xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c (revision 52c2bb75163559a6e2866ad374a7de67a4ea1273)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
3  *
4  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "ipoib.h"
41 
42 static	int ipoib_resolvemulti(struct ifnet *, struct sockaddr **,
43 		struct sockaddr *);
44 
45 
46 #include <linux/module.h>
47 
48 #include <linux/slab.h>
49 #include <linux/kernel.h>
50 #include <linux/vmalloc.h>
51 
52 #include <linux/if_arp.h>	/* For ARPHRD_xxx */
53 #include <linux/if_vlan.h>
54 #include <net/ip.h>
55 #include <net/ipv6.h>
56 
57 #include <rdma/ib_cache.h>
58 
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
61 MODULE_LICENSE("Dual BSD/GPL");
62 
63 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
64 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
65 
66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
70 
71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
72 int ipoib_debug_level = 1;
73 
74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
76 #endif
77 
78 struct ipoib_path_iter {
79 	struct ipoib_dev_priv *priv;
80 	struct ipoib_path  path;
81 };
82 
83 static const u8 ipv4_bcast_addr[] = {
84 	0x00, 0xff, 0xff, 0xff,
85 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
86 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
87 };
88 
89 struct workqueue_struct *ipoib_workqueue;
90 
91 struct ib_sa_client ipoib_sa_client;
92 
93 static void ipoib_add_one(struct ib_device *device);
94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
95 static struct net_device *ipoib_get_net_dev_by_params(
96 		struct ib_device *dev, u8 port, u16 pkey,
97 		const union ib_gid *gid, const struct sockaddr *addr,
98 		void *client_data);
99 static void ipoib_start(struct ifnet *dev);
100 static int ipoib_output(struct ifnet *ifp, struct mbuf *m,
101 	    const struct sockaddr *dst, struct route *ro);
102 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
103 static void ipoib_input(struct ifnet *ifp, struct mbuf *m);
104 
105 #define	IPOIB_MTAP(_ifp, _m)					\
106 do {								\
107 	if (bpf_peers_present((_ifp)->if_bpf)) {		\
108 		M_ASSERTVALID(_m);				\
109 		ipoib_mtap_mb((_ifp), (_m));			\
110 	}							\
111 } while (0)
112 
113 static struct unrhdr *ipoib_unrhdr;
114 
115 static void
116 ipoib_unrhdr_init(void *arg)
117 {
118 
119 	ipoib_unrhdr = new_unrhdr(0, 65535, NULL);
120 }
121 SYSINIT(ipoib_unrhdr_init, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_init, NULL);
122 
123 static void
124 ipoib_unrhdr_uninit(void *arg)
125 {
126 
127 	if (ipoib_unrhdr != NULL) {
128 		struct unrhdr *hdr;
129 
130 		hdr = ipoib_unrhdr;
131 		ipoib_unrhdr = NULL;
132 
133 		delete_unrhdr(hdr);
134 	}
135 }
136 SYSUNINIT(ipoib_unrhdr_uninit, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_uninit, NULL);
137 
138 /*
139  * This is for clients that have an ipoib_header in the mbuf.
140  */
141 static void
142 ipoib_mtap_mb(struct ifnet *ifp, struct mbuf *mb)
143 {
144 	struct ipoib_header *ih;
145 	struct ether_header eh;
146 
147 	ih = mtod(mb, struct ipoib_header *);
148 	eh.ether_type = ih->proto;
149 	bcopy(ih->hwaddr, &eh.ether_dhost, ETHER_ADDR_LEN);
150 	bzero(&eh.ether_shost, ETHER_ADDR_LEN);
151 	mb->m_data += sizeof(struct ipoib_header);
152 	mb->m_len -= sizeof(struct ipoib_header);
153 	bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
154 	mb->m_data -= sizeof(struct ipoib_header);
155 	mb->m_len += sizeof(struct ipoib_header);
156 }
157 
158 void
159 ipoib_mtap_proto(struct ifnet *ifp, struct mbuf *mb, uint16_t proto)
160 {
161 	struct ether_header eh;
162 
163 	eh.ether_type = proto;
164 	bzero(&eh.ether_shost, ETHER_ADDR_LEN);
165 	bzero(&eh.ether_dhost, ETHER_ADDR_LEN);
166 	bpf_mtap2(ifp->if_bpf, &eh, sizeof(eh), mb);
167 }
168 
169 static struct ib_client ipoib_client = {
170 	.name   = "ipoib",
171 	.add    = ipoib_add_one,
172 	.remove = ipoib_remove_one,
173 	.get_net_dev_by_params = ipoib_get_net_dev_by_params,
174 };
175 
176 int
177 ipoib_open(struct ipoib_dev_priv *priv)
178 {
179 	struct ifnet *dev = priv->dev;
180 
181 	ipoib_dbg(priv, "bringing up interface\n");
182 
183 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
184 
185 	if (ipoib_pkey_dev_delay_open(priv))
186 		return 0;
187 
188 	if (ipoib_ib_dev_open(priv))
189 		goto err_disable;
190 
191 	if (ipoib_ib_dev_up(priv))
192 		goto err_stop;
193 
194 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
195 		struct ipoib_dev_priv *cpriv;
196 
197 		/* Bring up any child interfaces too */
198 		mutex_lock(&priv->vlan_mutex);
199 		list_for_each_entry(cpriv, &priv->child_intfs, list)
200 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
201 				ipoib_open(cpriv);
202 		mutex_unlock(&priv->vlan_mutex);
203 	}
204 	dev->if_drv_flags |= IFF_DRV_RUNNING;
205 	dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
206 
207 	return 0;
208 
209 err_stop:
210 	ipoib_ib_dev_stop(priv, 1);
211 
212 err_disable:
213 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
214 
215 	return -EINVAL;
216 }
217 
218 static void
219 ipoib_init(void *arg)
220 {
221 	struct ifnet *dev;
222 	struct ipoib_dev_priv *priv;
223 
224 	priv = arg;
225 	dev = priv->dev;
226 	if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
227 		ipoib_open(priv);
228 	queue_work(ipoib_workqueue, &priv->flush_light);
229 }
230 
231 
232 static int
233 ipoib_stop(struct ipoib_dev_priv *priv)
234 {
235 	struct ifnet *dev = priv->dev;
236 
237 	ipoib_dbg(priv, "stopping interface\n");
238 
239 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
240 
241 	dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
242 
243 	ipoib_ib_dev_down(priv, 0);
244 	ipoib_ib_dev_stop(priv, 0);
245 
246 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
247 		struct ipoib_dev_priv *cpriv;
248 
249 		/* Bring down any child interfaces too */
250 		mutex_lock(&priv->vlan_mutex);
251 		list_for_each_entry(cpriv, &priv->child_intfs, list)
252 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
253 				ipoib_stop(cpriv);
254 		mutex_unlock(&priv->vlan_mutex);
255 	}
256 
257 	return 0;
258 }
259 
