xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision b8bb76713ec50df2f11efee386e16f93d51e1076)
1 /*
2  * Copyright (c) 2004, 2005 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 <linux/skbuff.h>
36 #include <linux/rtnetlink.h>
37 #include <linux/ip.h>
38 #include <linux/in.h>
39 #include <linux/igmp.h>
40 #include <linux/inetdevice.h>
41 #include <linux/delay.h>
42 #include <linux/completion.h>
43 
44 #include <net/dst.h>
45 
46 #include "ipoib.h"
47 
48 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
49 static int mcast_debug_level;
50 
51 module_param(mcast_debug_level, int, 0644);
52 MODULE_PARM_DESC(mcast_debug_level,
53 		 "Enable multicast debug tracing if > 0");
54 #endif
55 
56 static DEFINE_MUTEX(mcast_mutex);
57 
58 struct ipoib_mcast_iter {
59 	struct net_device *dev;
60 	union ib_gid       mgid;
61 	unsigned long      created;
62 	unsigned int       queuelen;
63 	unsigned int       complete;
64 	unsigned int       send_only;
65 };
66 
67 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
68 {
69 	struct net_device *dev = mcast->dev;
70 	struct ipoib_dev_priv *priv = netdev_priv(dev);
71 	struct ipoib_neigh *neigh, *tmp;
72 	int tx_dropped = 0;
73 
74 	ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
75 			mcast->mcmember.mgid.raw);
76 
77 	spin_lock_irq(&priv->lock);
78 
79 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
80 		/*
81 		 * It's safe to call ipoib_put_ah() inside priv->lock
82 		 * here, because we know that mcast->ah will always
83 		 * hold one more reference, so ipoib_put_ah() will
84 		 * never do more than decrement the ref count.
85 		 */
86 		if (neigh->ah)
87 			ipoib_put_ah(neigh->ah);
88 		ipoib_neigh_free(dev, neigh);
89 	}
90 
91 	spin_unlock_irq(&priv->lock);
92 
93 	if (mcast->ah)
94 		ipoib_put_ah(mcast->ah);
95 
96 	while (!skb_queue_empty(&mcast->pkt_queue)) {
97 		++tx_dropped;
98 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
99 	}
100 
101 	netif_tx_lock_bh(dev);
102 	dev->stats.tx_dropped += tx_dropped;
103 	netif_tx_unlock_bh(dev);
104 
105 	kfree(mcast);
106 }
107 
108 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
109 					     int can_sleep)
110 {
111 	struct ipoib_mcast *mcast;
112 
113 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
114 	if (!mcast)
115 		return NULL;
116 
117 	mcast->dev = dev;
118 	mcast->created = jiffies;
119 	mcast->backoff = 1;
120 
121 	INIT_LIST_HEAD(&mcast->list);
122 	INIT_LIST_HEAD(&mcast->neigh_list);
123 	skb_queue_head_init(&mcast->pkt_queue);
124 
125 	return mcast;
126 }
127 
128 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
129 {
130 	struct ipoib_dev_priv *priv = netdev_priv(dev);
131 	struct rb_node *n = priv->multicast_tree.rb_node;
132 
133 	while (n) {
134 		struct ipoib_mcast *mcast;
135 		int ret;
136 
137 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
138 
139 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
140 			     sizeof (union ib_gid));
141 		if (ret < 0)
142 			n = n->rb_left;
143 		else if (ret > 0)
144 			n = n->rb_right;
145 		else
146 			return mcast;
147 	}
148 
149 	return NULL;
150 }
151 
152 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
153 {
154 	struct ipoib_dev_priv *priv = netdev_priv(dev);
155 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
156 
157 	while (*n) {
158 		struct ipoib_mcast *tmcast;
159 		int ret;
160 
161 		pn = *n;
162 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
163 
164 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
165 			     sizeof (union ib_gid));
166 		if (ret < 0)
167 			n = &pn->rb_left;
168 		else if (ret > 0)
169 			n = &pn->rb_right;
170 		else
171 			return -EEXIST;
172 	}
173 
174 	rb_link_node(&mcast->rb_node, pn, n);
175 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
176 
177 	return 0;
178 }
179 
180 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
181 				   struct ib_sa_mcmember_rec *mcmember)
182 {
183 	struct net_device *dev = mcast->dev;
184 	struct ipoib_dev_priv *priv = netdev_priv(dev);
185 	struct ipoib_ah *ah;
186 	int ret;
187 	int set_qkey = 0;
188 
189 	mcast->mcmember = *mcmember;
190 
191 	/* Set the cached Q_Key before we attach if it's the broadcast group */
192 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
193 		    sizeof (union ib_gid))) {
194 		spin_lock_irq(&priv->lock);
195 		if (!priv->broadcast) {
196 			spin_unlock_irq(&priv->lock);
197 			return -EAGAIN;
198 		}
199 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
200 		spin_unlock_irq(&priv->lock);
201 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
202 		set_qkey = 1;
203 	}
204 
205 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
206 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
