These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / infiniband / core / roce_gid_mgmt.c
diff --git a/kernel/drivers/infiniband/core/roce_gid_mgmt.c b/kernel/drivers/infiniband/core/roce_gid_mgmt.c
new file mode 100644 (file)
index 0000000..178f984
--- /dev/null
@@ -0,0 +1,747 @@
+/*
+ * Copyright (c) 2015, Mellanox Technologies inc.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "core_priv.h"
+
+#include <linux/in.h>
+#include <linux/in6.h>
+
+/* For in6_dev_get/in6_dev_put */
+#include <net/addrconf.h>
+#include <net/bonding.h>
+
+#include <rdma/ib_cache.h>
+#include <rdma/ib_addr.h>
+
+enum gid_op_type {
+       GID_DEL = 0,
+       GID_ADD
+};
+
+struct update_gid_event_work {
+       struct work_struct work;
+       union ib_gid       gid;
+       struct ib_gid_attr gid_attr;
+       enum gid_op_type gid_op;
+};
+
+#define ROCE_NETDEV_CALLBACK_SZ                3
+struct netdev_event_work_cmd {
+       roce_netdev_callback    cb;
+       roce_netdev_filter      filter;
+       struct net_device       *ndev;
+       struct net_device       *filter_ndev;
+};
+
+struct netdev_event_work {
+       struct work_struct              work;
+       struct netdev_event_work_cmd    cmds[ROCE_NETDEV_CALLBACK_SZ];
+};
+
+static void update_gid(enum gid_op_type gid_op, struct ib_device *ib_dev,
+                      u8 port, union ib_gid *gid,
+                      struct ib_gid_attr *gid_attr)
+{
+       switch (gid_op) {
+       case GID_ADD:
+               ib_cache_gid_add(ib_dev, port, gid, gid_attr);
+               break;
+       case GID_DEL:
+               ib_cache_gid_del(ib_dev, port, gid, gid_attr);
+               break;
+       }
+}
+
+enum bonding_slave_state {
+       BONDING_SLAVE_STATE_ACTIVE      = 1UL << 0,
+       BONDING_SLAVE_STATE_INACTIVE    = 1UL << 1,
+       /* No primary slave or the device isn't a slave in bonding */
+       BONDING_SLAVE_STATE_NA          = 1UL << 2,
+};
+
+static enum bonding_slave_state is_eth_active_slave_of_bonding_rcu(struct net_device *dev,
+                                                                  struct net_device *upper)
+{
+       if (upper && netif_is_bond_master(upper)) {
+               struct net_device *pdev =
+                       bond_option_active_slave_get_rcu(netdev_priv(upper));
+
+               if (pdev)
+                       return dev == pdev ? BONDING_SLAVE_STATE_ACTIVE :
+                               BONDING_SLAVE_STATE_INACTIVE;
+       }
+
+       return BONDING_SLAVE_STATE_NA;
+}
+
+static bool is_upper_dev_rcu(struct net_device *dev, struct net_device *upper)
+{
+       struct net_device *_upper = NULL;
+       struct list_head *iter;
+
+       netdev_for_each_all_upper_dev_rcu(dev, _upper, iter)
+               if (_upper == upper)
+                       break;
+
+       return _upper == upper;
+}
+
+#define REQUIRED_BOND_STATES           (BONDING_SLAVE_STATE_ACTIVE |   \
+                                        BONDING_SLAVE_STATE_NA)
+static int is_eth_port_of_netdev(struct ib_device *ib_dev, u8 port,
+                                struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *event_ndev = (struct net_device *)cookie;
+       struct net_device *real_dev;
+       int res;
+
+       if (!rdma_ndev)
+               return 0;
+
+       rcu_read_lock();
+       real_dev = rdma_vlan_dev_real_dev(event_ndev);
+       if (!real_dev)
+               real_dev = event_ndev;
+
+       res = ((is_upper_dev_rcu(rdma_ndev, event_ndev) &&
+              (is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) &
+               REQUIRED_BOND_STATES)) ||
+              real_dev == rdma_ndev);
+
+       rcu_read_unlock();
+       return res;
+}
+
+static int is_eth_port_inactive_slave(struct ib_device *ib_dev, u8 port,
+                                     struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *master_dev;
+       int res;
+
+       if (!rdma_ndev)
+               return 0;
