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 static struct workqueue_struct *gid_cache_wq;
  46
  47 enum gid_op_type {
  48         GID_DEL = 0,
  49         GID_ADD
  50 };
  51
  52 struct update_gid_event_work {
  53         struct work_struct work;
  54         union ib_gid       gid;
  55         struct ib_gid_attr gid_attr;
  56         enum gid_op_type gid_op;
  57 };
  58
  59 #define ROCE_NETDEV_CALLBACK_SZ         3
  60 struct netdev_event_work_cmd {
  61         roce_netdev_callback    cb;
  62         roce_netdev_filter      filter;
  63         struct net_device       *ndev;
  64         struct net_device       *filter_ndev;
  65 };
  66
  67 struct netdev_event_work {
  68         struct work_struct              work;
  69         struct netdev_event_work_cmd    cmds[ROCE_NETDEV_CALLBACK_SZ];
  70 };
  71
  72 static const struct {
  73         bool (*is_supported)(const struct ib_device *device, u8 port_num);
  74         enum ib_gid_type gid_type;
  75 } PORT_CAP_TO_GID_TYPE[] = {
  76         {rdma_protocol_roce_eth_encap, IB_GID_TYPE_ROCE},
  77         {rdma_protocol_roce_udp_encap, IB_GID_TYPE_ROCE_UDP_ENCAP},
  78 };
  79
  80 #define CAP_TO_GID_TABLE_SIZE   ARRAY_SIZE(PORT_CAP_TO_GID_TYPE)
  81
  82 unsigned long roce_gid_type_mask_support(struct ib_device *ib_dev, u8 port)
  83 {
  84         int i;
  85         unsigned int ret_flags = 0;
  86
  87         if (!rdma_protocol_roce(ib_dev, port))
  88                 return 1UL << IB_GID_TYPE_IB;
  89
  90         for (i = 0; i < CAP_TO_GID_TABLE_SIZE; i++)
  91                 if (PORT_CAP_TO_GID_TYPE[i].is_supported(ib_dev, port))
  92                         ret_flags |= 1UL << PORT_CAP_TO_GID_TYPE[i].gid_type;
  93
  94         return ret_flags;
  95 }
  96 EXPORT_SYMBOL(roce_gid_type_mask_support);
  97
  98 static void update_gid(enum gid_op_type gid_op, struct ib_device *ib_dev,
  99                        u8 port, union ib_gid *gid,
 100                        struct ib_gid_attr *gid_attr)
 101 {
 102         int i;
 103         unsigned long gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 104
 105         for (i = 0; i < IB_GID_TYPE_SIZE; i++) {
 106                 if ((1UL << i) & gid_type_mask) {
 107                         gid_attr->gid_type = i;
 108                         switch (gid_op) {
 109                         case GID_ADD:
 110                                 ib_cache_gid_add(ib_dev, port,
 111                                                  gid, gid_attr);
 112                                 break;
 113                         case GID_DEL:
 114                                 ib_cache_gid_del(ib_dev, port,
 115                                                  gid, gid_attr);
 116                                 break;
 117                         }
 118                 }
 119         }
 120 }
 121
 122 enum bonding_slave_state {
 123         BONDING_SLAVE_STATE_ACTIVE      = 1UL << 0,
 124         BONDING_SLAVE_STATE_INACTIVE    = 1UL << 1,
 125         /* No primary slave or the device isn't a slave in bonding */
 126         BONDING_SLAVE_STATE_NA          = 1UL << 2,
 127 };
 128
 129 static enum bonding_slave_state is_eth_active_slave_of_bonding_rcu(struct net_device *dev,
 130                                                                    struct net_device *upper)
 131 {
 132         if (upper && netif_is_bond_master(upper)) {
 133                 struct net_device *pdev =
 134                         bond_option_active_slave_get_rcu(netdev_priv(upper));
 135
 136                 if (pdev)
 137                         return dev == pdev ? BONDING_SLAVE_STATE_ACTIVE :
 138                                 BONDING_SLAVE_STATE_INACTIVE;
 139         }
 140
 141         return BONDING_SLAVE_STATE_NA;
 142 }
 143
 144 #define REQUIRED_BOND_STATES            (BONDING_SLAVE_STATE_ACTIVE |   \
