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 
52 #define DRV_VERSION "1.0.0"
53 
54 const char ipoib_driver_version[] = DRV_VERSION;
55 
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
58 MODULE_LICENSE("Dual BSD/GPL");
59 MODULE_VERSION(DRV_VERSION);
60 
61 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
62 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
63 
64 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
65 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
66 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
67 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
68 
69 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
70 int ipoib_debug_level;
71 
72 module_param_named(debug_level, ipoib_debug_level, int, 0644);
73 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
74 #endif
75 
76 struct ipoib_path_iter {
77 	struct net_device *dev;
78 	struct ipoib_path  path;
79 };
80 
81 static const u8 ipv4_bcast_addr[] = {
82 	0x00, 0xff, 0xff, 0xff,
83 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
84 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
85 };
86 
87 struct workqueue_struct *ipoib_workqueue;
88 
89 struct ib_sa_client ipoib_sa_client;
90 
91 static void ipoib_add_one(struct ib_device *device);
92 static void ipoib_remove_one(struct ib_device *device);
93 static void ipoib_neigh_reclaim(struct rcu_head *rp);
94 
95 static struct ib_client ipoib_client = {
96 	.name   = "ipoib",
97 	.add    = ipoib_add_one,
98 	.remove = ipoib_remove_one
99 };
100 
101 int ipoib_open(struct net_device *dev)
102 {
103 	struct ipoib_dev_priv *priv = netdev_priv(dev);
104 
105 	ipoib_dbg(priv, "bringing up interface\n");
106 
107 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
108 
109 	if (ipoib_pkey_dev_delay_open(dev))
110 		return 0;
111 
112 	if (ipoib_ib_dev_open(dev))
113 		goto err_disable;
114 
115 	if (ipoib_ib_dev_up(dev))
116 		goto err_stop;
117 
118 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
119 		struct ipoib_dev_priv *cpriv;
120 
121 		/* Bring up any child interfaces too */
122 		down_read(&priv->vlan_rwsem);
123 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
124 			int flags;
125 
126 			flags = cpriv->dev->flags;
127 			if (flags & IFF_UP)
128 				continue;
129 
130 			dev_change_flags(cpriv->dev, flags | IFF_UP);
131 		}
132 		up_read(&priv->vlan_rwsem);
133 	}
134 
135 	netif_start_queue(dev);
136 
137 	return 0;
138 
139 err_stop:
140 	ipoib_ib_dev_stop(dev, 1);
141 
142 err_disable:
143 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
144 
145 	return -EINVAL;
146 }
147 
148 static int ipoib_stop(struct net_device *dev)
149 {
150 	struct ipoib_dev_priv *priv = netdev_priv(dev);
151 
152 	ipoib_dbg(priv, "stopping interface\n");
153 
154 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
155 
156 	netif_stop_queue(dev);
157 
158 	ipoib_ib_dev_down(dev, 1);
159 	ipoib_ib_dev_stop(dev, 0);
160 
161 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
162 		struct ipoib_dev_priv *cpriv;
163 
164 		/* Bring down any child interfaces too */
165 		down_read(&priv->vlan_rwsem);
166 		list_for_each_entry(cpriv, &priv->child_intfs, list) {
167 			int flags;
168 
169 			flags = cpriv->dev->flags;
170 			if (!(flags & IFF_UP))
171 				continue;
172 
173 			dev_change_flags(cpriv->dev, flags & ~IFF_UP);
174 		}
175 		up_read(&priv->vlan_rwsem);
176 	}
177 
178 	return 0;
179 }
180 
181 static void ipoib_uninit(struct net_device *dev)
182 {
183 	ipoib_dev_cleanup(dev);
184 }
185 
186 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
187 {
188 	struct ipoib_dev_priv *priv = netdev_priv(dev);
189 
190 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
191 		features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
192 
193 	return features;
194 }
195 
196 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
197 {
198 	struct ipoib_dev_priv *priv = netdev_priv(dev);
199 
200 	/* dev->mtu > 2K ==> connected mode */
201 	if (ipoib_cm_admin_enabled(dev)) {
202 		if (new_mtu > ipoib_cm_max_mtu(dev))
203 			return -EINVAL;
204 
205 		if (new_mtu > priv->mcast_mtu)
206 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
207 				   priv->mcast_mtu);
208 
209 		dev->mtu = new_mtu;
210 		return 0;
211 	}
212 
213 	if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
214 		return -EINVAL;
215 
216 	priv->admin_mtu = new_mtu;
217 
218 	dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
219 
220 	return 0;
221 }
222 
223 int ipoib_set_mode(struct net_device *dev, const char *buf)
224 {
225 	struct ipoib_dev_priv *priv = netdev_priv(dev);
226 
227 	/* flush paths if we switch modes so that connections are restarted */
228 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
229 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
230 		ipoib_warn(priv, "enabling connected mode "
231 			   "will cause multicast packet drops\n");
232 		netdev_update_features(dev);
233 		rtnl_unlock();
234 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
235 
236 		ipoib_flush_paths(dev);
237 		rtnl_lock();
238 		return 0;
239 	}
240 
241 	if (!strcmp(buf, "datagram\n")) {
242 		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
243 		netdev_update_features(dev);
244 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
245 		rtnl_unlock();
246 		ipoib_flush_paths(dev);
247 		rtnl_lock();
248 		return 0;
249 	}
250 
251 	return -EINVAL;
252 }
253 
254 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
255 {
256 	struct ipoib_dev_priv *priv = netdev_priv(dev);
257 	struct rb_node *n = priv->path_tree.rb_node;
258 	struct ipoib_path *path;
259 	int ret;
260 
261 	while (n) {
262 		path = rb_entry(n, struct ipoib_path, rb_node);
263 
264 		ret = memcmp(gid, path->pathrec.dgid.raw,
