1 /* 2 * Copyright (c) 2004 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include "ipoib.h" 36 37 #include <linux/module.h> 38 39 #include <linux/init.h> 40 #include <linux/slab.h> 41 #include <linux/kernel.h> 42 #include <linux/vmalloc.h> 43 44 #include <linux/if_arp.h> /* For ARPHRD_xxx */ 45 46 #include <linux/ip.h> 47 #include <linux/in.h> 48 49 #include <linux/jhash.h> 50 #include <net/arp.h> 51 #include <net/addrconf.h> 52 #include <linux/inetdevice.h> 53 #include <rdma/ib_cache.h> 54 55 #define DRV_VERSION "1.0.0" 56 57 const char ipoib_driver_version[] = DRV_VERSION; 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 __read_mostly = IPOIB_TX_RING_SIZE; 64 int ipoib_recvq_size __read_mostly = 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; 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 net_device *dev; 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 void ipoib_neigh_reclaim(struct rcu_head *rp); 96 static struct net_device *ipoib_get_net_dev_by_params( 97 struct ib_device *dev, u8 port, u16 pkey, 98 const union ib_gid *gid, const struct sockaddr *addr, 99 void *client_data); 100 static int ipoib_set_mac(struct net_device *dev, void *addr); 101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr, 102 int cmd); 103 104 static struct ib_client ipoib_client = { 105 .name = "ipoib", 106 .add = ipoib_add_one, 107 .remove = ipoib_remove_one, 108 .get_net_dev_by_params = ipoib_get_net_dev_by_params, 109 }; 110 111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 112 static int ipoib_netdev_event(struct notifier_block *this, 113 unsigned long event, void *ptr) 114 { 115 struct netdev_notifier_info *ni = ptr; 116 struct net_device *dev = ni->dev; 117 118 if (dev->netdev_ops->ndo_open != ipoib_open) 119 return NOTIFY_DONE; 120 121 switch (event) { 122 case NETDEV_REGISTER: 123 ipoib_create_debug_files(dev); 124 break; 125 case NETDEV_CHANGENAME: 126 ipoib_delete_debug_files(dev); 127 ipoib_create_debug_files(dev); 128 break; 129 case NETDEV_UNREGISTER: 130 ipoib_delete_debug_files(dev); 131 break; 132 } 133 134 return NOTIFY_DONE; 135 } 136 #endif 137 138 int ipoib_open(struct net_device *dev) 139 { 140 struct ipoib_dev_priv *priv = ipoib_priv(dev); 141 142 ipoib_dbg(priv, "bringing up interface\n"); 143 144 netif_carrier_off(dev); 145 146 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 147 148 priv->sm_fullmember_sendonly_support = false; 149 150 if (ipoib_ib_dev_open(dev)) { 151 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) 152 return 0; 153 goto err_disable; 154 } 155 156 ipoib_ib_dev_up(dev); 157 158 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 159 struct ipoib_dev_priv *cpriv; 160 161 /* Bring up any child interfaces too */ 162 down_read(&priv->vlan_rwsem); 163 list_for_each_entry(cpriv, &priv->child_intfs, list) { 164 int flags; 165 166 flags = cpriv->dev->flags; 167 if (flags & IFF_UP) 168 continue; 169 170 dev_change_flags(cpriv->dev, flags | IFF_UP); 171 } 172 up_read(&priv->vlan_rwsem); 173 } 174 175 netif_start_queue(dev); 176 177 return 0; 178 179 err_disable: 180 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 181 182 return -EINVAL; 183 } 184 185 static int ipoib_stop(struct net_device *dev) 186 { 187 struct ipoib_dev_priv *priv = ipoib_priv(dev); 188 189 ipoib_dbg(priv, "stopping interface\n"); 190 191 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 192 193 netif_stop_queue(dev); 194 195 ipoib_ib_dev_down(dev); 196 ipoib_ib_dev_stop(dev); 197 198 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 199 struct ipoib_dev_priv *cpriv; 200 201 /* Bring down any child interfaces too */ 202 down_read(&priv->vlan_rwsem); 203 list_for_each_entry(cpriv, &priv->child_intfs, list) { 204 int flags; 205 206 flags = cpriv->dev->flags; 207 if (!(flags & IFF_UP)) 208 continue; 209 210 dev_change_flags(cpriv->dev, flags & ~IFF_UP); 211 } 212 up_read(&priv->vlan_rwsem); 213 } 214 215 return 0; 216 } 217 218 static void ipoib_uninit(struct net_device *dev) 219 { 220 ipoib_dev_cleanup(dev); 221 } 222 223 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features) 224 { 225 struct ipoib_dev_priv *priv = ipoib_priv(dev); 226 227 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags)) 228 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO); 229 230 return features; 231 } 232 233 static int ipoib_change_mtu(struct net_device *dev, int new_mtu) 234 { 235 struct ipoib_dev_priv *priv = ipoib_priv(dev); 236 int ret = 0; 237 238 /* dev->mtu > 2K ==> connected mode */ 239 if (ipoib_cm_admin_enabled(dev)) { 240 if (new_mtu > ipoib_cm_max_mtu(dev)) 241 return -EINVAL; 242 243 if (new_mtu > priv->mcast_mtu) 244 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n", 245 priv->mcast_mtu); 246 247 dev->mtu = new_mtu; 248 return 0; 249 } 250 251 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu)) 252 return -EINVAL; 253 254 priv->admin_mtu = new_mtu; 255 256 if (priv->mcast_mtu < priv->admin_mtu) 257 ipoib_dbg(priv, "MTU must be smaller than the underlying " 258 "link layer MTU - 4 (%u)\n", priv->mcast_mtu); 259 260 new_mtu = min(priv->mcast_mtu, priv->admin_mtu); 261 262 if (priv->rn_ops->ndo_change_mtu) { 263 bool carrier_status = netif_carrier_ok(dev); 264 265 netif_carrier_off(dev); 266 267 /* notify lower level on the real mtu */ 268 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu); 269 270 if (carrier_status) 271 netif_carrier_on(dev); 272 } else { 273 dev->mtu = new_mtu; 274 } 275 276 return ret; 277 } 278 279 static void ipoib_get_stats(struct net_device *dev, 280 struct rtnl_link_stats64 *stats) 281 { 282 struct ipoib_dev_priv *priv = ipoib_priv(dev); 283 284 if (priv->rn_ops->ndo_get_stats64) 285 priv->rn_ops->ndo_get_stats64(dev, stats); 286 else 287 netdev_stats_to_stats64(stats, &dev->stats); 288 } 289 290 /* Called with an RCU read lock taken */ 291 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr, 292 struct net_device *dev) 293 { 294 struct net *net = dev_net(dev); 295 struct in_device *in_dev; 296 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr; 297 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr; 298 __be32 ret_addr; 299 300 switch (addr->sa_family) { 301 case AF_INET: 302 in_dev = in_dev_get(dev); 303 if (!in_dev) 304 return false; 305 306 ret_addr = inet_confirm_addr(net, in_dev, 0, 307 addr_in->sin_addr.s_addr, 308 RT_SCOPE_HOST); 309 in_dev_put(in_dev); 310 if (ret_addr) 311 return true; 312 313 break; 314 case AF_INET6: 315 if (IS_ENABLED(CONFIG_IPV6) && 316 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1)) 317 return true; 318 319 break; 320 } 321 return false; 322 } 323 324 /** 325 * Find the master net_device on top of the given net_device. 326 * @dev: base IPoIB net_device 327 * 328 * Returns the master net_device with a reference held, or the same net_device 329 * if no master exists. 330 */ 331 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev) 332 { 333 struct net_device *master; 334 335 rcu_read_lock(); 336 master = netdev_master_upper_dev_get_rcu(dev); 337 if (master) 338 dev_hold(master); 339 rcu_read_unlock(); 340 341 if (master) 342 return master; 343 344 dev_hold(dev); 345 return dev; 346 } 347 348 struct ipoib_walk_data { 349 const struct sockaddr *addr; 350 struct net_device *result; 351 }; 352 353 static int ipoib_upper_walk(struct net_device *upper, void *_data) 354 { 355 struct ipoib_walk_data *data = _data; 356 int ret = 0; 357 358 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) { 359 dev_hold(upper); 360 data->result = upper; 361 ret = 1; 362 } 363 364 return ret; 365 } 366 367 /** 368 * Find a net_device matching the given address, which is an upper device of 369 * the given net_device. 370 * @addr: IP address to look for. 371 * @dev: base IPoIB net_device 372 * 373 * If found, returns the net_device with a reference held. Otherwise return 374 * NULL. 375 */ 376 static struct net_device *ipoib_get_net_dev_match_addr( 377 const struct sockaddr *addr, struct net_device *dev) 378 { 379 struct ipoib_walk_data data = { 380 .addr = addr, 381 }; 382 383 rcu_read_lock(); 384 if (ipoib_is_dev_match_addr_rcu(addr, dev)) { 385 dev_hold(dev); 386 data.result = dev; 387 goto out; 388 } 389 390 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data); 391 out: 392 rcu_read_unlock(); 393 return data.result; 394 } 395 396 /* returns the number of IPoIB netdevs on top a given ipoib device matching a 397 * pkey_index and address, if one exists. 