207 			ipoib_warn(priv, "multicast group %pI6 already attached\n",
208 				   mcast->mcmember.mgid.raw);
209 
210 			return 0;
211 		}
212 
213 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
214 					 &mcast->mcmember.mgid, set_qkey);
215 		if (ret < 0) {
216 			ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
217 				   mcast->mcmember.mgid.raw);
218 
219 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
220 			return ret;
221 		}
222 	}
223 
224 	{
225 		struct ib_ah_attr av = {
226 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
227 			.port_num      = priv->port,
228 			.sl	       = mcast->mcmember.sl,
229 			.ah_flags      = IB_AH_GRH,
230 			.static_rate   = mcast->mcmember.rate,
231 			.grh	       = {
232 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
233 				.hop_limit     = mcast->mcmember.hop_limit,
234 				.sgid_index    = 0,
235 				.traffic_class = mcast->mcmember.traffic_class
236 			}
237 		};
238 		av.grh.dgid = mcast->mcmember.mgid;
239 
240 		ah = ipoib_create_ah(dev, priv->pd, &av);
241 		if (!ah) {
242 			ipoib_warn(priv, "ib_address_create failed\n");
243 		} else {
244 			spin_lock_irq(&priv->lock);
245 			mcast->ah = ah;
246 			spin_unlock_irq(&priv->lock);
247 
248 			ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
249 					mcast->mcmember.mgid.raw,
250 					mcast->ah->ah,
251 					be16_to_cpu(mcast->mcmember.mlid),
252 					mcast->mcmember.sl);
253 		}
254 	}
255 
256 	/* actually send any queued packets */
257 	netif_tx_lock_bh(dev);
258 	while (!skb_queue_empty(&mcast->pkt_queue)) {
259 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
260 		netif_tx_unlock_bh(dev);
261 
262 		skb->dev = dev;
263 
264 		if (!skb->dst || !skb->dst->neighbour) {
265 			/* put pseudoheader back on for next time */
266 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
267 		}
268 
269 		if (dev_queue_xmit(skb))
270 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
271 		netif_tx_lock_bh(dev);
272 	}
273 	netif_tx_unlock_bh(dev);
274 
275 	return 0;
276 }
277 
278 static int
279 ipoib_mcast_sendonly_join_complete(int status,
280 				   struct ib_sa_multicast *multicast)
281 {
282 	struct ipoib_mcast *mcast = multicast->context;
283 	struct net_device *dev = mcast->dev;
284 
285 	/* We trap for port events ourselves. */
286 	if (status == -ENETRESET)
287 		return 0;
288 
289 	if (!status)
290 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
291 
292 	if (status) {
293 		if (mcast->logcount++ < 20)
294 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
295 					mcast->mcmember.mgid.raw, status);
296 
297 		/* Flush out any queued packets */
298 		netif_tx_lock_bh(dev);
299 		while (!skb_queue_empty(&mcast->pkt_queue)) {
300 			++dev->stats.tx_dropped;
301 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
302 		}
303 		netif_tx_unlock_bh(dev);
304 
305 		/* Clear the busy flag so we try again */
306 		status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
307 					    &mcast->flags);
308 	}
309 	return status;
310 }
311 
312 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
313 {
314 	struct net_device *dev = mcast->dev;
315 	struct ipoib_dev_priv *priv = netdev_priv(dev);
316 	struct ib_sa_mcmember_rec rec = {
317 #if 0				/* Some SMs don't support send-only yet */
318 		.join_state = 4
319 #else
320 		.join_state = 1
321 #endif
322 	};
323 	int ret = 0;
324 
325 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
326 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
327 		return -ENODEV;
328 	}
329 
330 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
331 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
332 		return -EBUSY;
333 	}
334 
335 	rec.mgid     = mcast->mcmember.mgid;
336 	rec.port_gid = priv->local_gid;
337 	rec.pkey     = cpu_to_be16(priv->pkey);
338 
339 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
340 					 priv->port, &rec,
341 					 IB_SA_MCMEMBER_REC_MGID	|
342 					 IB_SA_MCMEMBER_REC_PORT_GID	|
343 					 IB_SA_MCMEMBER_REC_PKEY	|
344 					 IB_SA_MCMEMBER_REC_JOIN_STATE,
345 					 GFP_ATOMIC,
346 					 ipoib_mcast_sendonly_join_complete,
347 					 mcast);
348 	if (IS_ERR(mcast->mc)) {
349 		ret = PTR_ERR(mcast->mc);
350 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
351 		ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
352 			   ret);
353 	} else {
354 		ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
355 				mcast->mcmember.mgid.raw);
356 	}
357 
358 	return ret;
359 }
360 
361 void ipoib_mcast_carrier_on_task(struct work_struct *work)
362 {
363 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
364 						   carrier_on_task);
365 
366 	/*
367 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
368 	 * turning the carrier back on while a device is being
369 	 * removed.
