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         rcu_read_lock();
 271         if (netif_is_bond_master(cookie_ndev) &&
 272             rdma_is_upper_dev_rcu(rdma_ndev, cookie_ndev))
 273                 match = true;
 274         rcu_read_unlock();
 275         return match;
 276 }
 277
 278 static void update_gid_ip(enum gid_op_type gid_op,
 279                           struct ib_device *ib_dev,
 280                           u8 port, struct net_device *ndev,
 281                           struct sockaddr *addr)
 282 {
 283         union ib_gid gid;
 284         struct ib_gid_attr gid_attr;
 285
 286         rdma_ip2gid(addr, &gid);
 287         memset(&gid_attr, 0, sizeof(gid_attr));
 288         gid_attr.ndev = ndev;
 289
 290         update_gid(gid_op, ib_dev, port, &gid, &gid_attr);
 291 }
 292
 293 static void bond_delete_netdev_default_gids(struct ib_device *ib_dev,
 294                                             u8 port,
 295                                             struct net_device *rdma_ndev,
 296                                             struct net_device *event_ndev)
 297 {
 298         struct net_device *real_dev = rdma_vlan_dev_real_dev(event_ndev);
 299         unsigned long gid_type_mask;
 300
 301         if (!rdma_ndev)
 302                 return;
 303
 304         if (!real_dev)
 305                 real_dev = event_ndev;
 306
 307         rcu_read_lock();
 308
 309         if (((rdma_ndev != event_ndev &&
 310               !rdma_is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
 311              is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev)
 312                                                  ==
 313              BONDING_SLAVE_STATE_INACTIVE)) {
 314                 rcu_read_unlock();
 315                 return;
 316         }
 317
 318         rcu_read_unlock();
 319
 320         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 321
 322         ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
 323                                      gid_type_mask,
 324                                      IB_CACHE_GID_DEFAULT_MODE_DELETE);
 325 }
 326
 327 static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
 328                                  u8 port, struct net_device *ndev)
 329 {
 330         struct in_device *in_dev;
 331         struct sin_list {
 332                 struct list_head        list;
 333                 struct sockaddr_in      ip;
 334         };
 335         struct sin_list *sin_iter;
 336         struct sin_list *sin_temp;
 337
 338         LIST_HEAD(sin_list);
 339         if (ndev->reg_state >= NETREG_UNREGISTERING)
 340                 return;
 341
 342         rcu_read_lock();
 343         in_dev = __in_dev_get_rcu(ndev);
 344         if (!in_dev) {
 345                 rcu_read_unlock();
 346                 return;
 347         }
 348
 349         for_ifa(in_dev) {
 350                 struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 351
 352                 if (!entry)
 353                         continue;
 354
 355                 entry->ip.sin_family = AF_INET;
 356                 entry->ip.sin_addr.s_addr = ifa->ifa_address;
 357                 list_add_tail(&entry->list, &sin_list);
 358         }
 359         endfor_ifa(in_dev);
 360         rcu_read_unlock();
 361
 362         list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
 363                 update_gid_ip(GID_ADD, ib_dev, port, ndev,
 364                               (struct sockaddr *)&sin_iter->ip);
 365                 list_del(&sin_iter->list);
 366                 kfree(sin_iter);
 367         }
 368 }
 369
 370 static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
 371                                  u8 port, struct net_device *ndev)
 372 {
 373         struct inet6_ifaddr *ifp;
 374         struct inet6_dev *in6_dev;
 375         struct sin6_list {
 376                 struct list_head        list;
 377                 struct sockaddr_in6     sin6;
 378         };
 379         struct sin6_list *sin6_iter;
 380         struct sin6_list *sin6_temp;
 381         struct ib_gid_attr gid_attr = {.ndev = ndev};
