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 	unsigned long gid_type_mask;
259 
260 	if (!rdma_ndev)
261 		return;
262 
263 	if (!real_dev)
264 		real_dev = event_ndev;
265 
266 	rcu_read_lock();
267 
268 	if (((rdma_ndev != event_ndev &&
269 	      !rdma_is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
270 	     is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev)
271 						 ==
272 	     BONDING_SLAVE_STATE_INACTIVE)) {
273 		rcu_read_unlock();
274 		return;
275 	}
276 
277 	rcu_read_unlock();
278 
279 	gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
280 
281 	ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
282 				     gid_type_mask,
283 				     IB_CACHE_GID_DEFAULT_MODE_DELETE);
284 }
285 
286 static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
287 				 u8 port, struct net_device *ndev)
288 {
289 	struct in_device *in_dev;
290 	struct sin_list {
291 		struct list_head	list;
292 		struct sockaddr_in	ip;
293 	};
294 	struct sin_list *sin_iter;
295 	struct sin_list *sin_temp;
296 
297 	LIST_HEAD(sin_list);
298 	if (ndev->reg_state >= NETREG_UNREGISTERING)
299 		return;
300 
301 	rcu_read_lock();
302 	in_dev = __in_dev_get_rcu(ndev);
303 	if (!in_dev) {
304 		rcu_read_unlock();
305 		return;
306 	}
307 
308 	for_ifa(in_dev) {
309 		struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
310 
311 		if (!entry)
312 			continue;
313 
314 		entry->ip.sin_family = AF_INET;
315 		entry->ip.sin_addr.s_addr = ifa->ifa_address;
316 		list_add_tail(&entry->list, &sin_list);
317 	}
318 	endfor_ifa(in_dev);
319 	rcu_read_unlock();
320 
321 	list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
322 		update_gid_ip(GID_ADD, ib_dev, port, ndev,
323 			      (struct sockaddr *)&sin_iter->ip);
324 		list_del(&sin_iter->list);
325 		kfree(sin_iter);
326 	}
327 }
328 
329 static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
330 				 u8 port, struct net_device *ndev)
331 {
332 	struct inet6_ifaddr *ifp;
333 	struct inet6_dev *in6_dev;
334 	struct sin6_list {
335 		struct list_head	list;
336 		struct sockaddr_in6	sin6;
337 	};
338 	struct sin6_list *sin6_iter;
339 	struct sin6_list *sin6_temp;
340 	struct ib_gid_attr gid_attr = {.ndev = ndev};
341 	LIST_HEAD(sin6_list);
342 
343 	if (ndev->reg_state >= NETREG_UNREGISTERING)
344 		return;
345 
346 	in6_dev = in6_dev_get(ndev);
347 	if (!in6_dev)
348 		return;
349 
350 	read_lock_bh(&in6_dev->lock);
351 	list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
352 		struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
353 
354 		if (!entry)
355 			continue;
356 
357 		entry->sin6.sin6_family = AF_INET6;
358 		entry->sin6.sin6_addr = ifp->addr;
359 		list_add_tail(&entry->list, &sin6_list);
360 	}
361 	read_unlock_bh(&in6_dev->lock);
362 
363 	in6_dev_put(in6_dev);
364 
365 	list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
366 		union ib_gid	gid;
367 
368 		rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
369 		update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
370 		list_del(&sin6_iter->list);
371 		kfree(sin6_iter);
372 	}
373 }
374 
375 static void _add_netdev_ips(struct ib_device *ib_dev, u8 port,
376 			    struct net_device *ndev)
377 {
378 	enum_netdev_ipv4_ips(ib_dev, port, ndev);
379 	if (IS_ENABLED(CONFIG_IPV6))
380 		enum_netdev_ipv6_ips(ib_dev, port, ndev);
381 }
382 
383 static void add_netdev_ips(struct ib_device *ib_dev, u8 port,
384 			   struct net_device *rdma_ndev, void *cookie)
385 {
386 	enum_netdev_default_gids(ib_dev, port, cookie, rdma_ndev);
387 	_add_netdev_ips(ib_dev, port, cookie);
388 }
389 
390 static void del_netdev_ips(struct ib_device *ib_dev, u8 port,
391 			   struct net_device *rdma_ndev, void *cookie)
392 {
393 	ib_cache_gid_del_all_netdev_gids(ib_dev, port, cookie);
394 }
395 
396 static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
397 				    u8 port,
398 				    struct net_device *rdma_ndev,
399 				    void *cookie)
400 {
401 	struct net *net;
402 	struct net_device *ndev;
403 
404 	/* Lock the rtnl to make sure the netdevs does not move under
405 	 * our feet
406 	 */
407 	rtnl_lock();
408 	down_read(&net_rwsem);
409 	for_each_net(net)
410 		for_each_netdev(net, ndev)
411 			if (is_eth_port_of_netdev(ib_dev, port, rdma_ndev, ndev))
412 				add_netdev_ips(ib_dev, port, rdma_ndev, ndev);
413 	up_read(&net_rwsem);
414 	rtnl_unlock();
415 }
416 
417 /**
418  * rdma_roce_rescan_device - Rescan all of the network devices in the system
419  * and add their gids, as needed, to the relevant RoCE devices.
