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 #include <linux/pci.h> 55 56 #define DRV_VERSION "1.0.0" 57 58 const char ipoib_driver_version[] = DRV_VERSION; 59 60 MODULE_AUTHOR("Roland Dreier"); 61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver"); 62 MODULE_LICENSE("Dual BSD/GPL"); 63 MODULE_VERSION(DRV_VERSION); 64 65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE; 66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE; 67 68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444); 69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue"); 70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444); 71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue"); 72 73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 74 int ipoib_debug_level; 75 76 module_param_named(debug_level, ipoib_debug_level, int, 0644); 77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0"); 78 #endif 79 80 struct ipoib_path_iter { 81 struct net_device *dev; 82 struct ipoib_path path; 83 }; 84 85 static const u8 ipv4_bcast_addr[] = { 86 0x00, 0xff, 0xff, 0xff, 87 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00, 88 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff 89 }; 90 91 struct workqueue_struct *ipoib_workqueue; 92 93 struct ib_sa_client ipoib_sa_client; 94 95 static void ipoib_add_one(struct ib_device *device); 96 static void ipoib_remove_one(struct ib_device *device, void *client_data); 97 static void ipoib_neigh_reclaim(struct rcu_head *rp); 98 static struct net_device *ipoib_get_net_dev_by_params( 99 struct ib_device *dev, u8 port, u16 pkey, 100 const union ib_gid *gid, const struct sockaddr *addr, 101 void *client_data); 102 static int ipoib_set_mac(struct net_device *dev, void *addr); 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 int ipoib_open(struct net_device *dev) 112 { 113 struct ipoib_dev_priv *priv = netdev_priv(dev); 114 115 ipoib_dbg(priv, "bringing up interface\n"); 116 117 netif_carrier_off(dev); 118 119 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 120 121 priv->sm_fullmember_sendonly_support = false; 122 123 if (ipoib_ib_dev_open(dev)) { 124 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) 125 return 0; 126 goto err_disable; 127 } 128 129 if (ipoib_ib_dev_up(dev)) 130 goto err_stop; 131 132 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 133 struct ipoib_dev_priv *cpriv; 134 135 /* Bring up any child interfaces too */ 136 down_read(&priv->vlan_rwsem); 137 list_for_each_entry(cpriv, &priv->child_intfs, list) { 138 int flags; 139 140 flags = cpriv->dev->flags; 141 if (flags & IFF_UP) 142 continue; 143 144 dev_change_flags(cpriv->dev, flags | IFF_UP); 145 } 146 up_read(&priv->vlan_rwsem); 147 } 148 149 netif_start_queue(dev); 150 151 return 0; 152 153 err_stop: 154 ipoib_ib_dev_stop(dev); 155 156 err_disable: 157 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 158 159 return -EINVAL; 160 } 161 162 static int ipoib_stop(struct net_device *dev) 163 { 164 struct ipoib_dev_priv *priv = netdev_priv(dev); 165 166 ipoib_dbg(priv, "stopping interface\n"); 167 168 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags); 169 170 netif_stop_queue(dev); 171 172 ipoib_ib_dev_down(dev); 173 ipoib_ib_dev_stop(dev); 174 175 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 176 struct ipoib_dev_priv *cpriv; 177 178 /* Bring down any child interfaces too */ 179 down_read(&priv->vlan_rwsem); 180 list_for_each_entry(cpriv, &priv->child_intfs, list) { 181 int flags; 182 183 flags = cpriv->dev->flags; 184 if (!(flags & IFF_UP)) 185 continue; 186 187 dev_change_flags(cpriv->dev, flags & ~IFF_UP); 188 } 189 up_read(&priv->vlan_rwsem); 190 } 191 192 return 0; 193 } 194 195 static void ipoib_uninit(struct net_device *dev) 196 { 197 ipoib_dev_cleanup(dev); 198 } 199 200 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features) 201 { 202 struct ipoib_dev_priv *priv = netdev_priv(dev); 203 204 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags)) 205 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO); 206 207 return features; 208 } 209 210 static int ipoib_change_mtu(struct net_device *dev, int new_mtu) 211 { 212 struct ipoib_dev_priv *priv = netdev_priv(dev); 213 214 /* dev->mtu > 2K ==> connected mode */ 215 if (ipoib_cm_admin_enabled(dev)) { 216 if (new_mtu > ipoib_cm_max_mtu(dev)) 217 return -EINVAL; 218 219 if (new_mtu > priv->mcast_mtu) 220 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n", 221 priv->mcast_mtu); 222 223 dev->mtu = new_mtu; 224 return 0; 225 } 226 227 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu)) 228 return -EINVAL; 229 230 priv->admin_mtu = new_mtu; 231 232 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu); 233 234 return 0; 235 } 236 237 /* Called with an RCU read lock taken */ 238 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr, 239 struct net_device *dev) 240 { 241 struct net *net = dev_net(dev); 242 struct in_device *in_dev; 243 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr; 244 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr; 245 __be32 ret_addr; 246 247 switch (addr->sa_family) { 248 case AF_INET: 249 in_dev = in_dev_get(dev); 250 if (!in_dev) 251 return false; 252 253 ret_addr = inet_confirm_addr(net, in_dev, 0, 254 addr_in->sin_addr.s_addr, 255 RT_SCOPE_HOST); 256 in_dev_put(in_dev); 257 if (ret_addr) 258 return true; 259 260 break; 261 case AF_INET6: 262 if (IS_ENABLED(CONFIG_IPV6) && 263 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1)) 264 return true; 265 266 break; 267 } 268 return false; 269 } 270 271 /** 272 * Find the master net_device on top of the given net_device. 273 * @dev: base IPoIB net_device 274 * 275 * Returns the master net_device with a reference held, or the same net_device 276 * if no master exists. 277 */ 278 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev) 279 { 280 struct net_device *master; 281 282 rcu_read_lock(); 283 master = netdev_master_upper_dev_get_rcu(dev); 284 if (master) 285 dev_hold(master); 286 rcu_read_unlock(); 287 288 if (master) 289 return master; 290 291 dev_hold(dev); 292 return dev; 293 } 294 295 struct ipoib_walk_data { 296 const struct sockaddr *addr; 297 struct net_device *result; 298 }; 299 300 static int ipoib_upper_walk(struct net_device *upper, void *_data) 301 { 302 struct ipoib_walk_data *data = _data; 303 int ret = 0; 304 305 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) { 306 dev_hold(upper); 307 data->result = upper; 308 ret = 1; 309 } 310 311 return ret; 312 } 313 314 /** 315 * Find a net_device matching the given address, which is an upper device of 316 * the given net_device. 317 * @addr: IP address to look for. 318 * @dev: base IPoIB net_device 319 * 320 * If found, returns the net_device with a reference held. Otherwise return 321 * NULL. 322 */ 323 static struct net_device *ipoib_get_net_dev_match_addr( 324 const struct sockaddr *addr, struct net_device *dev) 325 { 326 struct ipoib_walk_data data = { 327 .addr = addr, 328 }; 329 330 rcu_read_lock(); 331 if (ipoib_is_dev_match_addr_rcu(addr, dev)) { 332 dev_hold(dev); 333 data.result = dev; 334 goto out; 335 } 336 337 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data); 338 out: 339 rcu_read_unlock(); 340 return data.result; 341 } 342 343 /* returns the number of IPoIB netdevs on top a given ipoib device matching a 344 * pkey_index and address, if one exists. 