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 <sys/cdefs.h>
  38 __FBSDID("$FreeBSD$");
  39
  40 #include "ipoib.h"
  41
  42 static  int ipoib_resolvemulti(struct ifnet *, struct sockaddr **,
  43                 struct sockaddr *);
  44
  45
  46 #include <linux/module.h>
  47
  48 #include <linux/slab.h>
  49 #include <linux/kernel.h>
  50 #include <linux/vmalloc.h>
  51
  52 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
  53 #include <linux/if_vlan.h>
  54 #include <net/ip.h>
  55 #include <net/ipv6.h>
  56
  57 #include <rdma/ib_cache.h>
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
  64 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
  65
  66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
  67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
  68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
  69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
  70
  71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  72 int ipoib_debug_level = 1;
  73
  74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
  75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  76 #endif
  77
  78 struct ipoib_path_iter {
  79         struct ipoib_dev_priv *priv;
  80         struct ipoib_path  path;
  81 };
  82
  83 static const u8 ipv4_bcast_addr[] = {
  84         0x00, 0xff, 0xff, 0xff,
  85         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
  86         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
  87 };
  88
  89 struct workqueue_struct *ipoib_workqueue;
  90
  91 struct ib_sa_client ipoib_sa_client;
  92
  93 static void ipoib_add_one(struct ib_device *device);
  94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
  95 static struct net_device *ipoib_get_net_dev_by_params(
  96                 struct ib_device *dev, u8 port, u16 pkey,
  97                 const union ib_gid *gid, const struct sockaddr *addr,
  98                 void *client_data);
  99 static void ipoib_start(struct ifnet *dev);
 100 static int ipoib_output(struct ifnet *ifp, struct mbuf *m,
 101             const struct sockaddr *dst, struct route *ro);
 102 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
 103 static void ipoib_input(struct ifnet *ifp, struct mbuf *m);
 104
 105 #define IPOIB_MTAP(_ifp, _m)                                    \
 106 do {                                                            \
 107         if (bpf_peers_present((_ifp)->if_bpf)) {                \
 108                 M_ASSERTVALID(_m);                              \
 109                 ipoib_mtap_mb((_ifp), (_m));                    \
 110         }                                                       \
 111 } while (0)
 112
 113 static struct unrhdr *ipoib_unrhdr;
 114
 115 static void
 116 ipoib_unrhdr_init(void *arg)
 117 {
 118
 119         ipoib_unrhdr = new_unrhdr(0, 65535, NULL);
 120 }
 121 SYSINIT(ipoib_unrhdr_init, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_init, NULL);
 122
 123 static void
 124 ipoib_unrhdr_uninit(void *arg)
 125 {
 126
 127         if (ipoib_unrhdr != NULL) {
 128                 struct unrhdr *hdr;
 129
 130                 hdr = ipoib_unrhdr;
 131                 ipoib_unrhdr = NULL;
 132
 133                 delete_unrhdr(hdr);
 134         }
 135 }
 136 SYSUNINIT(ipoib_unrhdr_uninit, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_uninit, NULL);
 137
 138 /*
 139  * This is for clients that have an ipoib_header in the mbuf.
 140  */
 141 static void
 142 ipoib_mtap_mb(struct ifnet *ifp, struct mbuf *mb)
 143 {
 144         struct ipoib_header *ih;
 145         struct ether_header eh;
 146
 147         ih = mtod(mb, struct ipoib_header *);
 148         eh.ether_type = ih->proto;
 149         bcopy(ih->hwaddr, &eh.ether_dhost, ETHER_ADDR_LEN);
 150         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 151         mb->m_data += sizeof(struct ipoib_header);
 152         mb->m_len -= sizeof(struct ipoib_header);
 153         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 154         mb->m_data -= sizeof(struct ipoib_header);
 155         mb->m_len += sizeof(struct ipoib_header);
 156 }
 157
 158 void
 159 ipoib_mtap_proto(struct ifnet *ifp, struct mbuf *mb, uint16_t proto)
 160 {
 161         struct ether_header eh;
 162
 163         eh.ether_type = proto;
 164         bzero(&eh.ether_shost, ETHER_ADDR_LEN);
 165         bzero(&eh.ether_dhost, ETHER_ADDR_LEN);
 166         bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
 167 }
 168
 169 static struct ib_client ipoib_client = {
 170         .name   = "ipoib",
 171         .add    = ipoib_add_one,
 172         .remove = ipoib_remove_one,
 173         .get_net_dev_by_params = ipoib_get_net_dev_by_params,
 174 };
 175
 176 int
 177 ipoib_open(struct ipoib_dev_priv *priv)
 178 {
 179         struct ifnet *dev = priv->dev;
 180
 181         ipoib_dbg(priv, "bringing up interface\n");
 182
 183         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 184
 185         if (ipoib_pkey_dev_delay_open(priv))
 186                 return 0;
 187
 188         if (ipoib_ib_dev_open(priv))
 189                 goto err_disable;
 190
 191         if (ipoib_ib_dev_up(priv))
 192                 goto err_stop;
 193
 194         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 195                 struct ipoib_dev_priv *cpriv;
 196
 197                 /* Bring up any child interfaces too */
 198                 mutex_lock(&priv->vlan_mutex);
 199                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 200                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 201                                 ipoib_open(cpriv);
 202                 mutex_unlock(&priv->vlan_mutex);
 203         }
 204         dev->if_drv_flags |= IFF_DRV_RUNNING;
 205         dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
 206
 207         return 0;
 208
 209 err_stop:
 210         ipoib_ib_dev_stop(priv, 1);
 211
 212 err_disable:
 213         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 214
 215         return -EINVAL;
 216 }
 217
 218 static void
 219 ipoib_init(void *arg)
 220 {
 221         struct ifnet *dev;
 222         struct ipoib_dev_priv *priv;
 223
 224         priv = arg;
 225         dev = priv->dev;
 226         if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
 227                 ipoib_open(priv);
 228         queue_work(ipoib_workqueue, &priv->flush_light);
 229 }
 230
 231
 232 static int
 233 ipoib_stop(struct ipoib_dev_priv *priv)
 234 {
 235         struct ifnet *dev = priv->dev;
 236
 237         ipoib_dbg(priv, "stopping interface\n");
 238
 239         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
 240
 241         dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
 242
 243         ipoib_ib_dev_down(priv, 0);
 244         ipoib_ib_dev_stop(priv, 0);
 245
 246         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 247                 struct ipoib_dev_priv *cpriv;
 248
 249                 /* Bring down any child interfaces too */
 250                 mutex_lock(&priv->vlan_mutex);
 251                 list_for_each_entry(cpriv, &priv->child_intfs, list)
 252                         if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
 253                                 ipoib_stop(cpriv);
 254                 mutex_unlock(&priv->vlan_mutex);
 255         }
 256
 257         return 0;
 258 }
 259
 260 int
 261 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu)
 262 {
 263         struct ifnet *dev = priv->dev;
 264
 265         /* dev->if_mtu > 2K ==> connected mode */
 266         if (ipoib_cm_admin_enabled(priv)) {
