xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c (revision d6b92ffa990dc57d8a969a98e3ca4e25ba39cbb2)
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 static	int ipoib_resolvemulti(struct ifnet *, struct sockaddr **,
38 		struct sockaddr *);
39 
40 
41 #include <linux/module.h>
42 
43 #include <linux/slab.h>
44 #include <linux/kernel.h>
45 #include <linux/vmalloc.h>
46 
47 #include <linux/if_arp.h>	/* For ARPHRD_xxx */
48 #include <linux/if_vlan.h>
49 #include <net/ip.h>
50 #include <net/ipv6.h>
51 
52 MODULE_AUTHOR("Roland Dreier");
53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
54 MODULE_LICENSE("Dual BSD/GPL");
55 
56 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
57 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
58 
59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
63 
64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
65 int ipoib_debug_level = 1;
66 
67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
69 #endif
70 
71 struct ipoib_path_iter {
72 	struct ipoib_dev_priv *priv;
73 	struct ipoib_path  path;
74 };
75 
76 static const u8 ipv4_bcast_addr[] = {
77 	0x00, 0xff, 0xff, 0xff,
78 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
79 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
80 };
81 
82 struct workqueue_struct *ipoib_workqueue;
83 
84 struct ib_sa_client ipoib_sa_client;
85 
86 static void ipoib_add_one(struct ib_device *device);
87 static void ipoib_remove_one(struct ib_device *device, void *client_data);
88 static void ipoib_start(struct ifnet *dev);
89 static int ipoib_output(struct ifnet *ifp, struct mbuf *m,
90 	    const struct sockaddr *dst, struct route *ro);
91 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
92 static void ipoib_input(struct ifnet *ifp, struct mbuf *m);
93 
94 #define	IPOIB_MTAP(_ifp, _m)					\
95 do {								\
96 	if (bpf_peers_present((_ifp)->if_bpf)) {		\
97 		M_ASSERTVALID(_m);				\
98 		ipoib_mtap_mb((_ifp), (_m));			\
99 	}							\
100 } while (0)
101 
102 static struct unrhdr *ipoib_unrhdr;
103 
104 static void
105 ipoib_unrhdr_init(void *arg)
106 {
107 
108 	ipoib_unrhdr = new_unrhdr(0, 65535, NULL);
109 }
110 SYSINIT(ipoib_unrhdr_init, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_init, NULL);
111 
112 static void
113 ipoib_unrhdr_uninit(void *arg)
114 {
115 
116 	if (ipoib_unrhdr != NULL) {
117 		struct unrhdr *hdr;
118 
119 		hdr = ipoib_unrhdr;
120 		ipoib_unrhdr = NULL;
121 
122 		delete_unrhdr(hdr);
123 	}
124 }
125 SYSUNINIT(ipoib_unrhdr_uninit, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_uninit, NULL);
126 
127 /*
128  * This is for clients that have an ipoib_header in the mbuf.
129  */
130 static void
131 ipoib_mtap_mb(struct ifnet *ifp, struct mbuf *mb)
132 {
133 	struct ipoib_header *ih;
134 	struct ether_header eh;
135 
136 	ih = mtod(mb, struct ipoib_header *);
137 	eh.ether_type = ih->proto;
138 	bcopy(ih->hwaddr, &eh.ether_dhost, ETHER_ADDR_LEN);
139 	bzero(&eh.ether_shost, ETHER_ADDR_LEN);
140 	mb->m_data += sizeof(struct ipoib_header);
141 	mb->m_len -= sizeof(struct ipoib_header);
142 	bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
143 	mb->m_data -= sizeof(struct ipoib_header);
144 	mb->m_len += sizeof(struct ipoib_header);
145 }
146 
147 void
148 ipoib_mtap_proto(struct ifnet *ifp, struct mbuf *mb, uint16_t proto)
149 {
150 	struct ether_header eh;
151 
152 	eh.ether_type = proto;
153 	bzero(&eh.ether_shost, ETHER_ADDR_LEN);
154 	bzero(&eh.ether_dhost, ETHER_ADDR_LEN);
155 	bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
156 }
157 
158 static struct ib_client ipoib_client = {
159 	.name   = "ipoib",
160 	.add    = ipoib_add_one,
161 	.remove = ipoib_remove_one
162 };
163 
164 int
165 ipoib_open(struct ipoib_dev_priv *priv)
166 {
167 	struct ifnet *dev = priv->dev;
168 
169 	ipoib_dbg(priv, "bringing up interface\n");
170 
171 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
172 
173 	if (ipoib_pkey_dev_delay_open(priv))
174 		return 0;
175 
176 	if (ipoib_ib_dev_open(priv))
177 		goto err_disable;
178 
179 	if (ipoib_ib_dev_up(priv))
180 		goto err_stop;
181 
182 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
183 		struct ipoib_dev_priv *cpriv;
184 
185 		/* Bring up any child interfaces too */
186 		mutex_lock(&priv->vlan_mutex);
187 		list_for_each_entry(cpriv, &priv->child_intfs, list)
188 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
189 				ipoib_open(cpriv);
190 		mutex_unlock(&priv->vlan_mutex);
191 	}
192 	dev->if_drv_flags |= IFF_DRV_RUNNING;
193 	dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
194 
195 	return 0;
196 
197 err_stop:
198 	ipoib_ib_dev_stop(priv, 1);
199 
200 err_disable:
201 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
202 
203 	return -EINVAL;
204 }
205 
206 static void
207 ipoib_init(void *arg)
208 {
209 	struct ifnet *dev;
210 	struct ipoib_dev_priv *priv;
211 
212 	priv = arg;
213 	dev = priv->dev;
214 	if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
215 		ipoib_open(priv);
216 	queue_work(ipoib_workqueue, &priv->flush_light);
217 }
218 
219 
220 static int
221 ipoib_stop(struct ipoib_dev_priv *priv)