265 			     sizeof (union ib_gid));
266 
267 		if (ret < 0)
268 			n = n->rb_left;
269 		else if (ret > 0)
270 			n = n->rb_right;
271 		else
272 			return path;
273 	}
274 
275 	return NULL;
276 }
277 
278 static int __path_add(struct net_device *dev, struct ipoib_path *path)
279 {
280 	struct ipoib_dev_priv *priv = netdev_priv(dev);
281 	struct rb_node **n = &priv->path_tree.rb_node;
282 	struct rb_node *pn = NULL;
283 	struct ipoib_path *tpath;
284 	int ret;
285 
286 	while (*n) {
287 		pn = *n;
288 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
289 
290 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
291 			     sizeof (union ib_gid));
292 		if (ret < 0)
293 			n = &pn->rb_left;
294 		else if (ret > 0)
295 			n = &pn->rb_right;
296 		else
297 			return -EEXIST;
298 	}
299 
300 	rb_link_node(&path->rb_node, pn, n);
301 	rb_insert_color(&path->rb_node, &priv->path_tree);
302 
303 	list_add_tail(&path->list, &priv->path_list);
304 
305 	return 0;
306 }
307 
308 static void path_free(struct net_device *dev, struct ipoib_path *path)
309 {
310 	struct sk_buff *skb;
311 
312 	while ((skb = __skb_dequeue(&path->queue)))
313 		dev_kfree_skb_irq(skb);
314 
315 	ipoib_dbg(netdev_priv(dev), "path_free\n");
316 
317 	/* remove all neigh connected to this path */
318 	ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
319 
320 	if (path->ah)
321 		ipoib_put_ah(path->ah);
322 
323 	kfree(path);
324 }
325 
326 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
327 
328 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
329 {
330 	struct ipoib_path_iter *iter;
331 
332 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
333 	if (!iter)
334 		return NULL;
335 
336 	iter->dev = dev;
337 	memset(iter->path.pathrec.dgid.raw, 0, 16);
338 
339 	if (ipoib_path_iter_next(iter)) {
340 		kfree(iter);
341 		return NULL;
342 	}
343 
344 	return iter;
345 }
346 
347 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
348 {
349 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
350 	struct rb_node *n;
351 	struct ipoib_path *path;
352 	int ret = 1;
353 
354 	spin_lock_irq(&priv->lock);
355 
356 	n = rb_first(&priv->path_tree);
357 
358 	while (n) {
359 		path = rb_entry(n, struct ipoib_path, rb_node);
360 
361 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
362 			   sizeof (union ib_gid)) < 0) {
363 			iter->path = *path;
364 			ret = 0;
365 			break;
366 		}
367 
368 		n = rb_next(n);
369 	}
370 
371 	spin_unlock_irq(&priv->lock);
372 
373 	return ret;
374 }
375 
376 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
377 			  struct ipoib_path *path)
378 {
379 	*path = iter->path;
380 }
381 
382 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
383 
384 void ipoib_mark_paths_invalid(struct net_device *dev)
385 {
386 	struct ipoib_dev_priv *priv = netdev_priv(dev);
387 	struct ipoib_path *path, *tp;
388 
389 	spin_lock_irq(&priv->lock);
390 
391 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
392 		ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
393 			be16_to_cpu(path->pathrec.dlid),
394 			path->pathrec.dgid.raw);
395 		path->valid =  0;
396 	}
397 
398 	spin_unlock_irq(&priv->lock);
399 }
400 
401 void ipoib_flush_paths(struct net_device *dev)
402 {
403 	struct ipoib_dev_priv *priv = netdev_priv(dev);
404 	struct ipoib_path *path, *tp;
405 	LIST_HEAD(remove_list);
406 	unsigned long flags;
407 
408 	netif_tx_lock_bh(dev);
409 	spin_lock_irqsave(&priv->lock, flags);
410 
411 	list_splice_init(&priv->path_list, &remove_list);
412 
413 	list_for_each_entry(path, &remove_list, list)
414 		rb_erase(&path->rb_node, &priv->path_tree);
415 
416 	list_for_each_entry_safe(path, tp, &remove_list, list) {
417 		if (path->query)
418 			ib_sa_cancel_query(path->query_id, path->query);
419 		spin_unlock_irqrestore(&priv->lock, flags);
420 		netif_tx_unlock_bh(dev);
421 		wait_for_completion(&path->done);
422 		path_free(dev, path);
423 		netif_tx_lock_bh(dev);
424 		spin_lock_irqsave(&priv->lock, flags);
425 	}
426 
427 	spin_unlock_irqrestore(&priv->lock, flags);
428 	netif_tx_unlock_bh(dev);
429 }
430 
431 static void path_rec_completion(int status,
432 				struct ib_sa_path_rec *pathrec,
433 				void *path_ptr)
434 {
435 	struct ipoib_path *path = path_ptr;
436 	struct net_device *dev = path->dev;
437 	struct ipoib_dev_priv *priv = netdev_priv(dev);
438 	struct ipoib_ah *ah = NULL;
439 	struct ipoib_ah *old_ah = NULL;
440 	struct ipoib_neigh *neigh, *tn;
441 	struct sk_buff_head skqueue;
442 	struct sk_buff *skb;
443 	unsigned long flags;
444 
445 	if (!status)
446 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
447 			  be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
448 	else
449 		ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
450 			  status, path->pathrec.dgid.raw);
451 
452 	skb_queue_head_init(&skqueue);
453 
454 	if (!status) {
455 		struct ib_ah_attr av;
456 
457 		if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
458 			ah = ipoib_create_ah(dev, priv->pd, &av);
459 	}
460 
461 	spin_lock_irqsave(&priv->lock, flags);
462 
463 	if (!IS_ERR_OR_NULL(ah)) {
464 		path->pathrec = *pathrec;
465 
466 		old_ah   = path->ah;
467 		path->ah = ah;
468 
469 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
470 			  ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
471 
472 		while ((skb = __skb_dequeue(&path->queue)))
473 			__skb_queue_tail(&skqueue, skb);
474 
475 		list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
476 			if (neigh->ah) {
477 				WARN_ON(neigh->ah != old_ah);
478 				/*
479 				 * Dropping the ah reference inside
480 				 * priv->lock is safe here, because we
481 				 * will hold one more reference from
482 				 * the original value of path->ah (ie
483 				 * old_ah).