398 * 399 * @found_net_dev: contains a matching net_device if the return value >= 1, 400 * with a reference held. */ 401 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv, 402 const union ib_gid *gid, 403 u16 pkey_index, 404 const struct sockaddr *addr, 405 int nesting, 406 struct net_device **found_net_dev) 407 { 408 struct ipoib_dev_priv *child_priv; 409 struct net_device *net_dev = NULL; 410 int matches = 0; 411 412 if (priv->pkey_index == pkey_index && 413 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) { 414 if (!addr) { 415 net_dev = ipoib_get_master_net_dev(priv->dev); 416 } else { 417 /* Verify the net_device matches the IP address, as 418 * IPoIB child devices currently share a GID. */ 419 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev); 420 } 421 if (net_dev) { 422 if (!*found_net_dev) 423 *found_net_dev = net_dev; 424 else 425 dev_put(net_dev); 426 ++matches; 427 } 428 } 429 430 /* Check child interfaces */ 431 down_read_nested(&priv->vlan_rwsem, nesting); 432 list_for_each_entry(child_priv, &priv->child_intfs, list) { 433 matches += ipoib_match_gid_pkey_addr(child_priv, gid, 434 pkey_index, addr, 435 nesting + 1, 436 found_net_dev); 437 if (matches > 1) 438 break; 439 } 440 up_read(&priv->vlan_rwsem); 441 442 return matches; 443 } 444 445 /* Returns the number of matching net_devs found (between 0 and 2). Also 446 * return the matching net_device in the @net_dev parameter, holding a 447 * reference to the net_device, if the number of matches >= 1 */ 448 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port, 449 u16 pkey_index, 450 const union ib_gid *gid, 451 const struct sockaddr *addr, 452 struct net_device **net_dev) 453 { 454 struct ipoib_dev_priv *priv; 455 int matches = 0; 456 457 *net_dev = NULL; 458 459 list_for_each_entry(priv, dev_list, list) { 460 if (priv->port != port) 461 continue; 462 463 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index, 464 addr, 0, net_dev); 465 if (matches > 1) 466 break; 467 } 468 469 return matches; 470 } 471 472 static struct net_device *ipoib_get_net_dev_by_params( 473 struct ib_device *dev, u8 port, u16 pkey, 474 const union ib_gid *gid, const struct sockaddr *addr, 475 void *client_data) 476 { 477 struct net_device *net_dev; 478 struct list_head *dev_list = client_data; 479 u16 pkey_index; 480 int matches; 481 int ret; 482 483 if (!rdma_protocol_ib(dev, port)) 484 return NULL; 485 486 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index); 487 if (ret) 488 return NULL; 489 490 if (!dev_list) 491 return NULL; 492 493 /* See if we can find a unique device matching the L2 parameters */ 494 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 495 gid, NULL, &net_dev); 496 497 switch (matches) { 498 case 0: 499 return NULL; 500 case 1: 501 return net_dev; 502 } 503 504 dev_put(net_dev); 505 506 /* Couldn't find a unique device with L2 parameters only. Use L3 507 * address to uniquely match the net device */ 508 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 509 gid, addr, &net_dev); 510 switch (matches) { 511 case 0: 512 return NULL; 513 default: 514 dev_warn_ratelimited(&dev->dev, 515 "duplicate IP address detected\n"); 516 /* Fall through */ 517 case 1: 518 return net_dev; 519 } 520 } 521 522 int ipoib_set_mode(struct net_device *dev, const char *buf) 523 { 524 struct ipoib_dev_priv *priv = ipoib_priv(dev); 525 526 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) && 527 !strcmp(buf, "connected\n")) || 528 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) && 529 !strcmp(buf, "datagram\n"))) { 530 return 0; 531 } 532 533 /* flush paths if we switch modes so that connections are restarted */ 534 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) { 535 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 536 ipoib_warn(priv, "enabling connected mode " 537 "will cause multicast packet drops\n"); 538 netdev_update_features(dev); 539 dev_set_mtu(dev, ipoib_cm_max_mtu(dev)); 540 rtnl_unlock(); 541 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM; 542 543 ipoib_flush_paths(dev); 544 return (!rtnl_trylock()) ? -EBUSY : 0; 545 } 546 547 if (!strcmp(buf, "datagram\n")) { 548 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 549 netdev_update_features(dev); 550 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu)); 551 rtnl_unlock(); 552 ipoib_flush_paths(dev); 553 return (!rtnl_trylock()) ? -EBUSY : 0; 554 } 555 556 return -EINVAL; 557 } 558 559 struct ipoib_path *__path_find(struct net_device *dev, void *gid) 560 { 561 struct ipoib_dev_priv *priv = ipoib_priv(dev); 562 struct rb_node *n = priv->path_tree.rb_node; 563 struct ipoib_path *path; 564 int ret; 565 566 while (n) { 567 path = rb_entry(n, struct ipoib_path, rb_node); 568 569 ret = memcmp(gid, path->pathrec.dgid.raw, 570 sizeof (union ib_gid)); 571 572 if (ret < 0) 573 n = n->rb_left; 574 else if (ret > 0) 575 n = n->rb_right; 576 else 577 return path; 578 } 579 580 return NULL; 581 } 582 583 static int __path_add(struct net_device *dev, struct ipoib_path *path) 584 { 585 struct ipoib_dev_priv *priv = ipoib_priv(dev); 586 struct rb_node **n = &priv->path_tree.rb_node; 587 struct rb_node *pn = NULL; 588 struct ipoib_path *tpath; 589 int ret; 590 591 while (*n) { 592 pn = *n; 593 tpath = rb_entry(pn, struct ipoib_path, rb_node); 594 595 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw, 596 sizeof (union ib_gid)); 597 if (ret < 0) 598 n = &pn->rb_left; 599 else if (ret > 0) 600 n = &pn->rb_right; 601 else 602 return -EEXIST; 603 } 604 605 rb_link_node(&path->rb_node, pn, n); 606 rb_insert_color(&path->rb_node, &priv->path_tree); 607 608 list_add_tail(&path->list, &priv->path_list); 609 610 return 0; 611 } 612 613 static void path_free(struct net_device *dev, struct ipoib_path *path) 614 { 615 struct sk_buff *skb; 616 617 while ((skb = __skb_dequeue(&path->queue))) 618 dev_kfree_skb_irq(skb); 619 620 ipoib_dbg(ipoib_priv(dev), "path_free\n"); 621 622 /* remove all neigh connected to this path */ 623 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 624 625 if (path->ah) 626 ipoib_put_ah(path->ah); 627 628 kfree(path); 629 } 630 631 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 632 633 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev) 634 { 635 struct ipoib_path_iter *iter; 636 637 iter = kmalloc(sizeof *iter, GFP_KERNEL); 638 if (!iter) 639 return NULL; 640 641 iter->dev = dev; 642 memset(iter->path.pathrec.dgid.raw, 0, 16); 643 644 if (ipoib_path_iter_next(iter)) { 645 kfree(iter); 646 return NULL; 647 } 648 649 return iter; 650 } 651 652 int ipoib_path_iter_next(struct ipoib_path_iter *iter) 653 { 654 struct ipoib_dev_priv *priv = ipoib_priv(iter->dev); 655 struct rb_node *n; 656 struct ipoib_path *path; 657 int ret = 1; 658 659 spin_lock_irq(&priv->lock); 660 661 n = rb_first(&priv->path_tree); 662 663 while (n) { 664 path = rb_entry(n, struct ipoib_path, rb_node); 665 666 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw, 667 sizeof (union ib_gid)) < 0) { 668 iter->path = *path; 669 ret = 0; 670 break; 671 } 672 673 n = rb_next(n); 674 } 675 676 spin_unlock_irq(&priv->lock); 677 678 return ret; 679 } 680 681 void ipoib_path_iter_read(struct ipoib_path_iter *iter, 682 struct ipoib_path *path) 683 { 684 *path = iter->path; 685 } 686 687 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */ 688 689 void ipoib_mark_paths_invalid(struct net_device *dev) 690 { 691 struct ipoib_dev_priv *priv = ipoib_priv(dev); 692 struct ipoib_path *path, *tp; 693 694 spin_lock_irq(&priv->lock); 695 696 list_for_each_entry_safe(path, tp, &priv->path_list, list) { 697 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n", 698 be32_to_cpu(sa_path_get_dlid(&path->pathrec)), 699 path->pathrec.dgid.raw); 700 if (path->ah) 701 path->ah->valid = 0; 702 } 703 704 spin_unlock_irq(&priv->lock); 705 } 706 707 static void push_pseudo_header(struct sk_buff *skb, const char *daddr) 708 { 709 struct ipoib_pseudo_header *phdr; 710 711 phdr = skb_push(skb, sizeof(*phdr)); 712 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN); 713 } 714 715 void ipoib_flush_paths(struct net_device *dev) 716 { 717 struct ipoib_dev_priv *priv = ipoib_priv(dev); 718 struct ipoib_path *path, *tp; 719 LIST_HEAD(remove_list); 720 unsigned long flags; 721 722 netif_tx_lock_bh(dev); 723 spin_lock_irqsave(&priv->lock, flags); 724 725 list_splice_init(&priv->path_list, &remove_list); 726 727 list_for_each_entry(path, &remove_list, list) 728 rb_erase(&path->rb_node, &priv->path_tree); 729 730 list_for_each_entry_safe(path, tp, &remove_list, list) { 731 if (path->query) 732 ib_sa_cancel_query(path->query_id, path->query); 733 