 267                 if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
 268                         return -EINVAL;
 269
 270                 if (new_mtu > priv->mcast_mtu)
 271                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
 272                                    priv->mcast_mtu);
 273
 274                 dev->if_mtu = new_mtu;
 275                 return 0;
 276         }
 277
 278         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
 279                 return -EINVAL;
 280
 281         priv->admin_mtu = new_mtu;
 282
 283         dev->if_mtu = min(priv->mcast_mtu, priv->admin_mtu);
 284
 285         queue_work(ipoib_workqueue, &priv->flush_light);
 286
 287         return 0;
 288 }
 289
 290 static int
 291 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
 292 {
 293         struct ipoib_dev_priv *priv = ifp->if_softc;
 294         struct ifaddr *ifa = (struct ifaddr *) data;
 295         struct ifreq *ifr = (struct ifreq *) data;
 296         int error = 0;
 297
 298         /* check if detaching */
 299         if (priv == NULL || priv->gone != 0)
 300                 return (ENXIO);
 301
 302         switch (command) {
 303         case SIOCSIFFLAGS:
 304                 if (ifp->if_flags & IFF_UP) {
 305                         if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
 306                                 error = -ipoib_open(priv);
 307                 } else
 308                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 309                                 ipoib_stop(priv);
 310                 break;
 311         case SIOCADDMULTI:
 312         case SIOCDELMULTI:
 313                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 314                         queue_work(ipoib_workqueue, &priv->restart_task);
 315                 break;
 316         case SIOCSIFADDR:
 317                 ifp->if_flags |= IFF_UP;
 318
 319                 switch (ifa->ifa_addr->sa_family) {
 320 #ifdef INET
 321                 case AF_INET:
 322                         ifp->if_init(ifp->if_softc);    /* before arpwhohas */
 323                         arp_ifinit(ifp, ifa);
 324                         break;
 325 #endif
 326                 default:
 327                         ifp->if_init(ifp->if_softc);
 328                         break;
 329                 }
 330                 break;
 331
 332         case SIOCGIFADDR:
 333                         bcopy(IF_LLADDR(ifp), &ifr->ifr_addr.sa_data[0],
 334                             INFINIBAND_ALEN);
 335                 break;
 336
 337         case SIOCSIFMTU:
 338                 /*
 339                  * Set the interface MTU.
 340                  */
 341                 error = -ipoib_change_mtu(priv, ifr->ifr_mtu);
 342                 break;
 343         default:
 344                 error = EINVAL;
 345                 break;
 346         }
 347         return (error);
 348 }
 349
 350
 351 static struct ipoib_path *
 352 __path_find(struct ipoib_dev_priv *priv, void *gid)
 353 {
 354         struct rb_node *n = priv->path_tree.rb_node;
 355         struct ipoib_path *path;
 356         int ret;
 357
 358         while (n) {
 359                 path = rb_entry(n, struct ipoib_path, rb_node);
 360
 361                 ret = memcmp(gid, path->pathrec.dgid.raw,
 362                              sizeof (union ib_gid));
 363
 364                 if (ret < 0)
 365                         n = n->rb_left;
 366                 else if (ret > 0)
 367                         n = n->rb_right;
 368                 else
 369                         return path;
 370         }
 371
 372         return NULL;
 373 }
 374
 375 static int
 376 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 377 {
 378         struct rb_node **n = &priv->path_tree.rb_node;
 379         struct rb_node *pn = NULL;
 380         struct ipoib_path *tpath;
 381         int ret;
 382
 383         while (*n) {
 384                 pn = *n;
 385                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 386
 387                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 388                              sizeof (union ib_gid));
 389                 if (ret < 0)
 390                         n = &pn->rb_left;
 391                 else if (ret > 0)
 392                         n = &pn->rb_right;
 393                 else
 394                         return -EEXIST;
 395         }
 396
 397         rb_link_node(&path->rb_node, pn, n);
 398         rb_insert_color(&path->rb_node, &priv->path_tree);
 399
 400         list_add_tail(&path->list, &priv->path_list);
 401
 402         return 0;
 403 }
 404
 405 void
 406 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 407 {
 408
 409         _IF_DRAIN(&path->queue);
 410
 411         if (path->ah)
 412                 ipoib_put_ah(path->ah);
 413         if (ipoib_cm_get(path))
 414                 ipoib_cm_destroy_tx(ipoib_cm_get(path));
 415
 416         kfree(path);
 417 }
 418
 419 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 420
 421 struct ipoib_path_iter *
 422 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
 423 {
 424         struct ipoib_path_iter *iter;
 425
 426         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 427         if (!iter)
 428                 return NULL;
 429
 430         iter->priv = priv;
 431         memset(iter->path.pathrec.dgid.raw, 0, 16);
 432
 433         if (ipoib_path_iter_next(iter)) {
 434                 kfree(iter);
 435                 return NULL;
 436         }
 437
 438         return iter;
 439 }
 440
 441 int
 442 ipoib_path_iter_next(struct ipoib_path_iter *iter)
 443 {
 444         struct ipoib_dev_priv *priv = iter->priv;
 445         struct rb_node *n;
 446         struct ipoib_path *path;
 447         int ret = 1;
 448
 449         spin_lock_irq(&priv->lock);
 450
 451         n = rb_first(&priv->path_tree);
 452
 453         while (n) {
 454                 path = rb_entry(n, struct ipoib_path, rb_node);
 455
 456                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 457                            sizeof (union ib_gid)) < 0) {
 458                         iter->path = *path;
 459                         ret = 0;
 460                         break;
 461                 }
 462
 463                 n = rb_next(n);
 464         }
 465
 466         spin_unlock_irq(&priv->lock);
 467
 468         return ret;
 469 }
 470
 471 void
 472 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
 473 {
 474         *path = iter->path;
 475 }
 476
 477 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 478
 479 void
 480 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
 481 {
 482         struct ipoib_path *path, *tp;
 483
 484         spin_lock_irq(&priv->lock);
 485
 486         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 487                 ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
 488                         be16_to_cpu(path->pathrec.dlid),
 489                         path->pathrec.dgid.raw, ":");
 490                 path->valid =  0;
 491         }
 492
 493         spin_unlock_irq(&priv->lock);
 494 }
 495
 496 void
 497 ipoib_flush_paths(struct ipoib_dev_priv *priv)
 498 {
 499         struct ipoib_path *path, *tp;
 500         LIST_HEAD(remove_list);
 501         unsigned long flags;
 502
 503         spin_lock_irqsave(&priv->lock, flags);
 504
 505         list_splice_init(&priv->path_list, &remove_list);
 506
 507         list_for_each_entry(path, &remove_list, list)
 508                 rb_erase(&path->rb_node, &priv->path_tree);
 509