484 				 */
485 				ipoib_put_ah(neigh->ah);
486 			}
487 			kref_get(&path->ah->ref);
488 			neigh->ah = path->ah;
489 
490 			if (ipoib_cm_enabled(dev, neigh->daddr)) {
491 				if (!ipoib_cm_get(neigh))
492 					ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
493 									       path,
494 									       neigh));
495 				if (!ipoib_cm_get(neigh)) {
496 					ipoib_neigh_free(neigh);
497 					continue;
498 				}
499 			}
500 
501 			while ((skb = __skb_dequeue(&neigh->queue)))
502 				__skb_queue_tail(&skqueue, skb);
503 		}
504 		path->valid = 1;
505 	}
506 
507 	path->query = NULL;
508 	complete(&path->done);
509 
510 	spin_unlock_irqrestore(&priv->lock, flags);
511 
512 	if (IS_ERR_OR_NULL(ah))
513 		ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
514 
515 	if (old_ah)
516 		ipoib_put_ah(old_ah);
517 
518 	while ((skb = __skb_dequeue(&skqueue))) {
519 		skb->dev = dev;
520 		if (dev_queue_xmit(skb))
521 			ipoib_warn(priv, "dev_queue_xmit failed "
522 				   "to requeue packet\n");
523 	}
524 }
525 
526 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
527 {
528 	struct ipoib_dev_priv *priv = netdev_priv(dev);
529 	struct ipoib_path *path;
530 
531 	if (!priv->broadcast)
532 		return NULL;
533 
534 	path = kzalloc(sizeof *path, GFP_ATOMIC);
535 	if (!path)
536 		return NULL;
537 
538 	path->dev = dev;
539 
540 	skb_queue_head_init(&path->queue);
541 
542 	INIT_LIST_HEAD(&path->neigh_list);
543 
544 	memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
545 	path->pathrec.sgid	    = priv->local_gid;
546 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
547 	path->pathrec.numb_path     = 1;
548 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
549 
550 	return path;
551 }
552 
553 static int path_rec_start(struct net_device *dev,
554 			  struct ipoib_path *path)
555 {
556 	struct ipoib_dev_priv *priv = netdev_priv(dev);
557 
558 	ipoib_dbg(priv, "Start path record lookup for %pI6\n",
559 		  path->pathrec.dgid.raw);
560 
561 	init_completion(&path->done);
562 
563 	path->query_id =
564 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
565 				   &path->pathrec,
566 				   IB_SA_PATH_REC_DGID		|
567 				   IB_SA_PATH_REC_SGID		|
568 				   IB_SA_PATH_REC_NUMB_PATH	|
569 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
570 				   IB_SA_PATH_REC_PKEY,
571 				   1000, GFP_ATOMIC,
572 				   path_rec_completion,
573 				   path, &path->query);
574 	if (path->query_id < 0) {
575 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
576 		path->query = NULL;
577 		complete(&path->done);
578 		return path->query_id;
579 	}
580 
581 	return 0;
582 }
583 
584 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
585 			   struct net_device *dev)
586 {
587 	struct ipoib_dev_priv *priv = netdev_priv(dev);
588 	struct ipoib_path *path;
589 	struct ipoib_neigh *neigh;
590 	unsigned long flags;
591 
592 	spin_lock_irqsave(&priv->lock, flags);
593 	neigh = ipoib_neigh_alloc(daddr, dev);
594 	if (!neigh) {
595 		spin_unlock_irqrestore(&priv->lock, flags);
596 		++dev->stats.tx_dropped;
597 		dev_kfree_skb_any(skb);
598 		return;
599 	}
600 
601 	path = __path_find(dev, daddr + 4);
602 	if (!path) {
603 		path = path_rec_create(dev, daddr + 4);
604 		if (!path)
605 			goto err_path;
606 
607 		__path_add(dev, path);
608 	}
609 
610 	list_add_tail(&neigh->list, &path->neigh_list);
611 
612 	if (path->ah) {
613 		kref_get(&path->ah->ref);
614 		neigh->ah = path->ah;
615 
616 		if (ipoib_cm_enabled(dev, neigh->daddr)) {
617 			if (!ipoib_cm_get(neigh))
618 				ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
619 			if (!ipoib_cm_get(neigh)) {
620 				ipoib_neigh_free(neigh);
621 				goto err_drop;
622 			}
623 			if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
624 				__skb_queue_tail(&neigh->queue, skb);
625 			else {
626 				ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
627 					   skb_queue_len(&neigh->queue));
628 				goto err_drop;
629 			}
630 		} else {
631 			spin_unlock_irqrestore(&priv->lock, flags);
632 			ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
633 			ipoib_neigh_put(neigh);
634 			return;
635 		}
636 	} else {
637 		neigh->ah  = NULL;
638 
639 		if (!path->query && path_rec_start(dev, path))
640 			goto err_path;
641 
642 		__skb_queue_tail(&neigh->queue, skb);
643 	}
644 
645 	spin_unlock_irqrestore(&priv->lock, flags);
646 	ipoib_neigh_put(neigh);
647 	return;
648 
649 err_path:
650 	ipoib_neigh_free(neigh);
651 err_drop:
652 	++dev->stats.tx_dropped;
653 	dev_kfree_skb_any(skb);
654 
655 	spin_unlock_irqrestore(&priv->lock, flags);
656 	ipoib_neigh_put(neigh);
657 }
658 
659 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
660 			     struct ipoib_cb *cb)
661 {
662 	struct ipoib_dev_priv *priv = netdev_priv(dev);
663 	struct ipoib_path *path;
664 	unsigned long flags;
665 
666 	spin_lock_irqsave(&priv->lock, flags);
667 
668 	path = __path_find(dev, cb->hwaddr + 4);
669 	if (!path || !path->valid) {
670 		int new_path = 0;
671 
672 		if (!path) {
673 			path = path_rec_create(dev, cb->hwaddr + 4);
674 			new_path = 1;
675 		}
676 		if (path) {
677 			__skb_queue_tail(&path->queue, skb);
678 
679 			if (!path->query && path_rec_start(dev, path)) {
680 				spin_unlock_irqrestore(&priv->lock, flags);
681 				if (new_path)
682 					path_free(dev, path);
683 				return;
684 			} else
685 				__path_add(dev, path);
686 		} else {
687 			++dev->stats.tx_dropped;
688 			dev_kfree_skb_any(skb);
689 		}
690 
691 		spin_unlock_irqrestore(&priv->lock, flags);
692 		return;
693 	}
694 
695 	if (path->ah) {
696 		ipoib_dbg(priv, "Send unicast ARP to %04x\n",
697 			  be16_to_cpu(path->pathrec.dlid));
698 
699 		spin_unlock_irqrestore(&priv->lock, flags);
700 		ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
701 		return;
702 	} else if ((path->query || !path_rec_start(dev, path)) &&
703 		   skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
704 		__skb_queue_tail(&path->queue, skb);
705 	} else {
706 		++dev->stats.tx_dropped;
707 		dev_kfree_skb_any(skb);
708 	}
709 
710 	spin_unlock_irqrestore(&priv->lock, flags);
711 }
712 
713 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
714 {
715 	struct ipoib_dev_priv *priv = netdev_priv(dev);
716 	struct ipoib_neigh *neigh;
717 	struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
718 	struct ipoib_header *header;
719 	unsigned long flags;
720 
721 	header = (struct ipoib_header *) skb->data;
722 
723 	if (unlikely(cb->hwaddr[4] == 0xff)) {
724 		/* multicast, arrange "if" according to probability */
725 		if ((header->proto != htons(ETH_P_IP)) &&
726 		    (header->proto != htons(ETH_P_IPV6)) &&
727 		    (header->proto != htons(ETH_P_ARP)) &&
728 		    (header->proto != htons(ETH_P_RARP)) &&
729 		    (header->proto != htons(ETH_P_TIPC))) {
730 			/* ethertype not supported by IPoIB */
731 			++dev->stats.tx_dropped;
732 			dev_kfree_skb_any(skb);
733 			return NETDEV_TX_OK;
734 		}
735 		/* Add in the P_Key for multicast*/
736 		cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
737 		cb->hwaddr[9] = priv->pkey & 0xff;
738 
739 		neigh = ipoib_neigh_get(dev, cb->hwaddr);
740 		if (likely(neigh))
741 			goto send_using_neigh;
742 		ipoib_mcast_send(dev, cb->hwaddr, skb);
743 		return NETDEV_TX_OK;
744 	}
745 
746 	/* unicast, arrange "switch" according to probability */
