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 #include <linux/pci.h>
55 
56 #define DRV_VERSION "1.0.0"
57 
58 const char ipoib_driver_version[] = DRV_VERSION;
59 
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
62 MODULE_LICENSE("Dual BSD/GPL");
63 MODULE_VERSION(DRV_VERSION);
64 
65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
67 
68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
72 
73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
74 int ipoib_debug_level;
75 
76 module_param_named(debug_level, ipoib_debug_level, int, 0644);
77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
78 #endif
79 
80 struct ipoib_path_iter {
81 	struct net_device *dev;
82 	struct ipoib_path  path;
83 };
84 
85 static const u8 ipv4_bcast_addr[] = {
86 	0x00, 0xff, 0xff, 0xff,
87 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
88 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
89 };
90 
91 struct workqueue_struct *ipoib_workqueue;
92 
93 struct ib_sa_client ipoib_sa_client;
94 
95 static void ipoib_add_one(struct ib_device *device);
96 static void ipoib_remove_one(struct ib_device *device, void *client_data);
97 static void ipoib_neigh_reclaim(struct rcu_head *rp);
98 static struct net_device *ipoib_get_net_dev_by_params(
99 		struct ib_device *dev, u8 port, u16 pkey,
100 		const union ib_gid *gid, const struct sockaddr *addr,
101 		void *client_data);
102 static int ipoib_set_mac(struct net_device *dev, void *addr);
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 int ipoib_open(struct net_device *dev)
112 {
113 	struct ipoib_dev_priv *priv = netdev_priv(dev);
114 
115 	ipoib_dbg(priv, "bringing up interface\n");
116 
117 	netif_carrier_off(dev);
118 
119 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
120 
121 	priv->sm_fullmember_sendonly_support = false;
122 
123 	if (ipoib_ib_dev_open(dev)) {
124 		if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
125 			return 0;
126 		goto err_disable;
127 	}
128 
129 	ipoib_ib_dev_up(dev);
130 
131 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
132 		struct ipoib_dev_priv *cpriv;
133 
134 		/* Bring up any child interfaces too */
135 		down_read(&priv->vlan_rwsem);
136 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
137 			int flags;
138 
139 			flags = cpriv->dev->flags;
140 			if (flags & IFF_UP)
141 				continue;
142 
143 			dev_change_flags(cpriv->dev, flags | IFF_UP);
144 		}
145 		up_read(&priv->vlan_rwsem);
146 	}
147 
148 	netif_start_queue(dev);
149 
150 	return 0;
151 
152 err_disable:
153 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
154 
155 	return -EINVAL;
156 }
157 
158 static int ipoib_stop(struct net_device *dev)
159 {
160 	struct ipoib_dev_priv *priv = netdev_priv(dev);
161 
162 	ipoib_dbg(priv, "stopping interface\n");
163 
164 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
165 
166 	netif_stop_queue(dev);
167 
168 	ipoib_ib_dev_down(dev);
169 	ipoib_ib_dev_stop(dev);
170 
171 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
172 		struct ipoib_dev_priv *cpriv;
173 
174 		/* Bring down any child interfaces too */
175 		down_read(&priv->vlan_rwsem);
176 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
177 			int flags;
178 
179 			flags = cpriv->dev->flags;
180 			if (!(flags & IFF_UP))
181 				continue;
182 
183 			dev_change_flags(cpriv->dev, flags & ~IFF_UP);
184 		}
185 		up_read(&priv->vlan_rwsem);
186 	}
187 
188 	return 0;
189 }
190 
191 static void ipoib_uninit(struct net_device *dev)
192 {
193 	ipoib_dev_cleanup(dev);
194 }
195 
196 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
197 {
198 	struct ipoib_dev_priv *priv = netdev_priv(dev);
199 
200 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
201 		features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
202 
203 	return features;
204 }
205 
206 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
207 {
208 	struct ipoib_dev_priv *priv = netdev_priv(dev);
209 
210 	/* dev->mtu > 2K ==> connected mode */
211 	if (ipoib_cm_admin_enabled(dev)) {
212 		if (new_mtu > ipoib_cm_max_mtu(dev))
213 			return -EINVAL;
214 
215 		if (new_mtu > priv->mcast_mtu)
216 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
217 				   priv->mcast_mtu);
218 
219 		dev->mtu = new_mtu;
220 		return 0;
221 	}
222 
223 	if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
224 		return -EINVAL;
225 
226 	priv->admin_mtu = new_mtu;
227 
228 	if (priv->mcast_mtu < priv->admin_mtu)
229 		ipoib_dbg(priv, "MTU must be smaller than the underlying "
230 				"link layer MTU - 4 (%u)\n", priv->mcast_mtu);
231 
232 	dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
233 
234 	return 0;
235 }
236 
237 /* Called with an RCU read lock taken */
238 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
239 					struct net_device *dev)
240 {
241 	struct net *net = dev_net(dev);
242 	struct in_device *in_dev;
243 	struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
244 	struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
245 	__be32 ret_addr;
246 
247 	switch (addr->sa_family) {
248 	case AF_INET:
249 		in_dev = in_dev_get(dev);
250 		if (!in_dev)
251 			return false;
252 
253 		ret_addr = inet_confirm_addr(net, in_dev, 0,
254 					     addr_in->sin_addr.s_addr,
255 					     RT_SCOPE_HOST);
256 		in_dev_put(in_dev);
257 		if (ret_addr)
258 			return true;
259 
260 		break;
261 	case AF_INET6:
262 		if (IS_ENABLED(CONFIG_IPV6) &&
263 		    ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
264 			return true;
265 
266 		break;
267 	}
268 	return false;
269 }
270 
271 /**
272  * Find the master net_device on top of the given net_device.
273  * @dev: base IPoIB net_device
274  *
275  * Returns the master net_device with a reference held, or the same net_device
276  * if no master exists.
277  */
278 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
279 {
280 	struct net_device *master;
281 
282 	rcu_read_lock();
283 	master = netdev_master_upper_dev_get_rcu(dev);
284 	if (master)
285 		dev_hold(master);
286 	rcu_read_unlock();
287 
288 	if (master)
289 		return master;
290 
291 	dev_hold(dev);
292 	return dev;
293 }
294 
295 struct ipoib_walk_data {
296 	const struct sockaddr *addr;
297 	struct net_device *result;
298 };
299 
300 static int ipoib_upper_walk(struct net_device *upper, void *_data)
301 {
302 	struct ipoib_walk_data *data = _data;
303 	int ret = 0;
304 
305 	if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
306 		dev_hold(upper);
307 		data->result = upper;
308 		ret = 1;
309 	}
310 
311 	return ret;
312 }
313 
314 /**
315  * Find a net_device matching the given address, which is an upper device of
316  * the given net_device.
317  * @addr: IP address to look for.
318  * @dev: base IPoIB net_device
319  *
320  * If found, returns the net_device with a reference held. Otherwise return
321  * NULL.
322  */
323 static struct net_device *ipoib_get_net_dev_match_addr(
324 		const struct sockaddr *addr, struct net_device *dev)
325 {
326 	struct ipoib_walk_data data = {
327 		.addr = addr,
328 	};
329 
330 	rcu_read_lock();
331 	if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
332 		dev_hold(dev);
333 		data.result = dev;
334 		goto out;
335 	}
336 
337 	netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
338 out:
339 	rcu_read_unlock();
340 	return data.result;
341 }
342 
343 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
344  * pkey_index and address, if one exists.