 510         list_for_each_entry_safe(path, tp, &remove_list, list) {
 511                 if (path->query)
 512                         ib_sa_cancel_query(path->query_id, path->query);
 513                 spin_unlock_irqrestore(&priv->lock, flags);
 514                 wait_for_completion(&path->done);
 515                 ipoib_path_free(priv, path);
 516                 spin_lock_irqsave(&priv->lock, flags);
 517         }
 518
 519         spin_unlock_irqrestore(&priv->lock, flags);
 520 }
 521
 522 static void
 523 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
 524 {
 525         struct ipoib_path *path = path_ptr;
 526         struct ipoib_dev_priv *priv = path->priv;
 527         struct ifnet *dev = priv->dev;
 528         struct ipoib_ah *ah = NULL;
 529         struct ipoib_ah *old_ah = NULL;
 530         struct ifqueue mbqueue;
 531         struct mbuf *mb;
 532         unsigned long flags;
 533
 534         if (!status)
 535                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
 536                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
 537         else
 538                 ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
 539                           status, path->pathrec.dgid.raw, ":");
 540
 541         bzero(&mbqueue, sizeof(mbqueue));
 542
 543         if (!status) {
 544                 struct ib_ah_attr av;
 545
 546                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 547                         ah = ipoib_create_ah(priv, priv->pd, &av);
 548         }
 549
 550         spin_lock_irqsave(&priv->lock, flags);
 551
 552         if (ah) {
 553                 path->pathrec = *pathrec;
 554
 555                 old_ah   = path->ah;
 556                 path->ah = ah;
 557
 558                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 559                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 560
 561                 for (;;) {
 562                         _IF_DEQUEUE(&path->queue, mb);
 563                         if (mb == NULL)
 564                                 break;
 565                         _IF_ENQUEUE(&mbqueue, mb);
 566                 }
 567
 568 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 569                 if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
 570                         ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
 571 #endif
 572
 573                 path->valid = 1;
 574         }
 575
 576         path->query = NULL;
 577         complete(&path->done);
 578
 579         spin_unlock_irqrestore(&priv->lock, flags);
 580
 581         if (old_ah)
 582                 ipoib_put_ah(old_ah);
 583
 584         for (;;) {
 585                 _IF_DEQUEUE(&mbqueue, mb);
 586                 if (mb == NULL)
 587                         break;
 588                 mb->m_pkthdr.rcvif = dev;
 589                 if (dev->if_transmit(dev, mb))
 590                         ipoib_warn(priv, "dev_queue_xmit failed "
 591                                    "to requeue packet\n");
 592         }
 593 }
 594
 595 static struct ipoib_path *
 596 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
 597 {
 598         struct ipoib_path *path;
 599
 600         if (!priv->broadcast)
 601                 return NULL;
 602
 603         path = kzalloc(sizeof *path, GFP_ATOMIC);
 604         if (!path)
 605                 return NULL;
 606
 607         path->priv = priv;
 608
 609         bzero(&path->queue, sizeof(path->queue));
 610
 611 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 612         memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
 613 #endif
 614         memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
 615         path->pathrec.sgid          = priv->local_gid;
 616         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 617         path->pathrec.numb_path     = 1;
 618         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 619
 620         return path;
 621 }
 622
 623 static int
 624 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 625 {
 626         struct ifnet *dev = priv->dev;
 627
 628         ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
 629         struct ib_sa_path_rec p_rec;
 630
 631         p_rec = path->pathrec;
 632         p_rec.mtu_selector = IB_SA_GT;
 633
 634         switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
 635         case 512:
 636                 p_rec.mtu = IB_MTU_256;
 637                 break;
 638         case 1024:
 639                 p_rec.mtu = IB_MTU_512;
 640                 break;
 641         case 2048:
 642                 p_rec.mtu = IB_MTU_1024;
 643                 break;
 644         case 4096:
 645                 p_rec.mtu = IB_MTU_2048;
 646                 break;
 647         default:
 648                 /* Wildcard everything */
 649                 comp_mask = 0;
 650                 p_rec.mtu = 0;
 651                 p_rec.mtu_selector = 0;
 652         }
 653
 654         ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
 655                   p_rec.dgid.raw, ":",
 656                   comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
 657
 658         init_completion(&path->done);
 659
 660         path->query_id =
 661                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 662                                    &p_rec, comp_mask            |
 663                                    IB_SA_PATH_REC_DGID          |
 664                                    IB_SA_PATH_REC_SGID          |
 665                                    IB_SA_PATH_REC_NUMB_PATH     |
 666                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 667                                    IB_SA_PATH_REC_PKEY,
 668                                    1000, GFP_ATOMIC,
 669                                    path_rec_completion,
 670                                    path, &path->query);
 671         if (path->query_id < 0) {
 672                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 673                 path->query = NULL;
 674                 complete(&path->done);
 675                 return path->query_id;
 676         }
 677
 678         return 0;
 679 }
 680
 681 static void
 682 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
 683 {
 684         struct ipoib_path *path;
 685
 686         path = __path_find(priv, eh->hwaddr + 4);
 687         if (!path || !path->valid) {
 688                 int new_path = 0;
 689
 690                 if (!path) {
 691                         path = path_rec_create(priv, eh->hwaddr);
 692                         new_path = 1;
 693                 }
 694                 if (path) {
 695                         if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 696                                 _IF_ENQUEUE(&path->queue, mb);
 697                         else {
 698                                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 699                                 m_freem(mb);
 700                         }
 701
 702                         if (!path->query && path_rec_start(priv, path)) {
 703                                 spin_unlock_irqrestore(&priv->lock, flags);
 704                                 if (new_path)
 705                                         ipoib_path_free(priv, path);
 706                                 return;
 707                         } else
 708                                 __path_add(priv, path);
 709                 } else {
 710                         if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 711                         m_freem(mb);
 712                 }
 713