spin_unlock_irqrestore(&priv->lock, flags); 734 netif_tx_unlock_bh(dev); 735 wait_for_completion(&path->done); 736 path_free(dev, path); 737 netif_tx_lock_bh(dev); 738 spin_lock_irqsave(&priv->lock, flags); 739 } 740 741 spin_unlock_irqrestore(&priv->lock, flags); 742 netif_tx_unlock_bh(dev); 743 } 744 745 static void path_rec_completion(int status, 746 struct sa_path_rec *pathrec, 747 void *path_ptr) 748 { 749 struct ipoib_path *path = path_ptr; 750 struct net_device *dev = path->dev; 751 struct ipoib_dev_priv *priv = ipoib_priv(dev); 752 struct ipoib_ah *ah = NULL; 753 struct ipoib_ah *old_ah = NULL; 754 struct ipoib_neigh *neigh, *tn; 755 struct sk_buff_head skqueue; 756 struct sk_buff *skb; 757 unsigned long flags; 758 759 if (!status) 760 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n", 761 be32_to_cpu(sa_path_get_dlid(pathrec)), 762 pathrec->dgid.raw); 763 else 764 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n", 765 status, path->pathrec.dgid.raw); 766 767 skb_queue_head_init(&skqueue); 768 769 if (!status) { 770 struct rdma_ah_attr av; 771 772 if (!ib_init_ah_attr_from_path(priv->ca, priv->port, 773 pathrec, &av)) 774 ah = ipoib_create_ah(dev, priv->pd, &av); 775 } 776 777 spin_lock_irqsave(&priv->lock, flags); 778 779 if (!IS_ERR_OR_NULL(ah)) { 780 /* 781 * pathrec.dgid is used as the database key from the LLADDR, 782 * it must remain unchanged even if the SA returns a different 783 * GID to use in the AH. 784 */ 785 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw, 786 sizeof(union ib_gid))) { 787 ipoib_dbg( 788 priv, 789 "%s got PathRec for gid %pI6 while asked for %pI6\n", 790 dev->name, pathrec->dgid.raw, 791 path->pathrec.dgid.raw); 792 memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw, 793 sizeof(union ib_gid)); 794 } 795 796 path->pathrec = *pathrec; 797 798 old_ah = path->ah; 799 path->ah = ah; 800 801 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n", 802 ah, be32_to_cpu(sa_path_get_dlid(pathrec)), 803 pathrec->sl); 804 805 while ((skb = __skb_dequeue(&path->queue))) 806 __skb_queue_tail(&skqueue, skb); 807 808 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) { 809 if (neigh->ah) { 810 WARN_ON(neigh->ah != old_ah); 811 /* 812 * Dropping the ah reference inside 813 * priv->lock is safe here, because we 814 * will hold one more reference from 815 * the original value of path->ah (ie 816 * old_ah). 817 */ 818 ipoib_put_ah(neigh->ah); 819 } 820 kref_get(&path->ah->ref); 821 neigh->ah = path->ah; 822 823 if (ipoib_cm_enabled(dev, neigh->daddr)) { 824 if (!ipoib_cm_get(neigh)) 825 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, 826 path, 827 neigh)); 828 if (!ipoib_cm_get(neigh)) { 829 ipoib_neigh_free(neigh); 830 continue; 831 } 832 } 833 834 while ((skb = __skb_dequeue(&neigh->queue))) 835 __skb_queue_tail(&skqueue, skb); 836 } 837 path->ah->valid = 1; 838 } 839 840 path->query = NULL; 841 complete(&path->done); 842 843 spin_unlock_irqrestore(&priv->lock, flags); 844 845 if (IS_ERR_OR_NULL(ah)) 846 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 847 848 if (old_ah) 849 ipoib_put_ah(old_ah); 850 851 while ((skb = __skb_dequeue(&skqueue))) { 852 int ret; 853 skb->dev = dev; 854 ret = dev_queue_xmit(skb); 855 if (ret) 856 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n", 857 __func__, ret); 858 } 859 } 860 861 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path, 862 void *gid) 863 { 864 path->dev = priv->dev; 865 866 if (rdma_cap_opa_ah(priv->ca, priv->port)) 867 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA; 868 else 869 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB; 870 871 memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid)); 872 path->pathrec.sgid = priv->local_gid; 873 path->pathrec.pkey = cpu_to_be16(priv->pkey); 874 path->pathrec.numb_path = 1; 875 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class; 876 } 877 878 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid) 879 { 880 struct ipoib_dev_priv *priv = ipoib_priv(dev); 881 struct ipoib_path *path; 882 883 if (!priv->broadcast) 884 return NULL; 885 886 path = kzalloc(sizeof *path, GFP_ATOMIC); 887 if (!path) 888 return NULL; 889 890 skb_queue_head_init(&path->queue); 891 892 INIT_LIST_HEAD(&path->neigh_list); 893 894 init_path_rec(priv, path, gid); 895 896 return path; 897 } 898 899 static int path_rec_start(struct net_device *dev, 900 struct ipoib_path *path) 901 { 902 struct ipoib_dev_priv *priv = ipoib_priv(dev); 903 904 ipoib_dbg(priv, "Start path record lookup for %pI6\n", 905 path->pathrec.dgid.raw); 906 907 init_completion(&path->done); 908 909 path->query_id = 910 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port, 911 &path->pathrec, 912 IB_SA_PATH_REC_DGID | 913 IB_SA_PATH_REC_SGID | 914 IB_SA_PATH_REC_NUMB_PATH | 915 IB_SA_PATH_REC_TRAFFIC_CLASS | 916 IB_SA_PATH_REC_PKEY, 917 1000, GFP_ATOMIC, 918 path_rec_completion, 919 path, &path->query); 920 if (path->query_id < 0) { 921 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id); 922 path->query = NULL; 923 complete(&path->done); 924 return path->query_id; 925 } 926 927 return 0; 928 } 929 930 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr, 931 struct net_device *dev) 932 { 933 struct ipoib_dev_priv *priv = ipoib_priv(dev); 934 struct ipoib_path *path; 935 unsigned long flags; 936 937 spin_lock_irqsave(&priv->lock, flags); 938 939 path = __path_find(dev, daddr + 4); 940 if (!path) 941 goto out; 942 if (!path->query) 943 path_rec_start(dev, path); 944 out: 945 spin_unlock_irqrestore(&priv->lock, flags); 946 } 947 948 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr, 949 struct net_device *dev) 950 { 951 struct ipoib_dev_priv *priv = ipoib_priv(dev); 952 struct rdma_netdev *rn = netdev_priv(dev); 953 struct ipoib_path *path; 954 struct ipoib_neigh *neigh; 955 unsigned long flags; 956 957 spin_lock_irqsave(&priv->lock, flags); 958 neigh = ipoib_neigh_alloc(daddr, dev); 959 if (!neigh) { 960 spin_unlock_irqrestore(&priv->lock, flags); 961 ++dev->stats.tx_dropped; 962 dev_kfree_skb_any(skb); 963 return NULL; 964 } 965 966 /* To avoid race condition, make sure that the 967 * neigh will be added only once. 968 */ 969 if (unlikely(!list_empty(&neigh->list))) { 970 spin_unlock_irqrestore(&priv->lock, flags); 971 return neigh; 972 } 973 974 path = __path_find(dev, daddr + 4); 975 if (!path) { 976 path = path_rec_create(dev, daddr + 4); 977 if (!path) 978 goto err_path; 979 980 __path_add(dev, path); 981 } 982 983 list_add_tail(&neigh->list, &path->neigh_list); 984 985 if (path->ah && path->ah->valid) { 986 kref_get(&path->ah->ref); 987 neigh->ah = path->ah; 988 989 if (ipoib_cm_enabled(dev, neigh->daddr)) { 990 if (!ipoib_cm_get(neigh)) 991 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh)); 992 if (!ipoib_cm_get(neigh)) { 993 ipoib_neigh_free(neigh); 994 goto err_drop; 995 } 996 if (skb_queue_len(&neigh->queue) < 997 IPOIB_MAX_PATH_REC_QUEUE) { 998 push_pseudo_header(skb, neigh->daddr); 999 __skb_queue_tail(&neigh->queue, skb); 1000 } else { 1001 ipoib_warn(priv, "queue length limit %d. Packet drop.\n", 1002 skb_queue_len(&neigh->queue)); 1003 goto err_drop; 1004 } 1005 } else { 1006 spin_unlock_irqrestore(&priv->lock, flags); 1007 path->ah->last_send = rn->send(dev, skb, path->ah->ah, 1008 IPOIB_QPN(daddr)); 1009 ipoib_neigh_put(neigh); 1010 return NULL; 1011 } 1012 } else { 1013 neigh->ah = NULL; 1014 1015 if (!path->query && path_rec_start(dev, path)) 1016 goto err_path; 1017 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1018 push_pseudo_header(skb, neigh->daddr); 1019 __skb_queue_tail(&neigh->queue, skb); 1020 } else { 1021 goto err_drop; 1022 } 1023 } 1024 1025 spin_unlock_irqrestore(&priv->lock, flags); 1026 ipoib_neigh_put(neigh); 1027 return NULL; 1028 1029 err_path: 1030 ipoib_neigh_free(neigh); 1031 err_drop: 1032 ++dev->stats.tx_dropped; 1033 dev_kfree_skb_any(skb); 1034 1035 spin_unlock_irqrestore(&priv->lock, flags); 1036 ipoib_neigh_put(neigh); 1037 1038 return NULL; 1039 } 1040 1041 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev, 1042 struct ipoib_pseudo_header *phdr) 1043 { 1044 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1045 struct rdma_netdev *rn = netdev_priv(dev); 1046 struct ipoib_path *path; 1047 unsigned long flags; 1048 1049 spin_lock_irqsave(&priv->lock, flags); 1050 1051 /* no broadcast means that all paths are (going to be) not valid */ 1052 if (!priv->broadcast) 1053 goto drop_and_unlock; 1054 1055 path = __path_find(dev, phdr->hwaddr + 4); 1056 if (!path || !path->ah || !path->ah->valid) { 1057 if (!path) { 1058 path = path_rec_create(dev, phdr->hwaddr + 4); 1059 if (!path) 1060 goto drop_and_unlock; 1061 __path_add(dev, path); 1062 } else { 1063 /* 1064 * make sure there are no changes in the existing 1065 * path record 1066 */ 1067 init_path_rec(priv, path, phdr->hwaddr + 4); 1068 } 1069 if (!path->query && path_rec_start(dev, path)) { 1070 goto drop_and_unlock; 1071 } 1072 1073 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1074 push_pseudo_header(skb, phdr->hwaddr); 1075 __skb_queue_tail(&path->queue, skb); 1076 goto unlock; 1077 } else { 1078 goto drop_and_unlock; 1079 } 1080 } 1081 1082 spin_unlock_irqrestore(&priv->lock, flags); 1083 ipoib_dbg(priv, "Send unicast ARP to %08x\n", 1084 be32_to_cpu(sa_path_get_dlid(&path->pathrec))); 1085 path->ah->last_send = rn->send(dev, skb, path->ah->ah, 1086 IPOIB_QPN(phdr->hwaddr)); 1087 return; 1088 1089 drop_and_unlock: 1090 ++dev->stats.tx_dropped; 1091 dev_kfree_skb_any(skb); 1092 unlock: 1093 spin_unlock_irqrestore(&priv->lock, flags); 1094 } 1095 1096 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev) 1097 { 1098 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1099 struct rdma_netdev *rn = netdev_priv(dev); 1100 struct ipoib_neigh *neigh; 1101 struct ipoib_pseudo_header *phdr; 1102 struct ipoib_header *header; 1103 unsigned long flags; 1104 1105 phdr = (struct ipoib_pseudo_header *) skb->data; 1106 skb_pull(skb, sizeof(*phdr)); 1107 header = (struct ipoib_header *) skb->data; 1108 1109 if (unlikely(phdr->hwaddr[4] == 0xff)) { 1110 /* multicast, arrange "if" according to probability */ 1111 if ((header->proto != htons(ETH_P_IP)) && 1112 (header->proto != htons(ETH_P_IPV6)) && 1113 (header->proto != htons(ETH_P_ARP)) && 1114 (header->proto != htons(ETH_P_RARP)) && 1115 (header->proto != htons(ETH_P_TIPC))) { 1116 /* ethertype not supported by IPoIB */ 1117 ++dev->stats.tx_dropped; 1118 dev_kfree_skb_any(skb); 1119 return NETDEV_TX_OK; 1120 } 1121 /* Add in the P_Key for multicast*/ 1122 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff; 1123 phdr->hwaddr[9] = priv->pkey & 0xff; 1124 1125 neigh = ipoib_neigh_get(dev, phdr->hwaddr); 1126 if (likely(neigh)) 1127 goto send_using_neigh; 1128 ipoib_mcast_send(dev, phdr->hwaddr, skb); 1129 return NETDEV_TX_OK; 1130 } 1131 1132 /* unicast, arrange "switch" according to probability */ 1133 switch (header->proto) { 1134 case htons(ETH_P_IP): 1135 case htons(ETH_P_IPV6): 1136 case htons(ETH_P_TIPC): 1137 neigh = ipoib_neigh_get(dev, phdr->hwaddr); 1138 if (unlikely(!neigh)) { 1139 neigh = neigh_add_path(skb, phdr->hwaddr, dev); 1140 if (likely(!neigh)) 1141 return NETDEV_TX_OK; 1142 } 1143 break; 1144 case htons(ETH_P_ARP): 1145 case htons(ETH_P_RARP): 1146 /* for unicast ARP and RARP should always perform path find */ 1147 unicast_arp_send(skb, dev, phdr); 1148 return NETDEV_TX_OK; 1149 default: 1150 /* ethertype not supported by IPoIB */ 1151 ++dev->stats.tx_dropped; 1152 dev_kfree_skb_any(skb); 1153 return NETDEV_TX_OK; 1154 } 1155 1156 send_using_neigh: 1157 /* note we now hold a ref to neigh */ 1158 if (ipoib_cm_get(neigh)) { 1159 if (ipoib_cm_up(neigh)) { 1160 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh)); 1161 goto unref; 1162 } 1163 } else if (neigh->ah && neigh->ah->valid) { 1164 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah, 1165 IPOIB_QPN(phdr->hwaddr)); 1166 goto unref; 1167 } else if (neigh->ah) { 1168 neigh_refresh_path(neigh, phdr->hwaddr, dev); 1169 } 1170 1171 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1172 push_pseudo_header(skb, phdr->hwaddr); 1173 spin_lock_irqsave(&priv->lock, flags); 1174 __skb_queue_tail(&neigh->queue, skb); 1175 spin_unlock_irqrestore(&priv->lock, flags); 1176 } else { 1177 ++dev->stats.tx_dropped; 1178 dev_kfree_skb_any(skb); 1179 } 1180 1181 unref: 1182 ipoib_neigh_put(neigh); 1183 1184 return NETDEV_TX_OK; 1185 } 1186 1187 static void ipoib_timeout(struct net_device *dev) 1188 { 1189 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1190 1191 ipoib_warn(priv, "transmit timeout: latency %d msecs\n", 1192 jiffies_to_msecs(jiffies - dev_trans_start(dev))); 1193 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n", 1194 netif_queue_stopped(dev), 1195 priv->tx_head, priv->tx_tail); 1196 /* XXX reset QP, etc. */ 1197 } 1198 1199 static int ipoib_hard_header(struct sk_buff *skb, 1200 struct net_device *dev, 1201 unsigned short type, 1202 const void *daddr, const void *saddr, unsigned len) 1203 { 1204 struct ipoib_header *header; 1205 1206 header = skb_push(skb, sizeof *header); 1207 1208 header->proto = htons(type); 1209 header->reserved = 0; 1210 1211 /* 1212 * we don't rely on dst_entry structure, always stuff the 1213 * destination address into skb hard header so we can figure out where 1214 * to send the packet later. 1215 */ 1216 push_pseudo_header(skb, daddr); 1217 1218 return IPOIB_HARD_LEN; 1219 } 1220 1221 static void ipoib_set_mcast_list(struct net_device *dev) 1222 { 1223 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1224 1225 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) { 1226 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set"); 1227 return; 1228 } 1229 1230 queue_work(priv->wq, &priv->restart_task); 1231 } 1232 1233 static int ipoib_get_iflink(const struct net_device *dev) 1234 { 1235 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1236 1237 /* parent interface */ 1238 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) 1239 return dev->ifindex; 1240 1241 /* child/vlan interface */ 1242 return priv->parent->ifindex; 1243 } 1244 1245 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr) 1246 { 1247 /* 1248 * Use only the address parts that contributes to spreading 1249 * The subnet prefix is not used as one can not connect to 1250 * same remote port (GUID) using the same remote QPN via two 1251 * different subnets. 1252 */ 1253 /* qpn octets[1:4) & port GUID octets[12:20) */ 1254 u32 *d32 = (u32 *) daddr; 1255 u32 hv; 1256 1257 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0); 1258 return hv & htbl->mask; 1259 } 1260 1261 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr) 1262 { 1263 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1264 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1265 struct ipoib_neigh_hash *htbl; 1266 struct ipoib_neigh *neigh = NULL; 1267 u32 hash_val; 1268 1269 rcu_read_lock_bh(); 1270 1271 htbl = rcu_dereference_bh(ntbl->htbl); 1272 1273 if (!htbl) 1274 goto out_unlock; 1275 1276 hash_val = ipoib_addr_hash(htbl, daddr); 1277 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]); 1278 neigh != NULL; 1279 neigh = rcu_dereference_bh(neigh->hnext)) { 1280 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1281 /* found, take one ref on behalf of the caller */ 1282 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1283 /* deleted */ 1284 neigh = NULL; 1285 goto out_unlock; 1286 } 1287 1288 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)) 1289 neigh->alive = jiffies; 1290 goto out_unlock; 1291 } 1292 } 1293 1294 out_unlock: 1295 rcu_read_unlock_bh(); 1296 return neigh; 1297 } 1298 1299 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv) 1300 { 1301 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1302 struct ipoib_neigh_hash *htbl; 1303 unsigned long neigh_obsolete; 1304 unsigned long dt; 1305 unsigned long flags; 1306 int i; 1307 LIST_HEAD(remove_list); 1308 1309 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1310 return; 1311 1312 spin_lock_irqsave(&priv->lock, flags); 1313 1314 htbl = rcu_dereference_protected(ntbl->htbl, 1315 lockdep_is_held(&priv->lock)); 1316 1317 if (!htbl) 1318 goto out_unlock; 1319 1320 /* neigh is obsolete if it was idle for two GC periods */ 1321 dt = 2 * arp_tbl.gc_interval; 1322 neigh_obsolete = jiffies - dt; 1323 /* handle possible race condition */ 1324 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1325 goto out_unlock; 1326 1327 for (i = 0; i < htbl->size; i++) { 1328 struct ipoib_neigh *neigh; 1329 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1330 1331 while ((neigh = rcu_dereference_protected(*np, 1332 lockdep_is_held(&priv->lock))) != NULL) { 1333 /* was the neigh idle for two GC periods */ 1334 if (time_after(neigh_obsolete, neigh->alive)) { 1335 1336 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list); 1337 1338 rcu_assign_pointer(*np, 1339 rcu_dereference_protected(neigh->hnext, 1340 lockdep_is_held(&priv->lock))); 1341 /* remove from path/mc list */ 1342 list_del_init(&neigh->list); 