747 	switch (header->proto) {
748 	case htons(ETH_P_IP):
749 	case htons(ETH_P_IPV6):
750 	case htons(ETH_P_TIPC):
751 		neigh = ipoib_neigh_get(dev, cb->hwaddr);
752 		if (unlikely(!neigh)) {
753 			neigh_add_path(skb, cb->hwaddr, dev);
754 			return NETDEV_TX_OK;
755 		}
756 		break;
757 	case htons(ETH_P_ARP):
758 	case htons(ETH_P_RARP):
759 		/* for unicast ARP and RARP should always perform path find */
760 		unicast_arp_send(skb, dev, cb);
761 		return NETDEV_TX_OK;
762 	default:
763 		/* ethertype not supported by IPoIB */
764 		++dev->stats.tx_dropped;
765 		dev_kfree_skb_any(skb);
766 		return NETDEV_TX_OK;
767 	}
768 
769 send_using_neigh:
770 	/* note we now hold a ref to neigh */
771 	if (ipoib_cm_get(neigh)) {
772 		if (ipoib_cm_up(neigh)) {
773 			ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
774 			goto unref;
775 		}
776 	} else if (neigh->ah) {
777 		ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
778 		goto unref;
779 	}
780 
781 	if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
782 		spin_lock_irqsave(&priv->lock, flags);
783 		__skb_queue_tail(&neigh->queue, skb);
784 		spin_unlock_irqrestore(&priv->lock, flags);
785 	} else {
786 		++dev->stats.tx_dropped;
787 		dev_kfree_skb_any(skb);
788 	}
789 
790 unref:
791 	ipoib_neigh_put(neigh);
792 
793 	return NETDEV_TX_OK;
794 }
795 
796 static void ipoib_timeout(struct net_device *dev)
797 {
798 	struct ipoib_dev_priv *priv = netdev_priv(dev);
799 
800 	ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
801 		   jiffies_to_msecs(jiffies - dev->trans_start));
802 	ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
803 		   netif_queue_stopped(dev),
804 		   priv->tx_head, priv->tx_tail);
805 	/* XXX reset QP, etc. */
806 }
807 
808 static int ipoib_hard_header(struct sk_buff *skb,
809 			     struct net_device *dev,
810 			     unsigned short type,
811 			     const void *daddr, const void *saddr, unsigned len)
812 {
813 	struct ipoib_header *header;
814 	struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
815 
816 	header = (struct ipoib_header *) skb_push(skb, sizeof *header);
817 
818 	header->proto = htons(type);
819 	header->reserved = 0;
820 
821 	/*
822 	 * we don't rely on dst_entry structure,  always stuff the
823 	 * destination address into skb->cb so we can figure out where
824 	 * to send the packet later.
825 	 */
826 	memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
827 
828 	return sizeof *header;
829 }
830 
831 static void ipoib_set_mcast_list(struct net_device *dev)
832 {
833 	struct ipoib_dev_priv *priv = netdev_priv(dev);
834 
835 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
836 		ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
837 		return;
838 	}
839 
840 	queue_work(ipoib_workqueue, &priv->restart_task);
841 }
842 
843 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
844 {
845 	/*
846 	 * Use only the address parts that contributes to spreading
847 	 * The subnet prefix is not used as one can not connect to
848 	 * same remote port (GUID) using the same remote QPN via two
849 	 * different subnets.
850 	 */
851 	 /* qpn octets[1:4) & port GUID octets[12:20) */
852 	u32 *d32 = (u32 *) daddr;
853 	u32 hv;
854 
855 	hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
856 	return hv & htbl->mask;
857 }
858 
859 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
860 {
861 	struct ipoib_dev_priv *priv = netdev_priv(dev);
862 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
863 	struct ipoib_neigh_hash *htbl;
864 	struct ipoib_neigh *neigh = NULL;
865 	u32 hash_val;
866 
867 	rcu_read_lock_bh();
868 
869 	htbl = rcu_dereference_bh(ntbl->htbl);
870 
871 	if (!htbl)
872 		goto out_unlock;
873 
874 	hash_val = ipoib_addr_hash(htbl, daddr);
875 	for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
876 	     neigh != NULL;
877 	     neigh = rcu_dereference_bh(neigh->hnext)) {
878 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
879 			/* found, take one ref on behalf of the caller */
880 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
881 				/* deleted */
882 				neigh = NULL;
883 				goto out_unlock;
884 			}
885 			neigh->alive = jiffies;
886 			goto out_unlock;
887 		}
888 	}
889 
890 out_unlock:
891 	rcu_read_unlock_bh();
892 	return neigh;
893 }
894 
895 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
896 {
897 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
898 	struct ipoib_neigh_hash *htbl;
899 	unsigned long neigh_obsolete;
900 	unsigned long dt;
901 	unsigned long flags;
902 	int i;
903 
904 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
905 		return;
906 
907 	spin_lock_irqsave(&priv->lock, flags);
908 
909 	htbl = rcu_dereference_protected(ntbl->htbl,
910 					 lockdep_is_held(&priv->lock));
911 
912 	if (!htbl)
913 		goto out_unlock;
914 
915 	/* neigh is obsolete if it was idle for two GC periods */
916 	dt = 2 * arp_tbl.gc_interval;
917 	neigh_obsolete = jiffies - dt;
918 	/* handle possible race condition */
919 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
920 		goto out_unlock;
921 
922 	for (i = 0; i < htbl->size; i++) {
923 		struct ipoib_neigh *neigh;
924 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
925 
926 		while ((neigh = rcu_dereference_protected(*np,
927 							  lockdep_is_held(&priv->lock))) != NULL) {
928 			/* was the neigh idle for two GC periods */
929 			if (time_after(neigh_obsolete, neigh->alive)) {
930 				rcu_assign_pointer(*np,
931 						   rcu_dereference_protected(neigh->hnext,
932 									     lockdep_is_held(&priv->lock)));
933 				/* remove from path/mc list */
934 				list_del(&neigh->list);
935 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
936 			} else {
937 				np = &neigh->hnext;
938 			}
939 
940 		}
941 	}
942 
943 out_unlock:
944 	spin_unlock_irqrestore(&priv->lock, flags);
945 }
946 
947 static void ipoib_reap_neigh(struct work_struct *work)
948 {
949 	struct ipoib_dev_priv *priv =
950 		container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
951 
952 	__ipoib_reap_neigh(priv);
953 
954 	if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
955 		queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
956 				   arp_tbl.gc_interval);
957 }
958 
959 
960 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
961 				      struct net_device *dev)
962 {
963 	struct ipoib_neigh *neigh;
964 
965 	neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
966 	if (!neigh)
967 		return NULL;
968 
969 	neigh->dev = dev;
970 	memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
971 	skb_queue_head_init(&neigh->queue);
972 	INIT_LIST_HEAD(&neigh->list);
973 	ipoib_cm_set(neigh, NULL);
974 	/* one ref on behalf of the caller */
975 	atomic_set(&neigh->refcnt, 1);
976 
977 	return neigh;
978 }
979 
980 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
981 				      struct net_device *dev)
982 {
983 	struct ipoib_dev_priv *priv = netdev_priv(dev);
984 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
985 	struct ipoib_neigh_hash *htbl;
986 	struct ipoib_neigh *neigh;
987 	u32 hash_val;
988 
989 	htbl = rcu_dereference_protected(ntbl->htbl,
990 					 lockdep_is_held(&priv->lock));
991 	if (!htbl) {
992 		neigh = NULL;
993 		goto out_unlock;
994 	}
995 
996 	/* need to add a new neigh, but maybe some other thread succeeded?