+
+       rcu_read_lock();
+       master_dev = netdev_master_upper_dev_get_rcu(rdma_ndev);
+       res = is_eth_active_slave_of_bonding_rcu(rdma_ndev, master_dev) ==
+               BONDING_SLAVE_STATE_INACTIVE;
+       rcu_read_unlock();
+
+       return res;
+}
+
+static int pass_all_filter(struct ib_device *ib_dev, u8 port,
+                          struct net_device *rdma_ndev, void *cookie)
+{
+       return 1;
+}
+
+static int upper_device_filter(struct ib_device *ib_dev, u8 port,
+                              struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *event_ndev = (struct net_device *)cookie;
+       int res;
+
+       if (!rdma_ndev)
+               return 0;
+
+       if (rdma_ndev == event_ndev)
+               return 1;
+
+       rcu_read_lock();
+       res = is_upper_dev_rcu(rdma_ndev, event_ndev);
+       rcu_read_unlock();
+
+       return res;
+}
+
+static void update_gid_ip(enum gid_op_type gid_op,
+                         struct ib_device *ib_dev,
+                         u8 port, struct net_device *ndev,
+                         struct sockaddr *addr)
+{
+       union ib_gid gid;
+       struct ib_gid_attr gid_attr;
+
+       rdma_ip2gid(addr, &gid);
+       memset(&gid_attr, 0, sizeof(gid_attr));
+       gid_attr.ndev = ndev;
+
+       update_gid(gid_op, ib_dev, port, &gid, &gid_attr);
+}
+
+static void enum_netdev_default_gids(struct ib_device *ib_dev,
+                                    u8 port, struct net_device *event_ndev,
+                                    struct net_device *rdma_ndev)
+{
+       rcu_read_lock();
+       if (!rdma_ndev ||
+           ((rdma_ndev != event_ndev &&
+             !is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
+            is_eth_active_slave_of_bonding_rcu(rdma_ndev,
+                                               netdev_master_upper_dev_get_rcu(rdma_ndev)) ==
+            BONDING_SLAVE_STATE_INACTIVE)) {
+               rcu_read_unlock();
+               return;
+       }
+       rcu_read_unlock();
+
+       ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
+                                    IB_CACHE_GID_DEFAULT_MODE_SET);
+}
+
+static void bond_delete_netdev_default_gids(struct ib_device *ib_dev,
+                                           u8 port,
+                                           struct net_device *event_ndev,
+                                           struct net_device *rdma_ndev)
+{
+       struct net_device *real_dev = rdma_vlan_dev_real_dev(event_ndev);
+
+       if (!rdma_ndev)
+               return;
+
+       if (!real_dev)
+               real_dev = event_ndev;
+
+       rcu_read_lock();
+
+       if (is_upper_dev_rcu(rdma_ndev, event_ndev) &&
+           is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) ==
+           BONDING_SLAVE_STATE_INACTIVE) {
+               rcu_read_unlock();
+
+               ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
+                                            IB_CACHE_GID_DEFAULT_MODE_DELETE);
+       } else {
+               rcu_read_unlock();
+       }
+}
+
+static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
+                                u8 port, struct net_device *ndev)
+{
+       struct in_device *in_dev;
+       struct sin_list {
+               struct list_head        list;
+               struct sockaddr_in      ip;
+       };
+       struct sin_list *sin_iter;
+       struct sin_list *sin_temp;
+
+       LIST_HEAD(sin_list);
+       if (ndev->reg_state >= NETREG_UNREGISTERING)
+               return;
+
+       rcu_read_lock();
+       in_dev = __in_dev_get_rcu(ndev);
+       if (!in_dev) {
+               rcu_read_unlock();
+               return;
+       }
+
+       for_ifa(in_dev) {
+               struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
+
+               if (!entry) {
+                       pr_warn("roce_gid_mgmt: couldn't allocate entry for IPv4 update\n");
