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                 {
 324                         struct sockaddr *sa;
 325
 326                         sa = (struct sockaddr *) & ifr->ifr_data;
 327                         bcopy(IF_LLADDR(ifp),
 328                               (caddr_t) sa->sa_data, INFINIBAND_ALEN);
 329                 }
 330                 break;
 331
 332         case SIOCSIFMTU:
 333                 /*
 334                  * Set the interface MTU.
 335                  */
 336                 error = -ipoib_change_mtu(priv, ifr->ifr_mtu);
 337                 break;
 338         default:
 339                 error = EINVAL;
 340                 break;
 341         }
 342         return (error);
 343 }
 344
 345
 346 static struct ipoib_path *
 347 __path_find(struct ipoib_dev_priv *priv, void *gid)
 348 {
 349         struct rb_node *n = priv->path_tree.rb_node;
 350         struct ipoib_path *path;
 351         int ret;
 352
 353         while (n) {
 354                 path = rb_entry(n, struct ipoib_path, rb_node);
 355
 356                 ret = memcmp(gid, path->pathrec.dgid.raw,
 357                              sizeof (union ib_gid));
 358
 359                 if (ret < 0)
 360                         n = n->rb_left;
 361                 else if (ret > 0)
 362                         n = n->rb_right;
 363                 else
 364                         return path;
 365         }
 366
 367         return NULL;
 368 }
 369
 370 static int
 371 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 372 {
 373         struct rb_node **n = &priv->path_tree.rb_node;
 374         struct rb_node *pn = NULL;
 375         struct ipoib_path *tpath;
 376         int ret;
 377
 378         while (*n) {
 379                 pn = *n;
 380                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
 381
 382                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
 383                              sizeof (union ib_gid));
 384                 if (ret < 0)
 385                         n = &pn->rb_left;
 386                 else if (ret > 0)
 387                         n = &pn->rb_right;
 388                 else
 389                         return -EEXIST;
 390         }
 391
 392         rb_link_node(&path->rb_node, pn, n);
 393         rb_insert_color(&path->rb_node, &priv->path_tree);
 394
 395         list_add_tail(&path->list, &priv->path_list);
 396
 397         return 0;
 398 }
 399
 400 void
 401 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 402 {
 403
 404         _IF_DRAIN(&path->queue);
 405
 406         if (path->ah)
 407                 ipoib_put_ah(path->ah);
 408         if (ipoib_cm_get(path))
 409                 ipoib_cm_destroy_tx(ipoib_cm_get(path));
 410
 411         kfree(path);
 412 }
 413
 414 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 415
 416 struct ipoib_path_iter *
 417 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
 418 {
 419         struct ipoib_path_iter *iter;
 420
 421         iter = kmalloc(sizeof *iter, GFP_KERNEL);
 422         if (!iter)
 423                 return NULL;
 424
 425         iter->priv = priv;
 426         memset(iter->path.pathrec.dgid.raw, 0, 16);
 427
 428         if (ipoib_path_iter_next(iter)) {
 429                 kfree(iter);
 430                 return NULL;
 431         }
 432
 433         return iter;
 434 }
 435
 436 int
 437 ipoib_path_iter_next(struct ipoib_path_iter *iter)
 438 {
 439         struct ipoib_dev_priv *priv = iter->priv;
 440         struct rb_node *n;
 441         struct ipoib_path *path;
 442         int ret = 1;
 443
 444         spin_lock_irq(&priv->lock);
 445
 446         n = rb_first(&priv->path_tree);
 447
 448         while (n) {
 449                 path = rb_entry(n, struct ipoib_path, rb_node);
 450
 451                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
 452                            sizeof (union ib_gid)) < 0) {
 453                         iter->path = *path;
 454                         ret = 0;
 455                         break;
 456                 }
 457
 458                 n = rb_next(n);
 459         }
 460
 461         spin_unlock_irq(&priv->lock);
 462
 463         return ret;
 464 }
 465
 466 void
 467 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
 468 {
 469         *path = iter->path;
 470 }
 471
 472 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
 473
 474 void
 475 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
 476 {
 477         struct ipoib_path *path, *tp;
 478
 479         spin_lock_irq(&priv->lock);
 480
 481         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
 482                 ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
 483                         be16_to_cpu(path->pathrec.dlid),
 484                         path->pathrec.dgid.raw, ":");
 485                 path->valid =  0;
 486         }
 487
 488         spin_unlock_irq(&priv->lock);
 489 }
 490
 491 void
 492 ipoib_flush_paths(struct ipoib_dev_priv *priv)
 493 {
 494         struct ipoib_path *path, *tp;
 495         LIST_HEAD(remove_list);
 496         unsigned long flags;
 497
 498         spin_lock_irqsave(&priv->lock, flags);
 499
 500         list_splice_init(&priv->path_list, &remove_list);
 501
 502         list_for_each_entry(path, &remove_list, list)
 503                 rb_erase(&path->rb_node, &priv->path_tree);
 504
 505         list_for_each_entry_safe(path, tp, &remove_list, list) {
 506                 if (path->query)
 507                         ib_sa_cancel_query(path->query_id, path->query);
 508                 spin_unlock_irqrestore(&priv->lock, flags);
 509                 wait_for_completion(&path->done);
 510                 ipoib_path_free(priv, path);
 511                 spin_lock_irqsave(&priv->lock, flags);
 512         }
 513
 514         spin_unlock_irqrestore(&priv->lock, flags);
