xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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 #include <sys/eventhandler.h>
42 
43 #include <linux/module.h>
44 
45 #include <linux/slab.h>
46 #include <linux/kernel.h>
47 #include <linux/vmalloc.h>
48 
49 #include <linux/if_vlan.h>
50 
51 #include <net/infiniband.h>
52 
53 #include <rdma/ib_addr.h>
54 #include <rdma/ib_cache.h>
55 
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
58 MODULE_LICENSE("Dual BSD/GPL");
59 
60 int ipoib_sendq_size = IPOIB_TX_RING_SIZE;
61 int ipoib_recvq_size = IPOIB_RX_RING_SIZE;
62 
63 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
64 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
65 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
66 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
67 
68 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
69 int ipoib_debug_level = 1;
70 
71 module_param_named(debug_level, ipoib_debug_level, int, 0644);
72 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
73 #endif
74 
75 struct ipoib_path_iter {
76 	struct ipoib_dev_priv *priv;
77 	struct ipoib_path  path;
78 };
79 
80 static const u8 ipv4_bcast_addr[] = {
81 	0x00, 0xff, 0xff, 0xff,
82 	0xff, 0x12, 0x40, 0x1b,	0x00, 0x00, 0x00, 0x00,
83 	0x00, 0x00, 0x00, 0x00,	0xff, 0xff, 0xff, 0xff
84 };
85 
86 struct workqueue_struct *ipoib_workqueue;
87 
88 struct ib_sa_client ipoib_sa_client;
89 
90 static void ipoib_add_one(struct ib_device *device);
91 static void ipoib_remove_one(struct ib_device *device, void *client_data);
92 static struct ifnet *ipoib_get_net_dev_by_params(
93 		struct ib_device *dev, u8 port, u16 pkey,
94 		const union ib_gid *gid, const struct sockaddr *addr,
95 		void *client_data);
96 static void ipoib_start(struct ifnet *dev);
97 static int ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
98 
99 static struct unrhdr *ipoib_unrhdr;
100 
101 static void
102 ipoib_unrhdr_init(void *arg)
103 {
104 
105 	ipoib_unrhdr = new_unrhdr(0, 65535, NULL);
106 }
107 SYSINIT(ipoib_unrhdr_init, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_init, NULL);
108 
109 static void
110 ipoib_unrhdr_uninit(void *arg)
111 {
112 
113 	if (ipoib_unrhdr != NULL) {
114 		struct unrhdr *hdr;
115 
116 		hdr = ipoib_unrhdr;
117 		ipoib_unrhdr = NULL;
118 
119 		delete_unrhdr(hdr);
120 	}
121 }
122 SYSUNINIT(ipoib_unrhdr_uninit, SI_SUB_KLD - 1, SI_ORDER_ANY, ipoib_unrhdr_uninit, NULL);
123 
124 static struct ib_client ipoib_client = {
125 	.name   = "ipoib",
126 	.add    = ipoib_add_one,
127 	.remove = ipoib_remove_one,
128 	.get_net_dev_by_params = ipoib_get_net_dev_by_params,
129 };
130 
131 int
132 ipoib_open(struct ipoib_dev_priv *priv)
133 {
134 	struct ifnet *dev = priv->dev;
135 
136 	ipoib_dbg(priv, "bringing up interface\n");
137 
138 	set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
139 
140 	if (ipoib_pkey_dev_delay_open(priv))
141 		return 0;
142 
143 	if (ipoib_ib_dev_open(priv))
144 		goto err_disable;
145 
146 	if (ipoib_ib_dev_up(priv))
147 		goto err_stop;
148 
149 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
150 		struct ipoib_dev_priv *cpriv;
151 
152 		/* Bring up any child interfaces too */
153 		mutex_lock(&priv->vlan_mutex);
154 		list_for_each_entry(cpriv, &priv->child_intfs, list)
155 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
156 				ipoib_open(cpriv);
157 		mutex_unlock(&priv->vlan_mutex);
158 	}
159 	dev->if_drv_flags |= IFF_DRV_RUNNING;
160 	dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
161 
162 	return 0;
163 
164 err_stop:
165 	ipoib_ib_dev_stop(priv, 1);
166 
167 err_disable:
168 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
169 
170 	return -EINVAL;
171 }
172 
173 static void
174 ipoib_init(void *arg)
175 {
176 	struct ifnet *dev;
177 	struct ipoib_dev_priv *priv;
178 
179 	priv = arg;
180 	dev = priv->dev;
181 	if ((dev->if_drv_flags & IFF_DRV_RUNNING) == 0)
182 		ipoib_open(priv);
183 	queue_work(ipoib_workqueue, &priv->flush_light);
184 }
185 
186 
187 static int
188 ipoib_stop(struct ipoib_dev_priv *priv)
189 {
190 	struct ifnet *dev = priv->dev;
191 
192 	ipoib_dbg(priv, "stopping interface\n");
193 
194 	clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
195 
196 	dev->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
197 
198 	ipoib_ib_dev_down(priv, 0);
199 	ipoib_ib_dev_stop(priv, 0);
200 
201 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
202 		struct ipoib_dev_priv *cpriv;
203 
204 		/* Bring down any child interfaces too */
205 		mutex_lock(&priv->vlan_mutex);
206 		list_for_each_entry(cpriv, &priv->child_intfs, list)
207 			if ((cpriv->dev->if_drv_flags & IFF_DRV_RUNNING) != 0)
208 				ipoib_stop(cpriv);
209 		mutex_unlock(&priv->vlan_mutex);
210 	}
211 
212 	return 0;
213 }
214 
215 static int
216 ipoib_propagate_ifnet_mtu(struct ipoib_dev_priv *priv, int new_mtu,
217     bool propagate)
218 {
219 	struct ifnet *ifp;
220 	struct ifreq ifr;
221 	int error;
222 
223 	ifp = priv->dev;
224 	if (ifp->if_mtu == new_mtu)
