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 		mutex_lock(&priv->vlan_mutex);
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 		mutex_unlock(&priv->vlan_mutex);
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 		mutex_lock(&priv->vlan_mutex);
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 		mutex_unlock(&priv->vlan_mutex);
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 					list_del(&neigh->list);
497 					ipoib_neigh_free(neigh);
498 					continue;
499 				}
500 			}
501 
502 			while ((skb = __skb_dequeue(&neigh->queue)))
503 				__skb_queue_tail(&skqueue, skb);
504 		}
505 		path->valid = 1;
506 	}
507 
508 	path->query = NULL;
509 	complete(&path->done);
510 
511 	spin_unlock_irqrestore(&priv->lock, flags);
512 
513 	if (IS_ERR_OR_NULL(ah))
514 		ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
515 
516 	if (old_ah)
517 		ipoib_put_ah(old_ah);
518 
519 	while ((skb = __skb_dequeue(&skqueue))) {
520 		skb->dev = dev;
521 		if (dev_queue_xmit(skb))
522 			ipoib_warn(priv, "dev_queue_xmit failed "
523 				   "to requeue packet\n");
524 	}
525 }
526 
527 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
528 {
529 	struct ipoib_dev_priv *priv = netdev_priv(dev);
530 	struct ipoib_path *path;
531 
532 	if (!priv->broadcast)
533 		return NULL;
534 
535 	path = kzalloc(sizeof *path, GFP_ATOMIC);
536 	if (!path)
537 		return NULL;
538 
539 	path->dev = dev;
540 
541 	skb_queue_head_init(&path->queue);
542 
543 	INIT_LIST_HEAD(&path->neigh_list);
544 
545 	memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
546 	path->pathrec.sgid	    = priv->local_gid;
547 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
548 	path->pathrec.numb_path     = 1;
549 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
550 
551 	return path;
552 }
553 
554 static int path_rec_start(struct net_device *dev,
555 			  struct ipoib_path *path)
556 {
557 	struct ipoib_dev_priv *priv = netdev_priv(dev);
558 
559 	ipoib_dbg(priv, "Start path record lookup for %pI6\n",
560 		  path->pathrec.dgid.raw);
561 
562 	init_completion(&path->done);
563 
564 	path->query_id =
565 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
566 				   &path->pathrec,
567 				   IB_SA_PATH_REC_DGID		|
568 				   IB_SA_PATH_REC_SGID		|
569 				   IB_SA_PATH_REC_NUMB_PATH	|
570 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
571 				   IB_SA_PATH_REC_PKEY,
572 				   1000, GFP_ATOMIC,
573 				   path_rec_completion,
574 				   path, &path->query);
575 	if (path->query_id < 0) {
576 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
577 		path->query = NULL;
578 		complete(&path->done);
579 		return path->query_id;
580 	}
581 
582 	return 0;
583 }
584 
585 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
586 			   struct net_device *dev)
587 {
588 	struct ipoib_dev_priv *priv = netdev_priv(dev);
589 	struct ipoib_path *path;
590 	struct ipoib_neigh *neigh;
591 	unsigned long flags;
592 
593 	spin_lock_irqsave(&priv->lock, flags);
594 	neigh = ipoib_neigh_alloc(daddr, dev);
595 	if (!neigh) {
596 		spin_unlock_irqrestore(&priv->lock, flags);
597 		++dev->stats.tx_dropped;
598 		dev_kfree_skb_any(skb);
599 		return;
600 	}
601 
602 	path = __path_find(dev, daddr + 4);
603 	if (!path) {
604 		path = path_rec_create(dev, daddr + 4);
605 		if (!path)
606 			goto err_path;
607 
608 		__path_add(dev, path);
609 	}
610 
611 	list_add_tail(&neigh->list, &path->neigh_list);
612 
613 	if (path->ah) {
614 		kref_get(&path->ah->ref);
615 		neigh->ah = path->ah;
616 
617 		if (ipoib_cm_enabled(dev, neigh->daddr)) {
618 			if (!ipoib_cm_get(neigh))
619 				ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
620 			if (!ipoib_cm_get(neigh)) {
621 				list_del(&neigh->list);
622 				ipoib_neigh_free(neigh);
623 				goto err_drop;
624 			}
625 			if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
626 				__skb_queue_tail(&neigh->queue, skb);
627 			else {
628 				ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
629 					   skb_queue_len(&neigh->queue));
630 				goto err_drop;
631 			}
632 		} else {
633 			spin_unlock_irqrestore(&priv->lock, flags);
634 			ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
635 			ipoib_neigh_put(neigh);
636 			return;
637 		}
638 	} else {
639 		neigh->ah  = NULL;
640 
641 		if (!path->query && path_rec_start(dev, path))
642 			goto err_list;
643 
644 		__skb_queue_tail(&neigh->queue, skb);
645 	}
646 
647 	spin_unlock_irqrestore(&priv->lock, flags);
648 	ipoib_neigh_put(neigh);
649 	return;
650 
651 err_list:
652 	list_del(&neigh->list);
653 
654 err_path:
655 	ipoib_neigh_free(neigh);
656 err_drop:
657 	++dev->stats.tx_dropped;
658 	dev_kfree_skb_any(skb);
659 
660 	spin_unlock_irqrestore(&priv->lock, flags);
661 	ipoib_neigh_put(neigh);
662 }
663 
664 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
665 			     struct ipoib_cb *cb)
666 {
667 	struct ipoib_dev_priv *priv = netdev_priv(dev);
