xref: /linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
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  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
35  */
36 
37 #include <linux/skbuff.h>
38 #include <linux/rtnetlink.h>
39 #include <linux/ip.h>
40 #include <linux/in.h>
41 #include <linux/igmp.h>
42 #include <linux/inetdevice.h>
43 #include <linux/delay.h>
44 #include <linux/completion.h>
45 
46 #include <net/dst.h>
47 
48 #include "ipoib.h"
49 
50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
51 static int mcast_debug_level;
52 
53 module_param(mcast_debug_level, int, 0644);
54 MODULE_PARM_DESC(mcast_debug_level,
55 		 "Enable multicast debug tracing if > 0");
56 #endif
57 
58 static DEFINE_MUTEX(mcast_mutex);
59 
60 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
61 struct ipoib_mcast {
62 	struct ib_sa_mcmember_rec mcmember;
63 	struct ib_sa_multicast	 *mc;
64 	struct ipoib_ah          *ah;
65 
66 	struct rb_node    rb_node;
67 	struct list_head  list;
68 
69 	unsigned long created;
70 	unsigned long backoff;
71 
72 	unsigned long flags;
73 	unsigned char logcount;
74 
75 	struct list_head  neigh_list;
76 
77 	struct sk_buff_head pkt_queue;
78 
79 	struct net_device *dev;
80 };
81 
82 struct ipoib_mcast_iter {
83 	struct net_device *dev;
84 	union ib_gid       mgid;
85 	unsigned long      created;
86 	unsigned int       queuelen;
87 	unsigned int       complete;
88 	unsigned int       send_only;
89 };
90 
91 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
92 {
93 	struct net_device *dev = mcast->dev;
94 	struct ipoib_dev_priv *priv = netdev_priv(dev);
95 	struct ipoib_neigh *neigh, *tmp;
96 	unsigned long flags;
97 	int tx_dropped = 0;
98 
99 	ipoib_dbg_mcast(netdev_priv(dev),
100 			"deleting multicast group " IPOIB_GID_FMT "\n",
101 			IPOIB_GID_ARG(mcast->mcmember.mgid));
102 
103 	spin_lock_irqsave(&priv->lock, flags);
104 
105 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
106 		/*
107 		 * It's safe to call ipoib_put_ah() inside priv->lock
108 		 * here, because we know that mcast->ah will always
109 		 * hold one more reference, so ipoib_put_ah() will
110 		 * never do more than decrement the ref count.
111 		 */
112 		if (neigh->ah)
113 			ipoib_put_ah(neigh->ah);
114 		ipoib_neigh_free(dev, neigh);
115 	}
116 
117 	spin_unlock_irqrestore(&priv->lock, flags);
118 
119 	if (mcast->ah)
120 		ipoib_put_ah(mcast->ah);
121 
122 	while (!skb_queue_empty(&mcast->pkt_queue)) {
123 		++tx_dropped;
124 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
125 	}
126 
127 	spin_lock_irqsave(&priv->tx_lock, flags);
128 	priv->stats.tx_dropped += tx_dropped;
129 	spin_unlock_irqrestore(&priv->tx_lock, flags);
130 
131 	kfree(mcast);
132 }
133 
134 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
135 					     int can_sleep)
136 {
137 	struct ipoib_mcast *mcast;
138 
139 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
140 	if (!mcast)
141 		return NULL;
142 
143 	mcast->dev = dev;
144 	mcast->created = jiffies;
145 	mcast->backoff = 1;
146 
147 	INIT_LIST_HEAD(&mcast->list);
148 	INIT_LIST_HEAD(&mcast->neigh_list);
149 	skb_queue_head_init(&mcast->pkt_queue);
150 
151 	return mcast;
152 }
153 
154 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
155 {
156 	struct ipoib_dev_priv *priv = netdev_priv(dev);
157 	struct rb_node *n = priv->multicast_tree.rb_node;
158 
159 	while (n) {
160 		struct ipoib_mcast *mcast;
161 		int ret;
162 
163 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
164 
165 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
166 			     sizeof (union ib_gid));
167 		if (ret < 0)
168 			n = n->rb_left;
169 		else if (ret > 0)
170 			n = n->rb_right;
171 		else
172 			return mcast;
173 	}
174 
175 	return NULL;
176 }
177 
178 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
179 {
180 	struct ipoib_dev_priv *priv = netdev_priv(dev);
181 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
182 
183 	while (*n) {
184 		struct ipoib_mcast *tmcast;
185 		int ret;
186 
187 		pn = *n;
188 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
189 
190 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
191 			     sizeof (union ib_gid));
192 		if (ret < 0)
193 			n = &pn->rb_left;
194 		else if (ret > 0)
195 			n = &pn->rb_right;
196 		else
197 			return -EEXIST;
198 	}
199 
200 	rb_link_node(&mcast->rb_node, pn, n);
201 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
202 
203 	return 0;
204 }
205 
206 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
207 				   struct ib_sa_mcmember_rec *mcmember)
208 {
209 	struct net_device *dev = mcast->dev;
210 	struct ipoib_dev_priv *priv = netdev_priv(dev);
211 	struct ipoib_ah *ah;
212 	int ret;