225 		return (0);
226 	if (propagate) {
227 		strlcpy(ifr.ifr_name, if_name(ifp), IFNAMSIZ);
228 		ifr.ifr_mtu = new_mtu;
229 		CURVNET_SET(ifp->if_vnet);
230 		error = ifhwioctl(SIOCSIFMTU, ifp, (caddr_t)&ifr, curthread);
231 		CURVNET_RESTORE();
232 	} else {
233 		ifp->if_mtu = new_mtu;
234 		error = 0;
235 	}
236 	return (error);
237 }
238 
239 int
240 ipoib_change_mtu(struct ipoib_dev_priv *priv, int new_mtu, bool propagate)
241 {
242 	int error, prev_admin_mtu;
243 
244 	/* dev->if_mtu > 2K ==> connected mode */
245 	if (ipoib_cm_admin_enabled(priv)) {
246 		if (new_mtu > IPOIB_CM_MTU(ipoib_cm_max_mtu(priv)))
247 			return -EINVAL;
248 
249 		if (new_mtu > priv->mcast_mtu)
250 			ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
251 				   priv->mcast_mtu);
252 
253 		return (ipoib_propagate_ifnet_mtu(priv, new_mtu, propagate));
254 	}
255 
256 	if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
257 		return -EINVAL;
258 
259 	prev_admin_mtu = priv->admin_mtu;
260 	priv->admin_mtu = new_mtu;
261 	error = ipoib_propagate_ifnet_mtu(priv, min(priv->mcast_mtu,
262 	    priv->admin_mtu), propagate);
263 	if (error == 0) {
264 		/* check for MTU change to avoid infinite loop */
265 		if (prev_admin_mtu != new_mtu)
266 			queue_work(ipoib_workqueue, &priv->flush_light);
267 	} else
268 		priv->admin_mtu = prev_admin_mtu;
269 	return (error);
270 }
271 
272 static int
273 ipoib_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
274 {
275 	struct ipoib_dev_priv *priv = ifp->if_softc;
276 	struct ifaddr *ifa = (struct ifaddr *) data;
277 	struct ifreq *ifr = (struct ifreq *) data;
278 	int error = 0;
279 
280 	/* check if detaching */
281 	if (priv == NULL)
282 		return (ENXIO);
283 	/* wait for device to become ready, if any */
284 	while (priv->gone == 2)
285 		pause("W", 1);
286 	/* check for device gone */
287 	if (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 			bcopy(IF_LLADDR(ifp), &ifr->ifr_addr.sa_data[0],
322                             INFINIBAND_ALEN);
323 		break;
324 
325 	case SIOCSIFMTU:
326 		/*
327 		 * Set the interface MTU.
328 		 */
329 		error = -ipoib_change_mtu(priv, ifr->ifr_mtu, false);
330 		break;
331 	default:
332 		error = EINVAL;
333 		break;
334 	}
335 	return (error);
336 }
337 
338 
339 static struct ipoib_path *
340 __path_find(struct ipoib_dev_priv *priv, void *gid)
341 {
342 	struct rb_node *n = priv->path_tree.rb_node;
343 	struct ipoib_path *path;
344 	int ret;
345 
346 	while (n) {
347 		path = rb_entry(n, struct ipoib_path, rb_node);
348 
349 		ret = memcmp(gid, path->pathrec.dgid.raw,
350 			     sizeof (union ib_gid));
351 
352 		if (ret < 0)
353 			n = n->rb_left;
354 		else if (ret > 0)
355 			n = n->rb_right;
356 		else
357 			return path;
358 	}
359 
360 	return NULL;
361 }
362 
363 static int
364 __path_add(struct ipoib_dev_priv *priv, struct ipoib_path *path)
365 {
366 	struct rb_node **n = &priv->path_tree.rb_node;
367 	struct rb_node *pn = NULL;
368 	struct ipoib_path *tpath;
369 	int ret;
370 
371 	while (*n) {
372 		pn = *n;
373 		tpath = rb_entry(pn, struct ipoib_path, rb_node);
374 
375 		ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
376 			     sizeof (union ib_gid));
377 		if (ret < 0)
378 			n = &pn->rb_left;
379 		else if (ret > 0)
380 			n = &pn->rb_right;
381 		else
382 			return -EEXIST;
383 	}
384 
385 	rb_link_node(&path->rb_node, pn, n);
386 	rb_insert_color(&path->rb_node, &priv->path_tree);
387 
388 	list_add_tail(&path->list, &priv->path_list);
389 
390 	return 0;
391 }
392 
393 void
394 ipoib_path_free(struct ipoib_dev_priv *priv, struct ipoib_path *path)
395 {
396 
397 	_IF_DRAIN(&path->queue);
398 
399 	if (path->ah)
400 		ipoib_put_ah(path->ah);
401 	if (ipoib_cm_get(path))
402 		ipoib_cm_destroy_tx(ipoib_cm_get(path));
403 
404 	kfree(path);
405 }
406 
407 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
408 
409 struct ipoib_path_iter *
410 ipoib_path_iter_init(struct ipoib_dev_priv *priv)
411 {
412 	struct ipoib_path_iter *iter;
413 
414 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
415 	if (!iter)
416 		return NULL;
417 
418 	iter->priv = priv;
419 	memset(iter->path.pathrec.dgid.raw, 0, 16);
420 
421 	if (ipoib_path_iter_next(iter)) {
422 		kfree(iter);
423 		return NULL;
424 	}
425 
426 	return iter;
427 }
428 
429 int
430 ipoib_path_iter_next(struct ipoib_path_iter *iter)
431 {
432 	struct ipoib_dev_priv *priv = iter->priv;
433 	struct rb_node *n;
434 	struct ipoib_path *path;
435 	int ret = 1;
436 
437 	spin_lock_irq(&priv->lock);
438 
439 	n = rb_first(&priv->path_tree);
440 
441 	while (n) {
442 		path = rb_entry(n, struct ipoib_path, rb_node);
443 
444 		if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
445 			   sizeof (union ib_gid)) < 0) {
446 			iter->path = *path;
447 			ret = 0;
448 			break;
449 		}
450 
451 		n = rb_next(n);
452 	}
453 
454 	spin_unlock_irq(&priv->lock);
455 
456 	return ret;
457 }
458 
459 void
460 ipoib_path_iter_read(struct ipoib_path_iter *iter, struct ipoib_path *path)