1343 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1344 } else { 1345 np = &neigh->hnext; 1346 } 1347 1348 } 1349 } 1350 1351 out_unlock: 1352 spin_unlock_irqrestore(&priv->lock, flags); 1353 ipoib_mcast_remove_list(&remove_list); 1354 } 1355 1356 static void ipoib_reap_neigh(struct work_struct *work) 1357 { 1358 struct ipoib_dev_priv *priv = 1359 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work); 1360 1361 __ipoib_reap_neigh(priv); 1362 1363 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1364 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1365 arp_tbl.gc_interval); 1366 } 1367 1368 1369 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr, 1370 struct net_device *dev) 1371 { 1372 struct ipoib_neigh *neigh; 1373 1374 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC); 1375 if (!neigh) 1376 return NULL; 1377 1378 neigh->dev = dev; 1379 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr)); 1380 skb_queue_head_init(&neigh->queue); 1381 INIT_LIST_HEAD(&neigh->list); 1382 ipoib_cm_set(neigh, NULL); 1383 /* one ref on behalf of the caller */ 1384 atomic_set(&neigh->refcnt, 1); 1385 1386 return neigh; 1387 } 1388 1389 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr, 1390 struct net_device *dev) 1391 { 1392 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1393 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1394 struct ipoib_neigh_hash *htbl; 1395 struct ipoib_neigh *neigh; 1396 u32 hash_val; 1397 1398 htbl = rcu_dereference_protected(ntbl->htbl, 1399 lockdep_is_held(&priv->lock)); 1400 if (!htbl) { 1401 neigh = NULL; 1402 goto out_unlock; 1403 } 1404 1405 /* need to add a new neigh, but maybe some other thread succeeded? 1406 * recalc hash, maybe hash resize took place so we do a search 1407 */ 1408 hash_val = ipoib_addr_hash(htbl, daddr); 1409 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val], 1410 lockdep_is_held(&priv->lock)); 1411 neigh != NULL; 1412 neigh = rcu_dereference_protected(neigh->hnext, 1413 lockdep_is_held(&priv->lock))) { 1414 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1415 /* found, take one ref on behalf of the caller */ 1416 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1417 /* deleted */ 1418 neigh = NULL; 1419 break; 1420 } 1421 neigh->alive = jiffies; 1422 goto out_unlock; 1423 } 1424 } 1425 1426 neigh = ipoib_neigh_ctor(daddr, dev); 1427 if (!neigh) 1428 goto out_unlock; 1429 1430 /* one ref on behalf of the hash table */ 1431 atomic_inc(&neigh->refcnt); 1432 neigh->alive = jiffies; 1433 /* put in hash */ 1434 rcu_assign_pointer(neigh->hnext, 1435 rcu_dereference_protected(htbl->buckets[hash_val], 1436 lockdep_is_held(&priv->lock))); 1437 rcu_assign_pointer(htbl->buckets[hash_val], neigh); 1438 atomic_inc(&ntbl->entries); 1439 1440 out_unlock: 1441 1442 return neigh; 1443 } 1444 1445 void ipoib_neigh_dtor(struct ipoib_neigh *neigh) 1446 { 1447 /* neigh reference count was dropprd to zero */ 1448 struct net_device *dev = neigh->dev; 1449 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1450 struct sk_buff *skb; 1451 if (neigh->ah) 1452 ipoib_put_ah(neigh->ah); 1453 while ((skb = __skb_dequeue(&neigh->queue))) { 1454 ++dev->stats.tx_dropped; 1455 dev_kfree_skb_any(skb); 1456 } 1457 if (ipoib_cm_get(neigh)) 1458 ipoib_cm_destroy_tx(ipoib_cm_get(neigh)); 1459 ipoib_dbg(ipoib_priv(dev), 1460 "neigh free for %06x %pI6\n", 1461 IPOIB_QPN(neigh->daddr), 1462 neigh->daddr + 4); 1463 kfree(neigh); 1464 if (atomic_dec_and_test(&priv->ntbl.entries)) { 1465 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags)) 1466 complete(&priv->ntbl.flushed); 1467 } 1468 } 1469 1470 static void ipoib_neigh_reclaim(struct rcu_head *rp) 1471 { 1472 /* Called as a result of removal from hash table */ 1473 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu); 1474 /* note TX context may hold another ref */ 1475 ipoib_neigh_put(neigh); 1476 } 1477 1478 void ipoib_neigh_free(struct ipoib_neigh *neigh) 1479 { 1480 struct net_device *dev = neigh->dev; 1481 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1482 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1483 struct ipoib_neigh_hash *htbl; 1484 struct ipoib_neigh __rcu **np; 1485 struct ipoib_neigh *n; 1486 u32 hash_val; 1487 1488 htbl = rcu_dereference_protected(ntbl->htbl, 1489 lockdep_is_held(&priv->lock)); 1490 if (!htbl) 1491 return; 1492 1493 hash_val = ipoib_addr_hash(htbl, neigh->daddr); 1494 np = &htbl->buckets[hash_val]; 1495 for (n = rcu_dereference_protected(*np, 1496 lockdep_is_held(&priv->lock)); 1497 n != NULL; 1498 n = rcu_dereference_protected(*np, 1499 lockdep_is_held(&priv->lock))) { 1500 if (n == neigh) { 1501 /* found */ 1502 rcu_assign_pointer(*np, 1503 rcu_dereference_protected(neigh->hnext, 1504 lockdep_is_held(&priv->lock))); 1505 /* remove from parent list */ 1506 list_del_init(&neigh->list); 1507 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1508 return; 1509 } else { 1510 np = &n->hnext; 1511 } 1512 } 1513 } 1514 1515 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv) 1516 { 1517 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1518 struct ipoib_neigh_hash *htbl; 1519 struct ipoib_neigh __rcu **buckets; 1520 u32 size; 1521 1522 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1523 ntbl->htbl = NULL; 1524 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL); 1525 if (!htbl) 1526 return -ENOMEM; 1527 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1528 size = roundup_pow_of_two(arp_tbl.gc_thresh3); 1529 buckets = kcalloc(size, sizeof(*buckets), GFP_KERNEL); 1530 if (!buckets) { 1531 kfree(htbl); 1532 return -ENOMEM; 1533 } 1534 htbl->size = size; 1535 htbl->mask = (size - 1); 1536 htbl->buckets = buckets; 1537 RCU_INIT_POINTER(ntbl->htbl, htbl); 1538 htbl->ntbl = ntbl; 1539 atomic_set(&ntbl->entries, 0); 1540 1541 /* start garbage collection */ 1542 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1543 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1544 arp_tbl.gc_interval); 1545 1546 return 0; 1547 } 1548 1549 static void neigh_hash_free_rcu(struct rcu_head *head) 1550 { 1551 struct ipoib_neigh_hash *htbl = container_of(head, 1552 struct ipoib_neigh_hash, 1553 rcu); 1554 struct ipoib_neigh __rcu **buckets = htbl->buckets; 1555 struct ipoib_neigh_table *ntbl = htbl->ntbl; 1556 1557 kfree(buckets); 1558 kfree(htbl); 1559 complete(&ntbl->deleted); 1560 } 1561 1562 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid) 1563 { 1564 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1565 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1566 struct ipoib_neigh_hash *htbl; 1567 unsigned long flags; 1568 int i; 1569 1570 /* remove all neigh connected to a given path or mcast */ 1571 spin_lock_irqsave(&priv->lock, flags); 1572 1573 htbl = rcu_dereference_protected(ntbl->htbl, 1574 lockdep_is_held(&priv->lock)); 1575 1576 if (!htbl) 1577 goto out_unlock; 1578 1579 for (i = 0; i < htbl->size; i++) { 1580 struct ipoib_neigh *neigh; 1581 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1582 1583 while ((neigh = rcu_dereference_protected(*np, 1584 lockdep_is_held(&priv->lock))) != NULL) { 1585 /* delete neighs belong to this parent */ 1586 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) { 1587 rcu_assign_pointer(*np, 1588 rcu_dereference_protected(neigh->hnext, 1589 lockdep_is_held(&priv->lock))); 1590 /* remove from parent list */ 1591 list_del_init(&neigh->list); 1592 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1593 } else { 1594 np = &neigh->hnext; 1595 } 1596 1597 } 1598 } 1599 out_unlock: 1600 spin_unlock_irqrestore(&priv->lock, flags); 1601 } 1602 1603 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv) 1604 { 1605 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1606 struct ipoib_neigh_hash *htbl; 1607 unsigned long flags; 1608 int i, wait_flushed = 0; 1609 1610 init_completion(&priv->ntbl.flushed); 1611 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1612 1613 spin_lock_irqsave(&priv->lock, flags); 1614 1615 htbl = rcu_dereference_protected(ntbl->htbl, 1616 lockdep_is_held(&priv->lock)); 1617 if (!htbl) 1618 goto out_unlock; 1619 1620 wait_flushed = atomic_read(&priv->ntbl.entries); 1621 if (!wait_flushed) 1622 goto free_htbl; 1623 1624 for (i = 0; i < htbl->size; i++) { 1625 struct ipoib_neigh *neigh; 1626 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1627 1628 while ((neigh = rcu_dereference_protected(*np, 1629 lockdep_is_held(&priv->lock))) != NULL) { 1630 rcu_assign_pointer(*np, 1631 rcu_dereference_protected(neigh->hnext, 1632 