668 	struct ipoib_path *path;
669 	unsigned long flags;
670 
671 	spin_lock_irqsave(&priv->lock, flags);
672 
673 	path = __path_find(dev, cb->hwaddr + 4);
674 	if (!path || !path->valid) {
675 		int new_path = 0;
676 
677 		if (!path) {
678 			path = path_rec_create(dev, cb->hwaddr + 4);
679 			new_path = 1;
680 		}
681 		if (path) {
682 			__skb_queue_tail(&path->queue, skb);
683 
684 			if (!path->query && path_rec_start(dev, path)) {
685 				spin_unlock_irqrestore(&priv->lock, flags);
686 				if (new_path)
687 					path_free(dev, path);
688 				return;
689 			} else
690 				__path_add(dev, path);
691 		} else {
692 			++dev->stats.tx_dropped;
693 			dev_kfree_skb_any(skb);
694 		}
695 
696 		spin_unlock_irqrestore(&priv->lock, flags);
697 		return;
698 	}
699 
700 	if (path->ah) {
701 		ipoib_dbg(priv, "Send unicast ARP to %04x\n",
702 			  be16_to_cpu(path->pathrec.dlid));
703 
704 		spin_unlock_irqrestore(&priv->lock, flags);
705 		ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
706 		return;
707 	} else if ((path->query || !path_rec_start(dev, path)) &&
708 		   skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
709 		__skb_queue_tail(&path->queue, skb);
710 	} else {
711 		++dev->stats.tx_dropped;
712 		dev_kfree_skb_any(skb);
713 	}
714 
715 	spin_unlock_irqrestore(&priv->lock, flags);
716 }
717 
718 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
719 {
720 	struct ipoib_dev_priv *priv = netdev_priv(dev);
721 	struct ipoib_neigh *neigh;
722 	struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
723 	struct ipoib_header *header;
724 	unsigned long flags;
725 
726 	header = (struct ipoib_header *) skb->data;
727 
728 	if (unlikely(cb->hwaddr[4] == 0xff)) {
729 		/* multicast, arrange "if" according to probability */
730 		if ((header->proto != htons(ETH_P_IP)) &&
731 		    (header->proto != htons(ETH_P_IPV6)) &&
732 		    (header->proto != htons(ETH_P_ARP)) &&
733 		    (header->proto != htons(ETH_P_RARP)) &&
734 		    (header->proto != htons(ETH_P_TIPC))) {
735 			/* ethertype not supported by IPoIB */
736 			++dev->stats.tx_dropped;
737 			dev_kfree_skb_any(skb);
738 			return NETDEV_TX_OK;
739 		}
740 		/* Add in the P_Key for multicast*/
741 		cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
742 		cb->hwaddr[9] = priv->pkey & 0xff;
743 
744 		neigh = ipoib_neigh_get(dev, cb->hwaddr);
745 		if (likely(neigh))
746 			goto send_using_neigh;
747 		ipoib_mcast_send(dev, cb->hwaddr, skb);
748 		return NETDEV_TX_OK;
749 	}
750 
751 	/* unicast, arrange "switch" according to probability */
752 	switch (header->proto) {
753 	case htons(ETH_P_IP):
754 	case htons(ETH_P_IPV6):
755 	case htons(ETH_P_TIPC):
756 		neigh = ipoib_neigh_get(dev, cb->hwaddr);
757 		if (unlikely(!neigh)) {
758 			neigh_add_path(skb, cb->hwaddr, dev);
759 			return NETDEV_TX_OK;
760 		}
761 		break;
762 	case htons(ETH_P_ARP):
763 	case htons(ETH_P_RARP):
764 		/* for unicast ARP and RARP should always perform path find */
765 		unicast_arp_send(skb, dev, cb);
766 		return NETDEV_TX_OK;
767 	default:
768 		/* ethertype not supported by IPoIB */
769 		++dev->stats.tx_dropped;
770 		dev_kfree_skb_any(skb);
771 		return NETDEV_TX_OK;
772 	}
773 
774 send_using_neigh:
775 	/* note we now hold a ref to neigh */
776 	if (ipoib_cm_get(neigh)) {
777 		if (ipoib_cm_up(neigh)) {
778 			ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
779 			goto unref;
780 		}
781 	} else if (neigh->ah) {
782 		ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
783 		goto unref;
784 	}
785 
786 	if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
787 		spin_lock_irqsave(&priv->lock, flags);
788 		__skb_queue_tail(&neigh->queue, skb);
789 		spin_unlock_irqrestore(&priv->lock, flags);
790 	} else {
791 		++dev->stats.tx_dropped;
792 		dev_kfree_skb_any(skb);
793 	}
794 
795 unref:
796 	ipoib_neigh_put(neigh);
797 
798 	return NETDEV_TX_OK;
799 }
800 
801 static void ipoib_timeout(struct net_device *dev)
802 {
803 	struct ipoib_dev_priv *priv = netdev_priv(dev);
804 
805 	ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
806 		   jiffies_to_msecs(jiffies - dev->trans_start));
807 	ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
808 		   netif_queue_stopped(dev),
809 		   priv->tx_head, priv->tx_tail);
810 	/* XXX reset QP, etc. */
811 }
812 
813 static int ipoib_hard_header(struct sk_buff *skb,
814 			     struct net_device *dev,
815 			     unsigned short type,
816 			     const void *daddr, const void *saddr, unsigned len)
817 {
818 	struct ipoib_header *header;
819 	struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
820 
821 	header = (struct ipoib_header *) skb_push(skb, sizeof *header);
822 
823 	header->proto = htons(type);
824 	header->reserved = 0;
825 
826 	/*
827 	 * we don't rely on dst_entry structure,  always stuff the
828 	 * destination address into skb->cb so we can figure out where
829 	 * to send the packet later.