345 * 346 * @found_net_dev: contains a matching net_device if the return value >= 1, 347 * with a reference held. */ 348 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv, 349 const union ib_gid *gid, 350 u16 pkey_index, 351 const struct sockaddr *addr, 352 int nesting, 353 struct net_device **found_net_dev) 354 { 355 struct ipoib_dev_priv *child_priv; 356 struct net_device *net_dev = NULL; 357 int matches = 0; 358 359 if (priv->pkey_index == pkey_index && 360 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) { 361 if (!addr) { 362 net_dev = ipoib_get_master_net_dev(priv->dev); 363 } else { 364 /* Verify the net_device matches the IP address, as 365 * IPoIB child devices currently share a GID. */ 366 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev); 367 } 368 if (net_dev) { 369 if (!*found_net_dev) 370 *found_net_dev = net_dev; 371 else 372 dev_put(net_dev); 373 ++matches; 374 } 375 } 376 377 /* Check child interfaces */ 378 down_read_nested(&priv->vlan_rwsem, nesting); 379 list_for_each_entry(child_priv, &priv->child_intfs, list) { 380 matches += ipoib_match_gid_pkey_addr(child_priv, gid, 381 pkey_index, addr, 382 nesting + 1, 383 found_net_dev); 384 if (matches > 1) 385 break; 386 } 387 up_read(&priv->vlan_rwsem); 388 389 return matches; 390 } 391 392 /* Returns the number of matching net_devs found (between 0 and 2). Also 393 * return the matching net_device in the @net_dev parameter, holding a 394 * reference to the net_device, if the number of matches >= 1 */ 395 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port, 396 u16 pkey_index, 397 const union ib_gid *gid, 398 const struct sockaddr *addr, 399 struct net_device **net_dev) 400 { 401 struct ipoib_dev_priv *priv; 402 int matches = 0; 403 404 *net_dev = NULL; 405 406 list_for_each_entry(priv, dev_list, list) { 407 if (priv->port != port) 408 continue; 409 410 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index, 411 addr, 0, net_dev); 412 if (matches > 1) 413 break; 414 } 415 416 return matches; 417 } 418 419 static struct net_device *ipoib_get_net_dev_by_params( 420 struct ib_device *dev, u8 port, u16 pkey, 421 const union ib_gid *gid, const struct sockaddr *addr, 422 void *client_data) 423 { 424 struct net_device *net_dev; 425 struct list_head *dev_list = client_data; 426 u16 pkey_index; 427 int matches; 428 int ret; 429 430 if (!rdma_protocol_ib(dev, port)) 431 return NULL; 432 433 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index); 434 if (ret) 435 return NULL; 436 437 if (!dev_list) 438 return NULL; 439 440 /* See if we can find a unique device matching the L2 parameters */ 441 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 442 gid, NULL, &net_dev); 443 444 switch (matches) { 445 case 0: 446 return NULL; 447 case 1: 448 return net_dev; 449 } 450 451 dev_put(net_dev); 452 453 /* Couldn't find a unique device with L2 parameters only. Use L3 454 * address to uniquely match the net device */ 455 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 456 gid, addr, &net_dev); 457 switch (matches) { 458 case 0: 459 return NULL; 460 default: 461 dev_warn_ratelimited(&dev->dev, 462 "duplicate IP address detected\n"); 463 /* Fall through */ 464 case 1: 465 return net_dev; 466 } 467 } 468 469 int ipoib_set_mode(struct net_device *dev, const char *buf) 470 { 471 struct ipoib_dev_priv *priv = netdev_priv(dev); 472 473 /* flush paths if we switch modes so that connections are restarted */ 474 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) { 475 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 476 ipoib_warn(priv, "enabling connected mode " 477 "will cause multicast packet drops\n"); 478 netdev_update_features(dev); 479 dev_set_mtu(dev, ipoib_cm_max_mtu(dev)); 480 rtnl_unlock(); 481 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM; 482 483 ipoib_flush_paths(dev); 484 rtnl_lock(); 485 return 0; 486 } 487 488 if (!strcmp(buf, "datagram\n")) { 489 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 490 netdev_update_features(dev); 491 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu)); 492 rtnl_unlock(); 493 ipoib_flush_paths(dev); 494 rtnl_lock(); 495 return 0; 496 } 497 498 return -EINVAL; 499 } 500 501 struct ipoib_path *__path_find(struct net_device *dev, void *gid) 502 { 503 struct ipoib_dev_priv *priv = netdev_priv(dev); 504 struct rb_node *n = priv->path_tree.rb_node; 505 struct ipoib_path *path; 506 int ret; 507 508 while (n) { 509 path = rb_entry(n, struct ipoib_path, rb_node); 510 511 ret = memcmp(gid, path->pathrec.dgid.raw, 512 sizeof (union ib_gid)); 513 514 if (ret < 0) 515 n = n->rb_left; 516 else if (ret > 0) 517 n = n->rb_right; 518 else 519 return path; 520 } 521 522 return NULL; 523 } 524 525 static int __path_add(struct net_device *dev, struct ipoib_path *path) 526 { 527 struct ipoib_dev_priv *priv = netdev_priv(dev); 528 struct rb_node **n = &priv->path_tree.rb_node; 529 struct rb_node *pn = NULL; 530 struct ipoib_path *tpath; 531 int ret; 532 533 while (*n) { 534 pn = *n; 535 tpath = rb_entry(pn, struct ipoib_path, rb_node); 536 537 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw, 538 sizeof (union ib_gid)); 539 if (ret < 0) 540 n = &pn->rb_left; 541 else if (ret > 0) 542 n = &pn->rb_right; 543 else 544 return -EEXIST; 545 } 546 547 rb_link_node(&path->rb_node, pn, n); 548 rb_insert_color(&path->rb_node, &priv->path_tree); 549 550 list_add_tail(&path->list, &priv->path_list); 551 552 return 0; 553 } 554 555 static void path_free(struct net_device *dev, struct ipoib_path *path) 556 { 557 struct sk_buff *skb; 558 559 while ((skb = __skb_dequeue(&path->queue))) 560 dev_kfree_skb_irq(skb); 561 562 ipoib_dbg(netdev_priv(dev), "path_free\n"); 563 564 /* remove all neigh connected to this path */ 565 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 566 567 if (path->ah) 568 ipoib_put_ah(path->ah); 569 570 kfree(path); 571 } 572 573 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 574 575 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev) 576 { 577 struct ipoib_path_iter *iter; 578 579 iter = kmalloc(sizeof *iter, GFP_KERNEL); 580 if (!iter) 581 return NULL; 582 583 iter->dev = dev; 584 memset(iter->path.pathrec.dgid.raw, 0, 16); 585 586 if (ipoib_path_iter_next(iter)) { 587 kfree(iter); 588 return NULL; 589 } 590 591 return iter; 592 } 593 594 int ipoib_path_iter_next(struct ipoib_path_iter *iter) 595 { 596 struct ipoib_dev_priv *priv = netdev_priv(iter->dev); 597 struct rb_node *n; 598 struct ipoib_path *path; 599 int ret = 1; 600 601 spin_lock_irq(&priv->lock); 602 603 n = rb_first(&priv->path_tree); 604 605 while (n) { 606 path = rb_entry(n, struct ipoib_path, rb_node); 607 608 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw, 609 sizeof (union ib_gid)) < 0) { 610 iter->path = *path; 611 ret = 0; 612 break; 613 } 614 615 n = rb_next(n); 616 } 617 618 spin_unlock_irq(&priv->lock); 619 620 return ret; 621 } 622 623 void ipoib_path_iter_read(struct ipoib_path_iter *iter, 624 struct ipoib_path *path) 625 { 626 *path = iter->path; 627 } 628 629 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */ 630 631 void ipoib_mark_paths_invalid(struct net_device *dev) 632 { 633 struct ipoib_dev_priv *priv = netdev_priv(dev); 634 struct ipoib_path *path, *tp; 635 636 spin_lock_irq(&priv->lock); 637 638 list_for_each_entry_safe(path, tp, &priv->path_list, list) { 639 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n", 640 be16_to_cpu(path->pathrec.dlid), 641 path->pathrec.dgid.raw); 642 path->valid = 0; 643 } 644 645 spin_unlock_irq(&priv->lock); 646 } 647 648 struct classport_info_context { 649 struct ipoib_dev_priv *priv; 650 struct completion done; 651 struct ib_sa_query *sa_query; 652 }; 653 654 static void classport_info_query_cb(int status, struct ib_class_port_info *rec, 655 