 515 }
 516
 517 static void
 518 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
 519 {
 520         struct ipoib_path *path = path_ptr;
 521         struct ipoib_dev_priv *priv = path->priv;
 522         struct ifnet *dev = priv->dev;
 523         struct ipoib_ah *ah = NULL;
 524         struct ipoib_ah *old_ah = NULL;
 525         struct ifqueue mbqueue;
 526         struct mbuf *mb;
 527         unsigned long flags;
 528
 529         if (!status)
 530                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
 531                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
 532         else
 533                 ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
 534                           status, path->pathrec.dgid.raw, ":");
 535
 536         bzero(&mbqueue, sizeof(mbqueue));
 537
 538         if (!status) {
 539                 struct ib_ah_attr av;
 540
 541                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
 542                         ah = ipoib_create_ah(priv, priv->pd, &av);
 543         }
 544
 545         spin_lock_irqsave(&priv->lock, flags);
 546
 547         if (ah) {
 548                 path->pathrec = *pathrec;
 549
 550                 old_ah   = path->ah;
 551                 path->ah = ah;
 552
 553                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
 554                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
 555
 556                 for (;;) {
 557                         _IF_DEQUEUE(&path->queue, mb);
 558                         if (mb == NULL)
 559                                 break;
 560                         _IF_ENQUEUE(&mbqueue, mb);
 561                 }
 562
 563 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 564                 if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
 565                         ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
 566 #endif
 567
 568                 path->valid = 1;
 569         }
 570
 571         path->query = NULL;
 572         complete(&path->done);
 573
 574         spin_unlock_irqrestore(&priv->lock, flags);
 575
 576         if (old_ah)
 577                 ipoib_put_ah(old_ah);
 578
 579         for (;;) {
 580                 _IF_DEQUEUE(&mbqueue, mb);
 581                 if (mb == NULL)
 582                         break;
 583                 mb->m_pkthdr.rcvif = dev;
 584                 if (dev->if_transmit(dev, mb))
 585                         ipoib_warn(priv, "dev_queue_xmit failed "
 586                                    "to requeue packet\n");
 587         }
 588 }
 589
 590 static struct ipoib_path *
 591 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
 592 {
 593         struct ipoib_path *path;
 594
 595         if (!priv->broadcast)
 596                 return NULL;
 597
 598         path = kzalloc(sizeof *path, GFP_ATOMIC);
 599         if (!path)
 600                 return NULL;
 601
 602         path->priv = priv;
 603
 604         bzero(&path->queue, sizeof(path->queue));
 605
 606 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 607         memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
 608 #endif
 609         memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
 610         path->pathrec.sgid          = priv->local_gid;
 611         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
 612         path->pathrec.numb_path     = 1;
 613         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
 614
 615         return path;
 616 }
 617
 618 static int
 619 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
 620 {
 621         struct ifnet *dev = priv->dev;
 622
 623         ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
 624         struct ib_sa_path_rec p_rec;
 625
 626         p_rec = path->pathrec;
 627         p_rec.mtu_selector = IB_SA_GT;
 628
 629         switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
 630         case 512:
 631                 p_rec.mtu = IB_MTU_256;
 632                 break;
 633         case 1024:
 634                 p_rec.mtu = IB_MTU_512;
 635                 break;
 636         case 2048:
 637                 p_rec.mtu = IB_MTU_1024;
 638                 break;
 639         case 4096:
 640                 p_rec.mtu = IB_MTU_2048;
 641                 break;
 642         default:
 643                 /* Wildcard everything */
 644                 comp_mask = 0;
 645                 p_rec.mtu = 0;
 646                 p_rec.mtu_selector = 0;
 647         }
 648
 649         ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
 650                   p_rec.dgid.raw, ":",
 651                   comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
 652
 653         init_completion(&path->done);
 654
 655         path->query_id =
 656                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
 657                                    &p_rec, comp_mask            |
 658                                    IB_SA_PATH_REC_DGID          |
 659                                    IB_SA_PATH_REC_SGID          |
 660                                    IB_SA_PATH_REC_NUMB_PATH     |
 661                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
 662                                    IB_SA_PATH_REC_PKEY,
 663                                    1000, GFP_ATOMIC,
 664                                    path_rec_completion,
 665                                    path, &path->query);
 666         if (path->query_id < 0) {
 667                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
 668                 path->query = NULL;
 669                 complete(&path->done);
 670                 return path->query_id;
 671         }
 672
 673         return 0;
 674 }
 675
 676 static void
 677 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
 678 {
 679         struct ipoib_path *path;