lockdep_is_held(&priv->lock))); 1633 /* remove from path/mc list */ 1634 list_del_init(&neigh->list); 1635 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1636 } 1637 } 1638 1639 free_htbl: 1640 rcu_assign_pointer(ntbl->htbl, NULL); 1641 call_rcu(&htbl->rcu, neigh_hash_free_rcu); 1642 1643 out_unlock: 1644 spin_unlock_irqrestore(&priv->lock, flags); 1645 if (wait_flushed) 1646 wait_for_completion(&priv->ntbl.flushed); 1647 } 1648 1649 static void ipoib_neigh_hash_uninit(struct net_device *dev) 1650 { 1651 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1652 int stopped; 1653 1654 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n"); 1655 init_completion(&priv->ntbl.deleted); 1656 1657 /* Stop GC if called at init fail need to cancel work */ 1658 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1659 if (!stopped) 1660 cancel_delayed_work(&priv->neigh_reap_task); 1661 1662 ipoib_flush_neighs(priv); 1663 1664 wait_for_completion(&priv->ntbl.deleted); 1665 } 1666 1667 static void ipoib_napi_add(struct net_device *dev) 1668 { 1669 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1670 1671 netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC); 1672 netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE); 1673 } 1674 1675 static void ipoib_napi_del(struct net_device *dev) 1676 { 1677 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1678 1679 netif_napi_del(&priv->recv_napi); 1680 netif_napi_del(&priv->send_napi); 1681 } 1682 1683 static void ipoib_dev_uninit_default(struct net_device *dev) 1684 { 1685 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1686 1687 ipoib_transport_dev_cleanup(dev); 1688 1689 ipoib_napi_del(dev); 1690 1691 ipoib_cm_dev_cleanup(dev); 1692 1693 kfree(priv->rx_ring); 1694 vfree(priv->tx_ring); 1695 1696 priv->rx_ring = NULL; 1697 priv->tx_ring = NULL; 1698 } 1699 1700 static int ipoib_dev_init_default(struct net_device *dev) 1701 { 1702 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1703 1704 ipoib_napi_add(dev); 1705 1706 /* Allocate RX/TX "rings" to hold queued skbs */ 1707 priv->rx_ring = kcalloc(ipoib_recvq_size, 1708 sizeof(*priv->rx_ring), 1709 GFP_KERNEL); 1710 if (!priv->rx_ring) 1711 goto out; 1712 1713 priv->tx_ring = vzalloc(array_size(ipoib_sendq_size, 1714 sizeof(*priv->tx_ring))); 1715 if (!priv->tx_ring) { 1716 pr_warn("%s: failed to allocate TX ring (%d entries)\n", 1717 priv->ca->name, ipoib_sendq_size); 1718 goto out_rx_ring_cleanup; 1719 } 1720 1721 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */ 1722 1723 if (ipoib_transport_dev_init(dev, priv->ca)) { 1724 pr_warn("%s: ipoib_transport_dev_init failed\n", 1725 priv->ca->name); 1726 goto out_tx_ring_cleanup; 1727 } 1728 1729 /* after qp created set dev address */ 1730 priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff; 1731 priv->dev->dev_addr[2] = (priv->qp->qp_num >> 8) & 0xff; 1732 priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff; 1733 1734 return 0; 1735 1736 out_tx_ring_cleanup: 1737 vfree(priv->tx_ring); 1738 1739 out_rx_ring_cleanup: 1740 kfree(priv->rx_ring); 1741 1742 out: 1743 ipoib_napi_del(dev); 1744 return -ENOMEM; 1745 } 1746 1747 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr, 1748 int cmd) 1749 { 1750 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1751 1752 if (!priv->rn_ops->ndo_do_ioctl) 1753 return -EOPNOTSUPP; 1754 1755 return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd); 1756 } 1757 1758 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port) 1759 { 1760 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1761 int ret = -ENOMEM; 1762 1763 priv->ca = ca; 1764 priv->port = port; 1765 priv->qp = NULL; 1766 1767 /* 1768 * the various IPoIB tasks assume they will never race against 1769 * themselves, so always use a single thread workqueue 1770 */ 1771 priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM); 1772 if (!priv->wq) { 1773 pr_warn("%s: failed to allocate device WQ\n", dev->name); 1774 goto out; 1775 } 1776 1777 /* create pd, which used both for control and datapath*/ 1778 priv->pd = ib_alloc_pd(priv->ca, 0); 1779 if (IS_ERR(priv->pd)) { 1780 pr_warn("%s: failed to allocate PD\n", ca->name); 1781 goto clean_wq; 1782 } 1783 1784 ret = priv->rn_ops->ndo_init(dev); 1785 if (ret) { 1786 pr_warn("%s failed to init HW resource\n", dev->name); 1787 goto out_free_pd; 1788 } 1789 1790 if (ipoib_neigh_hash_init(priv) < 0) { 1791 pr_warn("%s failed to init neigh hash\n", dev->name); 1792 goto out_dev_uninit; 1793 } 1794 1795 if (dev->flags & IFF_UP) { 1796 if (ipoib_ib_dev_open(dev)) { 1797 pr_warn("%s failed to open device\n", dev->name); 1798 ret = -ENODEV; 1799 goto out_dev_uninit; 1800 } 1801 } 1802 1803 return 0; 1804 1805 out_dev_uninit: 1806 ipoib_ib_dev_cleanup(dev); 1807 1808 out_free_pd: 1809 if (priv->pd) { 1810 ib_dealloc_pd(priv->pd); 1811 priv->pd = NULL; 1812 } 1813 1814 clean_wq: 1815 if (priv->wq) { 1816 destroy_workqueue(priv->wq); 1817 priv->wq = NULL; 1818 } 1819 1820 out: 1821 return ret; 1822 } 1823 1824 void ipoib_dev_cleanup(struct net_device *dev) 1825 { 1826 struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv; 1827 LIST_HEAD(head); 1828 1829 ASSERT_RTNL(); 1830 1831 /* Delete any child interfaces first */ 1832 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) { 1833 /* Stop GC on child */ 1834 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags); 1835 cancel_delayed_work(&cpriv->neigh_reap_task); 1836 unregister_netdevice_queue(cpriv->dev, &head); 1837 } 1838 unregister_netdevice_many(&head); 1839 1840 ipoib_neigh_hash_uninit(dev); 1841 1842 ipoib_ib_dev_cleanup(dev); 1843 1844 /* no more works over the priv->wq */ 1845 if (priv->wq) { 1846 flush_workqueue(priv->wq); 1847 destroy_workqueue(priv->wq); 1848 priv->wq = NULL; 1849 } 1850 } 1851 1852 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state) 1853 { 1854 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1855 1856 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state); 1857 } 1858 1859 static int ipoib_get_vf_config(struct net_device *dev, int vf, 1860 struct ifla_vf_info *ivf) 1861 { 1862 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1863 int err; 1864 1865 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf); 1866 if (err) 1867 return err; 1868 1869 ivf->vf = vf; 1870 1871 return 0; 1872 } 1873 1874 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type) 1875 { 1876 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1877 1878 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID) 1879 return -EINVAL; 1880 1881 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type); 1882 } 1883 1884 static int ipoib_get_vf_stats(struct net_device *dev, int vf, 1885 struct ifla_vf_stats *vf_stats) 1886 { 1887 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1888 1889 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats); 1890 } 1891 1892 static const struct header_ops ipoib_header_ops = { 1893 .create = ipoib_hard_header, 1894 }; 1895 1896 static const struct net_device_ops ipoib_netdev_ops_pf = { 1897 .ndo_uninit = ipoib_uninit, 1898 .ndo_open = ipoib_open, 1899 .ndo_stop = ipoib_stop, 1900 .ndo_change_mtu = ipoib_change_mtu, 1901 .ndo_fix_features = ipoib_fix_features, 1902 .ndo_start_xmit = ipoib_start_xmit, 1903 .ndo_tx_timeout = ipoib_timeout, 1904 .ndo_set_rx_mode = ipoib_set_mcast_list, 1905 .ndo_get_iflink = ipoib_get_iflink, 1906 .ndo_set_vf_link_state = ipoib_set_vf_link_state, 1907 .ndo_get_vf_config = ipoib_get_vf_config, 1908 .ndo_get_vf_stats = ipoib_get_vf_stats, 1909 .ndo_set_vf_guid = ipoib_set_vf_guid, 1910 .ndo_set_mac_address = ipoib_set_mac, 1911 .ndo_get_stats64 = ipoib_get_stats, 1912 .ndo_do_ioctl = ipoib_ioctl, 1913 }; 1914 1915 static const struct net_device_ops ipoib_netdev_ops_vf = { 1916 .ndo_uninit = ipoib_uninit, 1917 .ndo_open = ipoib_open, 1918 .ndo_stop = ipoib_stop, 1919 .ndo_change_mtu = ipoib_change_mtu, 1920 .ndo_fix_features = ipoib_fix_features, 1921 .ndo_start_xmit = ipoib_start_xmit, 1922 .ndo_tx_timeout = ipoib_timeout, 1923 .ndo_set_rx_mode = ipoib_set_mcast_list, 1924 .ndo_get_iflink = ipoib_get_iflink, 1925 .ndo_get_stats64 = ipoib_get_stats, 1926 .ndo_do_ioctl = ipoib_ioctl, 1927 }; 1928 1929 void ipoib_setup_common(struct net_device *dev) 1930 { 1931 dev->header_ops = &ipoib_header_ops; 1932 1933 ipoib_set_ethtool_ops(dev); 1934 1935 dev->watchdog_timeo = HZ; 1936 1937 dev->flags |= IFF_BROADCAST | IFF_MULTICAST; 1938 1939 dev->hard_header_len = IPOIB_HARD_LEN; 1940 dev->addr_len = INFINIBAND_ALEN; 1941 