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 if (priv->mcast_mtu < priv->admin_mtu) 233 ipoib_dbg(priv, "MTU must be smaller than the underlying " 234 "link layer MTU - 4 (%u)\n", priv->mcast_mtu); 235 236 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu); 237 238 return 0; 239 } 240 241 /* Called with an RCU read lock taken */ 242 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr, 243 struct net_device *dev) 244 { 245 struct net *net = dev_net(dev); 246 struct in_device *in_dev; 247 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr; 248 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr; 249 __be32 ret_addr; 250 251 switch (addr->sa_family) { 252 case AF_INET: 253 in_dev = in_dev_get(dev); 254 if (!in_dev) 255 return false; 256 257 ret_addr = inet_confirm_addr(net, in_dev, 0, 258 addr_in->sin_addr.s_addr, 259 RT_SCOPE_HOST); 260 in_dev_put(in_dev); 261 if (ret_addr) 262 return true; 263 264 break; 265 case AF_INET6: 266 if (IS_ENABLED(CONFIG_IPV6) && 267 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1)) 268 return true; 269 270 break; 271 } 272 return false; 273 } 274 275 /** 276 * Find the master net_device on top of the given net_device. 277 * @dev: base IPoIB net_device 278 * 279 * Returns the master net_device with a reference held, or the same net_device 280 * if no master exists. 281 */ 282 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev) 283 { 284 struct net_device *master; 285 286 rcu_read_lock(); 287 master = netdev_master_upper_dev_get_rcu(dev); 288 if (master) 289 dev_hold(master); 290 rcu_read_unlock(); 291 292 if (master) 293 return master; 294 295 dev_hold(dev); 296 return dev; 297 } 298 299 struct ipoib_walk_data { 300 const struct sockaddr *addr; 301 struct net_device *result; 302 }; 303 304 static int ipoib_upper_walk(struct net_device *upper, void *_data) 305 { 306 struct ipoib_walk_data *data = _data; 307 int ret = 0; 308 309 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) { 310 dev_hold(upper); 311 data->result = upper; 312 ret = 1; 313 } 314 315 return ret; 316 } 317 318 /** 319 * Find a net_device matching the given address, which is an upper device of 320 * the given net_device. 321 * @addr: IP address to look for. 322 * @dev: base IPoIB net_device 323 * 324 * If found, returns the net_device with a reference held. Otherwise return 325 * NULL. 326 */ 327 static struct net_device *ipoib_get_net_dev_match_addr( 328 const struct sockaddr *addr, struct net_device *dev) 329 { 330 struct ipoib_walk_data data = { 331 .addr = addr, 332 }; 333 334 rcu_read_lock(); 335 if (ipoib_is_dev_match_addr_rcu(addr, dev)) { 336 dev_hold(dev); 337 data.result = dev; 338 goto out; 339 } 340 341 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data); 342 out: 343 rcu_read_unlock(); 344 return data.result; 345 } 346 347 /* returns the number of IPoIB netdevs on top a given ipoib device matching a 348 * pkey_index and address, if one exists. 349 * 350 * @found_net_dev: contains a matching net_device if the return value >= 1, 351 * with a reference held. */ 352 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv, 353 const union ib_gid *gid, 354 u16 pkey_index, 355 const struct sockaddr *addr, 356 int nesting, 357 struct net_device **found_net_dev) 358 { 359 struct ipoib_dev_priv *child_priv; 360 struct net_device *net_dev = NULL; 361 int matches = 0; 362 363 if (priv->pkey_index == pkey_index && 364 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) { 365 if (!addr) { 366 net_dev = ipoib_get_master_net_dev(priv->dev); 367 } else { 368 /* Verify the net_device matches the IP address, as 369 * IPoIB child devices currently share a GID. */ 370 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev); 371 } 372 if (net_dev) { 373 if (!*found_net_dev) 374 *found_net_dev = net_dev; 375 else 376 dev_put(net_dev); 377 ++matches; 378 } 379 } 380 381 /* Check child interfaces */ 382 down_read_nested(&priv->vlan_rwsem, nesting); 383 list_for_each_entry(child_priv, &priv->child_intfs, list) { 384 matches += ipoib_match_gid_pkey_addr(child_priv, gid, 385 pkey_index, addr, 386 nesting + 1, 387 found_net_dev); 388 if (matches > 1) 389 break; 390 } 391 up_read(&priv->vlan_rwsem); 392 393 return matches; 394 } 395 396 /* Returns the number of matching net_devs found (between 0 and 2). Also 397 * return the matching net_device in the @net_dev parameter, holding a 398 * reference to the net_device, if the number of matches >= 1 */ 399 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port, 400 u16 pkey_index, 401 const union ib_gid *gid, 402 const struct sockaddr *addr, 403 struct net_device **net_dev) 404 { 405 struct ipoib_dev_priv *priv; 406 int matches = 0; 407 408 *net_dev = NULL; 409 410 list_for_each_entry(priv, dev_list, list) { 411 if (priv->port != port) 412 continue; 413 414 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index, 415 addr, 0, net_dev); 416 if (matches > 1) 417 break; 418 } 419 420 return matches; 421 } 422 423 static struct net_device *ipoib_get_net_dev_by_params( 424 struct ib_device *dev, u8 port, u16 pkey, 425 const union ib_gid *gid, const struct sockaddr *addr, 426 void *client_data) 427 { 428 struct net_device *net_dev; 429 struct list_head *dev_list = client_data; 430 u16 pkey_index; 431 int matches; 432 int ret; 433 434 if (!rdma_protocol_ib(dev, port)) 435 return NULL; 436 437 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index); 438 if (ret) 439 return NULL; 440 441 if (!dev_list) 442 return NULL; 443 444 /* See if we can find a unique device matching the L2 parameters */ 445 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 446 gid, NULL, &net_dev); 447 448 switch (matches) { 449 case 0: 450 return NULL; 451 case 1: 452 return net_dev; 453 } 454 455 dev_put(net_dev); 456 457 /* Couldn't find a unique device with L2 parameters only. Use L3 458 * address to uniquely match the net device */ 459 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index, 460 gid, addr, &net_dev); 461 switch (matches) { 462 case 0: 463 return NULL; 464 default: 465 dev_warn_ratelimited(&dev->dev, 466 "duplicate IP address detected\n"); 467 /* Fall through */ 468 case 1: 469 return net_dev; 470 } 471 } 472 473 int ipoib_set_mode(struct net_device *dev, const char *buf) 474 { 475 struct ipoib_dev_priv *priv = netdev_priv(dev); 476 477 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) && 478 !strcmp(buf, "connected\n")) || 479 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) && 480 !strcmp(buf, "datagram\n"))) { 481 return 0; 482 } 483 484 /* flush paths if we switch modes so that connections are restarted */ 485 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) { 486 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 487 ipoib_warn(priv, "enabling connected mode " 488 "will cause multicast packet drops\n"); 489 netdev_update_features(dev); 490 dev_set_mtu(dev, ipoib_cm_max_mtu(dev)); 491 rtnl_unlock(); 492 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM; 493 494 ipoib_flush_paths(dev); 495 rtnl_lock(); 496 return 0; 497 } 498 499 if (!strcmp(buf, "datagram\n")) { 500 