 145                                          BONDING_SLAVE_STATE_NA)
 146 static bool
 147 is_eth_port_of_netdev_filter(struct ib_device *ib_dev, u8 port,
 148                              struct net_device *rdma_ndev, void *cookie)
 149 {
 150         struct net_device *real_dev;
 151         bool res;
 152
 153         if (!rdma_ndev)
 154                 return false;
 155
 156         rcu_read_lock();
 157         real_dev = rdma_vlan_dev_real_dev(cookie);
 158         if (!real_dev)
 159                 real_dev = cookie;
 160
 161         res = ((rdma_is_upper_dev_rcu(rdma_ndev, cookie) &&
 162                (is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) &
 163                 REQUIRED_BOND_STATES)) ||
 164                real_dev == rdma_ndev);
 165
 166         rcu_read_unlock();
 167         return res;
 168 }
 169
 170 static bool
 171 is_eth_port_inactive_slave_filter(struct ib_device *ib_dev, u8 port,
 172                                   struct net_device *rdma_ndev, void *cookie)
 173 {
 174         struct net_device *master_dev;
 175         bool res;
 176
 177         if (!rdma_ndev)
 178                 return false;
 179
 180         rcu_read_lock();
 181         master_dev = netdev_master_upper_dev_get_rcu(rdma_ndev);
 182         res = is_eth_active_slave_of_bonding_rcu(rdma_ndev, master_dev) ==
 183                 BONDING_SLAVE_STATE_INACTIVE;
 184         rcu_read_unlock();
 185
 186         return res;
 187 }
 188
 189 /** is_ndev_for_default_gid_filter - Check if a given netdevice
 190  * can be considered for default GIDs or not.
 191  * @ib_dev:             IB device to check
 192  * @port:               Port to consider for adding default GID
 193  * @rdma_ndev:          rdma netdevice pointer
 194  * @cookie_ndev:        Netdevice to consider to form a default GID
 195  *
 196  * is_ndev_for_default_gid_filter() returns true if a given netdevice can be
 197  * considered for deriving default RoCE GID, returns false otherwise.
 198  */
 199 static bool
 200 is_ndev_for_default_gid_filter(struct ib_device *ib_dev, u8 port,
 201                                struct net_device *rdma_ndev, void *cookie)
 202 {
 203         struct net_device *cookie_ndev = cookie;
 204         bool res;
 205
 206         if (!rdma_ndev)
 207                 return false;
 208
 209         rcu_read_lock();
 210
 211         /*
 212          * When rdma netdevice is used in bonding, bonding master netdevice
 213          * should be considered for default GIDs. Therefore, ignore slave rdma
 214          * netdevices when bonding is considered.
 215          * Additionally when event(cookie) netdevice is bond master device,
 216          * make sure that it the upper netdevice of rdma netdevice.
 217          */
 218         res = ((cookie_ndev == rdma_ndev && !netif_is_bond_slave(rdma_ndev)) ||
 219                (netif_is_bond_master(cookie_ndev) &&
 220                 rdma_is_upper_dev_rcu(rdma_ndev, cookie_ndev)));
 221
 222         rcu_read_unlock();
 223         return res;
 224 }
 225
 226 static bool pass_all_filter(struct ib_device *ib_dev, u8 port,
 227                             struct net_device *rdma_ndev, void *cookie)
 228 {
 229         return true;
 230 }
 231
 232 static bool upper_device_filter(struct ib_device *ib_dev, u8 port,
 233                                 struct net_device *rdma_ndev, void *cookie)
 234 {
 235         bool res;
 236
 237         if (!rdma_ndev)
 238                 return false;
 239
 240         if (rdma_ndev == cookie)
 241                 return true;
 242
 243         rcu_read_lock();
 244         res = rdma_is_upper_dev_rcu(rdma_ndev, cookie);
 245         rcu_read_unlock();
 246
 247         return res;
 248 }
 249
 250 /**
 251  * is_upper_ndev_bond_master_filter - Check if a given netdevice
 252  * is bond master device of netdevice of the the RDMA device of port.