461 {
462 	*path = iter->path;
463 }
464 
465 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
466 
467 void
468 ipoib_mark_paths_invalid(struct ipoib_dev_priv *priv)
469 {
470 	struct ipoib_path *path, *tp;
471 
472 	spin_lock_irq(&priv->lock);
473 
474 	list_for_each_entry_safe(path, tp, &priv->path_list, list) {
475 		ipoib_dbg(priv, "mark path LID 0x%04x GID %16D invalid\n",
476 			be16_to_cpu(path->pathrec.dlid),
477 			path->pathrec.dgid.raw, ":");
478 		path->valid =  0;
479 	}
480 
481 	spin_unlock_irq(&priv->lock);
482 }
483 
484 void
485 ipoib_flush_paths(struct ipoib_dev_priv *priv)
486 {
487 	struct ipoib_path *path, *tp;
488 	LIST_HEAD(remove_list);
489 	unsigned long flags;
490 
491 	spin_lock_irqsave(&priv->lock, flags);
492 
493 	list_splice_init(&priv->path_list, &remove_list);
494 
495 	list_for_each_entry(path, &remove_list, list)
496 		rb_erase(&path->rb_node, &priv->path_tree);
497 
498 	list_for_each_entry_safe(path, tp, &remove_list, list) {
499 		if (path->query)
500 			ib_sa_cancel_query(path->query_id, path->query);
501 		spin_unlock_irqrestore(&priv->lock, flags);
502 		wait_for_completion(&path->done);
503 		ipoib_path_free(priv, path);
504 		spin_lock_irqsave(&priv->lock, flags);
505 	}
506 
507 	spin_unlock_irqrestore(&priv->lock, flags);
508 }
509 
510 static void
511 path_rec_completion(int status, struct ib_sa_path_rec *pathrec, void *path_ptr)
512 {
513 	struct ipoib_path *path = path_ptr;
514 	struct ipoib_dev_priv *priv = path->priv;
515 	struct ifnet *dev = priv->dev;
516 	struct ipoib_ah *ah = NULL;
517 	struct ipoib_ah *old_ah = NULL;
518 	struct epoch_tracker et;
519 	struct ifqueue mbqueue;
520 	struct mbuf *mb;
521 	unsigned long flags;
522 
523 	if (!status)
524 		ipoib_dbg(priv, "PathRec LID 0x%04x for GID %16D\n",
525 			  be16_to_cpu(pathrec->dlid), pathrec->dgid.raw, ":");
526 	else
527 		ipoib_dbg(priv, "PathRec status %d for GID %16D\n",
528 			  status, path->pathrec.dgid.raw, ":");
529 
530 	bzero(&mbqueue, sizeof(mbqueue));
531 
532 	if (!status) {
533 		struct ib_ah_attr av;
534 
535 		if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
536 			ah = ipoib_create_ah(priv, priv->pd, &av);
537 	}
538 
539 	spin_lock_irqsave(&priv->lock, flags);
540 
541 	if (ah) {
542 		path->pathrec = *pathrec;
543 
544 		old_ah   = path->ah;
545 		path->ah = ah;
546 
547 		ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
548 			  ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
549 
550 		for (;;) {
551 			_IF_DEQUEUE(&path->queue, mb);
552 			if (mb == NULL)
553 				break;
554 			_IF_ENQUEUE(&mbqueue, mb);
555 		}
556 
557 #ifdef CONFIG_INFINIBAND_IPOIB_CM
558 		if (ipoib_cm_enabled(priv, path->hwaddr) && !ipoib_cm_get(path))
559 			ipoib_cm_set(path, ipoib_cm_create_tx(priv, path));
560 #endif
561 
562 		path->valid = 1;
563 	}
564 
565 	path->query = NULL;
566 	complete(&path->done);
567 
568 	spin_unlock_irqrestore(&priv->lock, flags);
569 
570 	if (old_ah)
571 		ipoib_put_ah(old_ah);
572 
573 	NET_EPOCH_ENTER(et);
574 	for (;;) {
575 		_IF_DEQUEUE(&mbqueue, mb);
576 		if (mb == NULL)
577 			break;
578 		mb->m_pkthdr.rcvif = dev;
579 		if (dev->if_transmit(dev, mb))
580 			ipoib_warn(priv, "dev_queue_xmit failed "
581 				   "to requeue packet\n");
582 	}
583 	NET_EPOCH_EXIT(et);
584 }
585 
586 static struct ipoib_path *
587 path_rec_create(struct ipoib_dev_priv *priv, uint8_t *hwaddr)
588 {
589 	struct ipoib_path *path;
590 
591 	if (!priv->broadcast)
592 		return NULL;
593 
594 	path = kzalloc(sizeof *path, GFP_ATOMIC);
595 	if (!path)
596 		return NULL;
597 
598 	path->priv = priv;
599 
600 	bzero(&path->queue, sizeof(path->queue));
601 
602 #ifdef CONFIG_INFINIBAND_IPOIB_CM
603 	memcpy(&path->hwaddr, hwaddr, INFINIBAND_ALEN);
604 #endif
605 	memcpy(path->pathrec.dgid.raw, &hwaddr[4], sizeof (union ib_gid));
606 	path->pathrec.sgid	    = priv->local_gid;
607 	path->pathrec.pkey	    = cpu_to_be16(priv->pkey);
608 	path->pathrec.numb_path     = 1;
609 	path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
610 
611 	return path;
612 }
613 
614 static int
615 path_rec_start(struct ipoib_dev_priv *priv, struct ipoib_path *path)
616 {
617 	struct ifnet *dev = priv->dev;
618 
619 	ib_sa_comp_mask comp_mask = IB_SA_PATH_REC_MTU_SELECTOR | IB_SA_PATH_REC_MTU;
620 	struct ib_sa_path_rec p_rec;
621 
622 	p_rec = path->pathrec;
623 	p_rec.mtu_selector = IB_SA_GT;
624 
625 	switch (roundup_pow_of_two(dev->if_mtu + IPOIB_ENCAP_LEN)) {
626 	case 512:
627 		p_rec.mtu = IB_MTU_256;
628 		break;
629 	case 1024:
630 		p_rec.mtu = IB_MTU_512;
631 		break;
632 	case 2048:
633 		p_rec.mtu = IB_MTU_1024;
634 		break;
635 	case 4096:
636 		p_rec.mtu = IB_MTU_2048;
637 		break;
638 	default:
639 		/* Wildcard everything */
640 		comp_mask = 0;
641 		p_rec.mtu = 0;
642 		p_rec.mtu_selector = 0;
643 	}
644 
645 	ipoib_dbg(priv, "Start path record lookup for %16D MTU > %d\n",
646 		  p_rec.dgid.raw, ":",
647 		  comp_mask ? ib_mtu_enum_to_int(p_rec.mtu) : 0);