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 501 netdev_update_features(dev); 502 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu)); 503 rtnl_unlock(); 504 ipoib_flush_paths(dev); 505 rtnl_lock(); 506 return 0; 507 } 508 509 return -EINVAL; 510 } 511 512 struct ipoib_path *__path_find(struct net_device *dev, void *gid) 513 { 514 struct ipoib_dev_priv *priv = netdev_priv(dev); 515 struct rb_node *n = priv->path_tree.rb_node; 516 struct ipoib_path *path; 517 int ret; 518 519 while (n) { 520 path = rb_entry(n, struct ipoib_path, rb_node); 521 522 ret = memcmp(gid, path->pathrec.dgid.raw, 523 sizeof (union ib_gid)); 524 525 if (ret < 0) 526 n = n->rb_left; 527 else if (ret > 0) 528 n = n->rb_right; 529 else 530 return path; 531 } 532 533 return NULL; 534 } 535 536 static int __path_add(struct net_device *dev, struct ipoib_path *path) 537 { 538 struct ipoib_dev_priv *priv = netdev_priv(dev); 539 struct rb_node **n = &priv->path_tree.rb_node; 540 struct rb_node *pn = NULL; 541 struct ipoib_path *tpath; 542 int ret; 543 544 while (*n) { 545 pn = *n; 546 tpath = rb_entry(pn, struct ipoib_path, rb_node); 547 548 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw, 549 sizeof (union ib_gid)); 550 if (ret < 0) 551 n = &pn->rb_left; 552 else if (ret > 0) 553 n = &pn->rb_right; 554 else 555 return -EEXIST; 556 } 557 558 rb_link_node(&path->rb_node, pn, n); 559 rb_insert_color(&path->rb_node, &priv->path_tree); 560 561 list_add_tail(&path->list, &priv->path_list); 562 563 return 0; 564 } 565 566 static void path_free(struct net_device *dev, struct ipoib_path *path) 567 { 568 struct sk_buff *skb; 569 570 while ((skb = __skb_dequeue(&path->queue))) 571 dev_kfree_skb_irq(skb); 572 573 ipoib_dbg(netdev_priv(dev), "path_free\n"); 574 575 /* remove all neigh connected to this path */ 576 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 577 578 if (path->ah) 579 ipoib_put_ah(path->ah); 580 581 kfree(path); 582 } 583 584 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG 585 586 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev) 587 { 588 struct ipoib_path_iter *iter; 589 590 iter = kmalloc(sizeof *iter, GFP_KERNEL); 591 if (!iter) 592 return NULL; 593 594 iter->dev = dev; 595 memset(iter->path.pathrec.dgid.raw, 0, 16); 596 597 if (ipoib_path_iter_next(iter)) { 598 kfree(iter); 599 return NULL; 600 } 601 602 return iter; 603 } 604 605 int ipoib_path_iter_next(struct ipoib_path_iter *iter) 606 { 607 struct ipoib_dev_priv *priv = netdev_priv(iter->dev); 608 struct rb_node *n; 609 struct ipoib_path *path; 610 int ret = 1; 611 612 spin_lock_irq(&priv->lock); 613 614 n = rb_first(&priv->path_tree); 615 616 while (n) { 617 path = rb_entry(n, struct ipoib_path, rb_node); 618 619 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw, 620 sizeof (union ib_gid)) < 0) { 621 iter->path = *path; 622 ret = 0; 623 break; 624 } 625 626 n = rb_next(n); 627 } 628 629 spin_unlock_irq(&priv->lock); 630 631 return ret; 632 } 633 634 void ipoib_path_iter_read(struct ipoib_path_iter *iter, 635 struct ipoib_path *path) 636 { 637 *path = iter->path; 638 } 639 640 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */ 641 642 void ipoib_mark_paths_invalid(struct net_device *dev) 643 { 644 struct ipoib_dev_priv *priv = netdev_priv(dev); 645 struct ipoib_path *path, *tp; 646 647 spin_lock_irq(&priv->lock); 648 649 list_for_each_entry_safe(path, tp, &priv->path_list, list) { 650 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n", 651 be16_to_cpu(path->pathrec.dlid), 652 path->pathrec.dgid.raw); 653 path->valid = 0; 654 } 655 656 spin_unlock_irq(&priv->lock); 657 } 658 659 struct classport_info_context { 660 struct ipoib_dev_priv *priv; 661 struct completion done; 662 struct ib_sa_query *sa_query; 663 }; 664 665 static void classport_info_query_cb(int status, struct ib_class_port_info *rec, 666 void *context) 667 { 668 struct classport_info_context *cb_ctx = context; 669 struct ipoib_dev_priv *priv; 670 671 WARN_ON(!context); 672 673 priv = cb_ctx->priv; 674 675 if (status || !rec) { 676 pr_debug("device: %s failed query classport_info status: %d\n", 677 priv->dev->name, status); 678 /* keeps the default, will try next mcast_restart */ 679 priv->sm_fullmember_sendonly_support = false; 680 goto out; 681 } 682 683 if (ib_get_cpi_capmask2(rec) & 684 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) { 685 pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n", 686 priv->dev->name); 687 priv->sm_fullmember_sendonly_support = true; 688 } else { 689 pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n", 690 priv->dev->name); 691 priv->sm_fullmember_sendonly_support = false; 692 } 693 694 out: 695 complete(&cb_ctx->done); 696 } 697 698 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv) 699 { 700 struct classport_info_context *callback_context; 701 int ret; 702 703 callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL); 704 if (!callback_context) 705 return -ENOMEM; 706 707 callback_context->priv = priv; 708 init_completion(&callback_context->done); 709 710 ret = ib_sa_classport_info_rec_query(&ipoib_sa_client, 711 priv->ca, priv->port, 3000, 712 GFP_KERNEL, 713 classport_info_query_cb, 714 callback_context, 715 &callback_context->sa_query); 716 if (ret < 0) { 717 pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n", 718 priv->dev->name, ret); 719 kfree(callback_context); 720 return ret; 721 } 722 723 /* waiting for the callback to finish before returnning */ 724 wait_for_completion(&callback_context->done); 725 kfree(callback_context); 726 727 return ret; 728 } 729 730 void ipoib_flush_paths(struct net_device *dev) 731 { 732 struct ipoib_dev_priv *priv = netdev_priv(dev); 733 struct ipoib_path *path, *tp; 734 LIST_HEAD(remove_list); 735 unsigned long flags; 736 737 netif_tx_lock_bh(dev); 738 spin_lock_irqsave(&priv->lock, flags); 739 740 list_splice_init(&priv->path_list, &remove_list); 741 742 list_for_each_entry(path, &remove_list, list) 743 rb_erase(&path->rb_node, &priv->path_tree); 744 745 list_for_each_entry_safe(path, tp, &remove_list, list) { 746 if (path->query) 747 ib_sa_cancel_query(path->query_id, path->query); 748 spin_unlock_irqrestore(&priv->lock, flags); 749 netif_tx_unlock_bh(dev); 750 wait_for_completion(&path->done); 751 path_free(dev, path); 752 netif_tx_lock_bh(dev); 753 spin_lock_irqsave(&priv->lock, flags); 754 } 755 756 spin_unlock_irqrestore(&priv->lock, flags); 757 netif_tx_unlock_bh(dev); 758 } 759 760 static void path_rec_completion(int status, 761 struct ib_sa_path_rec *pathrec, 762 void *path_ptr) 763 { 764 struct ipoib_path *path = path_ptr; 765 struct net_device *dev = path->dev; 766 struct ipoib_dev_priv *priv = netdev_priv(dev); 767 struct ipoib_ah *ah = NULL; 768 struct ipoib_ah *old_ah = NULL; 769 struct ipoib_neigh *neigh, *tn; 770 struct sk_buff_head skqueue; 771 struct sk_buff *skb; 772 unsigned long flags; 773 774 if (!status) 775 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n", 776 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw); 777 else 778 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n", 779 status, path->pathrec.dgid.raw); 780 781 skb_queue_head_init(&skqueue); 782 783 if (!status) { 784 struct ib_ah_attr av; 785 786 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av)) 787 ah = ipoib_create_ah(dev, priv->pd, &av); 788 } 789 790 spin_lock_irqsave(&priv->lock, flags); 791 792 if (!IS_ERR_OR_NULL(ah)) { 793 path->pathrec = *pathrec; 794 795 old_ah = path->ah; 796 path->ah = ah; 797 798 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n", 799 ah, be16_to_cpu(pathrec->dlid), pathrec->sl); 800 801 while ((skb = __skb_dequeue(&path->queue))) 802 __skb_queue_tail(&skqueue, skb); 803 804 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) { 805 if (neigh->ah) { 806 WARN_ON(neigh->ah != old_ah); 807 /* 808 * Dropping the ah reference inside 809 * priv->lock is safe here, because we 810 * will hold one more reference from 811 * the original value of path->ah (ie 812 * old_ah). 