xref: /linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
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/moduleparam.h>
38 #include <linux/ip.h>
39 #include <linux/in.h>
40 #include <linux/igmp.h>
41 #include <linux/inetdevice.h>
42 #include <linux/delay.h>
43 #include <linux/completion.h>
44 #include <linux/slab.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 struct ipoib_mcast_iter {
61 	struct net_device *dev;
62 	union ib_gid       mgid;
63 	unsigned long      created;
64 	unsigned int       queuelen;
65 	unsigned int       complete;
66 	unsigned int       send_only;
67 };
68 
69 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
70 {
71 	struct net_device *dev = mcast->dev;
72 	struct ipoib_dev_priv *priv = netdev_priv(dev);
73 	struct ipoib_neigh *neigh, *tmp;
74 	int tx_dropped = 0;
75 
76 	ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
77 			mcast->mcmember.mgid.raw);
78 
79 	spin_lock_irq(&priv->lock);
80 
81 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
82 		/*
83 		 * It's safe to call ipoib_put_ah() inside priv->lock
84 		 * here, because we know that mcast->ah will always
85 		 * hold one more reference, so ipoib_put_ah() will
86 		 * never do more than decrement the ref count.
87 		 */
88 		if (neigh->ah)
89 			ipoib_put_ah(neigh->ah);
90 		ipoib_neigh_free(dev, neigh);
91 	}
92 
93 	spin_unlock_irq(&priv->lock);
94 
95 	if (mcast->ah)
96 		ipoib_put_ah(mcast->ah);
97 
98 	while (!skb_queue_empty(&mcast->pkt_queue)) {
99 		++tx_dropped;
100 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
101 	}
102 
103 	netif_tx_lock_bh(dev);
104 	dev->stats.tx_dropped += tx_dropped;
105 	netif_tx_unlock_bh(dev);
106 
107 	kfree(mcast);
108 }
109 
110 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
111 					     int can_sleep)
112 {
113 	struct ipoib_mcast *mcast;
114 
115 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
116 	if (!mcast)
117 		return NULL;
118 
119 	mcast->dev = dev;
120 	mcast->created = jiffies;
121 	mcast->backoff = 1;
122 
123 	INIT_LIST_HEAD(&mcast->list);
124 	INIT_LIST_HEAD(&mcast->neigh_list);
125 	skb_queue_head_init(&mcast->pkt_queue);
126 
127 	return mcast;
128 }
129 
130 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
131 {
132 	struct ipoib_dev_priv *priv = netdev_priv(dev);
133 	struct rb_node *n = priv->multicast_tree.rb_node;
134 
135 	while (n) {
136 		struct ipoib_mcast *mcast;
137 		int ret;
138 
139 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
140 
141 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
142 			     sizeof (union ib_gid));
143 		if (ret < 0)
144 			n = n->rb_left;
145 		else if (ret > 0)
146 			n = n->rb_right;
147 		else
148 			return mcast;
149 	}
150 
151 	return NULL;
152 }
153 
154 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
155 {
156 	struct ipoib_dev_priv *priv = netdev_priv(dev);
157 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
158 
159 	while (*n) {
160 		struct ipoib_mcast *tmcast;
161 		int ret;
162 
163 		pn = *n;
164 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
165 
166 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
167 			     sizeof (union ib_gid));
168 		if (ret < 0)
169 			n = &pn->rb_left;
170 		else if (ret > 0)
171 			n = &pn->rb_right;
172 		else
173 			return -EEXIST;
174 	}
175 
176 	rb_link_node(&mcast->rb_node, pn, n);
177 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
178 
179 	return 0;
180 }
181 
182 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
183 				   struct ib_sa_mcmember_rec *mcmember)
184 {
185 	struct net_device *dev = mcast->dev;
186 	struct ipoib_dev_priv *priv = netdev_priv(dev);
187 	struct ipoib_ah *ah;
188 	int ret;
189 	int set_qkey = 0;
190 
191 	mcast->mcmember = *mcmember;
192 
193 	/* Set the cached Q_Key before we attach if it's the broadcast group */
194 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
195 		    sizeof (union ib_gid))) {
196 		spin_lock_irq(&priv->lock);
197 		if (!priv->broadcast) {
198 			spin_unlock_irq(&priv->lock);
199 			return -EAGAIN;
200 		}
201 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
202 		spin_unlock_irq(&priv->lock);
203 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
204 		set_qkey = 1;
205 	}
206 
207 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
208 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
209 			ipoib_warn(priv, "multicast group %pI6 already attached\n",
210 				   mcast->mcmember.mgid.raw);
211 
212 			return 0;
213 		}
214 
215 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
216 					 &mcast->mcmember.mgid, set_qkey);
217 		if (ret < 0) {
218 			ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
219 				   mcast->mcmember.mgid.raw);