 253  * @ib_dev:             IB device to check
 254  * @port:               Port to consider for adding default GID
 255  * @rdma_ndev:          Pointer to rdma netdevice
 256  * @cookie:             Netdevice to consider to form a default GID
 257  *
 258  * is_upper_ndev_bond_master_filter() returns true if a cookie_netdev
 259  * is bond master device and rdma_ndev is its lower netdevice. It might
 260  * not have been established as slave device yet.
 261  */
 262 static bool
 263 is_upper_ndev_bond_master_filter(struct ib_device *ib_dev, u8 port,
 264                                  struct net_device *rdma_ndev,
 265                                  void *cookie)
 266 {
 267         struct net_device *cookie_ndev = cookie;
 268         bool match = false;
 269
 270         if (!rdma_ndev)
 271                 return false;
 272
 273         rcu_read_lock();
 274         if (netif_is_bond_master(cookie_ndev) &&
 275             rdma_is_upper_dev_rcu(rdma_ndev, cookie_ndev))
 276                 match = true;
 277         rcu_read_unlock();
 278         return match;
 279 }
 280
 281 static void update_gid_ip(enum gid_op_type gid_op,
 282                           struct ib_device *ib_dev,
 283                           u8 port, struct net_device *ndev,
 284                           struct sockaddr *addr)
 285 {
 286         union ib_gid gid;
 287         struct ib_gid_attr gid_attr;
 288
 289         rdma_ip2gid(addr, &gid);
 290         memset(&gid_attr, 0, sizeof(gid_attr));
 291         gid_attr.ndev = ndev;
 292
 293         update_gid(gid_op, ib_dev, port, &gid, &gid_attr);
 294 }
 295
 296 static void bond_delete_netdev_default_gids(struct ib_device *ib_dev,
 297                                             u8 port,
 298                                             struct net_device *rdma_ndev,
 299                                             struct net_device *event_ndev)
 300 {
 301         struct net_device *real_dev = rdma_vlan_dev_real_dev(event_ndev);
 302         unsigned long gid_type_mask;
 303
 304         if (!rdma_ndev)
 305                 return;
 306
 307         if (!real_dev)
 308                 real_dev = event_ndev;
 309
 310         rcu_read_lock();
 311
 312         if (((rdma_ndev != event_ndev &&
 313               !rdma_is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
 314              is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev)
 315                                                  ==
 316              BONDING_SLAVE_STATE_INACTIVE)) {
 317                 rcu_read_unlock();
 318                 return;
 319         }
 320
 321         rcu_read_unlock();
 322
 323         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 324
 325         ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
 326                                      gid_type_mask,
 327                                      IB_CACHE_GID_DEFAULT_MODE_DELETE);
 328 }
 329
 330 static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
 331                                  u8 port, struct net_device *ndev)
 332 {
 333         struct in_device *in_dev;
 334         struct sin_list {
 335                 struct list_head        list;
 336                 struct sockaddr_in      ip;
 337         };
 338         struct sin_list *sin_iter;
 339         struct sin_list *sin_temp;
 340
 341         LIST_HEAD(sin_list);
 342         if (ndev->reg_state >= NETREG_UNREGISTERING)
 343                 return;
 344
 345         rcu_read_lock();
 346         in_dev = __in_dev_get_rcu(ndev);
 347         if (!in_dev) {
 348                 rcu_read_unlock();
 349                 return;
 350         }
 351
 352         for_ifa(in_dev) {
 353                 struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 354
 355                 if (!entry)
 356                         continue;
 357
 358                 entry->ip.sin_family = AF_INET;
 359                 entry->ip.sin_addr.s_addr = ifa->ifa_address;
 360                 list_add_tail(&entry->list, &sin_list);
 361         }
 362         endfor_ifa(in_dev);
 363         rcu_read_unlock();
 364
 365         list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
 366                 update_gid_ip(GID_ADD, ib_dev, port, ndev,
 367                               (struct sockaddr *)&sin_iter->ip);
 368                 list_del(&sin_iter->list);
 369                 kfree(sin_iter);
 370         }
 371 }
 372
 373 static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
 374                                  u8 port, struct net_device *ndev)