648 
649 	init_completion(&path->done);
650 
651 	path->query_id =
652 		ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
653 				   &p_rec, comp_mask		|
654 				   IB_SA_PATH_REC_DGID		|
655 				   IB_SA_PATH_REC_SGID		|
656 				   IB_SA_PATH_REC_NUMB_PATH	|
657 				   IB_SA_PATH_REC_TRAFFIC_CLASS |
658 				   IB_SA_PATH_REC_PKEY,
659 				   1000, GFP_ATOMIC,
660 				   path_rec_completion,
661 				   path, &path->query);
662 	if (path->query_id < 0) {
663 		ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
664 		path->query = NULL;
665 		complete(&path->done);
666 		return path->query_id;
667 	}
668 
669 	return 0;
670 }
671 
672 static void
673 ipoib_unicast_send(struct mbuf *mb, struct ipoib_dev_priv *priv, struct ipoib_header *eh)
674 {
675 	struct ipoib_path *path;
676 
677 	path = __path_find(priv, eh->hwaddr + 4);
678 	if (!path || !path->valid) {
679 		int new_path = 0;
680 
681 		if (!path) {
682 			path = path_rec_create(priv, eh->hwaddr);
683 			new_path = 1;
684 		}
685 		if (path) {
686 			if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
687 				_IF_ENQUEUE(&path->queue, mb);
688 			else {
689 				if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
690 				m_freem(mb);
691 			}
692 
693 			if (!path->query && path_rec_start(priv, path)) {
694 				if (new_path)
695 					ipoib_path_free(priv, path);
696 				return;
697 			} else
698 				__path_add(priv, path);
699 		} else {
700 			if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
701 			m_freem(mb);
702 		}
703 
704 		return;
705 	}
706 
707 	if (ipoib_cm_get(path) && ipoib_cm_up(path)) {
708 		ipoib_cm_send(priv, mb, ipoib_cm_get(path));
709 	} else if (path->ah) {
710 		ipoib_send(priv, mb, path->ah, IPOIB_QPN(eh->hwaddr));
711 	} else if ((path->query || !path_rec_start(priv, path)) &&
712 		    path->queue.ifq_len < IPOIB_MAX_PATH_REC_QUEUE) {
713 		_IF_ENQUEUE(&path->queue, mb);
714 	} else {
715 		if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
716 		m_freem(mb);
717 	}
718 }
719 
720 static int
721 ipoib_send_one(struct ipoib_dev_priv *priv, struct mbuf *mb)
722 {
723 	struct ipoib_header *eh;
724 
725 	eh = mtod(mb, struct ipoib_header *);
726 	if (IPOIB_IS_MULTICAST(eh->hwaddr)) {
727 		/* Add in the P_Key for multicast*/
728 		eh->hwaddr[8] = (priv->pkey >> 8) & 0xff;
729 		eh->hwaddr[9] = priv->pkey & 0xff;
730 
731 		ipoib_mcast_send(priv, eh->hwaddr + 4, mb);
732 	} else
733 		ipoib_unicast_send(mb, priv, eh);
734 
735 	return 0;
736 }
737 
738 void
739 ipoib_start_locked(struct ifnet *dev, struct ipoib_dev_priv *priv)
740 {
741 	struct mbuf *mb;
742 
743 	assert_spin_locked(&priv->lock);
744 
745 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd) &&
746 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) == 0) {
747 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
748 		if (mb == NULL)
749 			break;
750 		INFINIBAND_BPF_MTAP(dev, mb);
751 		ipoib_send_one(priv, mb);
752 	}
753 }
754 
755 static void
756 _ipoib_start(struct ifnet *dev, struct ipoib_dev_priv *priv)
757 {
758 
759 	if ((dev->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
760 	    IFF_DRV_RUNNING)
761 		return;
762 
763 	spin_lock(&priv->lock);
764 	ipoib_start_locked(dev, priv);
765 	spin_unlock(&priv->lock);
766 }
767 
768 static void
769 ipoib_start(struct ifnet *dev)
770 {
771 	_ipoib_start(dev, dev->if_softc);
772 }
773 
774 static void
775 ipoib_vlan_start(struct ifnet *dev)
776 {
777 	struct ipoib_dev_priv *priv;
778 	struct mbuf *mb;
779 
780 	priv = VLAN_COOKIE(dev);
781 	if (priv != NULL)
782 		return _ipoib_start(dev, priv);
783 	while (!IFQ_DRV_IS_EMPTY(&dev->if_snd)) {
784 		IFQ_DRV_DEQUEUE(&dev->if_snd, mb);
785 		if (mb == NULL)
786 			break;
787 		m_freem(mb);
788 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
789 	}
790 }
791 
792 int
793 ipoib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
794 {
795 
796 	/* Allocate RX/TX "rings" to hold queued mbs */
797 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
798 				GFP_KERNEL);
799 	if (!priv->rx_ring) {
800 		printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
801 		       ca->name, ipoib_recvq_size);
802 		goto out;
803 	}
804 
805 	priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring, GFP_KERNEL);
806 	if (!priv->tx_ring) {
807 		printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
808 		       ca->name, ipoib_sendq_size);
809 		goto out_rx_ring_cleanup;
810 	}
811 	memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
812 
813 	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
814 
815 	if (ipoib_ib_dev_init(priv, ca, port))
816 		goto out_tx_ring_cleanup;
817 
818 	return 0;
819 
820 out_tx_ring_cleanup:
821 	kfree(priv->tx_ring);
822 
823 out_rx_ring_cleanup:
824 	kfree(priv->rx_ring);
825 
826 out:
827 	return -ENOMEM;
828 }
829 
830 static void
831 ipoib_ifdetach(struct ipoib_dev_priv *priv)
832 {
833 	struct ifnet *dev;
834 
835 	dev = priv->dev;
836 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