370 	 */
371 	rtnl_lock();
372 	netif_carrier_on(priv->dev);
373 	rtnl_unlock();
374 }
375 
376 static int ipoib_mcast_join_complete(int status,
377 				     struct ib_sa_multicast *multicast)
378 {
379 	struct ipoib_mcast *mcast = multicast->context;
380 	struct net_device *dev = mcast->dev;
381 	struct ipoib_dev_priv *priv = netdev_priv(dev);
382 
383 	ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
384 			mcast->mcmember.mgid.raw, status);
385 
386 	/* We trap for port events ourselves. */
387 	if (status == -ENETRESET)
388 		return 0;
389 
390 	if (!status)
391 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
392 
393 	if (!status) {
394 		mcast->backoff = 1;
395 		mutex_lock(&mcast_mutex);
396 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
397 			queue_delayed_work(ipoib_workqueue,
398 					   &priv->mcast_task, 0);
399 		mutex_unlock(&mcast_mutex);
400 
401 		/*
402 		 * Defer carrier on work to ipoib_workqueue to avoid a
403 		 * deadlock on rtnl_lock here.
404 		 */
405 		if (mcast == priv->broadcast)
406 			queue_work(ipoib_workqueue, &priv->carrier_on_task);
407 
408 		return 0;
409 	}
410 
411 	if (mcast->logcount++ < 20) {
412 		if (status == -ETIMEDOUT || status == -EAGAIN) {
413 			ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
414 					mcast->mcmember.mgid.raw, status);
415 		} else {
416 			ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
417 				   mcast->mcmember.mgid.raw, status);
418 		}
419 	}
420 
421 	mcast->backoff *= 2;
422 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
423 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
424 
425 	/* Clear the busy flag so we try again */
426 	status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
427 
428 	mutex_lock(&mcast_mutex);
429 	spin_lock_irq(&priv->lock);
430 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
431 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
432 				   mcast->backoff * HZ);
433 	spin_unlock_irq(&priv->lock);
434 	mutex_unlock(&mcast_mutex);
435 
436 	return status;
437 }
438 
439 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
440 			     int create)
441 {
442 	struct ipoib_dev_priv *priv = netdev_priv(dev);
443 	struct ib_sa_mcmember_rec rec = {
444 		.join_state = 1
445 	};
446 	ib_sa_comp_mask comp_mask;
447 	int ret = 0;
448 
449 	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
450 
451 	rec.mgid     = mcast->mcmember.mgid;
452 	rec.port_gid = priv->local_gid;
453 	rec.pkey     = cpu_to_be16(priv->pkey);
454 
455 	comp_mask =
456 		IB_SA_MCMEMBER_REC_MGID		|
457 		IB_SA_MCMEMBER_REC_PORT_GID	|
458 		IB_SA_MCMEMBER_REC_PKEY		|
459 		IB_SA_MCMEMBER_REC_JOIN_STATE;
460 
461 	if (create) {
462 		comp_mask |=
463 			IB_SA_MCMEMBER_REC_QKEY			|
464 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
465 			IB_SA_MCMEMBER_REC_MTU			|
466 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
467 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
468 			IB_SA_MCMEMBER_REC_RATE			|
469 			IB_SA_MCMEMBER_REC_SL			|
470 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
471 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
472 
473 		rec.qkey	  = priv->broadcast->mcmember.qkey;
474 		rec.mtu_selector  = IB_SA_EQ;
475 		rec.mtu		  = priv->broadcast->mcmember.mtu;
476 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
477 		rec.rate_selector = IB_SA_EQ;
478 		rec.rate	  = priv->broadcast->mcmember.rate;
479 		rec.sl		  = priv->broadcast->mcmember.sl;
480 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
481 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
482 	}
483 
484 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
485 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
486 					 &rec, comp_mask, GFP_KERNEL,
487 					 ipoib_mcast_join_complete, mcast);
488 	if (IS_ERR(mcast->mc)) {
489 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
490 		ret = PTR_ERR(mcast->mc);
491 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
492 
493 		mcast->backoff *= 2;
494 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
495 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
496 
497 		mutex_lock(&mcast_mutex);
498 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