 382         LIST_HEAD(sin6_list);
 383
 384         if (ndev->reg_state >= NETREG_UNREGISTERING)
 385                 return;
 386
 387         in6_dev = in6_dev_get(ndev);
 388         if (!in6_dev)
 389                 return;
 390
 391         read_lock_bh(&in6_dev->lock);
 392         list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
 393                 struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
 394
 395                 if (!entry)
 396                         continue;
 397
 398                 entry->sin6.sin6_family = AF_INET6;
 399                 entry->sin6.sin6_addr = ifp->addr;
 400                 list_add_tail(&entry->list, &sin6_list);
 401         }
 402         read_unlock_bh(&in6_dev->lock);
 403
 404         in6_dev_put(in6_dev);
 405
 406         list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
 407                 union ib_gid    gid;
 408
 409                 rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
 410                 update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
 411                 list_del(&sin6_iter->list);
 412                 kfree(sin6_iter);
 413         }
 414 }
 415
 416 static void _add_netdev_ips(struct ib_device *ib_dev, u8 port,
 417                             struct net_device *ndev)
 418 {
 419         enum_netdev_ipv4_ips(ib_dev, port, ndev);
 420         if (IS_ENABLED(CONFIG_IPV6))
 421                 enum_netdev_ipv6_ips(ib_dev, port, ndev);
 422 }
 423
 424 static void add_netdev_ips(struct ib_device *ib_dev, u8 port,
 425                            struct net_device *rdma_ndev, void *cookie)
 426 {
 427         _add_netdev_ips(ib_dev, port, cookie);
 428 }
 429
 430 static void del_netdev_ips(struct ib_device *ib_dev, u8 port,
 431                            struct net_device *rdma_ndev, void *cookie)
 432 {
 433         ib_cache_gid_del_all_netdev_gids(ib_dev, port, cookie);
 434 }
 435
 436 /**
 437  * del_default_gids - Delete default GIDs of the event/cookie netdevice
 438  * @ib_dev:     RDMA device pointer
 439  * @port:       Port of the RDMA device whose GID table to consider
 440  * @rdma_ndev:  Unused rdma netdevice
 441  * @cookie:     Pointer to event netdevice
 442  *
 443  * del_default_gids() deletes the default GIDs of the event/cookie netdevice.
 444  */
 445 static void del_default_gids(struct ib_device *ib_dev, u8 port,
 446                              struct net_device *rdma_ndev, void *cookie)
 447 {
 448         struct net_device *cookie_ndev = cookie;
 449         unsigned long gid_type_mask;
 450
 451         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 452
 453         ib_cache_gid_set_default_gid(ib_dev, port, cookie_ndev, gid_type_mask,
 454                                      IB_CACHE_GID_DEFAULT_MODE_DELETE);
 455 }
 456
 457 static void add_default_gids(struct ib_device *ib_dev, u8 port,
 458                              struct net_device *rdma_ndev, void *cookie)
 459 {
 460         struct net_device *event_ndev = cookie;
 461         unsigned long gid_type_mask;
 462
 463         gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
 464         ib_cache_gid_set_default_gid(ib_dev, port, event_ndev, gid_type_mask,
 465                                      IB_CACHE_GID_DEFAULT_MODE_SET);
 466 }
 467
 468 static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
 469                                     u8 port,
 470                                     struct net_device *rdma_ndev,
 471                                     void *cookie)
 472 {
 473         struct net *net;
 474         struct net_device *ndev;
 475
 476         /* Lock the rtnl to make sure the netdevs does not move under
 477          * our feet
 478          */
 479         rtnl_lock();
 480         down_read(&net_rwsem);
 481         for_each_net(net)
 482                 for_each_netdev(net, ndev) {
 483                         /*
 484                          * Filter and add default GIDs of the primary netdevice
 485                          * when not in bonding mode, or add default GIDs
 486                          * of bond master device, when in bonding mode.