420  *
421  * @device:         the rdma device
422  */
423 void rdma_roce_rescan_device(struct ib_device *ib_dev)
424 {
425 	ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
426 			    enum_all_gids_of_dev_cb, NULL);
427 }
428 EXPORT_SYMBOL(rdma_roce_rescan_device);
429 
430 static void callback_for_addr_gid_device_scan(struct ib_device *device,
431 					      u8 port,
432 					      struct net_device *rdma_ndev,
433 					      void *cookie)
434 {
435 	struct update_gid_event_work *parsed = cookie;
436 
437 	return update_gid(parsed->gid_op, device,
438 			  port, &parsed->gid,
439 			  &parsed->gid_attr);
440 }
441 
442 struct upper_list {
443 	struct list_head list;
444 	struct net_device *upper;
445 };
446 
447 static int netdev_upper_walk(struct net_device *upper, void *data)
448 {
449 	struct upper_list *entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
450 	struct list_head *upper_list = data;
451 
452 	if (!entry)
453 		return 0;
454 
455 	list_add_tail(&entry->list, upper_list);
456 	dev_hold(upper);
457 	entry->upper = upper;
458 
459 	return 0;
460 }
461 
462 static void handle_netdev_upper(struct ib_device *ib_dev, u8 port,
463 				void *cookie,
464 				void (*handle_netdev)(struct ib_device *ib_dev,
465 						      u8 port,
466 						      struct net_device *ndev))
467 {
468 	struct net_device *ndev = cookie;
469 	struct upper_list *upper_iter;
470 	struct upper_list *upper_temp;
471 	LIST_HEAD(upper_list);
472 
473 	rcu_read_lock();
474 	netdev_walk_all_upper_dev_rcu(ndev, netdev_upper_walk, &upper_list);
475 	rcu_read_unlock();
476 
477 	handle_netdev(ib_dev, port, ndev);
478 	list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
479 				 list) {
480 		handle_netdev(ib_dev, port, upper_iter->upper);
481 		dev_put(upper_iter->upper);
482 		list_del(&upper_iter->list);
483 		kfree(upper_iter);
484 	}
485 }
486 
487 static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
488 				      struct net_device *event_ndev)
489 {
490 	ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
491 }
492 
493 static void del_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
494 				 struct net_device *rdma_ndev, void *cookie)
495 {
496 	handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
497 }
498 
499 static void add_netdev_upper_ips(struct ib_device *ib_dev, u8 port,
500 				 struct net_device *rdma_ndev, void *cookie)
501 {
502 	handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
503 }
504 
505 static void del_netdev_default_ips_join(struct ib_device *ib_dev, u8 port,
506 					struct net_device *rdma_ndev,
507 					void *cookie)
508 {
509 	struct net_device *master_ndev;
510 
511 	rcu_read_lock();
512 	master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
513 	if (master_ndev)
514 		dev_hold(master_ndev);
515 	rcu_read_unlock();
516 
517 	if (master_ndev) {
518 		bond_delete_netdev_default_gids(ib_dev, port, master_ndev,
519 						rdma_ndev);
520 		dev_put(master_ndev);
521 	}
522 }
523 
524 static void del_netdev_default_ips(struct ib_device *ib_dev, u8 port,
525 				   struct net_device *rdma_ndev, void *cookie)
526 {
527 	bond_delete_netdev_default_gids(ib_dev, port, cookie, rdma_ndev);
528 }
529 
530 /* The following functions operate on all IB devices. netdevice_event and
531  * addr_event execute ib_enum_all_roce_netdevs through a work.
532  * ib_enum_all_roce_netdevs iterates through all IB devices.
