drivers/infiniband/core/roce_gid_mgmt.c

   1 /*
   2  * Copyright (c) 2015, Mellanox Technologies inc.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include "core_priv.h"
  34
  35 #include <linux/in.h>
  36 #include <linux/in6.h>
  37
  38 /* For in6_dev_get/in6_dev_put */
  39 #include <net/addrconf.h>
  40 #include <net/bonding.h>
  41
  42 #include <rdma/ib_cache.h>
  43 #include <rdma/ib_addr.h>
  44
  45 enum gid_op_type {
  46         GID_DEL = 0,
  47         GID_ADD
  48 };
  49
  50 struct update_gid_event_work {
  51         struct work_struct work;
  52         union ib_gid       gid;
  53         struct ib_gid_attr gid_attr;
  54         enum gid_op_type gid_op;
  55 };
  56
  57 #define ROCE_NETDEV_CALLBACK_SZ         3
  58 struct netdev_event_work_cmd {
  59         roce_netdev_callback    cb;
  60         roce_netdev_filter      filter;
  61         struct net_device       *ndev;
  62         struct net_device       *filter_ndev;
  63 };
  64
  65 struct netdev_event_work {
  66         struct work_struct              work;
  67         struct netdev_event_work_cmd    cmds[ROCE_NETDEV_CALLBACK_SZ];
  68 };
  69
  70 static const struct {
  71         bool (*is_supported)(const struct ib_device *device, u8 port_num);
  72         enum ib_gid_type gid_type;
  73 } PORT_CAP_TO_GID_TYPE[] = {
  74         {rdma_protocol_roce_eth_encap, IB_GID_TYPE_ROCE},
  75         {rdma_protocol_roce_udp_encap, IB_GID_TYPE_ROCE_UDP_ENCAP},
  76 };
  77
  78 #define CAP_TO_GID_TABLE_SIZE   ARRAY_SIZE(PORT_CAP_TO_GID_TYPE)
  79
  80 unsigned long roce_gid_type_mask_support(struct ib_device *ib_dev, u8 port)
  81 {
  82         int i;
  83         unsigned int ret_flags = 0;
  84
  85         if (!rdma_protocol_roce(ib_dev, port))
  86                 return 1UL << IB_GID_TYPE_IB;
  87
  88         for (i = 0; i < CAP_TO_GID_TABLE_SIZE; i++)
  89                 if (PORT_CAP_TO_GID_TYPE[i].is_supported(ib_dev, port))
  90                         ret_flags |= 1UL << PORT_CAP_TO_GID_TYPE[i].gid_type;
  91
  92         return ret_flags;
  93 }
  94 EXPORT_SYMBOL(roce_gid_type_mask_support);
  95
  96 static void update_gid(enum gid_op_type gid_op, struct ib_device *ib_dev,
  97                        u8 port, union ib_gid *gid,
  98                        struct ib_gid_attr *gid_attr)
  99 {
 100         int i;
 101         unsigned long gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 102
 103         for (i = 0; i < IB_GID_TYPE_SIZE; i++) {
 104                 if ((1UL << i) & gid_type_mask) {
 105                         gid_attr->gid_type = i;
 106                         switch (gid_op) {
 107                         case GID_ADD:
 108                                 ib_cache_gid_add(ib_dev, port,
 109                                                  gid, gid_attr);
 110                                 break;
 111                         case GID_DEL:
 112                                 ib_cache_gid_del(ib_dev, port,
 113                                                  gid, gid_attr);
 114                                 break;
 115                         }
 116                 }
 117         }
 118 }
 119
 120 enum bonding_slave_state {
 121         BONDING_SLAVE_STATE_ACTIVE      = 1UL << 0,
 122         BONDING_SLAVE_STATE_INACTIVE    = 1UL << 1,
 123         /* No primary slave or the device isn't a slave in bonding */
 124         BONDING_SLAVE_STATE_NA          = 1UL << 2,
 125 };
 126
 127 static enum bonding_slave_state is_eth_active_slave_of_bonding_rcu(struct net_device *dev,
 128                                                                    struct net_device *upper)
 129 {
 130         if (upper && netif_is_bond_master(upper)) {
 131                 struct net_device *pdev =
 132                         bond_option_active_slave_get_rcu(netdev_priv(upper));
 133
 134                 if (pdev)
 135                         return dev == pdev ? BONDING_SLAVE_STATE_ACTIVE :
 136                                 BONDING_SLAVE_STATE_INACTIVE;
 137         }
 138
 139         return BONDING_SLAVE_STATE_NA;
 140 }
 141
 142 static bool is_upper_dev_rcu(struct net_device *dev, struct net_device *upper)