 375 {
 376         struct inet6_ifaddr *ifp;
 377         struct inet6_dev *in6_dev;
 378         struct sin6_list {
 379                 struct list_head        list;
 380                 struct sockaddr_in6     sin6;
 381         };
 382         struct sin6_list *sin6_iter;
 383         struct sin6_list *sin6_temp;
 384         struct ib_gid_attr gid_attr = {.ndev = ndev};
 385         LIST_HEAD(sin6_list);
 386
 387         if (ndev->reg_state >= NETREG_UNREGISTERING)
 388                 return;
 389
 390         in6_dev = in6_dev_get(ndev);
 391         if (!in6_dev)
 392                 return;
 393
 394         read_lock_bh(&in6_dev->lock);
 395         list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
 396                 struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 397
 398                 if (!entry)
 399                         continue;
 400
 401                 entry->sin6.sin6_family = AF_INET6;
 402                 entry->sin6.sin6_addr = ifp->addr;
 403                 list_add_tail(&entry->list, &sin6_list);
 404         }
 405         read_unlock_bh(&in6_dev->lock);
 406
 407         in6_dev_put(in6_dev);
 408
 409         list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
 410                 union ib_gid    gid;
 411
 412                 rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
 413                 update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
 414                 list_del(&sin6_iter->list);
 415                 kfree(sin6_iter);
 416         }
 417 }
 418
 419 static void _add_netdev_ips(struct ib_device *ib_dev, u8 port,
 420                             struct net_device *ndev)
 421 {
 422         enum_netdev_ipv4_ips(ib_dev, port, ndev);
 423         if (IS_ENABLED(CONFIG_IPV6))
 424                 enum_netdev_ipv6_ips(ib_dev, port, ndev);
 425 }
 426
 427 static void add_netdev_ips(struct ib_device *ib_dev, u8 port,
 428                            struct net_device *rdma_ndev, void *cookie)
 429 {
 430         _add_netdev_ips(ib_dev, port, cookie);
 431 }
 432
 433 static void del_netdev_ips(struct ib_device *ib_dev, u8 port,
 434                            struct net_device *rdma_ndev, void *cookie)
 435 {
 436         ib_cache_gid_del_all_netdev_gids(ib_dev, port, cookie);
 437 }
 438
 439 /**
 440  * del_default_gids - Delete default GIDs of the event/cookie netdevice
 441  * @ib_dev:     RDMA device pointer
 442  * @port:       Port of the RDMA device whose GID table to consider
 443  * @rdma_ndev:  Unused rdma netdevice
 444  * @cookie:     Pointer to event netdevice
 445  *
 446  * del_default_gids() deletes the default GIDs of the event/cookie netdevice.
 447  */
 448 static void del_default_gids(struct ib_device *ib_dev, u8 port,
 449                              struct net_device *rdma_ndev, void *cookie)
 450 {
 451         struct net_device *cookie_ndev = cookie;
 452         unsigned long gid_type_mask;
 453
 454         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 455
 456         ib_cache_gid_set_default_gid(ib_dev, port, cookie_ndev, gid_type_mask,
 457                                      IB_CACHE_GID_DEFAULT_MODE_DELETE);
 458 }
 459
 460 static void add_default_gids(struct ib_device *ib_dev, u8 port,
 461                              struct net_device *rdma_ndev, void *cookie)
 462 {
 463         struct net_device *event_ndev = cookie;
 464         unsigned long gid_type_mask;
 465
 466         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 467         ib_cache_gid_set_default_gid(ib_dev, port, event_ndev, gid_type_mask,
 468                                      IB_CACHE_GID_DEFAULT_MODE_SET);
 469 }
 470
 471 static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
 472                                     u8 port,
 473                                     struct net_device *rdma_ndev,
 474                                     void *cookie)
 475 {
 476         struct net *net;
 477         struct net_device *ndev;
 478
 479         /* Lock the rtnl to make sure the netdevs does not move under
 480          * our feet
 481          */
 482         rtnl_lock();
 483         down_read(&net_rwsem);
 484         for_each_net(net)
 485                 for_each_netdev(net, ndev) {
 486                         /*
 487                          * Filter and add default GIDs of the primary netdevice
 488                          * when not in bonding mode, or add default GIDs
 489                          * of bond master device, when in bonding mode.
