1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include "ipoib.h"
36 
37 #include <linux/module.h>
38 
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
43 
44 #include <linux/if_arp.h>	/* For ARPHRD_xxx */
45 
46 #include <linux/ip.h>
47 #include <linux/in.h>
48 
49 #include <linux/jhash.h>
50 #include <net/arp.h>
51 #include <net/addrconf.h>
52 #include <linux/inetdevice.h>
53 #include <rdma/ib_cache.h>
54 
55 #define DRV_VERSION "1.0.0"
56 
57 const char ipoib_driver_version[] = DRV_VERSION;
58 
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
61 MODULE_LICENSE("Dual BSD/GPL");
62 
63 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
64 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
65 
66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
70 
71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
72 int ipoib_debug_level;
73 
74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
76 #endif
77 
78 struct ipoib_path_iter {
79 	struct net_device *dev;
80 	struct ipoib_path  path;
81 };
82 
83 static const u8 ipv4_bcast_addr[] = {
84 	0x00, 0xff, 0xff, 0xff,
85 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
86 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
87 };
88 
89 struct workqueue_struct *ipoib_workqueue;
90 
91 struct ib_sa_client ipoib_sa_client;
92 
93 static void ipoib_add_one(struct ib_device *device);
94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
95 static void ipoib_neigh_reclaim(struct rcu_head *rp);
96 static struct net_device *ipoib_get_net_dev_by_params(
97 		struct ib_device *dev, u8 port, u16 pkey,
98 		const union ib_gid *gid, const struct sockaddr *addr,
99 		void *client_data);
100 static int ipoib_set_mac(struct net_device *dev, void *addr);
101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
102 		       int cmd);
103 
104 static struct ib_client ipoib_client = {
105 	.name   = "ipoib",
106 	.add    = ipoib_add_one,
107 	.remove = ipoib_remove_one,
108 	.get_net_dev_by_params = ipoib_get_net_dev_by_params,
109 };
110 
111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
112 static int ipoib_netdev_event(struct notifier_block *this,
113 			      unsigned long event, void *ptr)
114 {
115 	struct netdev_notifier_info *ni = ptr;
116 	struct net_device *dev = ni->dev;
117 
118 	if (dev->netdev_ops->ndo_open != ipoib_open)
119 		return NOTIFY_DONE;
120 
121 	switch (event) {
122 	case NETDEV_REGISTER:
123 		ipoib_create_debug_files(dev);
124 		break;
125 	case NETDEV_CHANGENAME:
126 		ipoib_delete_debug_files(dev);
127 		ipoib_create_debug_files(dev);
128 		break;
129 	case NETDEV_UNREGISTER:
130 		ipoib_delete_debug_files(dev);
131 		break;
132 	}
133 
134 	return NOTIFY_DONE;
135 }
136 #endif
137 
138 int ipoib_open(struct net_device *dev)
139 {
140 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
141 
142 	ipoib_dbg(priv, "bringing up interface\n");
143 
144 	netif_carrier_off(dev);
145 
146 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
147 
148 	priv->sm_fullmember_sendonly_support = false;
149 
150 	if (ipoib_ib_dev_open(dev)) {
151 		if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
152 			return 0;
153 		goto err_disable;
154 	}
155 
156 	ipoib_ib_dev_up(dev);
157 
158 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
159 		struct ipoib_dev_priv *cpriv;
160 
161 		/* Bring up any child interfaces too */
162 		down_read(&priv->vlan_rwsem);
163 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
164 			int flags;
165 
166 			flags = cpriv->dev->flags;
167 			if (flags & IFF_UP)
168 				continue;
169 
170 			dev_change_flags(cpriv->dev, flags | IFF_UP, NULL);
171 		}
172 		up_read(&priv->vlan_rwsem);
173 	}
174 
175 	netif_start_queue(dev);
176 
177 	return 0;
178 
179 err_disable:
180 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
181 
182 	return -EINVAL;
183 }
184 
185 static int ipoib_stop(struct net_device *dev)
186 {
187 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
188 
189 	ipoib_dbg(priv, "stopping interface\n");
190 
191 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
192 
193 	netif_stop_queue(dev);
194 
195 	ipoib_ib_dev_down(dev);
196 	ipoib_ib_dev_stop(dev);
197 
198 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
199 		struct ipoib_dev_priv *cpriv;
200 
201 		/* Bring down any child interfaces too */
202 		down_read(&priv->vlan_rwsem);
203 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
204 			int flags;
205 
206 			flags = cpriv->dev->flags;
207 			if (!(flags & IFF_UP))
208 				continue;
209 
210 			dev_change_flags(cpriv->dev, flags & ~IFF_UP, NULL);
211 		}
212 		up_read(&priv->vlan_rwsem);
213 	}
214 
215 	return 0;
216 }
217 
218 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
219 {
220 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
221 
222 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
223 		features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
224 
225 	return features;
226 }
227 
228 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
229 {
230 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
231 	int ret = 0;
232 
233 	/* dev->mtu > 2K ==> connected mode */
234 	if (ipoib_cm_admin_enabled(dev)) {
235 		if (new_mtu > ipoib_cm_max_mtu(dev))
236 			return -EINVAL;
237 
238 		if (new_mtu > priv->mcast_mtu)
239 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
240 				   priv->mcast_mtu);
241 
242 		dev->mtu = new_mtu;
243 		return 0;
244 	}
245 
246 	if (new_mtu < (ETH_MIN_MTU + IPOIB_ENCAP_LEN) ||
247 	    new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
248 		return -EINVAL;
249 
250 	priv->admin_mtu = new_mtu;
251 
252 	if (priv->mcast_mtu < priv->admin_mtu)
253 		ipoib_dbg(priv, "MTU must be smaller than the underlying "
254 				"link layer MTU - 4 (%u)\n", priv->mcast_mtu);
255 
256 	new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
257 
258 	if (priv->rn_ops->ndo_change_mtu) {
259 		bool carrier_status = netif_carrier_ok(dev);
260 
261 		netif_carrier_off(dev);
262 
263 		/* notify lower level on the real mtu */
264 		ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
265 
266 		if (carrier_status)
267 			netif_carrier_on(dev);
268 	} else {
269 		dev->mtu = new_mtu;
270 	}
271 
272 	return ret;
273 }
274 
275 static void ipoib_get_stats(struct net_device *dev,
276 			    struct rtnl_link_stats64 *stats)
277 {
278 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
279 
280 	if (priv->rn_ops->ndo_get_stats64)
281 		priv->rn_ops->ndo_get_stats64(dev, stats);
282 	else
283 		netdev_stats_to_stats64(stats, &dev->stats);
284 }
285 
286 /* Called with an RCU read lock taken */
287 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
288 					struct net_device *dev)
289 {
290 	struct net *net = dev_net(dev);
291 	struct in_device *in_dev;
292 	struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
293 	struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
294 	__be32 ret_addr;
295 
296 	switch (addr->sa_family) {
297 	case AF_INET:
298 		in_dev = in_dev_get(dev);
299 		if (!in_dev)
300 			return false;
301 
302 		ret_addr = inet_confirm_addr(net, in_dev, 0,
303 					     addr_in->sin_addr.s_addr,
304 					     RT_SCOPE_HOST);
305 		in_dev_put(in_dev);
306 		if (ret_addr)
307 			return true;
308 
309 		break;
310 	case AF_INET6:
311 		if (IS_ENABLED(CONFIG_IPV6) &&
312 		    ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
313 			return true;
314 
315 		break;
316 	}
317 	return false;
318 }
319 
320 /**
321  * Find the master net_device on top of the given net_device.
322  * @dev: base IPoIB net_device
323  *
324  * Returns the master net_device with a reference held, or the same net_device
325  * if no master exists.
326  */
327 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
328 {
329 	struct net_device *master;
330 
331 	rcu_read_lock();
332 	master = netdev_master_upper_dev_get_rcu(dev);
333 	if (master)
334 		dev_hold(master);
335 	rcu_read_unlock();
336 
337 	if (master)
338 		return master;
339 
340 	dev_hold(dev);
341 	return dev;
342 }
343 
344 struct ipoib_walk_data {
345 	const struct sockaddr *addr;
346 	struct net_device *result;
347 };
348 
349 static int ipoib_upper_walk(struct net_device *upper, void *_data)
350 {
351 	struct ipoib_walk_data *data = _data;
352 	int ret = 0;
353 
354 	if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
355 		dev_hold(upper);
356 		data->result = upper;
357 		ret = 1;
358 	}
359 
360 	return ret;
361 }
362 
363 /**
364  * Find a net_device matching the given address, which is an upper device of
365  * the given net_device.
366  * @addr: IP address to look for.
367  * @dev: base IPoIB net_device
368  *
369  * If found, returns the net_device with a reference held. Otherwise return
370  * NULL.
371  */
372 static struct net_device *ipoib_get_net_dev_match_addr(
373 		const struct sockaddr *addr, struct net_device *dev)
374 {
375 	struct ipoib_walk_data data = {
376 		.addr = addr,
377 	};
378 
379 	rcu_read_lock();
380 	if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
381 		dev_hold(dev);
382 		data.result = dev;
383 		goto out;
384 	}
385 
386 	netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
387 out:
388 	rcu_read_unlock();
389 	return data.result;
390 }
391 
392 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
393  * pkey_index and address, if one exists.