 143 {
 144         struct net_device *_upper = NULL;
 145         struct list_head *iter;
 146
 147         netdev_for_each_all_upper_dev_rcu(dev, _upper, iter)
 148                 if (_upper == upper)
 149                         break;
 150
 151         return _upper == upper;
 152 }
 153
 154 #define REQUIRED_BOND_STATES            (BONDING_SLAVE_STATE_ACTIVE |   \
 155                                          BONDING_SLAVE_STATE_NA)
 156 static int is_eth_port_of_netdev(struct ib_device *ib_dev, u8 port,
 157                                  struct net_device *rdma_ndev, void *cookie)
 158 {
 159         struct net_device *event_ndev = (struct net_device *)cookie;
 160         struct net_device *real_dev;
 161         int res;
 162
 163         if (!rdma_ndev)
 164                 return 0;
 165
 166         rcu_read_lock();
 167         real_dev = rdma_vlan_dev_real_dev(event_ndev);
 168         if (!real_dev)
 169                 real_dev = event_ndev;
 170
 171         res = ((is_upper_dev_rcu(rdma_ndev, event_ndev) &&
 172                (is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) &
 173                 REQUIRED_BOND_STATES)) ||
 174                real_dev == rdma_ndev);
 175
 176         rcu_read_unlock();
 177         return res;
 178 }
 179
 180 static int is_eth_port_inactive_slave(struct ib_device *ib_dev, u8 port,
 181                                       struct net_device *rdma_ndev, void *cookie)
 182 {
 183         struct net_device *master_dev;
 184         int res;
 185
 186         if (!rdma_ndev)
 187                 return 0;
 188
 189         rcu_read_lock();
 190         master_dev = netdev_master_upper_dev_get_rcu(rdma_ndev);
 191         res = is_eth_active_slave_of_bonding_rcu(rdma_ndev, master_dev) ==
 192                 BONDING_SLAVE_STATE_INACTIVE;
 193         rcu_read_unlock();
 194
 195         return res;
 196 }
 197
 198 static int pass_all_filter(struct ib_device *ib_dev, u8 port,
 199                            struct net_device *rdma_ndev, void *cookie)
 200 {
 201         return 1;
 202 }
 203
 204 static int upper_device_filter(struct ib_device *ib_dev, u8 port,
 205                                struct net_device *rdma_ndev, void *cookie)
 206 {
 207         struct net_device *event_ndev = (struct net_device *)cookie;
 208         int res;
 209
 210         if (!rdma_ndev)
 211                 return 0;
 212
 213         if (rdma_ndev == event_ndev)
 214                 return 1;
 215
 216         rcu_read_lock();
 217         res = is_upper_dev_rcu(rdma_ndev, event_ndev);
 218         rcu_read_unlock();
 219
 220         return res;
 221 }
 222
 223 static void update_gid_ip(enum gid_op_type gid_op,
 224                           struct ib_device *ib_dev,
 225                           u8 port, struct net_device *ndev,
 226                           struct sockaddr *addr)
 227 {
 228         union ib_gid gid;
 229         struct ib_gid_attr gid_attr;
 230
 231         rdma_ip2gid(addr, &gid);
 232         memset(&gid_attr, 0, sizeof(gid_attr));
 233         gid_attr.ndev = ndev;
 234
 235         update_gid(gid_op, ib_dev, port, &gid, &gid_attr);
 236 }
 237
 238 static void enum_netdev_default_gids(struct ib_device *ib_dev,
 239                                      u8 port, struct net_device *event_ndev,
 240                                      struct net_device *rdma_ndev)
 241 {
 242         unsigned long gid_type_mask;
 243
 244         rcu_read_lock();
 245         if (!rdma_ndev ||
 246             ((rdma_ndev != event_ndev &&
 247               !is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
 248              is_eth_active_slave_of_bonding_rcu(rdma_ndev,
 249                                                 netdev_master_upper_dev_get_rcu(rdma_ndev)) ==
 250              BONDING_SLAVE_STATE_INACTIVE)) {
 251                 rcu_read_unlock();
 252                 return;
 253         }
 254         rcu_read_unlock();
 255
 256         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 257
 258         ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev, gid_type_mask,
 259                                      IB_CACHE_GID_DEFAULT_MODE_SET);