void *context) 656 { 657 struct classport_info_context *cb_ctx = context; 658 struct ipoib_dev_priv *priv; 659 660 WARN_ON(!context); 661 662 priv = cb_ctx->priv; 663 664 if (status || !rec) { 665 pr_debug("device: %s failed query classport_info status: %d\n", 666 priv->dev->name, status); 667 /* keeps the default, will try next mcast_restart */ 668 priv->sm_fullmember_sendonly_support = false; 669 goto out; 670 } 671 672 if (ib_get_cpi_capmask2(rec) & 673 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) { 674 pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n", 675 priv->dev->name); 676 priv->sm_fullmember_sendonly_support = true; 677 } else { 678 pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n", 679 priv->dev->name); 680 priv->sm_fullmember_sendonly_support = false; 681 } 682 683 out: 684 complete(&cb_ctx->done); 685 } 686 687 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv) 688 { 689 struct classport_info_context *callback_context; 690 int ret; 691 692 callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL); 693 if (!callback_context) 694 return -ENOMEM; 695 696 callback_context->priv = priv; 697 init_completion(&callback_context->done); 698 699 ret = ib_sa_classport_info_rec_query(&ipoib_sa_client, 700 priv->ca, priv->port, 3000, 701 GFP_KERNEL, 702 classport_info_query_cb, 703 callback_context, 704 &callback_context->sa_query); 705 if (ret < 0) { 706 pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n", 707 priv->dev->name, ret); 708 kfree(callback_context); 709 return ret; 710 } 711 712 /* waiting for the callback to finish before returnning */ 713 wait_for_completion(&callback_context->done); 714 kfree(callback_context); 715 716 return ret; 717 } 718 719 void ipoib_flush_paths(struct net_device *dev) 720 { 721 struct ipoib_dev_priv *priv = netdev_priv(dev); 722 struct ipoib_path *path, *tp; 723 LIST_HEAD(remove_list); 724 unsigned long flags; 725 726 netif_tx_lock_bh(dev); 727 spin_lock_irqsave(&priv->lock, flags); 728 729 list_splice_init(&priv->path_list, &remove_list); 730 731 list_for_each_entry(path, &remove_list, list) 732 rb_erase(&path->rb_node, &priv->path_tree); 733 734 list_for_each_entry_safe(path, tp, &remove_list, list) { 735 if (path->query) 736 ib_sa_cancel_query(path->query_id, path->query); 737 spin_unlock_irqrestore(&priv->lock, flags); 738 netif_tx_unlock_bh(dev); 739 wait_for_completion(&path->done); 740 path_free(dev, path); 741 netif_tx_lock_bh(dev); 742 spin_lock_irqsave(&priv->lock, flags); 743 } 744 745 spin_unlock_irqrestore(&priv->lock, flags); 746 netif_tx_unlock_bh(dev); 747 } 748 749 static void path_rec_completion(int status, 750 struct ib_sa_path_rec *pathrec, 751 void *path_ptr) 752 { 753 struct ipoib_path *path = path_ptr; 754 struct net_device *dev = path->dev; 755 struct ipoib_dev_priv *priv = netdev_priv(dev); 756 struct ipoib_ah *ah = NULL; 757 struct ipoib_ah *old_ah = NULL; 758 struct ipoib_neigh *neigh, *tn; 759 struct sk_buff_head skqueue; 760 struct sk_buff *skb; 761 unsigned long flags; 762 763 if (!status) 764 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n", 765 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw); 766 else 767 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n", 768 status, path->pathrec.dgid.raw); 769 770 skb_queue_head_init(&skqueue); 771 772 if (!status) { 773 struct ib_ah_attr av; 774 775 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av)) 776 ah = ipoib_create_ah(dev, priv->pd, &av); 777 } 778 779 spin_lock_irqsave(&priv->lock, flags); 780 781 if (!IS_ERR_OR_NULL(ah)) { 782 path->pathrec = *pathrec; 783 784 old_ah = path->ah; 785 path->ah = ah; 786 787 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n", 788 ah, be16_to_cpu(pathrec->dlid), pathrec->sl); 789 790 while ((skb = __skb_dequeue(&path->queue))) 791 __skb_queue_tail(&skqueue, skb); 792 793 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) { 794 if (neigh->ah) { 795 WARN_ON(neigh->ah != old_ah); 796 /* 797 * Dropping the ah reference inside 798 * priv->lock is safe here, because we 799 * will hold one more reference from 800 * the original value of path->ah (ie 801 * old_ah). 802 */ 803 ipoib_put_ah(neigh->ah); 804 } 805 kref_get(&path->ah->ref); 806 neigh->ah = path->ah; 807 808 if (ipoib_cm_enabled(dev, neigh->daddr)) { 809 if (!ipoib_cm_get(neigh)) 810 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, 811 path, 812 neigh)); 813 if (!ipoib_cm_get(neigh)) { 814 ipoib_neigh_free(neigh); 815 continue; 816 } 817 } 818 819 while ((skb = __skb_dequeue(&neigh->queue))) 820 __skb_queue_tail(&skqueue, skb); 821 } 822 path->valid = 1; 823 } 824 825 path->query = NULL; 826 complete(&path->done); 827 828 spin_unlock_irqrestore(&priv->lock, flags); 829 830 if (IS_ERR_OR_NULL(ah)) 831 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 832 833 if (old_ah) 834 ipoib_put_ah(old_ah); 835 836 while ((skb = __skb_dequeue(&skqueue))) { 837 skb->dev = dev; 838 if (dev_queue_xmit(skb)) 839 ipoib_warn(priv, "dev_queue_xmit failed " 840 "to requeue packet\n"); 841 } 842 } 843 844 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid) 845 { 846 struct ipoib_dev_priv *priv = netdev_priv(dev); 847 struct ipoib_path *path; 848 849 if (!priv->broadcast) 850 return NULL; 851 852 path = kzalloc(sizeof *path, GFP_ATOMIC); 853 if (!path) 854 return NULL; 855 856 path->dev = dev; 857 858 skb_queue_head_init(&path->queue); 859 860 INIT_LIST_HEAD(&path->neigh_list); 861 862 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid)); 863 path->pathrec.sgid = priv->local_gid; 864 path->pathrec.pkey = cpu_to_be16(priv->pkey); 865 path->pathrec.numb_path = 1; 866 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class; 867 868 return path; 869 } 870 871 static int path_rec_start(struct net_device *dev, 872 struct ipoib_path *path) 873 { 874 struct ipoib_dev_priv *priv = netdev_priv(dev); 875 876 ipoib_dbg(priv, "Start path record lookup for %pI6\n", 877 path->pathrec.dgid.raw); 878 879 init_completion(&path->done); 880 881 path->query_id = 882 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port, 883 &path->pathrec, 884 IB_SA_PATH_REC_DGID | 885 IB_SA_PATH_REC_SGID | 886 IB_SA_PATH_REC_NUMB_PATH | 887 IB_SA_PATH_REC_TRAFFIC_CLASS | 888 IB_SA_PATH_REC_PKEY, 889 1000, GFP_ATOMIC, 890 path_rec_completion, 891 path, &path->query); 892 if (path->query_id < 0) { 893 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id); 894 path->query = NULL; 895 complete(&path->done); 896 return path->query_id; 897 } 898 899 return 0; 900 } 901 902 static void neigh_add_path(struct sk_buff *skb, u8 *daddr, 903 struct net_device *dev) 904 { 905 struct ipoib_dev_priv *priv = netdev_priv(dev); 906 struct ipoib_path *path; 907 struct ipoib_neigh *neigh; 908 unsigned long flags; 909 910 spin_lock_irqsave(&priv->lock, flags); 911 neigh = ipoib_neigh_alloc(daddr, dev); 912 if (!neigh) { 913 spin_unlock_irqrestore(&priv->lock, flags); 914 ++dev->stats.tx_dropped; 915 dev_kfree_skb_any(skb); 916 return; 917 } 918 919 path = __path_find(dev, daddr + 4); 920 if (!path) { 921 path = path_rec_create(dev, daddr + 4); 922 if (!path) 923 goto err_path; 924 925 __path_add(dev, path); 926 } 927 928 list_add_tail(&neigh->list, &path->neigh_list); 929 930 if (path->ah) { 931 kref_get(&path->ah->ref); 932 neigh->ah = path->ah; 933 934 if (ipoib_cm_enabled(dev, neigh->daddr)) { 935 if (!ipoib_cm_get(neigh)) 936 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh)); 937 if (!ipoib_cm_get(neigh)) { 938 ipoib_neigh_free(neigh); 939 goto err_drop; 940 } 941 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) 942 __skb_queue_tail(&neigh->queue, skb); 943 else { 944 ipoib_warn(priv, "queue length limit %d. Packet drop.