345  *
346  * @found_net_dev: contains a matching net_device if the return value >= 1,
347  * with a reference held. */
348 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
349 				     const union ib_gid *gid,
350 				     u16 pkey_index,
351 				     const struct sockaddr *addr,
352 				     int nesting,
353 				     struct net_device **found_net_dev)
354 {
355 	struct ipoib_dev_priv *child_priv;
356 	struct net_device *net_dev = NULL;
357 	int matches = 0;
358 
359 	if (priv->pkey_index == pkey_index &&
360 	    (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
361 		if (!addr) {
362 			net_dev = ipoib_get_master_net_dev(priv->dev);
363 		} else {
364 			/* Verify the net_device matches the IP address, as
365 			 * IPoIB child devices currently share a GID. */
366 			net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
367 		}
368 		if (net_dev) {
369 			if (!*found_net_dev)
370 				*found_net_dev = net_dev;
371 			else
372 				dev_put(net_dev);
373 			++matches;
374 		}
375 	}
376 
377 	/* Check child interfaces */
378 	down_read_nested(&priv->vlan_rwsem, nesting);
379 	list_for_each_entry(child_priv, &priv->child_intfs, list) {
380 		matches += ipoib_match_gid_pkey_addr(child_priv, gid,
381 						    pkey_index, addr,
382 						    nesting + 1,
383 						    found_net_dev);
384 		if (matches > 1)
385 			break;
386 	}
387 	up_read(&priv->vlan_rwsem);
388 
389 	return matches;
390 }
391 
392 /* Returns the number of matching net_devs found (between 0 and 2). Also
393  * return the matching net_device in the @net_dev parameter, holding a
394  * reference to the net_device, if the number of matches >= 1 */
395 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
396 					 u16 pkey_index,
397 					 const union ib_gid *gid,
398 					 const struct sockaddr *addr,
399 					 struct net_device **net_dev)
400 {
401 	struct ipoib_dev_priv *priv;
402 	int matches = 0;
403 
404 	*net_dev = NULL;
405 
406 	list_for_each_entry(priv, dev_list, list) {
407 		if (priv->port != port)
408 			continue;
409 
410 		matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
411 						     addr, 0, net_dev);
412 		if (matches > 1)
413 			break;
414 	}
415 
416 	return matches;
417 }
418 
419 static struct net_device *ipoib_get_net_dev_by_params(
420 		struct ib_device *dev, u8 port, u16 pkey,
421 		const union ib_gid *gid, const struct sockaddr *addr,
422 		void *client_data)
423 {
424 	struct net_device *net_dev;
425 	struct list_head *dev_list = client_data;
426 	u16 pkey_index;
427 	int matches;
428 	int ret;
429 
430 	if (!rdma_protocol_ib(dev, port))
431 		return NULL;
432 
433 	ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
434 	if (ret)
435 		return NULL;
436 
437 	if (!dev_list)
438 		return NULL;
439 
440 	/* See if we can find a unique device matching the L2 parameters */
441 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
442 						gid, NULL, &net_dev);
443 
444 	switch (matches) {
445 	case 0:
446 		return NULL;
447 	case 1:
448 		return net_dev;
449 	}
450 
451 	dev_put(net_dev);
452 
453 	/* Couldn't find a unique device with L2 parameters only. Use L3
454 	 * address to uniquely match the net device */
455 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
456 						gid, addr, &net_dev);
457 	switch (matches) {
458 	case 0:
459 		return NULL;
460 	default:
461 		dev_warn_ratelimited(&dev->dev,
462 				     "duplicate IP address detected\n");
463 		/* Fall through */
464 	case 1:
465 		return net_dev;
466 	}
467 }
468 
469 int ipoib_set_mode(struct net_device *dev, const char *buf)
470 {
471 	struct ipoib_dev_priv *priv = netdev_priv(dev);
472 
473 	if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
474 	     !strcmp(buf, "connected\n")) ||
475 	     (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
476 	     !strcmp(buf, "datagram\n"))) {
477 		return 0;
478 	}
479 
480 	/* flush paths if we switch modes so that connections are restarted */
481 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
482 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
483 		ipoib_warn(priv, "enabling connected mode "
484 			   "will cause multicast packet drops\n");
485 		netdev_update_features(dev);
486 		dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
487 		rtnl_unlock();
488 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
489 
490 		ipoib_flush_paths(dev);
491 		return (!rtnl_trylock()) ? -EBUSY : 0;
492 	}
493 
494 	if (!strcmp(buf, "datagram\n")) {
495 		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
496 		netdev_update_features(dev);
497 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
498 		rtnl_unlock();
499 		ipoib_flush_paths(dev);
500 		return (!rtnl_trylock()) ? -EBUSY : 0;
501 	}
502 
503 	return -EINVAL;
504 }
505 
506 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
507 {
508 	struct ipoib_dev_priv *priv = netdev_priv(dev);
509 	struct rb_node *n = priv->path_tree.rb_node;
510 	struct ipoib_path *path;
511 	int ret;
512 
513 	while (n) {
514 		path = rb_entry(n, struct ipoib_path, rb_node);
515 
516 		ret = memcmp(gid, path->pathrec.dgid.raw,
517 			     sizeof (union ib_gid));
518 
519 		if (ret < 0)
520 			n = n->rb_left;
521 		else if (ret > 0)
522 			n = n->rb_right;
523 		else
524 			return path;
525 	}
526 
527 	return NULL;
528 }
529 
530 static int __path_add(struct net_device *dev, struct ipoib_path *path)
531 {
532 	struct ipoib_dev_priv *priv = netdev_priv(dev);
533 	struct rb_node **n = &priv->path_tree.rb_node;
534 	struct rb_node *pn = NULL;
535 	struct ipoib_path *tpath;
536 	int ret;
537 
538 	while (*n) {
539 		pn = *n;
540 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
541 
542 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
543 			     sizeof (union ib_gid));
544 		if (ret < 0)
545 			n = &pn->rb_left;
546 		else if (ret > 0)
547 			n = &pn->rb_right;
548 		else
549 			return -EEXIST;
550 	}
551 
552 	rb_link_node(&path->rb_node, pn, n);
553 	rb_insert_color(&path->rb_node, &priv->path_tree);
554 
555 	list_add_tail(&path->list, &priv->path_list);
556 
557 	return 0;
558 }
559 
560 static void path_free(struct net_device *dev, struct ipoib_path *path)
561 {
562 	struct sk_buff *skb;
563 
564 	while ((skb = __skb_dequeue(&path->queue)))
565 		dev_kfree_skb_irq(skb);
566 
567 	ipoib_dbg(netdev_priv(dev), "path_free\n");
568 
569 	/* remove all neigh connected to this path */
570 	ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
571 
572 	if (path->ah)
573 		ipoib_put_ah(path->ah);
574 
575 	kfree(path);
576 }
577 
578 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
579 
580 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
581 {
582 	struct ipoib_path_iter *iter;
583 
584 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
585 	if (!iter)
586 		return NULL;
587 
588 	iter->dev = dev;
589 	memset(iter->path.pathrec.dgid.raw, 0, 16);
590 
591 	if (ipoib_path_iter_next(iter)) {
592 		kfree(iter);
593 		return NULL;
594 	}
595 
596 	return iter;
597 }
598 
599 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
600 {
601 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
602 	struct rb_node *n;
603 	struct ipoib_path *path;
604 	int ret = 1;
605 
606 	spin_lock_irq(&priv->lock);
607 
608 	n = rb_first(&priv->path_tree);
609 
610 	while (n) {
611 		path = rb_entry(n, struct ipoib_path, rb_node);
612 
613 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
614 			   sizeof (union ib_gid)) < 0) {
615 			iter->path = *path;
616 			ret = 0;
617 			break;
618 		}
619 
620 		n = rb_next(n);
621 	}
622 
623 	spin_unlock_irq(&priv->lock);
624 
625 	return ret;
626 }
627 
628 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
629 			  struct ipoib_path *path)
630 {
631 	*path = iter->path;
632 }
633 
634 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
