
1474                 else
1475                         m->m_flags |= M_MCAST;
1476                 if_inc_counter(ifp, IFCOUNTER_IMCASTS, 1);
1477         }
1478
1479         ipoib_demux(ifp, m, ntohs(eh->proto));
1480         CURVNET_RESTORE();
1481 }
1482
1483 static int
1484 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1485         struct sockaddr *sa)
1486 {
1487         struct sockaddr_dl *sdl;
1488 #ifdef INET
1489         struct sockaddr_in *sin;
1490 #endif
1491 #ifdef INET6
1492         struct sockaddr_in6 *sin6;
1493 #endif
1494         u_char *e_addr;
1495
1496         switch(sa->sa_family) {
1497         case AF_LINK:
1498                 /*
1499                  * No mapping needed. Just check that it's a valid MC address.
1500                  */
1501                 sdl = (struct sockaddr_dl *)sa;
1502                 e_addr = LLADDR(sdl);
1503                 if (!IPOIB_IS_MULTICAST(e_addr))
1504                         return EADDRNOTAVAIL;
1505                 *llsa = NULL;
1506                 return 0;
1507
1508 #ifdef INET
1509         case AF_INET:
1510                 sin = (struct sockaddr_in *)sa;
1511                 if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1512                         return EADDRNOTAVAIL;
1513                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1514                 sdl->sdl_alen = INFINIBAND_ALEN;
1515                 e_addr = LLADDR(sdl);
1516                 ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1517                     e_addr);
1518                 *llsa = (struct sockaddr *)sdl;
1519                 return 0;
1520 #endif
1521 #ifdef INET6
1522         case AF_INET6:
1523                 sin6 = (struct sockaddr_in6 *)sa;
1524                 /*
1525                  * An IP6 address of 0 means listen to all
1526                  * of the multicast address used for IP6.
1527                  * This has no meaning in ipoib.
1528                  */
1529                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1530                         return EADDRNOTAVAIL;
1531                 if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1532                         return EADDRNOTAVAIL;
1533                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1534                 sdl->sdl_alen = INFINIBAND_ALEN;
1535                 e_addr = LLADDR(sdl);
1536                 ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1537                 *llsa = (struct sockaddr *)sdl;
1538                 return 0;
1539 #endif
1540
1541         default:
1542                 return EAFNOSUPPORT;
1543         }
1544 }
1545
1546 module_init(ipoib_init_module);
1547 module_exit(ipoib_cleanup_module);
1548
1549 static int
1550 ipoib_evhand(module_t mod, int event, void *arg)
1551 {
1552                         return (0);
1553 }
1554
1555 static moduledata_t ipoib_mod = {
1556                         .name = "ipoib",
1557                                         .evhand = ipoib_evhand,
1558 };
1559
1560 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1561 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1562 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);