220 
221 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
222 			return ret;
223 		}
224 	}
225 
226 	{
227 		struct ib_ah_attr av = {
228 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
229 			.port_num      = priv->port,
230 			.sl	       = mcast->mcmember.sl,
231 			.ah_flags      = IB_AH_GRH,
232 			.static_rate   = mcast->mcmember.rate,
233 			.grh	       = {
234 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
235 				.hop_limit     = mcast->mcmember.hop_limit,
236 				.sgid_index    = 0,
237 				.traffic_class = mcast->mcmember.traffic_class
238 			}
239 		};
240 		av.grh.dgid = mcast->mcmember.mgid;
241 
242 		ah = ipoib_create_ah(dev, priv->pd, &av);
243 		if (IS_ERR(ah)) {
244 			ipoib_warn(priv, "ib_address_create failed %ld\n",
245 				-PTR_ERR(ah));
246 			/* use original error */
247 			return PTR_ERR(ah);
248 		} else {
249 			spin_lock_irq(&priv->lock);
250 			mcast->ah = ah;
251 			spin_unlock_irq(&priv->lock);
252 
253 			ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
254 					mcast->mcmember.mgid.raw,
255 					mcast->ah->ah,
256 					be16_to_cpu(mcast->mcmember.mlid),
257 					mcast->mcmember.sl);
258 		}
259 	}
260 
261 	/* actually send any queued packets */
262 	netif_tx_lock_bh(dev);
263 	while (!skb_queue_empty(&mcast->pkt_queue)) {
264 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
265 		struct dst_entry *dst = skb_dst(skb);
266 		struct neighbour *n = NULL;
267 
268 		netif_tx_unlock_bh(dev);
269 
270 		skb->dev = dev;
271 		if (dst)
272 			n = dst_get_neighbour_noref_raw(dst);
273 		if (!dst || !n) {
274 			/* put pseudoheader back on for next time */
275 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
276 		}
277 
278 		if (dev_queue_xmit(skb))
279 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
280 		netif_tx_lock_bh(dev);
281 	}
282 	netif_tx_unlock_bh(dev);
283 
284 	return 0;
285 }
286 
287 static int
288 ipoib_mcast_sendonly_join_complete(int status,
289 				   struct ib_sa_multicast *multicast)
290 {
291 	struct ipoib_mcast *mcast = multicast->context;
292 	struct net_device *dev = mcast->dev;
293 
294 	/* We trap for port events ourselves. */
295 	if (status == -ENETRESET)
296 		return 0;
297 
298 	if (!status)
299 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
300 
301 	if (status) {
302 		if (mcast->logcount++ < 20)
303 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for %pI6, status %d\n",
304 					mcast->mcmember.mgid.raw, status);
305 
306 		/* Flush out any queued packets */
307 		netif_tx_lock_bh(dev);
308 		while (!skb_queue_empty(&mcast->pkt_queue)) {
309 			++dev->stats.tx_dropped;
310 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
311 		}
312 		netif_tx_unlock_bh(dev);
313 
314 		/* Clear the busy flag so we try again */
315 		status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
316 					    &mcast->flags);
317 	}
318 	return status;
319 }
320 
321 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
322 {
323 	struct net_device *dev = mcast->dev;
324 	struct ipoib_dev_priv *priv = netdev_priv(dev);
325 	struct ib_sa_mcmember_rec rec = {
326 #if 0				/* Some SMs don't support send-only yet */
327 		.join_state = 4
328 #else
329 		.join_state = 1
330 #endif
331 	};
332 	int ret = 0;
333 
334 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
335 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
336 		return -ENODEV;
337 	}
338 
339 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
340 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
341 		return -EBUSY;
342 	}
343 
344 	rec.mgid     = mcast->mcmember.mgid;
345 	rec.port_gid = priv->local_gid;
346 	rec.pkey     = cpu_to_be16(priv->pkey);
347 
348 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
349 					 priv->port, &rec,
350 					 IB_SA_MCMEMBER_REC_MGID	|
351 					 IB_SA_MCMEMBER_REC_PORT_GID	|
352 					 IB_SA_MCMEMBER_REC_PKEY	|
353 					 IB_SA_MCMEMBER_REC_JOIN_STATE,
354 					 GFP_ATOMIC,
355 					 ipoib_mcast_sendonly_join_complete,
356 					 mcast);
357 	if (IS_ERR(mcast->mc)) {
358 		ret = PTR_ERR(mcast->mc);
359 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
360 		ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
361 			   ret);
362 	} else {
363 		ipoib_dbg_mcast(priv, "no multicast record for %pI6, starting join\n",
364 				mcast->mcmember.mgid.raw);
365 	}
366 
367 	return ret;
368 }
369 
370 void ipoib_mcast_carrier_on_task(struct work_struct *work)
371 {
372 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
373 						   carrier_on_task);
374 	struct ib_port_attr attr;
375 
376 	/*
377 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
378 	 * turning the carrier back on while a device is being
379 	 * removed.