+                       continue;
+               }
+               entry->ip.sin_family = AF_INET;
+               entry->ip.sin_addr.s_addr = ifa->ifa_address;
+               list_add_tail(&entry->list, &sin_list);
+       }
+       endfor_ifa(in_dev);
+       rcu_read_unlock();
+
+       list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
+               update_gid_ip(GID_ADD, ib_dev, port, ndev,
+                             (struct sockaddr *)&sin_iter->ip);
+               list_del(&sin_iter->list);
+               kfree(sin_iter);
+       }
+}
+
+static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
+                                u8 port, struct net_device *ndev)
+{
+       struct inet6_ifaddr *ifp;
+       struct inet6_dev *in6_dev;
+       struct sin6_list {
+               struct list_head        list;
+               struct sockaddr_in6     sin6;
+       };
+       struct sin6_list *sin6_iter;
+       struct sin6_list *sin6_temp;
+       struct ib_gid_attr gid_attr = {.ndev = ndev};
+       LIST_HEAD(sin6_list);
+
+       if (ndev->reg_state >= NETREG_UNREGISTERING)
+               return;
+
+       in6_dev = in6_dev_get(ndev);
+       if (!in6_dev)
+               return;
+
+       read_lock_bh(&in6_dev->lock);
+       list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
+               struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
+
+               if (!entry) {
+                       pr_warn("roce_gid_mgmt: couldn't allocate entry for IPv6 update\n");
+                       continue;
+               }
+
+               entry->sin6.sin6_family = AF_INET6;
+               entry->sin6.sin6_addr = ifp->addr;
+               list_add_tail(&entry->list, &sin6_list);
+       }
+       read_unlock_bh(&in6_dev->lock);
+
+       in6_dev_put(in6_dev);
+
+       list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
+               union ib_gid    gid;
+
+               rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
+               update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
+               list_del(&sin6_iter->list);
+               kfree(sin6_iter);
+       }
+}
+
+static void _add_netdev_ips(struct ib_device *ib_dev, u8 port,
+                           struct net_device *ndev)
+{
+       enum_netdev_ipv4_ips(ib_dev, port, ndev);
+       if (IS_ENABLED(CONFIG_IPV6))
+               enum_netdev_ipv6_ips(ib_dev, port, ndev);
+}
+
+static void add_netdev_ips(struct ib_device *ib_dev, u8 port,
+                          struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *event_ndev = (struct net_device *)cookie;
+
+       enum_netdev_default_gids(ib_dev, port, event_ndev, rdma_ndev);
+       _add_netdev_ips(ib_dev, port, event_ndev);
+}
+
+static void del_netdev_ips(struct ib_device *ib_dev, u8 port,
+                          struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *event_ndev = (struct net_device *)cookie;
+
+       ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
+}
+
+static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
+                                   u8 port,
+                                   struct net_device *rdma_ndev,
+                                   void *cookie)
+{
+       struct net *net;
+       struct net_device *ndev;
+
+       /* Lock the rtnl to make sure the netdevs does not move under
+        * our feet
+        */
+       rtnl_lock();
+       for_each_net(net)
+               for_each_netdev(net, ndev)
+                       if (is_eth_port_of_netdev(ib_dev, port, rdma_ndev, ndev))
+                               add_netdev_ips(ib_dev, port, rdma_ndev, ndev);
+       rtnl_unlock();
+}
+
+/* This function will rescan all of the network devices in the system
+ * and add their gids, as needed, to the relevant RoCE devices. */
+int roce_rescan_device(struct ib_device *ib_dev)
+{
+       ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