 714                 return;
 715         }
 716
 717         if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
 718                 ipoib_cm_send(priv, mb, ipoib_cm_get(path));
 719         } else if (path->ah) {
 720                 ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
 721         } else if ((path->query || !path_rec_start(priv, path)) &&
 722                     path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
 723                 _IF_ENQUEUE(&path->queue, mb);
 724         } else {
 725                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 726                 m_freem(mb);
 727         }
 728 }
 729
 730 static int
 731 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
 732 {
 733         struct ipoib_header *eh;
 734
 735         eh = mtod(mb, struct ipoib_header *);
 736         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
 737                 /* Add in the P_Key for multicast*/
 738                 eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 739                 eh->hwaddr[9] = priv->pkey & 0xff;
 740
 741                 ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
 742         } else
 743                 ipoib_unicast_send(mb, priv, eh);
 744
 745         return 0;
 746 }
 747
 748
 749 static void
 750 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
 751 {
 752         struct mbuf *mb;
 753
 754         if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
 755             IFF_DRV_RUNNING)
 756                 return;
 757
 758         spin_lock(&priv->lock);
 759         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
 760             (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
 761                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 762                 if (mb == NULL)
 763                         break;
 764                 IPOIB_MTAP(dev, mb);
 765                 ipoib_send_one(priv, mb);
 766         }
 767         spin_unlock(&priv->lock);
 768 }
 769
 770 static void
 771 ipoib_start(struct ifnet *dev)
 772 {
 773         _ipoib_start(dev, dev->if_softc);
 774 }
 775
 776 static void
 777 ipoib_vlan_start(struct ifnet *dev)
 778 {
 779         struct ipoib_dev_priv *priv;
 780         struct mbuf *mb;
 781
 782         priv = VLAN_COOKIE(dev);
 783         if (priv != NULL)
 784                 return _ipoib_start(dev, priv);
 785         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
 786                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 787                 if (mb == NULL)
 788                         break;
 789                 m_freem(mb);
 790                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 791         }
 792 }
 793
 794 int
 795 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
 796 {
 797
 798         /* Allocate RX/TX "rings" to hold queued mbs */
 799         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 800                                 GFP_KERNEL);
 801         if (!priv->rx_ring) {
 802                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 803                        ca->name, ipoib_recvq_size);
 804                 goto out;
 805         }
 806
 807         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
 808         if (!priv->tx_ring) {
 809                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 810                        ca->name, ipoib_sendq_size);
 811                 goto out_rx_ring_cleanup;
 812         }
 813         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 814
 815         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 816
 817         if (ipoib_ib_dev_init(priv, ca, port))
 818                 goto out_tx_ring_cleanup;
 819
 820         return 0;
 821
 822 out_tx_ring_cleanup:
 823         kfree(priv->tx_ring);
 824
 825 out_rx_ring_cleanup:
 826         kfree(priv->rx_ring);
 827
 828 out:
 829         return -ENOMEM;
 830 }
 831
 832 static void
 833 ipoib_detach(struct ipoib_dev_priv *priv)
 834 {
 835         struct ifnet *dev;
 836
 837         dev = priv->dev;
 838         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 839                 priv->gone = 1;
 840                 bpfdetach(dev);
 841                 if_detach(dev);
 842                 if_free(dev);
 843                 free_unr(ipoib_unrhdr, priv->unit);
 844         } else
 845                 VLAN_SETCOOKIE(priv->dev, NULL);
 846
 847         free(priv, M_TEMP);
 848 }
 849
 850 void
 851 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
 852 {
 853         struct ipoib_dev_priv *cpriv, *tcpriv;
 854
 855         /* Delete any child interfaces first */
 856         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 857                 ipoib_dev_cleanup(cpriv);
 858                 ipoib_detach(cpriv);
 859         }
 860
 861         ipoib_ib_dev_cleanup(priv);
 862
 863         kfree(priv->rx_ring);
 864         kfree(priv->tx_ring);
 865
 866         priv->rx_ring = NULL;
 867         priv->tx_ring = NULL;
 868 }
 869
 870 static struct ipoib_dev_priv *
 871 ipoib_priv_alloc(void)
 872 {
 873         struct ipoib_dev_priv *priv;
 874
 875         priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
 876         spin_lock_init(&priv->lock);
 877         spin_lock_init(&priv->drain_lock);
 878         mutex_init(&priv->vlan_mutex);
 879         INIT_LIST_HEAD(&priv->path_list);
 880         INIT_LIST_HEAD(&priv->child_intfs);
 881         INIT_LIST_HEAD(&priv->dead_ahs);
 882         INIT_LIST_HEAD(&priv->multicast_list);
 883         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 884         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 885         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
 886         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
 887         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
 888         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
 889         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 890         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 891         memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
 892
 893         return (priv);
 894 }
 895
 896 struct ipoib_dev_priv *
 897 ipoib_intf_alloc(const char *name)
 898 {
 899         struct ipoib_dev_priv *priv;
 900         struct sockaddr_dl *sdl;
 901         struct ifnet *dev;
 902
 903         priv = ipoib_priv_alloc();
 904         dev = priv->dev = if_alloc(IFT_INFINIBAND);
 905         if (!dev) {
 906                 free(priv, M_TEMP);
 907                 return NULL;
 908         }
 909         dev->if_softc = priv;
 910         priv->unit = alloc_unr(ipoib_unrhdr);
 911         if (priv->unit == -1) {
 912                 if_free(dev);
 913                 free(priv, M_TEMP);
 914                 return NULL;
 915         }
 916         if_initname(dev, name, priv->unit);
 917         dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
 918         dev->if_addrlen = INFINIBAND_ALEN;
 919         dev->if_hdrlen = IPOIB_HEADER_LEN;
 920         if_attach(dev);
 921         dev->if_init = ipoib_init;
 922         dev->if_ioctl = ipoib_ioctl;
 923         dev->if_start = ipoib_start;
 924         dev->if_output = ipoib_output;
 925         dev->if_input = ipoib_input;
 926         dev->if_resolvemulti = ipoib_resolvemulti;
 927         dev->if_baudrate = IF_Gbps(10);
 928         dev->if_broadcastaddr = priv->broadcastaddr;
 929         dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
 930         sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
 931         sdl->sdl_type = IFT_INFINIBAND;