394  *
395  * @found_net_dev: contains a matching net_device if the return value >= 1,
396  * with a reference held. */
397 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
398 				     const union ib_gid *gid,
399 				     u16 pkey_index,
400 				     const struct sockaddr *addr,
401 				     int nesting,
402 				     struct net_device **found_net_dev)
403 {
404 	struct ipoib_dev_priv *child_priv;
405 	struct net_device *net_dev = NULL;
406 	int matches = 0;
407 
408 	if (priv->pkey_index == pkey_index &&
409 	    (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
410 		if (!addr) {
411 			net_dev = ipoib_get_master_net_dev(priv->dev);
412 		} else {
413 			/* Verify the net_device matches the IP address, as
414 			 * IPoIB child devices currently share a GID. */
415 			net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
416 		}
417 		if (net_dev) {
418 			if (!*found_net_dev)
419 				*found_net_dev = net_dev;
420 			else
421 				dev_put(net_dev);
422 			++matches;
423 		}
424 	}
425 
426 	/* Check child interfaces */
427 	down_read_nested(&priv->vlan_rwsem, nesting);
428 	list_for_each_entry(child_priv, &priv->child_intfs, list) {
429 		matches += ipoib_match_gid_pkey_addr(child_priv, gid,
430 						    pkey_index, addr,
431 						    nesting + 1,
432 						    found_net_dev);
433 		if (matches > 1)
434 			break;
435 	}
436 	up_read(&priv->vlan_rwsem);
437 
438 	return matches;
439 }
440 
441 /* Returns the number of matching net_devs found (between 0 and 2). Also
442  * return the matching net_device in the @net_dev parameter, holding a
443  * reference to the net_device, if the number of matches >= 1 */
444 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
445 					 u16 pkey_index,
446 					 const union ib_gid *gid,
447 					 const struct sockaddr *addr,
448 					 struct net_device **net_dev)
449 {
450 	struct ipoib_dev_priv *priv;
451 	int matches = 0;
452 
453 	*net_dev = NULL;
454 
455 	list_for_each_entry(priv, dev_list, list) {
456 		if (priv->port != port)
457 			continue;
458 
459 		matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
460 						     addr, 0, net_dev);
461 		if (matches > 1)
462 			break;
463 	}
464 
465 	return matches;
466 }
467 
468 static struct net_device *ipoib_get_net_dev_by_params(
469 		struct ib_device *dev, u8 port, u16 pkey,
470 		const union ib_gid *gid, const struct sockaddr *addr,
471 		void *client_data)
472 {
473 	struct net_device *net_dev;
474 	struct list_head *dev_list = client_data;
475 	u16 pkey_index;
476 	int matches;
477 	int ret;
478 
479 	if (!rdma_protocol_ib(dev, port))
480 		return NULL;
481 
482 	ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
483 	if (ret)
484 		return NULL;
485 
486 	if (!dev_list)
487 		return NULL;
488 
489 	/* See if we can find a unique device matching the L2 parameters */
490 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
491 						gid, NULL, &net_dev);
492 
493 	switch (matches) {
494 	case 0:
495 		return NULL;
496 	case 1:
497 		return net_dev;
498 	}
499 
500 	dev_put(net_dev);
501 
502 	/* Couldn't find a unique device with L2 parameters only. Use L3
503 	 * address to uniquely match the net device */
504 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
505 						gid, addr, &net_dev);
506 	switch (matches) {
507 	case 0:
508 		return NULL;
509 	default:
510 		dev_warn_ratelimited(&dev->dev,
511 				     "duplicate IP address detected\n");
512 		/* Fall through */
513 	case 1:
514 		return net_dev;
515 	}
516 }
517 
518 int ipoib_set_mode(struct net_device *dev, const char *buf)
519 {
520 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
521 
522 	if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
523 	     !strcmp(buf, "connected\n")) ||
524 	     (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
525 	     !strcmp(buf, "datagram\n"))) {
526 		return 0;
527 	}
528 
529 	/* flush paths if we switch modes so that connections are restarted */
530 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
531 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
532 		ipoib_warn(priv, "enabling connected mode "
533 			   "will cause multicast packet drops\n");
534 		netdev_update_features(dev);
535 		dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
536 		rtnl_unlock();
537 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
538 
539 		ipoib_flush_paths(dev);
540 		return (!rtnl_trylock()) ? -EBUSY : 0;
541 	}
542 
543 	if (!strcmp(buf, "datagram\n")) {
544 		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
545 		netdev_update_features(dev);
546 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
547 		rtnl_unlock();
548 		ipoib_flush_paths(dev);
549 		return (!rtnl_trylock()) ? -EBUSY : 0;
550 	}
551 
552 	return -EINVAL;
553 }
554 
555 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
556 {
557 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
558 	struct rb_node *n = priv->path_tree.rb_node;
559 	struct ipoib_path *path;
560 	int ret;
561 
562 	while (n) {
563 		path = rb_entry(n, struct ipoib_path, rb_node);
564 
565 		ret = memcmp(gid, path->pathrec.dgid.raw,
566 			     sizeof (union ib_gid));
567 
568 		if (ret < 0)
569 			n = n->rb_left;
570 		else if (ret > 0)
571 			n = n->rb_right;
572 		else
573 			return path;
574 	}
575 
576 	return NULL;
577 }
578 
579 static int __path_add(struct net_device *dev, struct ipoib_path *path)
580 {
581 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
582 	struct rb_node **n = &priv->path_tree.rb_node;
583 	struct rb_node *pn = NULL;
584 	struct ipoib_path *tpath;
585 	int ret;
586 
587 	while (*n) {
588 		pn = *n;
589 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
590 
591 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
592 			     sizeof (union ib_gid));
593 		if (ret < 0)
594 			n = &pn->rb_left;
595 		else if (ret > 0)
596 			n = &pn->rb_right;
597 		else
598 			return -EEXIST;
599 	}
600 
601 	rb_link_node(&path->rb_node, pn, n);
602 	rb_insert_color(&path->rb_node, &priv->path_tree);
603 
604 	list_add_tail(&path->list, &priv->path_list);
605 
606 	return 0;
607 }
608 
609 static void path_free(struct net_device *dev, struct ipoib_path *path)
610 {
611 	struct sk_buff *skb;
612 
613 	while ((skb = __skb_dequeue(&path->queue)))
614 		dev_kfree_skb_irq(skb);
615 
616 	ipoib_dbg(ipoib_priv(dev), "%s\n", __func__);
617 
618 	/* remove all neigh connected to this path */
619 	ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
620 
621 	if (path->ah)
622 		ipoib_put_ah(path->ah);
623 
624 	kfree(path);
625 }
626 
627 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
628 
629 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
630 {
631 	struct ipoib_path_iter *iter;
632 
633 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
634 	if (!iter)
635 		return NULL;
636 
637 	iter->dev = dev;
638 	memset(iter->path.pathrec.dgid.raw, 0, 16);
639 
640 	if (ipoib_path_iter_next(iter)) {
641 		kfree(iter);
642 		return NULL;
643 	}
644 
645 	return iter;
646 }
647 
648 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
649 {
650 	struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
651 	struct rb_node *n;
652 	struct ipoib_path *path;
653 	int ret = 1;
654 
655 	spin_lock_irq(&priv->lock);
656 
657 	n = rb_first(&priv->path_tree);
658 
659 	while (n) {
660 		path = rb_entry(n, struct ipoib_path, rb_node);
661 
662 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
663 			   sizeof (union ib_gid)) < 0) {
664 			iter->path = *path;
665 			ret = 0;
666 			break;
667 		}
668 
669 		n = rb_next(n);
670 	}
671 
672 	spin_unlock_irq(&priv->lock);
673 
674 	return ret;
675 }
676 
677 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
678 			  struct ipoib_path *path)
679 {
680 	*path = iter->path;
681 }
682 
683 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
684 
685 void ipoib_mark_paths_invalid(struct net_device *dev)
686 {
687 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
688 	struct ipoib_path *path, *tp;
689 
690 	spin_lock_irq(&priv->lock);
691 
692 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
693 		ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
694 			  be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
695 			  path->pathrec.dgid.raw);
696 		if (path->ah)
697 			path->ah->valid = 0;
698 	}
699 
700 	spin_unlock_irq(&priv->lock);
701 }
702 
703 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
704 {
705 	struct ipoib_pseudo_header *phdr;
706 
707 	phdr = skb_push(skb, sizeof(*phdr));
708 	memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
709 }
710 
711 void ipoib_flush_paths(struct net_device *dev)
712 {
713 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
714 	struct ipoib_path *path, *tp;
715 	LIST_HEAD(remove_list);
716 	unsigned long flags;
717 
718 	netif_tx_lock_bh(dev);
719 	spin_lock_irqsave(&priv->lock, flags);
720 
721 	list_splice_init(&priv->path_list, &remove_list);
722 
723 	list_for_each_entry(path, &remove_list, list)
724 		rb_erase(&path->rb_node, &priv->path_tree);
725 
726 	list_for_each_entry_safe(path, tp, &remove_list, list) {
727 		if (path->query)
728 			ib_sa_cancel_query(path->query_id, path->query);
729 		spin_unlock_irqrestore(&priv->lock, flags);
730 		netif_tx_unlock_bh(dev);
731 		wait_for_completion(&path->done);
732 		path_free(dev, path);
733 		netif_tx_lock_bh(dev);
734 		spin_lock_irqsave(&priv->lock, flags);
735 	}
736 
737 	spin_unlock_irqrestore(&priv->lock, flags);
738 	netif_tx_unlock_bh(dev);
739 }
740 
741 static void path_rec_completion(int status,
742 				struct sa_path_rec *pathrec,
743 				void *path_ptr)
744 {
745 	struct ipoib_path *path = path_ptr;
746 	struct net_device *dev = path->dev;
747 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
748 	struct ipoib_ah *ah = NULL;
749 	struct ipoib_ah *old_ah = NULL;
750 	struct ipoib_neigh *neigh, *tn;
751 	struct sk_buff_head skqueue;
752 	struct sk_buff *skb;
753 	unsigned long flags;
754 
755 	if (!status)
756 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
757 			  be32_to_cpu(sa_path_get_dlid(pathrec)),
758 			  pathrec->dgid.raw);
759 	else
760 		ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
761 			  status, path->pathrec.dgid.raw);
762 
763 	skb_queue_head_init(&skqueue);
764 
765 	if (!status) {
766 		struct rdma_ah_attr av;
767 
768 		if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
769 					       pathrec, &av, NULL)) {
770 			ah = ipoib_create_ah(dev, priv->pd, &av);
771 			rdma_destroy_ah_attr(&av);
772 		}
773 	}
774 
775 	spin_lock_irqsave(&priv->lock, flags);
776 
777 	if (!IS_ERR_OR_NULL(ah)) {
778 		/*
779 		 * pathrec.dgid is used as the database key from the LLADDR,
780 		 * it must remain unchanged even if the SA returns a different
781 		 * GID to use in the AH.