+                           enum_all_gids_of_dev_cb, NULL);
+
+       return 0;
+}
+
+static void callback_for_addr_gid_device_scan(struct ib_device *device,
+                                             u8 port,
+                                             struct net_device *rdma_ndev,
+                                             void *cookie)
+{
+       struct update_gid_event_work *parsed = cookie;
+
+       return update_gid(parsed->gid_op, device,
+                         port, &parsed->gid,
+                         &parsed->gid_attr);
+}
+
+static void handle_netdev_upper(struct ib_device *ib_dev, u8 port,
+                               void *cookie,
+                               void (*handle_netdev)(struct ib_device *ib_dev,
+                                                     u8 port,
+                                                     struct net_device *ndev))
+{
+       struct net_device *ndev = (struct net_device *)cookie;
+       struct upper_list {
+               struct list_head list;
+               struct net_device *upper;
+       };
+       struct net_device *upper;
+       struct list_head *iter;
+       struct upper_list *upper_iter;
+       struct upper_list *upper_temp;
+       LIST_HEAD(upper_list);
+
+       rcu_read_lock();
+       netdev_for_each_all_upper_dev_rcu(ndev, upper, iter) {
+               struct upper_list *entry = kmalloc(sizeof(*entry),
+                                                  GFP_ATOMIC);
+
+               if (!entry) {
+                       pr_info("roce_gid_mgmt: couldn't allocate entry to delete ndev\n");
+                       continue;
+               }
+
+               list_add_tail(&entry->list, &upper_list);
+               dev_hold(upper);
+               entry->upper = upper;
+       }
+       rcu_read_unlock();
+
+       handle_netdev(ib_dev, port, ndev);
+       list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
+                                list) {
+               handle_netdev(ib_dev, port, upper_iter->upper);
+               dev_put(upper_iter->upper);
+               list_del(&upper_iter->list);
+               kfree(upper_iter);
+       }
+}
+
+static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
+                                     struct net_device *event_ndev)
+{
+       ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
+}
+
+static void del_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
+                                struct net_device *rdma_ndev, void *cookie)
+{
+       handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
+}
+
+static void add_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
+                                struct net_device *rdma_ndev, void *cookie)
+{
+       handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
+}
+
+static void del_netdev_default_ips_join(struct ib_device *ib_dev, u8 port,
+                                       struct net_device *rdma_ndev,
+                                       void *cookie)
+{
+       struct net_device *master_ndev;
+
+       rcu_read_lock();
+       master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
+       if (master_ndev)
+               dev_hold(master_ndev);
+       rcu_read_unlock();
+
+       if (master_ndev) {
+               bond_delete_netdev_default_gids(ib_dev, port, master_ndev,
+                                               rdma_ndev);
+               dev_put(master_ndev);
+       }
+}
+
+static void del_netdev_default_ips(struct ib_device *ib_dev, u8 port,
+                                  struct net_device *rdma_ndev, void *cookie)
+{
+       struct net_device *event_ndev = (struct net_device *)cookie;
+
+       bond_delete_netdev_default_gids(ib_dev, port, event_ndev, rdma_ndev);
+}
+
+/* The following functions operate on all IB devices. netdevice_event and
+ * addr_event execute ib_enum_all_roce_netdevs through a work.
+ * ib_enum_all_roce_netdevs iterates through all IB devices.