222 {
223 	struct ifnet *dev = priv->dev;
224 
225 	ipoib_dbg(priv, "stopping interface\n");
226 
227 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
228 
229 	dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
230 
231 	ipoib_ib_dev_down(priv, 0);
232 	ipoib_ib_dev_stop(priv, 0);
233 
234 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
235 		struct ipoib_dev_priv *cpriv;
236 
237 		/* Bring down any child interfaces too */
238 		mutex_lock(&priv->vlan_mutex);
239 		list_for_each_entry(cpriv, &priv->child_intfs, list)
240 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
241 				ipoib_stop(cpriv);
242 		mutex_unlock(&priv->vlan_mutex);
243 	}
244 
245 	return 0;
246 }
247 
248 int
249 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu)
250 {
251 	struct ifnet *dev = priv->dev;
252 
253 	/* dev->if_mtu > 2K ==> connected mode */
254 	if (ipoib_cm_admin_enabled(priv)) {
255 		if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
256 			return -EINVAL;
257 
258 		if (new_mtu > priv->mcast_mtu)
259 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
260 				   priv->mcast_mtu);
261 
262 		dev->if_mtu = new_mtu;
263 		return 0;
264 	}
265 
266 	if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
267 		return -EINVAL;
268 
269 	priv->admin_mtu = new_mtu;
270 
271 	dev->if_mtu = min(priv->mcast_mtu, priv->admin_mtu);
272 
273 	queue_work(ipoib_workqueue, &priv->flush_light);
274 
275 	return 0;
276 }
277 
278 static int
279 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
280 {
281 	struct ipoib_dev_priv *priv = ifp->if_softc;
282 	struct ifaddr *ifa = (struct ifaddr *) data;
283 	struct ifreq *ifr = (struct ifreq *) data;
284 	int error = 0;
285 
286 	/* check if detaching */
287 	if (priv == NULL || priv->gone != 0)
288 		return (ENXIO);
289 
290 	switch (command) {
291 	case SIOCSIFFLAGS:
292 		if (ifp->if_flags & IFF_UP) {
293 			if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
294 				error = -ipoib_open(priv);
295 		} else
296 			if (ifp->if_drv_flags & IFF_DRV_RUNNING)
297 				ipoib_stop(priv);
298 		break;
299 	case SIOCADDMULTI:
300 	case SIOCDELMULTI:
301 		if (ifp->if_drv_flags & IFF_DRV_RUNNING)
302 			queue_work(ipoib_workqueue, &priv->restart_task);
303 		break;
304 	case SIOCSIFADDR:
305 		ifp->if_flags |= IFF_UP;
306 
307 		switch (ifa->ifa_addr->sa_family) {
308 #ifdef INET
309 		case AF_INET:
310 			ifp->if_init(ifp->if_softc);	/* before arpwhohas */
311 			arp_ifinit(ifp, ifa);
312 			break;
313 #endif
314 		default:
315 			ifp->if_init(ifp->if_softc);
316 			break;
317 		}
318 		break;
319 
320 	case SIOCGIFADDR:
321 		{
322 			struct sockaddr *sa;
323 
324 			sa = (struct sockaddr *) & ifr->ifr_data;
325 			bcopy(IF_LLADDR(ifp),
326 			      (caddr_t) sa->sa_data, INFINIBAND_ALEN);
327 		}
328 		break;
329 
330 	case SIOCSIFMTU:
331 		/*
332 		 * Set the interface MTU.
333 		 */
334 		error = -ipoib_change_mtu(priv, ifr->ifr_mtu);
335 		break;
336 	default:
337 		error = EINVAL;
338 		break;
339 	}
340 	return (error);
341 }
342 
343 
344 static struct ipoib_path *
345 __path_find(struct ipoib_dev_priv *priv, void *gid)
346 {
347 	struct rb_node *n = priv->path_tree.rb_node;
348 	struct ipoib_path *path;
349 	int ret;
350 
351 	while (n) {
352 		path = rb_entry(n, struct ipoib_path, rb_node);
353 
354 		ret = memcmp(gid, path->pathrec.dgid.raw,
355 			     sizeof (union ib_gid));
356 
357 		if (ret < 0)
358 			n = n->rb_left;
359 		else if (ret > 0)
360 			n = n->rb_right;
361 		else
362 			return path;
363 	}
364 
365 	return NULL;
366 }
367 
368 static int
369 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
370 {
371 	struct rb_node **n = &priv->path_tree.rb_node;
372 	struct rb_node *pn = NULL;
373 	struct ipoib_path *tpath;
374 	int ret;
375 
376 	while (*n) {
377 		pn = *n;
378 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
379 
380 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
381 			     sizeof (union ib_gid));
382 		if (ret < 0)
383 			n = &pn->rb_left;
384 		else if (ret > 0)
385 			n = &pn->rb_right;
386 		else
387 			return -EEXIST;
388 	}
389 
390 	rb_link_node(&path->rb_node, pn, n);
391 	rb_insert_color(&path->rb_node, &priv->path_tree);
392 
393 	list_add_tail(&path->list, &priv->path_list);
394 
395 	return 0;
396 }
397 
398 void
399 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
400 {
401 
402 	_IF_DRAIN(&path->queue);
403 
404 	if (path->ah)
405 		ipoib_put_ah(path->ah);
406 	if (ipoib_cm_get(path))
407 		ipoib_cm_destroy_tx(ipoib_cm_get(path));
408 
409 	kfree(path);
410 }
411 
412 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
413 
414 struct ipoib_path_iter *
415 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
416 {
417 	struct ipoib_path_iter *iter;
418 
419 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
420 	if (!iter)
421 		return NULL;
422 
423 	iter->priv = priv;
424 	memset(iter->path.pathrec.dgid.raw, 0, 16);
425 
426 	if (ipoib_path_iter_next(iter)) {
427 		kfree(iter);
428 		return NULL;
429 	}
430 
431 	return iter;
432 }
433 
434 int
435 ipoib_path_iter_next(struct ipoib_path_iter *iter)