782 		 */
783 		if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
784 			   sizeof(union ib_gid))) {
785 			ipoib_dbg(
786 				priv,
787 				"%s got PathRec for gid %pI6 while asked for %pI6\n",
788 				dev->name, pathrec->dgid.raw,
789 				path->pathrec.dgid.raw);
790 			memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
791 			       sizeof(union ib_gid));
792 		}
793 
794 		path->pathrec = *pathrec;
795 
796 		old_ah   = path->ah;
797 		path->ah = ah;
798 
799 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
800 			  ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
801 			  pathrec->sl);
802 
803 		while ((skb = __skb_dequeue(&path->queue)))
804 			__skb_queue_tail(&skqueue, skb);
805 
806 		list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
807 			if (neigh->ah) {
808 				WARN_ON(neigh->ah != old_ah);
809 				/*
810 				 * Dropping the ah reference inside
811 				 * priv->lock is safe here, because we
812 				 * will hold one more reference from
813 				 * the original value of path->ah (ie
814 				 * old_ah).
815 				 */
816 				ipoib_put_ah(neigh->ah);
817 			}
818 			kref_get(&path->ah->ref);
819 			neigh->ah = path->ah;
820 
821 			if (ipoib_cm_enabled(dev, neigh->daddr)) {
822 				if (!ipoib_cm_get(neigh))
823 					ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
824 									       path,
825 									       neigh));
826 				if (!ipoib_cm_get(neigh)) {
827 					ipoib_neigh_free(neigh);
828 					continue;
829 				}
830 			}
831 
832 			while ((skb = __skb_dequeue(&neigh->queue)))
833 				__skb_queue_tail(&skqueue, skb);
834 		}
835 		path->ah->valid = 1;
836 	}
837 
838 	path->query = NULL;
839 	complete(&path->done);
840 
841 	spin_unlock_irqrestore(&priv->lock, flags);
842 
843 	if (IS_ERR_OR_NULL(ah))
844 		ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
845 
846 	if (old_ah)
847 		ipoib_put_ah(old_ah);
848 
849 	while ((skb = __skb_dequeue(&skqueue))) {
850 		int ret;
851 		skb->dev = dev;
852 		ret = dev_queue_xmit(skb);
853 		if (ret)
854 			ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
855 				   __func__, ret);
856 	}
857 }
858 
859 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
860 			  void *gid)
861 {
862 	path->dev = priv->dev;
863 
864 	if (rdma_cap_opa_ah(priv->ca, priv->port))
865 		path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
866 	else
867 		path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
868 
869 	memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
870 	path->pathrec.sgid	    = priv->local_gid;
871 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
872 	path->pathrec.numb_path     = 1;
873 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
874 }
875 
876 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
877 {
878 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
879 	struct ipoib_path *path;
880 
881 	if (!priv->broadcast)
882 		return NULL;
883 
884 	path = kzalloc(sizeof(*path), GFP_ATOMIC);
885 	if (!path)
886 		return NULL;
887 
888 	skb_queue_head_init(&path->queue);
889 
890 	INIT_LIST_HEAD(&path->neigh_list);
891 
892 	init_path_rec(priv, path, gid);
893 
894 	return path;
895 }
896 
897 static int path_rec_start(struct net_device *dev,
898 			  struct ipoib_path *path)
899 {
900 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
901 
902 	ipoib_dbg(priv, "Start path record lookup for %pI6\n",
903 		  path->pathrec.dgid.raw);
904 
905 	init_completion(&path->done);
906 
907 	path->query_id =
908 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
909 				   &path->pathrec,
910 				   IB_SA_PATH_REC_DGID		|
911 				   IB_SA_PATH_REC_SGID		|
912 				   IB_SA_PATH_REC_NUMB_PATH	|
913 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
914 				   IB_SA_PATH_REC_PKEY,
915 				   1000, GFP_ATOMIC,
916 				   path_rec_completion,
917 				   path, &path->query);
918 	if (path->query_id < 0) {
919 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
920 		path->query = NULL;
921 		complete(&path->done);
922 		return path->query_id;
923 	}
924 
925 	return 0;
926 }
927 
928 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
929 			       struct net_device *dev)
930 {
931 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
932 	struct ipoib_path *path;
933 	unsigned long flags;
934 
935 	spin_lock_irqsave(&priv->lock, flags);
936 
937 	path = __path_find(dev, daddr + 4);
938 	if (!path)
939 		goto out;
940 	if (!path->query)
941 		path_rec_start(dev, path);
942 out:
943 	spin_unlock_irqrestore(&priv->lock, flags);
944 }
945 
946 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
947 					  struct net_device *dev)
948 {
949 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
950 	struct rdma_netdev *rn = netdev_priv(dev);
951 	struct ipoib_path *path;
952 	struct ipoib_neigh *neigh;
953 	unsigned long flags;
954 
955 	spin_lock_irqsave(&priv->lock, flags);
956 	neigh = ipoib_neigh_alloc(daddr, dev);
957 	if (!neigh) {
958 		spin_unlock_irqrestore(&priv->lock, flags);
959 		++dev->stats.tx_dropped;
960 		dev_kfree_skb_any(skb);
961 		return NULL;
962 	}
963 
964 	/* To avoid race condition, make sure that the
965 	 * neigh will be added only once.
966 	 */
967 	if (unlikely(!list_empty(&neigh->list))) {
968 		spin_unlock_irqrestore(&priv->lock, flags);
969 		return neigh;
970 	}
971 
972 	path = __path_find(dev, daddr + 4);
973 	if (!path) {
974 		path = path_rec_create(dev, daddr + 4);
975 		if (!path)
976 			goto err_path;
977 
978 		__path_add(dev, path);
979 	}
980 
981 	list_add_tail(&neigh->list, &path->neigh_list);
982 
983 	if (path->ah && path->ah->valid) {
984 		kref_get(&path->ah->ref);
985 		neigh->ah = path->ah;
986 
987 		if (ipoib_cm_enabled(dev, neigh->daddr)) {
988 			if (!ipoib_cm_get(neigh))
989 				ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
990 			if (!ipoib_cm_get(neigh)) {
991 				ipoib_neigh_free(neigh);
992 				goto err_drop;
993 			}
994 			if (skb_queue_len(&neigh->queue) <
995 			    IPOIB_MAX_PATH_REC_QUEUE) {
996 				push_pseudo_header(skb, neigh->daddr);
997 				__skb_queue_tail(&neigh->queue, skb);
998 			} else {
999 				ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
1000 					   skb_queue_len(&neigh->queue));
1001 				goto err_drop;
1002 			}
1003 		} else {
1004 			spin_unlock_irqrestore(&priv->lock, flags);
1005 			path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1006 						       IPOIB_QPN(daddr));
1007 			ipoib_neigh_put(neigh);
1008 			return NULL;
1009 		}
1010 	} else {
1011 		neigh->ah  = NULL;
1012 
1013 		if (!path->query && path_rec_start(dev, path))
1014 			goto err_path;
1015 		if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1016 			push_pseudo_header(skb, neigh->daddr);
1017 			__skb_queue_tail(&neigh->queue, skb);
1018 		} else {
1019 			goto err_drop;
1020 		}
1021 	}
1022 
1023 	spin_unlock_irqrestore(&priv->lock, flags);
1024 	ipoib_neigh_put(neigh);
1025 	return NULL;
1026 
1027 err_path:
1028 	ipoib_neigh_free(neigh);
1029 err_drop:
1030 	++dev->stats.tx_dropped;
1031 	dev_kfree_skb_any(skb);
1032 
1033 	spin_unlock_irqrestore(&priv->lock, flags);
1034 	ipoib_neigh_put(neigh);
1035 
1036 	return NULL;
1037 }
1038 
1039 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1040 			     struct ipoib_pseudo_header *phdr)
1041 {
1042 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1043 	struct rdma_netdev *rn = netdev_priv(dev);
1044 	struct ipoib_path *path;
1045 	unsigned long flags;
1046 
1047 	spin_lock_irqsave(&priv->lock, flags);
1048 
1049 	/* no broadcast means that all paths are (going to be) not valid */
1050 	if (!priv->broadcast)
1051 		goto drop_and_unlock;
1052 
1053 	path = __path_find(dev, phdr->hwaddr + 4);
1054 	if (!path || !path->ah || !path->ah->valid) {
1055 		if (!path) {
1056 			path = path_rec_create(dev, phdr->hwaddr + 4);
1057 			if (!path)
1058 				goto drop_and_unlock;
1059 			__path_add(dev, path);
1060 		} else {
1061 			/*
1062 			 * make sure there are no changes in the existing
1063 			 * path record
1064 			 */
1065 			init_path_rec(priv, path, phdr->hwaddr + 4);
1066 		}
1067 		if (!path->query && path_rec_start(dev, path)) {
1068 			goto drop_and_unlock;
1069 		}
1070 
1071 		if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1072 			push_pseudo_header(skb, phdr->hwaddr);
1073 			__skb_queue_tail(&path->queue, skb);
1074 			goto unlock;
1075 		} else {
1076 			goto drop_and_unlock;
1077 		}
1078 	}
1079 
1080 	spin_unlock_irqrestore(&priv->lock, flags);
1081 	ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1082 		  be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1083 	path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1084 				       IPOIB_QPN(phdr->hwaddr));
1085 	return;
1086 
1087 drop_and_unlock:
1088 	++dev->stats.tx_dropped;
1089 	dev_kfree_skb_any(skb);
1090 unlock:
1091 	spin_unlock_irqrestore(&priv->lock, flags);
1092 }
1093 
1094 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1095 {
1096 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1097 	struct rdma_netdev *rn = netdev_priv(dev);
1098 	struct ipoib_neigh *neigh;
1099 	struct ipoib_pseudo_header *phdr;
1100 	struct ipoib_header *header;
