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