499 			queue_delayed_work(ipoib_workqueue,
500 					   &priv->mcast_task,
501 					   mcast->backoff * HZ);
502 		mutex_unlock(&mcast_mutex);
503 	}
504 }
505 
506 void ipoib_mcast_join_task(struct work_struct *work)
507 {
508 	struct ipoib_dev_priv *priv =
509 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
510 	struct net_device *dev = priv->dev;
511 
512 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
513 		return;
514 
515 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
516 		ipoib_warn(priv, "ib_query_gid() failed\n");
517 	else
518 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
519 
520 	{
521 		struct ib_port_attr attr;
522 
523 		if (!ib_query_port(priv->ca, priv->port, &attr))
524 			priv->local_lid = attr.lid;
525 		else
526 			ipoib_warn(priv, "ib_query_port failed\n");
527 	}
528 
529 	if (!priv->broadcast) {
530 		struct ipoib_mcast *broadcast;
531 
532 		if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
533 			return;
534 
535 		broadcast = ipoib_mcast_alloc(dev, 1);
536 		if (!broadcast) {
537 			ipoib_warn(priv, "failed to allocate broadcast group\n");
538 			mutex_lock(&mcast_mutex);
539 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
540 				queue_delayed_work(ipoib_workqueue,
541 						   &priv->mcast_task, HZ);
542 			mutex_unlock(&mcast_mutex);
543 			return;
544 		}
545 
546 		spin_lock_irq(&priv->lock);
547 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
548 		       sizeof (union ib_gid));
549 		priv->broadcast = broadcast;
550 
551 		__ipoib_mcast_add(dev, priv->broadcast);
552 		spin_unlock_irq(&priv->lock);
553 	}
554 
555 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
556 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
557 			ipoib_mcast_join(dev, priv->broadcast, 0);
558 		return;
559 	}
560 
561 	while (1) {
562 		struct ipoib_mcast *mcast = NULL;
563 
564 		spin_lock_irq(&priv->lock);
565 		list_for_each_entry(mcast, &priv->multicast_list, list) {
566 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
567 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
568 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
569 				/* Found the next unjoined group */
570 				break;
571 			}
572 		}
573 		spin_unlock_irq(&priv->lock);
574 
575 		if (&mcast->list == &priv->multicast_list) {
576 			/* All done */
577 			break;
578 		}
579 
580 		ipoib_mcast_join(dev, mcast, 1);
581 		return;
582 	}
583 
584 	priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
585 
586 	if (!ipoib_cm_admin_enabled(dev)) {
587 		rtnl_lock();
588 		dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
589 		rtnl_unlock();
590 	}
591 
592 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
593 
594 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
595 }
596 
597 int ipoib_mcast_start_thread(struct net_device *dev)
598 {
599 	struct ipoib_dev_priv *priv = netdev_priv(dev);
600 
601 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
602 
603 	mutex_lock(&mcast_mutex);
604 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
605 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
606 	mutex_unlock(&mcast_mutex);
607 
608 	return 0;
609 }
610 
611 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
612 {
613 	struct ipoib_dev_priv *priv = netdev_priv(dev);
614 
615 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
616 
617 	mutex_lock(&mcast_mutex);
618 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
619 	cancel_delayed_work(&priv->mcast_task);
620 	mutex_unlock(&mcast_mutex);
621 
622 	if (flush)
623 		flush_workqueue(ipoib_workqueue);
624 
625 	return 0;
626 }
627 
628 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
629 {
630 	struct ipoib_dev_priv *priv = netdev_priv(dev);
631 	int ret = 0;
632 
633 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
634 		ib_sa_free_multicast(mcast->mc);
635 
636 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
637 		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
638 				mcast->mcmember.mgid.raw);
639 
640 		/* Remove ourselves from the multicast group */
641 		ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
642 				      be16_to_cpu(mcast->mcmember.mlid));
643 		if (ret)