436 {
437 	struct ipoib_dev_priv *priv = iter->priv;
438 	struct rb_node *n;
439 	struct ipoib_path *path;
440 	int ret = 1;
441 
442 	spin_lock_irq(&priv->lock);
443 
444 	n = rb_first(&priv->path_tree);
445 
446 	while (n) {
447 		path = rb_entry(n, struct ipoib_path, rb_node);
448 
449 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
450 			   sizeof (union ib_gid)) < 0) {
451 			iter->path = *path;
452 			ret = 0;
453 			break;
454 		}
455 
456 		n = rb_next(n);
457 	}
458 
459 	spin_unlock_irq(&priv->lock);
460 
461 	return ret;
462 }
463 
464 void
465 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
466 {
467 	*path = iter->path;
468 }
469 
470 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
471 
472 void
473 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
474 {
475 	struct ipoib_path *path, *tp;
476 
477 	spin_lock_irq(&priv->lock);
478 
479 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
480 		ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
481 			be16_to_cpu(path->pathrec.dlid),
482 			path->pathrec.dgid.raw, ":");
483 		path->valid =  0;
484 	}
485 
486 	spin_unlock_irq(&priv->lock);
487 }
488 
489 void
490 ipoib_flush_paths(struct ipoib_dev_priv *priv)
491 {
492 	struct ipoib_path *path, *tp;
493 	LIST_HEAD(remove_list);
494 	unsigned long flags;
495 
496 	spin_lock_irqsave(&priv->lock, flags);
497 
498 	list_splice_init(&priv->path_list, &remove_list);
499 
500 	list_for_each_entry(path, &remove_list, list)
501 		rb_erase(&path->rb_node, &priv->path_tree);
502 
503 	list_for_each_entry_safe(path, tp, &remove_list, list) {
504 		if (path->query)
505 			ib_sa_cancel_query(path->query_id, path->query);
506 		spin_unlock_irqrestore(&priv->lock, flags);
507 		wait_for_completion(&path->done);
508 		ipoib_path_free(priv, path);
509 		spin_lock_irqsave(&priv->lock, flags);
510 	}
511 
512 	spin_unlock_irqrestore(&priv->lock, flags);
513 }
514 
515 static void
516 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
517 {
518 	struct ipoib_path *path = path_ptr;
519 	struct ipoib_dev_priv *priv = path->priv;
520 	struct ifnet *dev = priv->dev;
521 	struct ipoib_ah *ah = NULL;
522 	struct ipoib_ah *old_ah = NULL;
523 	struct ifqueue mbqueue;
524 	struct mbuf *mb;
525 	unsigned long flags;
526 
527 	if (!status)
528 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
529 			  be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
530 	else
531 		ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
532 			  status, path->pathrec.dgid.raw, ":");
533 
534 	bzero(&mbqueue, sizeof(mbqueue));
535 
536 	if (!status) {
537 		struct ib_ah_attr av;
538 
539 		if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
540 			ah = ipoib_create_ah(priv, priv->pd, &av);
541 	}
542 
543 	spin_lock_irqsave(&priv->lock, flags);
544 
545 	if (ah) {
546 		path->pathrec = *pathrec;
547 
548 		old_ah   = path->ah;
549 		path->ah = ah;
550 
551 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
552 			  ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
553 
554 		for (;;) {
555 			_IF_DEQUEUE(&path->queue, mb);
556 			if (mb == NULL)
557 				break;
558 			_IF_ENQUEUE(&mbqueue, mb);
559 		}
560 
561 #ifdef CONFIG_INFINIBAND_IPOIB_CM
562 		if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
563 			ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
564 #endif
565 
566 		path->valid = 1;
567 	}
568 
569 	path->query = NULL;
570 	complete(&path->done);
571 
572 	spin_unlock_irqrestore(&priv->lock, flags);
573 
574 	if (old_ah)
575 		ipoib_put_ah(old_ah);
576 
577 	for (;;) {
578 		_IF_DEQUEUE(&mbqueue, mb);
579 		if (mb == NULL)
580 			break;
581 		mb->m_pkthdr.rcvif = dev;
582 		if (dev->if_transmit(dev, mb))
583 			ipoib_warn(priv, "dev_queue_xmit failed "
584 				   "to requeue packet\n");
585 	}
586 }
587 
588 static struct ipoib_path *
589 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
590 {
591 	struct ipoib_path *path;
592 
593 	if (!priv->broadcast)
594 		return NULL;
595 
596 	path = kzalloc(sizeof *path, GFP_ATOMIC);
597 	if (!path)
598 		return NULL;
599 
600 	path->priv = priv;
601 
602 	bzero(&path->queue, sizeof(path->queue));
603 
604 #ifdef CONFIG_INFINIBAND_IPOIB_CM
605 	memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
606 #endif
607 	memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
608 	path->pathrec.sgid	    = priv->local_gid;
609 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
610 	path->pathrec.numb_path     = 1;
611 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
612 
613 	return path;
614 }
615 
616 static int
617 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
618 {
619 	struct ifnet *dev = priv->dev;
620 
621 	ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
622 	struct ib_sa_path_rec p_rec;
623 
624 	p_rec = path->pathrec;
625 	p_rec.mtu_selector = IB_SA_GT;
626 
627 	switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
628 	case 512:
629 		p_rec.mtu = IB_MTU_256;
630 		break;
631 	case 1024:
632 		p_rec.mtu = IB_MTU_512;
633 		break;
634 	case 2048:
635 		p_rec.mtu = IB_MTU_1024;
636 		break;
637 	case 4096:
638 		p_rec.mtu = IB_MTU_2048;
