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