635 
636 void ipoib_mark_paths_invalid(struct net_device *dev)
637 {
638 	struct ipoib_dev_priv *priv = netdev_priv(dev);
639 	struct ipoib_path *path, *tp;
640 
641 	spin_lock_irq(&priv->lock);
642 
643 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
644 		ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
645 			be16_to_cpu(path->pathrec.dlid),
646 			path->pathrec.dgid.raw);
647 		path->valid =  0;
648 	}
649 
650 	spin_unlock_irq(&priv->lock);
651 }
652 
653 struct classport_info_context {
654 	struct ipoib_dev_priv	*priv;
655 	struct completion	done;
656 	struct ib_sa_query	*sa_query;
657 };
658 
659 static void classport_info_query_cb(int status, struct ib_class_port_info *rec,
660 				    void *context)
661 {
662 	struct classport_info_context *cb_ctx = context;
663 	struct ipoib_dev_priv *priv;
664 
665 	WARN_ON(!context);
666 
667 	priv = cb_ctx->priv;
668 
669 	if (status || !rec) {
670 		pr_debug("device: %s failed query classport_info status: %d\n",
671 			 priv->dev->name, status);
672 		/* keeps the default, will try next mcast_restart */
673 		priv->sm_fullmember_sendonly_support = false;
674 		goto out;
675 	}
676 
677 	if (ib_get_cpi_capmask2(rec) &
678 	    IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) {
679 		pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n",
680 			 priv->dev->name);
681 		priv->sm_fullmember_sendonly_support = true;
682 	} else {
683 		pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n",
684 			 priv->dev->name);
685 		priv->sm_fullmember_sendonly_support = false;
686 	}
687 
688 out:
689 	complete(&cb_ctx->done);
690 }
691 
692 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv)
693 {
694 	struct classport_info_context *callback_context;
695 	int ret;
696 
697 	callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL);
698 	if (!callback_context)
699 		return -ENOMEM;
700 
701 	callback_context->priv = priv;
702 	init_completion(&callback_context->done);
703 
704 	ret = ib_sa_classport_info_rec_query(&ipoib_sa_client,
705 					     priv->ca, priv->port, 3000,
706 					     GFP_KERNEL,
707 					     classport_info_query_cb,
708 					     callback_context,
709 					     &callback_context->sa_query);
710 	if (ret < 0) {
711 		pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n",
712 			priv->dev->name, ret);
713 		kfree(callback_context);
714 		return ret;
715 	}
716 
717 	/* waiting for the callback to finish before returnning */
718 	wait_for_completion(&callback_context->done);
719 	kfree(callback_context);
720 
721 	return ret;
722 }
723 
724 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
725 {
726 	struct ipoib_pseudo_header *phdr;
727 
728 	phdr = (struct ipoib_pseudo_header *)skb_push(skb, sizeof(*phdr));
729 	memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
730 }
731 
732 void ipoib_flush_paths(struct net_device *dev)
733 {
734 	struct ipoib_dev_priv *priv = netdev_priv(dev);
735 	struct ipoib_path *path, *tp;
736 	LIST_HEAD(remove_list);
737 	unsigned long flags;
738 
739 	netif_tx_lock_bh(dev);
740 	spin_lock_irqsave(&priv->lock, flags);
741 
742 	list_splice_init(&priv->path_list, &remove_list);
743 
744 	list_for_each_entry(path, &remove_list, list)
745 		rb_erase(&path->rb_node, &priv->path_tree);
746 
747 	list_for_each_entry_safe(path, tp, &remove_list, list) {
748 		if (path->query)
749 			ib_sa_cancel_query(path->query_id, path->query);
750 		spin_unlock_irqrestore(&priv->lock, flags);
751 		netif_tx_unlock_bh(dev);
752 		wait_for_completion(&path->done);
753 		path_free(dev, path);
754 		netif_tx_lock_bh(dev);
755 		spin_lock_irqsave(&priv->lock, flags);
756 	}
757 
758 	spin_unlock_irqrestore(&priv->lock, flags);
759 	netif_tx_unlock_bh(dev);
760 }
761 
762 static void path_rec_completion(int status,
763 				struct ib_sa_path_rec *pathrec,
764 				void *path_ptr)
765 {
766 	struct ipoib_path *path = path_ptr;
767 	struct net_device *dev = path->dev;
768 	struct ipoib_dev_priv *priv = netdev_priv(dev);
769 	struct ipoib_ah *ah = NULL;
770 	struct ipoib_ah *old_ah = NULL;
771 	struct ipoib_neigh *neigh, *tn;
772 	struct sk_buff_head skqueue;
773 	struct sk_buff *skb;
774 	unsigned long flags;
775 
776 	if (!status)
777 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
778 			  be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
779 	else
780 		ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
781 			  status, path->pathrec.dgid.raw);
782 
783 	skb_queue_head_init(&skqueue);
784 
785 	if (!status) {
786 		struct ib_ah_attr av;
787 
788 		if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
789 			ah = ipoib_create_ah(dev, priv->pd, &av);
790 	}
791 
792 	spin_lock_irqsave(&priv->lock, flags);
793 
794 	if (!IS_ERR_OR_NULL(ah)) {
795 		path->pathrec = *pathrec;
796 
797 		old_ah   = path->ah;
798 		path->ah = ah;
799 
800 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
801 			  ah, be16_to_cpu(pathrec->dlid), 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->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 struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
860 {
861 	struct ipoib_dev_priv *priv = netdev_priv(dev);
862 	struct ipoib_path *path;
863 
864 	if (!priv->broadcast)
865 		return NULL;
866 
867 	path = kzalloc(sizeof *path, GFP_ATOMIC);
868 	if (!path)
869 		return NULL;
870 
871 	path->dev = dev;
872 
873 	skb_queue_head_init(&path->queue);
874 
875 	INIT_LIST_HEAD(&path->neigh_list);
876 
877 	memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
878 	path->pathrec.sgid	    = priv->local_gid;
879 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
880 	path->pathrec.numb_path     = 1;
881 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
882 
883 	return path;
884 }
885 
886 static int path_rec_start(struct net_device *dev,
887 			  struct ipoib_path *path)
888 {
889 	struct ipoib_dev_priv *priv = netdev_priv(dev);
890 
891 	ipoib_dbg(priv, "Start path record lookup for %pI6\n",
892 		  path->pathrec.dgid.raw);
893 
894 	init_completion(&path->done);
895 
896 	path->query_id =
897 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
898 				   &path->pathrec,
899 				   IB_SA_PATH_REC_DGID		|
900 				   IB_SA_PATH_REC_SGID		|
901 				   IB_SA_PATH_REC_NUMB_PATH	|
902 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
903 				   IB_SA_PATH_REC_PKEY,
904 				   1000, GFP_ATOMIC,
905 				   path_rec_completion,
906 				   path, &path->query);
907 	if (path->query_id < 0) {
908 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
909 		path->query = NULL;
910 		complete(&path->done);
911 		return path->query_id;
912 	}
913 
914 	return 0;
915 }
916 
917 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
918 			   struct net_device *dev)
919 {
920 	struct ipoib_dev_priv *priv = netdev_priv(dev);
921 	struct ipoib_path *path;
922 	struct ipoib_neigh *neigh;
923 	unsigned long flags;
924 
925 	spin_lock_irqsave(&priv->lock, flags);
926 	neigh = ipoib_neigh_alloc(daddr, dev);
927 	if (!neigh) {
928 		spin_unlock_irqrestore(&priv->lock, flags);
929 		++dev->stats.tx_dropped;
930 		dev_kfree_skb_any(skb);
931 		return;
932 	}
933 
934 	path = __path_find(dev, daddr + 4);
935 	if (!path) {
936 		path = path_rec_create(dev, daddr + 4);
937 		if (!path)
938 			goto err_path;
939 
940 		__path_add(dev, path);
941 	}
942 
943 	list_add_tail(&neigh->list, &path->neigh_list);
944 
945 	if (path->ah) {
946 		kref_get(&path->ah->ref);
947 		neigh->ah = path->ah;
948 
949 		if (ipoib_cm_enabled(dev, neigh->daddr)) {
950 			if (!ipoib_cm_get(neigh))
951 				ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
952 			if (!ipoib_cm_get(neigh)) {
953 				ipoib_neigh_free(neigh);
954 				goto err_drop;
955 			}
956 			if (skb_queue_len(&neigh->queue) <
