xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision 8fa5723aa7e053d498336b48448b292fc2e0458b)
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),
75 			"deleting multicast group " IPOIB_GID_FMT "\n",
76 			IPOIB_GID_ARG(mcast->mcmember.mgid));
77 
78 	spin_lock_irq(&priv->lock);
79 
80 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
81 		/*
82 		 * It's safe to call ipoib_put_ah() inside priv->lock
83 		 * here, because we know that mcast->ah will always
84 		 * hold one more reference, so ipoib_put_ah() will
85 		 * never do more than decrement the ref count.
86 		 */
87 		if (neigh->ah)
88 			ipoib_put_ah(neigh->ah);
89 		ipoib_neigh_free(dev, neigh);
90 	}
91 
92 	spin_unlock_irq(&priv->lock);
93 
94 	if (mcast->ah)
95 		ipoib_put_ah(mcast->ah);
96 
97 	while (!skb_queue_empty(&mcast->pkt_queue)) {
98 		++tx_dropped;
99 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
100 	}
101 
102 	netif_tx_lock_bh(dev);
103 	dev->stats.tx_dropped += tx_dropped;
104 	netif_tx_unlock_bh(dev);
105 
106 	kfree(mcast);
107 }
108 
109 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
110 					     int can_sleep)
111 {
112 	struct ipoib_mcast *mcast;
113 
114 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
115 	if (!mcast)
116 		return NULL;
117 
118 	mcast->dev = dev;
119 	mcast->created = jiffies;
120 	mcast->backoff = 1;
121 
122 	INIT_LIST_HEAD(&mcast->list);
123 	INIT_LIST_HEAD(&mcast->neigh_list);
124 	skb_queue_head_init(&mcast->pkt_queue);
125 
126 	return mcast;
127 }
128 
129 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
130 {
131 	struct ipoib_dev_priv *priv = netdev_priv(dev);
132 	struct rb_node *n = priv->multicast_tree.rb_node;
133 
134 	while (n) {
135 		struct ipoib_mcast *mcast;
136 		int ret;
137 
138 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
139 
140 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
141 			     sizeof (union ib_gid));
142 		if (ret < 0)
143 			n = n->rb_left;
144 		else if (ret > 0)
145 			n = n->rb_right;
146 		else
147 			return mcast;
148 	}
149 
150 	return NULL;
151 }
152 
153 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
154 {
155 	struct ipoib_dev_priv *priv = netdev_priv(dev);
156 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
157 
158 	while (*n) {
159 		struct ipoib_mcast *tmcast;
160 		int ret;
161 
162 		pn = *n;
163 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
164 
165 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
166 			     sizeof (union ib_gid));
167 		if (ret < 0)
168 			n = &pn->rb_left;
169 		else if (ret > 0)
170 			n = &pn->rb_right;
171 		else
172 			return -EEXIST;
173 	}
174 
175 	rb_link_node(&mcast->rb_node, pn, n);
176 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
177 
178 	return 0;
179 }
180 
181 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
182 				   struct ib_sa_mcmember_rec *mcmember)
183 {
184 	struct net_device *dev = mcast->dev;
185 	struct ipoib_dev_priv *priv = netdev_priv(dev);
186 	struct ipoib_ah *ah;
187 	int ret;
188 	int set_qkey = 0;
189 
190 	mcast->mcmember = *mcmember;
191 
192 	/* Set the cached Q_Key before we attach if it's the broadcast group */
193 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
194 		    sizeof (union ib_gid))) {
195 		spin_lock_irq(&priv->lock);
196 		if (!priv->broadcast) {
197 			spin_unlock_irq(&priv->lock);
198 			return -EAGAIN;
199 		}
200 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
201 		spin_unlock_irq(&priv->lock);
202 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
203 		set_qkey = 1;
204 	}
205 
206 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
207 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
208 			ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
209 				   " already attached\n",
210 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
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 "
219 				   IPOIB_GID_FMT "\n",
220 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
221 
222 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
223 			return ret;
224 		}
225 	}
226 
227 	{
228 		struct ib_ah_attr av = {
229 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
230 			.port_num      = priv->port,
231 			.sl	       = mcast->mcmember.sl,