813 */ 814 ipoib_put_ah(neigh->ah); 815 } 816 kref_get(&path->ah->ref); 817 neigh->ah = path->ah; 818 819 if (ipoib_cm_enabled(dev, neigh->daddr)) { 820 if (!ipoib_cm_get(neigh)) 821 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, 822 path, 823 neigh)); 824 if (!ipoib_cm_get(neigh)) { 825 ipoib_neigh_free(neigh); 826 continue; 827 } 828 } 829 830 while ((skb = __skb_dequeue(&neigh->queue))) 831 __skb_queue_tail(&skqueue, skb); 832 } 833 path->valid = 1; 834 } 835 836 path->query = NULL; 837 complete(&path->done); 838 839 spin_unlock_irqrestore(&priv->lock, flags); 840 841 if (IS_ERR_OR_NULL(ah)) 842 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw); 843 844 if (old_ah) 845 ipoib_put_ah(old_ah); 846 847 while ((skb = __skb_dequeue(&skqueue))) { 848 skb->dev = dev; 849 if (dev_queue_xmit(skb)) 850 ipoib_warn(priv, "dev_queue_xmit failed " 851 "to requeue packet\n"); 852 } 853 } 854 855 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid) 856 { 857 struct ipoib_dev_priv *priv = netdev_priv(dev); 858 struct ipoib_path *path; 859 860 if (!priv->broadcast) 861 return NULL; 862 863 path = kzalloc(sizeof *path, GFP_ATOMIC); 864 if (!path) 865 return NULL; 866 867 path->dev = dev; 868 869 skb_queue_head_init(&path->queue); 870 871 INIT_LIST_HEAD(&path->neigh_list); 872 873 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid)); 874 path->pathrec.sgid = priv->local_gid; 875 path->pathrec.pkey = cpu_to_be16(priv->pkey); 876 path->pathrec.numb_path = 1; 877 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class; 878 879 return path; 880 } 881 882 static int path_rec_start(struct net_device *dev, 883 struct ipoib_path *path) 884 { 885 struct ipoib_dev_priv *priv = netdev_priv(dev); 886 887 ipoib_dbg(priv, "Start path record lookup for %pI6\n", 888 path->pathrec.dgid.raw); 889 890 init_completion(&path->done); 891 892 path->query_id = 893 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port, 894 &path->pathrec, 895 IB_SA_PATH_REC_DGID | 896 IB_SA_PATH_REC_SGID | 897 IB_SA_PATH_REC_NUMB_PATH | 898 IB_SA_PATH_REC_TRAFFIC_CLASS | 899 IB_SA_PATH_REC_PKEY, 900 1000, GFP_ATOMIC, 901 path_rec_completion, 902 path, &path->query); 903 if (path->query_id < 0) { 904 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id); 905 path->query = NULL; 906 complete(&path->done); 907 return path->query_id; 908 } 909 910 return 0; 911 } 912 913 static void neigh_add_path(struct sk_buff *skb, u8 *daddr, 914 struct net_device *dev) 915 { 916 struct ipoib_dev_priv *priv = netdev_priv(dev); 917 struct ipoib_path *path; 918 struct ipoib_neigh *neigh; 919 unsigned long flags; 920 921 spin_lock_irqsave(&priv->lock, flags); 922 neigh = ipoib_neigh_alloc(daddr, dev); 923 if (!neigh) { 924 spin_unlock_irqrestore(&priv->lock, flags); 925 ++dev->stats.tx_dropped; 926 dev_kfree_skb_any(skb); 927 return; 928 } 929 930 path = __path_find(dev, daddr + 4); 931 if (!path) { 932 path = path_rec_create(dev, daddr + 4); 933 if (!path) 934 goto err_path; 935 936 __path_add(dev, path); 937 } 938 939 list_add_tail(&neigh->list, &path->neigh_list); 940 941 if (path->ah) { 942 kref_get(&path->ah->ref); 943 neigh->ah = path->ah; 944 945 if (ipoib_cm_enabled(dev, neigh->daddr)) { 946 if (!ipoib_cm_get(neigh)) 947 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh)); 948 if (!ipoib_cm_get(neigh)) { 949 ipoib_neigh_free(neigh); 950 goto err_drop; 951 } 952 if (skb_queue_len(&neigh->queue) < 953 IPOIB_MAX_PATH_REC_QUEUE) { 954 /* put pseudoheader back on for next time */ 955 skb_push(skb, IPOIB_PSEUDO_LEN); 956 __skb_queue_tail(&neigh->queue, skb); 957 } else { 958 ipoib_warn(priv, "queue length limit %d. Packet drop.\n", 959 skb_queue_len(&neigh->queue)); 960 goto err_drop; 961 } 962 } else { 963 spin_unlock_irqrestore(&priv->lock, flags); 964 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr)); 965 ipoib_neigh_put(neigh); 966 return; 967 } 968 } else { 969 neigh->ah = NULL; 970 971 if (!path->query && path_rec_start(dev, path)) 972 goto err_path; 973 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) 974 __skb_queue_tail(&neigh->queue, skb); 975 else 976 goto err_drop; 977 } 978 979 spin_unlock_irqrestore(&priv->lock, flags); 980 ipoib_neigh_put(neigh); 981 return; 982 983 err_path: 984 ipoib_neigh_free(neigh); 985 err_drop: 986 ++dev->stats.tx_dropped; 987 dev_kfree_skb_any(skb); 988 989 spin_unlock_irqrestore(&priv->lock, flags); 990 ipoib_neigh_put(neigh); 991 } 992 993 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev, 994 struct ipoib_pseudo_header *phdr) 995 { 996 struct ipoib_dev_priv *priv = netdev_priv(dev); 997 struct ipoib_path *path; 998 unsigned long flags; 999 1000 spin_lock_irqsave(&priv->lock, flags); 1001 1002 path = __path_find(dev, phdr->hwaddr + 4); 1003 if (!path || !path->valid) { 1004 int new_path = 0; 1005 1006 if (!path) { 1007 path = path_rec_create(dev, phdr->hwaddr + 4); 1008 new_path = 1; 1009 } 1010 if (path) { 1011 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1012 /* put pseudoheader back on for next time */ 1013 skb_push(skb, IPOIB_PSEUDO_LEN); 1014 __skb_queue_tail(&path->queue, skb); 1015 } else { 1016 ++dev->stats.tx_dropped; 1017 dev_kfree_skb_any(skb); 1018 } 1019 1020 if (!path->query && path_rec_start(dev, path)) { 1021 spin_unlock_irqrestore(&priv->lock, flags); 1022 if (new_path) 1023 path_free(dev, path); 1024 return; 1025 } else 1026 __path_add(dev, path); 1027 } else { 1028 ++dev->stats.tx_dropped; 1029 dev_kfree_skb_any(skb); 1030 } 1031 1032 spin_unlock_irqrestore(&priv->lock, flags); 1033 return; 1034 } 1035 1036 if (path->ah) { 1037 ipoib_dbg(priv, "Send unicast ARP to %04x\n", 1038 be16_to_cpu(path->pathrec.dlid)); 1039 1040 spin_unlock_irqrestore(&priv->lock, flags); 1041 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr)); 1042 return; 1043 } else if ((path->query || !path_rec_start(dev, path)) && 1044 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1045 /* put pseudoheader back on for next time */ 1046 skb_push(skb, IPOIB_PSEUDO_LEN); 1047 __skb_queue_tail(&path->queue, skb); 1048 } else { 1049 ++dev->stats.tx_dropped; 1050 dev_kfree_skb_any(skb); 1051 } 1052 1053 spin_unlock_irqrestore(&priv->lock, flags); 1054 } 1055 1056 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev) 