 487                          */
 488                         if (is_ndev_for_default_gid_filter(ib_dev, port,
 489                                                            rdma_ndev, ndev))
 490                                 add_default_gids(ib_dev, port, rdma_ndev, ndev);
 491
 492                         if (is_eth_port_of_netdev_filter(ib_dev, port,
 493                                                          rdma_ndev, ndev))
 494                                 _add_netdev_ips(ib_dev, port, ndev);
 495                 }
 496         up_read(&net_rwsem);
 497         rtnl_unlock();
 498 }
 499
 500 /**
 501  * rdma_roce_rescan_device - Rescan all of the network devices in the system
 502  * and add their gids, as needed, to the relevant RoCE devices.
 503  *
 504  * @device:         the rdma device
 505  */
 506 void rdma_roce_rescan_device(struct ib_device *ib_dev)
 507 {
 508         ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
 509                             enum_all_gids_of_dev_cb, NULL);
 510 }
 511 EXPORT_SYMBOL(rdma_roce_rescan_device);
 512
 513 static void callback_for_addr_gid_device_scan(struct ib_device *device,
 514                                               u8 port,
 515                                               struct net_device *rdma_ndev,
 516                                               void *cookie)
 517 {
 518         struct update_gid_event_work *parsed = cookie;
 519
 520         return update_gid(parsed->gid_op, device,
 521                           port, &parsed->gid,
 522                           &parsed->gid_attr);
 523 }
 524
 525 struct upper_list {
 526         struct list_head list;
 527         struct net_device *upper;
 528 };
 529
 530 static int netdev_upper_walk(struct net_device *upper, void *data)
 531 {
 532         struct upper_list *entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 533         struct list_head *upper_list = data;
 534
 535         if (!entry)
 536                 return 0;
 537
 538         list_add_tail(&entry->list, upper_list);
 539         dev_hold(upper);
 540         entry->upper = upper;
 541
 542         return 0;
 543 }
 544
 545 static void handle_netdev_upper(struct ib_device *ib_dev, u8 port,
 546                                 void *cookie,
 547                                 void (*handle_netdev)(struct ib_device *ib_dev,
 548                                                       u8 port,
 549                                                       struct net_device *ndev))
 550 {
 551         struct net_device *ndev = cookie;
 552         struct upper_list *upper_iter;
 553         struct upper_list *upper_temp;
 554         LIST_HEAD(upper_list);
 555
 556         rcu_read_lock();
 557         netdev_walk_all_upper_dev_rcu(ndev, netdev_upper_walk, &upper_list);
 558         rcu_read_unlock();
 559
 560         handle_netdev(ib_dev, port, ndev);
 561         list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
 562                                  list) {
 563                 handle_netdev(ib_dev, port, upper_iter->upper);
 564                 dev_put(upper_iter->upper);
 565                 list_del(&upper_iter->list);
 566                 kfree(upper_iter);
 567         }
 568 }
 569
 570 static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
 571                                       struct net_device *event_ndev)
 572 {
 573         ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
 574 }
 575
 576 static void del_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 577                                  struct net_device *rdma_ndev, void *cookie)
 578 {
 579         handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
 580 }
 581
 582 static void add_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
 583                                  struct net_device *rdma_ndev, void *cookie)
 584 {
 585         handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
 586 }
 587
 588 static void del_netdev_default_ips_join(struct ib_device *ib_dev, u8 port,
 589                                         struct net_device *rdma_ndev,
 590                                         void *cookie)
 591 {
 592         struct net_device *master_ndev;
 593
 594         rcu_read_lock();
 595         master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
 596         if (master_ndev)
 597                 dev_hold(master_ndev);
 598         rcu_read_unlock();
 599
 600         if (master_ndev) {
 601                 bond_delete_netdev_default_gids(ib_dev, port, rdma_ndev,
 602                                                 master_ndev);
 603                 dev_put(master_ndev);
 604         }
 605 }
 606
 607 /* The following functions operate on all IB devices. netdevice_event and
 608  * addr_event execute ib_enum_all_roce_netdevs through a work.
 609  * ib_enum_all_roce_netdevs iterates through all IB devices.