\n", 945 skb_queue_len(&neigh->queue)); 946 goto err_drop; 947 } 948 } else { 949 spin_unlock_irqrestore(&priv->lock, flags); 950 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr)); 951 ipoib_neigh_put(neigh); 952 return; 953 } 954 } else { 955 neigh->ah = NULL; 956 957 if (!path->query && path_rec_start(dev, path)) 958 goto err_path; 959 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) 960 __skb_queue_tail(&neigh->queue, skb); 961 else 962 goto err_drop; 963 } 964 965 spin_unlock_irqrestore(&priv->lock, flags); 966 ipoib_neigh_put(neigh); 967 return; 968 969 err_path: 970 ipoib_neigh_free(neigh); 971 err_drop: 972 ++dev->stats.tx_dropped; 973 dev_kfree_skb_any(skb); 974 975 spin_unlock_irqrestore(&priv->lock, flags); 976 ipoib_neigh_put(neigh); 977 } 978 979 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev, 980 struct ipoib_cb *cb) 981 { 982 struct ipoib_dev_priv *priv = netdev_priv(dev); 983 struct ipoib_path *path; 984 unsigned long flags; 985 986 spin_lock_irqsave(&priv->lock, flags); 987 988 path = __path_find(dev, cb->hwaddr + 4); 989 if (!path || !path->valid) { 990 int new_path = 0; 991 992 if (!path) { 993 path = path_rec_create(dev, cb->hwaddr + 4); 994 new_path = 1; 995 } 996 if (path) { 997 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 998 __skb_queue_tail(&path->queue, skb); 999 } else { 1000 ++dev->stats.tx_dropped; 1001 dev_kfree_skb_any(skb); 1002 } 1003 1004 if (!path->query && path_rec_start(dev, path)) { 1005 spin_unlock_irqrestore(&priv->lock, flags); 1006 if (new_path) 1007 path_free(dev, path); 1008 return; 1009 } else 1010 __path_add(dev, path); 1011 } else { 1012 ++dev->stats.tx_dropped; 1013 dev_kfree_skb_any(skb); 1014 } 1015 1016 spin_unlock_irqrestore(&priv->lock, flags); 1017 return; 1018 } 1019 1020 if (path->ah) { 1021 ipoib_dbg(priv, "Send unicast ARP to %04x\n", 1022 be16_to_cpu(path->pathrec.dlid)); 1023 1024 spin_unlock_irqrestore(&priv->lock, flags); 1025 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr)); 1026 return; 1027 } else if ((path->query || !path_rec_start(dev, path)) && 1028 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1029 __skb_queue_tail(&path->queue, skb); 1030 } else { 1031 ++dev->stats.tx_dropped; 1032 dev_kfree_skb_any(skb); 1033 } 1034 1035 spin_unlock_irqrestore(&priv->lock, flags); 1036 } 1037 1038 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev) 1039 { 1040 struct ipoib_dev_priv *priv = netdev_priv(dev); 1041 struct ipoib_neigh *neigh; 1042 struct ipoib_cb *cb = ipoib_skb_cb(skb); 1043 struct ipoib_header *header; 1044 unsigned long flags; 1045 1046 header = (struct ipoib_header *) skb->data; 1047 1048 if (unlikely(cb->hwaddr[4] == 0xff)) { 1049 /* multicast, arrange "if" according to probability */ 1050 if ((header->proto != htons(ETH_P_IP)) && 1051 (header->proto != htons(ETH_P_IPV6)) && 1052 (header->proto != htons(ETH_P_ARP)) && 1053 (header->proto != htons(ETH_P_RARP)) && 1054 (header->proto != htons(ETH_P_TIPC))) { 1055 /* ethertype not supported by IPoIB */ 1056 ++dev->stats.tx_dropped; 1057 dev_kfree_skb_any(skb); 1058 return NETDEV_TX_OK; 1059 } 1060 /* Add in the P_Key for multicast*/ 1061 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff; 1062 cb->hwaddr[9] = priv->pkey & 0xff; 1063 1064 neigh = ipoib_neigh_get(dev, cb->hwaddr); 1065 if (likely(neigh)) 1066 goto send_using_neigh; 1067 ipoib_mcast_send(dev, cb->hwaddr, skb); 1068 return NETDEV_TX_OK; 1069 } 1070 1071 /* unicast, arrange "switch" according to probability */ 1072 switch (header->proto) { 1073 case htons(ETH_P_IP): 1074 case htons(ETH_P_IPV6): 1075 case htons(ETH_P_TIPC): 1076 neigh = ipoib_neigh_get(dev, cb->hwaddr); 1077 if (unlikely(!neigh)) { 1078 neigh_add_path(skb, cb->hwaddr, dev); 1079 return NETDEV_TX_OK; 1080 } 1081 break; 1082 case htons(ETH_P_ARP): 1083 case htons(ETH_P_RARP): 1084 /* for unicast ARP and RARP should always perform path find */ 1085 unicast_arp_send(skb, dev, cb); 1086 return NETDEV_TX_OK; 1087 default: 1088 /* ethertype not supported by IPoIB */ 1089 ++dev->stats.tx_dropped; 1090 dev_kfree_skb_any(skb); 1091 return NETDEV_TX_OK; 1092 } 1093 1094 send_using_neigh: 1095 /* note we now hold a ref to neigh */ 1096 if (ipoib_cm_get(neigh)) { 1097 if (ipoib_cm_up(neigh)) { 1098 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh)); 1099 goto unref; 1100 } 1101 } else if (neigh->ah) { 1102 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr)); 1103 goto unref; 1104 } 1105 1106 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1107 spin_lock_irqsave(&priv->lock, flags); 1108 __skb_queue_tail(&neigh->queue, skb); 1109 spin_unlock_irqrestore(&priv->lock, flags); 1110 } else { 1111 ++dev->stats.tx_dropped; 1112 dev_kfree_skb_any(skb); 1113 } 1114 1115 unref: 1116 ipoib_neigh_put(neigh); 1117 1118 return NETDEV_TX_OK; 1119 } 1120 1121 static void ipoib_timeout(struct net_device *dev) 1122 { 1123 struct ipoib_dev_priv *priv = netdev_priv(dev); 1124 1125 ipoib_warn(priv, "transmit timeout: latency %d msecs\n", 1126 jiffies_to_msecs(jiffies - dev_trans_start(dev))); 1127 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n", 1128 netif_queue_stopped(dev), 1129 priv->tx_head, priv->tx_tail); 1130 /* XXX reset QP, etc. */ 1131 } 1132 1133 static int ipoib_hard_header(struct sk_buff *skb, 1134 struct net_device *dev, 1135 unsigned short type, 1136 const void *daddr, const void *saddr, unsigned len) 1137 { 1138 struct ipoib_header *header; 1139 struct ipoib_cb *cb = ipoib_skb_cb(skb); 1140 1141 header = (struct ipoib_header *) skb_push(skb, sizeof *header); 1142 1143 header->proto = htons(type); 1144 header->reserved = 0; 1145 1146 /* 1147 * we don't rely on dst_entry structure, always stuff the 1148 * destination address into skb->cb so we can figure out where 1149 * to send the packet later. 1150 */ 1151 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN); 1152 1153 return sizeof *header; 1154 } 1155 1156 static void ipoib_set_mcast_list(struct net_device *dev) 1157 { 1158 struct ipoib_dev_priv *priv = netdev_priv(dev); 1159 1160 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) { 1161 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set"); 1162 return; 1163 } 1164 1165 queue_work(priv->wq, &priv->restart_task); 1166 } 1167 1168 static int ipoib_get_iflink(const struct net_device *dev) 1169 { 1170 struct ipoib_dev_priv *priv = netdev_priv(dev); 1171 1172 /* parent interface */ 1173 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) 1174 return dev->ifindex; 1175 1176 /* child/vlan interface */ 1177 return priv->parent->ifindex; 1178 } 1179 1180 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr) 1181 { 1182 /* 1183 * Use only the address parts that contributes to spreading 1184 * The subnet prefix is not used as one can not connect to 1185 * same remote port (GUID) using the same remote QPN via two 1186 * different subnets. 