837 		priv->gone = 1;
838 		infiniband_ifdetach(dev);
839 	}
840 }
841 
842 static void
843 ipoib_detach(struct ipoib_dev_priv *priv)
844 {
845 	struct ifnet *dev;
846 
847 	dev = priv->dev;
848 	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
849 		if_free(dev);
850 		free_unr(ipoib_unrhdr, priv->unit);
851 	} else
852 		VLAN_SETCOOKIE(priv->dev, NULL);
853 
854 	free(priv, M_TEMP);
855 }
856 
857 void
858 ipoib_dev_cleanup(struct ipoib_dev_priv *priv)
859 {
860 	struct ipoib_dev_priv *cpriv, *tcpriv;
861 
862 	/* Delete any child interfaces first */
863 	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
864 		ipoib_ifdetach(cpriv);
865 		ipoib_dev_cleanup(cpriv);
866 		ipoib_detach(cpriv);
867 	}
868 
869 	ipoib_ib_dev_cleanup(priv);
870 
871 	kfree(priv->rx_ring);
872 	kfree(priv->tx_ring);
873 
874 	priv->rx_ring = NULL;
875 	priv->tx_ring = NULL;
876 }
877 
878 static struct ipoib_dev_priv *
879 ipoib_priv_alloc(void)
880 {
881 	struct ipoib_dev_priv *priv;
882 
883 	priv = malloc(sizeof(struct ipoib_dev_priv), M_TEMP, M_ZERO|M_WAITOK);
884 	spin_lock_init(&priv->lock);
885 	spin_lock_init(&priv->drain_lock);
886 	mutex_init(&priv->vlan_mutex);
887 	INIT_LIST_HEAD(&priv->path_list);
888 	INIT_LIST_HEAD(&priv->child_intfs);
889 	INIT_LIST_HEAD(&priv->dead_ahs);
890 	INIT_LIST_HEAD(&priv->multicast_list);
891 	INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
892 	INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
893 	INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
894 	INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
895 	INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
896 	INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
897 	INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
898 	INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
899 	memcpy(priv->broadcastaddr, ipv4_bcast_addr, INFINIBAND_ALEN);
900 
901 	return (priv);
902 }
903 
904 struct ipoib_dev_priv *
905 ipoib_intf_alloc(const char *name)
906 {
907 	struct ipoib_dev_priv *priv;
908 	struct ifnet *dev;
909 
910 	priv = ipoib_priv_alloc();
911 	dev = priv->dev = if_alloc(IFT_INFINIBAND);
912 	if (!dev) {
913 		free(priv, M_TEMP);
914 		return NULL;
915 	}
916 	dev->if_softc = priv;
917 	priv->gone = 2; /* initializing */
918 	priv->unit = alloc_unr(ipoib_unrhdr);
919 	if (priv->unit == -1) {
920 		if_free(dev);
921 		free(priv, M_TEMP);
922 		return NULL;
923 	}
924 	if_initname(dev, name, priv->unit);
925 	dev->if_flags = IFF_BROADCAST | IFF_MULTICAST;
926 
927 	infiniband_ifattach(priv->dev, NULL, priv->broadcastaddr);
928 
929 	dev->if_init = ipoib_init;
930 	dev->if_ioctl = ipoib_ioctl;
931 	dev->if_start = ipoib_start;
932 
933 	dev->if_snd.ifq_maxlen = ipoib_sendq_size * 2;
934 
935 	priv->dev = dev;
936 	if_link_state_change(priv->dev, LINK_STATE_DOWN);
937 
938 	return dev->if_softc;
939 }
940 
941 int
942 ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
943 {
944 	struct ib_device_attr *device_attr = &hca->attrs;
945 
946 	priv->hca_caps = device_attr->device_cap_flags;
947 
948 	priv->dev->if_hwassist = 0;
949 	priv->dev->if_capabilities = 0;
950 
951 #ifndef CONFIG_INFINIBAND_IPOIB_CM
952 	if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
953 		set_bit(IPOIB_FLAG_CSUM, &priv->flags);
954 		priv->dev->if_hwassist = CSUM_IP | CSUM_TCP | CSUM_UDP;
955 		priv->dev->if_capabilities = IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
956 	}
957 
958 #if 0
959 	if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO) {
960 		priv->dev->if_capabilities |= IFCAP_TSO4;
961 		priv->dev->if_hwassist |= CSUM_TSO;
962 	}
963 #endif
964 #endif
965 	priv->dev->if_capabilities |=
966 	    IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_LINKSTATE;
967 	priv->dev->if_capenable = priv->dev->if_capabilities;
968 
969 	return 0;
970 }
971 
972 
973 static struct ifnet *
974 ipoib_add_port(const char *format, struct ib_device *hca, u8 port)
975 {
976 	struct ipoib_dev_priv *priv;
977 	struct ib_port_attr attr;
978 	int result = -ENOMEM;
979 
980 	priv = ipoib_intf_alloc(format);
981 	if (!priv)
982 		goto alloc_mem_failed;
983 
984 	if (!ib_query_port(hca, port, &attr))
985 		priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
986 	else {
987 		printk(KERN_WARNING "%s: ib_query_port %d failed\n",
988 		       hca->name, port);
989 		goto device_init_failed;
990 	}
991 
992 	/* MTU will be reset when mcast join happens */
993 	priv->dev->if_mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
994 	priv->mcast_mtu = priv->admin_mtu = priv->dev->if_mtu;
995 
996 	result = ib_query_pkey(hca, port, 0, &priv->pkey);
997 	if (result) {
998 		printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
999 		       hca->name, port, result);
1000 		goto device_init_failed;
1001 	}
1002 
1003 	if (ipoib_set_dev_features(priv, hca))
1004 		goto device_init_failed;
1005 
1006 	/*
1007 	 * Set the full membership bit, so that we join the right
1008 	 * broadcast group, etc.