 680
 681         path = __path_find(priv, eh->hwaddr + 4);
 682         if (!path || !path->valid) {
 683                 int new_path = 0;
 684
 685                 if (!path) {
 686                         path = path_rec_create(priv, eh->hwaddr);
 687                         new_path = 1;
 688                 }
 689                 if (path) {
 690                         if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
 691                                 _IF_ENQUEUE(&path->queue, mb);
 692                         else {
 693                                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 694                                 m_freem(mb);
 695                         }
 696
 697                         if (!path->query && path_rec_start(priv, path)) {
 698                                 spin_unlock_irqrestore(&priv->lock, flags);
 699                                 if (new_path)
 700                                         ipoib_path_free(priv, path);
 701                                 return;
 702                         } else
 703                                 __path_add(priv, path);
 704                 } else {
 705                         if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 706                         m_freem(mb);
 707                 }
 708
 709                 return;
 710         }
 711
 712         if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
 713                 ipoib_cm_send(priv, mb, ipoib_cm_get(path));
 714         } else if (path->ah) {
 715                 ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
 716         } else if ((path->query || !path_rec_start(priv, path)) &&
 717                     path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
 718                 _IF_ENQUEUE(&path->queue, mb);
 719         } else {
 720                 if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
 721                 m_freem(mb);
 722         }
 723 }
 724
 725 static int
 726 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
 727 {
 728         struct ipoib_header *eh;
 729
 730         eh = mtod(mb, struct ipoib_header *);
 731         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
 732                 /* Add in the P_Key for multicast*/
 733                 eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
 734                 eh->hwaddr[9] = priv->pkey & 0xff;
 735
 736                 ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
 737         } else
 738                 ipoib_unicast_send(mb, priv, eh);
 739
 740         return 0;
 741 }
 742
 743
 744 static void
 745 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
 746 {
 747         struct mbuf *mb;
 748
 749         if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
 750             IFF_DRV_RUNNING)
 751                 return;
 752
 753         spin_lock(&priv->lock);
 754         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
 755             (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
 756                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 757                 if (mb == NULL)
 758                         break;
 759                 IPOIB_MTAP(dev, mb);
 760                 ipoib_send_one(priv, mb);
 761         }
 762         spin_unlock(&priv->lock);
 763 }
 764
 765 static void
 766 ipoib_start(struct ifnet *dev)
 767 {
 768         _ipoib_start(dev, dev->if_softc);
 769 }
 770
 771 static void
 772 ipoib_vlan_start(struct ifnet *dev)
 773 {
 774         struct ipoib_dev_priv *priv;
 775         struct mbuf *mb;
 776
 777         priv = VLAN_COOKIE(dev);
 778         if (priv != NULL)
 779                 return _ipoib_start(dev, priv);
 780         while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
 781                 IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
 782                 if (mb == NULL)
 783                         break;
 784                 m_freem(mb);
 785                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 786         }
 787 }
 788
 789 int
 790 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
 791 {
 792
 793         /* Allocate RX/TX "rings" to hold queued mbs */
 794         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
 795                                 GFP_KERNEL);
 796         if (!priv->rx_ring) {
 797                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
 798                        ca->name, ipoib_recvq_size);
 799                 goto out;
 800         }
 801
 802         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
 803         if (!priv->tx_ring) {
 804                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
 805                        ca->name, ipoib_sendq_size);
 806                 goto out_rx_ring_cleanup;
 807         }
 808         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
 809
 810         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
 811
 812         if (ipoib_ib_dev_init(priv, ca, port))
 813                 goto out_tx_ring_cleanup;
 814
 815         return 0;
 816
 817 out_tx_ring_cleanup:
 818         kfree(priv->tx_ring);
 819
 820 out_rx_ring_cleanup:
 821         kfree(priv->rx_ring);
 822
 823 out:
 824         return -ENOMEM;
 825 }
 826
 827 static void
 828 ipoib_detach(struct ipoib_dev_priv *priv)
 829 {
 830         struct ifnet *dev;
 831
 832         dev = priv->dev;
 833         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 834                 priv->gone = 1;
 835                 bpfdetach(dev);
 836                 if_detach(dev);
 837                 if_free(dev);
 838                 free_unr(ipoib_unrhdr, priv->unit);
 839         } else
 840                 VLAN_SETCOOKIE(priv->dev, NULL);
 841
 842         free(priv, M_TEMP);
 843 }
 844
 845 void