957 			    IPOIB_MAX_PATH_REC_QUEUE) {
958 				push_pseudo_header(skb, neigh->daddr);
959 				__skb_queue_tail(&neigh->queue, skb);
960 			} else {
961 				ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
962 					   skb_queue_len(&neigh->queue));
963 				goto err_drop;
964 			}
965 		} else {
966 			spin_unlock_irqrestore(&priv->lock, flags);
967 			ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
968 			ipoib_neigh_put(neigh);
969 			return;
970 		}
971 	} else {
972 		neigh->ah  = NULL;
973 
974 		if (!path->query && path_rec_start(dev, path))
975 			goto err_path;
976 		if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
977 			push_pseudo_header(skb, neigh->daddr);
978 			__skb_queue_tail(&neigh->queue, skb);
979 		} else {
980 			goto err_drop;
981 		}
982 	}
983 
984 	spin_unlock_irqrestore(&priv->lock, flags);
985 	ipoib_neigh_put(neigh);
986 	return;
987 
988 err_path:
989 	ipoib_neigh_free(neigh);
990 err_drop:
991 	++dev->stats.tx_dropped;
992 	dev_kfree_skb_any(skb);
993 
994 	spin_unlock_irqrestore(&priv->lock, flags);
995 	ipoib_neigh_put(neigh);
996 }
997 
998 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
999 			     struct ipoib_pseudo_header *phdr)
1000 {
1001 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1002 	struct ipoib_path *path;
1003 	unsigned long flags;
1004 
1005 	spin_lock_irqsave(&priv->lock, flags);
1006 
1007 	path = __path_find(dev, phdr->hwaddr + 4);
1008 	if (!path || !path->valid) {
1009 		int new_path = 0;
1010 
1011 		if (!path) {
1012 			path = path_rec_create(dev, phdr->hwaddr + 4);
1013 			new_path = 1;
1014 		}
1015 		if (path) {
1016 			if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1017 				push_pseudo_header(skb, phdr->hwaddr);
1018 				__skb_queue_tail(&path->queue, skb);
1019 			} else {
1020 				++dev->stats.tx_dropped;
1021 				dev_kfree_skb_any(skb);
1022 			}
1023 
1024 			if (!path->query && path_rec_start(dev, path)) {
1025 				spin_unlock_irqrestore(&priv->lock, flags);
1026 				if (new_path)
1027 					path_free(dev, path);
1028 				return;
1029 			} else
1030 				__path_add(dev, path);
1031 		} else {
1032 			++dev->stats.tx_dropped;
1033 			dev_kfree_skb_any(skb);
1034 		}
1035 
1036 		spin_unlock_irqrestore(&priv->lock, flags);
1037 		return;
1038 	}
1039 
1040 	if (path->ah) {
1041 		ipoib_dbg(priv, "Send unicast ARP to %04x\n",
1042 			  be16_to_cpu(path->pathrec.dlid));
1043 
1044 		spin_unlock_irqrestore(&priv->lock, flags);
1045 		ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
1046 		return;
1047 	} else if ((path->query || !path_rec_start(dev, path)) &&
1048 		   skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1049 		push_pseudo_header(skb, phdr->hwaddr);
1050 		__skb_queue_tail(&path->queue, skb);
1051 	} else {
1052 		++dev->stats.tx_dropped;
1053 		dev_kfree_skb_any(skb);
1054 	}
1055 
1056 	spin_unlock_irqrestore(&priv->lock, flags);
1057 }
1058 
1059 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1060 {
1061 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1062 	struct ipoib_neigh *neigh;
1063 	struct ipoib_pseudo_header *phdr;
1064 	struct ipoib_header *header;
1065 	unsigned long flags;
1066 
1067 	phdr = (struct ipoib_pseudo_header *) skb->data;
1068 	skb_pull(skb, sizeof(*phdr));
1069 	header = (struct ipoib_header *) skb->data;
1070 
1071 	if (unlikely(phdr->hwaddr[4] == 0xff)) {
1072 		/* multicast, arrange "if" according to probability */
1073 		if ((header->proto != htons(ETH_P_IP)) &&
1074 		    (header->proto != htons(ETH_P_IPV6)) &&
1075 		    (header->proto != htons(ETH_P_ARP)) &&
1076 		    (header->proto != htons(ETH_P_RARP)) &&
1077 		    (header->proto != htons(ETH_P_TIPC))) {
1078 			/* ethertype not supported by IPoIB */
1079 			++dev->stats.tx_dropped;
1080 			dev_kfree_skb_any(skb);
1081 			return NETDEV_TX_OK;
1082 		}
1083 		/* Add in the P_Key for multicast*/
1084 		phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1085 		phdr->hwaddr[9] = priv->pkey & 0xff;
1086 
1087 		neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1088 		if (likely(neigh))
1089 			goto send_using_neigh;
1090 		ipoib_mcast_send(dev, phdr->hwaddr, skb);
1091 		return NETDEV_TX_OK;
1092 	}
1093 
1094 	/* unicast, arrange "switch" according to probability */
1095 	switch (header->proto) {
1096 	case htons(ETH_P_IP):
1097 	case htons(ETH_P_IPV6):
1098 	case htons(ETH_P_TIPC):
1099 		neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1100 		if (unlikely(!neigh)) {
1101 			neigh_add_path(skb, phdr->hwaddr, dev);
1102 			return NETDEV_TX_OK;
1103 		}
1104 		break;
1105 	case htons(ETH_P_ARP):
1106 	case htons(ETH_P_RARP):
1107 		/* for unicast ARP and RARP should always perform path find */
1108 		unicast_arp_send(skb, dev, phdr);
1109 		return NETDEV_TX_OK;
1110 	default:
1111 		/* ethertype not supported by IPoIB */
1112 		++dev->stats.tx_dropped;
1113 		dev_kfree_skb_any(skb);
1114 		return NETDEV_TX_OK;
1115 	}
1116 
1117 send_using_neigh:
1118 	/* note we now hold a ref to neigh */
1119 	if (ipoib_cm_get(neigh)) {
1120 		if (ipoib_cm_up(neigh)) {
1121 			ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1122 			goto unref;
1123 		}
1124 	} else if (neigh->ah) {
1125 		ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(phdr->hwaddr));
1126 		goto unref;
1127 	}
1128 
1129 	if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1130 		push_pseudo_header(skb, phdr->hwaddr);
1131 		spin_lock_irqsave(&priv->lock, flags);
1132 		__skb_queue_tail(&neigh->queue, skb);
1133 		spin_unlock_irqrestore(&priv->lock, flags);
1134 	} else {
1135 		++dev->stats.tx_dropped;
1136 		dev_kfree_skb_any(skb);
1137 	}
1138 
1139 unref:
1140 	ipoib_neigh_put(neigh);
1141 
1142 	return NETDEV_TX_OK;
1143 }
1144 
1145 static void ipoib_timeout(struct net_device *dev)
1146 {
1147 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1148 
1149 	ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1150 		   jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1151 	ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1152 		   netif_queue_stopped(dev),
1153 		   priv->tx_head, priv->tx_tail);
1154 	/* XXX reset QP, etc. */
1155 }
1156 
1157 static int ipoib_hard_header(struct sk_buff *skb,
1158 			     struct net_device *dev,
1159 			     unsigned short type,
1160 			     const void *daddr, const void *saddr, unsigned len)
1161 {
1162 	struct ipoib_header *header;
1163 
1164 	header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1165 
1166 	header->proto = htons(type);
1167 	header->reserved = 0;
1168 
1169 	/*
1170 	 * we don't rely on dst_entry structure,  always stuff the
1171 	 * destination address into skb hard header so we can figure out where
1172 	 * to send the packet later.
1173 	 */
1174 	push_pseudo_header(skb, daddr);
1175 
1176 	return IPOIB_HARD_LEN;
1177 }
1178 
1179 static void ipoib_set_mcast_list(struct net_device *dev)
1180 {
1181 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1182 
1183 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1184 		ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1185 		return;
1186 	}
1187 
1188 	queue_work(priv->wq, &priv->restart_task);
1189 }
1190 
1191 static int ipoib_get_iflink(const struct net_device *dev)
1192 {
1193 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1194 
1195 	/* parent interface */
1196 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1197 		return dev->ifindex;
1198 
1199 	/* child/vlan interface */
1200 	return priv->parent->ifindex;
1201 }
1202 
1203 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1204 {
1205 	/*
1206 	 * Use only the address parts that contributes to spreading
1207 	 * The subnet prefix is not used as one can not connect to
1208 	 * same remote port (GUID) using the same remote QPN via two
1209 	 * different subnets.
