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, u32 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, u32 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 		       u32 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, u32 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, u32 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, u32 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, u32 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, u32 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, u32 port,
264 				 struct net_device *rdma_ndev,
265 				 void *cookie)
266 {
267 	struct net_device *cookie_ndev = cookie;
268 	bool match = false;
269 
270 	if (!rdma_ndev)
271 		return false;
272 
273 	rcu_read_lock();
274 	if (netif_is_bond_master(cookie_ndev) &&
275 	    rdma_is_upper_dev_rcu(rdma_ndev, cookie_ndev))
276 		match = true;
277 	rcu_read_unlock();
278 	return match;
279 }
280 
281 static void update_gid_ip(enum gid_op_type gid_op,
282 			  struct ib_device *ib_dev,
283 			  u32 port, struct net_device *ndev,
284 			  struct sockaddr *addr)
285 {
286 	union ib_gid gid;
287 	struct ib_gid_attr gid_attr;
288 
289 	rdma_ip2gid(addr, &gid);
290 	memset(&gid_attr, 0, sizeof(gid_attr));
291 	gid_attr.ndev = ndev;
292 
293 	update_gid(gid_op, ib_dev, port, &gid, &gid_attr);
294 }
295 
296 static void bond_delete_netdev_default_gids(struct ib_device *ib_dev,
297 					    u32 port,
298 					    struct net_device *rdma_ndev,
299 					    struct net_device *event_ndev)
300 {
301 	struct net_device *real_dev = rdma_vlan_dev_real_dev(event_ndev);
302 	unsigned long gid_type_mask;
303 
304 	if (!rdma_ndev)
305 		return;
306 
307 	if (!real_dev)
308 		real_dev = event_ndev;
309 
310 	rcu_read_lock();
311 
312 	if (((rdma_ndev != event_ndev &&
313 	      !rdma_is_upper_dev_rcu(rdma_ndev, event_ndev)) ||
314 	     is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev)
315 						 ==
316 	     BONDING_SLAVE_STATE_INACTIVE)) {
317 		rcu_read_unlock();
318 		return;
319 	}
320 
321 	rcu_read_unlock();
322 
323 	gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
324 
325 	ib_cache_gid_set_default_gid(ib_dev, port, rdma_ndev,
326 				     gid_type_mask,
327 				     IB_CACHE_GID_DEFAULT_MODE_DELETE);
328 }
329 
330 static void enum_netdev_ipv4_ips(struct ib_device *ib_dev,
331 				 u32 port, struct net_device *ndev)
332 {
333 	const struct in_ifaddr *ifa;
334 	struct in_device *in_dev;
335 	struct sin_list {
336 		struct list_head	list;
337 		struct sockaddr_in	ip;
338 	};
339 	struct sin_list *sin_iter;
340 	struct sin_list *sin_temp;
341 
342 	LIST_HEAD(sin_list);
343 	if (ndev->reg_state >= NETREG_UNREGISTERING)
344 		return;
345 
346 	rcu_read_lock();
347 	in_dev = __in_dev_get_rcu(ndev);
348 	if (!in_dev) {
349 		rcu_read_unlock();
350 		return;
351 	}
352 
353 	in_dev_for_each_ifa_rcu(ifa, in_dev) {
354 		struct sin_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
355 
356 		if (!entry)
357 			continue;
358 
359 		entry->ip.sin_family = AF_INET;
360 		entry->ip.sin_addr.s_addr = ifa->ifa_address;
361 		list_add_tail(&entry->list, &sin_list);
362 	}
363 
364 	rcu_read_unlock();
365 
366 	list_for_each_entry_safe(sin_iter, sin_temp, &sin_list, list) {
367 		update_gid_ip(GID_ADD, ib_dev, port, ndev,
368 			      (struct sockaddr *)&sin_iter->ip);
369 		list_del(&sin_iter->list);
370 		kfree(sin_iter);
371 	}
372 }
373 
374 static void enum_netdev_ipv6_ips(struct ib_device *ib_dev,
375 				 u32 port, struct net_device *ndev)
376 {
377 	struct inet6_ifaddr *ifp;