260 static int
261 ipoib_propagate_ifnet_mtu(struct ipoib_dev_priv *priv, int new_mtu,
262     bool propagate)
263 {
264 	struct ifnet *ifp;
265 	struct ifreq ifr;
266 	int error;
267 
268 	ifp = priv->dev;
269 	if (ifp->if_mtu == new_mtu)
270 		return (0);
271 	if (propagate) {
272 		strlcpy(ifr.ifr_name, if_name(ifp), IFNAMSIZ);
273 		ifr.ifr_mtu = new_mtu;
274 		CURVNET_SET(ifp->if_vnet);
275 		error = ifhwioctl(SIOCSIFMTU, ifp, (caddr_t)&ifr, curthread);
276 		CURVNET_RESTORE();
277 	} else {
278 		ifp->if_mtu = new_mtu;
279 		error = 0;
280 	}
281 	return (error);
282 }
283 
284 int
285 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu, bool propagate)
286 {
287 	int error, prev_admin_mtu;
288 
289 	/* dev->if_mtu > 2K ==> connected mode */
290 	if (ipoib_cm_admin_enabled(priv)) {
291 		if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
292 			return -EINVAL;
293 
294 		if (new_mtu > priv->mcast_mtu)
295 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
296 				   priv->mcast_mtu);
297 
298 		return (ipoib_propagate_ifnet_mtu(priv, new_mtu, propagate));
299 	}
300 
301 	if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
302 		return -EINVAL;
303 
304 	prev_admin_mtu = priv->admin_mtu;
305 	priv->admin_mtu = new_mtu;
306 	error = ipoib_propagate_ifnet_mtu(priv, min(priv->mcast_mtu,
307 	    priv->admin_mtu), propagate);
308 	if (error == 0) {
309 		/* check for MTU change to avoid infinite loop */
310 		if (prev_admin_mtu != new_mtu)
311 			queue_work(ipoib_workqueue, &priv->flush_light);
312 	} else
313 		priv->admin_mtu = prev_admin_mtu;
314 	return (error);
315 }
316 
317 static int
318 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
319 {
320 	struct ipoib_dev_priv *priv = ifp->if_softc;
321 	struct ifaddr *ifa = (struct ifaddr *) data;
322 	struct ifreq *ifr = (struct ifreq *) data;
323 	int error = 0;
324 
325 	/* check if detaching */
326 	if (priv == NULL || priv->gone != 0)
327 		return (ENXIO);
328 
329 	switch (command) {
330 	case SIOCSIFFLAGS:
331 		if (ifp->if_flags & IFF_UP) {
332 			if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
333 				error = -ipoib_open(priv);
334 		} else
335 			if (ifp->if_drv_flags & IFF_DRV_RUNNING)
336 				ipoib_stop(priv);
337 		break;
338 	case SIOCADDMULTI:
339 	case SIOCDELMULTI:
340 		if (ifp->if_drv_flags & IFF_DRV_RUNNING)
341 			queue_work(ipoib_workqueue, &priv->restart_task);
342 		break;
343 	case SIOCSIFADDR:
344 		ifp->if_flags |= IFF_UP;
345 
346 		switch (ifa->ifa_addr->sa_family) {
347 #ifdef INET
348 		case AF_INET:
349 			ifp->if_init(ifp->if_softc);	/* before arpwhohas */
350 			arp_ifinit(ifp, ifa);
351 			break;
352 #endif
353 		default:
354 			ifp->if_init(ifp->if_softc);
355 			break;
356 		}
357 		break;
358 
359 	case SIOCGIFADDR:
360 			bcopy(IF_LLADDR(ifp), &ifr->ifr_addr.sa_data[0],
361                             INFINIBAND_ALEN);
362 		break;
363 
364 	case SIOCSIFMTU:
365 		/*
366 		 * Set the interface MTU.
367 		 */
368 		error = -ipoib_change_mtu(priv, ifr->ifr_mtu, false);
369 		break;
370 	default:
371 		error = EINVAL;
372 		break;
373 	}
374 	return (error);
375 }
376 
377 
378 static struct ipoib_path *
379 __path_find(struct ipoib_dev_priv *priv, void *gid)
380 {
381 	struct rb_node *n = priv->path_tree.rb_node;
382 	struct ipoib_path *path;
383 	int ret;
384 
385 	while (n) {
386 		path = rb_entry(n, struct ipoib_path, rb_node);
387 
388 		ret = memcmp(gid, path->pathrec.dgid.raw,
389 			     sizeof (union ib_gid));
390 
391 		if (ret < 0)
392 			n = n->rb_left;
393 		else if (ret > 0)
394 			n = n->rb_right;
395 		else
396 			return path;
397 	}
398 
399 	return NULL;
400 }
401 
402 static int
403 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
404 {
405 	struct rb_node **n = &priv->path_tree.rb_node;
406 	struct rb_node *pn = NULL;
407 	struct ipoib_path *tpath;
408 	int ret;
409 
410 	while (*n) {
411 		pn = *n;
412 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
413 
414 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
415 			     sizeof (union ib_gid));
416 		if (ret < 0)
417 			n = &pn->rb_left;
418 		else if (ret > 0)
419 			n = &pn->rb_right;
420 		else
421 			return -EEXIST;
422 	}
423 
424 	rb_link_node(&path->rb_node, pn, n);
425 	rb_insert_color(&path->rb_node, &priv->path_tree);
426 
427 	list_add_tail(&path->list, &priv->path_list);
428 
429 	return 0;
430 }
431 
432 void
433 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
434 {
435 
436 	_IF_DRAIN(&path->queue);
437 
438 	if (path->ah)
439 		ipoib_put_ah(path->ah);
440 	if (ipoib_cm_get(path))
441 		ipoib_cm_destroy_tx(ipoib_cm_get(path));
442 
443 	kfree(path);
444 }
445 
446 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
447 
448 struct ipoib_path_iter *
449 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
450 {
451 	struct ipoib_path_iter *iter;
452 
453 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
454 	if (!iter)
455 		return NULL;
456 
457 	iter->priv = priv;
458 	memset(iter->path.pathrec.dgid.raw, 0, 16);
459 
460 	if (ipoib_path_iter_next(iter)) {
461 		kfree(iter);
462 		return NULL;
463 	}
464 
465 	return iter;
466 }
467 
468 int
469 ipoib_path_iter_next(struct ipoib_path_iter *iter)
470 {
471 	struct ipoib_dev_priv *priv = iter->priv;
472 	struct rb_node *n;
473 	struct ipoib_path *path;
474 	int ret = 1;
475 
476 	spin_lock_irq(&priv->lock);
477 
478 	n = rb_first(&priv->path_tree);
479 
480 	while (n) {
481 		path = rb_entry(n, struct ipoib_path, rb_node);
482 
483 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
484 			   sizeof (union ib_gid)) < 0) {
485 			iter->path = *path;
486 			ret = 0;
487 			break;
488 		}
489 
490 		n = rb_next(n);
491 	}
492 
493 	spin_unlock_irq(&priv->lock);
494 
495 	return ret;
496 }
497 
498 void
499 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
500 {
501 	*path = iter->path;
502 }
503 
504 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
505 
506 void
507 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
508 {
509 	struct ipoib_path *path, *tp;
510 
511 	spin_lock_irq(&priv->lock);
512 
513 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