 260 }
 261
 262 static void bond_delete_netdev_default_gids(struct ib_device *ib_dev,
 263                                             u8 port,
 264                                             struct net_device *event_ndev,
 265                                             struct net_device *rdma_ndev)
 266 {
 267         struct net_device *real_dev = rdma_vlan_dev_real_dev(event_ndev);
 268
 269         if (!rdma_ndev)
 270                 return;
 271
 272         if (!real_dev)
 273                 real_dev = event_ndev;
 274
 275         rcu_read_lock();
 276
 277         if (is_upper_dev_rcu(rdma_ndev, event_ndev) &&
 278             is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) ==
 279             BONDING_SLAVE_STATE_INACTIVE) {
 280                 unsigned long gid_type_mask;
 281
 282                 rcu_read_unlock();
 283
 284                 gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 285
 286                 ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
 287                                              gid_type_mask,
 288                                              IB_CACHE_GID_DEFAULT_MODE_DELETE);
 289         } else {
 290                 rcu_read_unlock();
 291         }
 292 }
 293
 294 static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
 295                                  u8 port, struct net_device *ndev)
 296 {
 297         struct in_device *in_dev;
 298         struct sin_list {
 299                 struct list_head        list;
 300                 struct sockaddr_in      ip;
 301         };
 302         struct sin_list *sin_iter;
 303         struct sin_list *sin_temp;
 304
 305         LIST_HEAD(sin_list);
 306         if (ndev->reg_state >= NETREG_UNREGISTERING)
 307                 return;
 308
 309         rcu_read_lock();
 310         in_dev = __in_dev_get_rcu(ndev);
 311         if (!in_dev) {
 312                 rcu_read_unlock();
 313                 return;
 314         }
 315
 316         for_ifa(in_dev) {
 317                 struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 318
 319                 if (!entry) {
 320                         pr_warn("roce_gid_mgmt: couldn't allocate entry for IPv4 update\n");
 321                         continue;
 322                 }
 323                 entry->ip.sin_family = AF_INET;
 324                 entry->ip.sin_addr.s_addr = ifa->ifa_address;
 325                 list_add_tail(&entry->list, &sin_list);
 326         }
 327         endfor_ifa(in_dev);
 328         rcu_read_unlock();
 329
 330         list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
 331                 update_gid_ip(GID_ADD, ib_dev, port, ndev,
 332                               (struct sockaddr *)&sin_iter->ip);
 333                 list_del(&sin_iter->list);
 334                 kfree(sin_iter);
 335         }
 336 }
 337
 338 static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
 339                                  u8 port, struct net_device *ndev)
 340 {
 341         struct inet6_ifaddr *ifp;
 342         struct inet6_dev *in6_dev;
 343         struct sin6_list {
 344                 struct list_head        list;
 345                 struct sockaddr_in6     sin6;
 346         };
 347         struct sin6_list *sin6_iter;
 348         struct sin6_list *sin6_temp;
 349         struct ib_gid_attr gid_attr = {.ndev = ndev};
 350         LIST_HEAD(sin6_list);
 351
 352         if (ndev->reg_state >= NETREG_UNREGISTERING)
 353                 return;
 354
 355         in6_dev = in6_dev_get(ndev);
 356         if (!in6_dev)
 357                 return;
 358
 359         read_lock_bh(&in6_dev->lock);
 360         list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
 361                 struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 362
 363                 if (!entry) {
 364                         pr_warn("roce_gid_mgmt: couldn't allocate entry for IPv6 update\n");
 365                         continue;
 366                 }
 367
 368                 entry->sin6.sin6_family = AF_INET6;
 369                 entry->sin6.sin6_addr = ifp->addr;