378 	struct inet6_dev *in6_dev;
379 	struct sin6_list {
380 		struct list_head	list;
381 		struct sockaddr_in6	sin6;
382 	};
383 	struct sin6_list *sin6_iter;
384 	struct sin6_list *sin6_temp;
385 	struct ib_gid_attr gid_attr = {.ndev = ndev};
386 	LIST_HEAD(sin6_list);
387 
388 	if (ndev->reg_state >= NETREG_UNREGISTERING)
389 		return;
390 
391 	in6_dev = in6_dev_get(ndev);
392 	if (!in6_dev)
393 		return;
394 
395 	read_lock_bh(&in6_dev->lock);
396 	list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
397 		struct sin6_list *entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
398 
399 		if (!entry)
400 			continue;
401 
402 		entry->sin6.sin6_family = AF_INET6;
403 		entry->sin6.sin6_addr = ifp->addr;
404 		list_add_tail(&entry->list, &sin6_list);
405 	}
406 	read_unlock_bh(&in6_dev->lock);
407 
408 	in6_dev_put(in6_dev);
409 
410 	list_for_each_entry_safe(sin6_iter, sin6_temp, &sin6_list, list) {
411 		union ib_gid	gid;
412 
413 		rdma_ip2gid((struct sockaddr *)&sin6_iter->sin6, &gid);
414 		update_gid(GID_ADD, ib_dev, port, &gid, &gid_attr);
415 		list_del(&sin6_iter->list);
416 		kfree(sin6_iter);
417 	}
418 }
419 
420 static void _add_netdev_ips(struct ib_device *ib_dev, u32 port,
421 			    struct net_device *ndev)
422 {
423 	enum_netdev_ipv4_ips(ib_dev, port, ndev);
424 	if (IS_ENABLED(CONFIG_IPV6))
425 		enum_netdev_ipv6_ips(ib_dev, port, ndev);
426 }
427 
428 static void add_netdev_ips(struct ib_device *ib_dev, u32 port,
429 			   struct net_device *rdma_ndev, void *cookie)
430 {
431 	_add_netdev_ips(ib_dev, port, cookie);
432 }
433 
434 static void del_netdev_ips(struct ib_device *ib_dev, u32 port,
435 			   struct net_device *rdma_ndev, void *cookie)
436 {
437 	ib_cache_gid_del_all_netdev_gids(ib_dev, port, cookie);
438 }
439 
440 /**
441  * del_default_gids - Delete default GIDs of the event/cookie netdevice
442  * @ib_dev:	RDMA device pointer
443  * @port:	Port of the RDMA device whose GID table to consider
444  * @rdma_ndev:	Unused rdma netdevice
445  * @cookie:	Pointer to event netdevice
446  *
447  * del_default_gids() deletes the default GIDs of the event/cookie netdevice.
448  */
449 static void del_default_gids(struct ib_device *ib_dev, u32 port,
450 			     struct net_device *rdma_ndev, void *cookie)
451 {
452 	struct net_device *cookie_ndev = cookie;
453 	unsigned long gid_type_mask;
454 
455 	gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
456 
457 	ib_cache_gid_set_default_gid(ib_dev, port, cookie_ndev, gid_type_mask,
458 				     IB_CACHE_GID_DEFAULT_MODE_DELETE);
459 }
460 
461 static void add_default_gids(struct ib_device *ib_dev, u32 port,
462 			     struct net_device *rdma_ndev, void *cookie)
463 {
464 	struct net_device *event_ndev = cookie;
465 	unsigned long gid_type_mask;
466 
467 	gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
468 	ib_cache_gid_set_default_gid(ib_dev, port, event_ndev, gid_type_mask,
469 				     IB_CACHE_GID_DEFAULT_MODE_SET);
470 }
471 
472 static void enum_all_gids_of_dev_cb(struct ib_device *ib_dev,
473 				    u32 port,
474 				    struct net_device *rdma_ndev,
475 				    void *cookie)
476 {
477 	struct net *net;
478 	struct net_device *ndev;
479 
480 	/* Lock the rtnl to make sure the netdevs does not move under
481 	 * our feet
482 	 */
483 	rtnl_lock();
484 	down_read(&net_rwsem);
485 	for_each_net(net)
486 		for_each_netdev(net, ndev) {
487 			/*
488 			 * Filter and add default GIDs of the primary netdevice
489 			 * when not in bonding mode, or add default GIDs
490 			 * of bond master device, when in bonding mode.