639 		break;
640 	default:
641 		/* Wildcard everything */
642 		comp_mask = 0;
643 		p_rec.mtu = 0;
644 		p_rec.mtu_selector = 0;
645 	}
646 
647 	ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
648 		  p_rec.dgid.raw, ":",
649 		  comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
650 
651 	init_completion(&path->done);
652 
653 	path->query_id =
654 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
655 				   &p_rec, comp_mask		|
656 				   IB_SA_PATH_REC_DGID		|
657 				   IB_SA_PATH_REC_SGID		|
658 				   IB_SA_PATH_REC_NUMB_PATH	|
659 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
660 				   IB_SA_PATH_REC_PKEY,
661 				   1000, GFP_ATOMIC,
662 				   path_rec_completion,
663 				   path, &path->query);
664 	if (path->query_id < 0) {
665 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
666 		path->query = NULL;
667 		complete(&path->done);
668 		return path->query_id;
669 	}
670 
671 	return 0;
672 }
673 
674 static void
675 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
676 {
677 	struct ipoib_path *path;
678 
679 	path = __path_find(priv, eh->hwaddr + 4);
680 	if (!path || !path->valid) {
681 		int new_path = 0;
682 
683 		if (!path) {
684 			path = path_rec_create(priv, eh->hwaddr);
685 			new_path = 1;
686 		}
687 		if (path) {
688 			if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
689 				_IF_ENQUEUE(&path->queue, mb);
690 			else {
691 				if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
692 				m_freem(mb);
693 			}
694 
695 			if (!path->query && path_rec_start(priv, path)) {
696 				spin_unlock_irqrestore(&priv->lock, flags);
697 				if (new_path)
698 					ipoib_path_free(priv, path);
699 				return;
700 			} else
701 				__path_add(priv, path);
702 		} else {
703 			if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
704 			m_freem(mb);
705 		}
706 
707 		return;
708 	}
709 
710 	if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
711 		ipoib_cm_send(priv, mb, ipoib_cm_get(path));
712 	} else if (path->ah) {
713 		ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
714 	} else if ((path->query || !path_rec_start(priv, path)) &&
715 		    path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
716 		_IF_ENQUEUE(&path->queue, mb);
717 	} else {
718 		if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
719 		m_freem(mb);
720 	}
721 }
722 
723 static int
724 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
725 {
726 	struct ipoib_header *eh;
727 
728 	eh = mtod(mb, struct ipoib_header *);
729 	if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
730 		/* Add in the P_Key for multicast*/
731 		eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
732 		eh->hwaddr[9] = priv->pkey & 0xff;
733 
734 		ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
735 	} else
736 		ipoib_unicast_send(mb, priv, eh);
737 
738 	return 0;
739 }
740 
741 
742 static void
743 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
744 {
745 	struct mbuf *mb;
746 
747 	if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
748 	    IFF_DRV_RUNNING)
749 		return;
750 
751 	spin_lock(&priv->lock);
752 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
753 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
754 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
755 		if (mb == NULL)
756 			break;
757 		IPOIB_MTAP(dev, mb);
758 		ipoib_send_one(priv, mb);
759 	}
760 	spin_unlock(&priv->lock);
761 }
762 
763 static void
764 ipoib_start(struct ifnet *dev)
765 {
766 	_ipoib_start(dev, dev->if_softc);
767 }
768 
769 static void
770 ipoib_vlan_start(struct ifnet *dev)
771 {
772 	struct ipoib_dev_priv *priv;
773 	struct mbuf *mb;
774 
775 	priv = VLAN_COOKIE(dev);
776 	if (priv != NULL)
777 		return _ipoib_start(dev, priv);
778 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
779 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
780 		if (mb == NULL)
781 			break;
782 		m_freem(mb);
783 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
784 	}
785 }
786 
787 int
788 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
789 {
790 
791 	/* Allocate RX/TX "rings" to hold queued mbs */
792 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
793 				GFP_KERNEL);
794 	if (!priv->rx_ring) {
795 		printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
796 		       ca->name, ipoib_recvq_size);
797 		goto out;
798 	}
799 
800 	priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
801 	if (!priv->tx_ring) {
802 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
803 		       ca->name, ipoib_sendq_size);
804 		goto out_rx_ring_cleanup;
805 	}
806 	memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
807 
808 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
809 
810 	if (ipoib_ib_dev_init(priv, ca, port))
811 		goto out_tx_ring_cleanup;
812 
813 	return 0;
814 
815 out_tx_ring_cleanup:
816 	kfree(priv->tx_ring);
817 
818 out_rx_ring_cleanup:
819 	kfree(priv->rx_ring);
820 
821 out:
822 	return -ENOMEM;
823 }
824 
825 static void
826 ipoib_detach(struct ipoib_dev_priv *priv)
827 {
828 	struct ifnet *dev;
829 
830 	dev = priv->dev;