514 		ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
515 			be16_to_cpu(path->pathrec.dlid),
516 			path->pathrec.dgid.raw, ":");
517 		path->valid =  0;
518 	}
519 
520 	spin_unlock_irq(&priv->lock);
521 }
522 
523 void
524 ipoib_flush_paths(struct ipoib_dev_priv *priv)
525 {
526 	struct ipoib_path *path, *tp;
527 	LIST_HEAD(remove_list);
528 	unsigned long flags;
529 
530 	spin_lock_irqsave(&priv->lock, flags);
531 
532 	list_splice_init(&priv->path_list, &remove_list);
533 
534 	list_for_each_entry(path, &remove_list, list)
535 		rb_erase(&path->rb_node, &priv->path_tree);
536 
537 	list_for_each_entry_safe(path, tp, &remove_list, list) {
538 		if (path->query)
539 			ib_sa_cancel_query(path->query_id, path->query);
540 		spin_unlock_irqrestore(&priv->lock, flags);
541 		wait_for_completion(&path->done);
542 		ipoib_path_free(priv, path);
543 		spin_lock_irqsave(&priv->lock, flags);
544 	}
545 
546 	spin_unlock_irqrestore(&priv->lock, flags);
547 }
548 
549 static void
550 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
551 {
552 	struct ipoib_path *path = path_ptr;
553 	struct ipoib_dev_priv *priv = path->priv;
554 	struct ifnet *dev = priv->dev;
555 	struct ipoib_ah *ah = NULL;
556 	struct ipoib_ah *old_ah = NULL;
557 	struct ifqueue mbqueue;
558 	struct mbuf *mb;
559 	unsigned long flags;
560 
561 	if (!status)
562 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
563 			  be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
564 	else
565 		ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
566 			  status, path->pathrec.dgid.raw, ":");
567 
568 	bzero(&mbqueue, sizeof(mbqueue));
569 
570 	if (!status) {
571 		struct ib_ah_attr av;
572 
573 		if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
574 			ah = ipoib_create_ah(priv, priv->pd, &av);
575 	}
576 
577 	spin_lock_irqsave(&priv->lock, flags);
578 
579 	if (ah) {
580 		path->pathrec = *pathrec;
581 
582 		old_ah   = path->ah;
583 		path->ah = ah;
584 
585 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
586 			  ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
587 
588 		for (;;) {
589 			_IF_DEQUEUE(&path->queue, mb);
590 			if (mb == NULL)
591 				break;
592 			_IF_ENQUEUE(&mbqueue, mb);
593 		}
594 
595 #ifdef CONFIG_INFINIBAND_IPOIB_CM
596 		if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
597 			ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
598 #endif
599 
600 		path->valid = 1;
601 	}
602 
603 	path->query = NULL;
604 	complete(&path->done);
605 
606 	spin_unlock_irqrestore(&priv->lock, flags);
607 
608 	if (old_ah)
609 		ipoib_put_ah(old_ah);
610 
611 	for (;;) {
612 		_IF_DEQUEUE(&mbqueue, mb);
613 		if (mb == NULL)
614 			break;
615 		mb->m_pkthdr.rcvif = dev;
616 		if (dev->if_transmit(dev, mb))
617 			ipoib_warn(priv, "dev_queue_xmit failed "
618 				   "to requeue packet\n");
619 	}
620 }
621 
622 static struct ipoib_path *
623 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
624 {
625 	struct ipoib_path *path;
626 
627 	if (!priv->broadcast)
628 		return NULL;
629 
630 	path = kzalloc(sizeof *path, GFP_ATOMIC);
631 	if (!path)
632 		return NULL;
633 
634 	path->priv = priv;
635 
636 	bzero(&path->queue, sizeof(path->queue));
637 
638 #ifdef CONFIG_INFINIBAND_IPOIB_CM
639 	memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
640 #endif
641 	memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
642 	path->pathrec.sgid	    = priv->local_gid;
643 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
644 	path->pathrec.numb_path     = 1;
645 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
646 
647 	return path;
648 }
649 
650 static int
651 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
652 {
653 	struct ifnet *dev = priv->dev;
654 
655 	ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
656 	struct ib_sa_path_rec p_rec;
657 
658 	p_rec = path->pathrec;
659 	p_rec.mtu_selector = IB_SA_GT;
660 
661 	switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
662 	case 512:
663 		p_rec.mtu = IB_MTU_256;
664 		break;
665 	case 1024:
666 		p_rec.mtu = IB_MTU_512;
667 		break;
668 	case 2048:
669 		p_rec.mtu = IB_MTU_1024;
670 		break;
671 	case 4096:
672 		p_rec.mtu = IB_MTU_2048;
673 		break;
674 	default:
675 		/* Wildcard everything */
676 		comp_mask = 0;
677 		p_rec.mtu = 0;
678 		p_rec.mtu_selector = 0;
679 	}
680 
681 	ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
682 		  p_rec.dgid.raw, ":",
683 		  comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
684 
685 	init_completion(&path->done);
686 
687 	path->query_id =
688 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
689 				   &p_rec, comp_mask		|
690 				   IB_SA_PATH_REC_DGID		|
691 				   IB_SA_PATH_REC_SGID		|
692 				   IB_SA_PATH_REC_NUMB_PATH	|
693 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
694 				   IB_SA_PATH_REC_PKEY,
695 				   1000, GFP_ATOMIC,
696 				   path_rec_completion,
697 				   path, &path->query);
698 	if (path->query_id < 0) {
699 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
700 		path->query = NULL;
701 		complete(&path->done);
702 		return path->query_id;
703 	}
704 
705 	return 0;
706 }
707 
708 static void
709 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
710 {
711 	struct ipoib_path *path;
712 
713 	path = __path_find(priv, eh->hwaddr + 4);
714 	if (!path || !path->valid) {
715 		int new_path = 0;
716 
717 		if (!path) {
718 			path = path_rec_create(priv, eh->hwaddr);
719 			new_path = 1;
720 		}
721 		if (path) {
722 			if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
723 				_IF_ENQUEUE(&path->queue, mb);
724 			else {
725 				if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
726 				m_freem(mb);
727 			}
728 
729 			if (!path->query && path_rec_start(priv, path)) {
730 				spin_unlock_irqrestore(&priv->lock, flags);
731 				if (new_path)
732 					ipoib_path_free(priv, path);
733 				return;
734 			} else
735 				__path_add(priv, path);
736 		} else {
737 			if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
738 			m_freem(mb);
739 		}
740 
741 		return;
742 	}
743 
744 	if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