 490                          */
 491                         if (is_ndev_for_default_gid_filter(ib_dev, port,
 492                                                            rdma_ndev, ndev))
 493                                 add_default_gids(ib_dev, port, rdma_ndev, ndev);
 494
 495                         if (is_eth_port_of_netdev_filter(ib_dev, port,
 496                                                          rdma_ndev, ndev))
 497                                 _add_netdev_ips(ib_dev, port, ndev);
 498                 }
 499         up_read(&net_rwsem);
 500         rtnl_unlock();
 501 }
 502
 503 /**
 504  * rdma_roce_rescan_device - Rescan all of the network devices in the system
 505  * and add their gids, as needed, to the relevant RoCE devices.
 506  *
 507  * @device:         the rdma device
 508  */
 509 void rdma_roce_rescan_device(struct ib_device *ib_dev)
 510 {
 511         ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
 512                             enum_all_gids_of_dev_cb, NULL);
 513 }
 514 EXPORT_SYMBOL(rdma_roce_rescan_device);
 515
 516 static void callback_for_addr_gid_device_scan(struct ib_device *device,
 517                                               u8 port,
 518                                               struct net_device *rdma_ndev,
 519                                               void *cookie)
 520 {
 521         struct update_gid_event_work *parsed = cookie;
 522
 523         return update_gid(parsed->gid_op, device,
 524                           port, &parsed->gid,
 525                           &parsed->gid_attr);
 526 }
 527
 528 struct upper_list {
 529         struct list_head list;
 530         struct net_device *upper;
 531 };
 532
 533 static int netdev_upper_walk(struct net_device *upper, void *data)
 534 {
 535         struct upper_list *entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 536         struct list_head *upper_list = data;
 537
 538         if (!entry)
 539                 return 0;
 540
 541         list_add_tail(&entry->list, upper_list);
 542         dev_hold(upper);
 543         entry->upper = upper;
 544
 545         return 0;
 546 }
 547
 548 static void handle_netdev_upper(struct ib_device *ib_dev, u8 port,
 549                                 void *cookie,
 550                                 void (*handle_netdev)(struct ib_device *ib_dev,
 551                                                       u8 port,
 552                                                       struct net_device *ndev))
 553 {
 554         struct net_device *ndev = cookie;
 555         struct upper_list *upper_iter;
 556         struct upper_list *upper_temp;
 557         LIST_HEAD(upper_list);
 558
 559         rcu_read_lock();
 560         netdev_walk_all_upper_dev_rcu(ndev, netdev_upper_walk, &upper_list);
 561         rcu_read_unlock();
 562
 563         handle_netdev(ib_dev, port, ndev);
 564         list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
 565                                  list) {
 566                 handle_netdev(ib_dev, port, upper_iter->upper);
 567                 dev_put(upper_iter->upper);
 568                 list_del(&upper_iter->list);
 569                 kfree(upper_iter);
 570         }
 571 }
 572
 573 static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
 574                                       struct net_device *event_ndev)
 575 {
 576         ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
 577 }
 578
 579 static void del_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 580                                  struct net_device *rdma_ndev, void *cookie)
 581 {
 582         handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
 583 }
 584
 585 static void add_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 586                                  struct net_device *rdma_ndev, void *cookie)
 587 {
 588         handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
 589 }
 590
 591 static void del_netdev_default_ips_join(struct ib_device *ib_dev, u8 port,
 592                                         struct net_device *rdma_ndev,
 593                                         void *cookie)
 594 {
 595         struct net_device *master_ndev;
 596
 597         rcu_read_lock();
 598         master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
 599         if (master_ndev)
 600                 dev_hold(master_ndev);
 601         rcu_read_unlock();
 602
 603         if (master_ndev) {
 604                 bond_delete_netdev_default_gids(ib_dev, port, rdma_ndev,
 605                                                 master_ndev);
 606                 dev_put(master_ndev);
 607         }
 608 }
 609
 610 /* The following functions operate on all IB devices. netdevice_event and
 611  * addr_event execute ib_enum_all_roce_netdevs through a work.