1101 	unsigned long flags;
1102 
1103 	phdr = (struct ipoib_pseudo_header *) skb->data;
1104 	skb_pull(skb, sizeof(*phdr));
1105 	header = (struct ipoib_header *) skb->data;
1106 
1107 	if (unlikely(phdr->hwaddr[4] == 0xff)) {
1108 		/* multicast, arrange "if" according to probability */
1109 		if ((header->proto != htons(ETH_P_IP)) &&
1110 		    (header->proto != htons(ETH_P_IPV6)) &&
1111 		    (header->proto != htons(ETH_P_ARP)) &&
1112 		    (header->proto != htons(ETH_P_RARP)) &&
1113 		    (header->proto != htons(ETH_P_TIPC))) {
1114 			/* ethertype not supported by IPoIB */
1115 			++dev->stats.tx_dropped;
1116 			dev_kfree_skb_any(skb);
1117 			return NETDEV_TX_OK;
1118 		}
1119 		/* Add in the P_Key for multicast*/
1120 		phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1121 		phdr->hwaddr[9] = priv->pkey & 0xff;
1122 
1123 		neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1124 		if (likely(neigh))
1125 			goto send_using_neigh;
1126 		ipoib_mcast_send(dev, phdr->hwaddr, skb);
1127 		return NETDEV_TX_OK;
1128 	}
1129 
1130 	/* unicast, arrange "switch" according to probability */
1131 	switch (header->proto) {
1132 	case htons(ETH_P_IP):
1133 	case htons(ETH_P_IPV6):
1134 	case htons(ETH_P_TIPC):
1135 		neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1136 		if (unlikely(!neigh)) {
1137 			neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1138 			if (likely(!neigh))
1139 				return NETDEV_TX_OK;
1140 		}
1141 		break;
1142 	case htons(ETH_P_ARP):
1143 	case htons(ETH_P_RARP):
1144 		/* for unicast ARP and RARP should always perform path find */
1145 		unicast_arp_send(skb, dev, phdr);
1146 		return NETDEV_TX_OK;
1147 	default:
1148 		/* ethertype not supported by IPoIB */
1149 		++dev->stats.tx_dropped;
1150 		dev_kfree_skb_any(skb);
1151 		return NETDEV_TX_OK;
1152 	}
1153 
1154 send_using_neigh:
1155 	/* note we now hold a ref to neigh */
1156 	if (ipoib_cm_get(neigh)) {
1157 		if (ipoib_cm_up(neigh)) {
1158 			ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1159 			goto unref;
1160 		}
1161 	} else if (neigh->ah && neigh->ah->valid) {
1162 		neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1163 						IPOIB_QPN(phdr->hwaddr));
1164 		goto unref;
1165 	} else if (neigh->ah) {
1166 		neigh_refresh_path(neigh, phdr->hwaddr, dev);
1167 	}
1168 
1169 	if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1170 		push_pseudo_header(skb, phdr->hwaddr);
1171 		spin_lock_irqsave(&priv->lock, flags);
1172 		__skb_queue_tail(&neigh->queue, skb);
1173 		spin_unlock_irqrestore(&priv->lock, flags);
1174 	} else {
1175 		++dev->stats.tx_dropped;
1176 		dev_kfree_skb_any(skb);
1177 	}
1178 
1179 unref:
1180 	ipoib_neigh_put(neigh);
1181 
1182 	return NETDEV_TX_OK;
1183 }
1184 
1185 static void ipoib_timeout(struct net_device *dev, unsigned int txqueue)
1186 {
1187 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1188 
1189 	ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1190 		   jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1191 	ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1192 		   netif_queue_stopped(dev),
1193 		   priv->tx_head, priv->tx_tail);
1194 	/* XXX reset QP, etc. */
1195 }
1196 
1197 static int ipoib_hard_header(struct sk_buff *skb,
1198 			     struct net_device *dev,
1199 			     unsigned short type,
1200 			     const void *daddr,
1201 			     const void *saddr,
1202 			     unsigned int len)
1203 {
1204 	struct ipoib_header *header;
1205 
1206 	header = skb_push(skb, sizeof(*header));
1207 
1208 	header->proto = htons(type);
1209 	header->reserved = 0;
1210 
1211 	/*
1212 	 * we don't rely on dst_entry structure,  always stuff the
1213 	 * destination address into skb hard header so we can figure out where
1214 	 * to send the packet later.
1215 	 */
1216 	push_pseudo_header(skb, daddr);
1217 
1218 	return IPOIB_HARD_LEN;
1219 }
1220 
1221 static void ipoib_set_mcast_list(struct net_device *dev)
1222 {
1223 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1224 
1225 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1226 		ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1227 		return;
1228 	}
1229 
1230 	queue_work(priv->wq, &priv->restart_task);
1231 }
1232 
1233 static int ipoib_get_iflink(const struct net_device *dev)
1234 {
1235 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1236 
1237 	/* parent interface */
1238 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1239 		return dev->ifindex;
1240 
1241 	/* child/vlan interface */
1242 	return priv->parent->ifindex;
1243 }
1244 
1245 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1246 {
1247 	/*
1248 	 * Use only the address parts that contributes to spreading
1249 	 * The subnet prefix is not used as one can not connect to
1250 	 * same remote port (GUID) using the same remote QPN via two
1251 	 * different subnets.
1252 	 */
1253 	 /* qpn octets[1:4) & port GUID octets[12:20) */
1254 	u32 *d32 = (u32 *) daddr;
1255 	u32 hv;
1256 
1257 	hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1258 	return hv & htbl->mask;
1259 }
1260 
1261 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1262 {
1263 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1264 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1265 	struct ipoib_neigh_hash *htbl;
1266 	struct ipoib_neigh *neigh = NULL;
1267 	u32 hash_val;
1268 
1269 	rcu_read_lock_bh();
1270 
1271 	htbl = rcu_dereference_bh(ntbl->htbl);
1272 
1273 	if (!htbl)
1274 		goto out_unlock;
1275 
1276 	hash_val = ipoib_addr_hash(htbl, daddr);
1277 	for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1278 	     neigh != NULL;
1279 	     neigh = rcu_dereference_bh(neigh->hnext)) {
1280 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1281 			/* found, take one ref on behalf of the caller */
1282 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1283 				/* deleted */
1284 				neigh = NULL;
1285 				goto out_unlock;
1286 			}
1287 
1288 			if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1289 				neigh->alive = jiffies;
1290 			goto out_unlock;
1291 		}
1292 	}
1293 
1294 out_unlock:
1295 	rcu_read_unlock_bh();
1296 	return neigh;
1297 }
1298 
1299 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1300 {
1301 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1302 	struct ipoib_neigh_hash *htbl;
1303 	unsigned long neigh_obsolete;
1304 	unsigned long dt;
1305 	unsigned long flags;
1306 	int i;
1307 	LIST_HEAD(remove_list);
1308 
1309 	spin_lock_irqsave(&priv->lock, flags);
1310 
1311 	htbl = rcu_dereference_protected(ntbl->htbl,
1312 					 lockdep_is_held(&priv->lock));
1313 
1314 	if (!htbl)
1315 		goto out_unlock;
1316 
1317 	/* neigh is obsolete if it was idle for two GC periods */
1318 	dt = 2 * arp_tbl.gc_interval;
1319 	neigh_obsolete = jiffies - dt;
1320 
1321 	for (i = 0; i < htbl->size; i++) {
1322 		struct ipoib_neigh *neigh;
1323 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1324 
1325 		while ((neigh = rcu_dereference_protected(*np,
1326 							  lockdep_is_held(&priv->lock))) != NULL) {
1327 			/* was the neigh idle for two GC periods */
1328 			if (time_after(neigh_obsolete, neigh->alive)) {
1329 
1330 				ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1331 
1332 				rcu_assign_pointer(*np,
1333 						   rcu_dereference_protected(neigh->hnext,
1334 									     lockdep_is_held(&priv->lock)));
1335 				/* remove from path/mc list */
1336 				list_del_init(&neigh->list);
1337 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1338 			} else {
1339 				np = &neigh->hnext;
1340 			}
1341 
1342 		}
1343 	}
1344 
1345 out_unlock:
1346 	spin_unlock_irqrestore(&priv->lock, flags);
1347 	ipoib_mcast_remove_list(&remove_list);
1348 }
1349 
1350 static void ipoib_reap_neigh(struct work_struct *work)
1351 {
1352 	struct ipoib_dev_priv *priv =
1353 		container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1354 
1355 	__ipoib_reap_neigh(priv);
1356 
1357 	queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1358 			   arp_tbl.gc_interval);
1359 }
1360 
1361 
1362 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1363 				      struct net_device *dev)
1364 {
1365 	struct ipoib_neigh *neigh;
1366 
1367 	neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1368 	if (!neigh)
1369 		return NULL;
1370 
1371 	neigh->dev = dev;
1372 	memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1373 	skb_queue_head_init(&neigh->queue);
1374 	INIT_LIST_HEAD(&neigh->list);
1375 	ipoib_cm_set(neigh, NULL);
1376 	/* one ref on behalf of the caller */
1377 	atomic_set(&neigh->refcnt, 1);
1378 
1379 	return neigh;
1380 }
1381 
1382 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1383 				      struct net_device *dev)
1384 {
1385 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1386 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1387 	struct ipoib_neigh_hash *htbl;
1388 	struct ipoib_neigh *neigh;
1389 	u32 hash_val;
1390 
1391 	htbl = rcu_dereference_protected(ntbl->htbl,
1392 					 lockdep_is_held(&priv->lock));
1393 	if (!htbl) {
1394 		neigh = NULL;
1395 		goto out_unlock;
1396 	}
1397 
1398 	/* need to add a new neigh, but maybe some other thread succeeded?