745 		ipoib_cm_send(priv, mb, ipoib_cm_get(path));
746 	} else if (path->ah) {
747 		ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
748 	} else if ((path->query || !path_rec_start(priv, path)) &&
749 		    path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
750 		_IF_ENQUEUE(&path->queue, mb);
751 	} else {
752 		if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
753 		m_freem(mb);
754 	}
755 }
756 
757 static int
758 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
759 {
760 	struct ipoib_header *eh;
761 
762 	eh = mtod(mb, struct ipoib_header *);
763 	if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
764 		/* Add in the P_Key for multicast*/
765 		eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
766 		eh->hwaddr[9] = priv->pkey & 0xff;
767 
768 		ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
769 	} else
770 		ipoib_unicast_send(mb, priv, eh);
771 
772 	return 0;
773 }
774 
775 
776 static void
777 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
778 {
779 	struct mbuf *mb;
780 
781 	if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
782 	    IFF_DRV_RUNNING)
783 		return;
784 
785 	spin_lock(&priv->lock);
786 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
787 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
788 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
789 		if (mb == NULL)
790 			break;
791 		IPOIB_MTAP(dev, mb);
792 		ipoib_send_one(priv, mb);
793 	}
794 	spin_unlock(&priv->lock);
795 }
796 
797 static void
798 ipoib_start(struct ifnet *dev)
799 {
800 	_ipoib_start(dev, dev->if_softc);
801 }
802 
803 static void
804 ipoib_vlan_start(struct ifnet *dev)
805 {
806 	struct ipoib_dev_priv *priv;
807 	struct mbuf *mb;
808 
809 	priv = VLAN_COOKIE(dev);
810 	if (priv != NULL)
811 		return _ipoib_start(dev, priv);
812 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
813 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
814 		if (mb == NULL)
815 			break;
816 		m_freem(mb);
817 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
818 	}
819 }
820 
821 int
822 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
823 {
824 
825 	/* Allocate RX/TX "rings" to hold queued mbs */
826 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
827 				GFP_KERNEL);
828 	if (!priv->rx_ring) {
829 		printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
830 		       ca->name, ipoib_recvq_size);
831 		goto out;
832 	}
833 
834 	priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
835 	if (!priv->tx_ring) {
836 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
837 		       ca->name, ipoib_sendq_size);
838 		goto out_rx_ring_cleanup;
839 	}
840 	memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
841 
842 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
843 
844 	if (ipoib_ib_dev_init(priv, ca, port))
845 		goto out_tx_ring_cleanup;
846 
847 	return 0;
848 
849 out_tx_ring_cleanup:
850 	kfree(priv->tx_ring);
851 
852 out_rx_ring_cleanup:
853 	kfree(priv->rx_ring);
854 
855 out:
856 	return -ENOMEM;
857 }
858 
859 static void
860 ipoib_detach(struct ipoib_dev_priv *priv)
861 {
862 	struct ifnet *dev;
863 
864 	dev = priv->dev;
865 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
866 		priv->gone = 1;
867 		bpfdetach(dev);
868 		if_detach(dev);
869 		if_free(dev);
870 		free_unr(ipoib_unrhdr, priv->unit);
871 	} else
872 		VLAN_SETCOOKIE(priv->dev, NULL);
873 
874 	free(priv, M_TEMP);
875 }
876 
877 void
878 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
879 {
880 	struct ipoib_dev_priv *cpriv, *tcpriv;
881 
882 	/* Delete any child interfaces first */
883 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
884 		ipoib_dev_cleanup(cpriv);
885 		ipoib_detach(cpriv);
886 	}
887 
888 	ipoib_ib_dev_cleanup(priv);
889 
890 	kfree(priv->rx_ring);
891 	kfree(priv->tx_ring);
892 
893 	priv->rx_ring = NULL;
894 	priv->tx_ring = NULL;
895 }
896 
897 static struct ipoib_dev_priv *
898 ipoib_priv_alloc(void)
899 {
900 	struct ipoib_dev_priv *priv;
901 
902 	priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
903 	spin_lock_init(&priv->lock);
904 	spin_lock_init(&priv->drain_lock);
905 	mutex_init(&priv->vlan_mutex);
906 	INIT_LIST_HEAD(&priv->path_list);
907 	INIT_LIST_HEAD(&priv->child_intfs);
908 	INIT_LIST_HEAD(&priv->dead_ahs);
909 	INIT_LIST_HEAD(&priv->multicast_list);
910 	INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
911 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
912 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
913 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
914 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
915 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
916 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
917 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
918 	memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
919 
920 	return (priv);
921 }
922 
923 struct ipoib_dev_priv *
924 ipoib_intf_alloc(const char *name)
925 {
926 	struct ipoib_dev_priv *priv;
927 	struct sockaddr_dl *sdl;
928 	struct ifnet *dev;
929 
930 	priv = ipoib_priv_alloc();
931 	dev = priv->dev = if_alloc(IFT_INFINIBAND);
932 	if (!dev) {
933 		free(priv, M_TEMP);
934 		return NULL;
935 	}
936 	dev->if_softc = priv;
937 	priv->unit = alloc_unr(ipoib_unrhdr);
938 	if (priv->unit == -1) {
939 		if_free(dev);
940 		free(priv, M_TEMP);
941 		return NULL;
942 	}
943 	if_initname(dev, name, priv->unit);
944 	dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
945 	dev->if_addrlen = INFINIBAND_ALEN;
946 	dev->if_hdrlen = IPOIB_HEADER_LEN;
947 	if_attach(dev);
948 	dev->if_init = ipoib_init;
949 	dev->if_ioctl = ipoib_ioctl;
950 	dev->if_start = ipoib_start;
951 	dev->if_output = ipoib_output;
952 	dev->if_input = ipoib_input;
953 	dev->if_resolvemulti = ipoib_resolvemulti;
954 	dev->if_baudrate = IF_Gbps(10);
955 	dev->if_broadcastaddr = priv->broadcastaddr;
956 	dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
957 	sdl = (struct sockaddr_dl *)dev->if_addr->ifa_addr;
958 	sdl->sdl_type = IFT_INFINIBAND;
959 	sdl->sdl_alen = dev->if_addrlen;
960 	priv->dev = dev;