 932         sdl->sdl_alen = dev->if_addrlen;
 933         priv->dev = dev;
 934         if_link_state_change(dev, LINK_STATE_DOWN);
 935         bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
 936
 937         return dev->if_softc;
 938 }
 939
 940 int
 941 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
 942 {
 943         struct ib_device_attr *device_attr = &hca->attrs;
 944
 945         priv->hca_caps = device_attr->device_cap_flags;
 946
 947         priv->dev->if_hwassist = 0;
 948         priv->dev->if_capabilities = 0;
 949
 950 #ifndef CONFIG_INFINIBAND_IPOIB_CM
 951         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
 952                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
 953                 priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
 954                 priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
 955         }
 956
 957 #if 0
 958         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
 959                 priv->dev->if_capabilities |= IFCAP_TSO4;
 960                 priv->dev->if_hwassist |= CSUM_TSO;
 961         }
 962 #endif
 963 #endif
 964         priv->dev->if_capabilities |=
 965             IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
 966         priv->dev->if_capenable = priv->dev->if_capabilities;
 967
 968         return 0;
 969 }
 970
 971
 972 static struct ifnet *
 973 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
 974 {
 975         struct ipoib_dev_priv *priv;
 976         struct ib_port_attr attr;
 977         int result = -ENOMEM;
 978
 979         priv = ipoib_intf_alloc(format);
 980         if (!priv)
 981                 goto alloc_mem_failed;
 982
 983         if (!ib_query_port(hca, port, &attr))
 984                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
 985         else {
 986                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
 987                        hca->name, port);
 988                 goto device_init_failed;
 989         }
 990
 991         /* MTU will be reset when mcast join happens */
 992         priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
 993         priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
 994
 995         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 996         if (result) {
 997                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 998                        hca->name, port, result);
 999                 goto device_init_failed;
1000         }
1001
1002         if (ipoib_set_dev_features(priv, hca))
1003                 goto device_init_failed;
1004
1005         /*
1006          * Set the full membership bit, so that we join the right
1007          * broadcast group, etc.
1008          */
1009         priv->pkey |= 0x8000;
1010
1011         priv->broadcastaddr[8] = priv->pkey >> 8;
1012         priv->broadcastaddr[9] = priv->pkey & 0xff;
1013
1014         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1015         if (result) {
1016                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1017                        hca->name, port, result);
1018                 goto device_init_failed;
1019         }
1020         memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
1021
1022         result = ipoib_dev_init(priv, hca, port);
1023         if (result < 0) {
1024                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1025                        hca->name, port, result);
1026                 goto device_init_failed;
1027         }
1028         if (ipoib_cm_admin_enabled(priv))
1029                 priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1030
1031         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1032                               priv->ca, ipoib_event);
1033         result = ib_register_event_handler(&priv->event_handler);
1034         if (result < 0) {
1035                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1036                        "port %d (ret = %d)\n",
1037                        hca->name, port, result);
1038                 goto event_failed;
1039         }
1040         if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1041
1042         return priv->dev;
1043
1044 event_failed:
1045         ipoib_dev_cleanup(priv);
1046
1047 device_init_failed:
1048         ipoib_detach(priv);
1049
1050 alloc_mem_failed:
1051         return ERR_PTR(result);
1052 }
1053
1054 static void
1055 ipoib_add_one(struct ib_device *device)
1056 {
1057         struct list_head *dev_list;
1058         struct ifnet *dev;
1059         struct ipoib_dev_priv *priv;
1060         int s, e, p;
1061
1062         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1063                 return;
1064
1065         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1066         if (!dev_list)
1067                 return;
1068
1069         INIT_LIST_HEAD(dev_list);
1070
1071         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1072                 s = 0;
1073                 e = 0;
1074         } else {
1075                 s = 1;
1076                 e = device->phys_port_cnt;
1077         }
1078
1079         for (p = s; p <= e; ++p) {
1080                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1081                         continue;
1082                 dev = ipoib_add_port("ib", device, p);
1083                 if (!IS_ERR(dev)) {
1084                         priv = dev->if_softc;
1085                         list_add_tail(&priv->list, dev_list);
1086                 }
1087         }
1088
1089         ib_set_client_data(device, &ipoib_client, dev_list);
1090 }
1091
1092 static void
1093 ipoib_remove_one(struct ib_device *device, void *client_data)
1094 {
1095         struct ipoib_dev_priv *priv, *tmp;
1096         struct list_head *dev_list = client_data;
1097
1098         if (!dev_list)
1099                 return;
1100
1101         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1102                 return;
1103
1104         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1105                 if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1106                         continue;
1107
1108                 ipoib_stop(priv);
1109
1110                 ib_unregister_event_handler(&priv->event_handler);
1111
1112                 /* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1113
1114                 flush_workqueue(ipoib_workqueue);
1115
1116                 ipoib_dev_cleanup(priv);
1117                 ipoib_detach(priv);
1118         }
1119
1120         kfree(dev_list);
1121 }
1122
1123 static int
1124 ipoib_match_dev_addr(const struct sockaddr *addr, struct net_device *dev)
1125 {
1126         struct ifaddr *ifa;
1127         int retval = 0;
1128
1129         CURVNET_SET(dev->if_vnet);
1130         IF_ADDR_RLOCK(dev);
1131         CK_STAILQ_FOREACH(ifa, &dev->if_addrhead, ifa_link) {
1132                 if (ifa->ifa_addr == NULL ||
1133                     ifa->ifa_addr->sa_family != addr->sa_family ||
1134                     ifa->ifa_addr->sa_len != addr->sa_len) {
1135                         continue;
1136                 }
1137                 if (memcmp(ifa->ifa_addr, addr, addr->sa_len) == 0) {
1138                         retval = 1;
1139                         break;
1140                 }
1141         }
1142         IF_ADDR_RUNLOCK(dev);
1143         CURVNET_RESTORE();
1144
1145         return (retval);
1146 }
1147
1148 /*
1149  * ipoib_match_gid_pkey_addr - returns the number of IPoIB netdevs on
1150  * top a given ipoib device matching a pkey_index and address, if one
1151  * exists.