1399 	 * recalc hash, maybe hash resize took place so we do a search
1400 	 */
1401 	hash_val = ipoib_addr_hash(htbl, daddr);
1402 	for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1403 					       lockdep_is_held(&priv->lock));
1404 	     neigh != NULL;
1405 	     neigh = rcu_dereference_protected(neigh->hnext,
1406 					       lockdep_is_held(&priv->lock))) {
1407 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1408 			/* found, take one ref on behalf of the caller */
1409 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1410 				/* deleted */
1411 				neigh = NULL;
1412 				break;
1413 			}
1414 			neigh->alive = jiffies;
1415 			goto out_unlock;
1416 		}
1417 	}
1418 
1419 	neigh = ipoib_neigh_ctor(daddr, dev);
1420 	if (!neigh)
1421 		goto out_unlock;
1422 
1423 	/* one ref on behalf of the hash table */
1424 	atomic_inc(&neigh->refcnt);
1425 	neigh->alive = jiffies;
1426 	/* put in hash */
1427 	rcu_assign_pointer(neigh->hnext,
1428 			   rcu_dereference_protected(htbl->buckets[hash_val],
1429 						     lockdep_is_held(&priv->lock)));
1430 	rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1431 	atomic_inc(&ntbl->entries);
1432 
1433 out_unlock:
1434 
1435 	return neigh;
1436 }
1437 
1438 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1439 {
1440 	/* neigh reference count was dropprd to zero */
1441 	struct net_device *dev = neigh->dev;
1442 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1443 	struct sk_buff *skb;
1444 	if (neigh->ah)
1445 		ipoib_put_ah(neigh->ah);
1446 	while ((skb = __skb_dequeue(&neigh->queue))) {
1447 		++dev->stats.tx_dropped;
1448 		dev_kfree_skb_any(skb);
1449 	}
1450 	if (ipoib_cm_get(neigh))
1451 		ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1452 	ipoib_dbg(ipoib_priv(dev),
1453 		  "neigh free for %06x %pI6\n",
1454 		  IPOIB_QPN(neigh->daddr),
1455 		  neigh->daddr + 4);
1456 	kfree(neigh);
1457 	if (atomic_dec_and_test(&priv->ntbl.entries)) {
1458 		if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1459 			complete(&priv->ntbl.flushed);
1460 	}
1461 }
1462 
1463 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1464 {
1465 	/* Called as a result of removal from hash table */
1466 	struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1467 	/* note TX context may hold another ref */
1468 	ipoib_neigh_put(neigh);
1469 }
1470 
1471 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1472 {
1473 	struct net_device *dev = neigh->dev;
1474 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1475 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1476 	struct ipoib_neigh_hash *htbl;
1477 	struct ipoib_neigh __rcu **np;
1478 	struct ipoib_neigh *n;
1479 	u32 hash_val;
1480 
1481 	htbl = rcu_dereference_protected(ntbl->htbl,
1482 					lockdep_is_held(&priv->lock));
1483 	if (!htbl)
1484 		return;
1485 
1486 	hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1487 	np = &htbl->buckets[hash_val];
1488 	for (n = rcu_dereference_protected(*np,
1489 					    lockdep_is_held(&priv->lock));
1490 	     n != NULL;
1491 	     n = rcu_dereference_protected(*np,
1492 					lockdep_is_held(&priv->lock))) {
1493 		if (n == neigh) {
1494 			/* found */
1495 			rcu_assign_pointer(*np,
1496 					   rcu_dereference_protected(neigh->hnext,
1497 								     lockdep_is_held(&priv->lock)));
1498 			/* remove from parent list */
1499 			list_del_init(&neigh->list);
1500 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1501 			return;
1502 		} else {
1503 			np = &n->hnext;
1504 		}
1505 	}
1506 }
1507 
1508 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1509 {
1510 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1511 	struct ipoib_neigh_hash *htbl;
1512 	struct ipoib_neigh __rcu **buckets;
1513 	u32 size;
1514 
1515 	clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1516 	ntbl->htbl = NULL;
1517 	htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1518 	if (!htbl)
1519 		return -ENOMEM;
1520 	size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1521 	buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1522 	if (!buckets) {
1523 		kfree(htbl);
1524 		return -ENOMEM;
1525 	}
1526 	htbl->size = size;
1527 	htbl->mask = (size - 1);
1528 	htbl->buckets = buckets;
1529 	RCU_INIT_POINTER(ntbl->htbl, htbl);
1530 	htbl->ntbl = ntbl;
1531 	atomic_set(&ntbl->entries, 0);
1532 
1533 	/* start garbage collection */
1534 	queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1535 			   arp_tbl.gc_interval);
1536 
1537 	return 0;
1538 }
1539 
1540 static void neigh_hash_free_rcu(struct rcu_head *head)
1541 {
1542 	struct ipoib_neigh_hash *htbl = container_of(head,
1543 						    struct ipoib_neigh_hash,
1544 						    rcu);
1545 	struct ipoib_neigh __rcu **buckets = htbl->buckets;
1546 	struct ipoib_neigh_table *ntbl = htbl->ntbl;
1547 
1548 	kvfree(buckets);
1549 	kfree(htbl);
1550 	complete(&ntbl->deleted);
1551 }
1552 
1553 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1554 {
1555 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1556 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1557 	struct ipoib_neigh_hash *htbl;
1558 	unsigned long flags;
1559 	int i;
1560 
1561 	/* remove all neigh connected to a given path or mcast */
1562 	spin_lock_irqsave(&priv->lock, flags);
1563 
1564 	htbl = rcu_dereference_protected(ntbl->htbl,
1565 					 lockdep_is_held(&priv->lock));
1566 
1567 	if (!htbl)
1568 		goto out_unlock;
1569 
1570 	for (i = 0; i < htbl->size; i++) {
1571 		struct ipoib_neigh *neigh;
1572 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1573 
1574 		while ((neigh = rcu_dereference_protected(*np,
1575 							  lockdep_is_held(&priv->lock))) != NULL) {
1576 			/* delete neighs belong to this parent */
1577 			if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1578 				rcu_assign_pointer(*np,
1579 						   rcu_dereference_protected(neigh->hnext,
1580 									     lockdep_is_held(&priv->lock)));
1581 				/* remove from parent list */
1582 				list_del_init(&neigh->list);
1583 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1584 			} else {
1585 				np = &neigh->hnext;
1586 			}
1587 
1588 		}
1589 	}
1590 out_unlock:
1591 	spin_unlock_irqrestore(&priv->lock, flags);
1592 }
1593 
1594 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1595 {
1596 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1597 	struct ipoib_neigh_hash *htbl;
1598 	unsigned long flags;
1599 	int i, wait_flushed = 0;
1600 
1601 	init_completion(&priv->ntbl.flushed);
1602 	set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1603 
1604 	spin_lock_irqsave(&priv->lock, flags);
1605 
1606 	htbl = rcu_dereference_protected(ntbl->htbl,
1607 					lockdep_is_held(&priv->lock));
1608 	if (!htbl)
1609 		goto out_unlock;
1610 
1611 	wait_flushed = atomic_read(&priv->ntbl.entries);
1612 	if (!wait_flushed)
1613 		goto free_htbl;
1614 
1615 	for (i = 0; i < htbl->size; i++) {
1616 		struct ipoib_neigh *neigh;
1617 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1618 
1619 		while ((neigh = rcu_dereference_protected(*np,
1620 				       lockdep_is_held(&priv->lock))) != NULL) {
1621 			rcu_assign_pointer(*np,
1622 					   rcu_dereference_protected(neigh->hnext,
1623 								     lockdep_is_held(&priv->lock)));
1624 			/* remove from path/mc list */
1625 			list_del_init(&neigh->list);
1626 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1627 		}
1628 	}
1629 
1630 free_htbl:
1631 	rcu_assign_pointer(ntbl->htbl, NULL);
1632 	call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1633 
1634 out_unlock:
1635 	spin_unlock_irqrestore(&priv->lock, flags);
1636 	if (wait_flushed)
1637 		wait_for_completion(&priv->ntbl.flushed);
1638 }
1639 
1640 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1641 {
1642 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1643 
1644 	ipoib_dbg(priv, "%s\n", __func__);
1645 	init_completion(&priv->ntbl.deleted);
1646 
1647 	cancel_delayed_work_sync(&priv->neigh_reap_task);
1648 
1649 	ipoib_flush_neighs(priv);
1650 
1651 	wait_for_completion(&priv->ntbl.deleted);
1652 }
1653 
1654 static void ipoib_napi_add(struct net_device *dev)
1655 {
1656 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1657 
1658 	netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1659 	netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1660 }
1661 
1662 static void ipoib_napi_del(struct net_device *dev)
1663 {
1664 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1665 
1666 	netif_napi_del(&priv->recv_napi);
1667 	netif_napi_del(&priv->send_napi);
1668 }
1669 
1670 static void ipoib_dev_uninit_default(struct net_device *dev)
1671 {
1672 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1673 
1674 	ipoib_transport_dev_cleanup(dev);
1675 
1676 	ipoib_napi_del(dev);
1677 
1678 	ipoib_cm_dev_cleanup(dev);
1679 
1680 	kfree(priv->rx_ring);
1681 	vfree(priv->tx_ring);
1682 
1683 	priv->rx_ring = NULL;
1684 	priv->tx_ring = NULL;
1685 }
1686 
1687 static int ipoib_dev_init_default(struct net_device *dev)
1688 {
1689 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1690 
1691 	ipoib_napi_add(dev);
1692 
1693 	/* Allocate RX/TX "rings" to hold queued skbs */
1694 	priv->rx_ring =	kcalloc(ipoib_recvq_size,
1695 				       sizeof(*priv->rx_ring),
1696 				       GFP_KERNEL);
1697 	if (!priv->rx_ring)
1698 		goto out;
1699 
1700 	priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1701 					   sizeof(*priv->tx_ring)));
1702 	if (!priv->tx_ring) {
1703 		pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1704 			priv->ca->name, ipoib_sendq_size);
1705 		goto out_rx_ring_cleanup;
1706 	}
1707 
1708 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1709 
1710 	if (ipoib_transport_dev_init(dev, priv->ca)) {
1711 		pr_warn("%s: ipoib_transport_dev_init failed\n",
1712 			priv->ca->name);
1713 		goto out_tx_ring_cleanup;
1714 	}
1715 
1716 	/* after qp created set dev address */
1717 	priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1718 	priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
1719 	priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1720 
1721 	return 0;
1722 
1723 out_tx_ring_cleanup:
1724 	vfree(priv->tx_ring);
1725 
1726 out_rx_ring_cleanup:
1727 	kfree(priv->rx_ring);
1728 
1729 out:
1730 	ipoib_napi_del(dev);
1731 	return -ENOMEM;
1732 }
1733 
1734 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1735 		       int cmd)
1736 {
1737 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1738 
1739 	if (!priv->rn_ops->ndo_do_ioctl)
1740 		return -EOPNOTSUPP;
1741 
1742 	return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1743 }
1744 
1745 static int ipoib_dev_init(struct net_device *dev)
1746 {
1747 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1748 	int ret = -ENOMEM;
1749 
1750 	priv->qp = NULL;
1751 
1752 	/*
1753 	 * the various IPoIB tasks assume they will never race against
1754 	 * themselves, so always use a single thread workqueue
1755 	 */
1756 	priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1757 	if (!priv->wq) {
1758 		pr_warn("%s: failed to allocate device WQ\n", dev->name);
1759 		goto out;
1760 	}
1761 
1762 	/* create pd, which used both for control and datapath*/
1763 	priv->pd = ib_alloc_pd(priv->ca, 0);
1764 	if (IS_ERR(priv->pd)) {
1765 		pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1766 		goto clean_wq;
1767 	}
1768 
1769 	ret = priv->rn_ops->ndo_init(dev);
1770 	if (ret) {
1771 		pr_warn("%s failed to init HW resource\n", dev->name);
1772 		goto out_free_pd;
1773 	}
1774 
1775 	ret = ipoib_neigh_hash_init(priv);
1776 	if (ret) {
1777 		pr_warn("%s failed to init neigh hash\n", dev->name);
1778 		goto out_dev_uninit;
1779 	}
1780 
1781 	if (dev->flags & IFF_UP) {
1782 		if (ipoib_ib_dev_open(dev)) {
1783 			pr_warn("%s failed to open device\n", dev->name);
1784 			ret = -ENODEV;
1785 			goto out_hash_uninit;
1786 		}
1787 	}
1788 
1789 	return 0;
1790 
1791 out_hash_uninit:
1792 	ipoib_neigh_hash_uninit(dev);
1793 
1794 out_dev_uninit:
1795 	ipoib_ib_dev_cleanup(dev);
1796 
1797 out_free_pd:
1798 	if (priv->pd) {
1799 		ib_dealloc_pd(priv->pd);
1800 		priv->pd = NULL;
1801 	}
1802 
1803 clean_wq:
1804 	if (priv->wq) {
1805 		destroy_workqueue(priv->wq);
1806 		priv->wq = NULL;
1807 	}
1808 
1809 out:
1810 	return ret;
1811 }
1812 
1813 /*
1814  * This must be called before doing an unregister_netdev on a parent device to
1815  * shutdown the IB event handler.