961 	if_link_state_change(dev, LINK_STATE_DOWN);
962 	bpfattach(dev, DLT_EN10MB, ETHER_HDR_LEN);
963 
964 	return dev->if_softc;
965 }
966 
967 int
968 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
969 {
970 	struct ib_device_attr *device_attr = &hca->attrs;
971 
972 	priv->hca_caps = device_attr->device_cap_flags;
973 
974 	priv->dev->if_hwassist = 0;
975 	priv->dev->if_capabilities = 0;
976 
977 #ifndef CONFIG_INFINIBAND_IPOIB_CM
978 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
979 		set_bit(IPOIB_FLAG_CSUM, &priv->flags);
980 		priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
981 		priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
982 	}
983 
984 #if 0
985 	if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
986 		priv->dev->if_capabilities |= IFCAP_TSO4;
987 		priv->dev->if_hwassist |= CSUM_TSO;
988 	}
989 #endif
990 #endif
991 	priv->dev->if_capabilities |=
992 	    IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
993 	priv->dev->if_capenable = priv->dev->if_capabilities;
994 
995 	return 0;
996 }
997 
998 
999 static struct ifnet *
1000 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
1001 {
1002 	struct ipoib_dev_priv *priv;
1003 	struct ib_port_attr attr;
1004 	int result = -ENOMEM;
1005 
1006 	priv = ipoib_intf_alloc(format);
1007 	if (!priv)
1008 		goto alloc_mem_failed;
1009 
1010 	if (!ib_query_port(hca, port, &attr))
1011 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1012 	else {
1013 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1014 		       hca->name, port);
1015 		goto device_init_failed;
1016 	}
1017 
1018 	/* MTU will be reset when mcast join happens */
1019 	priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1020 	priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
1021 
1022 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
1023 	if (result) {
1024 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1025 		       hca->name, port, result);
1026 		goto device_init_failed;
1027 	}
1028 
1029 	if (ipoib_set_dev_features(priv, hca))
1030 		goto device_init_failed;
1031 
1032 	/*
1033 	 * Set the full membership bit, so that we join the right
1034 	 * broadcast group, etc.
1035 	 */
1036 	priv->pkey |= 0x8000;
1037 
1038 	priv->broadcastaddr[8] = priv->pkey >> 8;
1039 	priv->broadcastaddr[9] = priv->pkey & 0xff;
1040 
1041 	result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1042 	if (result) {
1043 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1044 		       hca->name, port, result);
1045 		goto device_init_failed;
1046 	}
1047 	memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
1048 
1049 	result = ipoib_dev_init(priv, hca, port);
1050 	if (result < 0) {
1051 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1052 		       hca->name, port, result);
1053 		goto device_init_failed;
1054 	}
1055 	if (ipoib_cm_admin_enabled(priv))
1056 		priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1057 
1058 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
1059 			      priv->ca, ipoib_event);
1060 	result = ib_register_event_handler(&priv->event_handler);
1061 	if (result < 0) {
1062 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1063 		       "port %d (ret = %d)\n",
1064 		       hca->name, port, result);
1065 		goto event_failed;
1066 	}
1067 	if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1068 
1069 	return priv->dev;
1070 
1071 event_failed:
1072 	ipoib_dev_cleanup(priv);
1073 
1074 device_init_failed:
1075 	ipoib_detach(priv);
1076 
1077 alloc_mem_failed:
1078 	return ERR_PTR(result);
1079 }
1080 
1081 static void
1082 ipoib_add_one(struct ib_device *device)
1083 {
1084 	struct list_head *dev_list;
1085 	struct ifnet *dev;
1086 	struct ipoib_dev_priv *priv;
1087 	int s, e, p;
1088 
1089 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1090 		return;
1091 
1092 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1093 	if (!dev_list)
1094 		return;
1095 
1096 	INIT_LIST_HEAD(dev_list);
1097 
1098 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
1099 		s = 0;
1100 		e = 0;
1101 	} else {
1102 		s = 1;
1103 		e = device->phys_port_cnt;
1104 	}
1105 
1106 	for (p = s; p <= e; ++p) {
1107 		if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1108 			continue;
1109 		dev = ipoib_add_port("ib", device, p);
1110 		if (!IS_ERR(dev)) {
1111 			priv = dev->if_softc;
1112 			list_add_tail(&priv->list, dev_list);
1113 		}
1114 	}
1115 
1116 	ib_set_client_data(device, &ipoib_client, dev_list);
1117 }
1118 
1119 static void
1120 ipoib_remove_one(struct ib_device *device, void *client_data)
1121 {
1122 	struct ipoib_dev_priv *priv, *tmp;
1123 	struct list_head *dev_list = client_data;
1124 
1125 	if (!dev_list)
1126 		return;
1127 
1128 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1129 		return;
1130 
1131 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
1132 		if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1133 			continue;
1134 
1135 		ipoib_stop(priv);
1136 
1137 		ib_unregister_event_handler(&priv->event_handler);
1138 
1139 		/* dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP); */
1140 
1141 		flush_workqueue(ipoib_workqueue);
1142 
1143 		ipoib_dev_cleanup(priv);
1144 		ipoib_detach(priv);
1145 	}
1146 
1147 	kfree(dev_list);
1148 }
1149 
1150 static int
1151 ipoib_match_dev_addr(const struct sockaddr *addr, struct net_device *dev)
1152 {
1153 	struct epoch_tracker et;
1154 	struct ifaddr *ifa;
1155 	int retval = 0;
1156 
1157 	CURVNET_SET(dev->if_vnet);
1158 	NET_EPOCH_ENTER(et);
1159 	CK_STAILQ_FOREACH(ifa, &dev->if_addrhead, ifa_link) {
1160 		if (ifa->ifa_addr == NULL ||
1161 		    ifa->ifa_addr->sa_family != addr->sa_family ||
1162 		    ifa->ifa_addr->sa_len != addr->sa_len) {
1163 			continue;
1164 		}
1165 		if (memcmp(ifa->ifa_addr, addr, addr->sa_len) == 0) {
1166 			retval = 1;
1167 			break;
1168 		}
1169 	}
1170 	NET_EPOCH_EXIT(et);
1171 	CURVNET_RESTORE();
1172 
1173 	return (retval);
1174 }
1175 
1176 /*
1177  * ipoib_match_gid_pkey_addr - returns the number of IPoIB netdevs on
1178  * top a given ipoib device matching a pkey_index and address, if one
1179  * exists.
