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