997 	 * recalc hash, maybe hash resize took place so we do a search
998 	 */
999 	hash_val = ipoib_addr_hash(htbl, daddr);
1000 	for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1001 					       lockdep_is_held(&priv->lock));
1002 	     neigh != NULL;
1003 	     neigh = rcu_dereference_protected(neigh->hnext,
1004 					       lockdep_is_held(&priv->lock))) {
1005 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1006 			/* found, take one ref on behalf of the caller */
1007 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1008 				/* deleted */
1009 				neigh = NULL;
1010 				break;
1011 			}
1012 			neigh->alive = jiffies;
1013 			goto out_unlock;
1014 		}
1015 	}
1016 
1017 	neigh = ipoib_neigh_ctor(daddr, dev);
1018 	if (!neigh)
1019 		goto out_unlock;
1020 
1021 	/* one ref on behalf of the hash table */
1022 	atomic_inc(&neigh->refcnt);
1023 	neigh->alive = jiffies;
1024 	/* put in hash */
1025 	rcu_assign_pointer(neigh->hnext,
1026 			   rcu_dereference_protected(htbl->buckets[hash_val],
1027 						     lockdep_is_held(&priv->lock)));
1028 	rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1029 	atomic_inc(&ntbl->entries);
1030 
1031 out_unlock:
1032 
1033 	return neigh;
1034 }
1035 
1036 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1037 {
1038 	/* neigh reference count was dropprd to zero */
1039 	struct net_device *dev = neigh->dev;
1040 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1041 	struct sk_buff *skb;
1042 	if (neigh->ah)
1043 		ipoib_put_ah(neigh->ah);
1044 	while ((skb = __skb_dequeue(&neigh->queue))) {
1045 		++dev->stats.tx_dropped;
1046 		dev_kfree_skb_any(skb);
1047 	}
1048 	if (ipoib_cm_get(neigh))
1049 		ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1050 	ipoib_dbg(netdev_priv(dev),
1051 		  "neigh free for %06x %pI6\n",
1052 		  IPOIB_QPN(neigh->daddr),
1053 		  neigh->daddr + 4);
1054 	kfree(neigh);
1055 	if (atomic_dec_and_test(&priv->ntbl.entries)) {
1056 		if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1057 			complete(&priv->ntbl.flushed);
1058 	}
1059 }
1060 
1061 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1062 {
1063 	/* Called as a result of removal from hash table */
1064 	struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1065 	/* note TX context may hold another ref */
1066 	ipoib_neigh_put(neigh);
1067 }
1068 
1069 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1070 {
1071 	struct net_device *dev = neigh->dev;
1072 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1073 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1074 	struct ipoib_neigh_hash *htbl;
1075 	struct ipoib_neigh __rcu **np;
1076 	struct ipoib_neigh *n;
1077 	u32 hash_val;
1078 
1079 	htbl = rcu_dereference_protected(ntbl->htbl,
1080 					lockdep_is_held(&priv->lock));
1081 	if (!htbl)
1082 		return;
1083 
1084 	hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1085 	np = &htbl->buckets[hash_val];
1086 	for (n = rcu_dereference_protected(*np,
1087 					    lockdep_is_held(&priv->lock));
1088 	     n != NULL;
1089 	     n = rcu_dereference_protected(*np,
1090 					lockdep_is_held(&priv->lock))) {
1091 		if (n == neigh) {
1092 			/* found */
1093 			rcu_assign_pointer(*np,
1094 					   rcu_dereference_protected(neigh->hnext,
1095 								     lockdep_is_held(&priv->lock)));
1096 			/* remove from parent list */
1097 			list_del(&neigh->list);
1098 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1099 			return;
1100 		} else {
1101 			np = &n->hnext;
1102 		}
1103 	}
1104 }
1105 
1106 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1107 {
1108 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1109 	struct ipoib_neigh_hash *htbl;
1110 	struct ipoib_neigh **buckets;
1111 	u32 size;
1112 
1113 	clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1114 	ntbl->htbl = NULL;
1115 	htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1116 	if (!htbl)
1117 		return -ENOMEM;
1118 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1119 	size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1120 	buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1121 	if (!buckets) {
1122 		kfree(htbl);
1123 		return -ENOMEM;
1124 	}
1125 	htbl->size = size;
1126 	htbl->mask = (size - 1);
1127 	htbl->buckets = buckets;
1128 	ntbl->htbl = htbl;
1129 	htbl->ntbl = ntbl;
1130 	atomic_set(&ntbl->entries, 0);
1131 
1132 	/* start garbage collection */
1133 	clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1134 	queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
1135 			   arp_tbl.gc_interval);
1136 
1137 	return 0;
1138 }
1139 
1140 static void neigh_hash_free_rcu(struct rcu_head *head)
1141 {
1142 	struct ipoib_neigh_hash *htbl = container_of(head,
1143 						    struct ipoib_neigh_hash,
1144 						    rcu);
1145 	struct ipoib_neigh __rcu **buckets = htbl->buckets;
1146 	struct ipoib_neigh_table *ntbl = htbl->ntbl;
1147 
1148 	kfree(buckets);
1149 	kfree(htbl);
1150 	complete(&ntbl->deleted);
1151 }
1152 
1153 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1154 {
1155 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1156 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1157 	struct ipoib_neigh_hash *htbl;
1158 	unsigned long flags;
1159 	int i;
1160 
1161 	/* remove all neigh connected to a given path or mcast */
1162 	spin_lock_irqsave(&priv->lock, flags);
1163 
1164 	htbl = rcu_dereference_protected(ntbl->htbl,
1165 					 lockdep_is_held(&priv->lock));
1166 
1167 	if (!htbl)
1168 		goto out_unlock;
1169 
1170 	for (i = 0; i < htbl->size; i++) {
1171 		struct ipoib_neigh *neigh;
1172 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1173 
1174 		while ((neigh = rcu_dereference_protected(*np,
1175 							  lockdep_is_held(&priv->lock))) != NULL) {
1176 			/* delete neighs belong to this parent */
1177 			if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1178 				rcu_assign_pointer(*np,
1179 						   rcu_dereference_protected(neigh->hnext,