1180  *
1181  * @found_net_dev: contains a matching net_device if the return value
1182  * >= 1, with a reference held.
1183  */
1184 static int
1185 ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
1186     const union ib_gid *gid, u16 pkey_index, const struct sockaddr *addr,
1187     struct net_device **found_net_dev)
1188 {
1189 	struct ipoib_dev_priv *child_priv;
1190 	int matches = 0;
1191 
1192 	if (priv->pkey_index == pkey_index &&
1193 	    (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
1194 		if (addr == NULL || ipoib_match_dev_addr(addr, priv->dev) != 0) {
1195 			if (*found_net_dev == NULL) {
1196 				struct net_device *net_dev;
1197 
1198 				if (priv->parent != NULL)
1199 					net_dev = priv->parent;
1200 				else
1201 					net_dev = priv->dev;
1202 				*found_net_dev = net_dev;
1203 				dev_hold(net_dev);
1204 			}
1205 			matches++;
1206 		}
1207 	}
1208 
1209 	/* Check child interfaces */
1210 	mutex_lock(&priv->vlan_mutex);
1211 	list_for_each_entry(child_priv, &priv->child_intfs, list) {
1212 		matches += ipoib_match_gid_pkey_addr(child_priv, gid,
1213 		    pkey_index, addr, found_net_dev);
1214 		if (matches > 1)
1215 			break;
1216 	}
1217 	mutex_unlock(&priv->vlan_mutex);
1218 
1219 	return matches;
1220 }
1221 
1222 /*
1223  * __ipoib_get_net_dev_by_params - returns the number of matching
1224  * net_devs found (between 0 and 2). Also return the matching
1225  * net_device in the @net_dev parameter, holding a reference to the
1226  * net_device, if the number of matches >= 1
1227  */
1228 static int
1229 __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
1230     u16 pkey_index, const union ib_gid *gid,
1231     const struct sockaddr *addr, struct net_device **net_dev)
1232 {
1233 	struct ipoib_dev_priv *priv;
1234 	int matches = 0;
1235 
1236 	*net_dev = NULL;
1237 
1238 	list_for_each_entry(priv, dev_list, list) {
1239 		if (priv->port != port)
1240 			continue;
1241 
1242 		matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
1243 		    addr, net_dev);
1244 
1245 		if (matches > 1)
1246 			break;
1247 	}
1248 
1249 	return matches;
1250 }
1251 
1252 static struct net_device *
1253 ipoib_get_net_dev_by_params(struct ib_device *dev, u8 port, u16 pkey,
1254     const union ib_gid *gid, const struct sockaddr *addr, void *client_data)
1255 {
1256 	struct net_device *net_dev;
1257 	struct list_head *dev_list = client_data;
1258 	u16 pkey_index;
1259 	int matches;
1260 	int ret;
1261 
1262 	if (!rdma_protocol_ib(dev, port))
1263 		return NULL;
1264 
1265 	ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
1266 	if (ret)
1267 		return NULL;
1268 
1269 	if (!dev_list)
1270 		return NULL;
1271 
1272 	/* See if we can find a unique device matching the L2 parameters */
1273 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1274 						gid, NULL, &net_dev);
1275 
1276 	switch (matches) {
1277 	case 0:
1278 		return NULL;
1279 	case 1:
1280 		return net_dev;
1281 	}
1282 
1283 	dev_put(net_dev);
1284 
1285 	/* Couldn't find a unique device with L2 parameters only. Use L3
1286 	 * address to uniquely match the net device */
1287 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1288 						gid, addr, &net_dev);
1289 	switch (matches) {
1290 	case 0:
1291 		return NULL;
1292 	default:
1293 		dev_warn_ratelimited(&dev->dev,
1294 				     "duplicate IP address detected\n");
1295 		/* Fall through */
1296 	case 1:
1297 		return net_dev;
1298 	}
1299 }
1300 
1301 static void
1302 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1303 {
1304 	struct ipoib_dev_priv *parent;
1305 	struct ipoib_dev_priv *priv;
1306 	struct ifnet *dev;
1307 	uint16_t pkey;
1308 	int error;
1309 
1310 	if (ifp->if_type != IFT_INFINIBAND)
1311 		return;
1312 	dev = VLAN_DEVAT(ifp, vtag);
1313 	if (dev == NULL)
1314 		return;
1315 	priv = NULL;
1316 	error = 0;
1317 	parent = ifp->if_softc;
1318 	/* We only support 15 bits of pkey. */
1319 	if (vtag & 0x8000)
1320 		return;
1321 	pkey = vtag | 0x8000;	/* Set full membership bit. */
1322 	if (pkey == parent->pkey)
1323 		return;
1324 	/* Check for dups */
1325 	mutex_lock(&parent->vlan_mutex);
1326 	list_for_each_entry(priv, &parent->child_intfs, list) {
1327 		if (priv->pkey == pkey) {
1328 			priv = NULL;
1329 			error = EBUSY;
1330 			goto out;
1331 		}
1332 	}
1333 	priv = ipoib_priv_alloc();
1334 	priv->dev = dev;
1335 	priv->max_ib_mtu = parent->max_ib_mtu;
1336 	priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1337 	set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1338 	error = ipoib_set_dev_features(priv, parent->ca);
1339 	if (error)
1340 		goto out;
1341 	priv->pkey = pkey;
1342 	priv->broadcastaddr[8] = pkey >> 8;
1343 	priv->broadcastaddr[9] = pkey & 0xff;
1344 	dev->if_broadcastaddr = priv->broadcastaddr;
1345 	error = ipoib_dev_init(priv, parent->ca, parent->port);
1346 	if (error)
1347 		goto out;
1348 	priv->parent = parent->dev;
1349 	list_add_tail(&priv->list, &parent->child_intfs);
1350 	VLAN_SETCOOKIE(dev, priv);
1351 	dev->if_start = ipoib_vlan_start;
1352 	dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1353 	dev->if_hdrlen = IPOIB_HEADER_LEN;
1354 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1355 		ipoib_open(priv);
1356 	mutex_unlock(&parent->vlan_mutex);
1357 	return;
1358 out:
1359 	mutex_unlock(&parent->vlan_mutex);
1360 	if (priv)
1361 		free(priv, M_TEMP);
1362 	if (error)
1363 		ipoib_warn(parent,
1364 		    "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1365 		    parent->ca->name, parent->port, vtag);
1366 	return;
1367 }
1368 
1369 static void
1370 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1371 {
1372 	struct ipoib_dev_priv *parent;
1373 	struct ipoib_dev_priv *priv;
1374 	struct ifnet *dev;
1375 	uint16_t pkey;
1376 
1377 	if (ifp->if_type != IFT_INFINIBAND)
1378 		return;
1379 
1380 	dev = VLAN_DEVAT(ifp, vtag);
1381 	if (dev)
1382 		VLAN_SETCOOKIE(dev, NULL);
1383 	pkey = vtag | 0x8000;
1384 	parent = ifp->if_softc;
1385 	mutex_lock(&parent->vlan_mutex);
1386 	list_for_each_entry(priv, &parent->child_intfs, list) {
1387 		if (priv->pkey == pkey) {
1388 			ipoib_dev_cleanup(priv);
1389 			list_del(&priv->list);
1390 			break;
1391 		}
1392 	}
1393 	mutex_unlock(&parent->vlan_mutex);
1394 }
1395 
1396 eventhandler_tag ipoib_vlan_attach;
1397 eventhandler_tag ipoib_vlan_detach;
1398 
1399 static int __init
1400 ipoib_init_module(void)
1401 {
1402 	int ret;
1403 
1404 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1405 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1406 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1407 
1408 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1409 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1410 	ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1411 						     IPOIB_MIN_QUEUE_SIZE));
1412 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1413 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1414 #endif
1415 
1416 	ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1417 		ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1418 	ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1419 		ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1420 
1421 	/*
1422 	 * We create our own workqueue mainly because we want to be
1423 	 * able to flush it when devices are being removed.  We can't
1424 	 * use schedule_work()/flush_scheduled_work() because both
1425 	 * unregister_netdev() and linkwatch_event take the rtnl lock,
1426 	 * so flush_scheduled_work() can deadlock during device
1427 	 * removal.