1057 { 1058 struct ipoib_dev_priv *priv = netdev_priv(dev); 1059 struct ipoib_neigh *neigh; 1060 struct ipoib_pseudo_header *phdr; 1061 struct ipoib_header *header; 1062 unsigned long flags; 1063 1064 phdr = (struct ipoib_pseudo_header *) skb->data; 1065 skb_pull(skb, sizeof(*phdr)); 1066 header = (struct ipoib_header *) skb->data; 1067 1068 if (unlikely(phdr->hwaddr[4] == 0xff)) { 1069 /* multicast, arrange "if" according to probability */ 1070 if ((header->proto != htons(ETH_P_IP)) && 1071 (header->proto != htons(ETH_P_IPV6)) && 1072 (header->proto != htons(ETH_P_ARP)) && 1073 (header->proto != htons(ETH_P_RARP)) && 1074 (header->proto != htons(ETH_P_TIPC))) { 1075 /* ethertype not supported by IPoIB */ 1076 ++dev->stats.tx_dropped; 1077 dev_kfree_skb_any(skb); 1078 return NETDEV_TX_OK; 1079 } 1080 /* Add in the P_Key for multicast*/ 1081 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff; 1082 phdr->hwaddr[9] = priv->pkey & 0xff; 1083 1084 neigh = ipoib_neigh_get(dev, phdr->hwaddr); 1085 if (likely(neigh)) 1086 goto send_using_neigh; 1087 ipoib_mcast_send(dev, phdr->hwaddr, skb); 1088 return NETDEV_TX_OK; 1089 } 1090 1091 /* unicast, arrange "switch" according to probability */ 1092 switch (header->proto) { 1093 case htons(ETH_P_IP): 1094 case htons(ETH_P_IPV6): 1095 case htons(ETH_P_TIPC): 1096 neigh = ipoib_neigh_get(dev, phdr->hwaddr); 1097 if (unlikely(!neigh)) { 1098 neigh_add_path(skb, phdr->hwaddr, dev); 1099 return NETDEV_TX_OK; 1100 } 1101 break; 1102 case htons(ETH_P_ARP): 1103 case htons(ETH_P_RARP): 1104 /* for unicast ARP and RARP should always perform path find */ 1105 unicast_arp_send(skb, dev, phdr); 1106 return NETDEV_TX_OK; 1107 default: 1108 /* ethertype not supported by IPoIB */ 1109 ++dev->stats.tx_dropped; 1110 dev_kfree_skb_any(skb); 1111 return NETDEV_TX_OK; 1112 } 1113 1114 send_using_neigh: 1115 /* note we now hold a ref to neigh */ 1116 if (ipoib_cm_get(neigh)) { 1117 if (ipoib_cm_up(neigh)) { 1118 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh)); 1119 goto unref; 1120 } 1121 } else if (neigh->ah) { 1122 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(phdr->hwaddr)); 1123 goto unref; 1124 } 1125 1126 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) { 1127 /* put pseudoheader back on for next time */ 1128 skb_push(skb, sizeof(*phdr)); 1129 spin_lock_irqsave(&priv->lock, flags); 1130 __skb_queue_tail(&neigh->queue, skb); 1131 spin_unlock_irqrestore(&priv->lock, flags); 1132 } else { 1133 ++dev->stats.tx_dropped; 1134 dev_kfree_skb_any(skb); 1135 } 1136 1137 unref: 1138 ipoib_neigh_put(neigh); 1139 1140 return NETDEV_TX_OK; 1141 } 1142 1143 static void ipoib_timeout(struct net_device *dev) 1144 { 1145 struct ipoib_dev_priv *priv = netdev_priv(dev); 1146 1147 ipoib_warn(priv, "transmit timeout: latency %d msecs\n", 1148 jiffies_to_msecs(jiffies - dev_trans_start(dev))); 1149 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n", 1150 netif_queue_stopped(dev), 1151 priv->tx_head, priv->tx_tail); 1152 /* XXX reset QP, etc. */ 1153 } 1154 1155 static int ipoib_hard_header(struct sk_buff *skb, 1156 struct net_device *dev, 1157 unsigned short type, 1158 const void *daddr, const void *saddr, unsigned len) 1159 { 1160 struct ipoib_pseudo_header *phdr; 1161 struct ipoib_header *header; 1162 1163 header = (struct ipoib_header *) skb_push(skb, sizeof *header); 1164 1165 header->proto = htons(type); 1166 header->reserved = 0; 1167 1168 /* 1169 * we don't rely on dst_entry structure, always stuff the 1170 * destination address into skb hard header so we can figure out where 1171 * to send the packet later. 1172 */ 1173 phdr = (struct ipoib_pseudo_header *) skb_push(skb, sizeof(*phdr)); 1174 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN); 1175 1176 return IPOIB_HARD_LEN; 1177 } 1178 1179 static void ipoib_set_mcast_list(struct net_device *dev) 1180 { 1181 struct ipoib_dev_priv *priv = netdev_priv(dev); 1182 1183 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) { 1184 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set"); 1185 return; 1186 } 1187 1188 queue_work(priv->wq, &priv->restart_task); 1189 } 1190 1191 static int ipoib_get_iflink(const struct net_device *dev) 1192 { 1193 struct ipoib_dev_priv *priv = netdev_priv(dev); 1194 1195 /* parent interface */ 1196 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) 1197 return dev->ifindex; 1198 1199 /* child/vlan interface */ 1200 return priv->parent->ifindex; 1201 } 1202 1203 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr) 1204 { 1205 /* 1206 * Use only the address parts that contributes to spreading 1207 * The subnet prefix is not used as one can not connect to 1208 * same remote port (GUID) using the same remote QPN via two 1209 * different subnets. 1210 */ 1211 /* qpn octets[1:4) & port GUID octets[12:20) */ 1212 u32 *d32 = (u32 *) daddr; 1213 u32 hv; 1214 1215 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0); 1216 return hv & htbl->mask; 1217 } 1218 1219 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr) 1220 { 1221 struct ipoib_dev_priv *priv = netdev_priv(dev); 1222 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1223 struct ipoib_neigh_hash *htbl; 1224 struct ipoib_neigh *neigh = NULL; 1225 u32 hash_val; 1226 1227 rcu_read_lock_bh(); 1228 1229 htbl = rcu_dereference_bh(ntbl->htbl); 1230 1231 if (!htbl) 1232 goto out_unlock; 1233 1234 hash_val = ipoib_addr_hash(htbl, daddr); 1235 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]); 1236 neigh != NULL; 1237 neigh = rcu_dereference_bh(neigh->hnext)) { 1238 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1239 /* found, take one ref on behalf of the caller */ 1240 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1241 /* deleted */ 1242 neigh = NULL; 1243 goto out_unlock; 1244 } 1245 1246 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)) 1247 neigh->alive = jiffies; 1248 goto out_unlock; 1249 } 1250 } 1251 1252 out_unlock: 1253 rcu_read_unlock_bh(); 1254 return neigh; 1255 } 1256 1257 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv) 1258 { 1259 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1260 struct ipoib_neigh_hash *htbl; 1261 unsigned long neigh_obsolete; 1262 unsigned long dt; 1263 unsigned long flags; 1264 int i; 1265 LIST_HEAD(remove_list); 1266 1267 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1268 return; 1269 1270 spin_lock_irqsave(&priv->lock, flags); 1271 1272 htbl = rcu_dereference_protected(ntbl->htbl, 1273 lockdep_is_held(&priv->lock)); 1274 1275 if (!htbl) 1276 goto out_unlock; 1277 1278 /* neigh is obsolete if it was idle for two GC periods */ 1279 dt = 2 * arp_tbl.gc_interval; 1280 neigh_obsolete = jiffies - dt; 1281 /* handle possible race condition */ 1282 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1283 goto out_unlock; 1284 1285 for (i = 0; i < htbl->size; i++) { 1286 struct ipoib_neigh *neigh; 1287 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1288 1289 while ((neigh = rcu_dereference_protected(*np, 1290 lockdep_is_held(&priv->lock))) != NULL) { 1291 /* was the neigh idle for two GC periods */ 1292 if (time_after(neigh_obsolete, neigh->alive)) { 1293 1294 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list); 1295 1296 rcu_assign_pointer(*np, 1297 rcu_dereference_protected(neigh->hnext, 1298 lockdep_is_held(&priv->lock))); 1299 /* remove from path/mc list */ 1300 list_del(&neigh->list); 1301 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1302 } else { 1303 np = &neigh->hnext; 1304 } 1305 1306 } 1307 } 1308 1309 out_unlock: 1310 spin_unlock_irqrestore(&priv->lock, flags); 1311 ipoib_mcast_remove_list(&remove_list); 1312 } 1313 1314 static void ipoib_reap_neigh(struct work_struct *work) 1315 { 1316 struct ipoib_dev_priv *priv = 1317 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work); 1318 1319 __ipoib_reap_neigh(priv); 1320 1321 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags)) 1322 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1323 arp_tbl.gc_interval); 1324 } 1325 1326 1327 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr, 1328 struct net_device *dev) 1329 { 1330 struct ipoib_neigh *neigh; 1331 1332 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC); 1333 if (!neigh) 1334 return NULL; 1335 1336 neigh->dev = dev; 1337 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr)); 1338 skb_queue_head_init(&neigh->queue); 1339 INIT_LIST_HEAD(&neigh->list); 1340 ipoib_cm_set(neigh, NULL); 1341 /* one ref on behalf of the caller */ 1342 atomic_set(&neigh->refcnt, 1); 1343 1344 return neigh; 1345 } 1346 1347 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr, 1348 struct net_device *dev) 1349 { 1350 struct ipoib_dev_priv *priv = netdev_priv(dev); 1351 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1352 struct ipoib_neigh_hash *htbl; 1353 struct ipoib_neigh *neigh; 1354 u32 hash_val; 1355 1356 htbl = rcu_dereference_protected(ntbl->htbl, 1357 lockdep_is_held(&priv->lock)); 1358 if (!htbl) { 1359 neigh = NULL; 1360 goto out_unlock; 1361 } 1362 1363 /* need to add a new neigh, but maybe some other thread succeeded? 1364 * recalc hash, maybe hash resize took place so we do a search 1365 */ 1366 hash_val = ipoib_addr_hash(htbl, daddr); 1367 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val], 1368 lockdep_is_held(&priv->lock)); 1369 neigh != NULL; 1370 neigh = rcu_dereference_protected(neigh->hnext, 1371 lockdep_is_held(&priv->lock))) { 1372 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) { 1373 /* found, take one ref on behalf of the caller */ 1374 if (!atomic_inc_not_zero(&neigh->refcnt)) { 1375 /* deleted */ 1376 neigh = NULL; 1377 break; 1378 } 1379 neigh->alive = jiffies; 1380 goto out_unlock; 1381 } 1382 } 1383 1384 neigh = ipoib_neigh_ctor(daddr, dev); 1385 if (!neigh) 1386 goto out_unlock; 1387 1388 /* one ref on behalf of the hash table */ 1389 atomic_inc(&neigh->refcnt); 1390 neigh->alive = jiffies; 1391 /* put in hash */ 1392 rcu_assign_pointer(neigh->hnext, 1393 rcu_dereference_protected(htbl->buckets[hash_val], 1394 lockdep_is_held(&priv->lock))); 1395 rcu_assign_pointer(htbl->buckets[hash_val], neigh); 1396 atomic_inc(&ntbl->entries); 1397 1398 out_unlock: 1399 1400 return neigh; 1401 } 1402 1403 void ipoib_neigh_dtor(struct ipoib_neigh *neigh) 1404 { 1405 /* neigh reference count was dropprd to zero */ 1406 struct net_device *dev = neigh->dev; 1407 struct ipoib_dev_priv *priv = netdev_priv(dev); 1408 struct sk_buff *skb; 1409 if (neigh->ah) 1410 ipoib_put_ah(neigh->ah); 1411 while ((skb = __skb_dequeue(&neigh->queue))) { 1412 ++dev->stats.tx_dropped; 1413 dev_kfree_skb_any(skb); 1414 } 1415 if (ipoib_cm_get(neigh)) 1416 ipoib_cm_destroy_tx(ipoib_cm_get(neigh)); 1417 ipoib_dbg(netdev_priv(dev), 1418 "neigh free for %06x %pI6\n", 1419 IPOIB_QPN(neigh->daddr), 1420 neigh->daddr + 4); 1421 kfree(neigh); 1422 if (atomic_dec_and_test(&priv->ntbl.entries)) { 1423 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags)) 1424 complete(&priv->ntbl.flushed); 1425 } 1426 } 1427 1428 static void ipoib_neigh_reclaim(struct rcu_head *rp) 1429 { 1430 /* Called as a result of removal from hash table */ 1431 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu); 1432 /* note TX context may hold another ref */ 1433 ipoib_neigh_put(neigh); 1434 } 1435 1436 void ipoib_neigh_free(struct ipoib_neigh *neigh) 1437 { 1438 struct net_device *dev = neigh->dev; 1439 struct ipoib_dev_priv *priv = netdev_priv(dev); 1440 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1441 struct ipoib_neigh_hash *htbl; 1442 struct ipoib_neigh __rcu **np; 1443 struct ipoib_neigh *n; 1444 u32 hash_val; 1445 1446 htbl = rcu_dereference_protected(ntbl->htbl, 1447 lockdep_is_held(&priv->lock)); 1448 if (!htbl) 1449 return; 1450 1451 hash_val = ipoib_addr_hash(htbl, neigh->daddr); 1452 np = &htbl->buckets[hash_val]; 1453 for (n = rcu_dereference_protected(*np, 1454 lockdep_is_held(&priv->lock)); 1455 n != NULL; 1456 n = rcu_dereference_protected(*np, 1457 lockdep_is_held(&priv->lock))) { 1458 if (n == neigh) { 1459 /* found */ 1460 rcu_assign_pointer(*np, 1461 rcu_dereference_protected(neigh->hnext, 1462 lockdep_is_held(&priv->lock))); 1463 /* remove from parent list */ 1464 list_del(&neigh->list); 1465 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1466 return; 1467 } else { 1468 np = &n->hnext; 1469 } 1470 } 1471 } 1472 1473 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv) 1474 { 1475 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1476 struct ipoib_neigh_hash *htbl; 1477 struct ipoib_neigh __rcu **buckets; 1478 u32 size; 1479 1480 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1481 ntbl->htbl = NULL; 1482 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL); 1483 if (!htbl) 1484 return -ENOMEM; 1485 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1486 size = roundup_pow_of_two(arp_tbl.gc_thresh3); 1487 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL); 1488 if (!buckets) { 1489 kfree(htbl); 1490 return -ENOMEM; 1491 } 1492 htbl->size = size; 1493 htbl->mask = (size - 1); 1494 htbl->buckets = buckets; 1495 RCU_INIT_POINTER(ntbl->htbl, htbl); 1496 htbl->ntbl = ntbl; 1497 atomic_set(&ntbl->entries, 0); 1498 1499 /* start garbage collection */ 1500 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1501 queue_delayed_work(priv->wq, &priv->neigh_reap_task, 1502 arp_tbl.gc_interval); 1503 1504 return 0; 1505 } 1506 1507 static void neigh_hash_free_rcu(struct rcu_head *head) 1508 { 1509 struct ipoib_neigh_hash *htbl = container_of(head, 1510 struct ipoib_neigh_hash, 1511 rcu); 1512 struct ipoib_neigh __rcu **buckets = htbl->buckets; 1513 struct ipoib_neigh_table *ntbl = htbl->ntbl; 1514 1515 kfree(buckets); 1516 kfree(htbl); 1517 complete(&ntbl->deleted); 1518 } 1519 1520 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid) 1521 { 1522 struct ipoib_dev_priv *priv = netdev_priv(dev); 1523 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1524 struct ipoib_neigh_hash *htbl; 1525 unsigned long flags; 1526 int i; 1527 1528 /* remove all neigh connected to a given path or mcast */ 1529 spin_lock_irqsave(&priv->lock, flags); 1530 1531 htbl = rcu_dereference_protected(ntbl->htbl, 1532 lockdep_is_held(&priv->lock)); 1533 1534 