1152  *
1153  * @found_net_dev: contains a matching net_device if the return value
1154  * >= 1, with a reference held.
1155  */
1156 static int
1157 ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
1158     const union ib_gid *gid, u16 pkey_index, const struct sockaddr *addr,
1159     struct net_device **found_net_dev)
1160 {
1161         struct ipoib_dev_priv *child_priv;
1162         int matches = 0;
1163
1164         if (priv->pkey_index == pkey_index &&
1165             (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
1166                 if (addr == NULL || ipoib_match_dev_addr(addr, priv->dev) != 0) {
1167                         if (*found_net_dev == NULL) {
1168                                 struct net_device *net_dev;
1169
1170                                 if (priv->parent != NULL)
1171                                         net_dev = priv->parent;
1172                                 else
1173                                         net_dev = priv->dev;
1174                                 *found_net_dev = net_dev;
1175                                 dev_hold(net_dev);
1176                         }
1177                         matches++;
1178                 }
1179         }
1180
1181         /* Check child interfaces */
1182         mutex_lock(&priv->vlan_mutex);
1183         list_for_each_entry(child_priv, &priv->child_intfs, list) {
1184                 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
1185                     pkey_index, addr, found_net_dev);
1186                 if (matches > 1)
1187                         break;
1188         }
1189         mutex_unlock(&priv->vlan_mutex);
1190
1191         return matches;
1192 }
1193
1194 /*
1195  * __ipoib_get_net_dev_by_params - returns the number of matching
1196  * net_devs found (between 0 and 2). Also return the matching
1197  * net_device in the @net_dev parameter, holding a reference to the
1198  * net_device, if the number of matches >= 1
1199  */
1200 static int
1201 __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
1202     u16 pkey_index, const union ib_gid *gid,
1203     const struct sockaddr *addr, struct net_device **net_dev)
1204 {
1205         struct ipoib_dev_priv *priv;
1206         int matches = 0;
1207
1208         *net_dev = NULL;
1209
1210         list_for_each_entry(priv, dev_list, list) {
1211                 if (priv->port != port)
1212                         continue;
1213
1214                 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
1215                     addr, net_dev);
1216
1217                 if (matches > 1)
1218                         break;
1219         }
1220
1221         return matches;
1222 }
1223
1224 static struct net_device *
1225 ipoib_get_net_dev_by_params(struct ib_device *dev, u8 port, u16 pkey,
1226     const union ib_gid *gid, const struct sockaddr *addr, void *client_data)
1227 {
1228         struct net_device *net_dev;
1229         struct list_head *dev_list = client_data;
1230         u16 pkey_index;
1231         int matches;
1232         int ret;
1233
1234         if (!rdma_protocol_ib(dev, port))
1235                 return NULL;
1236
1237         ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
1238         if (ret)
1239                 return NULL;
1240
1241         if (!dev_list)
1242                 return NULL;
1243
1244         /* See if we can find a unique device matching the L2 parameters */
1245         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1246                                                 gid, NULL, &net_dev);
1247
1248         switch (matches) {
1249         case 0:
1250                 return NULL;
1251         case 1:
1252                 return net_dev;
1253         }
1254
1255         dev_put(net_dev);
1256
1257         /* Couldn't find a unique device with L2 parameters only. Use L3
1258          * address to uniquely match the net device */
1259         matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1260                                                 gid, addr, &net_dev);
1261         switch (matches) {
1262         case 0:
1263                 return NULL;
1264         default:
1265                 dev_warn_ratelimited(&dev->dev,
1266                                      "duplicate IP address detected\n");
1267                 /* Fall through */
1268         case 1:
1269                 return net_dev;
1270         }
1271 }
1272
1273 static void
1274 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1275 {
1276         struct ipoib_dev_priv *parent;
1277         struct ipoib_dev_priv *priv;
1278         struct ifnet *dev;
1279         uint16_t pkey;
1280         int error;
1281
1282         if (ifp->if_type != IFT_INFINIBAND)
1283                 return;
1284         dev = VLAN_DEVAT(ifp, vtag);
1285         if (dev == NULL)
1286                 return;
1287         priv = NULL;
1288         error = 0;
1289         parent = ifp->if_softc;
1290         /* We only support 15 bits of pkey. */
1291         if (vtag & 0x8000)
1292                 return;
1293         pkey = vtag | 0x8000;   /* Set full membership bit. */
1294         if (pkey == parent->pkey)
1295                 return;
1296         /* Check for dups */
1297         mutex_lock(&parent->vlan_mutex);
1298         list_for_each_entry(priv, &parent->child_intfs, list) {
1299                 if (priv->pkey == pkey) {
1300                         priv = NULL;
1301                         error = EBUSY;
1302                         goto out;
1303                 }
1304         }
1305         priv = ipoib_priv_alloc();
1306         priv->dev = dev;
1307         priv->max_ib_mtu = parent->max_ib_mtu;
1308         priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1309         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1310         error = ipoib_set_dev_features(priv, parent->ca);
1311         if (error)
1312                 goto out;
1313         priv->pkey = pkey;
1314         priv->broadcastaddr[8] = pkey >> 8;
1315         priv->broadcastaddr[9] = pkey & 0xff;
1316         dev->if_broadcastaddr = priv->broadcastaddr;
1317         error = ipoib_dev_init(priv, parent->ca, parent->port);
1318         if (error)
1319                 goto out;
1320         priv->parent = parent->dev;
1321         list_add_tail(&priv->list, &parent->child_intfs);
1322         VLAN_SETCOOKIE(dev, priv);
1323         dev->if_start = ipoib_vlan_start;
1324         dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1325         dev->if_hdrlen = IPOIB_HEADER_LEN;
1326         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1327                 ipoib_open(priv);
1328         mutex_unlock(&parent->vlan_mutex);
1329         return;
1330 out:
1331         mutex_unlock(&parent->vlan_mutex);
1332         if (priv)
1333                 free(priv, M_TEMP);
1334         if (error)
1335                 ipoib_warn(parent,
1336                     "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1337                     parent->ca->name, parent->port, vtag);