533  */
534 
535 static void netdevice_event_work_handler(struct work_struct *_work)
536 {
537 	struct netdev_event_work *work =
538 		container_of(_work, struct netdev_event_work, work);
539 	unsigned int i;
540 
541 	for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
542 		ib_enum_all_roce_netdevs(work->cmds[i].filter,
543 					 work->cmds[i].filter_ndev,
544 					 work->cmds[i].cb,
545 					 work->cmds[i].ndev);
546 		dev_put(work->cmds[i].ndev);
547 		dev_put(work->cmds[i].filter_ndev);
548 	}
549 
550 	kfree(work);
551 }
552 
553 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
554 				struct net_device *ndev)
555 {
556 	unsigned int i;
557 	struct netdev_event_work *ndev_work =
558 		kmalloc(sizeof(*ndev_work), GFP_KERNEL);
559 
560 	if (!ndev_work)
561 		return NOTIFY_DONE;
562 
563 	memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
564 	for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
565 		if (!ndev_work->cmds[i].ndev)
566 			ndev_work->cmds[i].ndev = ndev;
567 		if (!ndev_work->cmds[i].filter_ndev)
568 			ndev_work->cmds[i].filter_ndev = ndev;
569 		dev_hold(ndev_work->cmds[i].ndev);
570 		dev_hold(ndev_work->cmds[i].filter_ndev);
571 	}
572 	INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
573 
574 	queue_work(gid_cache_wq, &ndev_work->work);
575 
576 	return NOTIFY_DONE;
577 }
578 
579 static const struct netdev_event_work_cmd add_cmd = {
580 	.cb = add_netdev_ips, .filter = is_eth_port_of_netdev};
581 static const struct netdev_event_work_cmd add_cmd_upper_ips = {
582 	.cb = add_netdev_upper_ips, .filter = is_eth_port_of_netdev};
583 
584 static void netdevice_event_changeupper(struct netdev_notifier_changeupper_info *changeupper_info,
585 					struct netdev_event_work_cmd *cmds)
586 {
587 	static const struct netdev_event_work_cmd upper_ips_del_cmd = {
588 		.cb = del_netdev_upper_ips, .filter = upper_device_filter};
589 	static const struct netdev_event_work_cmd bonding_default_del_cmd = {
590 		.cb = del_netdev_default_ips, .filter = is_eth_port_inactive_slave};
591 
592 	if (changeupper_info->linking == false) {
593 		cmds[0] = upper_ips_del_cmd;
594 		cmds[0].ndev = changeupper_info->upper_dev;
595 		cmds[1] = add_cmd;
596 	} else {
597 		cmds[0] = bonding_default_del_cmd;
598 		cmds[0].ndev = changeupper_info->upper_dev;
599 		cmds[1] = add_cmd_upper_ips;
600 		cmds[1].ndev = changeupper_info->upper_dev;
601 		cmds[1].filter_ndev = changeupper_info->upper_dev;
602 	}
603 }
604 
605 static int netdevice_event(struct notifier_block *this, unsigned long event,
606 			   void *ptr)
607 {
608 	static const struct netdev_event_work_cmd del_cmd = {
609 		.cb = del_netdev_ips, .filter = pass_all_filter};
610 	static const struct netdev_event_work_cmd bonding_default_del_cmd_join = {
611 		.cb = del_netdev_default_ips_join, .filter = is_eth_port_inactive_slave};
612 	static const struct netdev_event_work_cmd default_del_cmd = {
613 		.cb = del_netdev_default_ips, .filter = pass_all_filter};
614 	static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
615 		.cb = del_netdev_upper_ips, .filter = upper_device_filter};
616 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
617 	struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
618 
619 	if (ndev->type != ARPHRD_ETHER)
620 		return NOTIFY_DONE;
621 
622 	switch (event) {
623 	case NETDEV_REGISTER:
624 	case NETDEV_UP:
625 		cmds[0] = bonding_default_del_cmd_join;
626 		cmds[1] = add_cmd;
627 		break;
628 
629 	case NETDEV_UNREGISTER:
630 		if (ndev->reg_state < NETREG_UNREGISTERED)
631 			cmds[0] = del_cmd;
632 		else
633 			return NOTIFY_DONE;
634 		break;
635 
636 	case NETDEV_CHANGEADDR:
637 		cmds[0] = default_del_cmd;
638 		cmds[1] = add_cmd;
639 		break;
640 
641 	case NETDEV_CHANGEUPPER:
642 		netdevice_event_changeupper(
643 			container_of(ptr, struct netdev_notifier_changeupper_info, info),