1210 	 */
1211 	 /* qpn octets[1:4) & port GUID octets[12:20) */
1212 	u32 *d32 = (u32 *) daddr;
1213 	u32 hv;
1214 
1215 	hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1216 	return hv & htbl->mask;
1217 }
1218 
1219 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1220 {
1221 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1222 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1223 	struct ipoib_neigh_hash *htbl;
1224 	struct ipoib_neigh *neigh = NULL;
1225 	u32 hash_val;
1226 
1227 	rcu_read_lock_bh();
1228 
1229 	htbl = rcu_dereference_bh(ntbl->htbl);
1230 
1231 	if (!htbl)
1232 		goto out_unlock;
1233 
1234 	hash_val = ipoib_addr_hash(htbl, daddr);
1235 	for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1236 	     neigh != NULL;
1237 	     neigh = rcu_dereference_bh(neigh->hnext)) {
1238 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1239 			/* found, take one ref on behalf of the caller */
1240 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1241 				/* deleted */
1242 				neigh = NULL;
1243 				goto out_unlock;
1244 			}
1245 
1246 			if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1247 				neigh->alive = jiffies;
1248 			goto out_unlock;
1249 		}
1250 	}
1251 
1252 out_unlock:
1253 	rcu_read_unlock_bh();
1254 	return neigh;
1255 }
1256 
1257 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1258 {
1259 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1260 	struct ipoib_neigh_hash *htbl;
1261 	unsigned long neigh_obsolete;
1262 	unsigned long dt;
1263 	unsigned long flags;
1264 	int i;
1265 	LIST_HEAD(remove_list);
1266 
1267 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1268 		return;
1269 
1270 	spin_lock_irqsave(&priv->lock, flags);
1271 
1272 	htbl = rcu_dereference_protected(ntbl->htbl,
1273 					 lockdep_is_held(&priv->lock));
1274 
1275 	if (!htbl)
1276 		goto out_unlock;
1277 
1278 	/* neigh is obsolete if it was idle for two GC periods */
1279 	dt = 2 * arp_tbl.gc_interval;
1280 	neigh_obsolete = jiffies - dt;
1281 	/* handle possible race condition */
1282 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1283 		goto out_unlock;
1284 
1285 	for (i = 0; i < htbl->size; i++) {
1286 		struct ipoib_neigh *neigh;
1287 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1288 
1289 		while ((neigh = rcu_dereference_protected(*np,
1290 							  lockdep_is_held(&priv->lock))) != NULL) {
1291 			/* was the neigh idle for two GC periods */
1292 			if (time_after(neigh_obsolete, neigh->alive)) {
1293 
1294 				ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1295 
1296 				rcu_assign_pointer(*np,
1297 						   rcu_dereference_protected(neigh->hnext,
1298 									     lockdep_is_held(&priv->lock)));
1299 				/* remove from path/mc list */
1300 				list_del_init(&neigh->list);
1301 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1302 			} else {
1303 				np = &neigh->hnext;
1304 			}
1305 
1306 		}
1307 	}
1308 
1309 out_unlock:
1310 	spin_unlock_irqrestore(&priv->lock, flags);
1311 	ipoib_mcast_remove_list(&remove_list);
1312 }
1313 
1314 static void ipoib_reap_neigh(struct work_struct *work)
1315 {
1316 	struct ipoib_dev_priv *priv =
1317 		container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1318 
1319 	__ipoib_reap_neigh(priv);
1320 
1321 	if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1322 		queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1323 				   arp_tbl.gc_interval);
1324 }
1325 
1326 
1327 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1328 				      struct net_device *dev)
1329 {
1330 	struct ipoib_neigh *neigh;
1331 
1332 	neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1333 	if (!neigh)
1334 		return NULL;
1335 
1336 	neigh->dev = dev;
1337 	memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1338 	skb_queue_head_init(&neigh->queue);
1339 	INIT_LIST_HEAD(&neigh->list);
1340 	ipoib_cm_set(neigh, NULL);
1341 	/* one ref on behalf of the caller */
1342 	atomic_set(&neigh->refcnt, 1);
1343 
1344 	return neigh;
1345 }
1346 
1347 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1348 				      struct net_device *dev)
1349 {
1350 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1351 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1352 	struct ipoib_neigh_hash *htbl;
1353 	struct ipoib_neigh *neigh;
1354 	u32 hash_val;
1355 
1356 	htbl = rcu_dereference_protected(ntbl->htbl,
1357 					 lockdep_is_held(&priv->lock));
1358 	if (!htbl) {
1359 		neigh = NULL;
1360 		goto out_unlock;
1361 	}
1362 
1363 	/* need to add a new neigh, but maybe some other thread succeeded?
1364 	 * recalc hash, maybe hash resize took place so we do a search
1365 	 */
1366 	hash_val = ipoib_addr_hash(htbl, daddr);
1367 	for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1368 					       lockdep_is_held(&priv->lock));
1369 	     neigh != NULL;
1370 	     neigh = rcu_dereference_protected(neigh->hnext,
1371 					       lockdep_is_held(&priv->lock))) {
1372 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1373 			/* found, take one ref on behalf of the caller */
1374 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1375 				/* deleted */
1376 				neigh = NULL;
1377 				break;
1378 			}
1379 			neigh->alive = jiffies;
1380 			goto out_unlock;
1381 		}
1382 	}
1383 
1384 	neigh = ipoib_neigh_ctor(daddr, dev);
1385 	if (!neigh)
1386 		goto out_unlock;
1387 
1388 	/* one ref on behalf of the hash table */
1389 	atomic_inc(&neigh->refcnt);
1390 	neigh->alive = jiffies;
1391 	/* put in hash */
1392 	rcu_assign_pointer(neigh->hnext,
1393 			   rcu_dereference_protected(htbl->buckets[hash_val],
1394 						     lockdep_is_held(&priv->lock)));
1395 	rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1396 	atomic_inc(&ntbl->entries);
1397 
1398 out_unlock:
1399 
1400 	return neigh;
1401 }
1402 
1403 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1404 {
1405 	/* neigh reference count was dropprd to zero */
1406 	struct net_device *dev = neigh->dev;
1407 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1408 	struct sk_buff *skb;
1409 	if (neigh->ah)
1410 		ipoib_put_ah(neigh->ah);
1411 	while ((skb = __skb_dequeue(&neigh->queue))) {
1412 		++dev->stats.tx_dropped;
1413 		dev_kfree_skb_any(skb);
1414 	}
1415 	if (ipoib_cm_get(neigh))
1416 		ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1417 	ipoib_dbg(netdev_priv(dev),
1418 		  "neigh free for %06x %pI6\n",
1419 		  IPOIB_QPN(neigh->daddr),
1420 		  neigh->daddr + 4);
1421 	kfree(neigh);
1422 	if (atomic_dec_and_test(&priv->ntbl.entries)) {
1423 		if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1424 			complete(&priv->ntbl.flushed);
1425 	}
1426 }
1427 
1428 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1429 {
1430 	/* Called as a result of removal from hash table */
1431 	struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1432 	/* note TX context may hold another ref */
1433 	ipoib_neigh_put(neigh);
1434 }
1435 
1436 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1437 {
1438 	struct net_device *dev = neigh->dev;
1439 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1440 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1441 	struct ipoib_neigh_hash *htbl;
1442 	struct ipoib_neigh __rcu **np;
1443 	struct ipoib_neigh *n;
1444 	u32 hash_val;
1445 
1446 	htbl = rcu_dereference_protected(ntbl->htbl,
1447 					lockdep_is_held(&priv->lock));
1448 	if (!htbl)
1449 		return;
1450 
1451 	hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1452 	np = &htbl->buckets[hash_val];
1453 	for (n = rcu_dereference_protected(*np,
1454 					    lockdep_is_held(&priv->lock));
1455 	     n != NULL;
1456 	     n = rcu_dereference_protected(*np,
1457 					lockdep_is_held(&priv->lock))) {