 370                 list_add_tail(&entry->list, &sin6_list);
 371         }
 372         read_unlock_bh(&in6_dev->lock);
 373
 374         in6_dev_put(in6_dev);
 375
 376         list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
 377                 union ib_gid    gid;
 378
 379                 rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
 380                 update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
 381                 list_del(&sin6_iter->list);
 382                 kfree(sin6_iter);
 383         }
 384 }
 385
 386 static void _add_netdev_ips(struct ib_device *ib_dev, u8 port,
 387                             struct net_device *ndev)
 388 {
 389         enum_netdev_ipv4_ips(ib_dev, port, ndev);
 390         if (IS_ENABLED(CONFIG_IPV6))
 391                 enum_netdev_ipv6_ips(ib_dev, port, ndev);
 392 }
 393
 394 static void add_netdev_ips(struct ib_device *ib_dev, u8 port,
 395                            struct net_device *rdma_ndev, void *cookie)
 396 {
 397         struct net_device *event_ndev = (struct net_device *)cookie;
 398
 399         enum_netdev_default_gids(ib_dev, port, event_ndev, rdma_ndev);
 400         _add_netdev_ips(ib_dev, port, event_ndev);
 401 }
 402
 403 static void del_netdev_ips(struct ib_device *ib_dev, u8 port,
 404                            struct net_device *rdma_ndev, void *cookie)
 405 {
 406         struct net_device *event_ndev = (struct net_device *)cookie;
 407
 408         ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
 409 }
 410
 411 static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
 412                                     u8 port,
 413                                     struct net_device *rdma_ndev,
 414                                     void *cookie)
 415 {
 416         struct net *net;
 417         struct net_device *ndev;
 418
 419         /* Lock the rtnl to make sure the netdevs does not move under
 420          * our feet
 421          */
 422         rtnl_lock();
 423         for_each_net(net)
 424                 for_each_netdev(net, ndev)
 425                         if (is_eth_port_of_netdev(ib_dev, port, rdma_ndev, ndev))
 426                                 add_netdev_ips(ib_dev, port, rdma_ndev, ndev);
 427         rtnl_unlock();
 428 }
 429
 430 /* This function will rescan all of the network devices in the system
 431  * and add their gids, as needed, to the relevant RoCE devices. */
 432 int roce_rescan_device(struct ib_device *ib_dev)
 433 {
 434         ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
 435                             enum_all_gids_of_dev_cb, NULL);
 436
 437         return 0;
 438 }
 439
 440 static void callback_for_addr_gid_device_scan(struct ib_device *device,
 441                                               u8 port,
 442                                               struct net_device *rdma_ndev,
 443                                               void *cookie)
 444 {
 445         struct update_gid_event_work *parsed = cookie;
 446
 447         return update_gid(parsed->gid_op, device,
 448                           port, &parsed->gid,
 449                           &parsed->gid_attr);
 450 }
 451
 452 static void handle_netdev_upper(struct ib_device *ib_dev, u8 port,
 453                                 void *cookie,
 454                                 void (*handle_netdev)(struct ib_device *ib_dev,
 455                                                       u8 port,
 456                                                       struct net_device *ndev))
 457 {
 458         struct net_device *ndev = (struct net_device *)cookie;
 459         struct upper_list {
 460                 struct list_head list;
 461                 struct net_device *upper;
 462         };
 463         struct net_device *upper;
 464         struct list_head *iter;
 465         struct upper_list *upper_iter;
 466         struct upper_list *upper_temp;
 467         LIST_HEAD(upper_list);
 468
 469         rcu_read_lock();
 470         netdev_for_each_all_upper_dev_rcu(ndev, upper, iter) {
 471                 struct upper_list *entry = kmalloc(sizeof(*entry),
 472                                                    GFP_ATOMIC);