830 	 */
831 	memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
832 
833 	return sizeof *header;
834 }
835 
836 static void ipoib_set_mcast_list(struct net_device *dev)
837 {
838 	struct ipoib_dev_priv *priv = netdev_priv(dev);
839 
840 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
841 		ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
842 		return;
843 	}
844 
845 	queue_work(ipoib_workqueue, &priv->restart_task);
846 }
847 
848 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
849 {
850 	/*
851 	 * Use only the address parts that contributes to spreading
852 	 * The subnet prefix is not used as one can not connect to
853 	 * same remote port (GUID) using the same remote QPN via two
854 	 * different subnets.
855 	 */
856 	 /* qpn octets[1:4) & port GUID octets[12:20) */
857 	u32 *d32 = (u32 *) daddr;
858 	u32 hv;
859 
860 	hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
861 	return hv & htbl->mask;
862 }
863 
864 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
865 {
866 	struct ipoib_dev_priv *priv = netdev_priv(dev);
867 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
868 	struct ipoib_neigh_hash *htbl;
869 	struct ipoib_neigh *neigh = NULL;
870 	u32 hash_val;
871 
872 	rcu_read_lock_bh();
873 
874 	htbl = rcu_dereference_bh(ntbl->htbl);
875 
876 	if (!htbl)
877 		goto out_unlock;
878 
879 	hash_val = ipoib_addr_hash(htbl, daddr);
880 	for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
881 	     neigh != NULL;
882 	     neigh = rcu_dereference_bh(neigh->hnext)) {
883 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
884 			/* found, take one ref on behalf of the caller */
885 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
886 				/* deleted */
887 				neigh = NULL;
888 				goto out_unlock;
889 			}
890 			neigh->alive = jiffies;
891 			goto out_unlock;
892 		}
893 	}
894 
895 out_unlock:
896 	rcu_read_unlock_bh();
897 	return neigh;
898 }
899 
900 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
901 {
902 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
903 	struct ipoib_neigh_hash *htbl;
904 	unsigned long neigh_obsolete;
905 	unsigned long dt;
906 	unsigned long flags;
907 	int i;
908 
909 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
910 		return;
911 
912 	spin_lock_irqsave(&priv->lock, flags);
913 
914 	htbl = rcu_dereference_protected(ntbl->htbl,
915 					 lockdep_is_held(&priv->lock));
916 
917 	if (!htbl)
918 		goto out_unlock;
919 
920 	/* neigh is obsolete if it was idle for two GC periods */
921 	dt = 2 * arp_tbl.gc_interval;
922 	neigh_obsolete = jiffies - dt;
923 	/* handle possible race condition */
924 	if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
925 		goto out_unlock;
926 
927 	for (i = 0; i < htbl->size; i++) {
928 		struct ipoib_neigh *neigh;
929 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
930 
931 		while ((neigh = rcu_dereference_protected(*np,
932 							  lockdep_is_held(&priv->lock))) != NULL) {
933 			/* was the neigh idle for two GC periods */
934 			if (time_after(neigh_obsolete, neigh->alive)) {
935 				rcu_assign_pointer(*np,
936 						   rcu_dereference_protected(neigh->hnext,
937 									     lockdep_is_held(&priv->lock)));
938 				/* remove from path/mc list */
939 				list_del(&neigh->list);
940 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
941 			} else {
942 				np = &neigh->hnext;
943 			}
944 
945 		}
946 	}
947 
948 out_unlock:
949 	spin_unlock_irqrestore(&priv->lock, flags);
950 }
951 
952 static void ipoib_reap_neigh(struct work_struct *work)
953 {
954 	struct ipoib_dev_priv *priv =
955 		container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
956 
957 	__ipoib_reap_neigh(priv);
958 
959 	if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
960 		queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
961 				   arp_tbl.gc_interval);
962 }
963 
964 
965 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
966 				      struct net_device *dev)
967 {
968 	struct ipoib_neigh *neigh;
969 
970 	neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
971 	if (!neigh)
972 		return NULL;
973 
974 	neigh->dev = dev;
975 	memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
976 	skb_queue_head_init(&neigh->queue);
977 	INIT_LIST_HEAD(&neigh->list);
978 	ipoib_cm_set(neigh, NULL);
979 	/* one ref on behalf of the caller */
980 	atomic_set(&neigh->refcnt, 1);
981 
982 	return neigh;
983 }
984 
985 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
986 				      struct net_device *dev)
987 {
988 	struct ipoib_dev_priv *priv = netdev_priv(dev);
989 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
990 	struct ipoib_neigh_hash *htbl;
991 	struct ipoib_neigh *neigh;
992 	u32 hash_val;
993 
994 	htbl = rcu_dereference_protected(ntbl->htbl,
995 					 lockdep_is_held(&priv->lock));
996 	if (!htbl) {
997 		neigh = NULL;
998 		goto out_unlock;
999 	}
1000 
1001 	/* need to add a new neigh, but maybe some other thread succeeded?