380 	 */
381 	if (ib_query_port(priv->ca, priv->port, &attr) ||
382 	    attr.state != IB_PORT_ACTIVE) {
383 		ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
384 		return;
385 	}
386 
387 	rtnl_lock();
388 	netif_carrier_on(priv->dev);
389 	rtnl_unlock();
390 }
391 
392 static int ipoib_mcast_join_complete(int status,
393 				     struct ib_sa_multicast *multicast)
394 {
395 	struct ipoib_mcast *mcast = multicast->context;
396 	struct net_device *dev = mcast->dev;
397 	struct ipoib_dev_priv *priv = netdev_priv(dev);
398 
399 	ipoib_dbg_mcast(priv, "join completion for %pI6 (status %d)\n",
400 			mcast->mcmember.mgid.raw, status);
401 
402 	/* We trap for port events ourselves. */
403 	if (status == -ENETRESET)
404 		return 0;
405 
406 	if (!status)
407 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
408 
409 	if (!status) {
410 		mcast->backoff = 1;
411 		mutex_lock(&mcast_mutex);
412 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
413 			queue_delayed_work(ipoib_workqueue,
414 					   &priv->mcast_task, 0);
415 		mutex_unlock(&mcast_mutex);
416 
417 		/*
418 		 * Defer carrier on work to ipoib_workqueue to avoid a
419 		 * deadlock on rtnl_lock here.
420 		 */
421 		if (mcast == priv->broadcast)
422 			queue_work(ipoib_workqueue, &priv->carrier_on_task);
423 
424 		return 0;
425 	}
426 
427 	if (mcast->logcount++ < 20) {
428 		if (status == -ETIMEDOUT || status == -EAGAIN) {
429 			ipoib_dbg_mcast(priv, "multicast join failed for %pI6, status %d\n",
430 					mcast->mcmember.mgid.raw, status);
431 		} else {
432 			ipoib_warn(priv, "multicast join failed for %pI6, status %d\n",
433 				   mcast->mcmember.mgid.raw, status);
434 		}
435 	}
436 
437 	mcast->backoff *= 2;
438 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
439 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
440 
441 	/* Clear the busy flag so we try again */
442 	status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
443 
444 	mutex_lock(&mcast_mutex);
445 	spin_lock_irq(&priv->lock);
446 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
447 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
448 				   mcast->backoff * HZ);
449 	spin_unlock_irq(&priv->lock);
450 	mutex_unlock(&mcast_mutex);
451 
452 	return status;
453 }
454 
455 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
456 			     int create)
457 {
458 	struct ipoib_dev_priv *priv = netdev_priv(dev);
459 	struct ib_sa_mcmember_rec rec = {
460 		.join_state = 1
461 	};
462 	ib_sa_comp_mask comp_mask;
463 	int ret = 0;
464 
465 	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
466 
467 	rec.mgid     = mcast->mcmember.mgid;
468 	rec.port_gid = priv->local_gid;
469 	rec.pkey     = cpu_to_be16(priv->pkey);
470 
471 	comp_mask =
472 		IB_SA_MCMEMBER_REC_MGID		|
473 		IB_SA_MCMEMBER_REC_PORT_GID	|
474 		IB_SA_MCMEMBER_REC_PKEY		|
475 		IB_SA_MCMEMBER_REC_JOIN_STATE;
476 
477 	if (create) {
478 		comp_mask |=
479 			IB_SA_MCMEMBER_REC_QKEY			|
480 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
481 			IB_SA_MCMEMBER_REC_MTU			|
482 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
483 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
484 			IB_SA_MCMEMBER_REC_RATE			|
485 			IB_SA_MCMEMBER_REC_SL			|
486 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
487 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
488 
489 		rec.qkey	  = priv->broadcast->mcmember.qkey;
490 		rec.mtu_selector  = IB_SA_EQ;
491 		rec.mtu		  = priv->broadcast->mcmember.mtu;
492 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
493 		rec.rate_selector = IB_SA_EQ;
494 		rec.rate	  = priv->broadcast->mcmember.rate;
495 		rec.sl		  = priv->broadcast->mcmember.sl;
496 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
497 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
498 	}
499 
500 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
501 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
502 					 &rec, comp_mask, GFP_KERNEL,
503 					 ipoib_mcast_join_complete, mcast);
504 	if (IS_ERR(mcast->mc)) {
505 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
506 		ret = PTR_ERR(mcast->mc);
507 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
508 
509 		mcast->backoff *= 2;
510 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
511 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
512 
513 		mutex_lock(&mcast_mutex);