491 			 */
492 			if (is_ndev_for_default_gid_filter(ib_dev, port,
493 							   rdma_ndev, ndev))
494 				add_default_gids(ib_dev, port, rdma_ndev, ndev);
495 
496 			if (is_eth_port_of_netdev_filter(ib_dev, port,
497 							 rdma_ndev, ndev))
498 				_add_netdev_ips(ib_dev, port, ndev);
499 		}
500 	up_read(&net_rwsem);
501 	rtnl_unlock();
502 }
503 
504 /**
505  * rdma_roce_rescan_device - Rescan all of the network devices in the system
506  * and add their gids, as needed, to the relevant RoCE devices.
507  *
508  * @ib_dev:         the rdma device
509  */
510 void rdma_roce_rescan_device(struct ib_device *ib_dev)
511 {
512 	ib_enum_roce_netdev(ib_dev, pass_all_filter, NULL,
513 			    enum_all_gids_of_dev_cb, NULL);
514 }
515 EXPORT_SYMBOL(rdma_roce_rescan_device);
516 
517 static void callback_for_addr_gid_device_scan(struct ib_device *device,
518 					      u32 port,
519 					      struct net_device *rdma_ndev,
520 					      void *cookie)
521 {
522 	struct update_gid_event_work *parsed = cookie;
523 
524 	return update_gid(parsed->gid_op, device,
525 			  port, &parsed->gid,
526 			  &parsed->gid_attr);
527 }
528 
529 struct upper_list {
530 	struct list_head list;
531 	struct net_device *upper;
532 };
533 
534 static int netdev_upper_walk(struct net_device *upper,
535 			     struct netdev_nested_priv *priv)
536 {
537 	struct upper_list *entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
538 	struct list_head *upper_list = (struct list_head *)priv->data;
539 
540 	if (!entry)
541 		return 0;
542 
543 	list_add_tail(&entry->list, upper_list);
544 	dev_hold(upper);
545 	entry->upper = upper;
546 
547 	return 0;
548 }
549 
550 static void handle_netdev_upper(struct ib_device *ib_dev, u32 port,
551 				void *cookie,
552 				void (*handle_netdev)(struct ib_device *ib_dev,
553 						      u32 port,
554 						      struct net_device *ndev))
555 {
556 	struct net_device *ndev = cookie;
557 	struct netdev_nested_priv priv;
558 	struct upper_list *upper_iter;
559 	struct upper_list *upper_temp;
560 	LIST_HEAD(upper_list);
561 
562 	priv.data = &upper_list;
563 	rcu_read_lock();
564 	netdev_walk_all_upper_dev_rcu(ndev, netdev_upper_walk, &priv);
565 	rcu_read_unlock();
566 
567 	handle_netdev(ib_dev, port, ndev);
568 	list_for_each_entry_safe(upper_iter, upper_temp, &upper_list,
569 				 list) {
570 		handle_netdev(ib_dev, port, upper_iter->upper);
571 		dev_put(upper_iter->upper);
572 		list_del(&upper_iter->list);
573 		kfree(upper_iter);
574 	}
575 }
576 
577 static void _roce_del_all_netdev_gids(struct ib_device *ib_dev, u32 port,
578 				      struct net_device *event_ndev)
579 {
580 	ib_cache_gid_del_all_netdev_gids(ib_dev, port, event_ndev);
581 }
582 
583 static void del_netdev_upper_ips(struct ib_device *ib_dev, u32 port,
584 				 struct net_device *rdma_ndev, void *cookie)
585 {
586 	handle_netdev_upper(ib_dev, port, cookie, _roce_del_all_netdev_gids);
587 }
588 
589 static void add_netdev_upper_ips(struct ib_device *ib_dev, u32 port,
590 				 struct net_device *rdma_ndev, void *cookie)
591 {
592 	handle_netdev_upper(ib_dev, port, cookie, _add_netdev_ips);
593 }
594 
595 static void del_netdev_default_ips_join(struct ib_device *ib_dev, u32 port,
596 					struct net_device *rdma_ndev,
597 					void *cookie)
598 {
599 	struct net_device *master_ndev;
600 
601 	rcu_read_lock();
602 	master_ndev = netdev_master_upper_dev_get_rcu(rdma_ndev);
603 	if (master_ndev)
604 		dev_hold(master_ndev);
605 	rcu_read_unlock();
606 
607 	if (master_ndev) {
608 		bond_delete_netdev_default_gids(ib_dev, port, rdma_ndev,
609 						master_ndev);
610 		dev_put(master_ndev);
611 	}
612 }
613 
614 /* The following functions operate on all IB devices. netdevice_event and
615  * addr_event execute ib_enum_all_roce_netdevs through a work.