 846 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
 847 {
 848         struct ipoib_dev_priv *cpriv, *tcpriv;
 849
 850         /* Delete any child interfaces first */
 851         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
 852                 ipoib_dev_cleanup(cpriv);
 853                 ipoib_detach(cpriv);
 854         }
 855
 856         ipoib_ib_dev_cleanup(priv);
 857
 858         kfree(priv->rx_ring);
 859         kfree(priv->tx_ring);
 860
 861         priv->rx_ring = NULL;
 862         priv->tx_ring = NULL;
 863 }
 864
 865 static struct ipoib_dev_priv *
 866 ipoib_priv_alloc(void)
 867 {
 868         struct ipoib_dev_priv *priv;
 869
 870         priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
 871         spin_lock_init(&priv->lock);
 872         spin_lock_init(&priv->drain_lock);
 873         mutex_init(&priv->vlan_mutex);
 874         INIT_LIST_HEAD(&priv->path_list);
 875         INIT_LIST_HEAD(&priv->child_intfs);
 876         INIT_LIST_HEAD(&priv->dead_ahs);
 877         INIT_LIST_HEAD(&priv->multicast_list);
 878         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
 879         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
 880         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
 881         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
 882         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
 883         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
 884         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
 885         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
 886         memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
 887
 888         return (priv);
 889 }
 890
 891 struct ipoib_dev_priv *
 892 ipoib_intf_alloc(const char *name)
 893 {
 894         struct ipoib_dev_priv *priv;
 895         struct sockaddr_dl *sdl;
 896         struct ifnet *dev;
 897
 898         priv = ipoib_priv_alloc();
 899         dev = priv->dev = if_alloc(IFT_INFINIBAND);
 900         if (!dev) {
 901                 free(priv, M_TEMP);
 902                 return NULL;
 903         }
 904         dev->if_softc = priv;
 905         priv->unit = alloc_unr(ipoib_unrhdr);
 906         if (priv->unit == -1) {
 907                 if_free(dev);
 908                 free(priv, M_TEMP);
 909                 return NULL;
 910         }
 911         if_initname(dev, name, priv->unit);
 912         dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
 913         dev->if_addrlen = INFINIBAND_ALEN;
 914         dev->if_hdrlen = IPOIB_HEADER_LEN;
 915         if_attach(dev);
 916         dev->if_init = ipoib_init;
 917         dev->if_ioctl = ipoib_ioctl;
 918         dev->if_start = ipoib_start;
 919         dev->if_output = ipoib_output;
 920         dev->if_input = ipoib_input;
 921         dev->if_resolvemulti = ipoib_resolvemulti;
 922         dev->if_baudrate = IF_Gbps(10);
 923         dev->if_broadcastaddr = priv->broadcastaddr;
 924         dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
 925         sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
 926         sdl->sdl_type = IFT_INFINIBAND;
 927         sdl->sdl_alen = dev->if_addrlen;
 928         priv->dev = dev;
 929         if_link_state_change(dev, LINK_STATE_DOWN);
 930         bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
 931
 932         return dev->if_softc;
 933 }
 934
 935 int
 936 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
 937 {
 938         struct ib_device_attr *device_attr = &hca->attrs;
 939
 940         priv->hca_caps = device_attr->device_cap_flags;
 941
 942         priv->dev->if_hwassist = 0;
 943         priv->dev->if_capabilities = 0;
 944
 945 #ifndef CONFIG_INFINIBAND_IPOIB_CM
 946         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
 947                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
 948                 priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
 949                 priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
 950         }
 951
 952 #if 0
 953         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
 954                 priv->dev->if_capabilities |= IFCAP_TSO4;
 955                 priv->dev->if_hwassist |= CSUM_TSO;
 956         }
 957 #endif
 958 #endif
 959         priv->dev->if_capabilities |=
 960             IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
 961         priv->dev->if_capenable = priv->dev->if_capabilities;
 962
 963         return 0;
 964 }
 965
 966
 967 static struct ifnet *
 968 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
 969 {
 970         struct ipoib_dev_priv *priv;
 971         struct ib_port_attr attr;
 972         int result = -ENOMEM;
 973
 974         priv = ipoib_intf_alloc(format);
 975         if (!priv)
 976                 goto alloc_mem_failed;
 977
 978         if (!ib_query_port(hca, port, &attr))
 979                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
 980         else {
 981                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
 982                        hca->name, port);
 983                 goto device_init_failed;
 984         }
 985
 986         /* MTU will be reset when mcast join happens */
 987         priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
 988         priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
 989
 990         result = ib_query_pkey(hca, port, 0, &priv->pkey);
 991         if (result) {
 992                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
 993                        hca->name, port, result);
 994                 goto device_init_failed;
 995         }
 996
 997         if (ipoib_set_dev_features(priv, hca))
 998                 goto device_init_failed;
 999
1000         /*
1001          * Set the full membership bit, so that we join the right
1002          * broadcast group, etc.