514 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
515 			queue_delayed_work(ipoib_workqueue,
516 					   &priv->mcast_task,
517 					   mcast->backoff * HZ);
518 		mutex_unlock(&mcast_mutex);
519 	}
520 }
521 
522 void ipoib_mcast_join_task(struct work_struct *work)
523 {
524 	struct ipoib_dev_priv *priv =
525 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
526 	struct net_device *dev = priv->dev;
527 
528 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
529 		return;
530 
531 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
532 		ipoib_warn(priv, "ib_query_gid() failed\n");
533 	else
534 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
535 
536 	{
537 		struct ib_port_attr attr;
538 
539 		if (!ib_query_port(priv->ca, priv->port, &attr))
540 			priv->local_lid = attr.lid;
541 		else
542 			ipoib_warn(priv, "ib_query_port failed\n");
543 	}
544 
545 	if (!priv->broadcast) {
546 		struct ipoib_mcast *broadcast;
547 
548 		if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
549 			return;
550 
551 		broadcast = ipoib_mcast_alloc(dev, 1);
552 		if (!broadcast) {
553 			ipoib_warn(priv, "failed to allocate broadcast group\n");
554 			mutex_lock(&mcast_mutex);
555 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
556 				queue_delayed_work(ipoib_workqueue,
557 						   &priv->mcast_task, HZ);
558 			mutex_unlock(&mcast_mutex);
559 			return;
560 		}
561 
562 		spin_lock_irq(&priv->lock);
563 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
564 		       sizeof (union ib_gid));
565 		priv->broadcast = broadcast;
566 
567 		__ipoib_mcast_add(dev, priv->broadcast);
568 		spin_unlock_irq(&priv->lock);
569 	}
570 
571 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
572 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
573 			ipoib_mcast_join(dev, priv->broadcast, 0);
574 		return;
575 	}
576 
577 	while (1) {
578 		struct ipoib_mcast *mcast = NULL;
579 
580 		spin_lock_irq(&priv->lock);
581 		list_for_each_entry(mcast, &priv->multicast_list, list) {
582 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
583 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
584 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
585 				/* Found the next unjoined group */
586 				break;
587 			}
588 		}
589 		spin_unlock_irq(&priv->lock);
590 
591 		if (&mcast->list == &priv->multicast_list) {
592 			/* All done */
593 			break;
594 		}
595 
596 		ipoib_mcast_join(dev, mcast, 1);
597 		return;
598 	}
599 
600 	priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
601 
602 	if (!ipoib_cm_admin_enabled(dev)) {
603 		rtnl_lock();
604 		dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
605 		rtnl_unlock();
606 	}
607 
608 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
609 
610 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
611 }
612 
613 int ipoib_mcast_start_thread(struct net_device *dev)
614 {
615 	struct ipoib_dev_priv *priv = netdev_priv(dev);
616 
617 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
618 
619 	mutex_lock(&mcast_mutex);
620 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
621 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
622 	mutex_unlock(&mcast_mutex);
623 
624 	return 0;
625 }
626 
627 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
628 {
629 	struct ipoib_dev_priv *priv = netdev_priv(dev);
630 
631 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
632 
633 	mutex_lock(&mcast_mutex);
634 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
635 	cancel_delayed_work(&priv->mcast_task);
636 	mutex_unlock(&mcast_mutex);
637 
638 	if (flush)
639 		flush_workqueue(ipoib_workqueue);
640 
641 	return 0;
642 }
643 
644 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
645 {
646 	struct ipoib_dev_priv *priv = netdev_priv(dev);
647 	int ret = 0;
648 
649 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
650 		ib_sa_free_multicast(mcast->mc);
651 
652 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
653 		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
654 				mcast->mcmember.mgid.raw);
655 
656 		/* Remove ourselves from the multicast group */
657 		ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
658 				      be16_to_cpu(mcast->mcmember.mlid));
659 		if (ret)
660 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
661 	}
662 
663 	return 0;
664 }
665 
666 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
667 {