if (!htbl) 1535 goto out_unlock; 1536 1537 for (i = 0; i < htbl->size; i++) { 1538 struct ipoib_neigh *neigh; 1539 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1540 1541 while ((neigh = rcu_dereference_protected(*np, 1542 lockdep_is_held(&priv->lock))) != NULL) { 1543 /* delete neighs belong to this parent */ 1544 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) { 1545 rcu_assign_pointer(*np, 1546 rcu_dereference_protected(neigh->hnext, 1547 lockdep_is_held(&priv->lock))); 1548 /* remove from parent list */ 1549 list_del(&neigh->list); 1550 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1551 } else { 1552 np = &neigh->hnext; 1553 } 1554 1555 } 1556 } 1557 out_unlock: 1558 spin_unlock_irqrestore(&priv->lock, flags); 1559 } 1560 1561 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv) 1562 { 1563 struct ipoib_neigh_table *ntbl = &priv->ntbl; 1564 struct ipoib_neigh_hash *htbl; 1565 unsigned long flags; 1566 int i, wait_flushed = 0; 1567 1568 init_completion(&priv->ntbl.flushed); 1569 1570 spin_lock_irqsave(&priv->lock, flags); 1571 1572 htbl = rcu_dereference_protected(ntbl->htbl, 1573 lockdep_is_held(&priv->lock)); 1574 if (!htbl) 1575 goto out_unlock; 1576 1577 wait_flushed = atomic_read(&priv->ntbl.entries); 1578 if (!wait_flushed) 1579 goto free_htbl; 1580 1581 for (i = 0; i < htbl->size; i++) { 1582 struct ipoib_neigh *neigh; 1583 struct ipoib_neigh __rcu **np = &htbl->buckets[i]; 1584 1585 while ((neigh = rcu_dereference_protected(*np, 1586 lockdep_is_held(&priv->lock))) != NULL) { 1587 rcu_assign_pointer(*np, 1588 rcu_dereference_protected(neigh->hnext, 1589 lockdep_is_held(&priv->lock))); 1590 /* remove from path/mc list */ 1591 list_del(&neigh->list); 1592 call_rcu(&neigh->rcu, ipoib_neigh_reclaim); 1593 } 1594 } 1595 1596 free_htbl: 1597 rcu_assign_pointer(ntbl->htbl, NULL); 1598 call_rcu(&htbl->rcu, neigh_hash_free_rcu); 1599 1600 out_unlock: 1601 spin_unlock_irqrestore(&priv->lock, flags); 1602 if (wait_flushed) 1603 wait_for_completion(&priv->ntbl.flushed); 1604 } 1605 1606 static void ipoib_neigh_hash_uninit(struct net_device *dev) 1607 { 1608 struct ipoib_dev_priv *priv = netdev_priv(dev); 1609 int stopped; 1610 1611 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n"); 1612 init_completion(&priv->ntbl.deleted); 1613 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags); 1614 1615 /* Stop GC if called at init fail need to cancel work */ 1616 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 1617 if (!stopped) 1618 cancel_delayed_work(&priv->neigh_reap_task); 1619 1620 ipoib_flush_neighs(priv); 1621 1622 wait_for_completion(&priv->ntbl.deleted); 1623 } 1624 1625 1626 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port) 1627 { 1628 struct ipoib_dev_priv *priv = netdev_priv(dev); 1629 1630 /* Allocate RX/TX "rings" to hold queued skbs */ 1631 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring, 1632 GFP_KERNEL); 1633 if (!priv->rx_ring) 1634 goto out; 1635 1636 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring); 1637 if (!priv->tx_ring) { 1638 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n", 1639 ca->name, ipoib_sendq_size); 1640 goto out_rx_ring_cleanup; 1641 } 1642 1643 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */ 1644 1645 if (ipoib_ib_dev_init(dev, ca, port)) 1646 goto out_tx_ring_cleanup; 1647 1648 /* 1649 * Must be after ipoib_ib_dev_init so we can allocate a per 1650 * device wq there and use it here 1651 */ 1652 if (ipoib_neigh_hash_init(priv) < 0) 1653 goto out_dev_uninit; 1654 1655 return 0; 1656 1657 out_dev_uninit: 1658 ipoib_ib_dev_cleanup(dev); 1659 1660 out_tx_ring_cleanup: 1661 vfree(priv->tx_ring); 1662 1663 out_rx_ring_cleanup: 1664 kfree(priv->rx_ring); 1665 1666 out: 1667 return -ENOMEM; 1668 } 1669 1670 void ipoib_dev_cleanup(struct net_device *dev) 1671 { 1672 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv; 1673 LIST_HEAD(head); 1674 1675 ASSERT_RTNL(); 1676 1677 ipoib_delete_debug_files(dev); 1678 1679 /* Delete any child interfaces first */ 1680 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) { 1681 /* Stop GC on child */ 1682 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags); 1683 cancel_delayed_work(&cpriv->neigh_reap_task); 1684 unregister_netdevice_queue(cpriv->dev, &head); 1685 } 1686 unregister_netdevice_many(&head); 1687 1688 /* 1689 * Must be before ipoib_ib_dev_cleanup or we delete an in use 1690 * work queue 1691 */ 1692 ipoib_neigh_hash_uninit(dev); 1693 1694 ipoib_ib_dev_cleanup(dev); 1695 1696 kfree(priv->rx_ring); 1697 vfree(priv->tx_ring); 1698 1699 priv->rx_ring = NULL; 1700 priv->tx_ring = NULL; 1701 } 1702 1703 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state) 1704 { 1705 struct ipoib_dev_priv *priv = netdev_priv(dev); 1706 1707 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state); 1708 } 1709 1710 static int ipoib_get_vf_config(struct net_device *dev, int vf, 1711 struct ifla_vf_info *ivf) 1712 { 1713 struct ipoib_dev_priv *priv = netdev_priv(dev); 1714 int err; 1715 1716 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf); 1717 if (err) 1718 return err; 1719 1720 ivf->vf = vf; 1721 1722 return 0; 1723 } 1724 1725 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type) 1726 { 1727 struct ipoib_dev_priv *priv = netdev_priv(dev); 1728 1729 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID) 1730 return -EINVAL; 1731 1732 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type); 1733 } 1734 1735 static int ipoib_get_vf_stats(struct net_device *dev, int vf, 1736 struct ifla_vf_stats *vf_stats) 1737 { 1738 struct ipoib_dev_priv *priv = netdev_priv(dev); 1739 1740 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats); 1741 } 1742 1743 static const struct header_ops ipoib_header_ops = { 1744 .create = ipoib_hard_header, 1745 }; 1746 1747 static const struct net_device_ops ipoib_netdev_ops_pf = { 1748 .ndo_uninit = ipoib_uninit, 1749 .ndo_open = ipoib_open, 1750 .ndo_stop = ipoib_stop, 1751 .ndo_change_mtu = ipoib_change_mtu, 1752 .ndo_fix_features = ipoib_fix_features, 1753 .ndo_start_xmit = ipoib_start_xmit, 1754 .ndo_tx_timeout = ipoib_timeout, 1755 .ndo_set_rx_mode = ipoib_set_mcast_list, 1756 .ndo_get_iflink = ipoib_get_iflink, 1757 .ndo_set_vf_link_state = ipoib_set_vf_link_state, 1758 .ndo_get_vf_config = ipoib_get_vf_config, 1759 .ndo_get_vf_stats = ipoib_get_vf_stats, 1760 .ndo_set_vf_guid = ipoib_set_vf_guid, 1761 .ndo_set_mac_address = ipoib_set_mac, 1762 }; 1763 1764 static const struct net_device_ops ipoib_netdev_ops_vf = { 1765 .ndo_uninit = ipoib_uninit, 1766 .ndo_open = ipoib_open, 1767 .ndo_stop = ipoib_stop, 1768 .ndo_change_mtu = ipoib_change_mtu, 1769 .ndo_fix_features = ipoib_fix_features, 1770 .ndo_start_xmit = ipoib_start_xmit, 1771 .ndo_tx_timeout = ipoib_timeout, 1772 .ndo_set_rx_mode = ipoib_set_mcast_list, 1773 .ndo_get_iflink = ipoib_get_iflink, 1774 }; 1775 1776 void ipoib_setup(struct net_device *dev) 1777 { 1778 struct ipoib_dev_priv *priv = netdev_priv(dev); 1779 1780 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION) 1781 dev->netdev_ops = &ipoib_netdev_ops_vf; 1782 