232 			.ah_flags      = IB_AH_GRH,
233 			.static_rate   = mcast->mcmember.rate,
234 			.grh	       = {
235 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
236 				.hop_limit     = mcast->mcmember.hop_limit,
237 				.sgid_index    = 0,
238 				.traffic_class = mcast->mcmember.traffic_class
239 			}
240 		};
241 		av.grh.dgid = mcast->mcmember.mgid;
242 
243 		ah = ipoib_create_ah(dev, priv->pd, &av);
244 		if (!ah) {
245 			ipoib_warn(priv, "ib_address_create failed\n");
246 		} else {
247 			spin_lock_irq(&priv->lock);
248 			mcast->ah = ah;
249 			spin_unlock_irq(&priv->lock);
250 
251 			ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
252 					" AV %p, LID 0x%04x, SL %d\n",
253 					IPOIB_GID_ARG(mcast->mcmember.mgid),
254 					mcast->ah->ah,
255 					be16_to_cpu(mcast->mcmember.mlid),
256 					mcast->mcmember.sl);
257 		}
258 	}
259 
260 	/* actually send any queued packets */
261 	netif_tx_lock_bh(dev);
262 	while (!skb_queue_empty(&mcast->pkt_queue)) {
263 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
264 		netif_tx_unlock_bh(dev);
265 
266 		skb->dev = dev;
267 
268 		if (!skb->dst || !skb->dst->neighbour) {
269 			/* put pseudoheader back on for next time */
270 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
271 		}
272 
273 		if (dev_queue_xmit(skb))
274 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
275 		netif_tx_lock_bh(dev);
276 	}
277 	netif_tx_unlock_bh(dev);
278 
279 	return 0;
280 }
281 
282 static int
283 ipoib_mcast_sendonly_join_complete(int status,
284 				   struct ib_sa_multicast *multicast)
285 {
286 	struct ipoib_mcast *mcast = multicast->context;
287 	struct net_device *dev = mcast->dev;
288 
289 	/* We trap for port events ourselves. */
290 	if (status == -ENETRESET)
291 		return 0;
292 
293 	if (!status)
294 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
295 
296 	if (status) {
297 		if (mcast->logcount++ < 20)
298 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
299 					IPOIB_GID_FMT ", status %d\n",
300 					IPOIB_GID_ARG(mcast->mcmember.mgid), status);
301 
302 		/* Flush out any queued packets */
303 		netif_tx_lock_bh(dev);
304 		while (!skb_queue_empty(&mcast->pkt_queue)) {
305 			++dev->stats.tx_dropped;
306 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
307 		}
308 		netif_tx_unlock_bh(dev);
309 
310 		/* Clear the busy flag so we try again */
311 		status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
312 					    &mcast->flags);
313 	}
314 	return status;
315 }
316 
317 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
318 {
319 	struct net_device *dev = mcast->dev;
320 	struct ipoib_dev_priv *priv = netdev_priv(dev);
321 	struct ib_sa_mcmember_rec rec = {
322 #if 0				/* Some SMs don't support send-only yet */
323 		.join_state = 4
324 #else
325 		.join_state = 1
326 #endif
327 	};
328 	int ret = 0;
329 
330 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
331 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
332 		return -ENODEV;
333 	}
334 
335 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
336 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
337 		return -EBUSY;
338 	}
339 
340 	rec.mgid     = mcast->mcmember.mgid;
341 	rec.port_gid = priv->local_gid;
342 	rec.pkey     = cpu_to_be16(priv->pkey);
343 
344 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
345 					 priv->port, &rec,
346 					 IB_SA_MCMEMBER_REC_MGID	|
347 					 IB_SA_MCMEMBER_REC_PORT_GID	|
348 					 IB_SA_MCMEMBER_REC_PKEY	|
349 					 IB_SA_MCMEMBER_REC_JOIN_STATE,
350 					 GFP_ATOMIC,
351 					 ipoib_mcast_sendonly_join_complete,
352 					 mcast);
353 	if (IS_ERR(mcast->mc)) {
354 		ret = PTR_ERR(mcast->mc);
355 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
356 		ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
357 			   ret);
358 	} else {
359 		ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
360 				", starting join\n",
361 				IPOIB_GID_ARG(mcast->mcmember.mgid));
362 	}
363 
364 	return ret;
365 }
366 
367 void ipoib_mcast_carrier_on_task(struct work_struct *work)
368 {
369 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
370 						   carrier_on_task);
371 
372 	/*
373 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
374 	 * turning the carrier back on while a device is being
375 	 * removed.