644 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
645 	}
646 
647 	return 0;
648 }
649 
650 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
651 {
652 	struct ipoib_dev_priv *priv = netdev_priv(dev);
653 	struct ipoib_mcast *mcast;
654 	unsigned long flags;
655 
656 	spin_lock_irqsave(&priv->lock, flags);
657 
658 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
659 	    !priv->broadcast					||
660 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
661 		++dev->stats.tx_dropped;
662 		dev_kfree_skb_any(skb);
663 		goto unlock;
664 	}
665 
666 	mcast = __ipoib_mcast_find(dev, mgid);
667 	if (!mcast) {
668 		/* Let's create a new send only group now */
669 		ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
670 				mgid);
671 
672 		mcast = ipoib_mcast_alloc(dev, 0);
673 		if (!mcast) {
674 			ipoib_warn(priv, "unable to allocate memory for "
675 				   "multicast structure\n");
676 			++dev->stats.tx_dropped;
677 			dev_kfree_skb_any(skb);
678 			goto out;
679 		}
680 
681 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
682 		memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
683 		__ipoib_mcast_add(dev, mcast);
684 		list_add_tail(&mcast->list, &priv->multicast_list);
685 	}
686 
687 	if (!mcast->ah) {
688 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
689 			skb_queue_tail(&mcast->pkt_queue, skb);
690 		else {
691 			++dev->stats.tx_dropped;
692 			dev_kfree_skb_any(skb);
693 		}
694 
695 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
696 			ipoib_dbg_mcast(priv, "no address vector, "
697 					"but multicast join already started\n");
698 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
699 			ipoib_mcast_sendonly_join(mcast);
700 
701 		/*
702 		 * If lookup completes between here and out:, don't
703 		 * want to send packet twice.
704 		 */
705 		mcast = NULL;
706 	}
707 
708 out:
709 	if (mcast && mcast->ah) {
710 		if (skb->dst		&&
711 		    skb->dst->neighbour &&
712 		    !*to_ipoib_neigh(skb->dst->neighbour)) {
713 			struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour,
714 									skb->dev);
715 
716 			if (neigh) {
717 				kref_get(&mcast->ah->ref);
718 				neigh->ah	= mcast->ah;
719 				list_add_tail(&neigh->list, &mcast->neigh_list);
720 			}
721 		}
722 
723 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
724 	}
725 
726 unlock:
727 	spin_unlock_irqrestore(&priv->lock, flags);
728 }
729 
730 void ipoib_mcast_dev_flush(struct net_device *dev)
731 {
732 	struct ipoib_dev_priv *priv = netdev_priv(dev);
733 	LIST_HEAD(remove_list);
734 	struct ipoib_mcast *mcast, *tmcast;
735 	unsigned long flags;
736 
737 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
738 
739 	spin_lock_irqsave(&priv->lock, flags);
740 
741 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
742 		list_del(&mcast->list);
743 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
744 		list_add_tail(&mcast->list, &remove_list);
745 	}
746 
747 	if (priv->broadcast) {
748 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
749 		list_add_tail(&priv->broadcast->list, &remove_list);
750 		priv->broadcast = NULL;
751 	}
752 
753 	spin_unlock_irqrestore(&priv->lock, flags);
754 
755 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
756 		ipoib_mcast_leave(dev, mcast);
757 		ipoib_mcast_free(mcast);
758 	}
759 }
760 
761 void ipoib_mcast_restart_task(struct work_struct *work)
762 {
763 	struct ipoib_dev_priv *priv =
764 		container_of(work, struct ipoib_dev_priv, restart_task);
765 	struct net_device *dev = priv->dev;
766 	struct dev_mc_list *mclist;
767 	struct ipoib_mcast *mcast, *tmcast;
768 	LIST_HEAD(remove_list);
769 	unsigned long flags;
770 	struct ib_sa_mcmember_rec rec;
771 
772 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
773 
774 	ipoib_mcast_stop_thread(dev, 0);
775 
776 	local_irq_save(flags);
777 	netif_addr_lock(dev);
778 	spin_lock(&priv->lock);
779 
780 	/*
781 	 * Unfortunately, the networking core only gives us a list of all of
782 	 * the multicast hardware addresses. We need to figure out which ones
783 	 * are new and which ones have been removed
784 	 */
785 
786 	/* Clear out the found flag */
787 	list_for_each_entry(mcast, &priv->multicast_list, list)
788 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