213 
214 	mcast->mcmember = *mcmember;
215 
216 	/* Set the cached Q_Key before we attach if it's the broadcast group */
217 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
218 		    sizeof (union ib_gid))) {
219 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
220 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
221 	}
222 
223 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
224 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
225 			ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
226 				   " already attached\n",
227 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
228 
229 			return 0;
230 		}
231 
232 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
233 					 &mcast->mcmember.mgid);
234 		if (ret < 0) {
235 			ipoib_warn(priv, "couldn't attach QP to multicast group "
236 				   IPOIB_GID_FMT "\n",
237 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
238 
239 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
240 			return ret;
241 		}
242 	}
243 
244 	{
245 		struct ib_ah_attr av = {
246 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
247 			.port_num      = priv->port,
248 			.sl	       = mcast->mcmember.sl,
249 			.ah_flags      = IB_AH_GRH,
250 			.static_rate   = mcast->mcmember.rate,
251 			.grh	       = {
252 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
253 				.hop_limit     = mcast->mcmember.hop_limit,
254 				.sgid_index    = 0,
255 				.traffic_class = mcast->mcmember.traffic_class
256 			}
257 		};
258 		av.grh.dgid = mcast->mcmember.mgid;
259 
260 		ah = ipoib_create_ah(dev, priv->pd, &av);
261 		if (!ah) {
262 			ipoib_warn(priv, "ib_address_create failed\n");
263 		} else {
264 			spin_lock_irq(&priv->lock);
265 			mcast->ah = ah;
266 			spin_unlock_irq(&priv->lock);
267 
268 			ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
269 					" AV %p, LID 0x%04x, SL %d\n",
270 					IPOIB_GID_ARG(mcast->mcmember.mgid),
271 					mcast->ah->ah,
272 					be16_to_cpu(mcast->mcmember.mlid),
273 					mcast->mcmember.sl);
274 		}
275 	}
276 
277 	/* actually send any queued packets */
278 	spin_lock_irq(&priv->tx_lock);
279 	while (!skb_queue_empty(&mcast->pkt_queue)) {
280 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
281 		spin_unlock_irq(&priv->tx_lock);
282 
283 		skb->dev = dev;
284 
285 		if (!skb->dst || !skb->dst->neighbour) {
286 			/* put pseudoheader back on for next time */
287 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
288 		}
289 
290 		if (dev_queue_xmit(skb))
291 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
292 		spin_lock_irq(&priv->tx_lock);
293 	}
294 	spin_unlock_irq(&priv->tx_lock);
295 
296 	return 0;
297 }
298 
299 static int
300 ipoib_mcast_sendonly_join_complete(int status,
301 				   struct ib_sa_multicast *multicast)
302 {
303 	struct ipoib_mcast *mcast = multicast->context;
304 	struct net_device *dev = mcast->dev;
305 	struct ipoib_dev_priv *priv = netdev_priv(dev);
306 
307 	/* We trap for port events ourselves. */
308 	if (status == -ENETRESET)
309 		return 0;
310 
311 	if (!status)
312 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
313 
314 	if (status) {
315 		if (mcast->logcount++ < 20)
316 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
317 					IPOIB_GID_FMT ", status %d\n",
318 					IPOIB_GID_ARG(mcast->mcmember.mgid), status);
319 
320 		/* Flush out any queued packets */
321 		spin_lock_irq(&priv->tx_lock);
322 		while (!skb_queue_empty(&mcast->pkt_queue)) {
323 			++priv->stats.tx_dropped;
324 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
325 		}
326 		spin_unlock_irq(&priv->tx_lock);
327 
328 		/* Clear the busy flag so we try again */
329 		status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
330 					    &mcast->flags);
331 	}
332 	return status;
333 }
334 
335 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
336 {
337 	struct net_device *dev = mcast->dev;
338 	struct ipoib_dev_priv *priv = netdev_priv(dev);
339 	struct ib_sa_mcmember_rec rec = {
340 #if 0				/* Some SMs don't support send-only yet */
341 		.join_state = 4
342 #else
343 		.join_state = 1
344 #endif
345 	};
346 	int ret = 0;
347 
348 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
349 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
350 		return -ENODEV;
351 	}
352 
353 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
354 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
355 		return -EBUSY;
356 	}
357 
358 	rec.mgid     = mcast->mcmember.mgid;
359 	rec.port_gid = priv->local_gid;
360 	rec.pkey     = cpu_to_be16(priv->pkey);
361 
362 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
363 					 priv->port, &rec,
364 					 IB_SA_MCMEMBER_REC_MGID	|