1428 	 */
1429 	ipoib_workqueue = create_singlethread_workqueue("ipoib");
1430 	if (!ipoib_workqueue) {
1431 		ret = -ENOMEM;
1432 		goto err_fs;
1433 	}
1434 
1435 	ib_sa_register_client(&ipoib_sa_client);
1436 
1437 	ret = ib_register_client(&ipoib_client);
1438 	if (ret)
1439 		goto err_sa;
1440 
1441 	return 0;
1442 
1443 err_sa:
1444 	ib_sa_unregister_client(&ipoib_sa_client);
1445 	destroy_workqueue(ipoib_workqueue);
1446 
1447 err_fs:
1448 	return ret;
1449 }
1450 
1451 static void __exit
1452 ipoib_cleanup_module(void)
1453 {
1454 
1455 	EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1456 	EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1457 	ib_unregister_client(&ipoib_client);
1458 	ib_sa_unregister_client(&ipoib_sa_client);
1459 	destroy_workqueue(ipoib_workqueue);
1460 }
1461 
1462 /*
1463  * Infiniband output routine.
1464  */
1465 static int
1466 ipoib_output(struct ifnet *ifp, struct mbuf *m,
1467 	const struct sockaddr *dst, struct route *ro)
1468 {
1469 	u_char edst[INFINIBAND_ALEN];
1470 #if defined(INET) || defined(INET6)
1471 	struct llentry *lle = NULL;
1472 #endif
1473 	struct ipoib_header *eh;
1474 	int error = 0, is_gw = 0;
1475 	short type;
1476 
1477 	if (ro != NULL)
1478 		is_gw = (ro->ro_flags & RT_HAS_GW) != 0;
1479 #ifdef MAC
1480 	error = mac_ifnet_check_transmit(ifp, m);
1481 	if (error)
1482 		goto bad;
1483 #endif
1484 
1485 	M_PROFILE(m);
1486 	if (ifp->if_flags & IFF_MONITOR) {
1487 		error = ENETDOWN;
1488 		goto bad;
1489 	}
1490 	if (!((ifp->if_flags & IFF_UP) &&
1491 	    (ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1492 		error = ENETDOWN;
1493 		goto bad;
1494 	}
1495 
1496 	switch (dst->sa_family) {
1497 #ifdef INET
1498 	case AF_INET:
1499 		if (lle != NULL && (lle->la_flags & LLE_VALID))
1500 			memcpy(edst, lle->ll_addr, sizeof(edst));
1501 		else if (m->m_flags & M_MCAST)
1502 			ip_ib_mc_map(((struct sockaddr_in *)dst)->sin_addr.s_addr, ifp->if_broadcastaddr, edst);
1503 		else
1504 			error = arpresolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1505 		if (error)
1506 			return (error == EWOULDBLOCK ? 0 : error);
1507 		type = htons(ETHERTYPE_IP);
1508 		break;
1509 	case AF_ARP:
1510 	{
1511 		struct arphdr *ah;
1512 		ah = mtod(m, struct arphdr *);
1513 		ah->ar_hrd = htons(ARPHRD_INFINIBAND);
1514 
1515 		switch(ntohs(ah->ar_op)) {
1516 		case ARPOP_REVREQUEST:
1517 		case ARPOP_REVREPLY:
1518 			type = htons(ETHERTYPE_REVARP);
1519 			break;
1520 		case ARPOP_REQUEST:
1521 		case ARPOP_REPLY:
1522 		default:
1523 			type = htons(ETHERTYPE_ARP);
1524 			break;
1525 		}
1526 
1527 		if (m->m_flags & M_BCAST)
1528 			bcopy(ifp->if_broadcastaddr, edst, INFINIBAND_ALEN);
1529 		else
1530 			bcopy(ar_tha(ah), edst, INFINIBAND_ALEN);
1531 
1532 	}
1533 	break;
1534 #endif
1535 #ifdef INET6
1536 	case AF_INET6:
1537 		if (lle != NULL && (lle->la_flags & LLE_VALID))
1538 			memcpy(edst, lle->ll_addr, sizeof(edst));
1539 		else if (m->m_flags & M_MCAST)
1540 			ipv6_ib_mc_map(&((struct sockaddr_in6 *)dst)->sin6_addr, ifp->if_broadcastaddr, edst);
1541 		else
1542 			error = nd6_resolve(ifp, is_gw, m, dst, edst, NULL, NULL);
1543 		if (error)
1544 			return error;
1545 		type = htons(ETHERTYPE_IPV6);
1546 		break;
1547 #endif
1548 
1549 	default:
1550 		if_printf(ifp, "can't handle af%d\n", dst->sa_family);
1551 		error = EAFNOSUPPORT;
1552 		goto bad;
1553 	}
1554 
1555 	/*
1556 	 * Add local net header.  If no space in first mbuf,
1557 	 * allocate another.