1003          */
1004         priv->pkey |= 0x8000;
1005
1006         priv->broadcastaddr[8] = priv->pkey >> 8;
1007         priv->broadcastaddr[9] = priv->pkey & 0xff;
1008
1009         result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1010         if (result) {
1011                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1012                        hca->name, port, result);
1013                 goto device_init_failed;
1014         }
1015         memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
1016
1017         result = ipoib_dev_init(priv, hca, port);
1018         if (result < 0) {
1019                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1020                        hca->name, port, result);
1021                 goto device_init_failed;
1022         }
1023         if (ipoib_cm_admin_enabled(priv))
1024                 priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1025
1026         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1027                               priv->ca, ipoib_event);
1028         result = ib_register_event_handler(&priv->event_handler);
1029         if (result < 0) {
1030                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1031                        "port %d (ret = %d)\n",
1032                        hca->name, port, result);
1033                 goto event_failed;
1034         }
1035         if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1036
1037         return priv->dev;
1038
1039 event_failed:
1040         ipoib_dev_cleanup(priv);
1041
1042 device_init_failed:
1043         ipoib_detach(priv);
1044
1045 alloc_mem_failed:
1046         return ERR_PTR(result);
1047 }
1048
1049 static void
1050 ipoib_add_one(struct ib_device *device)
1051 {
1052         struct list_head *dev_list;
1053         struct ifnet *dev;
1054         struct ipoib_dev_priv *priv;
1055         int s, e, p;
1056
1057         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1058                 return;
1059
1060         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1061         if (!dev_list)
1062                 return;
1063
1064         INIT_LIST_HEAD(dev_list);
1065
1066         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1067                 s = 0;
1068                 e = 0;
1069         } else {
1070                 s = 1;
1071                 e = device->phys_port_cnt;
1072         }
1073
1074         for (p = s; p <= e; ++p) {
1075                 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1076                         continue;
1077                 dev = ipoib_add_port("ib", device, p);
1078                 if (!IS_ERR(dev)) {
1079                         priv = dev->if_softc;
1080                         list_add_tail(&priv->list, dev_list);
1081                 }
1082         }
1083
1084         ib_set_client_data(device, &ipoib_client, dev_list);
1085 }
1086
1087 static void
1088 ipoib_remove_one(struct ib_device *device, void *client_data)
1089 {
1090         struct ipoib_dev_priv *priv, *tmp;
1091         struct list_head *dev_list = client_data;
1092
1093         if (!dev_list)
1094                 return;
1095
1096         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1097                 return;
1098
1099         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1100                 if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1101                         continue;
1102
1103                 ipoib_stop(priv);
1104
1105                 ib_unregister_event_handler(&priv->event_handler);
1106
1107                 /* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1108
1109                 flush_workqueue(ipoib_workqueue);
1110
1111                 ipoib_dev_cleanup(priv);
1112                 ipoib_detach(priv);
1113         }
1114
1115         kfree(dev_list);
1116 }
1117
1118 static void
1119 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1120 {
1121         struct ipoib_dev_priv *parent;
1122         struct ipoib_dev_priv *priv;
1123         struct ifnet *dev;
1124         uint16_t pkey;
1125         int error;
1126
1127         if (ifp->if_type != IFT_INFINIBAND)
1128                 return;
1129         dev = VLAN_DEVAT(ifp, vtag);
1130         if (dev == NULL)
1131                 return;
1132         priv = NULL;
1133         error = 0;
1134         parent = ifp->if_softc;
1135         /* We only support 15 bits of pkey. */
1136         if (vtag & 0x8000)
1137                 return;
1138         pkey = vtag | 0x8000;   /* Set full membership bit. */
1139         if (pkey == parent->pkey)
1140                 return;
1141         /* Check for dups */
1142         mutex_lock(&parent->vlan_mutex);
1143         list_for_each_entry(priv, &parent->child_intfs, list) {
1144                 if (priv->pkey == pkey) {
1145                         priv = NULL;
1146                         error = EBUSY;
1147                         goto out;