1187 */ 1188 /* qpn octets[1:4) & port GUID octets[12:20) */ 1189 u32 *d32 = (u32 *) daddr; 1190 u32 hv; 1191 1192 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0); 1193 return hv & htbl->mask; 1194 } 1195 1196 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr) 1197 { 1198 struct ipoib_dev_priv *priv = netdev_priv(dev); 1199 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1200 struct ipoib_neigh_hash *htbl; 1201 struct ipoib_neigh *neigh = NULL; 1202 u32 hash_val; 1203 1204 rcu_read_lock_bh(); 1205 1206 htbl = rcu_dereference_bh(ntbl->htbl); 1207 1208 if (!htbl) 1209 goto out_unlock; 1210 1211 hash_val = ipoib_addr_hash(htbl, daddr); 1212 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]); 1213 neigh != NULL; 1214 neigh = rcu_dereference_bh(neigh->hnext)) { 1215 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1216 /* found, take one ref on behalf of the caller */ 1217 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1218 /* deleted */ 1219 neigh = NULL; 1220 goto out_unlock; 1221 } 1222 1223 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)) 1224 neigh->alive = jiffies; 1225 goto out_unlock; 1226 } 1227 } 1228 1229 out_unlock: 1230 rcu_read_unlock_bh(); 1231 return neigh; 1232 } 1233 1234 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv) 1235 { 1236 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1237 struct ipoib_neigh_hash *htbl; 1238 unsigned long neigh_obsolete; 1239 unsigned long dt; 1240 unsigned long flags; 1241 int i; 1242 LIST_HEAD(remove_list); 1243 1244 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1245 return; 1246 1247 spin_lock_irqsave(&priv->lock, flags); 1248 1249 htbl = rcu_dereference_protected(ntbl->htbl, 1250 lockdep_is_held(&priv->lock)); 1251 1252 if (!htbl) 1253 goto out_unlock; 1254 1255 /* neigh is obsolete if it was idle for two GC periods */ 1256 dt = 2 * arp_tbl.gc_interval; 1257 neigh_obsolete = jiffies - dt; 1258 /* handle possible race condition */ 1259 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1260 goto out_unlock; 1261 1262 for (i = 0; i < htbl->size; i++) { 1263 struct ipoib_neigh *neigh; 1264 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1265 1266 while ((neigh = rcu_dereference_protected(*np, 1267 lockdep_is_held(&priv->lock))) != NULL) { 1268 /* was the neigh idle for two GC periods */ 1269 if (time_after(neigh_obsolete, neigh->alive)) { 1270 1271 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list); 1272 1273 rcu_assign_pointer(*np, 1274 rcu_dereference_protected(neigh->hnext, 1275 lockdep_is_held(&priv->lock))); 1276 /* remove from path/mc list */ 1277 list_del(&neigh->list); 1278 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1279 } else { 1280 np = &neigh->hnext; 1281 } 1282 1283 } 1284 } 1285 1286 out_unlock: 1287 spin_unlock_irqrestore(&priv->lock, flags); 1288 ipoib_mcast_remove_list(&remove_list); 1289 } 1290 1291 static void ipoib_reap_neigh(struct work_struct *work) 1292 { 1293 struct ipoib_dev_priv *priv = 1294 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work); 1295 1296 __ipoib_reap_neigh(priv); 1297 1298 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1299 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1300 arp_tbl.gc_interval); 1301 } 1302 1303 1304 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr, 1305 struct net_device *dev) 1306 { 1307 struct ipoib_neigh *neigh; 1308 1309 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC); 1310 if (!neigh) 1311 return NULL; 1312 1313 neigh->dev = dev; 1314 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr)); 1315 skb_queue_head_init(&neigh->queue); 1316 INIT_LIST_HEAD(&neigh->list); 1317 ipoib_cm_set(neigh, NULL); 1318 /* one ref on behalf of the caller */ 1319 atomic_set(&neigh->refcnt, 1); 1320 1321 return neigh; 1322 } 1323 1324 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr, 1325 struct net_device *dev) 1326 { 1327 struct ipoib_dev_priv *priv = netdev_priv(dev); 1328 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1329 struct ipoib_neigh_hash *htbl; 1330 struct ipoib_neigh *neigh; 1331 u32 hash_val; 1332 1333 htbl = rcu_dereference_protected(ntbl->htbl, 1334 lockdep_is_held(&priv->lock)); 1335 if (!htbl) { 1336 neigh = NULL; 1337 goto out_unlock; 1338 } 1339 1340 /* need to add a new neigh, but maybe some other thread succeeded? 1341 * recalc hash, maybe hash resize took place so we do a search 1342 */ 1343 hash_val = ipoib_addr_hash(htbl, daddr); 1344 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val], 1345 lockdep_is_held(&priv->lock)); 1346 neigh != NULL; 1347 neigh = rcu_dereference_protected(neigh->hnext, 1348 lockdep_is_held(&priv->lock))) { 1349 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1350 /* found, take one ref on behalf of the caller */ 1351 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1352 /* deleted */ 1353 neigh = NULL; 1354 break; 1355 } 1356 neigh->alive = jiffies; 1357 goto out_unlock; 1358 } 1359 } 1360 1361 neigh = ipoib_neigh_ctor(daddr, dev); 1362 if (!neigh) 1363 goto out_unlock; 1364 1365 /* one ref on behalf of the hash table */ 1366 atomic_inc(&neigh->refcnt); 1367 neigh->alive = jiffies; 1368 /* put in hash */ 1369 rcu_assign_pointer(neigh->hnext, 1370 rcu_dereference_protected(htbl->buckets[hash_val], 1371 lockdep_is_held(&priv->lock))); 1372 rcu_assign_pointer(htbl->buckets[hash_val], neigh); 1373 atomic_inc(&ntbl->entries); 1374 1375 out_unlock: 1376 1377 return neigh; 1378 } 1379 1380 void ipoib_neigh_dtor(struct ipoib_neigh *neigh) 1381 { 1382 /* neigh reference count was dropprd to zero */ 1383 struct net_device *dev = neigh->dev; 1384 struct ipoib_dev_priv *priv = netdev_priv(dev); 1385 struct sk_buff *skb; 1386 if (neigh->ah) 1387 ipoib_put_ah(neigh->ah); 1388 while ((skb = __skb_dequeue(&neigh->queue))) { 1389 ++dev->stats.tx_dropped; 1390 dev_kfree_skb_any(skb); 1391 } 1392 if (ipoib_cm_get(neigh)) 1393 ipoib_cm_destroy_tx(ipoib_cm_get(neigh)); 1394 ipoib_dbg(netdev_priv(dev), 1395 "neigh free for %06x %pI6\n", 1396 IPOIB_QPN(neigh->daddr), 1397 neigh->daddr + 4); 1398 kfree(neigh); 1399 if (atomic_dec_and_test(&priv->ntbl.entries)) { 1400 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags)) 1401 complete(&priv->ntbl.flushed); 1402 } 1403 } 1404 1405 static void ipoib_neigh_reclaim(struct rcu_head *rp) 1406 { 1407 /* Called as a result of removal from hash table */ 1408 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu); 1409 /* note TX context may hold another ref */ 1410 ipoib_neigh_put(neigh); 1411 } 1412 1413 void ipoib_neigh_free(struct ipoib_neigh *neigh) 1414 { 1415 struct net_device *dev = neigh->dev; 1416 struct ipoib_dev_priv *priv = netdev_priv(dev); 1417 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1418 struct ipoib_neigh_hash *htbl; 1419 struct ipoib_neigh __rcu **np; 1420 struct ipoib_neigh *n; 1421 u32 hash_val; 1422 1423 htbl = rcu_dereference_protected(ntbl->htbl, 1424 lockdep_is_held(&priv->lock)); 1425 if (!htbl) 1426 return; 1427 1428 hash_val = ipoib_addr_hash(htbl, neigh->daddr); 1429 np = &htbl->buckets[hash_val]; 1430 for (n = rcu_dereference_protected(*np, 1431 lockdep_is_held(&priv->lock)); 1432 n != NULL; 1433 n = rcu_dereference_protected(*np, 1434 lockdep_is_held(&priv->lock))) { 1435 if (n == neigh) { 1436 /* found */ 1437 rcu_assign_pointer(*np, 1438 rcu_dereference_protected(neigh->hnext, 1439 lockdep_is_held(&priv->lock))); 1440 /* remove from parent list */ 1441 list_del(&neigh->list); 1442 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1443 return; 1444 } else { 1445 np = &n->hnext; 1446 } 1447 } 1448 } 1449 1450 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv) 1451 { 1452 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1453 struct ipoib_neigh_hash *htbl; 1454 struct ipoib_neigh __rcu **buckets; 1455 u32 size; 1456 1457 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1458 ntbl->htbl = NULL; 1459 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL); 1460 if (!htbl) 1461 return -ENOMEM; 1462 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1463 size = roundup_pow_of_two(arp_tbl.gc_thresh3); 1464 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL); 1465 