1009 	 */
1010 	priv->pkey |= 0x8000;
1011 
1012 	priv->broadcastaddr[8] = priv->pkey >> 8;
1013 	priv->broadcastaddr[9] = priv->pkey & 0xff;
1014 
1015 	result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1016 	if (result) {
1017 		printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1018 		       hca->name, port, result);
1019 		goto device_init_failed;
1020 	}
1021 	memcpy(IF_LLADDR(priv->dev) + 4, priv->local_gid.raw, sizeof(union ib_gid));
1022 
1023 	result = ipoib_dev_init(priv, hca, port);
1024 	if (result < 0) {
1025 		printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1026 		       hca->name, port, result);
1027 		goto device_init_failed;
1028 	}
1029 	if (ipoib_cm_admin_enabled(priv))
1030 		priv->dev->if_mtu = IPOIB_CM_MTU(ipoib_cm_max_mtu(priv));
1031 
1032 	INIT_IB_EVENT_HANDLER(&priv->event_handler,
1033 			      priv->ca, ipoib_event);
1034 	result = ib_register_event_handler(&priv->event_handler);
1035 	if (result < 0) {
1036 		printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1037 		       "port %d (ret = %d)\n",
1038 		       hca->name, port, result);
1039 		goto event_failed;
1040 	}
1041 	if_printf(priv->dev, "Attached to %s port %d\n", hca->name, port);
1042 
1043 	priv->gone = 0;	/* ready */
1044 
1045 	return priv->dev;
1046 
1047 event_failed:
1048 	ipoib_dev_cleanup(priv);
1049 
1050 device_init_failed:
1051 	ipoib_ifdetach(priv);
1052 	ipoib_detach(priv);
1053 
1054 alloc_mem_failed:
1055 	return ERR_PTR(result);
1056 }
1057 
1058 static void
1059 ipoib_add_one(struct ib_device *device)
1060 {
1061 	struct list_head *dev_list;
1062 	struct ifnet *dev;
1063 	struct ipoib_dev_priv *priv;
1064 	int s, e, p;
1065 
1066 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1067 		return;
1068 
1069 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1070 	if (!dev_list)
1071 		return;
1072 
1073 	INIT_LIST_HEAD(dev_list);
1074 
1075 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
1076 		s = 0;
1077 		e = 0;
1078 	} else {
1079 		s = 1;
1080 		e = device->phys_port_cnt;
1081 	}
1082 
1083 	for (p = s; p <= e; ++p) {
1084 		if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1085 			continue;
1086 		dev = ipoib_add_port("ib", device, p);
1087 		if (!IS_ERR(dev)) {
1088 			priv = dev->if_softc;
1089 			list_add_tail(&priv->list, dev_list);
1090 		}
1091 	}
1092 
1093 	ib_set_client_data(device, &ipoib_client, dev_list);
1094 }
1095 
1096 static void
1097 ipoib_remove_one(struct ib_device *device, void *client_data)
1098 {
1099 	struct ipoib_dev_priv *priv, *tmp;
1100 	struct list_head *dev_list = client_data;
1101 
1102 	if (!dev_list)
1103 		return;
1104 
1105 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1106 		return;
1107 
1108 	list_for_each_entry_safe(priv, tmp, dev_list, list) {
1109 		if (rdma_port_get_link_layer(device, priv->port) != IB_LINK_LAYER_INFINIBAND)
1110 			continue;
1111 
1112 		ipoib_ifdetach(priv);
1113 		ipoib_stop(priv);
1114 
1115 		ib_unregister_event_handler(&priv->event_handler);
1116 
1117 		flush_workqueue(ipoib_workqueue);
1118 
1119 		ipoib_dev_cleanup(priv);
1120 		ipoib_detach(priv);
1121 	}
1122 
1123 	kfree(dev_list);
1124 }
1125 
1126 static int
1127 ipoib_match_dev_addr(const struct sockaddr *addr, struct ifnet *dev)
1128 {
1129 	struct epoch_tracker et;
1130 	struct ifaddr *ifa;
1131 	int retval = 0;
1132 
1133 	NET_EPOCH_ENTER(et);
1134 	CK_STAILQ_FOREACH(ifa, &dev->if_addrhead, ifa_link) {
1135 		if (ifa->ifa_addr == NULL ||
1136 		    ifa->ifa_addr->sa_family != addr->sa_family ||
1137 		    ifa->ifa_addr->sa_len != addr->sa_len) {
1138 			continue;
1139 		}
1140 		if (memcmp(ifa->ifa_addr, addr, addr->sa_len) == 0) {
1141 			retval = 1;
1142 			break;
1143 		}
1144 	}
1145 	NET_EPOCH_EXIT(et);
1146 
1147 	return (retval);
1148 }
1149 
1150 /*
1151  * ipoib_match_gid_pkey_addr - returns the number of IPoIB netdevs on
1152  * top a given ipoib device matching a pkey_index and address, if one
1153  * exists.