616  * ib_enum_all_roce_netdevs iterates through all IB devices.
617  */
618 
619 static void netdevice_event_work_handler(struct work_struct *_work)
620 {
621 	struct netdev_event_work *work =
622 		container_of(_work, struct netdev_event_work, work);
623 	unsigned int i;
624 
625 	for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {
626 		ib_enum_all_roce_netdevs(work->cmds[i].filter,
627 					 work->cmds[i].filter_ndev,
628 					 work->cmds[i].cb,
629 					 work->cmds[i].ndev);
630 		dev_put(work->cmds[i].ndev);
631 		dev_put(work->cmds[i].filter_ndev);
632 	}
633 
634 	kfree(work);
635 }
636 
637 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
638 				struct net_device *ndev)
639 {
640 	unsigned int i;
641 	struct netdev_event_work *ndev_work =
642 		kmalloc(sizeof(*ndev_work), GFP_KERNEL);
643 
644 	if (!ndev_work)
645 		return NOTIFY_DONE;
646 
647 	memcpy(ndev_work->cmds, cmds, sizeof(ndev_work->cmds));
648 	for (i = 0; i < ARRAY_SIZE(ndev_work->cmds) && ndev_work->cmds[i].cb; i++) {
649 		if (!ndev_work->cmds[i].ndev)
650 			ndev_work->cmds[i].ndev = ndev;
651 		if (!ndev_work->cmds[i].filter_ndev)
652 			ndev_work->cmds[i].filter_ndev = ndev;
653 		dev_hold(ndev_work->cmds[i].ndev);
654 		dev_hold(ndev_work->cmds[i].filter_ndev);
655 	}
656 	INIT_WORK(&ndev_work->work, netdevice_event_work_handler);
657 
658 	queue_work(gid_cache_wq, &ndev_work->work);
659 
660 	return NOTIFY_DONE;
661 }
662 
663 static const struct netdev_event_work_cmd add_cmd = {
664 	.cb	= add_netdev_ips,
665 	.filter	= is_eth_port_of_netdev_filter
666 };
667 
668 static const struct netdev_event_work_cmd add_cmd_upper_ips = {
669 	.cb	= add_netdev_upper_ips,
670 	.filter = is_eth_port_of_netdev_filter
671 };
672 
673 static void
674 ndev_event_unlink(struct netdev_notifier_changeupper_info *changeupper_info,
675 		  struct netdev_event_work_cmd *cmds)
676 {
677 	static const struct netdev_event_work_cmd
678 			upper_ips_del_cmd = {
679 				.cb	= del_netdev_upper_ips,
680 				.filter	= upper_device_filter
681 	};
682 
683 	cmds[0] = upper_ips_del_cmd;
684 	cmds[0].ndev = changeupper_info->upper_dev;
685 	cmds[1] = add_cmd;
686 }
687 
688 static const struct netdev_event_work_cmd bonding_default_add_cmd = {
689 	.cb	= add_default_gids,
690 	.filter	= is_upper_ndev_bond_master_filter
691 };
692 
693 static void
694 ndev_event_link(struct net_device *event_ndev,
695 		struct netdev_notifier_changeupper_info *changeupper_info,
696 		struct netdev_event_work_cmd *cmds)
697 {
698 	static const struct netdev_event_work_cmd
699 			bonding_default_del_cmd = {
700 				.cb	= del_default_gids,
701 				.filter	= is_upper_ndev_bond_master_filter
702 			};
703 	/*
704 	 * When a lower netdev is linked to its upper bonding
705 	 * netdev, delete lower slave netdev's default GIDs.