1338         return;
1339 }
1340
1341 static void
1342 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1343 {
1344         struct ipoib_dev_priv *parent;
1345         struct ipoib_dev_priv *priv;
1346         struct ifnet *dev;
1347         uint16_t pkey;
1348
1349         if (ifp->if_type != IFT_INFINIBAND)
1350                 return;
1351
1352         dev = VLAN_DEVAT(ifp, vtag);
1353         if (dev)
1354                 VLAN_SETCOOKIE(dev, NULL);
1355         pkey = vtag | 0x8000;
1356         parent = ifp->if_softc;
1357         mutex_lock(&parent->vlan_mutex);
1358         list_for_each_entry(priv, &parent->child_intfs, list) {
1359                 if (priv->pkey == pkey) {
1360                         ipoib_dev_cleanup(priv);
1361                         list_del(&priv->list);
1362                         break;
1363                 }
1364         }
1365         mutex_unlock(&parent->vlan_mutex);
1366 }
1367
1368 eventhandler_tag ipoib_vlan_attach;
1369 eventhandler_tag ipoib_vlan_detach;
1370
1371 static int __init
1372 ipoib_init_module(void)
1373 {
1374         int ret;
1375
1376         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1377         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1378         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1379
1380         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1381         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1382         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1383                                                      IPOIB_MIN_QUEUE_SIZE));
1384 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1385         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1386 #endif
1387
1388         ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1389                 ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1390         ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1391                 ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1392
1393         /*
1394          * We create our own workqueue mainly because we want to be
1395          * able to flush it when devices are being removed.  We can't
1396          * use schedule_work()/flush_scheduled_work() because both
1397          * unregister_netdev() and linkwatch_event take the rtnl lock,
1398          * so flush_scheduled_work() can deadlock during device
1399          * removal.
1400          */
1401         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1402         if (!ipoib_workqueue) {
1403                 ret = -ENOMEM;
1404                 goto err_fs;
1405         }
1406
1407         ib_sa_register_client(&ipoib_sa_client);
1408
1409         ret = ib_register_client(&ipoib_client);
1410         if (ret)
1411                 goto err_sa;
1412
1413         return 0;
1414
1415 err_sa:
1416         ib_sa_unregister_client(&ipoib_sa_client);
1417         destroy_workqueue(ipoib_workqueue);
1418
1419 err_fs:
1420         return ret;
1421 }
1422
1423 static void __exit
1424 ipoib_cleanup_module(void)
1425 {
1426
1427         EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1428         EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1429         ib_unregister_client(&ipoib_client);
1430         ib_sa_unregister_client(&ipoib_sa_client);
1431         destroy_workqueue(ipoib_workqueue);
1432 }
1433
1434 /*
1435  * Infiniband output routine.
1436  */
1437 static int
1438 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1439         const struct sockaddr *dst, struct route *ro)
1440 {
1441         u_char edst[INFINIBAND_ALEN];
1442 #if defined(INET) || defined(INET6)
1443         struct llentry *lle = NULL;
1444 #endif
1445         struct ipoib_header *eh;
1446         int error = 0, is_gw = 0;
1447         short type;
1448
1449         if (ro != NULL)
1450                 is_gw = (ro->ro_flags & RT_HAS_GW) != 0;
1451 #ifdef MAC
1452         error = mac_ifnet_check_transmit(ifp, m);
1453         if (error)
1454                 goto bad;
1455 #endif
1456
1457         M_PROFILE(m);
1458         if (ifp->if_flags & IFF_MONITOR) {
1459                 error = ENETDOWN;
1460                 goto bad;
1461         }
1462         if (!((ifp->if_flags & IFF_UP) &&
1463             (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1464                 error = ENETDOWN;
1465                 goto bad;
1466         }
1467
1468         switch (dst->sa_family) {
1469 #ifdef INET
1470         case AF_INET:
1471                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1472                         memcpy(edst, lle->ll_addr, sizeof(edst));
1473                 else if (m->m_flags & M_MCAST)
1474                         ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1475                 else
1476                         error = arpresolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1477                 if (error)
1478                         return (error == EWOULDBLOCK ? 0 : error);
1479                 type = htons(ETHERTYPE_IP);
1480                 break;
1481         case AF_ARP:
1482         {
1483                 struct arphdr *ah;
1484                 ah = mtod(m, struct arphdr *);
1485                 ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1486
1487                 switch(ntohs(ah->ar_op)) {
1488                 case ARPOP_REVREQUEST:
1489                 case ARPOP_REVREPLY:
1490                         type = htons(ETHERTYPE_REVARP);
1491                         break;
1492                 case ARPOP_REQUEST:
1493                 case ARPOP_REPLY:
1494                 default:
1495                         type = htons(ETHERTYPE_ARP);
1496                         break;
1497                 }
1498
1499                 if (m->m_flags & M_BCAST)
1500                         bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1501                 else
1502                         bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1503
1504         }
1505         break;
1506 #endif
1507 #ifdef INET6
1508         case AF_INET6:
1509                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1510                         memcpy(edst, lle->ll_addr, sizeof(edst));
1511                 else if (m->m_flags & M_MCAST)
1512                         ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1513                 else
1514                         error = nd6_resolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1515                 if (error)
1516                         return error;
1517                 type = htons(ETHERTYPE_IPV6);
1518                 break;
1519 #endif
1520
1521         default:
1522                 if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1523                 error = EAFNOSUPPORT;
1524                 goto bad;
1525         }
1526
1527         /*
1528          * Add local net header.  If no space in first mbuf,
1529          * allocate another.