1154  *
1155  * @found_net_dev: contains a matching net_device if the return value
1156  * >= 1, with a reference held.
1157  */
1158 static int
1159 ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
1160     const union ib_gid *gid, u16 pkey_index, const struct sockaddr *addr,
1161     struct ifnet **found_net_dev)
1162 {
1163 	struct ipoib_dev_priv *child_priv;
1164 	int matches = 0;
1165 
1166 	if (priv->pkey_index == pkey_index &&
1167 	    (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
1168 		if (addr == NULL || ipoib_match_dev_addr(addr, priv->dev) != 0) {
1169 			if (*found_net_dev == NULL) {
1170 				struct ifnet *net_dev;
1171 
1172 				if (priv->parent != NULL)
1173 					net_dev = priv->parent;
1174 				else
1175 					net_dev = priv->dev;
1176 				*found_net_dev = net_dev;
1177 				dev_hold(net_dev);
1178 			}
1179 			matches++;
1180 		}
1181 	}
1182 
1183 	/* Check child interfaces */
1184 	mutex_lock(&priv->vlan_mutex);
1185 	list_for_each_entry(child_priv, &priv->child_intfs, list) {
1186 		matches += ipoib_match_gid_pkey_addr(child_priv, gid,
1187 		    pkey_index, addr, found_net_dev);
1188 		if (matches > 1)
1189 			break;
1190 	}
1191 	mutex_unlock(&priv->vlan_mutex);
1192 
1193 	return matches;
1194 }
1195 
1196 /*
1197  * __ipoib_get_net_dev_by_params - returns the number of matching
1198  * net_devs found (between 0 and 2). Also return the matching
1199  * net_device in the @net_dev parameter, holding a reference to the
1200  * net_device, if the number of matches >= 1
1201  */
1202 static int
1203 __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
1204     u16 pkey_index, const union ib_gid *gid,
1205     const struct sockaddr *addr, struct ifnet **net_dev)
1206 {
1207 	struct ipoib_dev_priv *priv;
1208 	int matches = 0;
1209 
1210 	*net_dev = NULL;
1211 
1212 	list_for_each_entry(priv, dev_list, list) {
1213 		if (priv->port != port)
1214 			continue;
1215 
1216 		matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
1217 		    addr, net_dev);
1218 
1219 		if (matches > 1)
1220 			break;
1221 	}
1222 
1223 	return matches;
1224 }
1225 
1226 static struct ifnet *
1227 ipoib_get_net_dev_by_params(struct ib_device *dev, u8 port, u16 pkey,
1228     const union ib_gid *gid, const struct sockaddr *addr, void *client_data)
1229 {
1230 	struct ifnet *net_dev;
1231 	struct list_head *dev_list = client_data;
1232 	u16 pkey_index;
1233 	int matches;
1234 	int ret;
1235 
1236 	if (!rdma_protocol_ib(dev, port))
1237 		return NULL;
1238 
1239 	ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
1240 	if (ret)
1241 		return NULL;
1242 
1243 	if (!dev_list)
1244 		return NULL;
1245 
1246 	/* See if we can find a unique device matching the L2 parameters */
1247 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1248 						gid, NULL, &net_dev);
1249 
1250 	switch (matches) {
1251 	case 0:
1252 		return NULL;
1253 	case 1:
1254 		return net_dev;
1255 	}
1256 
1257 	dev_put(net_dev);
1258 
1259 	/* Couldn't find a unique device with L2 parameters only. Use L3
1260 	 * address to uniquely match the net device */
1261 	matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
1262 						gid, addr, &net_dev);
1263 	switch (matches) {
1264 	case 0:
1265 		return NULL;
1266 	default:
1267 		dev_warn_ratelimited(&dev->dev,
1268 				     "duplicate IP address detected\n");
1269 		/* Fall through */
1270 	case 1:
1271 		return net_dev;
1272 	}
1273 }
1274 
1275 static void
1276 ipoib_config_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1277 {
1278 	struct ipoib_dev_priv *parent;
1279 	struct ipoib_dev_priv *priv;
1280 	struct epoch_tracker et;
1281 	struct ifnet *dev;
1282 	uint16_t pkey;
1283 	int error;
1284 
1285 	if (ifp->if_type != IFT_INFINIBAND)
1286 		return;
1287 	NET_EPOCH_ENTER(et);
1288 	dev = VLAN_DEVAT(ifp, vtag);
1289 	NET_EPOCH_EXIT(et);
1290 	if (dev == NULL)
1291 		return;
1292 	priv = NULL;
1293 	error = 0;
1294 	parent = ifp->if_softc;
1295 	/* We only support 15 bits of pkey. */
1296 	if (vtag & 0x8000)
1297 		return;
1298 	pkey = vtag | 0x8000;	/* Set full membership bit. */
1299 	if (pkey == parent->pkey)
1300 		return;
1301 	/* Check for dups */
1302 	mutex_lock(&parent->vlan_mutex);
1303 	list_for_each_entry(priv, &parent->child_intfs, list) {
1304 		if (priv->pkey == pkey) {
1305 			priv = NULL;
1306 			error = EBUSY;
1307 			goto out;
1308 		}
1309 	}
1310 	priv = ipoib_priv_alloc();
1311 	priv->dev = dev;
1312 	priv->max_ib_mtu = parent->max_ib_mtu;
1313 	priv->mcast_mtu = priv->admin_mtu = parent->dev->if_mtu;