 610  */
 611
 612 static void netdevice_event_work_handler(struct work_struct *_work)
 613 {
 614         struct netdev_event_work *work =
 615                 container_of(_work, struct netdev_event_work, work);
 616         unsigned int i;
 617
 618         for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
 619                 ib_enum_all_roce_netdevs(work->cmds[i].filter,
 620                                          work->cmds[i].filter_ndev,
 621                                          work->cmds[i].cb,
 622                                          work->cmds[i].ndev);
 623                 dev_put(work->cmds[i].ndev);
 624                 dev_put(work->cmds[i].filter_ndev);
 625         }
 626
 627         kfree(work);
 628 }
 629
 630 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
 631                                 struct net_device *ndev)
 632 {
 633         unsigned int i;
 634         struct netdev_event_work *ndev_work =
 635                 kmalloc(sizeof(*ndev_work), GFP_KERNEL);
 636
 637         if (!ndev_work)
 638                 return NOTIFY_DONE;
 639
 640         memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
 641         for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
 642                 if (!ndev_work->cmds[i].ndev)
 643                         ndev_work->cmds[i].ndev = ndev;
 644                 if (!ndev_work->cmds[i].filter_ndev)
 645                         ndev_work->cmds[i].filter_ndev = ndev;
 646                 dev_hold(ndev_work->cmds[i].ndev);
 647                 dev_hold(ndev_work->cmds[i].filter_ndev);
 648         }
 649         INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
 650
 651         queue_work(gid_cache_wq, &ndev_work->work);
 652
 653         return NOTIFY_DONE;
 654 }
 655
 656 static const struct netdev_event_work_cmd add_cmd = {
 657         .cb     = add_netdev_ips,
 658         .filter = is_eth_port_of_netdev_filter
 659 };
 660
 661 static const struct netdev_event_work_cmd add_cmd_upper_ips = {
 662         .cb     = add_netdev_upper_ips,
 663         .filter = is_eth_port_of_netdev_filter
 664 };
 665
 666 static void
 667 ndev_event_unlink(struct netdev_notifier_changeupper_info *changeupper_info,
 668                   struct netdev_event_work_cmd *cmds)
 669 {
 670         static const struct netdev_event_work_cmd
 671                         upper_ips_del_cmd = {
 672                                 .cb     = del_netdev_upper_ips,
 673                                 .filter = upper_device_filter
 674         };
 675
 676         cmds[0] = upper_ips_del_cmd;
 677         cmds[0].ndev = changeupper_info->upper_dev;
 678         cmds[1] = add_cmd;
 679 }
 680
 681 static const struct netdev_event_work_cmd bonding_default_add_cmd = {
 682         .cb     = add_default_gids,
 683         .filter = is_upper_ndev_bond_master_filter
 684 };
 685
 686 static void
 687 ndev_event_link(struct net_device *event_ndev,
 688                 struct netdev_notifier_changeupper_info *changeupper_info,
 689                 struct netdev_event_work_cmd *cmds)
 690 {
 691         static const struct netdev_event_work_cmd
 692                         bonding_default_del_cmd = {
 693                                 .cb     = del_default_gids,
 694                                 .filter = is_upper_ndev_bond_master_filter
 695                         };
 696         /*
 697          * When a lower netdev is linked to its upper bonding
 698          * netdev, delete lower slave netdev's default GIDs.