1148                 }
1149         }
1150         priv = ipoib_priv_alloc();
1151         priv->dev = dev;
1152         priv->max_ib_mtu = parent->max_ib_mtu;
1153         priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1154         set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1155         error = ipoib_set_dev_features(priv, parent->ca);
1156         if (error)
1157                 goto out;
1158         priv->pkey = pkey;
1159         priv->broadcastaddr[8] = pkey >> 8;
1160         priv->broadcastaddr[9] = pkey & 0xff;
1161         dev->if_broadcastaddr = priv->broadcastaddr;
1162         error = ipoib_dev_init(priv, parent->ca, parent->port);
1163         if (error)
1164                 goto out;
1165         priv->parent = parent->dev;
1166         list_add_tail(&priv->list, &parent->child_intfs);
1167         VLAN_SETCOOKIE(dev, priv);
1168         dev->if_start = ipoib_vlan_start;
1169         dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1170         dev->if_hdrlen = IPOIB_HEADER_LEN;
1171         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1172                 ipoib_open(priv);
1173         mutex_unlock(&parent->vlan_mutex);
1174         return;
1175 out:
1176         mutex_unlock(&parent->vlan_mutex);
1177         if (priv)
1178                 free(priv, M_TEMP);
1179         if (error)
1180                 ipoib_warn(parent,
1181                     "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1182                     parent->ca->name, parent->port, vtag);
1183         return;
1184 }
1185
1186 static void
1187 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1188 {
1189         struct ipoib_dev_priv *parent;
1190         struct ipoib_dev_priv *priv;
1191         struct ifnet *dev;
1192         uint16_t pkey;
1193
1194         if (ifp->if_type != IFT_INFINIBAND)
1195                 return;
1196
1197         dev = VLAN_DEVAT(ifp, vtag);
1198         if (dev)
1199                 VLAN_SETCOOKIE(dev, NULL);
1200         pkey = vtag | 0x8000;
1201         parent = ifp->if_softc;
1202         mutex_lock(&parent->vlan_mutex);
1203         list_for_each_entry(priv, &parent->child_intfs, list) {
1204                 if (priv->pkey == pkey) {
1205                         ipoib_dev_cleanup(priv);
1206                         list_del(&priv->list);
1207                         break;
1208                 }
1209         }
1210         mutex_unlock(&parent->vlan_mutex);
1211 }
1212
1213 eventhandler_tag ipoib_vlan_attach;
1214 eventhandler_tag ipoib_vlan_detach;
1215
1216 static int __init
1217 ipoib_init_module(void)
1218 {
1219         int ret;
1220
1221         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1222         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1223         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1224
1225         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1226         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1227         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1228                                                      IPOIB_MIN_QUEUE_SIZE));
1229 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1230         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1231 #endif
1232
1233         ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1234                 ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1235         ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1236                 ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1237
1238         /*
1239          * We create our own workqueue mainly because we want to be
1240          * able to flush it when devices are being removed.  We can't
1241          * use schedule_work()/flush_scheduled_work() because both
1242          * unregister_netdev() and linkwatch_event take the rtnl lock,
1243          * so flush_scheduled_work() can deadlock during device
1244          * removal.
1245          */
1246         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1247         if (!ipoib_workqueue) {
1248                 ret = -ENOMEM;
1249                 goto err_fs;
1250         }
1251
1252         ib_sa_register_client(&ipoib_sa_client);
1253
1254         ret = ib_register_client(&ipoib_client);
1255         if (ret)
1256                 goto err_sa;
1257
1258         return 0;
1259
1260 err_sa:
1261         ib_sa_unregister_client(&ipoib_sa_client);
1262         destroy_workqueue(ipoib_workqueue);
1263
1264 err_fs:
1265         return ret;
1266 }
1267
1268 static void __exit
1269 ipoib_cleanup_module(void)
1270 {
1271
1272         EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1273         EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1274         ib_unregister_client(&ipoib_client);
1275         ib_sa_unregister_client(&ipoib_sa_client);
1276         destroy_workqueue(ipoib_workqueue);
1277 }
1278
1279 /*
1280  * Infiniband output routine.