else 1783 dev->netdev_ops = &ipoib_netdev_ops_pf; 1784 1785 dev->header_ops = &ipoib_header_ops; 1786 1787 ipoib_set_ethtool_ops(dev); 1788 1789 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT); 1790 1791 dev->watchdog_timeo = HZ; 1792 1793 dev->flags |= IFF_BROADCAST | IFF_MULTICAST; 1794 1795 dev->hard_header_len = IPOIB_HARD_LEN; 1796 dev->addr_len = INFINIBAND_ALEN; 1797 dev->type = ARPHRD_INFINIBAND; 1798 dev->tx_queue_len = ipoib_sendq_size * 2; 1799 dev->features = (NETIF_F_VLAN_CHALLENGED | 1800 NETIF_F_HIGHDMA); 1801 netif_keep_dst(dev); 1802 1803 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN); 1804 1805 priv->dev = dev; 1806 1807 spin_lock_init(&priv->lock); 1808 1809 init_rwsem(&priv->vlan_rwsem); 1810 1811 INIT_LIST_HEAD(&priv->path_list); 1812 INIT_LIST_HEAD(&priv->child_intfs); 1813 INIT_LIST_HEAD(&priv->dead_ahs); 1814 INIT_LIST_HEAD(&priv->multicast_list); 1815 1816 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task); 1817 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task); 1818 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light); 1819 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal); 1820 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy); 1821 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task); 1822 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah); 1823 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh); 1824 } 1825 1826 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name) 1827 { 1828 struct net_device *dev; 1829 1830 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name, 1831 NET_NAME_UNKNOWN, ipoib_setup); 1832 if (!dev) 1833 return NULL; 1834 1835 return netdev_priv(dev); 1836 } 1837 1838 static ssize_t show_pkey(struct device *dev, 1839 struct device_attribute *attr, char *buf) 1840 { 1841 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev)); 1842 1843 return sprintf(buf, "0x%04x\n", priv->pkey); 1844 } 1845 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL); 1846 1847 static ssize_t show_umcast(struct device *dev, 1848 struct device_attribute *attr, char *buf) 1849 { 1850 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev)); 1851 1852 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags)); 1853 } 1854 1855 void ipoib_set_umcast(struct net_device *ndev, int umcast_val) 1856 { 1857 struct ipoib_dev_priv *priv = netdev_priv(ndev); 1858 1859 if (umcast_val > 0) { 1860 set_bit(IPOIB_FLAG_UMCAST, &priv->flags); 1861 ipoib_warn(priv, "ignoring multicast groups joined directly " 1862 "by userspace\n"); 1863 } else 1864 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags); 1865 } 1866 1867 static ssize_t set_umcast(struct device *dev, 1868 struct device_attribute *attr, 1869 const char *buf, size_t count) 1870 { 1871 unsigned long umcast_val = simple_strtoul(buf, NULL, 0); 1872 1873 ipoib_set_umcast(to_net_dev(dev), umcast_val); 1874 1875 return count; 1876 } 1877 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast); 1878 1879 int ipoib_add_umcast_attr(struct net_device *dev) 1880 { 1881 return device_create_file(&dev->dev, &dev_attr_umcast); 1882 } 1883 1884 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid) 1885 { 1886 struct ipoib_dev_priv *child_priv; 1887 struct net_device *netdev = priv->dev; 1888 1889 netif_addr_lock_bh(netdev); 1890 1891 memcpy(&priv->local_gid.global.interface_id, 1892 &gid->global.interface_id, 1893 sizeof(gid->global.interface_id)); 1894 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid)); 1895 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 1896 1897 netif_addr_unlock_bh(netdev); 1898 1899 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 1900 down_read(&priv->vlan_rwsem); 1901 list_for_each_entry(child_priv, &priv->child_intfs, list) 1902 set_base_guid(child_priv, gid); 1903 up_read(&priv->vlan_rwsem); 1904 } 1905 } 1906 1907 static int ipoib_check_lladdr(struct net_device *dev, 1908 struct sockaddr_storage *ss) 1909 { 1910 union ib_gid *gid = (union ib_gid *)(ss->__data + 4); 1911 int ret = 0; 1912 1913 netif_addr_lock_bh(dev); 1914 1915 /* Make sure the QPN, reserved and subnet prefix match the current 1916 * lladdr, it also makes sure the lladdr is unicast. 1917 */ 1918 if (memcmp(dev->dev_addr, ss->__data, 1919 4 + sizeof(gid->global.subnet_prefix)) || 1920 gid->global.interface_id == 0) 1921 ret = -EINVAL; 1922 1923 netif_addr_unlock_bh(dev); 1924 1925 return ret; 1926 } 1927 1928 static int ipoib_set_mac(struct net_device *dev, void *addr) 1929 { 1930 struct ipoib_dev_priv *priv = netdev_priv(dev); 1931 struct sockaddr_storage *ss = addr; 1932 int ret; 1933 1934 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev)) 1935 return -EBUSY; 1936 1937 ret = ipoib_check_lladdr(dev, ss); 1938 if (ret) 1939 return ret; 1940 1941 set_base_guid(priv, (union ib_gid *)(ss->__data + 4)); 1942 1943 queue_work(ipoib_workqueue, &priv->flush_light); 1944 1945 return 0; 1946 } 1947 1948 static ssize_t create_child(struct device *dev, 1949 struct device_attribute *attr, 1950 const char *buf, size_t count) 1951 { 1952 int pkey; 1953 int ret; 1954 1955 if (sscanf(buf, "%i", &pkey) != 1) 1956 return -EINVAL; 1957 1958 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000) 1959 return -EINVAL; 1960 1961 /* 1962 * Set the full membership bit, so that we join the right 1963 * broadcast group, etc. 1964 */ 1965 pkey |= 0x8000; 1966 1967 ret = ipoib_vlan_add(to_net_dev(dev), pkey); 1968 1969 return ret ? ret : count; 1970 } 1971 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child); 1972 1973 static ssize_t delete_child(struct device *dev, 1974 struct device_attribute *attr, 1975 const char *buf, size_t count) 1976 { 1977 int pkey; 1978 int ret; 1979 1980 if (sscanf(buf, "%i", &pkey) != 1) 1981 return -EINVAL; 1982 1983 if (pkey < 0 || pkey > 0xffff) 1984 return -EINVAL; 1985 1986 ret = ipoib_vlan_delete(to_net_dev(dev), pkey); 1987 1988 return ret ? ret : count; 1989 1990 } 1991 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child); 1992 1993 int ipoib_add_pkey_attr(struct net_device *dev) 1994 { 1995 return device_create_file(&dev->dev, &dev_attr_pkey); 1996 } 1997 1998 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca) 1999 { 2000 priv->hca_caps = hca->attrs.device_cap_flags; 2001 2002 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) { 2003 priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM; 2004 2005 if (priv->hca_caps & IB_DEVICE_UD_TSO) 2006 priv->dev->hw_features |= NETIF_F_TSO; 2007 2008 priv->dev->features |= priv->dev->hw_features; 2009 } 2010 2011 return 0; 2012 } 2013 2014 static struct net_device *ipoib_add_port(const char *format, 2015 struct ib_device *hca, u8 port) 2016 { 2017 struct ipoib_dev_priv *priv; 2018 struct ib_port_attr attr; 2019 int result = -ENOMEM; 2020 2021 priv = ipoib_intf_alloc(format); 2022 if (!priv) 2023 goto alloc_mem_failed; 2024 2025 SET_NETDEV_DEV(priv->dev, hca->dma_device); 2026 priv->dev->dev_id = port - 1; 2027 2028 result = ib_query_port(hca, port, &attr); 2029 if (!result) 2030 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu); 2031 else { 2032 