1558 	 */
1559 	M_PREPEND(m, IPOIB_HEADER_LEN, M_NOWAIT);
1560 	if (m == NULL) {
1561 		error = ENOBUFS;
1562 		goto bad;
1563 	}
1564 	eh = mtod(m, struct ipoib_header *);
1565 	(void)memcpy(&eh->proto, &type, sizeof(eh->proto));
1566 	(void)memcpy(&eh->hwaddr, edst, sizeof (edst));
1567 
1568 	/*
1569 	 * Queue message on interface, update output statistics if
1570 	 * successful, and start output if interface not yet active.
1571 	 */
1572 	return ((ifp->if_transmit)(ifp, m));
1573 bad:
1574 	if (m != NULL)
1575 		m_freem(m);
1576 	return (error);
1577 }
1578 
1579 /*
1580  * Upper layer processing for a received Infiniband packet.
1581  */
1582 void
1583 ipoib_demux(struct ifnet *ifp, struct mbuf *m, u_short proto)
1584 {
1585 	int isr;
1586 
1587 #ifdef MAC
1588 	/*
1589 	 * Tag the mbuf with an appropriate MAC label before any other
1590 	 * consumers can get to it.
1591 	 */
1592 	mac_ifnet_create_mbuf(ifp, m);
1593 #endif
1594 	/* Allow monitor mode to claim this frame, after stats are updated. */
1595 	if (ifp->if_flags & IFF_MONITOR) {
1596 		if_printf(ifp, "discard frame at IFF_MONITOR\n");
1597 		m_freem(m);
1598 		return;
1599 	}
1600 	/*
1601 	 * Dispatch frame to upper layer.
1602 	 */
1603 	switch (proto) {
1604 #ifdef INET
1605 	case ETHERTYPE_IP:
1606 		isr = NETISR_IP;
1607 		break;
1608 
1609 	case ETHERTYPE_ARP:
1610 		if (ifp->if_flags & IFF_NOARP) {
1611 			/* Discard packet if ARP is disabled on interface */
1612 			m_freem(m);
1613 			return;
1614 		}
1615 		isr = NETISR_ARP;
1616 		break;
1617 #endif
1618 #ifdef INET6
1619 	case ETHERTYPE_IPV6:
1620 		isr = NETISR_IPV6;
1621 		break;
1622 #endif
1623 	default:
1624 		goto discard;
1625 	}
1626 	netisr_dispatch(isr, m);
1627 	return;
1628 
1629 discard:
1630 	m_freem(m);
1631 }
1632 
1633 /*
1634  * Process a received Infiniband packet.
1635  */
1636 static void
1637 ipoib_input(struct ifnet *ifp, struct mbuf *m)
1638 {
1639 	struct ipoib_header *eh;
1640 
1641 	if ((ifp->if_flags & IFF_UP) == 0) {
1642 		m_freem(m);
1643 		return;
1644 	}
1645 	CURVNET_SET_QUIET(ifp->if_vnet);
1646 
1647 	/* Let BPF have it before we strip the header. */
1648 	IPOIB_MTAP(ifp, m);
1649 	eh = mtod(m, struct ipoib_header *);
1650 	/*
1651 	 * Reset layer specific mbuf flags to avoid confusing upper layers.
1652 	 * Strip off Infiniband header.
1653 	 */
1654 	m->m_flags &= ~M_VLANTAG;
1655 	m_clrprotoflags(m);
1656 	m_adj(m, IPOIB_HEADER_LEN);
1657 
1658 	if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
1659 		if (memcmp(eh->hwaddr, ifp->if_broadcastaddr,
1660 		    ifp->if_addrlen) == 0)
1661 			m->m_flags |= M_BCAST;
1662 		else
1663 			m->m_flags |= M_MCAST;
1664 		if_inc_counter(ifp, IFCOUNTER_IMCASTS, 1);
1665 	}
1666 
1667 	ipoib_demux(ifp, m, ntohs(eh->proto));
1668 	CURVNET_RESTORE();
1669 }
1670 
1671 static int
1672 ipoib_resolvemulti(struct ifnet *ifp, struct sockaddr **llsa,
1673 	struct sockaddr *sa)
1674 {
1675 	struct sockaddr_dl *sdl;
1676 #ifdef INET
1677 	struct sockaddr_in *sin;
1678 #endif
1679 #ifdef INET6
1680 	struct sockaddr_in6 *sin6;
1681 #endif
1682 	u_char *e_addr;
1683 
1684 	switch(sa->sa_family) {
1685 	case AF_LINK:
1686 		/*
1687 		 * No mapping needed. Just check that it's a valid MC address.
1688 		 */
1689 		sdl = (struct sockaddr_dl *)sa;
1690 		e_addr = LLADDR(sdl);
1691 		if (!IPOIB_IS_MULTICAST(e_addr))
1692 			return EADDRNOTAVAIL;
1693 		*llsa = NULL;
1694 		return 0;
1695 
1696 #ifdef INET
1697 	case AF_INET:
1698 		sin = (struct sockaddr_in *)sa;
1699 		if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
1700 			return EADDRNOTAVAIL;
1701 		sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1702 		sdl->sdl_alen = INFINIBAND_ALEN;
1703 		e_addr = LLADDR(sdl);
1704 		ip_ib_mc_map(sin->sin_addr.s_addr, ifp->if_broadcastaddr,
1705 		    e_addr);
1706 		*llsa = (struct sockaddr *)sdl;
1707 		return 0;
1708 #endif
1709 #ifdef INET6
1710 	case AF_INET6:
1711 		sin6 = (struct sockaddr_in6 *)sa;
1712 		/*
1713 		 * An IP6 address of 0 means listen to all
1714 		 * of the multicast address used for IP6.
1715 		 * This has no meaning in ipoib.
1716 		 */
1717 		if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
1718 			return EADDRNOTAVAIL;
1719 		if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr))
1720 			return EADDRNOTAVAIL;
1721 		sdl = link_init_sdl(ifp, *llsa, IFT_INFINIBAND);
1722 		sdl->sdl_alen = INFINIBAND_ALEN;
1723 		e_addr = LLADDR(sdl);
1724 		ipv6_ib_mc_map(&sin6->sin6_addr, ifp->if_broadcastaddr, e_addr);
1725 		*llsa = (struct sockaddr *)sdl;
1726 		return 0;
1727 #endif
1728 
1729 	default:
1730 		return EAFNOSUPPORT;
1731 	}
1732 }
1733 
1734 module_init(ipoib_init_module);
1735 module_exit(ipoib_cleanup_module);
1736 
1737 static int
1738 ipoib_evhand(module_t mod, int event, void *arg)
1739 {
1740 	                return (0);
1741 }
1742 
1743 static moduledata_t ipoib_mod = {
1744 	                .name = "ipoib",
1745 			                .evhand = ipoib_evhand,
1746 };
1747 
1748 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1749 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1750 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);
1751