1458 		if (n == neigh) {
1459 			/* found */
1460 			rcu_assign_pointer(*np,
1461 					   rcu_dereference_protected(neigh->hnext,
1462 								     lockdep_is_held(&priv->lock)));
1463 			/* remove from parent list */
1464 			list_del_init(&neigh->list);
1465 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1466 			return;
1467 		} else {
1468 			np = &n->hnext;
1469 		}
1470 	}
1471 }
1472 
1473 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1474 {
1475 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1476 	struct ipoib_neigh_hash *htbl;
1477 	struct ipoib_neigh __rcu **buckets;
1478 	u32 size;
1479 
1480 	clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1481 	ntbl->htbl = NULL;
1482 	htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1483 	if (!htbl)
1484 		return -ENOMEM;
1485 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1486 	size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1487 	buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1488 	if (!buckets) {
1489 		kfree(htbl);
1490 		return -ENOMEM;
1491 	}
1492 	htbl->size = size;
1493 	htbl->mask = (size - 1);
1494 	htbl->buckets = buckets;
1495 	RCU_INIT_POINTER(ntbl->htbl, htbl);
1496 	htbl->ntbl = ntbl;
1497 	atomic_set(&ntbl->entries, 0);
1498 
1499 	/* start garbage collection */
1500 	clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1501 	queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1502 			   arp_tbl.gc_interval);
1503 
1504 	return 0;
1505 }
1506 
1507 static void neigh_hash_free_rcu(struct rcu_head *head)
1508 {
1509 	struct ipoib_neigh_hash *htbl = container_of(head,
1510 						    struct ipoib_neigh_hash,
1511 						    rcu);
1512 	struct ipoib_neigh __rcu **buckets = htbl->buckets;
1513 	struct ipoib_neigh_table *ntbl = htbl->ntbl;
1514 
1515 	kfree(buckets);
1516 	kfree(htbl);
1517 	complete(&ntbl->deleted);
1518 }
1519 
1520 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1521 {
1522 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1523 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1524 	struct ipoib_neigh_hash *htbl;
1525 	unsigned long flags;
1526 	int i;
1527 
1528 	/* remove all neigh connected to a given path or mcast */
1529 	spin_lock_irqsave(&priv->lock, flags);
1530 
1531 	htbl = rcu_dereference_protected(ntbl->htbl,
1532 					 lockdep_is_held(&priv->lock));
1533 
1534 	if (!htbl)
1535 		goto out_unlock;
1536 
1537 	for (i = 0; i < htbl->size; i++) {
1538 		struct ipoib_neigh *neigh;
1539 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1540 
1541 		while ((neigh = rcu_dereference_protected(*np,
1542 							  lockdep_is_held(&priv->lock))) != NULL) {
1543 			/* delete neighs belong to this parent */
1544 			if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1545 				rcu_assign_pointer(*np,
1546 						   rcu_dereference_protected(neigh->hnext,
1547 									     lockdep_is_held(&priv->lock)));
1548 				/* remove from parent list */
1549 				list_del_init(&neigh->list);
1550 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1551 			} else {
1552 				np = &neigh->hnext;
1553 			}
1554 
1555 		}
1556 	}
1557 out_unlock:
1558 	spin_unlock_irqrestore(&priv->lock, flags);
1559 }
1560 
1561 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1562 {
1563 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1564 	struct ipoib_neigh_hash *htbl;
1565 	unsigned long flags;
1566 	int i, wait_flushed = 0;
1567 
1568 	init_completion(&priv->ntbl.flushed);
1569 
1570 	spin_lock_irqsave(&priv->lock, flags);
1571 
1572 	htbl = rcu_dereference_protected(ntbl->htbl,
1573 					lockdep_is_held(&priv->lock));
1574 	if (!htbl)
1575 		goto out_unlock;
1576 
1577 	wait_flushed = atomic_read(&priv->ntbl.entries);
1578 	if (!wait_flushed)
1579 		goto free_htbl;
1580 
1581 	for (i = 0; i < htbl->size; i++) {
1582 		struct ipoib_neigh *neigh;
1583 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1584 
1585 		while ((neigh = rcu_dereference_protected(*np,
1586 				       lockdep_is_held(&priv->lock))) != NULL) {
1587 			rcu_assign_pointer(*np,
1588 					   rcu_dereference_protected(neigh->hnext,
1589 								     lockdep_is_held(&priv->lock)));
1590 			/* remove from path/mc list */
1591 			list_del_init(&neigh->list);
1592 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1593 		}
1594 	}
1595 
1596 free_htbl:
1597 	rcu_assign_pointer(ntbl->htbl, NULL);
1598 	call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1599 
1600 out_unlock:
1601 	spin_unlock_irqrestore(&priv->lock, flags);
1602 	if (wait_flushed)
1603 		wait_for_completion(&priv->ntbl.flushed);
1604 }
1605 
1606 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1607 {
1608 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1609 	int stopped;
1610 
1611 	ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1612 	init_completion(&priv->ntbl.deleted);
1613 	set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1614 
1615 	/* Stop GC if called at init fail need to cancel work */
1616 	stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1617 	if (!stopped)
1618 		cancel_delayed_work(&priv->neigh_reap_task);
1619 
1620 	ipoib_flush_neighs(priv);
1621 
1622 	wait_for_completion(&priv->ntbl.deleted);
1623 }
1624 
1625 
1626 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1627 {
1628 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1629 
1630 	/* Allocate RX/TX "rings" to hold queued skbs */
1631 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1632 				GFP_KERNEL);
1633 	if (!priv->rx_ring)
1634 		goto out;
1635 
1636 	priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1637 	if (!priv->tx_ring) {
1638 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1639 		       ca->name, ipoib_sendq_size);
1640 		goto out_rx_ring_cleanup;
1641 	}
1642 
1643 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1644 
1645 	if (ipoib_ib_dev_init(dev, ca, port))
1646 		goto out_tx_ring_cleanup;
1647 
1648 	/*
1649 	 * Must be after ipoib_ib_dev_init so we can allocate a per
1650 	 * device wq there and use it here
1651 	 */
1652 	if (ipoib_neigh_hash_init(priv) < 0)
1653 		goto out_dev_uninit;
1654 
1655 	return 0;
1656 
1657 out_dev_uninit:
1658 	ipoib_ib_dev_cleanup(dev);
1659 
1660 out_tx_ring_cleanup:
1661 	vfree(priv->tx_ring);
1662 
1663 out_rx_ring_cleanup:
1664 	kfree(priv->rx_ring);
1665 
1666 out:
1667 	return -ENOMEM;
1668 }
1669 
1670 void ipoib_dev_cleanup(struct net_device *dev)
1671 {
1672 	struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1673 	LIST_HEAD(head);
1674 
1675 	ASSERT_RTNL();
1676 
1677 	ipoib_delete_debug_files(dev);
1678 
1679 	/* Delete any child interfaces first */
1680 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1681 		/* Stop GC on child */
1682 		set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1683 		cancel_delayed_work(&cpriv->neigh_reap_task);
1684 		unregister_netdevice_queue(cpriv->dev, &head);
1685 	}
1686 	unregister_netdevice_many(&head);
1687 
1688 	/*
1689 	 * Must be before ipoib_ib_dev_cleanup or we delete an in use
1690 	 * work queue
1691 	 */
1692 	ipoib_neigh_hash_uninit(dev);
1693 
1694 	ipoib_ib_dev_cleanup(dev);
1695 
1696 	kfree(priv->rx_ring);
1697 	vfree(priv->tx_ring);
1698 
1699 	priv->rx_ring = NULL;
1700 	priv->tx_ring = NULL;
1701 }
1702 
1703 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1704 {
1705 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1706 
1707 	return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1708 }
1709 
1710 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1711 			       struct ifla_vf_info *ivf)
1712 {
1713 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1714 	int err;
1715 