printk(KERN_WARNING "%s: ib_query_port %d failed\n", 2033 hca->name, port); 2034 goto device_init_failed; 2035 } 2036 2037 /* MTU will be reset when mcast join happens */ 2038 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu); 2039 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu; 2040 priv->dev->max_mtu = IPOIB_CM_MTU; 2041 2042 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh); 2043 2044 result = ib_query_pkey(hca, port, 0, &priv->pkey); 2045 if (result) { 2046 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n", 2047 hca->name, port, result); 2048 goto device_init_failed; 2049 } 2050 2051 result = ipoib_set_dev_features(priv, hca); 2052 if (result) 2053 goto device_init_failed; 2054 2055 /* 2056 * Set the full membership bit, so that we join the right 2057 * broadcast group, etc. 2058 */ 2059 priv->pkey |= 0x8000; 2060 2061 priv->dev->broadcast[8] = priv->pkey >> 8; 2062 priv->dev->broadcast[9] = priv->pkey & 0xff; 2063 2064 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL); 2065 if (result) { 2066 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n", 2067 hca->name, port, result); 2068 goto device_init_failed; 2069 } else 2070 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid)); 2071 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 2072 2073 result = ipoib_dev_init(priv->dev, hca, port); 2074 if (result < 0) { 2075 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n", 2076 hca->name, port, result); 2077 goto device_init_failed; 2078 } 2079 2080 INIT_IB_EVENT_HANDLER(&priv->event_handler, 2081 priv->ca, ipoib_event); 2082 result = ib_register_event_handler(&priv->event_handler); 2083 if (result < 0) { 2084 printk(KERN_WARNING "%s: ib_register_event_handler failed for " 2085 "port %d (ret = %d)\n", 2086 hca->name, port, result); 2087 goto event_failed; 2088 } 2089 2090 result = register_netdev(priv->dev); 2091 if (result) { 2092 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n", 2093 hca->name, port, result); 2094 goto register_failed; 2095 } 2096 2097 ipoib_create_debug_files(priv->dev); 2098 2099 if (ipoib_cm_add_mode_attr(priv->dev)) 2100 goto sysfs_failed; 2101 if (ipoib_add_pkey_attr(priv->dev)) 2102 goto sysfs_failed; 2103 if (ipoib_add_umcast_attr(priv->dev)) 2104 goto sysfs_failed; 2105 if (device_create_file(&priv->dev->dev, &dev_attr_create_child)) 2106 goto sysfs_failed; 2107 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child)) 2108 goto sysfs_failed; 2109 2110 return priv->dev; 2111 2112 sysfs_failed: 2113 ipoib_delete_debug_files(priv->dev); 2114 unregister_netdev(priv->dev); 2115 2116 register_failed: 2117 ib_unregister_event_handler(&priv->event_handler); 2118 flush_workqueue(ipoib_workqueue); 2119 /* Stop GC if started before flush */ 2120 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2121 cancel_delayed_work(&priv->neigh_reap_task); 2122 flush_workqueue(priv->wq); 2123 2124 event_failed: 2125 ipoib_dev_cleanup(priv->dev); 2126 2127 device_init_failed: 2128 free_netdev(priv->dev); 2129 2130 alloc_mem_failed: 2131 return ERR_PTR(result); 2132 } 2133 2134 static void ipoib_add_one(struct ib_device *device) 2135 { 2136 struct list_head *dev_list; 2137 struct net_device *dev; 2138 struct ipoib_dev_priv *priv; 2139 int p; 2140 int count = 0; 2141 2142 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL); 2143 if (!dev_list) 2144 return; 2145 2146 INIT_LIST_HEAD(dev_list); 2147 2148 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) { 2149 if (!rdma_protocol_ib(device, p)) 2150 continue; 2151 dev = ipoib_add_port("ib%d", device, p); 2152 if (!IS_ERR(dev)) { 2153 priv = netdev_priv(dev); 2154 list_add_tail(&priv->list, dev_list); 2155 count++; 2156 } 2157 } 2158 2159 if (!count) { 2160 kfree(dev_list); 2161 return; 2162 } 2163 2164 ib_set_client_data(device, &ipoib_client, dev_list); 2165 } 2166 2167 static void ipoib_remove_one(struct ib_device *device, void *client_data) 2168 { 2169 struct ipoib_dev_priv *priv, *tmp; 2170 struct list_head *dev_list = client_data; 2171 2172 if (!dev_list) 2173 return; 2174 2175 list_for_each_entry_safe(priv, tmp, dev_list, list) { 2176 ib_unregister_event_handler(&priv->event_handler); 2177 flush_workqueue(ipoib_workqueue); 2178 2179 /* mark interface in the middle of destruction */ 2180 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags); 2181 2182 rtnl_lock(); 2183 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); 2184 rtnl_unlock(); 2185 2186 /* Stop GC */ 2187 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags); 2188 cancel_delayed_work(&priv->neigh_reap_task); 2189 flush_workqueue(priv->wq); 2190 2191 unregister_netdev(priv->dev); 2192 free_netdev(priv->dev); 2193 } 2194 2195 kfree(dev_list); 2196 } 2197 2198 static int __init ipoib_init_module(void) 2199 { 2200 int ret; 2201 2202 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size); 2203 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE); 2204 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE); 2205 2206 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size); 2207 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE); 2208 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE); 2209 #ifdef CONFIG_INFINIBAND_IPOIB_CM 2210 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP); 2211 #endif 2212 2213 /* 2214 * When copying small received packets, we only copy from the 2215 * linear data part of the SKB, so we rely on this condition. 2216 */ 2217 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE); 2218 2219 ret = ipoib_register_debugfs(); 2220 if (ret) 2221 return ret; 2222 2223 /* 2224 * We create a global workqueue here that is used for all flush 2225 * operations. However, if you attempt to flush a workqueue 2226 * from a task on that same workqueue, it deadlocks the system. 2227 * We want to be able to flush the tasks associated with a 2228 * specific net device, so we also create a workqueue for each 2229 * netdevice. We queue up the tasks for that device only on 2230 * its private workqueue, and we only queue up flush events 2231 * on our global flush workqueue. This avoids the deadlocks. 2232 */ 2233 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 2234 WQ_MEM_RECLAIM); 2235 if (!ipoib_workqueue) { 2236 ret = -ENOMEM; 2237 goto err_fs; 2238 } 2239 2240 ib_sa_register_client(&ipoib_sa_client); 2241 2242 ret = ib_register_client(&ipoib_client); 2243 if (ret) 2244 goto err_sa; 2245 2246 ret = ipoib_netlink_init(); 2247 if (ret) 2248 goto err_client; 2249 2250 return 0; 2251 2252 err_client: 2253 ib_unregister_client(&ipoib_client); 2254 2255 err_sa: 2256 ib_sa_unregister_client(&ipoib_sa_client); 2257 destroy_workqueue(ipoib_workqueue); 2258 2259 err_fs: 2260 ipoib_unregister_debugfs(); 2261 2262 return ret; 2263 } 2264 2265 static void __exit ipoib_cleanup_module(void) 2266 { 2267 ipoib_netlink_fini(); 2268 ib_unregister_client(&ipoib_client); 2269 ib_sa_unregister_client(&ipoib_sa_client); 2270 ipoib_unregister_debugfs(); 2271 destroy_workqueue(ipoib_workqueue); 2272 } 2273 2274 module_init(ipoib_init_module); 2275 module_exit(ipoib_cleanup_module); 2276