1816  */
1817 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1818 {
1819 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1820 
1821 	/*
1822 	 * ipoib_set_mac checks netif_running before pushing work, clearing
1823 	 * running ensures the it will not add more work.
1824 	 */
1825 	rtnl_lock();
1826 	dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP, NULL);
1827 	rtnl_unlock();
1828 
1829 	/* ipoib_event() cannot be running once this returns */
1830 	ib_unregister_event_handler(&priv->event_handler);
1831 
1832 	/*
1833 	 * Work on the queue grabs the rtnl lock, so this cannot be done while
1834 	 * also holding it.
1835 	 */
1836 	flush_workqueue(ipoib_workqueue);
1837 }
1838 
1839 static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1840 {
1841 	priv->hca_caps = priv->ca->attrs.device_cap_flags;
1842 
1843 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1844 		priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1845 
1846 		if (priv->hca_caps & IB_DEVICE_UD_TSO)
1847 			priv->dev->hw_features |= NETIF_F_TSO;
1848 
1849 		priv->dev->features |= priv->dev->hw_features;
1850 	}
1851 }
1852 
1853 static int ipoib_parent_init(struct net_device *ndev)
1854 {
1855 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1856 	struct ib_port_attr attr;
1857 	int result;
1858 
1859 	result = ib_query_port(priv->ca, priv->port, &attr);
1860 	if (result) {
1861 		pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1862 			priv->port);
1863 		return result;
1864 	}
1865 	priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1866 
1867 	result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1868 	if (result) {
1869 		pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1870 			priv->ca->name, priv->port, result);
1871 		return result;
1872 	}
1873 
1874 	result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1875 	if (result) {
1876 		pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1877 			priv->ca->name, priv->port, result);
1878 		return result;
1879 	}
1880 	memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
1881 	       sizeof(union ib_gid));
1882 
1883 	SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1884 	priv->dev->dev_port = priv->port - 1;
1885 	/* Let's set this one too for backwards compatibility. */
1886 	priv->dev->dev_id = priv->port - 1;
1887 
1888 	return 0;
1889 }
1890 
1891 static void ipoib_child_init(struct net_device *ndev)
1892 {
1893 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1894 	struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1895 
1896 	priv->max_ib_mtu = ppriv->max_ib_mtu;
1897 	set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1898 	memcpy(priv->dev->dev_addr, ppriv->dev->dev_addr, INFINIBAND_ALEN);
1899 	memcpy(&priv->local_gid, &ppriv->local_gid, sizeof(priv->local_gid));
1900 }
1901 
1902 static int ipoib_ndo_init(struct net_device *ndev)
1903 {
1904 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1905 	int rc;
1906 
1907 	if (priv->parent) {
1908 		ipoib_child_init(ndev);
1909 	} else {
1910 		rc = ipoib_parent_init(ndev);
1911 		if (rc)
1912 			return rc;
1913 	}
1914 
1915 	/* MTU will be reset when mcast join happens */
1916 	ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1917 	priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
1918 	ndev->max_mtu = IPOIB_CM_MTU;
1919 
1920 	ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
1921 
1922 	/*
1923 	 * Set the full membership bit, so that we join the right
1924 	 * broadcast group, etc.
1925 	 */
1926 	priv->pkey |= 0x8000;
1927 
1928 	ndev->broadcast[8] = priv->pkey >> 8;
1929 	ndev->broadcast[9] = priv->pkey & 0xff;
1930 	set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1931 
1932 	ipoib_set_dev_features(priv);
1933 
1934 	rc = ipoib_dev_init(ndev);
1935 	if (rc) {
1936 		pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1937 			priv->ca->name, priv->dev->name, priv->port, rc);
1938 		return rc;
1939 	}
1940 
1941 	if (priv->parent) {
1942 		struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1943 
1944 		dev_hold(priv->parent);
1945 
1946 		down_write(&ppriv->vlan_rwsem);
1947 		list_add_tail(&priv->list, &ppriv->child_intfs);
1948 		up_write(&ppriv->vlan_rwsem);
1949 	}
1950 
1951 	return 0;
1952 }
1953 
1954 static void ipoib_ndo_uninit(struct net_device *dev)
1955 {
1956 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1957 
1958 	ASSERT_RTNL();
1959 
1960 	/*
1961 	 * ipoib_remove_one guarantees the children are removed before the
1962 	 * parent, and that is the only place where a parent can be removed.
1963 	 */
1964 	WARN_ON(!list_empty(&priv->child_intfs));
1965 
1966 	if (priv->parent) {
1967 		struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1968 
1969 		down_write(&ppriv->vlan_rwsem);
1970 		list_del(&priv->list);
1971 		up_write(&ppriv->vlan_rwsem);
1972 	}
1973 
1974 	ipoib_neigh_hash_uninit(dev);
1975 
1976 	ipoib_ib_dev_cleanup(dev);
1977 
1978 	/* no more works over the priv->wq */
1979 	if (priv->wq) {
1980 		flush_workqueue(priv->wq);
1981 		destroy_workqueue(priv->wq);
1982 		priv->wq = NULL;
1983 	}
1984 
1985 	if (priv->parent)
1986 		dev_put(priv->parent);
1987 }
1988 
1989 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1990 {
1991 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1992 
1993 	return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1994 }
1995 
1996 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1997 			       struct ifla_vf_info *ivf)
1998 {
1999 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2000 	int err;
2001 
2002 	err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2003 	if (err)
2004 		return err;
2005 
2006 	ivf->vf = vf;
2007 	memcpy(ivf->mac, dev->dev_addr, dev->addr_len);
2008 
2009 	return 0;
2010 }
2011 
2012 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2013 {
2014 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2015 
2016 	if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2017 		return -EINVAL;
2018 
2019 	return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2020 }
2021 
2022 static int ipoib_get_vf_guid(struct net_device *dev, int vf,
2023 			     struct ifla_vf_guid *node_guid,
2024 			     struct ifla_vf_guid *port_guid)
2025 {
2026 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2027 
2028 	return ib_get_vf_guid(priv->ca, vf, priv->port, node_guid, port_guid);
2029 }
2030 
2031 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2032 			      struct ifla_vf_stats *vf_stats)
2033 {
2034 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2035 
2036 	return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2037 }
2038 
2039 static const struct header_ops ipoib_header_ops = {
2040 	.create	= ipoib_hard_header,
2041 };
2042 
2043 static const struct net_device_ops ipoib_netdev_ops_pf = {
2044 	.ndo_init		 = ipoib_ndo_init,
2045 	.ndo_uninit		 = ipoib_ndo_uninit,
2046 	.ndo_open		 = ipoib_open,
2047 	.ndo_stop		 = ipoib_stop,
2048 	.ndo_change_mtu		 = ipoib_change_mtu,
2049 	.ndo_fix_features	 = ipoib_fix_features,
2050 	.ndo_start_xmit		 = ipoib_start_xmit,
2051 	.ndo_tx_timeout		 = ipoib_timeout,
2052 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
2053 	.ndo_get_iflink		 = ipoib_get_iflink,
2054 	.ndo_set_vf_link_state	 = ipoib_set_vf_link_state,
2055 	.ndo_get_vf_config	 = ipoib_get_vf_config,
2056 	.ndo_get_vf_stats	 = ipoib_get_vf_stats,
2057 	.ndo_get_vf_guid	 = ipoib_get_vf_guid,
2058 	.ndo_set_vf_guid	 = ipoib_set_vf_guid,
2059 	.ndo_set_mac_address	 = ipoib_set_mac,
2060 	.ndo_get_stats64	 = ipoib_get_stats,
2061 	.ndo_do_ioctl		 = ipoib_ioctl,
2062 };
2063 
2064 static const struct net_device_ops ipoib_netdev_ops_vf = {
2065 	.ndo_init		 = ipoib_ndo_init,
2066 	.ndo_uninit		 = ipoib_ndo_uninit,
2067 	.ndo_open		 = ipoib_open,
2068 	.ndo_stop		 = ipoib_stop,
2069 	.ndo_change_mtu		 = ipoib_change_mtu,
2070 	.ndo_fix_features	 = ipoib_fix_features,
2071 	.ndo_start_xmit	 	 = ipoib_start_xmit,
2072 	.ndo_tx_timeout		 = ipoib_timeout,
2073 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
2074 	.ndo_get_iflink		 = ipoib_get_iflink,
2075 	.ndo_get_stats64	 = ipoib_get_stats,
2076 	.ndo_do_ioctl		 = ipoib_ioctl,
2077 };
2078 
2079 void ipoib_setup_common(struct net_device *dev)
2080 {
2081 	dev->header_ops		 = &ipoib_header_ops;
2082 
2083 	ipoib_set_ethtool_ops(dev);
2084 
2085 	dev->watchdog_timeo	 = HZ;
2086 
2087 	dev->flags		|= IFF_BROADCAST | IFF_MULTICAST;
2088 
2089 	dev->hard_header_len	 = IPOIB_HARD_LEN;
2090 	dev->addr_len		 = INFINIBAND_ALEN;
2091 	dev->type		 = ARPHRD_INFINIBAND;
2092 	dev->tx_queue_len	 = ipoib_sendq_size * 2;
2093 	dev->features		 = (NETIF_F_VLAN_CHALLENGED	|
2094 				    NETIF_F_HIGHDMA);
2095 	netif_keep_dst(dev);
2096 
2097 	memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2098 
2099 	/*
2100 	 * unregister_netdev always frees the netdev, we use this mode
2101 	 * consistently to unify all the various unregister paths, including
2102 	 * those connected to rtnl_link_ops which require it.