668 	struct ipoib_dev_priv *priv = netdev_priv(dev);
669 	struct ipoib_mcast *mcast;
670 	unsigned long flags;
671 
672 	spin_lock_irqsave(&priv->lock, flags);
673 
674 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
675 	    !priv->broadcast					||
676 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
677 		++dev->stats.tx_dropped;
678 		dev_kfree_skb_any(skb);
679 		goto unlock;
680 	}
681 
682 	mcast = __ipoib_mcast_find(dev, mgid);
683 	if (!mcast) {
684 		/* Let's create a new send only group now */
685 		ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
686 				mgid);
687 
688 		mcast = ipoib_mcast_alloc(dev, 0);
689 		if (!mcast) {
690 			ipoib_warn(priv, "unable to allocate memory for "
691 				   "multicast structure\n");
692 			++dev->stats.tx_dropped;
693 			dev_kfree_skb_any(skb);
694 			goto out;
695 		}
696 
697 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
698 		memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
699 		__ipoib_mcast_add(dev, mcast);
700 		list_add_tail(&mcast->list, &priv->multicast_list);
701 	}
702 
703 	if (!mcast->ah) {
704 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
705 			skb_queue_tail(&mcast->pkt_queue, skb);
706 		else {
707 			++dev->stats.tx_dropped;
708 			dev_kfree_skb_any(skb);
709 		}
710 
711 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
712 			ipoib_dbg_mcast(priv, "no address vector, "
713 					"but multicast join already started\n");
714 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
715 			ipoib_mcast_sendonly_join(mcast);
716 
717 		/*
718 		 * If lookup completes between here and out:, don't
719 		 * want to send packet twice.
720 		 */
721 		mcast = NULL;
722 	}
723 
724 out:
725 	if (mcast && mcast->ah) {
726 		struct dst_entry *dst = skb_dst(skb);
727 		struct neighbour *n = NULL;
728 
729 		rcu_read_lock();
730 		if (dst)
731 			n = dst_get_neighbour_noref(dst);
732 		if (n && !*to_ipoib_neigh(n)) {
733 			struct ipoib_neigh *neigh = ipoib_neigh_alloc(n,
734 								      skb->dev);
735 
736 			if (neigh) {
737 				kref_get(&mcast->ah->ref);
738 				neigh->ah	= mcast->ah;
739 				list_add_tail(&neigh->list, &mcast->neigh_list);
740 			}
741 		}
742 		rcu_read_unlock();
743 		spin_unlock_irqrestore(&priv->lock, flags);
744 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
745 		return;
746 	}
747 
748 unlock:
749 	spin_unlock_irqrestore(&priv->lock, flags);
750 }
751 
752 void ipoib_mcast_dev_flush(struct net_device *dev)
753 {
754 	struct ipoib_dev_priv *priv = netdev_priv(dev);
755 	LIST_HEAD(remove_list);
756 	struct ipoib_mcast *mcast, *tmcast;
757 	unsigned long flags;
758 
759 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
760 
761 	spin_lock_irqsave(&priv->lock, flags);
762 
763 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
764 		list_del(&mcast->list);
765 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
766 		list_add_tail(&mcast->list, &remove_list);
767 	}
768 
769 	if (priv->broadcast) {
770 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
771 		list_add_tail(&priv->broadcast->list, &remove_list);
772 		priv->broadcast = NULL;
773 	}
774 
775 	spin_unlock_irqrestore(&priv->lock, flags);
776 
777 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
778 		ipoib_mcast_leave(dev, mcast);
779 		ipoib_mcast_free(mcast);
780 	}
781 }
782 
783 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
784 {
785 	/* reserved QPN, prefix, scope */
786 	if (memcmp(addr, broadcast, 6))
787 		return 0;
788 	/* signature lower, pkey */
789 	if (memcmp(addr + 7, broadcast + 7, 3))
790 		return 0;
791 	return 1;
792 }
793 
794 void ipoib_mcast_restart_task(struct work_struct *work)
795 {
796 	struct ipoib_dev_priv *priv =
797 		container_of(work, struct ipoib_dev_priv, restart_task);
798 	struct net_device *dev = priv->dev;
799 	struct netdev_hw_addr *ha;
800 	struct ipoib_mcast *mcast, *tmcast;
801 	LIST_HEAD(remove_list);
802 	unsigned long flags;
803 	struct ib_sa_mcmember_rec rec;
804 
805 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
806 
807 	ipoib_mcast_stop_thread(dev, 0);
808 
809 	local_irq_save(flags);
810 	netif_addr_lock(dev);
811 	spin_lock(&priv->lock);
812 
813 	/*
814 	 * Unfortunately, the networking core only gives us a list of all of
815 	 * the multicast hardware addresses. We need to figure out which ones
816 	 * are new and which ones have been removed
817 	 */
818 