1002 	 * recalc hash, maybe hash resize took place so we do a search
1003 	 */
1004 	hash_val = ipoib_addr_hash(htbl, daddr);
1005 	for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1006 					       lockdep_is_held(&priv->lock));
1007 	     neigh != NULL;
1008 	     neigh = rcu_dereference_protected(neigh->hnext,
1009 					       lockdep_is_held(&priv->lock))) {
1010 		if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1011 			/* found, take one ref on behalf of the caller */
1012 			if (!atomic_inc_not_zero(&neigh->refcnt)) {
1013 				/* deleted */
1014 				neigh = NULL;
1015 				break;
1016 			}
1017 			neigh->alive = jiffies;
1018 			goto out_unlock;
1019 		}
1020 	}
1021 
1022 	neigh = ipoib_neigh_ctor(daddr, dev);
1023 	if (!neigh)
1024 		goto out_unlock;
1025 
1026 	/* one ref on behalf of the hash table */
1027 	atomic_inc(&neigh->refcnt);
1028 	neigh->alive = jiffies;
1029 	/* put in hash */
1030 	rcu_assign_pointer(neigh->hnext,
1031 			   rcu_dereference_protected(htbl->buckets[hash_val],
1032 						     lockdep_is_held(&priv->lock)));
1033 	rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1034 	atomic_inc(&ntbl->entries);
1035 
1036 out_unlock:
1037 
1038 	return neigh;
1039 }
1040 
1041 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1042 {
1043 	/* neigh reference count was dropprd to zero */
1044 	struct net_device *dev = neigh->dev;
1045 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1046 	struct sk_buff *skb;
1047 	if (neigh->ah)
1048 		ipoib_put_ah(neigh->ah);
1049 	while ((skb = __skb_dequeue(&neigh->queue))) {
1050 		++dev->stats.tx_dropped;
1051 		dev_kfree_skb_any(skb);
1052 	}
1053 	if (ipoib_cm_get(neigh))
1054 		ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1055 	ipoib_dbg(netdev_priv(dev),
1056 		  "neigh free for %06x %pI6\n",
1057 		  IPOIB_QPN(neigh->daddr),
1058 		  neigh->daddr + 4);
1059 	kfree(neigh);
1060 	if (atomic_dec_and_test(&priv->ntbl.entries)) {
1061 		if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1062 			complete(&priv->ntbl.flushed);
1063 	}
1064 }
1065 
1066 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1067 {
1068 	/* Called as a result of removal from hash table */
1069 	struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1070 	/* note TX context may hold another ref */
1071 	ipoib_neigh_put(neigh);
1072 }
1073 
1074 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1075 {
1076 	struct net_device *dev = neigh->dev;
1077 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1078 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1079 	struct ipoib_neigh_hash *htbl;
1080 	struct ipoib_neigh __rcu **np;
1081 	struct ipoib_neigh *n;
1082 	u32 hash_val;
1083 
1084 	htbl = rcu_dereference_protected(ntbl->htbl,
1085 					lockdep_is_held(&priv->lock));
1086 	if (!htbl)
1087 		return;
1088 
1089 	hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1090 	np = &htbl->buckets[hash_val];
1091 	for (n = rcu_dereference_protected(*np,
1092 					    lockdep_is_held(&priv->lock));
1093 	     n != NULL;
1094 	     n = rcu_dereference_protected(*np,
1095 					lockdep_is_held(&priv->lock))) {
1096 		if (n == neigh) {
1097 			/* found */
1098 			rcu_assign_pointer(*np,
1099 					   rcu_dereference_protected(neigh->hnext,
1100 								     lockdep_is_held(&priv->lock)));
1101 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1102 			return;
1103 		} else {
1104 			np = &n->hnext;
1105 		}
1106 	}
1107 }
1108 
1109 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1110 {
1111 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1112 	struct ipoib_neigh_hash *htbl;
1113 	struct ipoib_neigh **buckets;
1114 	u32 size;
1115 
1116 	clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1117 	ntbl->htbl = NULL;
1118 	htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1119 	if (!htbl)
1120 		return -ENOMEM;
1121 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1122 	size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1123 	buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1124 	if (!buckets) {
1125 		kfree(htbl);
1126 		return -ENOMEM;
1127 	}
1128 	htbl->size = size;
1129 	htbl->mask = (size - 1);
1130 	htbl->buckets = buckets;
1131 	ntbl->htbl = htbl;
1132 	htbl->ntbl = ntbl;
1133 	atomic_set(&ntbl->entries, 0);
1134 
1135 	/* start garbage collection */
1136 	clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1137 	queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
1138 			   arp_tbl.gc_interval);
1139 
1140 	return 0;
1141 }
1142 
1143 static void neigh_hash_free_rcu(struct rcu_head *head)
1144 {
1145 	struct ipoib_neigh_hash *htbl = container_of(head,
1146 						    struct ipoib_neigh_hash,
1147 						    rcu);
1148 	struct ipoib_neigh __rcu **buckets = htbl->buckets;
1149 	struct ipoib_neigh_table *ntbl = htbl->ntbl;
1150 
1151 	kfree(buckets);