1716 	err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1717 	if (err)
1718 		return err;
1719 
1720 	ivf->vf = vf;
1721 
1722 	return 0;
1723 }
1724 
1725 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1726 {
1727 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1728 
1729 	if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1730 		return -EINVAL;
1731 
1732 	return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1733 }
1734 
1735 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1736 			      struct ifla_vf_stats *vf_stats)
1737 {
1738 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1739 
1740 	return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1741 }
1742 
1743 static const struct header_ops ipoib_header_ops = {
1744 	.create	= ipoib_hard_header,
1745 };
1746 
1747 static const struct net_device_ops ipoib_netdev_ops_pf = {
1748 	.ndo_uninit		 = ipoib_uninit,
1749 	.ndo_open		 = ipoib_open,
1750 	.ndo_stop		 = ipoib_stop,
1751 	.ndo_change_mtu		 = ipoib_change_mtu,
1752 	.ndo_fix_features	 = ipoib_fix_features,
1753 	.ndo_start_xmit		 = ipoib_start_xmit,
1754 	.ndo_tx_timeout		 = ipoib_timeout,
1755 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
1756 	.ndo_get_iflink		 = ipoib_get_iflink,
1757 	.ndo_set_vf_link_state	 = ipoib_set_vf_link_state,
1758 	.ndo_get_vf_config	 = ipoib_get_vf_config,
1759 	.ndo_get_vf_stats	 = ipoib_get_vf_stats,
1760 	.ndo_set_vf_guid	 = ipoib_set_vf_guid,
1761 	.ndo_set_mac_address	 = ipoib_set_mac,
1762 };
1763 
1764 static const struct net_device_ops ipoib_netdev_ops_vf = {
1765 	.ndo_uninit		 = ipoib_uninit,
1766 	.ndo_open		 = ipoib_open,
1767 	.ndo_stop		 = ipoib_stop,
1768 	.ndo_change_mtu		 = ipoib_change_mtu,
1769 	.ndo_fix_features	 = ipoib_fix_features,
1770 	.ndo_start_xmit	 	 = ipoib_start_xmit,
1771 	.ndo_tx_timeout		 = ipoib_timeout,
1772 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
1773 	.ndo_get_iflink		 = ipoib_get_iflink,
1774 };
1775 
1776 void ipoib_setup(struct net_device *dev)
1777 {
1778 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1779 
1780 	if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1781 		dev->netdev_ops	= &ipoib_netdev_ops_vf;
1782 	else
1783 		dev->netdev_ops	= &ipoib_netdev_ops_pf;
1784 
1785 	dev->header_ops		 = &ipoib_header_ops;
1786 
1787 	ipoib_set_ethtool_ops(dev);
1788 
1789 	netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1790 
1791 	dev->watchdog_timeo	 = HZ;
1792 
1793 	dev->flags		|= IFF_BROADCAST | IFF_MULTICAST;
1794 
1795 	dev->hard_header_len	 = IPOIB_HARD_LEN;
1796 	dev->addr_len		 = INFINIBAND_ALEN;
1797 	dev->type		 = ARPHRD_INFINIBAND;
1798 	dev->tx_queue_len	 = ipoib_sendq_size * 2;
1799 	dev->features		 = (NETIF_F_VLAN_CHALLENGED	|
1800 				    NETIF_F_HIGHDMA);
1801 	netif_keep_dst(dev);
1802 
1803 	memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1804 
1805 	priv->dev = dev;
1806 
1807 	spin_lock_init(&priv->lock);
1808 
1809 	init_rwsem(&priv->vlan_rwsem);
1810 
1811 	INIT_LIST_HEAD(&priv->path_list);
1812 	INIT_LIST_HEAD(&priv->child_intfs);
1813 	INIT_LIST_HEAD(&priv->dead_ahs);
1814 	INIT_LIST_HEAD(&priv->multicast_list);
1815 
1816 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1817 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1818 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1819 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1820 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1821 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1822 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1823 	INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1824 }
1825 
1826 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1827 {
1828 	struct net_device *dev;
1829 
1830 	dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1831 			   NET_NAME_UNKNOWN, ipoib_setup);
1832 	if (!dev)
1833 		return NULL;
1834 
1835 	return netdev_priv(dev);
1836 }
1837 
1838 static ssize_t show_pkey(struct device *dev,
1839 			 struct device_attribute *attr, char *buf)
1840 {
1841 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1842 
1843 	return sprintf(buf, "0x%04x\n", priv->pkey);
1844 }
1845 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1846 
1847 static ssize_t show_umcast(struct device *dev,
1848 			   struct device_attribute *attr, char *buf)
1849 {
1850 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1851 
1852 	return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1853 }
1854 
1855 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1856 {
1857 	struct ipoib_dev_priv *priv = netdev_priv(ndev);
1858 
1859 	if (umcast_val > 0) {
1860 		set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1861 		ipoib_warn(priv, "ignoring multicast groups joined directly "
1862 				"by userspace\n");
1863 	} else
1864 		clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1865 }
1866 
1867 static ssize_t set_umcast(struct device *dev,
1868 			  struct device_attribute *attr,
1869 			  const char *buf, size_t count)
1870 {
1871 	unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1872 
1873 	ipoib_set_umcast(to_net_dev(dev), umcast_val);
1874 
1875 	return count;
1876 }
1877 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1878 
1879 int ipoib_add_umcast_attr(struct net_device *dev)
1880 {
1881 	return device_create_file(&dev->dev, &dev_attr_umcast);
1882 }
1883 
1884 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
1885 {
1886 	struct ipoib_dev_priv *child_priv;
1887 	struct net_device *netdev = priv->dev;
1888 
1889 	netif_addr_lock_bh(netdev);
1890 
1891 	memcpy(&priv->local_gid.global.interface_id,
1892 	       &gid->global.interface_id,
1893 	       sizeof(gid->global.interface_id));
1894 	memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
1895 	clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1896 
1897 	netif_addr_unlock_bh(netdev);
1898 
1899 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1900 		down_read(&priv->vlan_rwsem);
1901 		list_for_each_entry(child_priv, &priv->child_intfs, list)
1902 			set_base_guid(child_priv, gid);
1903 		up_read(&priv->vlan_rwsem);
1904 	}
1905 }
1906 
1907 static int ipoib_check_lladdr(struct net_device *dev,
1908 			      struct sockaddr_storage *ss)
1909 {
1910 	union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
1911 	int ret = 0;
1912 
1913 	netif_addr_lock_bh(dev);
1914 
1915 	/* Make sure the QPN, reserved and subnet prefix match the current
1916 	 * lladdr, it also makes sure the lladdr is unicast.
1917 	 */
1918 	if (memcmp(dev->dev_addr, ss->__data,
1919 		   4 + sizeof(gid->global.subnet_prefix)) ||
1920 	    gid->global.interface_id == 0)
1921 		ret = -EINVAL;
1922 
1923 	netif_addr_unlock_bh(dev);
1924 
1925 	return ret;
1926 }
1927 
1928 static int ipoib_set_mac(struct net_device *dev, void *addr)
1929 {
1930 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1931 	struct sockaddr_storage *ss = addr;
1932 	int ret;
1933 
1934 	if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
1935 		return -EBUSY;
1936 
1937 	ret = ipoib_check_lladdr(dev, ss);
1938 	if (ret)
1939 		return ret;
1940 
1941 	set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
1942 
1943 	queue_work(ipoib_workqueue, &priv->flush_light);
1944 
1945 	return 0;
1946 }
1947 
1948 static ssize_t create_child(struct device *dev,
1949 			    struct device_attribute *attr,
1950 			    const char *buf, size_t count)
1951 {
1952 	int pkey;
1953 	int ret;
1954 
1955 	if (sscanf(buf, "%i", &pkey) != 1)
1956 		return -EINVAL;
1957 
1958 	if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1959 		return -EINVAL;
1960 
1961 	/*
1962 	 * Set the full membership bit, so that we join the right
1963 	 * broadcast group, etc.