 473
 474                 if (!entry) {
 475                         pr_info("roce_gid_mgmt: couldn't allocate entry to delete ndev\n");
 476                         continue;
 477                 }
 478
 479                 list_add_tail(&entry->list, &upper_list);
 480                 dev_hold(upper);
 481                 entry->upper = upper;
 482         }
 483         rcu_read_unlock();
 484
 485         handle_netdev(ib_dev, port, ndev);
 486         list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
 487                                  list) {
 488                 handle_netdev(ib_dev, port, upper_iter->upper);
 489                 dev_put(upper_iter->upper);
 490                 list_del(&upper_iter->list);
 491                 kfree(upper_iter);
 492         }
 493 }
 494
 495 static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
 496                                       struct net_device *event_ndev)
 497 {
 498         ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
 499 }
 500
 501 static void del_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 502                                  struct net_device *rdma_ndev, void *cookie)
 503 {
 504         handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
 505 }
 506
 507 static void add_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 508                                  struct net_device *rdma_ndev, void *cookie)
 509 {
 510         handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
 511 }
 512
 513 static void del_netdev_default_ips_join(struct ib_device *ib_dev, u8 port,
 514                                         struct net_device *rdma_ndev,
 515                                         void *cookie)
 516 {
 517         struct net_device *master_ndev;
 518
 519         rcu_read_lock();
 520         master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
 521         if (master_ndev)
 522                 dev_hold(master_ndev);
 523         rcu_read_unlock();
 524
 525         if (master_ndev) {
 526                 bond_delete_netdev_default_gids(ib_dev, port, master_ndev,
 527                                                 rdma_ndev);
 528                 dev_put(master_ndev);
 529         }
 530 }
 531
 532 static void del_netdev_default_ips(struct ib_device *ib_dev, u8 port,
 533                                    struct net_device *rdma_ndev, void *cookie)
 534 {
 535         struct net_device *event_ndev = (struct net_device *)cookie;
 536
 537         bond_delete_netdev_default_gids(ib_dev, port, event_ndev, rdma_ndev);
 538 }
 539
 540 /* The following functions operate on all IB devices. netdevice_event and
 541  * addr_event execute ib_enum_all_roce_netdevs through a work.
 542  * ib_enum_all_roce_netdevs iterates through all IB devices.
 543  */
 544
 545 static void netdevice_event_work_handler(struct work_struct *_work)
 546 {
 547         struct netdev_event_work *work =
 548                 container_of(_work, struct netdev_event_work, work);
 549         unsigned int i;
 550
 551         for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
 552                 ib_enum_all_roce_netdevs(work->cmds[i].filter,
 553                                          work->cmds[i].filter_ndev,
 554                                          work->cmds[i].cb,
 555                                          work->cmds[i].ndev);
 556                 dev_put(work->cmds[i].ndev);
 557                 dev_put(work->cmds[i].filter_ndev);
 558         }
 559
 560         kfree(work);
 561 }
 562
 563 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
 564                                 struct net_device *ndev)
 565 {
 566         unsigned int i;
 567         struct netdev_event_work *ndev_work =
 568                 kmalloc(sizeof(*ndev_work), GFP_KERNEL);
 569
 570         if (!ndev_work) {
 571                 pr_warn("roce_gid_mgmt: can't allocate work for netdevice_event\n");
 572                 return NOTIFY_DONE;
 573         }
 574
 575         memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
 576         for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