1152 	kfree(htbl);
1153 	complete(&ntbl->deleted);
1154 }
1155 
1156 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1157 {
1158 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1159 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1160 	struct ipoib_neigh_hash *htbl;
1161 	unsigned long flags;
1162 	int i;
1163 
1164 	/* remove all neigh connected to a given path or mcast */
1165 	spin_lock_irqsave(&priv->lock, flags);
1166 
1167 	htbl = rcu_dereference_protected(ntbl->htbl,
1168 					 lockdep_is_held(&priv->lock));
1169 
1170 	if (!htbl)
1171 		goto out_unlock;
1172 
1173 	for (i = 0; i < htbl->size; i++) {
1174 		struct ipoib_neigh *neigh;
1175 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1176 
1177 		while ((neigh = rcu_dereference_protected(*np,
1178 							  lockdep_is_held(&priv->lock))) != NULL) {
1179 			/* delete neighs belong to this parent */
1180 			if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1181 				rcu_assign_pointer(*np,
1182 						   rcu_dereference_protected(neigh->hnext,
1183 									     lockdep_is_held(&priv->lock)));
1184 				/* remove from parent list */
1185 				list_del(&neigh->list);
1186 				call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1187 			} else {
1188 				np = &neigh->hnext;
1189 			}
1190 
1191 		}
1192 	}
1193 out_unlock:
1194 	spin_unlock_irqrestore(&priv->lock, flags);
1195 }
1196 
1197 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1198 {
1199 	struct ipoib_neigh_table *ntbl = &priv->ntbl;
1200 	struct ipoib_neigh_hash *htbl;
1201 	unsigned long flags;
1202 	int i, wait_flushed = 0;
1203 
1204 	init_completion(&priv->ntbl.flushed);
1205 
1206 	spin_lock_irqsave(&priv->lock, flags);
1207 
1208 	htbl = rcu_dereference_protected(ntbl->htbl,
1209 					lockdep_is_held(&priv->lock));
1210 	if (!htbl)
1211 		goto out_unlock;
1212 
1213 	wait_flushed = atomic_read(&priv->ntbl.entries);
1214 	if (!wait_flushed)
1215 		goto free_htbl;
1216 
1217 	for (i = 0; i < htbl->size; i++) {
1218 		struct ipoib_neigh *neigh;
1219 		struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1220 
1221 		while ((neigh = rcu_dereference_protected(*np,
1222 				       lockdep_is_held(&priv->lock))) != NULL) {
1223 			rcu_assign_pointer(*np,
1224 					   rcu_dereference_protected(neigh->hnext,
1225 								     lockdep_is_held(&priv->lock)));
1226 			/* remove from path/mc list */
1227 			list_del(&neigh->list);
1228 			call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1229 		}
1230 	}
1231 
1232 free_htbl:
1233 	rcu_assign_pointer(ntbl->htbl, NULL);
1234 	call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1235 
1236 out_unlock:
1237 	spin_unlock_irqrestore(&priv->lock, flags);
1238 	if (wait_flushed)
1239 		wait_for_completion(&priv->ntbl.flushed);
1240 }
1241 
1242 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1243 {
1244 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1245 	int stopped;
1246 
1247 	ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1248 	init_completion(&priv->ntbl.deleted);
1249 	set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1250 
1251 	/* Stop GC if called at init fail need to cancel work */
1252 	stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1253 	if (!stopped)
1254 		cancel_delayed_work(&priv->neigh_reap_task);
1255 
1256 	ipoib_flush_neighs(priv);
1257 
1258 	wait_for_completion(&priv->ntbl.deleted);
1259 }
1260 
1261 
1262 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1263 {
1264 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1265 
1266 	if (ipoib_neigh_hash_init(priv) < 0)
1267 		goto out;
1268 	/* Allocate RX/TX "rings" to hold queued skbs */
1269 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1270 				GFP_KERNEL);
1271 	if (!priv->rx_ring) {
1272 		printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1273 		       ca->name, ipoib_recvq_size);
1274 		goto out_neigh_hash_cleanup;
1275 	}
1276 
1277 	priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1278 	if (!priv->tx_ring) {
1279 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1280 		       ca->name, ipoib_sendq_size);
1281 		goto out_rx_ring_cleanup;
1282 	}
1283 
1284 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1285 
1286 	if (ipoib_ib_dev_init(dev, ca, port))
1287 		goto out_tx_ring_cleanup;
1288 
1289 	return 0;
1290 
1291 out_tx_ring_cleanup:
1292 	vfree(priv->tx_ring);
1293 
1294 out_rx_ring_cleanup:
1295 	kfree(priv->rx_ring);
1296 
1297 out_neigh_hash_cleanup:
1298 	ipoib_neigh_hash_uninit(dev);
1299 out:
1300 	return -ENOMEM;
1301 }
1302 
1303 void ipoib_dev_cleanup(struct net_device *dev)
1304 {
1305 	struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1306 	LIST_HEAD(head);
1307 
1308 	ASSERT_RTNL();
1309 
1310 	ipoib_delete_debug_files(dev);
1311 
1312 	/* Delete any child interfaces first */