+ */
+
+static void netdevice_event_work_handler(struct work_struct *_work)
+{
+       struct netdev_event_work *work =
+               container_of(_work, struct netdev_event_work, work);
+       unsigned int i;
+
+       for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
+               ib_enum_all_roce_netdevs(work->cmds[i].filter,
+                                        work->cmds[i].filter_ndev,
+                                        work->cmds[i].cb,
+                                        work->cmds[i].ndev);
+               dev_put(work->cmds[i].ndev);
+               dev_put(work->cmds[i].filter_ndev);
+       }
+
+       kfree(work);
+}
+
+static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
+                               struct net_device *ndev)
+{
+       unsigned int i;
+       struct netdev_event_work *ndev_work =
+               kmalloc(sizeof(*ndev_work), GFP_KERNEL);
+
+       if (!ndev_work) {
+               pr_warn("roce_gid_mgmt: can't allocate work for netdevice_event\n");
+               return NOTIFY_DONE;
+       }
+
+       memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
+       for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
+               if (!ndev_work->cmds[i].ndev)
+                       ndev_work->cmds[i].ndev = ndev;
+               if (!ndev_work->cmds[i].filter_ndev)
+                       ndev_work->cmds[i].filter_ndev = ndev;
+               dev_hold(ndev_work->cmds[i].ndev);
+               dev_hold(ndev_work->cmds[i].filter_ndev);
+       }
+       INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
+
+       queue_work(ib_wq, &ndev_work->work);
+
+       return NOTIFY_DONE;
+}
+
+static const struct netdev_event_work_cmd add_cmd = {
+       .cb = add_netdev_ips, .filter = is_eth_port_of_netdev};
+static const struct netdev_event_work_cmd add_cmd_upper_ips = {
+       .cb = add_netdev_upper_ips, .filter = is_eth_port_of_netdev};
+
+static void netdevice_event_changeupper(struct netdev_notifier_changeupper_info *changeupper_info,
+                                       struct netdev_event_work_cmd *cmds)
+{
+       static const struct netdev_event_work_cmd upper_ips_del_cmd = {
+               .cb = del_netdev_upper_ips, .filter = upper_device_filter};
+       static const struct netdev_event_work_cmd bonding_default_del_cmd = {
+               .cb = del_netdev_default_ips, .filter = is_eth_port_inactive_slave};
+
+       if (changeupper_info->linking == false) {
+               cmds[0] = upper_ips_del_cmd;
+               cmds[0].ndev = changeupper_info->upper_dev;
+               cmds[1] = add_cmd;
+       } else {
+               cmds[0] = bonding_default_del_cmd;
+               cmds[0].ndev = changeupper_info->upper_dev;
+               cmds[1] = add_cmd_upper_ips;
+               cmds[1].ndev = changeupper_info->upper_dev;
+               cmds[1].filter_ndev = changeupper_info->upper_dev;
+       }
+}
+
+static int netdevice_event(struct notifier_block *this, unsigned long event,
+                          void *ptr)
+{
+       static const struct netdev_event_work_cmd del_cmd = {
+               .cb = del_netdev_ips, .filter = pass_all_filter};
+       static const struct netdev_event_work_cmd bonding_default_del_cmd_join = {
+               .cb = del_netdev_default_ips_join, .filter = is_eth_port_inactive_slave};
+       static const struct netdev_event_work_cmd default_del_cmd = {
+               .cb = del_netdev_default_ips, .filter = pass_all_filter};
+       static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
+               .cb = del_netdev_upper_ips, .filter = upper_device_filter};
+       struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
+       struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
+
+       if (ndev->type != ARPHRD_ETHER)
+               return NOTIFY_DONE;
+
+       switch (event) {
+       case NETDEV_REGISTER:
+       case NETDEV_UP:
+               cmds[0] = bonding_default_del_cmd_join;
+               cmds[1] = add_cmd;
+               break;
+
+       case NETDEV_UNREGISTER:
+               if (ndev->reg_state < NETREG_UNREGISTERED)
+                       cmds[0] = del_cmd;
+               else
+                       return NOTIFY_DONE;
+               break;
+
+       case NETDEV_CHANGEADDR:
+               cmds[0] = default_del_cmd;
+               cmds[1] = add_cmd;
+               break;
+
+       case NETDEV_CHANGEUPPER:
+               netdevice_event_changeupper(
+                       container_of(ptr, struct netdev_notifier_changeupper_info, info),
+                       cmds);
+               break;
+
+       case NETDEV_BONDING_FAILOVER:
+               cmds[0] = bonding_event_ips_del_cmd;
+               cmds[1] = bonding_default_del_cmd_join;
+               cmds[2] = add_cmd_upper_ips;
+               break;
+
+       default:
+               return NOTIFY_DONE;
+       }
+
+       return netdevice_queue_work(cmds, ndev);
+}
+
+static void update_gid_event_work_handler(struct work_struct *_work)
+{
+       struct update_gid_event_work *work =
+               container_of(_work, struct update_gid_event_work, work);
+
+       ib_enum_all_roce_netdevs(is_eth_port_of_netdev, work->gid_attr.ndev,
+                                callback_for_addr_gid_device_scan, work);
+
+       dev_put(work->gid_attr.ndev);
+       kfree(work);
+}
+
+static int addr_event(struct notifier_block *this, unsigned long event,
+                     struct sockaddr *sa, struct net_device *ndev)
+{
+       struct update_gid_event_work *work;
+       enum gid_op_type gid_op;
+
+       if (ndev->type != ARPHRD_ETHER)
+               return NOTIFY_DONE;
+
+       switch (event) {
+       case NETDEV_UP:
+               gid_op = GID_ADD;
+               break;
+
+       case NETDEV_DOWN:
+               gid_op = GID_DEL;
+               break;
+
+       default:
+               return NOTIFY_DONE;
+       }
+
+       work = kmalloc(sizeof(*work), GFP_ATOMIC);
+       if (!work) {
+               pr_warn("roce_gid_mgmt: Couldn't allocate work for addr_event\n");
+               return NOTIFY_DONE;
+       }
+
+       INIT_WORK(&work->work, update_gid_event_work_handler);
+
+       rdma_ip2gid(sa, &work->gid);
+       work->gid_op = gid_op;
+
+       memset(&work->gid_attr, 0, sizeof(work->gid_attr));
+       dev_hold(ndev);
+       work->gid_attr.ndev   = ndev;
+
+       queue_work(ib_wq, &work->work);
+
+       return NOTIFY_DONE;
+}
+
+static int inetaddr_event(struct notifier_block *this, unsigned long event,
+                         void *ptr)
+{
+       struct sockaddr_in      in;
+       struct net_device       *ndev;
+       struct in_ifaddr        *ifa = ptr;
+
+       in.sin_family = AF_INET;
+       in.sin_addr.s_addr = ifa->ifa_address;
+       ndev = ifa->ifa_dev->dev;
+
+       return addr_event(this, event, (struct sockaddr *)&in, ndev);
+}
+
+static int inet6addr_event(struct notifier_block *this, unsigned long event,
+                          void *ptr)
+{
+       struct sockaddr_in6     in6;
+       struct net_device       *ndev;
+       struct inet6_ifaddr     *ifa6 = ptr;
+
+       in6.sin6_family = AF_INET6;
+       in6.sin6_addr = ifa6->addr;
+       ndev = ifa6->idev->dev;
+
+       return addr_event(this, event, (struct sockaddr *)&in6, ndev);
+}
+
+static struct notifier_block nb_netdevice = {
+       .notifier_call = netdevice_event
+};
+
+static struct notifier_block nb_inetaddr = {
+       .notifier_call = inetaddr_event
+};
+
+static struct notifier_block nb_inet6addr = {
+       .notifier_call = inet6addr_event
+};
+
+int __init roce_gid_mgmt_init(void)
+{
+       register_inetaddr_notifier(&nb_inetaddr);
+       if (IS_ENABLED(CONFIG_IPV6))
+               register_inet6addr_notifier(&nb_inet6addr);
+       /* We relay on the netdevice notifier to enumerate all
+        * existing devices in the system. Register to this notifier
+        * last to make sure we will not miss any IP add/del
+        * callbacks.
+        */
+       register_netdevice_notifier(&nb_netdevice);
+
+       return 0;
+}
+
+void __exit roce_gid_mgmt_cleanup(void)
+{
+       if (IS_ENABLED(CONFIG_IPV6))
+               unregister_inet6addr_notifier(&nb_inet6addr);
+       unregister_inetaddr_notifier(&nb_inetaddr);
+       unregister_netdevice_notifier(&nb_netdevice);
+       /* Ensure all gid deletion tasks complete before we go down,
+        * to avoid any reference to free'd memory. By the time
+        * ib-core is removed, all physical devices have been removed,
+        * so no issue with remaining hardware contexts.
+        */
+}