365 					 IB_SA_MCMEMBER_REC_PORT_GID	|
366 					 IB_SA_MCMEMBER_REC_PKEY	|
367 					 IB_SA_MCMEMBER_REC_JOIN_STATE,
368 					 GFP_ATOMIC,
369 					 ipoib_mcast_sendonly_join_complete,
370 					 mcast);
371 	if (IS_ERR(mcast->mc)) {
372 		ret = PTR_ERR(mcast->mc);
373 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
374 		ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
375 			   ret);
376 	} else {
377 		ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
378 				", starting join\n",
379 				IPOIB_GID_ARG(mcast->mcmember.mgid));
380 	}
381 
382 	return ret;
383 }
384 
385 static int ipoib_mcast_join_complete(int status,
386 				     struct ib_sa_multicast *multicast)
387 {
388 	struct ipoib_mcast *mcast = multicast->context;
389 	struct net_device *dev = mcast->dev;
390 	struct ipoib_dev_priv *priv = netdev_priv(dev);
391 
392 	ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
393 			" (status %d)\n",
394 			IPOIB_GID_ARG(mcast->mcmember.mgid), status);
395 
396 	/* We trap for port events ourselves. */
397 	if (status == -ENETRESET)
398 		return 0;
399 
400 	if (!status)
401 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
402 
403 	if (!status) {
404 		mcast->backoff = 1;
405 		mutex_lock(&mcast_mutex);
406 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
407 			queue_delayed_work(ipoib_workqueue,
408 					   &priv->mcast_task, 0);
409 		mutex_unlock(&mcast_mutex);
410 
411 		if (mcast == priv->broadcast)
412 			netif_carrier_on(dev);
413 
414 		return 0;
415 	}
416 
417 	if (mcast->logcount++ < 20) {
418 		if (status == -ETIMEDOUT) {
419 			ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
420 					", status %d\n",
421 					IPOIB_GID_ARG(mcast->mcmember.mgid),
422 					status);
423 		} else {
424 			ipoib_warn(priv, "multicast join failed for "
425 				   IPOIB_GID_FMT ", status %d\n",
426 				   IPOIB_GID_ARG(mcast->mcmember.mgid),
427 				   status);
428 		}
429 	}
430 
431 	mcast->backoff *= 2;
432 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
433 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
434 
435 	/* Clear the busy flag so we try again */
436 	status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
437 
438 	mutex_lock(&mcast_mutex);
439 	spin_lock_irq(&priv->lock);
440 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
441 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
442 				   mcast->backoff * HZ);
443 	spin_unlock_irq(&priv->lock);
444 	mutex_unlock(&mcast_mutex);
445 
446 	return status;
447 }
448 
449 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
450 			     int create)
451 {
452 	struct ipoib_dev_priv *priv = netdev_priv(dev);
453 	struct ib_sa_mcmember_rec rec = {
454 		.join_state = 1
455 	};
456 	ib_sa_comp_mask comp_mask;
457 	int ret = 0;
458 
459 	ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
460 			IPOIB_GID_ARG(mcast->mcmember.mgid));
461 
462 	rec.mgid     = mcast->mcmember.mgid;
463 	rec.port_gid = priv->local_gid;
464 	rec.pkey     = cpu_to_be16(priv->pkey);
465 
466 	comp_mask =
467 		IB_SA_MCMEMBER_REC_MGID		|
468 		IB_SA_MCMEMBER_REC_PORT_GID	|
469 		IB_SA_MCMEMBER_REC_PKEY		|
470 		IB_SA_MCMEMBER_REC_JOIN_STATE;
471 
472 	if (create) {
473 		comp_mask |=
474 			IB_SA_MCMEMBER_REC_QKEY			|
475 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
476 			IB_SA_MCMEMBER_REC_MTU			|
477 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
478 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
479 			IB_SA_MCMEMBER_REC_RATE			|
480 			IB_SA_MCMEMBER_REC_SL			|
481 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
482 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
483 
484 		rec.qkey	  = priv->broadcast->mcmember.qkey;
485 		rec.mtu_selector  = IB_SA_EQ;
486 		rec.mtu		  = priv->broadcast->mcmember.mtu;
487 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
488 		rec.rate_selector = IB_SA_EQ;
489 		rec.rate	  = priv->broadcast->mcmember.rate;
490 		rec.sl		  = priv->broadcast->mcmember.sl;
491 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
492 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
493 	}
494 
495 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
496 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
497 					 &rec, comp_mask, GFP_KERNEL,
498 					 ipoib_mcast_join_complete, mcast);
499 	if (IS_ERR(mcast->mc)) {
500 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
501 		ret = PTR_ERR(mcast->mc);
502 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
503 
504 		mcast->backoff *= 2;
505 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
506 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
507 
508 		mutex_lock(&mcast_mutex);
509 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