1313 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1314 		/* Stop GC on child */
1315 		set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1316 		cancel_delayed_work(&cpriv->neigh_reap_task);
1317 		unregister_netdevice_queue(cpriv->dev, &head);
1318 	}
1319 	unregister_netdevice_many(&head);
1320 
1321 	ipoib_ib_dev_cleanup(dev);
1322 
1323 	kfree(priv->rx_ring);
1324 	vfree(priv->tx_ring);
1325 
1326 	priv->rx_ring = NULL;
1327 	priv->tx_ring = NULL;
1328 
1329 	ipoib_neigh_hash_uninit(dev);
1330 }
1331 
1332 static const struct header_ops ipoib_header_ops = {
1333 	.create	= ipoib_hard_header,
1334 };
1335 
1336 static const struct net_device_ops ipoib_netdev_ops = {
1337 	.ndo_uninit		 = ipoib_uninit,
1338 	.ndo_open		 = ipoib_open,
1339 	.ndo_stop		 = ipoib_stop,
1340 	.ndo_change_mtu		 = ipoib_change_mtu,
1341 	.ndo_fix_features	 = ipoib_fix_features,
1342 	.ndo_start_xmit	 	 = ipoib_start_xmit,
1343 	.ndo_tx_timeout		 = ipoib_timeout,
1344 	.ndo_set_rx_mode	 = ipoib_set_mcast_list,
1345 };
1346 
1347 void ipoib_setup(struct net_device *dev)
1348 {
1349 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1350 
1351 	dev->netdev_ops		 = &ipoib_netdev_ops;
1352 	dev->header_ops		 = &ipoib_header_ops;
1353 
1354 	ipoib_set_ethtool_ops(dev);
1355 
1356 	netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1357 
1358 	dev->watchdog_timeo	 = HZ;
1359 
1360 	dev->flags		|= IFF_BROADCAST | IFF_MULTICAST;
1361 
1362 	dev->hard_header_len	 = IPOIB_ENCAP_LEN;
1363 	dev->addr_len		 = INFINIBAND_ALEN;
1364 	dev->type		 = ARPHRD_INFINIBAND;
1365 	dev->tx_queue_len	 = ipoib_sendq_size * 2;
1366 	dev->features		 = (NETIF_F_VLAN_CHALLENGED	|
1367 				    NETIF_F_HIGHDMA);
1368 	dev->priv_flags		&= ~IFF_XMIT_DST_RELEASE;
1369 
1370 	memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1371 
1372 	netif_carrier_off(dev);
1373 
1374 	priv->dev = dev;
1375 
1376 	spin_lock_init(&priv->lock);
1377 
1378 	mutex_init(&priv->vlan_mutex);
1379 
1380 	INIT_LIST_HEAD(&priv->path_list);
1381 	INIT_LIST_HEAD(&priv->child_intfs);
1382 	INIT_LIST_HEAD(&priv->dead_ahs);
1383 	INIT_LIST_HEAD(&priv->multicast_list);
1384 
1385 	INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1386 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1387 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1388 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1389 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1390 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1391 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1392 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1393 	INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1394 }
1395 
1396 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1397 {
1398 	struct net_device *dev;
1399 
1400 	dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1401 			   ipoib_setup);
1402 	if (!dev)
1403 		return NULL;
1404 
1405 	return netdev_priv(dev);
1406 }
1407 
1408 static ssize_t show_pkey(struct device *dev,
1409 			 struct device_attribute *attr, char *buf)
1410 {
1411 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1412 
1413 	return sprintf(buf, "0x%04x\n", priv->pkey);
1414 }
1415 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1416 
1417 static ssize_t show_umcast(struct device *dev,
1418 			   struct device_attribute *attr, char *buf)
1419 {
1420 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1421 
1422 	return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1423 }
1424 
1425 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1426 {
1427 	struct ipoib_dev_priv *priv = netdev_priv(ndev);
1428 
1429 	if (umcast_val > 0) {
1430 		set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1431 		ipoib_warn(priv, "ignoring multicast groups joined directly "
1432 				"by userspace\n");
1433 	} else
1434 		clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1435 }
1436 
1437 static ssize_t set_umcast(struct device *dev,
1438 			  struct device_attribute *attr,
1439 			  const char *buf, size_t count)
1440 {
1441 	unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1442 
1443 	ipoib_set_umcast(to_net_dev(dev), umcast_val);
1444 
1445 	return count;
1446 }
1447 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1448 
1449 int ipoib_add_umcast_attr(struct net_device *dev)
1450 {
1451 	return device_create_file(&dev->dev, &dev_attr_umcast);
1452 }
1453 
1454 static ssize_t create_child(struct device *dev,
1455 			    struct device_attribute *attr,
1456 			    const char *buf, size_t count)
1457 {
1458 	int pkey;
1459 	int ret;
1460 
1461 	if (sscanf(buf, "%i", &pkey) != 1)
1462 		return -EINVAL;
1463 
1464 	if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1465 		return -EINVAL;
1466 
1467 	/*
1468 	 * Set the full membership bit, so that we join the right
1469 	 * broadcast group, etc.