 699          */
 700         cmds[0] = bonding_default_del_cmd;
 701         cmds[0].ndev = event_ndev;
 702         cmds[0].filter_ndev = changeupper_info->upper_dev;
 703
 704         /* Now add bonding upper device default GIDs */
 705         cmds[1] = bonding_default_add_cmd;
 706         cmds[1].ndev = changeupper_info->upper_dev;
 707         cmds[1].filter_ndev = changeupper_info->upper_dev;
 708
 709         /* Now add bonding upper device IP based GIDs */
 710         cmds[2] = add_cmd_upper_ips;
 711         cmds[2].ndev = changeupper_info->upper_dev;
 712         cmds[2].filter_ndev = changeupper_info->upper_dev;
 713 }
 714
 715 static void netdevice_event_changeupper(struct net_device *event_ndev,
 716                 struct netdev_notifier_changeupper_info *changeupper_info,
 717                 struct netdev_event_work_cmd *cmds)
 718 {
 719         if (changeupper_info->linking)
 720                 ndev_event_link(event_ndev, changeupper_info, cmds);
 721         else
 722                 ndev_event_unlink(changeupper_info, cmds);
 723 }
 724
 725 static const struct netdev_event_work_cmd add_default_gid_cmd = {
 726         .cb     = add_default_gids,
 727         .filter = is_ndev_for_default_gid_filter,
 728 };
 729
 730 static int netdevice_event(struct notifier_block *this, unsigned long event,
 731                            void *ptr)
 732 {
 733         static const struct netdev_event_work_cmd del_cmd = {
 734                 .cb = del_netdev_ips, .filter = pass_all_filter};
 735         static const struct netdev_event_work_cmd
 736                         bonding_default_del_cmd_join = {
 737                                 .cb     = del_netdev_default_ips_join,
 738                                 .filter = is_eth_port_inactive_slave_filter
 739                         };
 740         static const struct netdev_event_work_cmd
 741                         netdev_del_cmd = {
 742                                 .cb     = del_netdev_ips,
 743                                 .filter = is_eth_port_of_netdev_filter
 744                         };
 745         static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
 746                 .cb = del_netdev_upper_ips, .filter = upper_device_filter};
 747         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 748         struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
 749
 750         if (ndev->type != ARPHRD_ETHER)
 751                 return NOTIFY_DONE;
 752
 753         switch (event) {
 754         case NETDEV_REGISTER:
 755         case NETDEV_UP:
 756                 cmds[0] = bonding_default_del_cmd_join;
 757                 cmds[1] = add_default_gid_cmd;
 758                 cmds[2] = add_cmd;
 759                 break;
 760
 761         case NETDEV_UNREGISTER:
 762                 if (ndev->reg_state < NETREG_UNREGISTERED)
 763                         cmds[0] = del_cmd;
 764                 else
 765                         return NOTIFY_DONE;
 766                 break;
 767
 768         case NETDEV_CHANGEADDR:
 769                 cmds[0] = netdev_del_cmd;
 770                 cmds[1] = add_default_gid_cmd;
 771                 cmds[2] = add_cmd;
 772                 break;
 773
 774         case NETDEV_CHANGEUPPER:
 775                 netdevice_event_changeupper(ndev,
 776                         container_of(ptr, struct netdev_notifier_changeupper_info, info),
 777                         cmds);
 778                 break;
 779
 780         case NETDEV_BONDING_FAILOVER:
 781                 cmds[0] = bonding_event_ips_del_cmd;
 782                 /* Add default GIDs of the bond device */
 783                 cmds[1] = bonding_default_add_cmd;
 784                 /* Add IP based GIDs of the bond device */
 785                 cmds[2] = add_cmd_upper_ips;
 786                 break;
 787
 788         default:
 789                 return NOTIFY_DONE;
 790         }
 791
 792         return netdevice_queue_work(cmds, ndev);
 793 }
 794
 795 static void update_gid_event_work_handler(struct work_struct *_work)
 796 {
 797         struct update_gid_event_work *work =