dev->type = ARPHRD_INFINIBAND; 1942 dev->tx_queue_len = ipoib_sendq_size * 2; 1943 dev->features = (NETIF_F_VLAN_CHALLENGED | 1944 NETIF_F_HIGHDMA); 1945 netif_keep_dst(dev); 1946 1947 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN); 1948 } 1949 1950 static void ipoib_build_priv(struct net_device *dev) 1951 { 1952 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1953 1954 priv->dev = dev; 1955 spin_lock_init(&priv->lock); 1956 init_rwsem(&priv->vlan_rwsem); 1957 mutex_init(&priv->mcast_mutex); 1958 mutex_init(&priv->sysfs_mutex); 1959 1960 INIT_LIST_HEAD(&priv->path_list); 1961 INIT_LIST_HEAD(&priv->child_intfs); 1962 INIT_LIST_HEAD(&priv->dead_ahs); 1963 INIT_LIST_HEAD(&priv->multicast_list); 1964 1965 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task); 1966 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task); 1967 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light); 1968 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal); 1969 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy); 1970 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task); 1971 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah); 1972 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh); 1973 } 1974 1975 static const struct net_device_ops ipoib_netdev_default_pf = { 1976 .ndo_init = ipoib_dev_init_default, 1977 .ndo_uninit = ipoib_dev_uninit_default, 1978 .ndo_open = ipoib_ib_dev_open_default, 1979 .ndo_stop = ipoib_ib_dev_stop_default, 1980 }; 1981 1982 static struct net_device 1983 *ipoib_create_netdev_default(struct ib_device *hca, 1984 const char *name, 1985 unsigned char name_assign_type, 1986 void (*setup)(struct net_device *)) 1987 { 1988 struct net_device *dev; 1989 struct rdma_netdev *rn; 1990 1991 dev = alloc_netdev((int)sizeof(struct rdma_netdev), 1992 name, 1993 name_assign_type, setup); 1994 if (!dev) 1995 return NULL; 1996 1997 rn = netdev_priv(dev); 1998 1999 rn->send = ipoib_send; 2000 rn->attach_mcast = ipoib_mcast_attach; 2001 rn->detach_mcast = ipoib_mcast_detach; 2002 rn->free_rdma_netdev = free_netdev; 2003 rn->hca = hca; 2004 2005 dev->netdev_ops = &ipoib_netdev_default_pf; 2006 2007 return dev; 2008 } 2009 2010 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port, 2011 const char *name) 2012 { 2013 struct net_device *dev; 2014 2015 if (hca->alloc_rdma_netdev) { 2016 dev = hca->alloc_rdma_netdev(hca, port, 2017 RDMA_NETDEV_IPOIB, name, 2018 NET_NAME_UNKNOWN, 2019 ipoib_setup_common); 2020 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP) 2021 return NULL; 2022 } 2023 2024 if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP) 2025 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN, 2026 ipoib_setup_common); 2027 2028 return dev; 2029 } 2030 2031 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port, 2032 const char *name) 2033 { 2034 struct net_device *dev; 2035 struct ipoib_dev_priv *priv; 2036 struct rdma_netdev *rn; 2037 2038 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 2039 if (!priv) 2040 return NULL; 2041 2042 dev = ipoib_get_netdev(hca, port, name); 2043 if (!dev) 2044 goto free_priv; 2045 2046 priv->rn_ops = dev->netdev_ops; 2047 2048 /* fixme : should be after the query_cap */ 2049 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION) 2050 dev->netdev_ops = &ipoib_netdev_ops_vf; 2051 else 2052 dev->netdev_ops = &ipoib_netdev_ops_pf; 2053 2054 rn = netdev_priv(dev); 2055 rn->clnt_priv = priv; 2056 ipoib_build_priv(dev); 2057 2058 return priv; 2059 free_priv: 2060 kfree(priv); 2061 return NULL; 2062 } 2063 2064 static ssize_t show_pkey(struct device *dev, 2065 struct device_attribute *attr, char *buf) 2066 { 2067 struct net_device *ndev = to_net_dev(dev); 2068 struct ipoib_dev_priv *priv = ipoib_priv(ndev); 2069 2070 return sprintf(buf, "0x%04x\n", priv->pkey); 2071 } 2072 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL); 2073 2074 static ssize_t show_umcast(struct device *dev, 2075 struct device_attribute *attr, char *buf) 2076 { 2077 struct net_device *ndev = to_net_dev(dev); 2078 struct ipoib_dev_priv *priv = ipoib_priv(ndev); 2079 2080 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags)); 2081 } 2082 2083 void ipoib_set_umcast(struct net_device *ndev, int umcast_val) 2084 { 2085 struct ipoib_dev_priv *priv = ipoib_priv(ndev); 2086 2087 if (umcast_val > 0) { 2088 set_bit(IPOIB_FLAG_UMCAST, &priv->flags); 2089 ipoib_warn(priv, "ignoring multicast groups joined directly " 2090 "by userspace\n"); 2091 } else 2092 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags); 2093 } 2094 2095 static ssize_t set_umcast(struct device *dev, 2096 struct device_attribute *attr, 2097 const char *buf, size_t count) 2098 { 2099 unsigned long umcast_val = simple_strtoul(buf, NULL, 0); 2100 2101 ipoib_set_umcast(to_net_dev(dev), umcast_val); 2102 2103 return count; 2104 } 2105 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast); 2106 2107 int ipoib_add_umcast_attr(struct net_device *dev) 2108 { 2109 return device_create_file(&dev->dev, &dev_attr_umcast); 2110 } 2111 2112 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid) 2113 { 2114 struct ipoib_dev_priv *child_priv; 2115 struct net_device *netdev = priv->dev; 2116 2117 netif_addr_lock_bh(netdev); 2118 2119 memcpy(&priv->local_gid.global.interface_id, 2120 &gid->global.interface_id, 2121 sizeof(gid->global.interface_id)); 2122 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid)); 2123 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 2124 2125 netif_addr_unlock_bh(netdev); 2126 2127 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 2128 down_read(&priv->vlan_rwsem); 2129 list_for_each_entry(child_priv, &priv->child_intfs, list) 2130 set_base_guid(child_priv, gid); 2131 up_read(&priv->vlan_rwsem); 2132 } 2133 } 2134 2135 static int ipoib_check_lladdr(struct net_device *dev, 2136 struct sockaddr_storage *ss) 2137 { 2138 union ib_gid *gid = (union ib_gid *)(ss->__data + 4); 2139 int ret = 0; 2140 2141 netif_addr_lock_bh(dev); 2142 2143 /* Make sure the QPN, reserved and subnet prefix match the current 2144 * lladdr, it also makes sure the lladdr is unicast. 2145 */ 2146 if (memcmp(dev->dev_addr, ss->__data, 2147 4 + sizeof(gid->global.subnet_prefix)) || 2148 gid->global.interface_id == 0) 2149 ret = -EINVAL; 2150 2151 netif_addr_unlock_bh(dev); 2152 2153 return ret; 2154 } 2155 2156 static int ipoib_set_mac(struct net_device *dev, void *addr) 2157 { 2158 struct ipoib_dev_priv *priv = ipoib_priv(dev); 2159 struct sockaddr_storage *ss = addr; 2160 int ret; 2161 2162 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev)) 2163 return -EBUSY; 2164 2165 ret = ipoib_check_lladdr(dev, ss); 2166 if (ret) 2167 return ret; 2168 2169 set_base_guid(priv, (union ib_gid *)(ss->__data + 4)); 2170 2171 queue_work(ipoib_workqueue, &priv->flush_light); 2172 2173 return 0; 2174 } 2175 2176 static ssize_t create_child(struct device *dev, 2177 struct device_attribute *attr, 2178 const char *buf, size_t count) 2179 { 2180 int pkey; 2181 int ret; 2182 2183 if (sscanf(buf, "%i", &pkey) != 1) 2184 return -EINVAL; 2185 2186 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000) 2187 return -EINVAL; 2188 2189 /* 2190 * Set the full membership bit, so that we join the right 2191 * broadcast group, etc. 2192 */ 2193 pkey |= 0x8000; 2194 2195 ret = ipoib_vlan_add(to_net_dev(dev), pkey); 2196 2197 return ret ? ret : count; 2198 } 2199 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child); 2200 2201 static ssize_t delete_child(struct device *dev, 2202 struct device_attribute *attr, 2203 const char *buf, size_t count) 2204 { 2205 int pkey; 2206 int ret; 2207 2208 if (sscanf(buf, "%i", &pkey) != 1) 2209 return -EINVAL; 2210 2211 if (pkey < 0 || pkey > 0xffff) 2212 return -EINVAL; 2213 2214 ret = ipoib_vlan_delete(to_net_dev(dev), pkey); 2215 2216 return ret ? ret : count; 2217 2218 } 2219 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child); 2220 2221 int ipoib_add_pkey_attr(struct net_device *dev) 2222 { 2223 return device_create_file(&dev->dev, &dev_attr_pkey); 2224 } 2225 2226 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca) 2227 { 2228 priv->hca_caps = hca->attrs.device_cap_flags; 2229 2230 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) { 2231 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM; 2232 2233 if (priv->hca_caps & IB_DEVICE_UD_TSO) 2234 priv->dev->hw_features |= NETIF_F_TSO; 2235 2236 priv->dev->features |= priv->dev->hw_features; 2237 } 2238 } 2239 2240 static struct net_device *ipoib_add_port(const