1470 	 */
1471 	pkey |= 0x8000;
1472 
1473 	ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1474 
1475 	return ret ? ret : count;
1476 }
1477 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1478 
1479 static ssize_t delete_child(struct device *dev,
1480 			    struct device_attribute *attr,
1481 			    const char *buf, size_t count)
1482 {
1483 	int pkey;
1484 	int ret;
1485 
1486 	if (sscanf(buf, "%i", &pkey) != 1)
1487 		return -EINVAL;
1488 
1489 	if (pkey < 0 || pkey > 0xffff)
1490 		return -EINVAL;
1491 
1492 	ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1493 
1494 	return ret ? ret : count;
1495 
1496 }
1497 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1498 
1499 int ipoib_add_pkey_attr(struct net_device *dev)
1500 {
1501 	return device_create_file(&dev->dev, &dev_attr_pkey);
1502 }
1503 
1504 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1505 {
1506 	struct ib_device_attr *device_attr;
1507 	int result = -ENOMEM;
1508 
1509 	device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1510 	if (!device_attr) {
1511 		printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1512 		       hca->name, sizeof *device_attr);
1513 		return result;
1514 	}
1515 
1516 	result = ib_query_device(hca, device_attr);
1517 	if (result) {
1518 		printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1519 		       hca->name, result);
1520 		kfree(device_attr);
1521 		return result;
1522 	}
1523 	priv->hca_caps = device_attr->device_cap_flags;
1524 
1525 	kfree(device_attr);
1526 
1527 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1528 		priv->dev->hw_features = NETIF_F_SG |
1529 			NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1530 
1531 		if (priv->hca_caps & IB_DEVICE_UD_TSO)
1532 			priv->dev->hw_features |= NETIF_F_TSO;
1533 
1534 		priv->dev->features |= priv->dev->hw_features;
1535 	}
1536 
1537 	return 0;
1538 }
1539 
1540 static struct net_device *ipoib_add_port(const char *format,
1541 					 struct ib_device *hca, u8 port)
1542 {
1543 	struct ipoib_dev_priv *priv;
1544 	struct ib_port_attr attr;
1545 	int result = -ENOMEM;
1546 
1547 	priv = ipoib_intf_alloc(format);
1548 	if (!priv)
1549 		goto alloc_mem_failed;
1550 
1551 	SET_NETDEV_DEV(priv->dev, hca->dma_device);
1552 	priv->dev->dev_id = port - 1;
1553 
1554 	if (!ib_query_port(hca, port, &attr))
1555 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1556 	else {
1557 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1558 		       hca->name, port);
1559 		goto device_init_failed;
1560 	}
1561 
1562 	/* MTU will be reset when mcast join happens */
1563 	priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1564 	priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1565 
1566 	priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1567 
1568 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
1569 	if (result) {
1570 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1571 		       hca->name, port, result);
1572 		goto device_init_failed;
1573 	}
1574 
1575 	if (ipoib_set_dev_features(priv, hca))
1576 		goto device_init_failed;
1577 
1578 	/*
1579 	 * Set the full membership bit, so that we join the right
1580 	 * broadcast group, etc.