510 			queue_delayed_work(ipoib_workqueue,
511 					   &priv->mcast_task,
512 					   mcast->backoff * HZ);
513 		mutex_unlock(&mcast_mutex);
514 	}
515 }
516 
517 void ipoib_mcast_join_task(struct work_struct *work)
518 {
519 	struct ipoib_dev_priv *priv =
520 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
521 	struct net_device *dev = priv->dev;
522 
523 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
524 		return;
525 
526 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
527 		ipoib_warn(priv, "ib_gid_entry_get() failed\n");
528 	else
529 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
530 
531 	{
532 		struct ib_port_attr attr;
533 
534 		if (!ib_query_port(priv->ca, priv->port, &attr))
535 			priv->local_lid = attr.lid;
536 		else
537 			ipoib_warn(priv, "ib_query_port failed\n");
538 	}
539 
540 	if (!priv->broadcast) {
541 		struct ipoib_mcast *broadcast;
542 
543 		broadcast = ipoib_mcast_alloc(dev, 1);
544 		if (!broadcast) {
545 			ipoib_warn(priv, "failed to allocate broadcast group\n");
546 			mutex_lock(&mcast_mutex);
547 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
548 				queue_delayed_work(ipoib_workqueue,
549 						   &priv->mcast_task, HZ);
550 			mutex_unlock(&mcast_mutex);
551 			return;
552 		}
553 
554 		spin_lock_irq(&priv->lock);
555 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
556 		       sizeof (union ib_gid));
557 		priv->broadcast = broadcast;
558 
559 		__ipoib_mcast_add(dev, priv->broadcast);
560 		spin_unlock_irq(&priv->lock);
561 	}
562 
563 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
564 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
565 			ipoib_mcast_join(dev, priv->broadcast, 0);
566 		return;
567 	}
568 
569 	while (1) {
570 		struct ipoib_mcast *mcast = NULL;
571 
572 		spin_lock_irq(&priv->lock);
573 		list_for_each_entry(mcast, &priv->multicast_list, list) {
574 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
575 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
576 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
577 				/* Found the next unjoined group */
578 				break;
579 			}
580 		}
581 		spin_unlock_irq(&priv->lock);
582 
583 		if (&mcast->list == &priv->multicast_list) {
584 			/* All done */
585 			break;
586 		}
587 
588 		ipoib_mcast_join(dev, mcast, 1);
589 		return;
590 	}
591 
592 	priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
593 		IPOIB_ENCAP_LEN;
594 
595 	if (!ipoib_cm_admin_enabled(dev))
596 		dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
597 
598 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
599 
600 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
601 }
602 
603 int ipoib_mcast_start_thread(struct net_device *dev)
604 {
605 	struct ipoib_dev_priv *priv = netdev_priv(dev);
606 
607 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
608 
609 	mutex_lock(&mcast_mutex);
610 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
611 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
612 	mutex_unlock(&mcast_mutex);
613 
614 	spin_lock_irq(&priv->lock);
615 	set_bit(IPOIB_MCAST_STARTED, &priv->flags);
616 	spin_unlock_irq(&priv->lock);
617 
618 	return 0;
619 }
620 
621 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
622 {
623 	struct ipoib_dev_priv *priv = netdev_priv(dev);
624 
625 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
626 
627 	spin_lock_irq(&priv->lock);
628 	clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
629 	spin_unlock_irq(&priv->lock);
630 
631 	mutex_lock(&mcast_mutex);
632 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
633 	cancel_delayed_work(&priv->mcast_task);
634 	mutex_unlock(&mcast_mutex);
635 
636 	if (flush)
637 		flush_workqueue(ipoib_workqueue);
638 
639 	return 0;
640 }
641 
642 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
643 {
644 	struct ipoib_dev_priv *priv = netdev_priv(dev);
645 	int ret = 0;
646 
647 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
648 		ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
649 				IPOIB_GID_ARG(mcast->mcmember.mgid));
650 
651 		/* Remove ourselves from the multicast group */
652 		ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
653 					 &mcast->mcmember.mgid);
654 		if (ret)
655 			ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
656 	}
657 
658 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
659 		ib_sa_free_multicast(mcast->mc);
660 
661 	return 0;
662 }
663 
664 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
665 {
666 	struct ipoib_dev_priv *priv = netdev_priv(dev);
667 	struct ipoib_mcast *mcast;
668 
669 	/*
670 	 * We can only be called from ipoib_start_xmit, so we're
671 	 * inside tx_lock -- no need to save/restore flags.