2103 	 */
2104 	dev->needs_free_netdev = true;
2105 }
2106 
2107 static void ipoib_build_priv(struct net_device *dev)
2108 {
2109 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2110 
2111 	priv->dev = dev;
2112 	spin_lock_init(&priv->lock);
2113 	init_rwsem(&priv->vlan_rwsem);
2114 	mutex_init(&priv->mcast_mutex);
2115 
2116 	INIT_LIST_HEAD(&priv->path_list);
2117 	INIT_LIST_HEAD(&priv->child_intfs);
2118 	INIT_LIST_HEAD(&priv->dead_ahs);
2119 	INIT_LIST_HEAD(&priv->multicast_list);
2120 
2121 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
2122 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2123 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
2124 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
2125 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
2126 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2127 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2128 	INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2129 }
2130 
2131 static const struct net_device_ops ipoib_netdev_default_pf = {
2132 	.ndo_init		 = ipoib_dev_init_default,
2133 	.ndo_uninit		 = ipoib_dev_uninit_default,
2134 	.ndo_open		 = ipoib_ib_dev_open_default,
2135 	.ndo_stop		 = ipoib_ib_dev_stop_default,
2136 };
2137 
2138 static struct net_device *ipoib_alloc_netdev(struct ib_device *hca, u8 port,
2139 					     const char *name)
2140 {
2141 	struct net_device *dev;
2142 
2143 	dev = rdma_alloc_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2144 				NET_NAME_UNKNOWN, ipoib_setup_common);
2145 	if (!IS_ERR(dev) || PTR_ERR(dev) != -EOPNOTSUPP)
2146 		return dev;
2147 
2148 	dev = alloc_netdev(sizeof(struct rdma_netdev), name, NET_NAME_UNKNOWN,
2149 			   ipoib_setup_common);
2150 	if (!dev)
2151 		return ERR_PTR(-ENOMEM);
2152 	return dev;
2153 }
2154 
2155 int ipoib_intf_init(struct ib_device *hca, u8 port, const char *name,
2156 		    struct net_device *dev)
2157 {
2158 	struct rdma_netdev *rn = netdev_priv(dev);
2159 	struct ipoib_dev_priv *priv;
2160 	int rc;
2161 
2162 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2163 	if (!priv)
2164 		return -ENOMEM;
2165 
2166 	priv->ca = hca;
2167 	priv->port = port;
2168 
2169 	rc = rdma_init_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2170 			      NET_NAME_UNKNOWN, ipoib_setup_common, dev);
2171 	if (rc) {
2172 		if (rc != -EOPNOTSUPP)
2173 			goto out;
2174 
2175 		dev->netdev_ops = &ipoib_netdev_default_pf;
2176 		rn->send = ipoib_send;
2177 		rn->attach_mcast = ipoib_mcast_attach;
2178 		rn->detach_mcast = ipoib_mcast_detach;
2179 		rn->hca = hca;
2180 	}
2181 
2182 	priv->rn_ops = dev->netdev_ops;
2183 
2184 	if (hca->attrs.device_cap_flags & IB_DEVICE_VIRTUAL_FUNCTION)
2185 		dev->netdev_ops	= &ipoib_netdev_ops_vf;
2186 	else
2187 		dev->netdev_ops	= &ipoib_netdev_ops_pf;
2188 
2189 	rn->clnt_priv = priv;
2190 	/*
2191 	 * Only the child register_netdev flows can handle priv_destructor
2192 	 * being set, so we force it to NULL here and handle manually until it
2193 	 * is safe to turn on.
2194 	 */
2195 	priv->next_priv_destructor = dev->priv_destructor;
2196 	dev->priv_destructor = NULL;
2197 
2198 	ipoib_build_priv(dev);
2199 
2200 	return 0;
2201 
2202 out:
2203 	kfree(priv);
2204 	return rc;
2205 }
2206 
2207 struct net_device *ipoib_intf_alloc(struct ib_device *hca, u8 port,
2208 				    const char *name)
2209 {
2210 	struct net_device *dev;
2211 	int rc;
2212 
2213 	dev = ipoib_alloc_netdev(hca, port, name);
2214 	if (IS_ERR(dev))
2215 		return dev;
2216 
2217 	rc = ipoib_intf_init(hca, port, name, dev);
2218 	if (rc) {
2219 		free_netdev(dev);
2220 		return ERR_PTR(rc);
2221 	}
2222 
2223 	/*
2224 	 * Upon success the caller must ensure ipoib_intf_free is called or
2225 	 * register_netdevice succeed'd and priv_destructor is set to
2226 	 * ipoib_intf_free.
2227 	 */
2228 	return dev;
2229 }
2230 
2231 void ipoib_intf_free(struct net_device *dev)
2232 {
2233 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2234 	struct rdma_netdev *rn = netdev_priv(dev);
2235 
2236 	dev->priv_destructor = priv->next_priv_destructor;
2237 	if (dev->priv_destructor)
2238 		dev->priv_destructor(dev);
2239 
2240 	/*
2241 	 * There are some error flows around register_netdev failing that may
2242 	 * attempt to call priv_destructor twice, prevent that from happening.
2243 	 */
2244 	dev->priv_destructor = NULL;
2245 
2246 	/* unregister/destroy is very complicated. Make bugs more obvious. */
2247 	rn->clnt_priv = NULL;
2248 
2249 	kfree(priv);
2250 }
2251 
2252 static ssize_t show_pkey(struct device *dev,
2253 			 struct device_attribute *attr, char *buf)
2254 {
2255 	struct net_device *ndev = to_net_dev(dev);
2256 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2257 
2258 	return sprintf(buf, "0x%04x\n", priv->pkey);
2259 }
2260 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2261 
2262 static ssize_t show_umcast(struct device *dev,
2263 			   struct device_attribute *attr, char *buf)
2264 {
2265 	struct net_device *ndev = to_net_dev(dev);
2266 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2267 
2268 	return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2269 }
2270 
2271 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2272 {
2273 	struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2274 
2275 	if (umcast_val > 0) {
2276 		set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2277 		ipoib_warn(priv, "ignoring multicast groups joined directly "
2278 				"by userspace\n");
2279 	} else
2280 		clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2281 }
2282 
2283 static ssize_t set_umcast(struct device *dev,
2284 			  struct device_attribute *attr,
2285 			  const char *buf, size_t count)
2286 {
2287 	unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2288 
2289 	ipoib_set_umcast(to_net_dev(dev), umcast_val);
2290 
2291 	return count;
2292 }
2293 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2294 
2295 int ipoib_add_umcast_attr(struct net_device *dev)
2296 {
2297 	return device_create_file(&dev->dev, &dev_attr_umcast);
2298 }
2299 
2300 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2301 {
2302 	struct ipoib_dev_priv *child_priv;
2303 	struct net_device *netdev = priv->dev;
2304 
2305 	netif_addr_lock_bh(netdev);
2306 
2307 	memcpy(&priv->local_gid.global.interface_id,
2308 	       &gid->global.interface_id,
2309 	       sizeof(gid->global.interface_id));
2310 	memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2311 	clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2312 
2313 	netif_addr_unlock_bh(netdev);
2314 
2315 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2316 		down_read(&priv->vlan_rwsem);
2317 		list_for_each_entry(child_priv, &priv->child_intfs, list)
2318 			set_base_guid(child_priv, gid);
2319 		up_read(&priv->vlan_rwsem);
2320 	}
2321 }
2322 
2323 static int ipoib_check_lladdr(struct net_device *dev,
2324 			      struct sockaddr_storage *ss)
2325 {
2326 	union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2327 	int ret = 0;
2328 
2329 	netif_addr_lock_bh(dev);
2330 
2331 	/* Make sure the QPN, reserved and subnet prefix match the current
2332 	 * lladdr, it also makes sure the lladdr is unicast.