831 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
832 		priv->gone = 1;
833 		bpfdetach(dev);
834 		if_detach(dev);
835 		if_free(dev);
836 		free_unr(ipoib_unrhdr, priv->unit);
837 	} else
838 		VLAN_SETCOOKIE(priv->dev, NULL);
839 
840 	free(priv, M_TEMP);
841 }
842 
843 void
844 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
845 {
846 	struct ipoib_dev_priv *cpriv, *tcpriv;
847 
848 	/* Delete any child interfaces first */
849 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
850 		ipoib_dev_cleanup(cpriv);
851 		ipoib_detach(cpriv);
852 	}
853 
854 	ipoib_ib_dev_cleanup(priv);
855 
856 	kfree(priv->rx_ring);
857 	kfree(priv->tx_ring);
858 
859 	priv->rx_ring = NULL;
860 	priv->tx_ring = NULL;
861 }
862 
863 static struct ipoib_dev_priv *
864 ipoib_priv_alloc(void)
865 {
866 	struct ipoib_dev_priv *priv;
867 
868 	priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
869 	spin_lock_init(&priv->lock);
870 	spin_lock_init(&priv->drain_lock);
871 	mutex_init(&priv->vlan_mutex);
872 	INIT_LIST_HEAD(&priv->path_list);
873 	INIT_LIST_HEAD(&priv->child_intfs);
874 	INIT_LIST_HEAD(&priv->dead_ahs);
875 	INIT_LIST_HEAD(&priv->multicast_list);
876 	INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
877 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
878 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
879 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
880 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
881 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
882 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
883 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
884 	memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
885 
886 	return (priv);
887 }
888 
889 struct ipoib_dev_priv *
890 ipoib_intf_alloc(const char *name)
891 {
892 	struct ipoib_dev_priv *priv;
893 	struct sockaddr_dl *sdl;
894 	struct ifnet *dev;
895 
896 	priv = ipoib_priv_alloc();
897 	dev = priv->dev = if_alloc(IFT_INFINIBAND);
898 	if (!dev) {
899 		free(priv, M_TEMP);
900 		return NULL;
901 	}
902 	dev->if_softc = priv;
903 	priv->unit = alloc_unr(ipoib_unrhdr);
904 	if (priv->unit == -1) {
905 		if_free(dev);
906 		free(priv, M_TEMP);
907 		return NULL;
908 	}
909 	if_initname(dev, name, priv->unit);
910 	dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
911 	dev->if_addrlen = INFINIBAND_ALEN;
912 	dev->if_hdrlen = IPOIB_HEADER_LEN;
913 	if_attach(dev);
914 	dev->if_init = ipoib_init;
915 	dev->if_ioctl = ipoib_ioctl;
916 	dev->if_start = ipoib_start;
917 	dev->if_output = ipoib_output;
918 	dev->if_input = ipoib_input;
919 	dev->if_resolvemulti = ipoib_resolvemulti;
920 	dev->if_baudrate = IF_Gbps(10);
921 	dev->if_broadcastaddr = priv->broadcastaddr;
922 	dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
923 	sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
924 	sdl->sdl_type = IFT_INFINIBAND;
925 	sdl->sdl_alen = dev->if_addrlen;
926 	priv->dev = dev;
927 	if_link_state_change(dev, LINK_STATE_DOWN);
928 	bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
929 
930 	return dev->if_softc;
931 }
932 
933 int
934 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
935 {
936 	struct ib_device_attr *device_attr = &hca->attrs;
937 
938 	priv->hca_caps = device_attr->device_cap_flags;
939 
940 	priv->dev->if_hwassist = 0;
941 	priv->dev->if_capabilities = 0;
942 
943 #ifndef CONFIG_INFINIBAND_IPOIB_CM
944 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
945 		set_bit(IPOIB_FLAG_CSUM, &priv->flags);
946 		priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
947 		priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
948 	}
949 
950 #if 0
951 	if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
952 		priv->dev->if_capabilities |= IFCAP_TSO4;
953 		priv->dev->if_hwassist |= CSUM_TSO;
954 	}
955 #endif
956 #endif
957 	priv->dev->if_capabilities |=
958 	    IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
959 	priv->dev->if_capenable = priv->dev->if_capabilities;
960 
961 	return 0;
962 }
963 
964 
965 static struct ifnet *
966 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
967 {
968 	struct ipoib_dev_priv *priv;
969 	struct ib_port_attr attr;
970 	int result = -ENOMEM;
971 
972 	priv = ipoib_intf_alloc(format);
973 	if (!priv)
974 		goto alloc_mem_failed;
975 
976 	if (!ib_query_port(hca, port, &attr))
977 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
978 	else {
979 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
980 		       hca->name, port);
981 		goto device_init_failed;
982 	}
983 
984 	/* MTU will be reset when mcast join happens */
985 	priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
986 	priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
987 
988 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
989 	if (result) {
990 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
991 		       hca->name, port, result);
992 		goto device_init_failed;
993 	}
994 
995 	if (ipoib_set_dev_features(priv, hca))
996 		goto device_init_failed;
997 
998 	/*
999 	 * Set the full membership bit, so that we join the right
1000 	 * broadcast group, etc.