376 	 */
377 	rtnl_lock();
378 	netif_carrier_on(priv->dev);
379 	rtnl_unlock();
380 }
381 
382 static int ipoib_mcast_join_complete(int status,
383 				     struct ib_sa_multicast *multicast)
384 {
385 	struct ipoib_mcast *mcast = multicast->context;
386 	struct net_device *dev = mcast->dev;
387 	struct ipoib_dev_priv *priv = netdev_priv(dev);
388 
389 	ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
390 			" (status %d)\n",
391 			IPOIB_GID_ARG(mcast->mcmember.mgid), status);
392 
393 	/* We trap for port events ourselves. */
394 	if (status == -ENETRESET)
395 		return 0;
396 
397 	if (!status)
398 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
399 
400 	if (!status) {
401 		mcast->backoff = 1;
402 		mutex_lock(&mcast_mutex);
403 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
404 			queue_delayed_work(ipoib_workqueue,
405 					   &priv->mcast_task, 0);
406 		mutex_unlock(&mcast_mutex);
407 
408 		/*
409 		 * Defer carrier on work to ipoib_workqueue to avoid a
410 		 * deadlock on rtnl_lock here.
411 		 */
412 		if (mcast == priv->broadcast)
413 			queue_work(ipoib_workqueue, &priv->carrier_on_task);
414 
415 		return 0;
416 	}
417 
418 	if (mcast->logcount++ < 20) {
419 		if (status == -ETIMEDOUT) {
420 			ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
421 					", status %d\n",
422 					IPOIB_GID_ARG(mcast->mcmember.mgid),
423 					status);
424 		} else {
425 			ipoib_warn(priv, "multicast join failed for "
426 				   IPOIB_GID_FMT ", status %d\n",
427 				   IPOIB_GID_ARG(mcast->mcmember.mgid),
428 				   status);
429 		}
430 	}
431 
432 	mcast->backoff *= 2;
433 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
434 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
435 
436 	/* Clear the busy flag so we try again */
437 	status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
438 
439 	mutex_lock(&mcast_mutex);
440 	spin_lock_irq(&priv->lock);
441 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
442 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
443 				   mcast->backoff * HZ);
444 	spin_unlock_irq(&priv->lock);
445 	mutex_unlock(&mcast_mutex);
446 
447 	return status;
448 }
449 
450 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
451 			     int create)
452 {
453 	struct ipoib_dev_priv *priv = netdev_priv(dev);
454 	struct ib_sa_mcmember_rec rec = {
455 		.join_state = 1
456 	};
457 	ib_sa_comp_mask comp_mask;
458 	int ret = 0;
459 
460 	ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
461 			IPOIB_GID_ARG(mcast->mcmember.mgid));
462 
463 	rec.mgid     = mcast->mcmember.mgid;
464 	rec.port_gid = priv->local_gid;
465 	rec.pkey     = cpu_to_be16(priv->pkey);
466 
467 	comp_mask =
468 		IB_SA_MCMEMBER_REC_MGID		|
469 		IB_SA_MCMEMBER_REC_PORT_GID	|
470 		IB_SA_MCMEMBER_REC_PKEY		|
471 		IB_SA_MCMEMBER_REC_JOIN_STATE;
472 
473 	if (create) {
474 		comp_mask |=
475 			IB_SA_MCMEMBER_REC_QKEY			|
476 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
477 			IB_SA_MCMEMBER_REC_MTU			|
478 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
479 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
480 			IB_SA_MCMEMBER_REC_RATE			|
481 			IB_SA_MCMEMBER_REC_SL			|
482 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
483 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
484 
485 		rec.qkey	  = priv->broadcast->mcmember.qkey;
486 		rec.mtu_selector  = IB_SA_EQ;
487 		rec.mtu		  = priv->broadcast->mcmember.mtu;
488 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
489 		rec.rate_selector = IB_SA_EQ;
490 		rec.rate	  = priv->broadcast->mcmember.rate;
491 		rec.sl		  = priv->broadcast->mcmember.sl;
492 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
493 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
494 	}
495 
496 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
497 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
498 					 &rec, comp_mask, GFP_KERNEL,
499 					 ipoib_mcast_join_complete, mcast);
500 	if (IS_ERR(mcast->mc)) {
501 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
502 		ret = PTR_ERR(mcast->mc);
503 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
504 
505 		mcast->backoff *= 2;
506 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