819 	/* Clear out the found flag */
820 	list_for_each_entry(mcast, &priv->multicast_list, list)
821 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
822 
823 	/* Mark all of the entries that are found or don't exist */
824 	netdev_for_each_mc_addr(ha, dev) {
825 		union ib_gid mgid;
826 
827 		if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
828 			continue;
829 
830 		memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
831 
832 		mcast = __ipoib_mcast_find(dev, &mgid);
833 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
834 			struct ipoib_mcast *nmcast;
835 
836 			/* ignore group which is directly joined by userspace */
837 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
838 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
839 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
840 						mgid.raw);
841 				continue;
842 			}
843 
844 			/* Not found or send-only group, let's add a new entry */
845 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
846 					mgid.raw);
847 
848 			nmcast = ipoib_mcast_alloc(dev, 0);
849 			if (!nmcast) {
850 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
851 				continue;
852 			}
853 
854 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
855 
856 			nmcast->mcmember.mgid = mgid;
857 
858 			if (mcast) {
859 				/* Destroy the send only entry */
860 				list_move_tail(&mcast->list, &remove_list);
861 
862 				rb_replace_node(&mcast->rb_node,
863 						&nmcast->rb_node,
864 						&priv->multicast_tree);
865 			} else
866 				__ipoib_mcast_add(dev, nmcast);
867 
868 			list_add_tail(&nmcast->list, &priv->multicast_list);
869 		}
870 
871 		if (mcast)
872 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
873 	}
874 
875 	/* Remove all of the entries don't exist anymore */
876 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
877 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
878 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
879 			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
880 					mcast->mcmember.mgid.raw);
881 
882 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
883 
884 			/* Move to the remove list */
885 			list_move_tail(&mcast->list, &remove_list);
886 		}
887 	}
888 
889 	spin_unlock(&priv->lock);
890 	netif_addr_unlock(dev);
891 	local_irq_restore(flags);
892 
893 	/* We have to cancel outside of the spinlock */
894 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
895 		ipoib_mcast_leave(mcast->dev, mcast);
896 		ipoib_mcast_free(mcast);
897 	}
898 
899 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
900 		ipoib_mcast_start_thread(dev);
901 }
902 
903 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
904 
905 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
906 {
907 	struct ipoib_mcast_iter *iter;
908 
909 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
910 	if (!iter)
911 		return NULL;
912 
913 	iter->dev = dev;
914 	memset(iter->mgid.raw, 0, 16);
915 
916 	if (ipoib_mcast_iter_next(iter)) {
917 		kfree(iter);
918 		return NULL;
919 	}
920 
921 	return iter;
922 }
923 
924 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
925 {
926 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
927 	struct rb_node *n;
928 	struct ipoib_mcast *mcast;
929 	int ret = 1;
930 
931 	spin_lock_irq(&priv->lock);
932 
933 	n = rb_first(&priv->multicast_tree);
934 
935 	while (n) {
936 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
937 
938 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
939 			   sizeof (union ib_gid)) < 0) {
940 			iter->mgid      = mcast->mcmember.mgid;
941 			iter->created   = mcast->created;
942 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
943 			iter->complete  = !!mcast->ah;
944 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
945 
946 			ret = 0;
947 
948 			break;
949 		}
950 
951 		n = rb_next(n);
952 	}
953 
954 	spin_unlock_irq(&priv->lock);
955 
956 	return ret;
957 }
958 
959 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
960 			   union ib_gid *mgid,
961 			   unsigned long *created,
962 			   unsigned int *queuelen,
963 			   unsigned int *complete,
964 			   unsigned int *send_only)
965 {
966 	*mgid      = iter->mgid;
967 	*created   = iter->created;
968 	*queuelen  = iter->queuelen;
969 	*complete  = iter->complete;
970 	*send_only = iter->send_only;
971 }
972 
973 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
974