if (!buckets) { 1466 kfree(htbl); 1467 return -ENOMEM; 1468 } 1469 htbl->size = size; 1470 htbl->mask = (size - 1); 1471 htbl->buckets = buckets; 1472 RCU_INIT_POINTER(ntbl->htbl, htbl); 1473 htbl->ntbl = ntbl; 1474 atomic_set(&ntbl->entries, 0); 1475 1476 /* start garbage collection */ 1477 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1478 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1479 arp_tbl.gc_interval); 1480 1481 return 0; 1482 } 1483 1484 static void neigh_hash_free_rcu(struct rcu_head *head) 1485 { 1486 struct ipoib_neigh_hash *htbl = container_of(head, 1487 struct ipoib_neigh_hash, 1488 rcu); 1489 struct ipoib_neigh __rcu **buckets = htbl->buckets; 1490 struct ipoib_neigh_table *ntbl = htbl->ntbl; 1491 1492 kfree(buckets); 1493 kfree(htbl); 1494 complete(&ntbl->deleted); 1495 } 1496 1497 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid) 1498 { 1499 struct ipoib_dev_priv *priv = netdev_priv(dev); 1500 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1501 struct ipoib_neigh_hash *htbl; 1502 unsigned long flags; 1503 int i; 1504 1505 /* remove all neigh connected to a given path or mcast */ 1506 spin_lock_irqsave(&priv->lock, flags); 1507 1508 htbl = rcu_dereference_protected(ntbl->htbl, 1509 lockdep_is_held(&priv->lock)); 1510 1511 if (!htbl) 1512 goto out_unlock; 1513 1514 for (i = 0; i < htbl->size; i++) { 1515 struct ipoib_neigh *neigh; 1516 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1517 1518 while ((neigh = rcu_dereference_protected(*np, 1519 lockdep_is_held(&priv->lock))) != NULL) { 1520 /* delete neighs belong to this parent */ 1521 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) { 1522 rcu_assign_pointer(*np, 1523 rcu_dereference_protected(neigh->hnext, 1524 lockdep_is_held(&priv->lock))); 1525 /* remove from parent list */ 1526 list_del(&neigh->list); 1527 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1528 } else { 1529 np = &neigh->hnext; 1530 } 1531 1532 } 1533 } 1534 out_unlock: 1535 spin_unlock_irqrestore(&priv->lock, flags); 1536 } 1537 1538 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv) 1539 { 1540 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1541 struct ipoib_neigh_hash *htbl; 1542 unsigned long flags; 1543 int i, wait_flushed = 0; 1544 1545 init_completion(&priv->ntbl.flushed); 1546 1547 spin_lock_irqsave(&priv->lock, flags); 1548 1549 htbl = rcu_dereference_protected(ntbl->htbl, 1550 lockdep_is_held(&priv->lock)); 1551 if (!htbl) 1552 goto out_unlock; 1553 1554 wait_flushed = atomic_read(&priv->ntbl.entries); 1555 if (!wait_flushed) 1556 goto free_htbl; 1557 1558 for (i = 0; i < htbl->size; i++) { 1559 struct ipoib_neigh *neigh; 1560 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1561 1562 while ((neigh = rcu_dereference_protected(*np, 1563 lockdep_is_held(&priv->lock))) != NULL) { 1564 rcu_assign_pointer(*np, 1565 rcu_dereference_protected(neigh->hnext, 1566 lockdep_is_held(&priv->lock))); 1567 /* remove from path/mc list */ 1568 list_del(&neigh->list); 1569 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1570 } 1571 } 1572 1573 free_htbl: 1574 rcu_assign_pointer(ntbl->htbl, NULL); 1575 call_rcu(&htbl->rcu, neigh_hash_free_rcu); 1576 1577 out_unlock: 1578 spin_unlock_irqrestore(&priv->lock, flags); 1579 if (wait_flushed) 1580 wait_for_completion(&priv->ntbl.flushed); 1581 } 1582 1583 static void ipoib_neigh_hash_uninit(struct net_device *dev) 1584 { 1585 struct ipoib_dev_priv *priv = netdev_priv(dev); 1586 int stopped; 1587 1588 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n"); 1589 init_completion(&priv->ntbl.deleted); 1590 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1591 1592 /* Stop GC if called at init fail need to cancel work */ 1593 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1594 if (!stopped) 1595 cancel_delayed_work(&priv->neigh_reap_task); 1596 1597 ipoib_flush_neighs(priv); 1598 1599 wait_for_completion(&priv->ntbl.deleted); 1600 } 1601 1602 1603 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port) 1604 { 1605 struct ipoib_dev_priv *priv = netdev_priv(dev); 1606 1607 /* Allocate RX/TX "rings" to hold queued skbs */ 1608 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring, 1609 GFP_KERNEL); 1610 if (!priv->rx_ring) { 1611 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n", 1612 ca->name, ipoib_recvq_size); 1613 goto out; 1614 } 1615 1616 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring); 1617 if (!priv->tx_ring) { 1618 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n", 1619 ca->name, ipoib_sendq_size); 1620 goto out_rx_ring_cleanup; 1621 } 1622 1623 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */ 1624 1625 if (ipoib_ib_dev_init(dev, ca, port)) 1626 goto out_tx_ring_cleanup; 1627 1628 /* 1629 * Must be after ipoib_ib_dev_init so we can allocate a per 1630 * device wq there and use it here 1631 */ 1632 if (ipoib_neigh_hash_init(priv) < 0) 1633 goto out_dev_uninit; 1634 1635 return 0; 1636 1637 out_dev_uninit: 1638 ipoib_ib_dev_cleanup(dev); 1639 1640 out_tx_ring_cleanup: 1641 vfree(priv->tx_ring); 1642 1643 out_rx_ring_cleanup: 1644 kfree(priv->rx_ring); 1645 1646 out: 1647 return -ENOMEM; 1648 } 1649 1650 void ipoib_dev_cleanup(struct net_device *dev) 1651 { 1652 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv; 1653 LIST_HEAD(head); 1654 1655 ASSERT_RTNL(); 1656 1657 ipoib_delete_debug_files(dev); 1658 1659 /* Delete any child interfaces first */ 1660 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) { 1661 /* Stop GC on child */ 1662 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags); 1663 cancel_delayed_work(&cpriv->neigh_reap_task); 1664 unregister_netdevice_queue(cpriv->dev, &head); 1665 } 1666 unregister_netdevice_many(&head); 1667 1668 /* 1669 * Must be before ipoib_ib_dev_cleanup or we delete an in use 1670 * work queue 1671 */ 1672 ipoib_neigh_hash_uninit(dev); 1673 1674 ipoib_ib_dev_cleanup(dev); 1675 1676 kfree(priv->rx_ring); 1677 vfree(priv->tx_ring); 1678 1679 priv->rx_ring = NULL; 1680 priv->tx_ring = NULL; 1681 } 1682 1683 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state) 1684 { 1685 struct ipoib_dev_priv *priv = netdev_priv(dev); 1686 1687 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state); 1688 } 1689 1690 static int ipoib_get_vf_config(struct net_device *dev, int vf, 1691 struct ifla_vf_info *ivf) 1692 { 1693 struct ipoib_dev_priv *priv = netdev_priv(dev); 1694 int err; 1695 1696 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf); 1697 if (err) 1698 return err; 1699 1700 ivf->vf = vf; 1701 1702 return 0; 1703 } 1704 1705 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type) 1706 { 1707 struct ipoib_dev_priv *priv = netdev_priv(dev); 1708 1709 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID) 1710 return -EINVAL; 1711 1712 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type); 1713 } 1714 1715 static int ipoib_get_vf_stats(struct net_device *dev, int vf, 1716 struct ifla_vf_stats *vf_stats) 1717 { 1718 struct ipoib_dev_priv *priv = netdev_priv(dev); 1719 1720 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats); 1721 } 1722 1723 static const struct header_ops ipoib_header_ops = { 1724 .create = ipoib_hard_header, 1725 }; 1726 1727 static const struct net_device_ops ipoib_netdev_ops_pf = { 1728 .ndo_uninit = ipoib_uninit, 1729 .ndo_open = ipoib_open, 1730 .ndo_stop = ipoib_stop, 1731 .ndo_change_mtu = ipoib_change_mtu, 1732 .ndo_fix_features = ipoib_fix_features, 1733 .ndo_start_xmit = ipoib_start_xmit, 1734 .ndo_tx_timeout = ipoib_timeout, 1735 .ndo_set_rx_mode = ipoib_set_mcast_list, 