644 			cmds);
645 		break;
646 
647 	case NETDEV_BONDING_FAILOVER:
648 		cmds[0] = bonding_event_ips_del_cmd;
649 		cmds[1] = bonding_default_del_cmd_join;
650 		cmds[2] = add_cmd_upper_ips;
651 		break;
652 
653 	default:
654 		return NOTIFY_DONE;
655 	}
656 
657 	return netdevice_queue_work(cmds, ndev);
658 }
659 
660 static void update_gid_event_work_handler(struct work_struct *_work)
661 {
662 	struct update_gid_event_work *work =
663 		container_of(_work, struct update_gid_event_work, work);
664 
665 	ib_enum_all_roce_netdevs(is_eth_port_of_netdev, work->gid_attr.ndev,
666 				 callback_for_addr_gid_device_scan, work);
667 
668 	dev_put(work->gid_attr.ndev);
669 	kfree(work);
670 }
671 
672 static int addr_event(struct notifier_block *this, unsigned long event,
673 		      struct sockaddr *sa, struct net_device *ndev)
674 {
675 	struct update_gid_event_work *work;
676 	enum gid_op_type gid_op;
677 
678 	if (ndev->type != ARPHRD_ETHER)
679 		return NOTIFY_DONE;
680 
681 	switch (event) {
682 	case NETDEV_UP:
683 		gid_op = GID_ADD;
684 		break;
685 
686 	case NETDEV_DOWN:
687 		gid_op = GID_DEL;
688 		break;
689 
690 	default:
691 		return NOTIFY_DONE;
692 	}
693 
694 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
695 	if (!work)
696 		return NOTIFY_DONE;
697 
698 	INIT_WORK(&work->work, update_gid_event_work_handler);
699 
700 	rdma_ip2gid(sa, &work->gid);
701 	work->gid_op = gid_op;
702 
703 	memset(&work->gid_attr, 0, sizeof(work->gid_attr));
704 	dev_hold(ndev);
705 	work->gid_attr.ndev   = ndev;
706 
707 	queue_work(gid_cache_wq, &work->work);
708 
709 	return NOTIFY_DONE;
710 }
711 
712 static int inetaddr_event(struct notifier_block *this, unsigned long event,
713 			  void *ptr)
714 {
715 	struct sockaddr_in	in;
716 	struct net_device	*ndev;
717 	struct in_ifaddr	*ifa = ptr;
718 
719 	in.sin_family = AF_INET;
720 	in.sin_addr.s_addr = ifa->ifa_address;
721 	ndev = ifa->ifa_dev->dev;
722 
723 	return addr_event(this, event, (struct sockaddr *)&in, ndev);
724 }
725 
726 static int inet6addr_event(struct notifier_block *this, unsigned long event,
727 			   void *ptr)
728 {
729 	struct sockaddr_in6	in6;
730 	struct net_device	*ndev;
731 	struct inet6_ifaddr	*ifa6 = ptr;
732 
733 	in6.sin6_family = AF_INET6;
734 	in6.sin6_addr = ifa6->addr;
735 	ndev = ifa6->idev->dev;
736 
737 	return addr_event(this, event, (struct sockaddr *)&in6, ndev);
738 }
739 
740 static struct notifier_block nb_netdevice = {
741 	.notifier_call = netdevice_event
742 };
743 
744 static struct notifier_block nb_inetaddr = {
745 	.notifier_call = inetaddr_event
746 };
747 
748 static struct notifier_block nb_inet6addr = {
749 	.notifier_call = inet6addr_event
750 };
751 
752 int __init roce_gid_mgmt_init(void)
753 {
754 	gid_cache_wq = alloc_ordered_workqueue("gid-cache-wq", 0);
755 	if (!gid_cache_wq)
756 		return -ENOMEM;
757 
758 	register_inetaddr_notifier(&nb_inetaddr);
759 	if (IS_ENABLED(CONFIG_IPV6))
760 		register_inet6addr_notifier(&nb_inet6addr);
761 	/* We relay on the netdevice notifier to enumerate all
762 	 * existing devices in the system. Register to this notifier
763 	 * last to make sure we will not miss any IP add/del
764 	 * callbacks.
765 	 */
766 	register_netdevice_notifier(&nb_netdevice);
767 
768 	return 0;
769 }
770 
771 void __exit roce_gid_mgmt_cleanup(void)
772 {
773 	if (IS_ENABLED(CONFIG_IPV6))
774 		unregister_inet6addr_notifier(&nb_inet6addr);
775 	unregister_inetaddr_notifier(&nb_inetaddr);
776 	unregister_netdevice_notifier(&nb_netdevice);
777 	/* Ensure all gid deletion tasks complete before we go down,
778 	 * to avoid any reference to free'd memory. By the time
779 	 * ib-core is removed, all physical devices have been removed,
780 	 * so no issue with remaining hardware contexts.
781 	 */
782 	destroy_workqueue(gid_cache_wq);
783 }
784