char *format, 2241 struct ib_device *hca, u8 port) 2242 { 2243 struct ipoib_dev_priv *priv; 2244 struct ib_port_attr attr; 2245 struct rdma_netdev *rn; 2246 int result = -ENOMEM; 2247 2248 priv = ipoib_intf_alloc(hca, port, format); 2249 if (!priv) { 2250 pr_warn("%s, %d: ipoib_intf_alloc failed\n", hca->name, port); 2251 goto alloc_mem_failed; 2252 } 2253 2254 SET_NETDEV_DEV(priv->dev, hca->dev.parent); 2255 priv->dev->dev_id = port - 1; 2256 2257 result = ib_query_port(hca, port, &attr); 2258 if (result) { 2259 pr_warn("%s: ib_query_port %d failed\n", hca->name, port); 2260 goto device_init_failed; 2261 } 2262 2263 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu); 2264 2265 /* MTU will be reset when mcast join happens */ 2266 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu); 2267 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu; 2268 priv->dev->max_mtu = IPOIB_CM_MTU; 2269 2270 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh); 2271 2272 result = ib_query_pkey(hca, port, 0, &priv->pkey); 2273 if (result) { 2274 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n", 2275 hca->name, port, result); 2276 goto device_init_failed; 2277 } 2278 2279 ipoib_set_dev_features(priv, hca); 2280 2281 /* 2282 * Set the full membership bit, so that we join the right 2283 * broadcast group, etc. 2284 */ 2285 priv->pkey |= 0x8000; 2286 2287 priv->dev->broadcast[8] = priv->pkey >> 8; 2288 priv->dev->broadcast[9] = priv->pkey & 0xff; 2289 2290 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL); 2291 if (result) { 2292 pr_warn("%s: ib_query_gid port %d failed (ret = %d)\n", 2293 hca->name, port, result); 2294 goto device_init_failed; 2295 } 2296 2297 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, 2298 sizeof(union ib_gid)); 2299 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 2300 2301 result = ipoib_dev_init(priv->dev, hca, port); 2302 if (result) { 2303 pr_warn("%s: failed to initialize port %d (ret = %d)\n", 2304 hca->name, port, result); 2305 goto device_init_failed; 2306 } 2307 2308 INIT_IB_EVENT_HANDLER(&priv->event_handler, 2309 priv->ca, ipoib_event); 2310 ib_register_event_handler(&priv->event_handler); 2311 2312 /* call event handler to ensure pkey in sync */ 2313 queue_work(ipoib_workqueue, &priv->flush_heavy); 2314 2315 result = register_netdev(priv->dev); 2316 if (result) { 2317 pr_warn("%s: couldn't register ipoib port %d; error %d\n", 2318 hca->name, port, result); 2319 goto register_failed; 2320 } 2321 2322 result = -ENOMEM; 2323 if (ipoib_cm_add_mode_attr(priv->dev)) 2324 goto sysfs_failed; 2325 if (ipoib_add_pkey_attr(priv->dev)) 2326 goto sysfs_failed; 2327 if (ipoib_add_umcast_attr(priv->dev)) 2328 goto sysfs_failed; 2329 if (device_create_file(&priv->dev->dev, &dev_attr_create_child)) 2330 goto sysfs_failed; 2331 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child)) 2332 goto sysfs_failed; 2333 2334 return priv->dev; 2335 2336 sysfs_failed: 2337 unregister_netdev(priv->dev); 2338 2339 register_failed: 2340 ib_unregister_event_handler(&priv->event_handler); 2341 flush_workqueue(ipoib_workqueue); 2342 /* Stop GC if started before flush */ 2343 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2344 cancel_delayed_work(&priv->neigh_reap_task); 2345 flush_workqueue(priv->wq); 2346 ipoib_dev_cleanup(priv->dev); 2347 2348 device_init_failed: 2349 rn = netdev_priv(priv->dev); 2350 rn->free_rdma_netdev(priv->dev); 2351 kfree(priv); 2352 2353 alloc_mem_failed: 2354 return ERR_PTR(result); 2355 } 2356 2357 static void ipoib_add_one(struct ib_device *device) 2358 { 2359 struct list_head *dev_list; 2360 struct net_device *dev; 2361 struct ipoib_dev_priv *priv; 2362 int p; 2363 int count = 0; 2364 2365 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL); 2366 if (!dev_list) 2367 return; 2368 2369 INIT_LIST_HEAD(dev_list); 2370 2371 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) { 2372 if (!rdma_protocol_ib(device, p)) 2373 continue; 2374 dev = ipoib_add_port("ib%d", device, p); 2375 if (!IS_ERR(dev)) { 2376 priv = ipoib_priv(dev); 2377 list_add_tail(&priv->list, dev_list); 2378 count++; 2379 } 2380 } 2381 2382 if (!count) { 2383 kfree(dev_list); 2384 return; 2385 } 2386 2387 ib_set_client_data(device, &ipoib_client, dev_list); 2388 } 2389 2390 static void ipoib_remove_one(struct ib_device *device, void *client_data) 2391 { 2392 struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv; 2393 struct list_head *dev_list = client_data; 2394 2395 if (!dev_list) 2396 return; 2397 2398 list_for_each_entry_safe(priv, tmp, dev_list, list) { 2399 struct rdma_netdev *parent_rn = netdev_priv(priv->dev); 2400 2401 ib_unregister_event_handler(&priv->event_handler); 2402 flush_workqueue(ipoib_workqueue); 2403 2404 /* mark interface in the middle of destruction */ 2405 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags); 2406 2407 rtnl_lock(); 2408 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); 2409 rtnl_unlock(); 2410 2411 /* Stop GC */ 2412 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2413 cancel_delayed_work(&priv->neigh_reap_task); 2414 flush_workqueue(priv->wq); 2415 2416 /* Wrap rtnl_lock/unlock with mutex to protect sysfs calls */ 2417 mutex_lock(&priv->sysfs_mutex); 2418 unregister_netdev(priv->dev); 2419 mutex_unlock(&priv->sysfs_mutex); 2420 2421 parent_rn->free_rdma_netdev(priv->dev); 2422 2423 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) { 2424 struct rdma_netdev *child_rn; 2425 2426 child_rn = netdev_priv(cpriv->dev); 2427 child_rn->free_rdma_netdev(cpriv->dev); 2428 kfree(cpriv); 2429 } 2430 2431 kfree(priv); 2432 } 2433 2434 kfree(dev_list); 2435 } 2436 2437 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 2438 static struct notifier_block ipoib_netdev_notifier = { 2439 .notifier_call = ipoib_netdev_event, 2440 }; 2441 #endif 2442 2443 static int __init ipoib_init_module(void) 2444 { 2445 int ret; 2446 2447 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size); 2448 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE); 2449 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE); 2450 2451 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size); 2452 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE); 2453 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE); 2454 #ifdef CONFIG_INFINIBAND_IPOIB_CM 2455 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP); 2456 ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0); 2457 #endif 2458 2459 /* 2460 * When copying small received packets, we only copy from the 2461 * linear data part of the SKB, so we rely on this condition. 2462 */ 2463 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE); 2464 2465 ret = ipoib_register_debugfs(); 2466 if (ret) 2467 return ret; 2468 2469 /* 2470 * We create a global workqueue here that is used for all flush 2471 * operations. However, if you attempt to flush a workqueue 2472 * from a task on that same workqueue, it deadlocks the system. 2473 * We want to be able to flush the tasks associated with a 2474 * specific net device, so we also create a workqueue for each 2475 * netdevice. We queue up the tasks for that device only on 2476 * its private workqueue, and we only queue up flush events 2477 * on our global flush workqueue. This avoids the deadlocks. 2478 */ 2479 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 2480 WQ_MEM_RECLAIM); 2481 if (!ipoib_workqueue) { 2482 ret = -ENOMEM; 2483 goto err_fs; 2484 } 2485 2486 ib_sa_register_client(&ipoib_sa_client); 2487 2488 ret = ib_register_client(&ipoib_client); 2489 if (ret) 2490 goto err_sa; 2491 2492 ret = ipoib_netlink_init(); 2493 if (ret) 2494 goto err_client; 2495 2496 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 2497 register_netdevice_notifier(&ipoib_netdev_notifier); 2498 #endif 2499 return 0; 2500 2501 err_client: 2502 ib_unregister_client(&ipoib_client); 2503 2504 err_sa: 2505 ib_sa_unregister_client(&ipoib_sa_client); 2506 destroy_workqueue(ipoib_workqueue); 2507 2508 err_fs: 2509 ipoib_unregister_debugfs(); 2510 2511 return ret; 2512 } 2513 2514 static void __exit ipoib_cleanup_module(void) 2515 { 2516 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 2517 unregister_netdevice_notifier(&ipoib_netdev_notifier); 2518 #endif 2519 ipoib_netlink_fini(); 2520 ib_unregister_client(&ipoib_client); 2521 ib_sa_unregister_client(&ipoib_sa_client); 2522 ipoib_unregister_debugfs(); 2523 destroy_workqueue(ipoib_workqueue); 2524 } 2525 2526 module_init(ipoib_init_module); 2527 module_exit(ipoib_cleanup_module); 2528