1736 .ndo_get_iflink = ipoib_get_iflink, 1737 .ndo_set_vf_link_state = ipoib_set_vf_link_state, 1738 .ndo_get_vf_config = ipoib_get_vf_config, 1739 .ndo_get_vf_stats = ipoib_get_vf_stats, 1740 .ndo_set_vf_guid = ipoib_set_vf_guid, 1741 .ndo_set_mac_address = ipoib_set_mac, 1742 }; 1743 1744 static const struct net_device_ops ipoib_netdev_ops_vf = { 1745 .ndo_uninit = ipoib_uninit, 1746 .ndo_open = ipoib_open, 1747 .ndo_stop = ipoib_stop, 1748 .ndo_change_mtu = ipoib_change_mtu, 1749 .ndo_fix_features = ipoib_fix_features, 1750 .ndo_start_xmit = ipoib_start_xmit, 1751 .ndo_tx_timeout = ipoib_timeout, 1752 .ndo_set_rx_mode = ipoib_set_mcast_list, 1753 .ndo_get_iflink = ipoib_get_iflink, 1754 }; 1755 1756 void ipoib_setup(struct net_device *dev) 1757 { 1758 struct ipoib_dev_priv *priv = netdev_priv(dev); 1759 1760 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION) 1761 dev->netdev_ops = &ipoib_netdev_ops_vf; 1762 else 1763 dev->netdev_ops = &ipoib_netdev_ops_pf; 1764 1765 dev->header_ops = &ipoib_header_ops; 1766 1767 ipoib_set_ethtool_ops(dev); 1768 1769 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT); 1770 1771 dev->watchdog_timeo = HZ; 1772 1773 dev->flags |= IFF_BROADCAST | IFF_MULTICAST; 1774 1775 dev->hard_header_len = IPOIB_ENCAP_LEN; 1776 dev->addr_len = INFINIBAND_ALEN; 1777 dev->type = ARPHRD_INFINIBAND; 1778 dev->tx_queue_len = ipoib_sendq_size * 2; 1779 dev->features = (NETIF_F_VLAN_CHALLENGED | 1780 NETIF_F_HIGHDMA); 1781 netif_keep_dst(dev); 1782 1783 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN); 1784 1785 priv->dev = dev; 1786 1787 spin_lock_init(&priv->lock); 1788 1789 init_rwsem(&priv->vlan_rwsem); 1790 1791 INIT_LIST_HEAD(&priv->path_list); 1792 INIT_LIST_HEAD(&priv->child_intfs); 1793 INIT_LIST_HEAD(&priv->dead_ahs); 1794 INIT_LIST_HEAD(&priv->multicast_list); 1795 1796 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task); 1797 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task); 1798 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light); 1799 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal); 1800 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy); 1801 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task); 1802 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah); 1803 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh); 1804 } 1805 1806 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name) 1807 { 1808 struct net_device *dev; 1809 1810 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name, 1811 NET_NAME_UNKNOWN, ipoib_setup); 1812 if (!dev) 1813 return NULL; 1814 1815 return netdev_priv(dev); 1816 } 1817 1818 static ssize_t show_pkey(struct device *dev, 1819 struct device_attribute *attr, char *buf) 1820 { 1821 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev)); 1822 1823 return sprintf(buf, "0x%04x\n", priv->pkey); 1824 } 1825 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL); 1826 1827 static ssize_t show_umcast(struct device *dev, 1828 struct device_attribute *attr, char *buf) 1829 { 1830 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev)); 1831 1832 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags)); 1833 } 1834 1835 void ipoib_set_umcast(struct net_device *ndev, int umcast_val) 1836 { 1837 struct ipoib_dev_priv *priv = netdev_priv(ndev); 1838 1839 if (umcast_val > 0) { 1840 set_bit(IPOIB_FLAG_UMCAST, &priv->flags); 1841 ipoib_warn(priv, "ignoring multicast groups joined directly " 1842 "by userspace\n"); 1843 } else 1844 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags); 1845 } 1846 1847 static ssize_t set_umcast(struct device *dev, 1848 struct device_attribute *attr, 1849 const char *buf, size_t count) 1850 { 1851 unsigned long umcast_val = simple_strtoul(buf, NULL, 0); 1852 1853 ipoib_set_umcast(to_net_dev(dev), umcast_val); 1854 1855 return count; 1856 } 1857 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast); 1858 1859 int ipoib_add_umcast_attr(struct net_device *dev) 1860 { 1861 return device_create_file(&dev->dev, &dev_attr_umcast); 1862 } 1863 1864 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid) 1865 { 1866 struct ipoib_dev_priv *child_priv; 1867 struct net_device *netdev = priv->dev; 1868 1869 netif_addr_lock_bh(netdev); 1870 1871 memcpy(&priv->local_gid.global.interface_id, 1872 &gid->global.interface_id, 1873 sizeof(gid->global.interface_id)); 1874 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid)); 1875 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 1876 1877 netif_addr_unlock_bh(netdev); 1878 1879 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 1880 down_read(&priv->vlan_rwsem); 1881 list_for_each_entry(child_priv, &priv->child_intfs, list) 1882 set_base_guid(child_priv, gid); 1883 up_read(&priv->vlan_rwsem); 1884 } 1885 } 1886 1887 static int ipoib_check_lladdr(struct net_device *dev, 1888 struct sockaddr_storage *ss) 1889 { 1890 union ib_gid *gid = (union ib_gid *)(ss->__data + 4); 1891 int ret = 0; 1892 1893 netif_addr_lock_bh(dev); 1894 1895 /* Make sure the QPN, reserved and subnet prefix match the current 1896 * lladdr, it also makes sure the lladdr is unicast. 1897 */ 1898 if (memcmp(dev->dev_addr, ss->__data, 1899 4 + sizeof(gid->global.subnet_prefix)) || 1900 gid->global.interface_id == 0) 1901 ret = -EINVAL; 1902 1903 netif_addr_unlock_bh(dev); 1904 1905 return ret; 1906 } 1907 1908 static int ipoib_set_mac(struct net_device *dev, void *addr) 1909 { 1910 struct ipoib_dev_priv *priv = netdev_priv(dev); 1911 struct sockaddr_storage *ss = addr; 1912 int ret; 1913 1914 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev)) 1915 return -EBUSY; 1916 1917 ret = ipoib_check_lladdr(dev, ss); 1918 if (ret) 1919 return ret; 1920 1921 set_base_guid(priv, (union ib_gid *)(ss->__data + 4)); 1922 1923 queue_work(ipoib_workqueue, &priv->flush_light); 1924 1925 return 0; 1926 } 1927 1928 static ssize_t create_child(struct device *dev, 1929 struct device_attribute *attr, 1930 const char *buf, size_t count) 1931 { 1932 int pkey; 1933 int ret; 1934 1935 if (sscanf(buf, "%i", &pkey) != 1) 1936 return -EINVAL; 1937 1938 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000) 1939 return -EINVAL; 1940 1941 /* 1942 * Set the full membership bit, so that we join the right 1943 * broadcast group, etc. 1944 */ 1945 pkey |= 0x8000; 1946 1947 ret = ipoib_vlan_add(to_net_dev(dev), pkey); 1948 1949 return ret ? ret : count; 1950 } 1951 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child); 1952 1953 static ssize_t delete_child(struct device *dev, 1954 struct device_attribute *attr, 1955 const char *buf, size_t count) 1956 { 1957 int pkey; 1958 int ret; 1959 1960 if (sscanf(buf, "%i", &pkey) != 1) 1961 return -EINVAL; 1962 1963 if (pkey < 0 || pkey > 0xffff) 1964 return -EINVAL; 1965 1966 ret = ipoib_vlan_delete(to_net_dev(dev), pkey); 1967 1968 return ret ? ret : count; 1969 1970 } 1971 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child); 1972 1973 int ipoib_add_pkey_attr(struct net_device *dev) 1974 { 1975 return device_create_file(&dev->dev, &dev_attr_pkey); 1976 } 1977 1978 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca) 1979 { 1980 priv->hca_caps = hca->attrs.device_cap_flags; 1981 1982 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) { 1983 priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM; 1984 1985 if (priv->hca_caps & IB_DEVICE_UD_TSO) 1986 priv->dev->hw_features |= NETIF_F_TSO; 1987 1988 priv->dev->features |= priv->dev->hw_features; 1989 } 1990 1991 return 0; 1992 } 1993 1994 static struct net_device *ipoib_add_port(const