1180 									     lockdep_is_held(&priv->lock)));
1181 				/* remove from parent list */
1182 				list_del(&neigh->list);
1183 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1184 			} else {
1185 				np = &neigh->hnext;
1186 			}
1187 
1188 		}
1189 	}
1190 out_unlock:
1191 	spin_unlock_irqrestore(&priv->lock, flags);
1192 }
1193 
1194 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1195 {
1196 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1197 	struct ipoib_neigh_hash *htbl;
1198 	unsigned long flags;
1199 	int i, wait_flushed = 0;
1200 
1201 	init_completion(&priv->ntbl.flushed);
1202 
1203 	spin_lock_irqsave(&priv->lock, flags);
1204 
1205 	htbl = rcu_dereference_protected(ntbl->htbl,
1206 					lockdep_is_held(&priv->lock));
1207 	if (!htbl)
1208 		goto out_unlock;
1209 
1210 	wait_flushed = atomic_read(&priv->ntbl.entries);
1211 	if (!wait_flushed)
1212 		goto free_htbl;
1213 
1214 	for (i = 0; i < htbl->size; i++) {
1215 		struct ipoib_neigh *neigh;
1216 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1217 
1218 		while ((neigh = rcu_dereference_protected(*np,
1219 				       lockdep_is_held(&priv->lock))) != NULL) {
1220 			rcu_assign_pointer(*np,
1221 					   rcu_dereference_protected(neigh->hnext,
1222 								     lockdep_is_held(&priv->lock)));
1223 			/* remove from path/mc list */
1224 			list_del(&neigh->list);
1225 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1226 		}
1227 	}
1228 
1229 free_htbl:
1230 	rcu_assign_pointer(ntbl->htbl, NULL);
1231 	call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1232 
1233 out_unlock:
1234 	spin_unlock_irqrestore(&priv->lock, flags);
1235 	if (wait_flushed)
1236 		wait_for_completion(&priv->ntbl.flushed);
1237 }
1238 
1239 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1240 {
1241 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1242 	int stopped;
1243 
1244 	ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1245 	init_completion(&priv->ntbl.deleted);
1246 	set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1247 
1248 	/* Stop GC if called at init fail need to cancel work */
1249 	stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1250 	if (!stopped)
1251 		cancel_delayed_work(&priv->neigh_reap_task);
1252 
1253 	ipoib_flush_neighs(priv);
1254 
1255 	wait_for_completion(&priv->ntbl.deleted);
1256 }
1257 
1258 
1259 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1260 {
1261 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1262 
1263 	if (ipoib_neigh_hash_init(priv) < 0)
1264 		goto out;
1265 	/* Allocate RX/TX "rings" to hold queued skbs */
1266 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1267 				GFP_KERNEL);
1268 	if (!priv->rx_ring) {
1269 		printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1270 		       ca->name, ipoib_recvq_size);
1271 		goto out_neigh_hash_cleanup;
1272 	}
1273 
1274 	priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1275 	if (!priv->tx_ring) {
1276 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1277 		       ca->name, ipoib_sendq_size);
1278 		goto out_rx_ring_cleanup;
1279 	}
1280 
1281 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1282 
1283 	if (ipoib_ib_dev_init(dev, ca, port))
1284 		goto out_tx_ring_cleanup;
1285 
1286 	return 0;
1287 
1288 out_tx_ring_cleanup:
1289 	vfree(priv->tx_ring);
1290 
1291 out_rx_ring_cleanup:
1292 	kfree(priv->rx_ring);
1293 
1294 out_neigh_hash_cleanup:
1295 	ipoib_neigh_hash_uninit(dev);
1296 out:
1297 	return -ENOMEM;
1298 }
1299 
1300 void ipoib_dev_cleanup(struct net_device *dev)
1301 {
1302 	struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1303 	LIST_HEAD(head);
1304 
1305 	ASSERT_RTNL();
1306 
1307 	ipoib_delete_debug_files(dev);
1308 
1309 	/* Delete any child interfaces first */
1310 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1311 		/* Stop GC on child */
1312 		set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1313 		cancel_delayed_work(&cpriv->neigh_reap_task);
1314 		unregister_netdevice_queue(cpriv->dev, &head);
1315 	}
1316 	unregister_netdevice_many(&head);
1317 
1318 	ipoib_ib_dev_cleanup(dev);
1319 
1320 	kfree(priv->rx_ring);
1321 	vfree(priv->tx_ring);
1322 
1323 	priv->rx_ring = NULL;
1324 	priv->tx_ring = NULL;
1325 
1326 	ipoib_neigh_hash_uninit(dev);
1327 }
1328 
1329 static const struct header_ops ipoib_header_ops = {
1330 	.create	= ipoib_hard_header,
1331 };
1332 
1333 static const struct net_device_ops ipoib_netdev_ops = {
1334 	.ndo_uninit		 = ipoib_uninit,
1335 	.ndo_open		 = ipoib_open,
1336 	.ndo_stop		 = ipoib_stop,
1337 	.ndo_change_mtu		 = ipoib_change_mtu,
1338 	.ndo_fix_features	 = ipoib_fix_features,
1339 	.ndo_start_xmit	 	 = ipoib_start_xmit,
1340 	.ndo_tx_timeout		 = ipoib_timeout,
1341 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
1342 };
1343 
1344 void ipoib_setup(struct net_device *dev)
1345 {
1346 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1347 
1348 	dev->netdev_ops		 = &ipoib_netdev_ops;
1349 	dev->header_ops		 = &ipoib_header_ops;
1350 
1351 	ipoib_set_ethtool_ops(dev);
1352 
1353 	netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1354 
1355 	dev->watchdog_timeo	 = HZ;
1356 
1357 	dev->flags		|= IFF_BROADCAST | IFF_MULTICAST;
1358 
1359 	dev->hard_header_len	 = IPOIB_ENCAP_LEN;
1360 	dev->addr_len		 = INFINIBAND_ALEN;
1361 	dev->type		 = ARPHRD_INFINIBAND;
1362 	dev->tx_queue_len	 = ipoib_sendq_size * 2;
1363 	dev->features		 = (NETIF_F_VLAN_CHALLENGED	|
1364 				    NETIF_F_HIGHDMA);
1365 	dev->priv_flags		&= ~IFF_XMIT_DST_RELEASE;
1366 
1367 	memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1368 
1369 	netif_carrier_off(dev);
1370 
1371 	priv->dev = dev;
1372 
1373 	spin_lock_init(&priv->lock);
1374 
1375 	init_rwsem(&priv->vlan_rwsem);
1376 
1377 	INIT_LIST_HEAD(&priv->path_list);
1378 	INIT_LIST_HEAD(&priv->child_intfs);
1379 	INIT_LIST_HEAD(&priv->dead_ahs);
1380 	INIT_LIST_HEAD(&priv->multicast_list);