1530          */
1531         M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1532         if (m == NULL) {
1533                 error = ENOBUFS;
1534                 goto bad;
1535         }
1536         eh = mtod(m, struct ipoib_header *);
1537         (void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1538         (void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1539
1540         /*
1541          * Queue message on interface, update output statistics if
1542          * successful, and start output if interface not yet active.
1543          */
1544         return ((ifp->if_transmit)(ifp, m));
1545 bad:
1546         if (m != NULL)
1547                 m_freem(m);
1548         return (error);
1549 }
1550
1551 /*
1552  * Upper layer processing for a received Infiniband packet.
1553  */
1554 void
1555 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1556 {
1557         int isr;
1558
1559 #ifdef MAC
1560         /*
1561          * Tag the mbuf with an appropriate MAC label before any other
1562          * consumers can get to it.
1563          */
1564         mac_ifnet_create_mbuf(ifp, m);
1565 #endif
1566         /* Allow monitor mode to claim this frame, after stats are updated. */
1567         if (ifp->if_flags & IFF_MONITOR) {
1568                 if_printf(ifp, "discard frame at IFF_MONITOR\n");
1569                 m_freem(m);
1570                 return;
1571         }
1572         /*
1573          * Dispatch frame to upper layer.
1574          */
1575         switch (proto) {
1576 #ifdef INET
1577         case ETHERTYPE_IP:
1578                 isr = NETISR_IP;
1579                 break;
1580
1581         case ETHERTYPE_ARP:
1582                 if (ifp->if_flags & IFF_NOARP) {
1583                         /* Discard packet if ARP is disabled on interface */
1584                         m_freem(m);
1585                         return;
1586                 }
1587                 isr = NETISR_ARP;
1588                 break;
1589 #endif
1590 #ifdef INET6
1591         case ETHERTYPE_IPV6:
1592                 isr = NETISR_IPV6;
1593                 break;
1594 #endif
1595         default:
1596                 goto discard;
1597         }
1598         netisr_dispatch(isr, m);
1599         return;
1600
1601 discard:
1602         m_freem(m);
1603 }
1604
1605 /*
1606  * Process a received Infiniband packet.
1607  */
1608 static void
1609 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1610 {
1611         struct ipoib_header *eh;
1612
1613         if ((ifp->if_flags & IFF_UP) == 0) {
1614                 m_freem(m);
1615                 return;
1616         }
1617         CURVNET_SET_QUIET(ifp->if_vnet);
1618
1619         /* Let BPF have it before we strip the header. */
1620         IPOIB_MTAP(ifp, m);
1621         eh = mtod(m, struct ipoib_header *);
1622         /*
1623          * Reset layer specific mbuf flags to avoid confusing upper layers.
1624          * Strip off Infiniband header.
1625          */
1626         m->m_flags &= ~M_VLANTAG;
1627         m_clrprotoflags(m);
1628         m_adj(m, IPOIB_HEADER_LEN);
1629
1630         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1631                 if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1632                     ifp->if_addrlen) == 0)
1633                         m->m_flags |= M_BCAST;
1634                 else
1635                         m->m_flags |= M_MCAST;
1636                 if_inc_counter(ifp, IFCOUNTER_IMCASTS, 1);
1637         }
1638
1639         ipoib_demux(ifp, m, ntohs(eh->proto));
1640         CURVNET_RESTORE();
1641 }
1642
1643 static int
1644 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1645         struct sockaddr *sa)
1646 {
1647         struct sockaddr_dl *sdl;
1648 #ifdef INET
1649         struct sockaddr_in *sin;
1650 #endif
1651 #ifdef INET6
1652         struct sockaddr_in6 *sin6;
1653 #endif
1654         u_char *e_addr;
1655
1656         switch(sa->sa_family) {
1657         case AF_LINK:
1658                 /*
1659                  * No mapping needed. Just check that it's a valid MC address.
1660                  */
1661                 sdl = (struct sockaddr_dl *)sa;
1662                 e_addr = LLADDR(sdl);
1663                 if (!IPOIB_IS_MULTICAST(e_addr))
1664                         return EADDRNOTAVAIL;
1665                 *llsa = NULL;
1666                 return 0;
1667
1668 #ifdef INET
1669         case AF_INET:
1670                 sin = (struct sockaddr_in *)sa;
1671                 if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1672                         return EADDRNOTAVAIL;
1673                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1674                 sdl->sdl_alen = INFINIBAND_ALEN;
1675                 e_addr = LLADDR(sdl);
1676                 ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1677                     e_addr);
1678                 *llsa = (struct sockaddr *)sdl;
1679                 return 0;
1680 #endif
1681 #ifdef INET6
1682         case AF_INET6:
1683                 sin6 = (struct sockaddr_in6 *)sa;
1684                 /*
1685                  * An IP6 address of 0 means listen to all
1686                  * of the multicast address used for IP6.
1687                  * This has no meaning in ipoib.
1688                  */
1689                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1690                         return EADDRNOTAVAIL;
1691                 if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1692                         return EADDRNOTAVAIL;
1693                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1694                 sdl->sdl_alen = INFINIBAND_ALEN;
1695                 e_addr = LLADDR(sdl);
1696                 ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1697                 *llsa = (struct sockaddr *)sdl;
1698                 return 0;
1699 #endif
1700
1701         default:
1702                 return EAFNOSUPPORT;
1703         }
1704 }
1705
1706 module_init(ipoib_init_module);
1707 module_exit(ipoib_cleanup_module);
1708
1709 static int
1710 ipoib_evhand(module_t mod, int event, void *arg)
1711 {
1712                         return (0);
1713 }
1714
1715 static moduledata_t ipoib_mod = {
1716                         .name = "ipoib",
1717                                         .evhand = ipoib_evhand,
1718 };
1719
1720 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1721 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1722 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);