706 	 */
707 	cmds[0] = bonding_default_del_cmd;
708 	cmds[0].ndev = event_ndev;
709 	cmds[0].filter_ndev = changeupper_info->upper_dev;
710 
711 	/* Now add bonding upper device default GIDs */
712 	cmds[1] = bonding_default_add_cmd;
713 	cmds[1].ndev = changeupper_info->upper_dev;
714 	cmds[1].filter_ndev = changeupper_info->upper_dev;
715 
716 	/* Now add bonding upper device IP based GIDs */
717 	cmds[2] = add_cmd_upper_ips;
718 	cmds[2].ndev = changeupper_info->upper_dev;
719 	cmds[2].filter_ndev = changeupper_info->upper_dev;
720 }
721 
722 static void netdevice_event_changeupper(struct net_device *event_ndev,
723 		struct netdev_notifier_changeupper_info *changeupper_info,
724 		struct netdev_event_work_cmd *cmds)
725 {
726 	if (changeupper_info->linking)
727 		ndev_event_link(event_ndev, changeupper_info, cmds);
728 	else
729 		ndev_event_unlink(changeupper_info, cmds);
730 }
731 
732 static const struct netdev_event_work_cmd add_default_gid_cmd = {
733 	.cb	= add_default_gids,
734 	.filter	= is_ndev_for_default_gid_filter,
735 };
736 
737 static int netdevice_event(struct notifier_block *this, unsigned long event,
738 			   void *ptr)
739 {
740 	static const struct netdev_event_work_cmd del_cmd = {
741 		.cb = del_netdev_ips, .filter = pass_all_filter};
742 	static const struct netdev_event_work_cmd
743 			bonding_default_del_cmd_join = {
744 				.cb	= del_netdev_default_ips_join,
745 				.filter	= is_eth_port_inactive_slave_filter
746 			};
747 	static const struct netdev_event_work_cmd
748 			netdev_del_cmd = {
749 				.cb	= del_netdev_ips,
750 				.filter = is_eth_port_of_netdev_filter
751 			};
752 	static const struct netdev_event_work_cmd bonding_event_ips_del_cmd = {
753 		.cb = del_netdev_upper_ips, .filter = upper_device_filter};
754 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
755 	struct netdev_event_work_cmd cmds[ROCE_NETDEV_CALLBACK_SZ] = { {NULL} };
756 
757 	if (ndev->type != ARPHRD_ETHER)
758 		return NOTIFY_DONE;
759 
760 	switch (event) {
761 	case NETDEV_REGISTER:
762 	case NETDEV_UP:
763 		cmds[0] = bonding_default_del_cmd_join;
764 		cmds[1] = add_default_gid_cmd;
765 		cmds[2] = add_cmd;
766 		break;
767 
768 	case NETDEV_UNREGISTER:
769 		if (ndev->reg_state < NETREG_UNREGISTERED)
770 			cmds[0] = del_cmd;
771 		else
772 			return NOTIFY_DONE;
773 		break;
774 
775 	case NETDEV_CHANGEADDR:
776 		cmds[0] = netdev_del_cmd;
777 		if (ndev->reg_state == NETREG_REGISTERED) {
778 			cmds[1] = add_default_gid_cmd;
779 			cmds[2] = add_cmd;
780 		}
781 		break;
782 
783 	case NETDEV_CHANGEUPPER:
784 		netdevice_event_changeupper(ndev,
785 			container_of(ptr, struct netdev_notifier_changeupper_info, info),
786 			cmds);
787 		break;
788 
789 	case NETDEV_BONDING_FAILOVER:
790 		cmds[0] = bonding_event_ips_del_cmd;
791 		/* Add default GIDs of the bond device */
792 		cmds[1] = bonding_default_add_cmd;
793 		/* Add IP based GIDs of the bond device */
794 		cmds[2] = add_cmd_upper_ips;
795 		break;
796 
797 	default:
798 		return NOTIFY_DONE;
799 	}
800 
801 	return netdevice_queue_work(cmds, ndev);
802 }
803 
804 static void update_gid_event_work_handler(struct work_struct *_work)
805 {
806 	struct update_gid_event_work *work =
807 		container_of(_work, struct update_gid_event_work, work);
808 