1381 
1382 	INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1383 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1384 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1385 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1386 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1387 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1388 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1389 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1390 	INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1391 }
1392 
1393 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1394 {
1395 	struct net_device *dev;
1396 
1397 	dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1398 			   ipoib_setup);
1399 	if (!dev)
1400 		return NULL;
1401 
1402 	return netdev_priv(dev);
1403 }
1404 
1405 static ssize_t show_pkey(struct device *dev,
1406 			 struct device_attribute *attr, char *buf)
1407 {
1408 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1409 
1410 	return sprintf(buf, "0x%04x\n", priv->pkey);
1411 }
1412 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1413 
1414 static ssize_t show_umcast(struct device *dev,
1415 			   struct device_attribute *attr, char *buf)
1416 {
1417 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1418 
1419 	return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1420 }
1421 
1422 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1423 {
1424 	struct ipoib_dev_priv *priv = netdev_priv(ndev);
1425 
1426 	if (umcast_val > 0) {
1427 		set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1428 		ipoib_warn(priv, "ignoring multicast groups joined directly "
1429 				"by userspace\n");
1430 	} else
1431 		clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1432 }
1433 
1434 static ssize_t set_umcast(struct device *dev,
1435 			  struct device_attribute *attr,
1436 			  const char *buf, size_t count)
1437 {
1438 	unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1439 
1440 	ipoib_set_umcast(to_net_dev(dev), umcast_val);
1441 
1442 	return count;
1443 }
1444 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1445 
1446 int ipoib_add_umcast_attr(struct net_device *dev)
1447 {
1448 	return device_create_file(&dev->dev, &dev_attr_umcast);
1449 }
1450 
1451 static ssize_t create_child(struct device *dev,
1452 			    struct device_attribute *attr,
1453 			    const char *buf, size_t count)
1454 {
1455 	int pkey;
1456 	int ret;
1457 
1458 	if (sscanf(buf, "%i", &pkey) != 1)
1459 		return -EINVAL;
1460 
1461 	if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1462 		return -EINVAL;
1463 
1464 	/*
1465 	 * Set the full membership bit, so that we join the right
1466 	 * broadcast group, etc.
1467 	 */
1468 	pkey |= 0x8000;
1469 
1470 	ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1471 
1472 	return ret ? ret : count;
1473 }
1474 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1475 
1476 static ssize_t delete_child(struct device *dev,
1477 			    struct device_attribute *attr,
1478 			    const char *buf, size_t count)
1479 {
1480 	int pkey;
1481 	int ret;
1482 
1483 	if (sscanf(buf, "%i", &pkey) != 1)
1484 		return -EINVAL;
1485 
1486 	if (pkey < 0 || pkey > 0xffff)
1487 		return -EINVAL;
1488 
1489 	ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1490 
1491 	return ret ? ret : count;
1492 
1493 }
1494 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1495 
1496 int ipoib_add_pkey_attr(struct net_device *dev)
1497 {
1498 	return device_create_file(&dev->dev, &dev_attr_pkey);
1499 }
1500 
1501 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1502 {
1503 	struct ib_device_attr *device_attr;
1504 	int result = -ENOMEM;
1505 
1506 	device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1507 	if (!device_attr) {
1508 		printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1509 		       hca->name, sizeof *device_attr);
1510 		return result;
1511 	}
1512 
1513 	result = ib_query_device(hca, device_attr);
1514 	if (result) {
1515 		printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1516 		       hca->name, result);
1517 		kfree(device_attr);
1518 		return result;
1519 	}
1520 	priv->hca_caps = device_attr->device_cap_flags;
1521 
1522 	kfree(device_attr);
1523 
1524 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1525 		priv->dev->hw_features = NETIF_F_SG |
1526 			NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1527 
1528 		if (priv->hca_caps & IB_DEVICE_UD_TSO)
1529 			priv->dev->hw_features |= NETIF_F_TSO;
1530 
1531 		priv->dev->features |= priv->dev->hw_features;
1532 	}
1533 
1534 	return 0;
1535 }
1536 
1537 static struct net_device *ipoib_add_port(const char *format,
1538 					 struct ib_device *hca, u8 port)
1539 {
1540 	struct ipoib_dev_priv *priv;
1541 	struct ib_port_attr attr;
1542 	int result = -ENOMEM;
1543 
1544 	priv = ipoib_intf_alloc(format);
1545 	if (!priv)
1546 		goto alloc_mem_failed;
1547 
1548 	SET_NETDEV_DEV(priv->dev, hca->dma_device);
1549 	priv->dev->dev_id = port - 1;
1550 
1551 	if (!ib_query_port(hca, port, &attr))
1552 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1553 	else {
1554 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1555 		       hca->name, port);
1556 		goto device_init_failed;
1557 	}
1558 
1559 	/* MTU will be reset when mcast join happens */
1560 	priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1561 	priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1562 
1563 	priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1564 
1565 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
1566 	if (result) {
1567 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1568 		       hca->name, port, result);
1569 		goto device_init_failed;
1570 	}
1571 
1572 	if (ipoib_set_dev_features(priv, hca))
1573 		goto device_init_failed;
1574 
1575 	/*
1576 	 * Set the full membership bit, so that we join the right
1577 	 * broadcast group, etc.