 577                 if (!ndev_work->cmds[i].ndev)
 578                         ndev_work->cmds[i].ndev = ndev;
 579                 if (!ndev_work->cmds[i].filter_ndev)
 580                         ndev_work->cmds[i].filter_ndev = ndev;
 581                 dev_hold(ndev_work->cmds[i].ndev);
 582                 dev_hold(ndev_work->cmds[i].filter_ndev);
 583         }
 584         INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
 585
 586         queue_work(ib_wq, &ndev_work->work);
 587
 588         return NOTIFY_DONE;
 589 }
 590
 591 static const struct netdev_event_work_cmd add_cmd = {
 592         .cb = add_netdev_ips, .filter = is_eth_port_of_netdev};
 593 static const struct netdev_event_work_cmd add_cmd_upper_ips = {
 594         .cb = add_netdev_upper_ips, .filter = is_eth_port_of_netdev};
 595
 596 static void netdevice_event_changeupper(struct netdev_notifier_changeupper_info *changeupper_info,
 597                                         struct netdev_event_work_cmd *cmds)
 598 {
 599         static const struct netdev_event_work_cmd upper_ips_del_cmd = {
 600                 .cb = del_netdev_upper_ips, .filter = upper_device_filter};
 601         static const struct netdev_event_work_cmd bonding_default_del_cmd = {
 602                 .cb = del_netdev_default_ips, .filter = is_eth_port_inactive_slave};
 603
 604         if (changeupper_info->linking == false) {
 605                 cmds[0] = upper_ips_del_cmd;
 606                 cmds[0].ndev = changeupper_info->upper_dev;
 607                 cmds[1] = add_cmd;
 608         } else {
 609                 cmds[0] = bonding_default_del_cmd;
 610                 cmds[0].ndev = changeupper_info->upper_dev;
 611                 cmds[1] = add_cmd_upper_ips;
 612                 cmds[1].ndev = changeupper_info->upper_dev;
 613                 cmds[1].filter_ndev = changeupper_info->upper_dev;
 614         }
 615 }
 616
 617 static int netdevice_event(struct notifier_block *this, unsigned long event,
 618                            void *ptr)
 619 {
 620         static const struct netdev_event_work_cmd del_cmd = {
 621                 .cb = del_netdev_ips, .filter = pass_all_filter};
 622         static const struct netdev_event_work_cmd bonding_default_del_cmd_join = {
 623                 .cb = del_netdev_default_ips_join, .filter = is_eth_port_inactive_slave};
 624         static const struct netdev_event_work_cmd default_del_cmd = {
 625                 .cb = del_netdev_default_ips, .filter = pass_all_filter};
 626         static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
 627                 .cb = del_netdev_upper_ips, .filter = upper_device_filter};
 628         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 629         struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
 630
 631         if (ndev->type != ARPHRD_ETHER)
 632                 return NOTIFY_DONE;
 633
 634         switch (event) {
 635         case NETDEV_REGISTER:
 636         case NETDEV_UP:
 637                 cmds[0] = bonding_default_del_cmd_join;
 638                 cmds[1] = add_cmd;
 639                 break;
 640
 641         case NETDEV_UNREGISTER:
 642                 if (ndev->reg_state < NETREG_UNREGISTERED)
 643                         cmds[0] = del_cmd;
 644                 else
 645                         return NOTIFY_DONE;
 646                 break;
 647
 648         case NETDEV_CHANGEADDR:
 649                 cmds[0] = default_del_cmd;
 650                 cmds[1] = add_cmd;
 651                 break;
 652
 653         case NETDEV_CHANGEUPPER:
 654                 netdevice_event_changeupper(
 655                         container_of(ptr, struct netdev_notifier_changeupper_info, info),
 656                         cmds);
 657                 break;
 658
 659         case NETDEV_BONDING_FAILOVER:
 660                 cmds[0] = bonding_event_ips_del_cmd;
 661                 cmds[1] = bonding_default_del_cmd_join;
 662                 cmds[2] = add_cmd_upper_ips;
 663                 break;
 664
 665         default:
 666                 return NOTIFY_DONE;
 667         }
 668
 669         return netdevice_queue_work(cmds, ndev);
 670 }
 671