672 	 */
673 	spin_lock(&priv->lock);
674 
675 	if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)	||
676 	    !priv->broadcast					||
677 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
678 		++priv->stats.tx_dropped;
679 		dev_kfree_skb_any(skb);
680 		goto unlock;
681 	}
682 
683 	mcast = __ipoib_mcast_find(dev, mgid);
684 	if (!mcast) {
685 		/* Let's create a new send only group now */
686 		ipoib_dbg_mcast(priv, "setting up send only multicast group for "
687 				IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
688 
689 		mcast = ipoib_mcast_alloc(dev, 0);
690 		if (!mcast) {
691 			ipoib_warn(priv, "unable to allocate memory for "
692 				   "multicast structure\n");
693 			++priv->stats.tx_dropped;
694 			dev_kfree_skb_any(skb);
695 			goto out;
696 		}
697 
698 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
699 		memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
700 		__ipoib_mcast_add(dev, mcast);
701 		list_add_tail(&mcast->list, &priv->multicast_list);
702 	}
703 
704 	if (!mcast->ah) {
705 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
706 			skb_queue_tail(&mcast->pkt_queue, skb);
707 		else {
708 			++priv->stats.tx_dropped;
709 			dev_kfree_skb_any(skb);
710 		}
711 
712 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
713 			ipoib_dbg_mcast(priv, "no address vector, "
714 					"but multicast join already started\n");
715 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
716 			ipoib_mcast_sendonly_join(mcast);
717 
718 		/*
719 		 * If lookup completes between here and out:, don't
720 		 * want to send packet twice.
721 		 */
722 		mcast = NULL;
723 	}
724 
725 out:
726 	if (mcast && mcast->ah) {
727 		if (skb->dst            &&
728 		    skb->dst->neighbour &&
729 		    !*to_ipoib_neigh(skb->dst->neighbour)) {
730 			struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour);
731 
732 			if (neigh) {
733 				kref_get(&mcast->ah->ref);
734 				neigh->ah  	= mcast->ah;
735 				list_add_tail(&neigh->list, &mcast->neigh_list);
736 			}
737 		}
738 
739 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
740 	}
741 
742 unlock:
743 	spin_unlock(&priv->lock);
744 }
745 
746 void ipoib_mcast_dev_flush(struct net_device *dev)
747 {
748 	struct ipoib_dev_priv *priv = netdev_priv(dev);
749 	LIST_HEAD(remove_list);
750 	struct ipoib_mcast *mcast, *tmcast;
751 	unsigned long flags;
752 
753 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
754 
755 	spin_lock_irqsave(&priv->lock, flags);
756 
757 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
758 		list_del(&mcast->list);
759 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
760 		list_add_tail(&mcast->list, &remove_list);
761 	}
762 
763 	if (priv->broadcast) {
764 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
765 		list_add_tail(&priv->broadcast->list, &remove_list);
766 		priv->broadcast = NULL;
767 	}
768 
769 	spin_unlock_irqrestore(&priv->lock, flags);
770 
771 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
772 		ipoib_mcast_leave(dev, mcast);
773 		ipoib_mcast_free(mcast);
774 	}
775 }
776 
777 void ipoib_mcast_restart_task(struct work_struct *work)
778 {
779 	struct ipoib_dev_priv *priv =
780 		container_of(work, struct ipoib_dev_priv, restart_task);
781 	struct net_device *dev = priv->dev;
782 	struct dev_mc_list *mclist;
783 	struct ipoib_mcast *mcast, *tmcast;
784 	LIST_HEAD(remove_list);
785 	unsigned long flags;
786 
787 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
788 
789 	ipoib_mcast_stop_thread(dev, 0);
790 
791 	local_irq_save(flags);
792 	netif_tx_lock(dev);
793 	spin_lock(&priv->lock);
794 
795 	/*
796 	 * Unfortunately, the networking core only gives us a list of all of
797 	 * the multicast hardware addresses. We need to figure out which ones
798 	 * are new and which ones have been removed
799 	 */
800 
801 	/* Clear out the found flag */
802 	list_for_each_entry(mcast, &priv->multicast_list, list)
803 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
804 
805 	/* Mark all of the entries that are found or don't exist */