1578 	 */
1579 	priv->pkey |= 0x8000;
1580 
1581 	priv->dev->broadcast[8] = priv->pkey >> 8;
1582 	priv->dev->broadcast[9] = priv->pkey & 0xff;
1583 
1584 	result = ib_query_gid(hca, port, 0, &priv->local_gid);
1585 	if (result) {
1586 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1587 		       hca->name, port, result);
1588 		goto device_init_failed;
1589 	} else
1590 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1591 
1592 	result = ipoib_dev_init(priv->dev, hca, port);
1593 	if (result < 0) {
1594 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1595 		       hca->name, port, result);
1596 		goto device_init_failed;
1597 	}
1598 
1599 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
1600 			      priv->ca, ipoib_event);
1601 	result = ib_register_event_handler(&priv->event_handler);
1602 	if (result < 0) {
1603 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1604 		       "port %d (ret = %d)\n",
1605 		       hca->name, port, result);
1606 		goto event_failed;
1607 	}
1608 
1609 	result = register_netdev(priv->dev);
1610 	if (result) {
1611 		printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1612 		       hca->name, port, result);
1613 		goto register_failed;
1614 	}
1615 
1616 	ipoib_create_debug_files(priv->dev);
1617 
1618 	if (ipoib_cm_add_mode_attr(priv->dev))
1619 		goto sysfs_failed;
1620 	if (ipoib_add_pkey_attr(priv->dev))
1621 		goto sysfs_failed;
1622 	if (ipoib_add_umcast_attr(priv->dev))
1623 		goto sysfs_failed;
1624 	if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1625 		goto sysfs_failed;
1626 	if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1627 		goto sysfs_failed;
1628 
1629 	return priv->dev;
1630 
1631 sysfs_failed:
1632 	ipoib_delete_debug_files(priv->dev);
1633 	unregister_netdev(priv->dev);
1634 
1635 register_failed:
1636 	ib_unregister_event_handler(&priv->event_handler);
1637 	/* Stop GC if started before flush */
1638 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1639 	cancel_delayed_work(&priv->neigh_reap_task);
1640 	flush_workqueue(ipoib_workqueue);
1641 
1642 event_failed:
1643 	ipoib_dev_cleanup(priv->dev);
1644 
1645 device_init_failed:
1646 	free_netdev(priv->dev);
1647 
1648 alloc_mem_failed:
1649 	return ERR_PTR(result);
1650 }
1651 
1652 static void ipoib_add_one(struct ib_device *device)
1653 {
1654 	struct list_head *dev_list;
1655 	struct net_device *dev;
1656 	struct ipoib_dev_priv *priv;
1657 	int s, e, p;
1658 
1659 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1660 		return;
1661 
1662 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1663 	if (!dev_list)
1664 		return;
1665 
1666 	INIT_LIST_HEAD(dev_list);
1667 
1668 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
1669 		s = 0;
1670 		e = 0;
1671 	} else {
1672 		s = 1;
1673 		e = device->phys_port_cnt;
1674 	}
1675 
1676 	for (p = s; p <= e; ++p) {
1677 		if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1678 			continue;
1679 		dev = ipoib_add_port("ib%d", device, p);
1680 		if (!IS_ERR(dev)) {
1681 			priv = netdev_priv(dev);
1682 			list_add_tail(&priv->list, dev_list);
1683 		}
1684 	}
1685 
1686 	ib_set_client_data(device, &ipoib_client, dev_list);
1687 }
1688 
1689 static void ipoib_remove_one(struct ib_device *device)
1690 {
1691 	struct ipoib_dev_priv *priv, *tmp;
1692 	struct list_head *dev_list;
1693 
1694 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1695 		return;
1696 
1697 	dev_list = ib_get_client_data(device, &ipoib_client);
1698 	if (!dev_list)
1699 		return;
1700 
1701 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
1702 		ib_unregister_event_handler(&priv->event_handler);
1703 
1704 		rtnl_lock();
1705 		dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1706 		rtnl_unlock();
1707 
1708 		/* Stop GC */
1709 		set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1710 		cancel_delayed_work(&priv->neigh_reap_task);
1711 		flush_workqueue(ipoib_workqueue);
1712 
1713 		unregister_netdev(priv->dev);
1714 		free_netdev(priv->dev);
1715 	}
1716 
1717 	kfree(dev_list);
1718 }
1719 
1720 static int __init ipoib_init_module(void)
1721 {
1722 	int ret;
1723 
1724 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1725 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1726 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1727 
1728 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1729 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1730 	ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1731 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1732 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1733 #endif
1734 
1735 	/*
1736 	 * When copying small received packets, we only copy from the
1737 	 * linear data part of the SKB, so we rely on this condition.
1738 	 */
1739 	BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1740 
1741 	ret = ipoib_register_debugfs();
1742 	if (ret)
1743 		return ret;
1744 
1745 	/*
1746 	 * We create our own workqueue mainly because we want to be
1747 	 * able to flush it when devices are being removed.  We can't
1748 	 * use schedule_work()/flush_scheduled_work() because both
1749 	 * unregister_netdev() and linkwatch_event take the rtnl lock,
1750 	 * so flush_scheduled_work() can deadlock during device
1751 	 * removal.
1752 	 */
1753 	ipoib_workqueue = create_singlethread_workqueue("ipoib");
1754 	if (!ipoib_workqueue) {
1755 		ret = -ENOMEM;
1756 		goto err_fs;
1757 	}
1758 
1759 	ib_sa_register_client(&ipoib_sa_client);
1760 
1761 	ret = ib_register_client(&ipoib_client);
1762 	if (ret)
1763 		goto err_sa;
1764 
1765 	ret = ipoib_netlink_init();
1766 	if (ret)
1767 		goto err_client;
1768 
1769 	return 0;
1770 
1771 err_client:
1772 	ib_unregister_client(&ipoib_client);
1773 
1774 err_sa:
1775 	ib_sa_unregister_client(&ipoib_sa_client);
1776 	destroy_workqueue(ipoib_workqueue);
1777 
1778 err_fs:
1779 	ipoib_unregister_debugfs();
1780 
1781 	return ret;
1782 }
1783 
1784 static void __exit ipoib_cleanup_module(void)
1785 {
1786 	ipoib_netlink_fini();
1787 	ib_unregister_client(&ipoib_client);
1788 	ib_sa_unregister_client(&ipoib_sa_client);
1789 	ipoib_unregister_debugfs();
1790 	destroy_workqueue(ipoib_workqueue);
1791 }
1792 
1793 module_init(ipoib_init_module);
1794 module_exit(ipoib_cleanup_module);
1795