 798                 container_of(_work, struct update_gid_event_work, work);
 799
 800         ib_enum_all_roce_netdevs(is_eth_port_of_netdev_filter,
 801                                  work->gid_attr.ndev,
 802                                  callback_for_addr_gid_device_scan, work);
 803
 804         dev_put(work->gid_attr.ndev);
 805         kfree(work);
 806 }
 807
 808 static int addr_event(struct notifier_block *this, unsigned long event,
 809                       struct sockaddr *sa, struct net_device *ndev)
 810 {
 811         struct update_gid_event_work *work;
 812         enum gid_op_type gid_op;
 813
 814         if (ndev->type != ARPHRD_ETHER)
 815                 return NOTIFY_DONE;
 816
 817         switch (event) {
 818         case NETDEV_UP:
 819                 gid_op = GID_ADD;
 820                 break;
 821
 822         case NETDEV_DOWN:
 823                 gid_op = GID_DEL;
 824                 break;
 825
 826         default:
 827                 return NOTIFY_DONE;
 828         }
 829
 830         work = kmalloc(sizeof(*work), GFP_ATOMIC);
 831         if (!work)
 832                 return NOTIFY_DONE;
 833
 834         INIT_WORK(&work->work, update_gid_event_work_handler);
 835
 836         rdma_ip2gid(sa, &work->gid);
 837         work->gid_op = gid_op;
 838
 839         memset(&work->gid_attr, 0, sizeof(work->gid_attr));
 840         dev_hold(ndev);
 841         work->gid_attr.ndev   = ndev;
 842
 843         queue_work(gid_cache_wq, &work->work);
 844
 845         return NOTIFY_DONE;
 846 }
 847
 848 static int inetaddr_event(struct notifier_block *this, unsigned long event,
 849                           void *ptr)
 850 {
 851         struct sockaddr_in      in;
 852         struct net_device       *ndev;
 853         struct in_ifaddr        *ifa = ptr;
 854
 855         in.sin_family = AF_INET;
 856         in.sin_addr.s_addr = ifa->ifa_address;
 857         ndev = ifa->ifa_dev->dev;
 858
 859         return addr_event(this, event, (struct sockaddr *)&in, ndev);
 860 }
 861
 862 static int inet6addr_event(struct notifier_block *this, unsigned long event,
 863                            void *ptr)
 864 {
 865         struct sockaddr_in6     in6;
 866         struct net_device       *ndev;
 867         struct inet6_ifaddr     *ifa6 = ptr;
 868
 869         in6.sin6_family = AF_INET6;
 870         in6.sin6_addr = ifa6->addr;
 871         ndev = ifa6->idev->dev;
 872
 873         return addr_event(this, event, (struct sockaddr *)&in6, ndev);
 874 }
 875
 876 static struct notifier_block nb_netdevice = {
 877         .notifier_call = netdevice_event
 878 };
 879
 880 static struct notifier_block nb_inetaddr = {
 881         .notifier_call = inetaddr_event
 882 };
 883
 884 static struct notifier_block nb_inet6addr = {
 885         .notifier_call = inet6addr_event
 886 };
 887
 888 int __init roce_gid_mgmt_init(void)
 889 {
 890         gid_cache_wq = alloc_ordered_workqueue("gid-cache-wq", 0);
 891         if (!gid_cache_wq)
 892                 return -ENOMEM;
 893
 894         register_inetaddr_notifier(&nb_inetaddr);
 895         if (IS_ENABLED(CONFIG_IPV6))
 896                 register_inet6addr_notifier(&nb_inet6addr);
 897         /* We relay on the netdevice notifier to enumerate all
 898          * existing devices in the system. Register to this notifier
 899          * last to make sure we will not miss any IP add/del
 900          * callbacks.
 901          */
 902         register_netdevice_notifier(&nb_netdevice);
 903
 904         return 0;
 905 }
 906
 907 void __exit roce_gid_mgmt_cleanup(void)
 908 {
 909         if (IS_ENABLED(CONFIG_IPV6))
 910                 unregister_inet6addr_notifier(&nb_inet6addr);
 911         unregister_inetaddr_notifier(&nb_inetaddr);
 912         unregister_netdevice_notifier(&nb_netdevice);
 913         /* Ensure all gid deletion tasks complete before we go down,
 914          * to avoid any reference to free'd memory. By the time
 915          * ib-core is removed, all physical devices have been removed,
 916          * so no issue with remaining hardware contexts.
 917          */
 918         destroy_workqueue(gid_cache_wq);
 919 }