806 	for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
807 		union ib_gid mgid;
808 
809 		memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
810 
811 		/* Add in the P_Key */
812 		mgid.raw[4] = (priv->pkey >> 8) & 0xff;
813 		mgid.raw[5] = priv->pkey & 0xff;
814 
815 		mcast = __ipoib_mcast_find(dev, &mgid);
816 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
817 			struct ipoib_mcast *nmcast;
818 
819 			/* Not found or send-only group, let's add a new entry */
820 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
821 					IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
822 
823 			nmcast = ipoib_mcast_alloc(dev, 0);
824 			if (!nmcast) {
825 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
826 				continue;
827 			}
828 
829 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
830 
831 			nmcast->mcmember.mgid = mgid;
832 
833 			if (mcast) {
834 				/* Destroy the send only entry */
835 				list_move_tail(&mcast->list, &remove_list);
836 
837 				rb_replace_node(&mcast->rb_node,
838 						&nmcast->rb_node,
839 						&priv->multicast_tree);
840 			} else
841 				__ipoib_mcast_add(dev, nmcast);
842 
843 			list_add_tail(&nmcast->list, &priv->multicast_list);
844 		}
845 
846 		if (mcast)
847 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
848 	}
849 
850 	/* Remove all of the entries don't exist anymore */
851 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
852 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
853 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
854 			ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
855 					IPOIB_GID_ARG(mcast->mcmember.mgid));
856 
857 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
858 
859 			/* Move to the remove list */
860 			list_move_tail(&mcast->list, &remove_list);
861 		}
862 	}
863 
864 	spin_unlock(&priv->lock);
865 	netif_tx_unlock(dev);
866 	local_irq_restore(flags);
867 
868 	/* We have to cancel outside of the spinlock */
869 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
870 		ipoib_mcast_leave(mcast->dev, mcast);
871 		ipoib_mcast_free(mcast);
872 	}
873 
874 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
875 		ipoib_mcast_start_thread(dev);
876 }
877 
878 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
879 
880 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
881 {
882 	struct ipoib_mcast_iter *iter;
883 
884 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
885 	if (!iter)
886 		return NULL;
887 
888 	iter->dev = dev;
889 	memset(iter->mgid.raw, 0, 16);
890 
891 	if (ipoib_mcast_iter_next(iter)) {
892 		kfree(iter);
893 		return NULL;
894 	}
895 
896 	return iter;
897 }
898 
899 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
900 {
901 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
902 	struct rb_node *n;
903 	struct ipoib_mcast *mcast;
904 	int ret = 1;
905 
906 	spin_lock_irq(&priv->lock);
907 
908 	n = rb_first(&priv->multicast_tree);
909 
910 	while (n) {
911 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
912 
913 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
914 			   sizeof (union ib_gid)) < 0) {
915 			iter->mgid      = mcast->mcmember.mgid;
916 			iter->created   = mcast->created;
917 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
918 			iter->complete  = !!mcast->ah;
919 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
920 
921 			ret = 0;
922 
923 			break;
924 		}
925 
926 		n = rb_next(n);
927 	}
928 
929 	spin_unlock_irq(&priv->lock);
930 
931 	return ret;
932 }
933 
934 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
935 			   union ib_gid *mgid,
936 			   unsigned long *created,
937 			   unsigned int *queuelen,
938 			   unsigned int *complete,
939 			   unsigned int *send_only)
940 {
941 	*mgid      = iter->mgid;
942 	*created   = iter->created;
943 	*queuelen  = iter->queuelen;
944 	*complete  = iter->complete;
945 	*send_only = iter->send_only;
946 }
947 
948 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
949