1581 	 */
1582 	priv->pkey |= 0x8000;
1583 
1584 	priv->dev->broadcast[8] = priv->pkey >> 8;
1585 	priv->dev->broadcast[9] = priv->pkey & 0xff;
1586 
1587 	result = ib_query_gid(hca, port, 0, &priv->local_gid);
1588 	if (result) {
1589 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1590 		       hca->name, port, result);
1591 		goto device_init_failed;
1592 	} else
1593 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1594 
1595 	result = ipoib_dev_init(priv->dev, hca, port);
1596 	if (result < 0) {
1597 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1598 		       hca->name, port, result);
1599 		goto device_init_failed;
1600 	}
1601 
1602 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
1603 			      priv->ca, ipoib_event);
1604 	result = ib_register_event_handler(&priv->event_handler);
1605 	if (result < 0) {
1606 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1607 		       "port %d (ret = %d)\n",
1608 		       hca->name, port, result);
1609 		goto event_failed;
1610 	}
1611 
1612 	result = register_netdev(priv->dev);
1613 	if (result) {
1614 		printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1615 		       hca->name, port, result);
1616 		goto register_failed;
1617 	}
1618 
1619 	ipoib_create_debug_files(priv->dev);
1620 
1621 	if (ipoib_cm_add_mode_attr(priv->dev))
1622 		goto sysfs_failed;
1623 	if (ipoib_add_pkey_attr(priv->dev))
1624 		goto sysfs_failed;
1625 	if (ipoib_add_umcast_attr(priv->dev))
1626 		goto sysfs_failed;
1627 	if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1628 		goto sysfs_failed;
1629 	if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1630 		goto sysfs_failed;
1631 
1632 	return priv->dev;
1633 
1634 sysfs_failed:
1635 	ipoib_delete_debug_files(priv->dev);
1636 	unregister_netdev(priv->dev);
1637 
1638 register_failed:
1639 	ib_unregister_event_handler(&priv->event_handler);
1640 	/* Stop GC if started before flush */
1641 	set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1642 	cancel_delayed_work(&priv->neigh_reap_task);
1643 	flush_workqueue(ipoib_workqueue);
1644 
1645 event_failed:
1646 	ipoib_dev_cleanup(priv->dev);
1647 
1648 device_init_failed:
1649 	free_netdev(priv->dev);
1650 
1651 alloc_mem_failed:
1652 	return ERR_PTR(result);
1653 }
1654 
1655 static void ipoib_add_one(struct ib_device *device)
1656 {
1657 	struct list_head *dev_list;
1658 	struct net_device *dev;
1659 	struct ipoib_dev_priv *priv;
1660 	int s, e, p;
1661 
1662 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1663 		return;
1664 
1665 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1666 	if (!dev_list)
1667 		return;
1668 
1669 	INIT_LIST_HEAD(dev_list);
1670 
1671 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
1672 		s = 0;
1673 		e = 0;
1674 	} else {
1675 		s = 1;
1676 		e = device->phys_port_cnt;
1677 	}
1678 
1679 	for (p = s; p <= e; ++p) {
1680 		if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1681 			continue;
1682 		dev = ipoib_add_port("ib%d", device, p);
1683 		if (!IS_ERR(dev)) {
1684 			priv = netdev_priv(dev);
1685 			list_add_tail(&priv->list, dev_list);
1686 		}
1687 	}
1688 
1689 	ib_set_client_data(device, &ipoib_client, dev_list);
1690 }
1691 
1692 static void ipoib_remove_one(struct ib_device *device)
1693 {
1694 	struct ipoib_dev_priv *priv, *tmp;
1695 	struct list_head *dev_list;
1696 
1697 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1698 		return;
1699 
1700 	dev_list = ib_get_client_data(device, &ipoib_client);
1701 	if (!dev_list)
1702 		return;
1703 
1704 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
1705 		ib_unregister_event_handler(&priv->event_handler);
1706 
1707 		rtnl_lock();
1708 		dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1709 		rtnl_unlock();
1710 
1711 		/* Stop GC */
1712 		set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1713 		cancel_delayed_work(&priv->neigh_reap_task);
1714 		flush_workqueue(ipoib_workqueue);
1715 
1716 		unregister_netdev(priv->dev);
1717 		free_netdev(priv->dev);
1718 	}
1719 
1720 	kfree(dev_list);
1721 }
1722 
1723 static int __init ipoib_init_module(void)
1724 {
1725 	int ret;
1726 
1727 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1728 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1729 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1730 
1731 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1732 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1733 	ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1734 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1735 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1736 #endif
1737 
1738 	/*
1739 	 * When copying small received packets, we only copy from the
1740 	 * linear data part of the SKB, so we rely on this condition.
1741 	 */
1742 	BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1743 
1744 	ret = ipoib_register_debugfs();
1745 	if (ret)
1746 		return ret;
1747 
1748 	/*
1749 	 * We create our own workqueue mainly because we want to be
1750 	 * able to flush it when devices are being removed.  We can't
1751 	 * use schedule_work()/flush_scheduled_work() because both
1752 	 * unregister_netdev() and linkwatch_event take the rtnl lock,
1753 	 * so flush_scheduled_work() can deadlock during device
1754 	 * removal.
1755 	 */
1756 	ipoib_workqueue = create_singlethread_workqueue("ipoib");
1757 	if (!ipoib_workqueue) {
1758 		ret = -ENOMEM;
1759 		goto err_fs;
1760 	}
1761 
1762 	ib_sa_register_client(&ipoib_sa_client);
1763 
1764 	ret = ib_register_client(&ipoib_client);
1765 	if (ret)
1766 		goto err_sa;
1767 
1768 	ret = ipoib_netlink_init();
1769 	if (ret)
1770 		goto err_client;
1771 
1772 	return 0;
1773 
1774 err_client:
1775 	ib_unregister_client(&ipoib_client);
1776 
1777 err_sa:
1778 	ib_sa_unregister_client(&ipoib_sa_client);
1779 	destroy_workqueue(ipoib_workqueue);
1780 
1781 err_fs:
1782 	ipoib_unregister_debugfs();
1783 
1784 	return ret;
1785 }
1786 
1787 static void __exit ipoib_cleanup_module(void)
1788 {
1789 	ipoib_netlink_fini();
1790 	ib_unregister_client(&ipoib_client);
1791 	ib_sa_unregister_client(&ipoib_sa_client);
1792 	ipoib_unregister_debugfs();
1793 	destroy_workqueue(ipoib_workqueue);
1794 }
1795 
1796 module_init(ipoib_init_module);
1797 module_exit(ipoib_cleanup_module);
1798