1314 	set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1315 	error = ipoib_set_dev_features(priv, parent->ca);
1316 	if (error)
1317 		goto out;
1318 	priv->pkey = pkey;
1319 	priv->broadcastaddr[8] = pkey >> 8;
1320 	priv->broadcastaddr[9] = pkey & 0xff;
1321 	dev->if_broadcastaddr = priv->broadcastaddr;
1322 	error = ipoib_dev_init(priv, parent->ca, parent->port);
1323 	if (error)
1324 		goto out;
1325 	priv->parent = parent->dev;
1326 	list_add_tail(&priv->list, &parent->child_intfs);
1327 	VLAN_SETCOOKIE(dev, priv);
1328 	dev->if_start = ipoib_vlan_start;
1329 	dev->if_drv_flags &= ~IFF_DRV_RUNNING;
1330 	dev->if_hdrlen = IPOIB_HEADER_LEN;
1331 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1332 		ipoib_open(priv);
1333 	mutex_unlock(&parent->vlan_mutex);
1334 	return;
1335 out:
1336 	mutex_unlock(&parent->vlan_mutex);
1337 	if (priv)
1338 		free(priv, M_TEMP);
1339 	if (error)
1340 		ipoib_warn(parent,
1341 		    "failed to initialize subinterface: device %s, port %d vtag 0x%X",
1342 		    parent->ca->name, parent->port, vtag);
1343 	return;
1344 }
1345 
1346 static void
1347 ipoib_unconfig_vlan(void *arg, struct ifnet *ifp, u_int16_t vtag)
1348 {
1349 	struct ipoib_dev_priv *parent;
1350 	struct ipoib_dev_priv *priv;
1351 	struct epoch_tracker et;
1352 	struct ifnet *dev;
1353 	uint16_t pkey;
1354 
1355 	if (ifp->if_type != IFT_INFINIBAND)
1356 		return;
1357 
1358 	NET_EPOCH_ENTER(et);
1359 	dev = VLAN_DEVAT(ifp, vtag);
1360 	NET_EPOCH_EXIT(et);
1361 	if (dev)
1362 		VLAN_SETCOOKIE(dev, NULL);
1363 	pkey = vtag | 0x8000;
1364 	parent = ifp->if_softc;
1365 	mutex_lock(&parent->vlan_mutex);
1366 	list_for_each_entry(priv, &parent->child_intfs, list) {
1367 		if (priv->pkey == pkey) {
1368 			ipoib_dev_cleanup(priv);
1369 			list_del(&priv->list);
1370 			break;
1371 		}
1372 	}
1373 	mutex_unlock(&parent->vlan_mutex);
1374 }
1375 
1376 eventhandler_tag ipoib_vlan_attach;
1377 eventhandler_tag ipoib_vlan_detach;
1378 
1379 static int __init
1380 ipoib_init_module(void)
1381 {
1382 	int ret;
1383 
1384 	ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1385 	ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1386 	ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1387 
1388 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1389 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1390 	ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1391 						     IPOIB_MIN_QUEUE_SIZE));
1392 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1393 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1394 #endif
1395 
1396 	ipoib_vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
1397 		ipoib_config_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1398 	ipoib_vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
1399 		ipoib_unconfig_vlan, NULL, EVENTHANDLER_PRI_FIRST);
1400 
1401 	/*
1402 	 * We create our own workqueue mainly because we want to be
1403 	 * able to flush it when devices are being removed.  We can't
1404 	 * use schedule_work()/flush_scheduled_work() because both
1405 	 * unregister_netdev() and linkwatch_event take the rtnl lock,
1406 	 * so flush_scheduled_work() can deadlock during device
1407 	 * removal.
1408 	 */
1409 	ipoib_workqueue = create_singlethread_workqueue("ipoib");
1410 	if (!ipoib_workqueue) {
1411 		ret = -ENOMEM;
1412 		goto err_fs;
1413 	}
1414 
1415 	ib_sa_register_client(&ipoib_sa_client);
1416 
1417 	ret = ib_register_client(&ipoib_client);
1418 	if (ret)
1419 		goto err_sa;
1420 
1421 	return 0;
1422 
1423 err_sa:
1424 	ib_sa_unregister_client(&ipoib_sa_client);
1425 	destroy_workqueue(ipoib_workqueue);
1426 
1427 err_fs:
1428 	return ret;
1429 }
1430 
1431 static void __exit
1432 ipoib_cleanup_module(void)
1433 {
1434 
1435 	EVENTHANDLER_DEREGISTER(vlan_config, ipoib_vlan_attach);
1436 	EVENTHANDLER_DEREGISTER(vlan_unconfig, ipoib_vlan_detach);
1437 	ib_unregister_client(&ipoib_client);
1438 	ib_sa_unregister_client(&ipoib_sa_client);
1439 	destroy_workqueue(ipoib_workqueue);
1440 }
1441 module_init_order(ipoib_init_module, SI_ORDER_FIFTH);
1442 module_exit_order(ipoib_cleanup_module, SI_ORDER_FIFTH);
1443 
1444 static int
1445 ipoib_evhand(module_t mod, int event, void *arg)
1446 {
1447 	return (0);
1448 }
1449 
1450 static moduledata_t ipoib_mod = {
1451 	.name = "ipoib",
1452 	.evhand = ipoib_evhand,
1453 };
1454 
1455 DECLARE_MODULE(ipoib, ipoib_mod, SI_SUB_LAST, SI_ORDER_ANY);
1456 MODULE_DEPEND(ipoib, ibcore, 1, 1, 1);
1457 MODULE_DEPEND(ipoib, if_infiniband, 1, 1, 1);
1458 MODULE_DEPEND(ipoib, linuxkpi, 1, 1, 1);
1459