 612  * ib_enum_all_roce_netdevs iterates through all IB devices.
 613  */
 614
 615 static void netdevice_event_work_handler(struct work_struct *_work)
 616 {
 617         struct netdev_event_work *work =
 618                 container_of(_work, struct netdev_event_work, work);
 619         unsigned int i;
 620
 621         for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
 622                 ib_enum_all_roce_netdevs(work->cmds[i].filter,
 623                                          work->cmds[i].filter_ndev,
 624                                          work->cmds[i].cb,
 625                                          work->cmds[i].ndev);
 626                 dev_put(work->cmds[i].ndev);
 627                 dev_put(work->cmds[i].filter_ndev);
 628         }
 629
 630         kfree(work);
 631 }
 632
 633 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
 634                                 struct net_device *ndev)
 635 {
 636         unsigned int i;
 637         struct netdev_event_work *ndev_work =
 638                 kmalloc(sizeof(*ndev_work), GFP_KERNEL);
 639
 640         if (!ndev_work)
 641                 return NOTIFY_DONE;
 642
 643         memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
 644         for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
 645                 if (!ndev_work->cmds[i].ndev)
 646                         ndev_work->cmds[i].ndev = ndev;
 647                 if (!ndev_work->cmds[i].filter_ndev)
 648                         ndev_work->cmds[i].filter_ndev = ndev;
 649                 dev_hold(ndev_work->cmds[i].ndev);
 650                 dev_hold(ndev_work->cmds[i].filter_ndev);
 651         }
 652         INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
 653
 654         queue_work(gid_cache_wq, &ndev_work->work);
 655
 656         return NOTIFY_DONE;
 657 }
 658
 659 static const struct netdev_event_work_cmd add_cmd = {
 660         .cb     = add_netdev_ips,
 661         .filter = is_eth_port_of_netdev_filter
 662 };
 663
 664 static const struct netdev_event_work_cmd add_cmd_upper_ips = {
 665         .cb     = add_netdev_upper_ips,
 666         .filter = is_eth_port_of_netdev_filter
 667 };
 668
 669 static void
 670 ndev_event_unlink(struct netdev_notifier_changeupper_info *changeupper_info,
 671                   struct netdev_event_work_cmd *cmds)
 672 {
 673         static const struct netdev_event_work_cmd
 674                         upper_ips_del_cmd = {
 675                                 .cb     = del_netdev_upper_ips,
 676                                 .filter = upper_device_filter
 677         };
 678
 679         cmds[0] = upper_ips_del_cmd;
 680         cmds[0].ndev = changeupper_info->upper_dev;
 681         cmds[1] = add_cmd;
 682 }
 683
 684 static const struct netdev_event_work_cmd bonding_default_add_cmd = {
 685         .cb     = add_default_gids,
 686         .filter = is_upper_ndev_bond_master_filter
 687 };
 688
 689 static void
 690 ndev_event_link(struct net_device *event_ndev,
 691                 struct netdev_notifier_changeupper_info *changeupper_info,
 692                 struct netdev_event_work_cmd *cmds)
 693 {
 694         static const struct netdev_event_work_cmd
 695                         bonding_default_del_cmd = {
 696                                 .cb     = del_default_gids,
 697                                 .filter = is_upper_ndev_bond_master_filter
 698                         };
 699         /*
 700          * When a lower netdev is linked to its upper bonding
 701          * netdev, delete lower slave netdev's default GIDs.