1001 	 */
1002 	priv->pkey |= 0x8000;
1003 
1004 	priv->broadcastaddr[8] = priv->pkey >> 8;
1005 	priv->broadcastaddr[9] = priv->pkey & 0xff;
1006 
1007 	result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1008 	if (result) {
1009 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1010 		       hca->name, port, result);
1011 		goto device_init_failed;
1012 	}
1013 	memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
1014 
1015 	result = ipoib_dev_init(priv, hca, port);
1016 	if (result < 0) {
1017 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1018 		       hca->name, port, result);
1019 		goto device_init_failed;
1020 	}
1021 	if (ipoib_cm_admin_enabled(priv))
1022 		priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1023 
1024 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
1025 			      priv->ca, ipoib_event);
1026 	result = ib_register_event_handler(&priv->event_handler);
1027 	if (result < 0) {
1028 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1029 		       "port %d (ret = %d)\n",
1030 		       hca->name, port, result);
1031 		goto event_failed;
1032 	}
1033 	if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1034 
1035 	return priv->dev;
1036 
1037 event_failed:
1038 	ipoib_dev_cleanup(priv);
1039 
1040 device_init_failed:
1041 	ipoib_detach(priv);
1042 
1043 alloc_mem_failed:
1044 	return ERR_PTR(result);
1045 }
1046 
1047 static void
1048 ipoib_add_one(struct ib_device *device)
1049 {
1050 	struct list_head *dev_list;
1051 	struct ifnet *dev;
1052 	struct ipoib_dev_priv *priv;
1053 	int s, e, p;
1054 
1055 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1056 		return;
1057 
1058 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1059 	if (!dev_list)
1060 		return;
1061 
1062 	INIT_LIST_HEAD(dev_list);
1063 
1064 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
1065 		s = 0;
1066 		e = 0;
1067 	} else {
1068 		s = 1;
1069 		e = device->phys_port_cnt;
1070 	}
1071 
1072 	for (p = s; p <= e; ++p) {
1073 		if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1074 			continue;
1075 		dev = ipoib_add_port("ib", device, p);
1076 		if (!IS_ERR(dev)) {
1077 			priv = dev->if_softc;
1078 			list_add_tail(&priv->list, dev_list);
1079 		}
1080 	}
1081 
1082 	ib_set_client_data(device, &ipoib_client, dev_list);
1083 }
1084 
1085 static void
1086 ipoib_remove_one(struct ib_device *device, void *client_data)
1087 {
1088 	struct ipoib_dev_priv *priv, *tmp;
1089 	struct list_head *dev_list = client_data;
1090 
1091 	if (!dev_list)
1092 		return;
1093 
1094 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1095 		return;
1096 
1097 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
1098 		if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1099 			continue;
1100 
1101 		ipoib_stop(priv);
1102 
1103 		ib_unregister_event_handler(&priv->event_handler);
1104 
1105 		/* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1106 
1107 		flush_workqueue(ipoib_workqueue);
1108 
1109 		ipoib_dev_cleanup(priv);
1110 		ipoib_detach(priv);
1111 	}
1112 
1113 	kfree(dev_list);
1114 }
1115 
1116 static void
1117 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1118 {
1119 	struct ipoib_dev_priv *parent;
1120 	struct ipoib_dev_priv *priv;
1121 	struct ifnet *dev;
1122 	uint16_t pkey;
1123 	int error;
1124 
1125 	if (ifp->if_type != IFT_INFINIBAND)
1126 		return;
1127 	dev = VLAN_DEVAT(ifp, vtag);
1128 	if (dev == NULL)
1129 		return;
1130 	priv = NULL;
1131 	error = 0;
1132 	parent = ifp->if_softc;
1133 	/* We only support 15 bits of pkey. */
1134 	if (vtag & 0x8000)
1135 		return;
1136 	pkey = vtag | 0x8000;	/* Set full membership bit. */
1137 	if (pkey == parent->pkey)
1138 		return;
1139 	/* Check for dups */
1140 	mutex_lock(&parent->vlan_mutex);
1141 	list_for_each_entry(priv, &parent->child_intfs, list) {
1142 		if (priv->pkey == pkey) {
1143 			priv = NULL;
1144 			error = EBUSY;
1145 			goto out;
1146 		}
1147 	}
1148 	priv = ipoib_priv_alloc();
1149 	priv->dev = dev;
1150 	priv->max_ib_mtu = parent->max_ib_mtu;
1151 	priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1152 	set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1153 	error = ipoib_set_dev_features(priv, parent->ca);
1154 	if (error)
1155 		goto out;
1156 	priv->pkey = pkey;
1157 	priv->broadcastaddr[8] = pkey >> 8;
1158 	priv->broadcastaddr[9] = pkey & 0xff;
1159 	dev->if_broadcastaddr = priv->broadcastaddr;
1160 	error = ipoib_dev_init(priv, parent->ca, parent->port);
1161 	if (error)
1162 		goto out;
1163 	priv->parent = parent->dev;
1164 	list_add_tail(&priv->list, &parent->child_intfs);
1165 	VLAN_SETCOOKIE(dev, priv);
1166 	dev->if_start = ipoib_vlan_start;
1167 	dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1168 	dev->if_hdrlen = IPOIB_HEADER_LEN;
1169 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1170 		ipoib_open(priv);
1171 	mutex_unlock(&parent->vlan_mutex);
1172 	return;
1173 out:
1174 	mutex_unlock(&parent->vlan_mutex);
1175 	if (priv)
1176 		free(priv, M_TEMP);
1177 	if (error)
1178 		ipoib_warn(parent,
1179 		    "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1180 		    parent->ca->name, parent->port, vtag);
1181 	return;
1182 }
1183 
1184 static void
1185 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1186 {
1187 	struct ipoib_dev_priv *parent;
1188 	struct ipoib_dev_priv *priv;
1189 	struct ifnet *dev;
1190 	uint16_t pkey;
1191 
1192 	if (ifp->if_type != IFT_INFINIBAND)
1193 		return;
1194 
1195 	dev = VLAN_DEVAT(ifp, vtag);
1196 	if (dev)
1197 		VLAN_SETCOOKIE(dev, NULL);
1198 	pkey = vtag | 0x8000;
1199 	parent = ifp->if_softc;
1200 	mutex_lock(&parent->vlan_mutex);
1201 	list_for_each_entry(priv, &parent->child_intfs, list) {
1202 		if (priv->pkey == pkey) {
1203 			ipoib_dev_cleanup(priv);
1204 			list_del(&priv->list);
1205 			break;
1206 		}
1207 	}
1208 	mutex_unlock(&parent->vlan_mutex);
1209 }
1210 
1211 eventhandler_tag ipoib_vlan_attach;
1212 eventhandler_tag ipoib_vlan_detach;
1213 
1214 static int __init
1215 ipoib_init_module(void)
1216 {
1217 	int ret;
1218 
1219 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1220 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1221 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1222 
1223 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1224 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1225 	ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1226 						     IPOIB_MIN_QUEUE_SIZE));
1227 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1228 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1229 #endif
1230 
1231 	ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1232 		ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1233 	ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1234 		ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1235 
1236 	/*
1237 	 * We create our own workqueue mainly because we want to be
1238 	 * able to flush it when devices are being removed.  We can't
1239 	 * use schedule_work()/flush_scheduled_work() because both
1240 	 * unregister_netdev() and linkwatch_event take the rtnl lock,
1241 	 * so flush_scheduled_work() can deadlock during device
1242 	 * removal.