2333 	 */
2334 	if (memcmp(dev->dev_addr, ss->__data,
2335 		   4 + sizeof(gid->global.subnet_prefix)) ||
2336 	    gid->global.interface_id == 0)
2337 		ret = -EINVAL;
2338 
2339 	netif_addr_unlock_bh(dev);
2340 
2341 	return ret;
2342 }
2343 
2344 static int ipoib_set_mac(struct net_device *dev, void *addr)
2345 {
2346 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
2347 	struct sockaddr_storage *ss = addr;
2348 	int ret;
2349 
2350 	if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2351 		return -EBUSY;
2352 
2353 	ret = ipoib_check_lladdr(dev, ss);
2354 	if (ret)
2355 		return ret;
2356 
2357 	set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2358 
2359 	queue_work(ipoib_workqueue, &priv->flush_light);
2360 
2361 	return 0;
2362 }
2363 
2364 static ssize_t create_child(struct device *dev,
2365 			    struct device_attribute *attr,
2366 			    const char *buf, size_t count)
2367 {
2368 	int pkey;
2369 	int ret;
2370 
2371 	if (sscanf(buf, "%i", &pkey) != 1)
2372 		return -EINVAL;
2373 
2374 	if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2375 		return -EINVAL;
2376 
2377 	ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2378 
2379 	return ret ? ret : count;
2380 }
2381 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2382 
2383 static ssize_t delete_child(struct device *dev,
2384 			    struct device_attribute *attr,
2385 			    const char *buf, size_t count)
2386 {
2387 	int pkey;
2388 	int ret;
2389 
2390 	if (sscanf(buf, "%i", &pkey) != 1)
2391 		return -EINVAL;
2392 
2393 	if (pkey < 0 || pkey > 0xffff)
2394 		return -EINVAL;
2395 
2396 	ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2397 
2398 	return ret ? ret : count;
2399 
2400 }
2401 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2402 
2403 int ipoib_add_pkey_attr(struct net_device *dev)
2404 {
2405 	return device_create_file(&dev->dev, &dev_attr_pkey);
2406 }
2407 
2408 /*
2409  * We erroneously exposed the iface's port number in the dev_id
2410  * sysfs field long after dev_port was introduced for that purpose[1],
2411  * and we need to stop everyone from relying on that.
2412  * Let's overload the shower routine for the dev_id file here
2413  * to gently bring the issue up.
2414  *
2415  * [1] https://www.spinics.net/lists/netdev/msg272123.html
2416  */
2417 static ssize_t dev_id_show(struct device *dev,
2418 			   struct device_attribute *attr, char *buf)
2419 {
2420 	struct net_device *ndev = to_net_dev(dev);
2421 
2422 	/*
2423 	 * ndev->dev_port will be equal to 0 in old kernel prior to commit
2424 	 * 9b8b2a323008 ("IB/ipoib: Use dev_port to expose network interface
2425 	 * port numbers") Zero was chosen as special case for user space
2426 	 * applications to fallback and query dev_id to check if it has
2427 	 * different value or not.
2428 	 *
2429 	 * Don't print warning in such scenario.
2430 	 *
2431 	 * https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c#L358
2432 	 */
2433 	if (ndev->dev_port && ndev->dev_id == ndev->dev_port)
2434 		netdev_info_once(ndev,
2435 			"\"%s\" wants to know my dev_id. Should it look at dev_port instead? See Documentation/ABI/testing/sysfs-class-net for more info.\n",
2436 			current->comm);
2437 
2438 	return sprintf(buf, "%#x\n", ndev->dev_id);
2439 }
2440 static DEVICE_ATTR_RO(dev_id);
2441 
2442 static int ipoib_intercept_dev_id_attr(struct net_device *dev)
2443 {
2444 	device_remove_file(&dev->dev, &dev_attr_dev_id);
2445 	return device_create_file(&dev->dev, &dev_attr_dev_id);
2446 }
2447 
2448 static struct net_device *ipoib_add_port(const char *format,
2449 					 struct ib_device *hca, u8 port)
2450 {
2451 	struct rtnl_link_ops *ops = ipoib_get_link_ops();
2452 	struct rdma_netdev_alloc_params params;
2453 	struct ipoib_dev_priv *priv;
2454 	struct net_device *ndev;
2455 	int result;
2456 
2457 	ndev = ipoib_intf_alloc(hca, port, format);
2458 	if (IS_ERR(ndev)) {
2459 		pr_warn("%s, %d: ipoib_intf_alloc failed %ld\n", hca->name, port,
2460 			PTR_ERR(ndev));
2461 		return ndev;
2462 	}
2463 	priv = ipoib_priv(ndev);
2464 
2465 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
2466 			      priv->ca, ipoib_event);
2467 	ib_register_event_handler(&priv->event_handler);
2468 
2469 	/* call event handler to ensure pkey in sync */
2470 	queue_work(ipoib_workqueue, &priv->flush_heavy);
2471 
2472 	result = register_netdev(ndev);
2473 	if (result) {
2474 		pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2475 			hca->name, port, result);
2476 
2477 		ipoib_parent_unregister_pre(ndev);
2478 		ipoib_intf_free(ndev);
2479 		free_netdev(ndev);
2480 
2481 		return ERR_PTR(result);
2482 	}
2483 
2484 	if (hca->ops.rdma_netdev_get_params) {
2485 		int rc = hca->ops.rdma_netdev_get_params(hca, port,
2486 						     RDMA_NETDEV_IPOIB,
2487 						     &params);
2488 
2489 		if (!rc && ops->priv_size < params.sizeof_priv)
2490 			ops->priv_size = params.sizeof_priv;
2491 	}
2492 	/*
2493 	 * We cannot set priv_destructor before register_netdev because we
2494 	 * need priv to be always valid during the error flow to execute
2495 	 * ipoib_parent_unregister_pre(). Instead handle it manually and only
2496 	 * enter priv_destructor mode once we are completely registered.
2497 	 */
2498 	ndev->priv_destructor = ipoib_intf_free;
2499 
2500 	if (ipoib_intercept_dev_id_attr(ndev))
2501 		goto sysfs_failed;
2502 	if (ipoib_cm_add_mode_attr(ndev))
2503 		goto sysfs_failed;
2504 	if (ipoib_add_pkey_attr(ndev))
2505 		goto sysfs_failed;
2506 	if (ipoib_add_umcast_attr(ndev))
2507 		goto sysfs_failed;
2508 	if (device_create_file(&ndev->dev, &dev_attr_create_child))
2509 		goto sysfs_failed;
2510 	if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2511 		goto sysfs_failed;
2512 
2513 	return ndev;
2514 
2515 sysfs_failed:
2516 	ipoib_parent_unregister_pre(ndev);
2517 	unregister_netdev(ndev);
2518 	return ERR_PTR(-ENOMEM);
2519 }
2520 
2521 static void ipoib_add_one(struct ib_device *device)
2522 {
2523 	struct list_head *dev_list;
2524 	struct net_device *dev;
2525 	struct ipoib_dev_priv *priv;
2526 	unsigned int p;
2527 	int count = 0;
2528 
2529 	dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2530 	if (!dev_list)
2531 		return;
2532 
2533 	INIT_LIST_HEAD(dev_list);
2534 
2535 	rdma_for_each_port (device, p) {
2536 		if (!rdma_protocol_ib(device, p))
2537 			continue;
2538 		dev = ipoib_add_port("ib%d", device, p);
2539 		if (!IS_ERR(dev)) {
2540 			priv = ipoib_priv(dev);
2541 			list_add_tail(&priv->list, dev_list);
2542 			count++;
2543 		}
2544 	}
2545 
2546 	if (!count) {
2547 		kfree(dev_list);
2548 		return;
2549 	}
2550 
2551 	ib_set_client_data(device, &ipoib_client, dev_list);
2552 }
2553 
2554 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2555 {
2556 	struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2557 	struct list_head *dev_list = client_data;
2558 
2559 	if (!dev_list)
2560 		return;
2561 
2562 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
2563 		LIST_HEAD(head);
2564 		ipoib_parent_unregister_pre(priv->dev);
2565 
2566 		rtnl_lock();
2567 
2568 		list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2569 					 list)
2570 			unregister_netdevice_queue(cpriv->dev, &head);
2571 		unregister_netdevice_queue(priv->dev, &head);
2572 		unregister_netdevice_many(&head);
2573 
2574 		rtnl_unlock();
2575 	}
2576 
2577 	kfree(dev_list);
2578 }
2579 
2580 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2581 static struct notifier_block ipoib_netdev_notifier = {
2582 	.notifier_call = ipoib_netdev_event,
2583 };
2584 #endif
2585 
2586 static int __init ipoib_init_module(void)
2587 {
2588 	int ret;
2589 
2590 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2591 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2592 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2593 
2594 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2595 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2596 	ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2597 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2598 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2599 	ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2600 #endif
2601 
2602 	/*
2603 	 * When copying small received packets, we only copy from the
2604 	 * linear data part of the SKB, so we rely on this condition.
2605 	 */
2606 	BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2607 
2608 	ipoib_register_debugfs();
2609 
2610 	/*
2611 	 * We create a global workqueue here that is used for all flush
2612 	 * operations.  However, if you attempt to flush a workqueue
2613 	 * from a task on that same workqueue, it deadlocks the system.
2614 	 * We want to be able to flush the tasks associated with a
2615 	 * specific net device, so we also create a workqueue for each
2616 	 * netdevice.  We queue up the tasks for that device only on
2617 	 * its private workqueue, and we only queue up flush events
2618 	 * on our global flush workqueue.  This avoids the deadlocks.
2619 	 */
2620 	ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2621 	if (!ipoib_workqueue) {
2622 		ret = -ENOMEM;
2623 		goto err_fs;
2624 	}
2625 
2626 	ib_sa_register_client(&ipoib_sa_client);
2627 
2628 	ret = ib_register_client(&ipoib_client);
2629 	if (ret)
2630 		goto err_sa;
2631 
2632 	ret = ipoib_netlink_init();
2633 	if (ret)
2634 		goto err_client;
2635 
2636 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2637 	register_netdevice_notifier(&ipoib_netdev_notifier);
2638 #endif
2639 	return 0;
2640 
2641 err_client:
2642 	ib_unregister_client(&ipoib_client);
2643 
2644 err_sa:
2645 	ib_sa_unregister_client(&ipoib_sa_client);
2646 	destroy_workqueue(ipoib_workqueue);
2647 
2648 err_fs:
2649 	ipoib_unregister_debugfs();
2650 
2651 	return ret;
2652 }
2653 
2654 static void __exit ipoib_cleanup_module(void)
2655 {
2656 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2657 	unregister_netdevice_notifier(&ipoib_netdev_notifier);
2658 #endif
2659 	ipoib_netlink_fini();
2660 	ib_unregister_client(&ipoib_client);
2661 	ib_sa_unregister_client(&ipoib_sa_client);
2662 	ipoib_unregister_debugfs();
2663 	destroy_workqueue(ipoib_workqueue);
2664 }
2665 
2666 module_init(ipoib_init_module);
2667 module_exit(ipoib_cleanup_module);
2668