 672 static void update_gid_event_work_handler(struct work_struct *_work)
 673 {
 674         struct update_gid_event_work *work =
 675                 container_of(_work, struct update_gid_event_work, work);
 676
 677         ib_enum_all_roce_netdevs(is_eth_port_of_netdev, work->gid_attr.ndev,
 678                                  callback_for_addr_gid_device_scan, work);
 679
 680         dev_put(work->gid_attr.ndev);
 681         kfree(work);
 682 }
 683
 684 static int addr_event(struct notifier_block *this, unsigned long event,
 685                       struct sockaddr *sa, struct net_device *ndev)
 686 {
 687         struct update_gid_event_work *work;
 688         enum gid_op_type gid_op;
 689
 690         if (ndev->type != ARPHRD_ETHER)
 691                 return NOTIFY_DONE;
 692
 693         switch (event) {
 694         case NETDEV_UP:
 695                 gid_op = GID_ADD;
 696                 break;
 697
 698         case NETDEV_DOWN:
 699                 gid_op = GID_DEL;
 700                 break;
 701
 702         default:
 703                 return NOTIFY_DONE;
 704         }
 705
 706         work = kmalloc(sizeof(*work), GFP_ATOMIC);
 707         if (!work) {
 708                 pr_warn("roce_gid_mgmt: Couldn't allocate work for addr_event\n");
 709                 return NOTIFY_DONE;
 710         }
 711
 712         INIT_WORK(&work->work, update_gid_event_work_handler);
 713
 714         rdma_ip2gid(sa, &work->gid);
 715         work->gid_op = gid_op;
 716
 717         memset(&work->gid_attr, 0, sizeof(work->gid_attr));
 718         dev_hold(ndev);
 719         work->gid_attr.ndev   = ndev;
 720
 721         queue_work(ib_wq, &work->work);
 722
 723         return NOTIFY_DONE;
 724 }
 725
 726 static int inetaddr_event(struct notifier_block *this, unsigned long event,
 727                           void *ptr)
 728 {
 729         struct sockaddr_in      in;
 730         struct net_device       *ndev;
 731         struct in_ifaddr        *ifa = ptr;
 732
 733         in.sin_family = AF_INET;
 734         in.sin_addr.s_addr = ifa->ifa_address;
 735         ndev = ifa->ifa_dev->dev;
 736
 737         return addr_event(this, event, (struct sockaddr *)&in, ndev);
 738 }
 739
 740 static int inet6addr_event(struct notifier_block *this, unsigned long event,
 741                            void *ptr)
 742 {
 743         struct sockaddr_in6     in6;
 744         struct net_device       *ndev;
 745         struct inet6_ifaddr     *ifa6 = ptr;
 746
 747         in6.sin6_family = AF_INET6;
 748         in6.sin6_addr = ifa6->addr;
 749         ndev = ifa6->idev->dev;
 750
 751         return addr_event(this, event, (struct sockaddr *)&in6, ndev);
 752 }
 753
 754 static struct notifier_block nb_netdevice = {
 755         .notifier_call = netdevice_event
 756 };
 757
 758 static struct notifier_block nb_inetaddr = {
 759         .notifier_call = inetaddr_event
 760 };
 761
 762 static struct notifier_block nb_inet6addr = {
 763         .notifier_call = inet6addr_event
 764 };
 765
 766 int __init roce_gid_mgmt_init(void)
 767 {
 768         register_inetaddr_notifier(&nb_inetaddr);
 769         if (IS_ENABLED(CONFIG_IPV6))
 770                 register_inet6addr_notifier(&nb_inet6addr);
 771         /* We relay on the netdevice notifier to enumerate all
 772          * existing devices in the system. Register to this notifier
 773          * last to make sure we will not miss any IP add/del
 774          * callbacks.
 775          */
 776         register_netdevice_notifier(&nb_netdevice);
 777
 778         return 0;
 779 }
 780
 781 void __exit roce_gid_mgmt_cleanup(void)
 782 {
 783         if (IS_ENABLED(CONFIG_IPV6))
 784                 unregister_inet6addr_notifier(&nb_inet6addr);
 785         unregister_inetaddr_notifier(&nb_inetaddr);
 786         unregister_netdevice_notifier(&nb_netdevice);
 787         /* Ensure all gid deletion tasks complete before we go down,
 788          * to avoid any reference to free'd memory. By the time
 789          * ib-core is removed, all physical devices have been removed,
 790          * so no issue with remaining hardware contexts.
 791          */
 792 }