1243 	 */
1244 	ipoib_workqueue = create_singlethread_workqueue("ipoib");
1245 	if (!ipoib_workqueue) {
1246 		ret = -ENOMEM;
1247 		goto err_fs;
1248 	}
1249 
1250 	ib_sa_register_client(&ipoib_sa_client);
1251 
1252 	ret = ib_register_client(&ipoib_client);
1253 	if (ret)
1254 		goto err_sa;
1255 
1256 	return 0;
1257 
1258 err_sa:
1259 	ib_sa_unregister_client(&ipoib_sa_client);
1260 	destroy_workqueue(ipoib_workqueue);
1261 
1262 err_fs:
1263 	return ret;
1264 }
1265 
1266 static void __exit
1267 ipoib_cleanup_module(void)
1268 {
1269 
1270 	EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1271 	EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1272 	ib_unregister_client(&ipoib_client);
1273 	ib_sa_unregister_client(&ipoib_sa_client);
1274 	destroy_workqueue(ipoib_workqueue);
1275 }
1276 
1277 /*
1278  * Infiniband output routine.
1279  */
1280 static int
1281 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1282 	const struct sockaddr *dst, struct route *ro)
1283 {
1284 	u_char edst[INFINIBAND_ALEN];
1285 #if defined(INET) || defined(INET6)
1286 	struct llentry *lle = NULL;
1287 #endif
1288 	struct ipoib_header *eh;
1289 	int error = 0, is_gw = 0;
1290 	short type;
1291 
1292 	if (ro != NULL)
1293 		is_gw = (ro->ro_flags & RT_HAS_GW) != 0;
1294 #ifdef MAC
1295 	error = mac_ifnet_check_transmit(ifp, m);
1296 	if (error)
1297 		goto bad;
1298 #endif
1299 
1300 	M_PROFILE(m);
1301 	if (ifp->if_flags & IFF_MONITOR) {
1302 		error = ENETDOWN;
1303 		goto bad;
1304 	}
1305 	if (!((ifp->if_flags & IFF_UP) &&
1306 	    (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1307 		error = ENETDOWN;
1308 		goto bad;
1309 	}
1310 
1311 	switch (dst->sa_family) {
1312 #ifdef INET
1313 	case AF_INET:
1314 		if (lle != NULL && (lle->la_flags & LLE_VALID))
1315 			memcpy(edst, lle->ll_addr, sizeof(edst));
1316 		else if (m->m_flags & M_MCAST)
1317 			ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1318 		else
1319 			error = arpresolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1320 		if (error)
1321 			return (error == EWOULDBLOCK ? 0 : error);
1322 		type = htons(ETHERTYPE_IP);
1323 		break;
1324 	case AF_ARP:
1325 	{
1326 		struct arphdr *ah;
1327 		ah = mtod(m, struct arphdr *);
1328 		ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1329 
1330 		switch(ntohs(ah->ar_op)) {
1331 		case ARPOP_REVREQUEST:
1332 		case ARPOP_REVREPLY:
1333 			type = htons(ETHERTYPE_REVARP);
1334 			break;
1335 		case ARPOP_REQUEST:
1336 		case ARPOP_REPLY:
1337 		default:
1338 			type = htons(ETHERTYPE_ARP);
1339 			break;
1340 		}
1341 
1342 		if (m->m_flags & M_BCAST)
1343 			bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1344 		else
1345 			bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1346 
1347 	}
1348 	break;
1349 #endif
1350 #ifdef INET6
1351 	case AF_INET6:
1352 		if (lle != NULL && (lle->la_flags & LLE_VALID))
1353 			memcpy(edst, lle->ll_addr, sizeof(edst));
1354 		else if (m->m_flags & M_MCAST)
1355 			ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1356 		else
1357 			error = nd6_resolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1358 		if (error)
1359 			return error;
1360 		type = htons(ETHERTYPE_IPV6);
1361 		break;
1362 #endif
1363 
1364 	default:
1365 		if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1366 		error = EAFNOSUPPORT;
1367 		goto bad;
1368 	}
1369 
1370 	/*
1371 	 * Add local net header.  If no space in first mbuf,
1372 	 * allocate another.
1373 	 */
1374 	M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1375 	if (m == NULL) {
1376 		error = ENOBUFS;
1377 		goto bad;
1378 	}
1379 	eh = mtod(m, struct ipoib_header *);
1380 	(void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1381 	(void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1382 
1383 	/*
1384 	 * Queue message on interface, update output statistics if
1385 	 * successful, and start output if interface not yet active.