809 	ib_enum_all_roce_netdevs(is_eth_port_of_netdev_filter,
810 				 work->gid_attr.ndev,
811 				 callback_for_addr_gid_device_scan, work);
812 
813 	dev_put(work->gid_attr.ndev);
814 	kfree(work);
815 }
816 
817 static int addr_event(struct notifier_block *this, unsigned long event,
818 		      struct sockaddr *sa, struct net_device *ndev)
819 {
820 	struct update_gid_event_work *work;
821 	enum gid_op_type gid_op;
822 
823 	if (ndev->type != ARPHRD_ETHER)
824 		return NOTIFY_DONE;
825 
826 	switch (event) {
827 	case NETDEV_UP:
828 		gid_op = GID_ADD;
829 		break;
830 
831 	case NETDEV_DOWN:
832 		gid_op = GID_DEL;
833 		break;
834 
835 	default:
836 		return NOTIFY_DONE;
837 	}
838 
839 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
840 	if (!work)
841 		return NOTIFY_DONE;
842 
843 	INIT_WORK(&work->work, update_gid_event_work_handler);
844 
845 	rdma_ip2gid(sa, &work->gid);
846 	work->gid_op = gid_op;
847 
848 	memset(&work->gid_attr, 0, sizeof(work->gid_attr));
849 	dev_hold(ndev);
850 	work->gid_attr.ndev   = ndev;
851 
852 	queue_work(gid_cache_wq, &work->work);
853 
854 	return NOTIFY_DONE;
855 }
856 
857 static int inetaddr_event(struct notifier_block *this, unsigned long event,
858 			  void *ptr)
859 {
860 	struct sockaddr_in	in;
861 	struct net_device	*ndev;
862 	struct in_ifaddr	*ifa = ptr;
863 
864 	in.sin_family = AF_INET;
865 	in.sin_addr.s_addr = ifa->ifa_address;
866 	ndev = ifa->ifa_dev->dev;
867 
868 	return addr_event(this, event, (struct sockaddr *)&in, ndev);
869 }
870 
871 static int inet6addr_event(struct notifier_block *this, unsigned long event,
872 			   void *ptr)
873 {
874 	struct sockaddr_in6	in6;
875 	struct net_device	*ndev;
876 	struct inet6_ifaddr	*ifa6 = ptr;
877 
878 	in6.sin6_family = AF_INET6;
879 	in6.sin6_addr = ifa6->addr;
880 	ndev = ifa6->idev->dev;
881 
882 	return addr_event(this, event, (struct sockaddr *)&in6, ndev);
883 }
884 
885 static struct notifier_block nb_netdevice = {
886 	.notifier_call = netdevice_event
887 };
888 
889 static struct notifier_block nb_inetaddr = {
890 	.notifier_call = inetaddr_event
891 };
892 
893 static struct notifier_block nb_inet6addr = {
894 	.notifier_call = inet6addr_event
895 };
896 
897 int __init roce_gid_mgmt_init(void)
898 {
899 	gid_cache_wq = alloc_ordered_workqueue("gid-cache-wq", 0);
900 	if (!gid_cache_wq)
901 		return -ENOMEM;
902 
903 	register_inetaddr_notifier(&nb_inetaddr);
904 	if (IS_ENABLED(CONFIG_IPV6))
905 		register_inet6addr_notifier(&nb_inet6addr);
906 	/* We relay on the netdevice notifier to enumerate all
907 	 * existing devices in the system. Register to this notifier
908 	 * last to make sure we will not miss any IP add/del
909 	 * callbacks.
910 	 */
911 	register_netdevice_notifier(&nb_netdevice);
912 
913 	return 0;
914 }
915 
916 void __exit roce_gid_mgmt_cleanup(void)
917 {
918 	if (IS_ENABLED(CONFIG_IPV6))
919 		unregister_inet6addr_notifier(&nb_inet6addr);
920 	unregister_inetaddr_notifier(&nb_inetaddr);
921 	unregister_netdevice_notifier(&nb_netdevice);
922 	/* Ensure all gid deletion tasks complete before we go down,
923 	 * to avoid any reference to free'd memory. By the time
924 	 * ib-core is removed, all physical devices have been removed,
925 	 * so no issue with remaining hardware contexts.
926 	 */
927 	destroy_workqueue(gid_cache_wq);
928 }
929