1964 	 */
1965 	pkey |= 0x8000;
1966 
1967 	ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1968 
1969 	return ret ? ret : count;
1970 }
1971 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1972 
1973 static ssize_t delete_child(struct device *dev,
1974 			    struct device_attribute *attr,
1975 			    const char *buf, size_t count)
1976 {
1977 	int pkey;
1978 	int ret;
1979 
1980 	if (sscanf(buf, "%i", &pkey) != 1)
1981 		return -EINVAL;
1982 
1983 	if (pkey < 0 || pkey > 0xffff)
1984 		return -EINVAL;
1985 
1986 	ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1987 
1988 	return ret ? ret : count;
1989 
1990 }
1991 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1992 
1993 int ipoib_add_pkey_attr(struct net_device *dev)
1994 {
1995 	return device_create_file(&dev->dev, &dev_attr_pkey);
1996 }
1997 
1998 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1999 {
2000 	priv->hca_caps = hca->attrs.device_cap_flags;
2001 
2002 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2003 		priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2004 
2005 		if (priv->hca_caps & IB_DEVICE_UD_TSO)
2006 			priv->dev->hw_features |= NETIF_F_TSO;
2007 
2008 		priv->dev->features |= priv->dev->hw_features;
2009 	}
2010 }
2011 
2012 static struct net_device *ipoib_add_port(const char *format,
2013 					 struct ib_device *hca, u8 port)
2014 {
2015 	struct ipoib_dev_priv *priv;
2016 	struct ib_port_attr attr;
2017 	int result = -ENOMEM;
2018 
2019 	priv = ipoib_intf_alloc(format);
2020 	if (!priv)
2021 		goto alloc_mem_failed;
2022 
2023 	SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2024 	priv->dev->dev_id = port - 1;
2025 
2026 	result = ib_query_port(hca, port, &attr);
2027 	if (!result)
2028 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2029 	else {
2030 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2031 		       hca->name, port);
2032 		goto device_init_failed;
2033 	}
2034 
2035 	/* MTU will be reset when mcast join happens */
2036 	priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
2037 	priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
2038 	priv->dev->max_mtu = IPOIB_CM_MTU;
2039 
2040 	priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2041 
2042 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
2043 	if (result) {
2044 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2045 		       hca->name, port, result);
2046 		goto device_init_failed;
2047 	}
2048 
2049 	ipoib_set_dev_features(priv, hca);
2050 
2051 	/*
2052 	 * Set the full membership bit, so that we join the right
2053 	 * broadcast group, etc.
2054 	 */
2055 	priv->pkey |= 0x8000;
2056 
2057 	priv->dev->broadcast[8] = priv->pkey >> 8;
2058 	priv->dev->broadcast[9] = priv->pkey & 0xff;
2059 
2060 	result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2061 	if (result) {
2062 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2063 		       hca->name, port, result);
2064 		goto device_init_failed;
2065 	} else
2066 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2067 	set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2068 
2069 	result = ipoib_dev_init(priv->dev, hca, port);
2070 	if (result < 0) {
2071 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2072 		       hca->name, port, result);
2073 		goto device_init_failed;
2074 	}
2075 
2076 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
2077 			      priv->ca, ipoib_event);
2078 	result = ib_register_event_handler(&priv->event_handler);
2079 	if (result < 0) {
2080 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2081 		       "port %d (ret = %d)\n",
2082 		       hca->name, port, result);
2083 		goto event_failed;
2084 	}
2085 
2086 	result = register_netdev(priv->dev);
2087 	if (result) {
2088 		printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2089 		       hca->name, port, result);
2090 		goto register_failed;
2091 	}
2092 
2093 	ipoib_create_debug_files(priv->dev);
2094 
2095 	if (ipoib_cm_add_mode_attr(priv->dev))
2096 		goto sysfs_failed;
2097 	if (ipoib_add_pkey_attr(priv->dev))
2098 		goto sysfs_failed;
2099 	if (ipoib_add_umcast_attr(priv->dev))
2100 		goto sysfs_failed;
2101 	if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2102 		goto sysfs_failed;
2103 	if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2104 		goto sysfs_failed;
2105 
2106 	return priv->dev;
2107 
2108 sysfs_failed:
2109 	ipoib_delete_debug_files(priv->dev);
2110 	unregister_netdev(priv->dev);
2111 
2112 register_failed:
2113 	ib_unregister_event_handler(&priv->event_handler);
2114 	flush_workqueue(ipoib_workqueue);
2115 	/* Stop GC if started before flush */
2116 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2117 	cancel_delayed_work(&priv->neigh_reap_task);
2118 	flush_workqueue(priv->wq);
2119 
2120 event_failed:
2121 	ipoib_dev_cleanup(priv->dev);
2122 
2123 device_init_failed:
2124 	free_netdev(priv->dev);
2125 
2126 alloc_mem_failed:
2127 	return ERR_PTR(result);
2128 }
2129 
2130 static void ipoib_add_one(struct ib_device *device)
2131 {
2132 	struct list_head *dev_list;
2133 	struct net_device *dev;
2134 	struct ipoib_dev_priv *priv;
2135 	int p;
2136 	int count = 0;
2137 
2138 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2139 	if (!dev_list)
2140 		return;
2141 
2142 	INIT_LIST_HEAD(dev_list);
2143 
2144 	for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2145 		if (!rdma_protocol_ib(device, p))
2146 			continue;
2147 		dev = ipoib_add_port("ib%d", device, p);
2148 		if (!IS_ERR(dev)) {
2149 			priv = netdev_priv(dev);
2150 			list_add_tail(&priv->list, dev_list);
2151 			count++;
2152 		}
2153 	}
2154 
2155 	if (!count) {
2156 		kfree(dev_list);
2157 		return;
2158 	}
2159 
2160 	ib_set_client_data(device, &ipoib_client, dev_list);
2161 }
2162 
2163 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2164 {
2165 	struct ipoib_dev_priv *priv, *tmp;
2166 	struct list_head *dev_list = client_data;
2167 
2168 	if (!dev_list)
2169 		return;
2170 
2171 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
2172 		ib_unregister_event_handler(&priv->event_handler);
2173 		flush_workqueue(ipoib_workqueue);
2174 
2175 		/* mark interface in the middle of destruction */
2176 		set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2177 
2178 		rtnl_lock();
2179 		dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2180 		rtnl_unlock();
2181 
2182 		/* Stop GC */
2183 		set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2184 		cancel_delayed_work(&priv->neigh_reap_task);
2185 		flush_workqueue(priv->wq);
2186 
2187 		unregister_netdev(priv->dev);
2188 		free_netdev(priv->dev);
2189 	}
2190 
2191 	kfree(dev_list);
2192 }
2193 
2194 static int __init ipoib_init_module(void)
2195 {
2196 	int ret;
2197 
2198 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2199 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2200 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2201 
2202 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2203 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2204 	ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2205 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2206 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2207 #endif
2208 
2209 	/*
2210 	 * When copying small received packets, we only copy from the
2211 	 * linear data part of the SKB, so we rely on this condition.
2212 	 */
2213 	BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2214 
2215 	ret = ipoib_register_debugfs();
2216 	if (ret)
2217 		return ret;
2218 
2219 	/*
2220 	 * We create a global workqueue here that is used for all flush
2221 	 * operations.  However, if you attempt to flush a workqueue
2222 	 * from a task on that same workqueue, it deadlocks the system.
2223 	 * We want to be able to flush the tasks associated with a
2224 	 * specific net device, so we also create a workqueue for each
2225 	 * netdevice.  We queue up the tasks for that device only on
2226 	 * its private workqueue, and we only queue up flush events
2227 	 * on our global flush workqueue.  This avoids the deadlocks.
2228 	 */
2229 	ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2230 						  WQ_MEM_RECLAIM);
2231 	if (!ipoib_workqueue) {
2232 		ret = -ENOMEM;
2233 		goto err_fs;
2234 	}
2235 
2236 	ib_sa_register_client(&ipoib_sa_client);
2237 
2238 	ret = ib_register_client(&ipoib_client);
2239 	if (ret)
2240 		goto err_sa;
2241 
2242 	ret = ipoib_netlink_init();
2243 	if (ret)
2244 		goto err_client;
2245 
2246 	return 0;
2247 
2248 err_client:
2249 	ib_unregister_client(&ipoib_client);
2250 
2251 err_sa:
2252 	ib_sa_unregister_client(&ipoib_sa_client);
2253 	destroy_workqueue(ipoib_workqueue);
2254 
2255 err_fs:
2256 	ipoib_unregister_debugfs();
2257 
2258 	return ret;
2259 }
2260 
2261 static void __exit ipoib_cleanup_module(void)
2262 {
2263 	ipoib_netlink_fini();
2264 	ib_unregister_client(&ipoib_client);
2265 	ib_sa_unregister_client(&ipoib_sa_client);
2266 	ipoib_unregister_debugfs();
2267 	destroy_workqueue(ipoib_workqueue);
2268 }
2269 
2270 module_init(ipoib_init_module);
2271 module_exit(ipoib_cleanup_module);
2272