 702          */
 703         cmds[0] = bonding_default_del_cmd;
 704         cmds[0].ndev = event_ndev;
 705         cmds[0].filter_ndev = changeupper_info->upper_dev;
 706
 707         /* Now add bonding upper device default GIDs */
 708         cmds[1] = bonding_default_add_cmd;
 709         cmds[1].ndev = changeupper_info->upper_dev;
 710         cmds[1].filter_ndev = changeupper_info->upper_dev;
 711
 712         /* Now add bonding upper device IP based GIDs */
 713         cmds[2] = add_cmd_upper_ips;
 714         cmds[2].ndev = changeupper_info->upper_dev;
 715         cmds[2].filter_ndev = changeupper_info->upper_dev;
 716 }
 717
 718 static void netdevice_event_changeupper(struct net_device *event_ndev,
 719                 struct netdev_notifier_changeupper_info *changeupper_info,
 720                 struct netdev_event_work_cmd *cmds)
 721 {
 722         if (changeupper_info->linking)
 723                 ndev_event_link(event_ndev, changeupper_info, cmds);
 724         else
 725                 ndev_event_unlink(changeupper_info, cmds);
 726 }
 727
 728 static const struct netdev_event_work_cmd add_default_gid_cmd = {
 729         .cb     = add_default_gids,
 730         .filter = is_ndev_for_default_gid_filter,
 731 };
 732
 733 static int netdevice_event(struct notifier_block *this, unsigned long event,
 734                            void *ptr)
 735 {
 736         static const struct netdev_event_work_cmd del_cmd = {
 737                 .cb = del_netdev_ips, .filter = pass_all_filter};
 738         static const struct netdev_event_work_cmd
 739                         bonding_default_del_cmd_join = {
 740                                 .cb     = del_netdev_default_ips_join,
 741                                 .filter = is_eth_port_inactive_slave_filter
 742                         };
 743         static const struct netdev_event_work_cmd
 744                         netdev_del_cmd = {
 745                                 .cb     = del_netdev_ips,
 746                                 .filter = is_eth_port_of_netdev_filter
 747                         };
 748         static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
 749                 .cb = del_netdev_upper_ips, .filter = upper_device_filter};
 750         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 751         struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
 752
 753         if (ndev->type != ARPHRD_ETHER)
 754                 return NOTIFY_DONE;
 755
 756         switch (event) {
 757         case NETDEV_REGISTER:
 758         case NETDEV_UP:
 759                 cmds[0] = bonding_default_del_cmd_join;
 760                 cmds[1] = add_default_gid_cmd;
 761                 cmds[2] = add_cmd;
 762                 break;
 763
 764         case NETDEV_UNREGISTER:
 765                 if (ndev->reg_state < NETREG_UNREGISTERED)
 766                         cmds[0] = del_cmd;
 767                 else
 768                         return NOTIFY_DONE;
 769                 break;
 770
 771         case NETDEV_CHANGEADDR:
 772                 cmds[0] = netdev_del_cmd;
 773                 if (ndev->reg_state == NETREG_REGISTERED) {
 774                         cmds[1] = add_default_gid_cmd;
 775                         cmds[2] = add_cmd;
 776                 }
 777                 break;
 778
 779         case NETDEV_CHANGEUPPER:
 780                 netdevice_event_changeupper(ndev,
 781                         container_of(ptr, struct netdev_notifier_changeupper_info, info),
 782                         cmds);
 783                 break;
 784
 785         case NETDEV_BONDING_FAILOVER:
 786                 cmds[0] = bonding_event_ips_del_cmd;
 787                 /* Add default GIDs of the bond device */
 788                 cmds[1] = bonding_default_add_cmd;
 789                 /* Add IP based GIDs of the bond device */
 790                 cmds[2] = add_cmd_upper_ips;
 791                 break;
 792
 793         default:
 794                 return NOTIFY_DONE;
 795         }
 796
 797         return netdevice_queue_work(cmds, ndev);
 798 }
 799
 800 static void update_gid_event_work_handler(struct work_struct *_work)
 801 {