1281  */
1282 static int
1283 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1284         const struct sockaddr *dst, struct route *ro)
1285 {
1286         u_char edst[INFINIBAND_ALEN];
1287 #if defined(INET) || defined(INET6)
1288         struct llentry *lle = NULL;
1289 #endif
1290         struct ipoib_header *eh;
1291         int error = 0, is_gw = 0;
1292         short type;
1293
1294         if (ro != NULL)
1295                 is_gw = (ro->ro_flags & RT_HAS_GW) != 0;
1296 #ifdef MAC
1297         error = mac_ifnet_check_transmit(ifp, m);
1298         if (error)
1299                 goto bad;
1300 #endif
1301
1302         M_PROFILE(m);
1303         if (ifp->if_flags & IFF_MONITOR) {
1304                 error = ENETDOWN;
1305                 goto bad;
1306         }
1307         if (!((ifp->if_flags & IFF_UP) &&
1308             (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1309                 error = ENETDOWN;
1310                 goto bad;
1311         }
1312
1313         switch (dst->sa_family) {
1314 #ifdef INET
1315         case AF_INET:
1316                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1317                         memcpy(edst, lle->ll_addr, sizeof(edst));
1318                 else if (m->m_flags & M_MCAST)
1319                         ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1320                 else
1321                         error = arpresolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1322                 if (error)
1323                         return (error == EWOULDBLOCK ? 0 : error);
1324                 type = htons(ETHERTYPE_IP);
1325                 break;
1326         case AF_ARP:
1327         {
1328                 struct arphdr *ah;
1329                 ah = mtod(m, struct arphdr *);
1330                 ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1331
1332                 switch(ntohs(ah->ar_op)) {
1333                 case ARPOP_REVREQUEST:
1334                 case ARPOP_REVREPLY:
1335                         type = htons(ETHERTYPE_REVARP);
1336                         break;
1337                 case ARPOP_REQUEST:
1338                 case ARPOP_REPLY:
1339                 default:
1340                         type = htons(ETHERTYPE_ARP);
1341                         break;
1342                 }
1343
1344                 if (m->m_flags & M_BCAST)
1345                         bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1346                 else
1347                         bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1348
1349         }
1350         break;
1351 #endif
1352 #ifdef INET6
1353         case AF_INET6:
1354                 if (lle != NULL && (lle->la_flags & LLE_VALID))
1355                         memcpy(edst, lle->ll_addr, sizeof(edst));
1356                 else if (m->m_flags & M_MCAST)
1357                         ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1358                 else
1359                         error = nd6_resolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1360                 if (error)
1361                         return error;
1362                 type = htons(ETHERTYPE_IPV6);
1363                 break;
1364 #endif
1365
1366         default:
1367                 if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1368                 error = EAFNOSUPPORT;
1369                 goto bad;
1370         }
1371
1372         /*
1373          * Add local net header.  If no space in first mbuf,
1374          * allocate another.
1375          */
1376         M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1377         if (m == NULL) {
1378                 error = ENOBUFS;
1379                 goto bad;
1380         }
1381         eh = mtod(m, struct ipoib_header *);
1382         (void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1383         (void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1384
1385         /*
1386          * Queue message on interface, update output statistics if
1387          * successful, and start output if interface not yet active.
1388          */
1389         return ((ifp->if_transmit)(ifp, m));
1390 bad:
1391         if (m != NULL)
1392                 m_freem(m);
1393         return (error);
1394 }
1395
1396 /*
1397  * Upper layer processing for a received Infiniband packet.
1398  */
1399 void
1400 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1401 {
1402         int isr;
1403
1404 #ifdef MAC
1405         /*
1406          * Tag the mbuf with an appropriate MAC label before any other
1407          * consumers can get to it.
1408          */
1409         mac_ifnet_create_mbuf(ifp, m);
1410 #endif
1411         /* Allow monitor mode to claim this frame, after stats are updated. */
1412         if (ifp->if_flags & IFF_MONITOR) {
1413                 if_printf(ifp, "discard frame at IFF_MONITOR\n");
1414                 m_freem(m);
1415                 return;
1416         }
1417         /*
1418          * Dispatch frame to upper layer.