char *format, 1995 struct ib_device *hca, u8 port) 1996 { 1997 struct ipoib_dev_priv *priv; 1998 struct ib_port_attr attr; 1999 int result = -ENOMEM; 2000 2001 priv = ipoib_intf_alloc(format); 2002 if (!priv) 2003 goto alloc_mem_failed; 2004 2005 SET_NETDEV_DEV(priv->dev, hca->dma_device); 2006 priv->dev->dev_id = port - 1; 2007 2008 result = ib_query_port(hca, port, &attr); 2009 if (!result) 2010 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu); 2011 else { 2012 printk(KERN_WARNING "%s: ib_query_port %d failed\n", 2013 hca->name, port); 2014 goto device_init_failed; 2015 } 2016 2017 /* MTU will be reset when mcast join happens */ 2018 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu); 2019 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu; 2020 priv->dev->max_mtu = IPOIB_CM_MTU; 2021 2022 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh); 2023 2024 result = ib_query_pkey(hca, port, 0, &priv->pkey); 2025 if (result) { 2026 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n", 2027 hca->name, port, result); 2028 goto device_init_failed; 2029 } 2030 2031 result = ipoib_set_dev_features(priv, hca); 2032 if (result) 2033 goto device_init_failed; 2034 2035 /* 2036 * Set the full membership bit, so that we join the right 2037 * broadcast group, etc. 2038 */ 2039 priv->pkey |= 0x8000; 2040 2041 priv->dev->broadcast[8] = priv->pkey >> 8; 2042 priv->dev->broadcast[9] = priv->pkey & 0xff; 2043 2044 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL); 2045 if (result) { 2046 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n", 2047 hca->name, port, result); 2048 goto device_init_failed; 2049 } else 2050 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid)); 2051 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 2052 2053 result = ipoib_dev_init(priv->dev, hca, port); 2054 if (result < 0) { 2055 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n", 2056 hca->name, port, result); 2057 goto device_init_failed; 2058 } 2059 2060 INIT_IB_EVENT_HANDLER(&priv->event_handler, 2061 priv->ca, ipoib_event); 2062 result = ib_register_event_handler(&priv->event_handler); 2063 if (result < 0) { 2064 printk(KERN_WARNING "%s: ib_register_event_handler failed for " 2065 "port %d (ret = %d)\n", 2066 hca->name, port, result); 2067 goto event_failed; 2068 } 2069 2070 result = register_netdev(priv->dev); 2071 if (result) { 2072 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n", 2073 hca->name, port, result); 2074 goto register_failed; 2075 } 2076 2077 ipoib_create_debug_files(priv->dev); 2078 2079 if (ipoib_cm_add_mode_attr(priv->dev)) 2080 goto sysfs_failed; 2081 if (ipoib_add_pkey_attr(priv->dev)) 2082 goto sysfs_failed; 2083 if (ipoib_add_umcast_attr(priv->dev)) 2084 goto sysfs_failed; 2085 if (device_create_file(&priv->dev->dev, &dev_attr_create_child)) 2086 goto sysfs_failed; 2087 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child)) 2088 goto sysfs_failed; 2089 2090 return priv->dev; 2091 2092 sysfs_failed: 2093 ipoib_delete_debug_files(priv->dev); 2094 unregister_netdev(priv->dev); 2095 2096 register_failed: 2097 ib_unregister_event_handler(&priv->event_handler); 2098 flush_workqueue(ipoib_workqueue); 2099 /* Stop GC if started before flush */ 2100 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2101 cancel_delayed_work(&priv->neigh_reap_task); 2102 flush_workqueue(priv->wq); 2103 2104 event_failed: 2105 ipoib_dev_cleanup(priv->dev); 2106 2107 device_init_failed: 2108 free_netdev(priv->dev); 2109 2110 alloc_mem_failed: 2111 return ERR_PTR(result); 2112 } 2113 2114 static void ipoib_add_one(struct ib_device *device) 2115 { 2116 struct list_head *dev_list; 2117 struct net_device *dev; 2118 struct ipoib_dev_priv *priv; 2119 int p; 2120 int count = 0; 2121 2122 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL); 2123 if (!dev_list) 2124 return; 2125 2126 INIT_LIST_HEAD(dev_list); 2127 2128 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) { 2129 if (!rdma_protocol_ib(device, p)) 2130 continue; 2131 dev = ipoib_add_port("ib%d", device, p); 2132 if (!IS_ERR(dev)) { 2133 priv = netdev_priv(dev); 2134 list_add_tail(&priv->list, dev_list); 2135 count++; 2136 } 2137 } 2138 2139 if (!count) { 2140 kfree(dev_list); 2141 return; 2142 } 2143 2144 ib_set_client_data(device, &ipoib_client, dev_list); 2145 } 2146 2147 static void ipoib_remove_one(struct ib_device *device, void *client_data) 2148 { 2149 struct ipoib_dev_priv *priv, *tmp; 2150 struct list_head *dev_list = client_data; 2151 2152 if (!dev_list) 2153 return; 2154 2155 list_for_each_entry_safe(priv, tmp, dev_list, list) { 2156 ib_unregister_event_handler(&priv->event_handler); 2157 flush_workqueue(ipoib_workqueue); 2158 2159 /* mark interface in the middle of destruction */ 2160 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags); 2161 2162 rtnl_lock(); 2163 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); 2164 rtnl_unlock(); 2165 2166 /* Stop GC */ 2167 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2168 cancel_delayed_work(&priv->neigh_reap_task); 2169 flush_workqueue(priv->wq); 2170 2171 unregister_netdev(priv->dev); 2172 free_netdev(priv->dev); 2173 } 2174 2175 kfree(dev_list); 2176 } 2177 2178 static int __init ipoib_init_module(void) 2179 { 2180 int ret; 2181 2182 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size); 2183 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE); 2184 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE); 2185 2186 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size); 2187 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE); 2188 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE); 2189 #ifdef CONFIG_INFINIBAND_IPOIB_CM 2190 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP); 2191 #endif 2192 2193 /* 2194 * When copying small received packets, we only copy from the 2195 * linear data part of the SKB, so we rely on this condition. 2196 */ 2197 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE); 2198 2199 ret = ipoib_register_debugfs(); 2200 if (ret) 2201 return ret; 2202 2203 /* 2204 * We create a global workqueue here that is used for all flush 2205 * operations. However, if you attempt to flush a workqueue 2206 * from a task on that same workqueue, it deadlocks the system. 2207 * We want to be able to flush the tasks associated with a 2208 * specific net device, so we also create a workqueue for each 2209 * netdevice. We queue up the tasks for that device only on 2210 * its private workqueue, and we only queue up flush events 2211 * on our global flush workqueue. This avoids the deadlocks. 2212 */ 2213 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 2214 WQ_MEM_RECLAIM); 2215 if (!ipoib_workqueue) { 2216 ret = -ENOMEM; 2217 goto err_fs; 2218 } 2219 2220 ib_sa_register_client(&ipoib_sa_client); 2221 2222 ret = ib_register_client(&ipoib_client); 2223 if (ret) 2224 goto err_sa; 2225 2226 ret = ipoib_netlink_init(); 2227 if (ret) 2228 goto err_client; 2229 2230 return 0; 2231 2232 err_client: 2233 ib_unregister_client(&ipoib_client); 2234 2235 err_sa: 2236 ib_sa_unregister_client(&ipoib_sa_client); 2237 destroy_workqueue(ipoib_workqueue); 2238 2239 err_fs: 2240 ipoib_unregister_debugfs(); 2241 2242 return ret; 2243 } 2244 2245 static void __exit ipoib_cleanup_module(void) 2246 { 2247 ipoib_netlink_fini(); 2248 ib_unregister_client(&ipoib_client); 2249 ib_sa_unregister_client(&ipoib_sa_client); 2250 ipoib_unregister_debugfs(); 2251 destroy_workqueue(ipoib_workqueue); 2252 } 2253 2254 module_init(ipoib_init_module); 2255 module_exit(ipoib_cleanup_module); 2256