789 
790 	/* Mark all of the entries that are found or don't exist */
791 	for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
792 		union ib_gid mgid;
793 
794 		memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
795 
796 		mcast = __ipoib_mcast_find(dev, &mgid);
797 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
798 			struct ipoib_mcast *nmcast;
799 
800 			/* ignore group which is directly joined by userspace */
801 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
802 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
803 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
804 						mgid.raw);
805 				continue;
806 			}
807 
808 			/* Not found or send-only group, let's add a new entry */
809 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
810 					mgid.raw);
811 
812 			nmcast = ipoib_mcast_alloc(dev, 0);
813 			if (!nmcast) {
814 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
815 				continue;
816 			}
817 
818 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
819 
820 			nmcast->mcmember.mgid = mgid;
821 
822 			if (mcast) {
823 				/* Destroy the send only entry */
824 				list_move_tail(&mcast->list, &remove_list);
825 
826 				rb_replace_node(&mcast->rb_node,
827 						&nmcast->rb_node,
828 						&priv->multicast_tree);
829 			} else
830 				__ipoib_mcast_add(dev, nmcast);
831 
832 			list_add_tail(&nmcast->list, &priv->multicast_list);
833 		}
834 
835 		if (mcast)
836 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
837 	}
838 
839 	/* Remove all of the entries don't exist anymore */
840 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
841 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
842 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
843 			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
844 					mcast->mcmember.mgid.raw);
845 
846 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
847 
848 			/* Move to the remove list */
849 			list_move_tail(&mcast->list, &remove_list);
850 		}
851 	}
852 
853 	spin_unlock(&priv->lock);
854 	netif_addr_unlock(dev);
855 	local_irq_restore(flags);
856 
857 	/* We have to cancel outside of the spinlock */
858 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
859 		ipoib_mcast_leave(mcast->dev, mcast);
860 		ipoib_mcast_free(mcast);
861 	}
862 
863 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
864 		ipoib_mcast_start_thread(dev);
865 }
866 
867 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
868 
869 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
870 {
871 	struct ipoib_mcast_iter *iter;
872 
873 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
874 	if (!iter)
875 		return NULL;
876 
877 	iter->dev = dev;
878 	memset(iter->mgid.raw, 0, 16);
879 
880 	if (ipoib_mcast_iter_next(iter)) {
881 		kfree(iter);
882 		return NULL;
883 	}
884 
885 	return iter;
886 }
887 
888 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
889 {
890 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
891 	struct rb_node *n;
892 	struct ipoib_mcast *mcast;
893 	int ret = 1;
894 
895 	spin_lock_irq(&priv->lock);
896 
897 	n = rb_first(&priv->multicast_tree);
898 
899 	while (n) {
900 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
901 
902 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
903 			   sizeof (union ib_gid)) < 0) {
904 			iter->mgid      = mcast->mcmember.mgid;
905 			iter->created   = mcast->created;
906 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
907 			iter->complete  = !!mcast->ah;
908 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
909 
910 			ret = 0;
911 
912 			break;
913 		}
914 
915 		n = rb_next(n);
916 	}
917 
918 	spin_unlock_irq(&priv->lock);
919 
920 	return ret;
921 }
922 
923 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
924 			   union ib_gid *mgid,
925 			   unsigned long *created,
926 			   unsigned int *queuelen,
927 			   unsigned int *complete,
928 			   unsigned int *send_only)
929 {
930 	*mgid      = iter->mgid;
931 	*created   = iter->created;
932 	*queuelen  = iter->queuelen;
933 	*complete  = iter->complete;
934 	*send_only = iter->send_only;
935 }
936 
937 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
938