1419          */
1420         switch (proto) {
1421 #ifdef INET
1422         case ETHERTYPE_IP:
1423                 isr = NETISR_IP;
1424                 break;
1425
1426         case ETHERTYPE_ARP:
1427                 if (ifp->if_flags & IFF_NOARP) {
1428                         /* Discard packet if ARP is disabled on interface */
1429                         m_freem(m);
1430                         return;
1431                 }
1432                 isr = NETISR_ARP;
1433                 break;
1434 #endif
1435 #ifdef INET6
1436         case ETHERTYPE_IPV6:
1437                 isr = NETISR_IPV6;
1438                 break;
1439 #endif
1440         default:
1441                 goto discard;
1442         }
1443         netisr_dispatch(isr, m);
1444         return;
1445
1446 discard:
1447         m_freem(m);
1448 }
1449
1450 /*
1451  * Process a received Infiniband packet.
1452  */
1453 static void
1454 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1455 {
1456         struct ipoib_header *eh;
1457
1458         if ((ifp->if_flags & IFF_UP) == 0) {
1459                 m_freem(m);
1460                 return;
1461         }
1462         CURVNET_SET_QUIET(ifp->if_vnet);
1463
1464         /* Let BPF have it before we strip the header. */
1465         IPOIB_MTAP(ifp, m);
1466         eh = mtod(m, struct ipoib_header *);
1467         /*
1468          * Reset layer specific mbuf flags to avoid confusing upper layers.
1469          * Strip off Infiniband header.
1470          */
1471         m->m_flags &= ~M_VLANTAG;
1472         m_clrprotoflags(m);
1473         m_adj(m, IPOIB_HEADER_LEN);
1474
1475         if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1476                 if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1477                     ifp->if_addrlen) == 0)
1478                         m->m_flags |= M_BCAST;
1479                 else
1480                         m->m_flags |= M_MCAST;
1481                 if_inc_counter(ifp, IFCOUNTER_IMCASTS, 1);
1482         }
1483
1484         ipoib_demux(ifp, m, ntohs(eh->proto));
1485         CURVNET_RESTORE();
1486 }
1487
1488 static int
1489 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1490         struct sockaddr *sa)
1491 {
1492         struct sockaddr_dl *sdl;
1493 #ifdef INET
1494         struct sockaddr_in *sin;
1495 #endif
1496 #ifdef INET6
1497         struct sockaddr_in6 *sin6;
1498 #endif
1499         u_char *e_addr;
1500
1501         switch(sa->sa_family) {
1502         case AF_LINK:
1503                 /*
1504                  * No mapping needed. Just check that it's a valid MC address.
1505                  */
1506                 sdl = (struct sockaddr_dl *)sa;
1507                 e_addr = LLADDR(sdl);
1508                 if (!IPOIB_IS_MULTICAST(e_addr))
1509                         return EADDRNOTAVAIL;
1510                 *llsa = NULL;
1511                 return 0;
1512
1513 #ifdef INET
1514         case AF_INET:
1515                 sin = (struct sockaddr_in *)sa;
1516                 if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1517                         return EADDRNOTAVAIL;
1518                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1519                 sdl->sdl_alen = INFINIBAND_ALEN;
1520                 e_addr = LLADDR(sdl);
1521                 ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1522                     e_addr);
1523                 *llsa = (struct sockaddr *)sdl;
1524                 return 0;
1525 #endif
1526 #ifdef INET6
1527         case AF_INET6:
1528                 sin6 = (struct sockaddr_in6 *)sa;
1529                 /*
1530                  * An IP6 address of 0 means listen to all
1531                  * of the multicast address used for IP6.
1532                  * This has no meaning in ipoib.
1533                  */
1534                 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1535                         return EADDRNOTAVAIL;
1536                 if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1537                         return EADDRNOTAVAIL;
1538                 sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1539                 sdl->sdl_alen = INFINIBAND_ALEN;
1540                 e_addr = LLADDR(sdl);
1541                 ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1542                 *llsa = (struct sockaddr *)sdl;
1543                 return 0;
1544 #endif
1545
1546         default:
1547                 return EAFNOSUPPORT;
1548         }
1549 }
1550
1551 module_init(ipoib_init_module);
1552 module_exit(ipoib_cleanup_module);
1553
1554 static int
1555 ipoib_evhand(module_t mod, int event, void *arg)
1556 {
1557                         return (0);
1558 }
1559
1560 static moduledata_t ipoib_mod = {
1561                         .name = "ipoib",
1562                                         .evhand = ipoib_evhand,
1563 };
1564
1565 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1566 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1567 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);