1386 	 */
1387 	return ((ifp->if_transmit)(ifp, m));
1388 bad:
1389 	if (m != NULL)
1390 		m_freem(m);
1391 	return (error);
1392 }
1393 
1394 /*
1395  * Upper layer processing for a received Infiniband packet.
1396  */
1397 void
1398 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1399 {
1400 	int isr;
1401 
1402 #ifdef MAC
1403 	/*
1404 	 * Tag the mbuf with an appropriate MAC label before any other
1405 	 * consumers can get to it.
1406 	 */
1407 	mac_ifnet_create_mbuf(ifp, m);
1408 #endif
1409 	/* Allow monitor mode to claim this frame, after stats are updated. */
1410 	if (ifp->if_flags & IFF_MONITOR) {
1411 		if_printf(ifp, "discard frame at IFF_MONITOR\n");
1412 		m_freem(m);
1413 		return;
1414 	}
1415 	/*
1416 	 * Dispatch frame to upper layer.
1417 	 */
1418 	switch (proto) {
1419 #ifdef INET
1420 	case ETHERTYPE_IP:
1421 		isr = NETISR_IP;
1422 		break;
1423 
1424 	case ETHERTYPE_ARP:
1425 		if (ifp->if_flags & IFF_NOARP) {
1426 			/* Discard packet if ARP is disabled on interface */
1427 			m_freem(m);
1428 			return;
1429 		}
1430 		isr = NETISR_ARP;
1431 		break;
1432 #endif
1433 #ifdef INET6
1434 	case ETHERTYPE_IPV6:
1435 		isr = NETISR_IPV6;
1436 		break;
1437 #endif
1438 	default:
1439 		goto discard;
1440 	}
1441 	netisr_dispatch(isr, m);
1442 	return;
1443 
1444 discard:
1445 	m_freem(m);
1446 }
1447 
1448 /*
1449  * Process a received Infiniband packet.
1450  */
1451 static void
1452 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1453 {
1454 	struct ipoib_header *eh;
1455 
1456 	if ((ifp->if_flags & IFF_UP) == 0) {
1457 		m_freem(m);
1458 		return;
1459 	}
1460 	CURVNET_SET_QUIET(ifp->if_vnet);
1461 
1462 	/* Let BPF have it before we strip the header. */
1463 	IPOIB_MTAP(ifp, m);
1464 	eh = mtod(m, struct ipoib_header *);
1465 	/*
1466 	 * Reset layer specific mbuf flags to avoid confusing upper layers.
1467 	 * Strip off Infiniband header.
1468 	 */
1469 	m->m_flags &= ~M_VLANTAG;
1470 	m_clrprotoflags(m);
1471 	m_adj(m, IPOIB_HEADER_LEN);
1472 
1473 	if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1474 		if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1475 		    ifp->if_addrlen) == 0)
1476 			m->m_flags |= M_BCAST;
1477 		else
1478 			m->m_flags |= M_MCAST;
1479 		if_inc_counter(ifp, IFCOUNTER_IMCASTS, 1);
1480 	}
1481 
1482 	ipoib_demux(ifp, m, ntohs(eh->proto));
1483 	CURVNET_RESTORE();
1484 }
1485 
1486 static int
1487 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1488 	struct sockaddr *sa)
1489 {
1490 	struct sockaddr_dl *sdl;
1491 #ifdef INET
1492 	struct sockaddr_in *sin;
1493 #endif
1494 #ifdef INET6
1495 	struct sockaddr_in6 *sin6;
1496 #endif
1497 	u_char *e_addr;
1498 
1499 	switch(sa->sa_family) {
1500 	case AF_LINK:
1501 		/*
1502 		 * No mapping needed. Just check that it's a valid MC address.
1503 		 */
1504 		sdl = (struct sockaddr_dl *)sa;
1505 		e_addr = LLADDR(sdl);
1506 		if (!IPOIB_IS_MULTICAST(e_addr))
1507 			return EADDRNOTAVAIL;
1508 		*llsa = NULL;
1509 		return 0;
1510 
1511 #ifdef INET
1512 	case AF_INET:
1513 		sin = (struct sockaddr_in *)sa;
1514 		if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1515 			return EADDRNOTAVAIL;
1516 		sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1517 		sdl->sdl_alen = INFINIBAND_ALEN;
1518 		e_addr = LLADDR(sdl);
1519 		ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1520 		    e_addr);
1521 		*llsa = (struct sockaddr *)sdl;
1522 		return 0;
1523 #endif
1524 #ifdef INET6
1525 	case AF_INET6:
1526 		sin6 = (struct sockaddr_in6 *)sa;
1527 		/*
1528 		 * An IP6 address of 0 means listen to all
1529 		 * of the multicast address used for IP6.
1530 		 * This has no meaning in ipoib.
1531 		 */
1532 		if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1533 			return EADDRNOTAVAIL;
1534 		if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1535 			return EADDRNOTAVAIL;
1536 		sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1537 		sdl->sdl_alen = INFINIBAND_ALEN;
1538 		e_addr = LLADDR(sdl);
1539 		ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1540 		*llsa = (struct sockaddr *)sdl;
1541 		return 0;
1542 #endif
1543 
1544 	default:
1545 		return EAFNOSUPPORT;
1546 	}
1547 }
1548 
1549 module_init(ipoib_init_module);
1550 module_exit(ipoib_cleanup_module);
1551 
1552 static int
1553 ipoib_evhand(module_t mod, int event, void *arg)
1554 {
1555 	                return (0);
1556 }
1557 
1558 static moduledata_t ipoib_mod = {
1559 	                .name = "ipoib",
1560 			                .evhand = ipoib_evhand,
1561 };
1562 
1563 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1564 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1565 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);
1566