 802         struct update_gid_event_work *work =
 803                 container_of(_work, struct update_gid_event_work, work);
 804
 805         ib_enum_all_roce_netdevs(is_eth_port_of_netdev_filter,
 806                                  work->gid_attr.ndev,
 807                                  callback_for_addr_gid_device_scan, work);
 808
 809         dev_put(work->gid_attr.ndev);
 810         kfree(work);
 811 }
 812
 813 static int addr_event(struct notifier_block *this, unsigned long event,
 814                       struct sockaddr *sa, struct net_device *ndev)
 815 {
 816         struct update_gid_event_work *work;
 817         enum gid_op_type gid_op;
 818
 819         if (ndev->type != ARPHRD_ETHER)
 820                 return NOTIFY_DONE;
 821
 822         switch (event) {
 823         case NETDEV_UP:
 824                 gid_op = GID_ADD;
 825                 break;
 826
 827         case NETDEV_DOWN:
 828                 gid_op = GID_DEL;
 829                 break;
 830
 831         default:
 832                 return NOTIFY_DONE;
 833         }
 834
 835         work = kmalloc(sizeof(*work), GFP_ATOMIC);
 836         if (!work)
 837                 return NOTIFY_DONE;
 838
 839         INIT_WORK(&work->work, update_gid_event_work_handler);
 840
 841         rdma_ip2gid(sa, &work->gid);
 842         work->gid_op = gid_op;
 843
 844         memset(&work->gid_attr, 0, sizeof(work->gid_attr));
 845         dev_hold(ndev);
 846         work->gid_attr.ndev   = ndev;
 847
 848         queue_work(gid_cache_wq, &work->work);
 849
 850         return NOTIFY_DONE;
 851 }
 852
 853 static int inetaddr_event(struct notifier_block *this, unsigned long event,
 854                           void *ptr)
 855 {
 856         struct sockaddr_in      in;
 857         struct net_device       *ndev;
 858         struct in_ifaddr        *ifa = ptr;
 859
 860         in.sin_family = AF_INET;
 861         in.sin_addr.s_addr = ifa->ifa_address;
 862         ndev = ifa->ifa_dev->dev;
 863
 864         return addr_event(this, event, (struct sockaddr *)&in, ndev);
 865 }
 866
 867 static int inet6addr_event(struct notifier_block *this, unsigned long event,
 868                            void *ptr)
 869 {
 870         struct sockaddr_in6     in6;
 871         struct net_device       *ndev;
 872         struct inet6_ifaddr     *ifa6 = ptr;
 873
 874         in6.sin6_family = AF_INET6;
 875         in6.sin6_addr = ifa6->addr;
 876         ndev = ifa6->idev->dev;
 877
 878         return addr_event(this, event, (struct sockaddr *)&in6, ndev);
 879 }
 880
 881 static struct notifier_block nb_netdevice = {
 882         .notifier_call = netdevice_event
 883 };
 884
 885 static struct notifier_block nb_inetaddr = {
 886         .notifier_call = inetaddr_event
 887 };
 888
 889 static struct notifier_block nb_inet6addr = {
 890         .notifier_call = inet6addr_event
 891 };
 892
 893 int __init roce_gid_mgmt_init(void)
 894 {
 895         gid_cache_wq = alloc_ordered_workqueue("gid-cache-wq", 0);
 896         if (!gid_cache_wq)
 897                 return -ENOMEM;
 898
 899         register_inetaddr_notifier(&nb_inetaddr);
 900         if (IS_ENABLED(CONFIG_IPV6))
 901                 register_inet6addr_notifier(&nb_inet6addr);
 902         /* We relay on the netdevice notifier to enumerate all
 903          * existing devices in the system. Register to this notifier
 904          * last to make sure we will not miss any IP add/del
 905          * callbacks.
 906          */
 907         register_netdevice_notifier(&nb_netdevice);
 908
 909         return 0;
 910 }
 911
 912 void __exit roce_gid_mgmt_cleanup(void)
 913 {
 914         if (IS_ENABLED(CONFIG_IPV6))
 915                 unregister_inet6addr_notifier(&nb_inet6addr);
 916         unregister_inetaddr_notifier(&nb_inetaddr);
 917         unregister_netdevice_notifier(&nb_netdevice);
 918         /* Ensure all gid deletion tasks complete before we go down,
 919          * to avoid any reference to free'd memory. By the time
 920          * ib-core is removed, all physical devices have been removed,
 921          * so no issue with remaining hardware contexts.
 922          */
 923         destroy_workqueue(gid_cache_wq);
 924 }