507 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
508 
509 		mutex_lock(&mcast_mutex);
510 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
511 			queue_delayed_work(ipoib_workqueue,
512 					   &priv->mcast_task,
513 					   mcast->backoff * HZ);
514 		mutex_unlock(&mcast_mutex);
515 	}
516 }
517 
518 void ipoib_mcast_join_task(struct work_struct *work)
519 {
520 	struct ipoib_dev_priv *priv =
521 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
522 	struct net_device *dev = priv->dev;
523 
524 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
525 		return;
526 
527 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
528 		ipoib_warn(priv, "ib_query_gid() failed\n");
529 	else
530 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
531 
532 	{
533 		struct ib_port_attr attr;
534 
535 		if (!ib_query_port(priv->ca, priv->port, &attr))
536 			priv->local_lid = attr.lid;
537 		else
538 			ipoib_warn(priv, "ib_query_port failed\n");
539 	}
540 
541 	if (!priv->broadcast) {
542 		struct ipoib_mcast *broadcast;
543 
544 		broadcast = ipoib_mcast_alloc(dev, 1);
545 		if (!broadcast) {
546 			ipoib_warn(priv, "failed to allocate broadcast group\n");
547 			mutex_lock(&mcast_mutex);
548 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
549 				queue_delayed_work(ipoib_workqueue,
550 						   &priv->mcast_task, HZ);
551 			mutex_unlock(&mcast_mutex);
552 			return;
553 		}
554 
555 		spin_lock_irq(&priv->lock);
556 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
557 		       sizeof (union ib_gid));
558 		priv->broadcast = broadcast;
559 
560 		__ipoib_mcast_add(dev, priv->broadcast);
561 		spin_unlock_irq(&priv->lock);
562 	}
563 
564 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
565 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
566 			ipoib_mcast_join(dev, priv->broadcast, 0);
567 		return;
568 	}
569 
570 	while (1) {
571 		struct ipoib_mcast *mcast = NULL;
572 
573 		spin_lock_irq(&priv->lock);
574 		list_for_each_entry(mcast, &priv->multicast_list, list) {
575 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
576 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
577 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
578 				/* Found the next unjoined group */
579 				break;
580 			}
581 		}
582 		spin_unlock_irq(&priv->lock);
583 
584 		if (&mcast->list == &priv->multicast_list) {
585 			/* All done */
586 			break;
587 		}
588 
589 		ipoib_mcast_join(dev, mcast, 1);
590 		return;
591 	}
592 
593 	priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
594 
595 	if (!ipoib_cm_admin_enabled(dev)) {
596 		rtnl_lock();
597 		dev_set_mtu(dev, min(priv->mcast_mtu, priv->admin_mtu));
598 		rtnl_unlock();
599 	}
600 
601 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
602 
603 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
604 }
605 
606 int ipoib_mcast_start_thread(struct net_device *dev)
607 {
608 	struct ipoib_dev_priv *priv = netdev_priv(dev);
609 
610 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
611 
612 	mutex_lock(&mcast_mutex);
613 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
614 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
615 	mutex_unlock(&mcast_mutex);
616 
617 	return 0;
618 }
619 
620 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
621 {
622 	struct ipoib_dev_priv *priv = netdev_priv(dev);
623 
624 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
625 
626 	mutex_lock(&mcast_mutex);
627 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
628 	cancel_delayed_work(&priv->mcast_task);
629 	mutex_unlock(&mcast_mutex);
630 
631 	if (flush)
632 		flush_workqueue(ipoib_workqueue);
633 
634 	return 0;
635 }
636 
637 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
638 {
639 	struct ipoib_dev_priv *priv = netdev_priv(dev);
640 	int ret = 0;
641 
642 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
643 		ib_sa_free_multicast(mcast->mc);
644 
645 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
646 		ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
647 				IPOIB_GID_ARG(mcast->mcmember.mgid));
648 
649 		/* Remove ourselves from the multicast group */
650 		ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
651 				      be16_to_cpu(mcast->mcmember.mlid));
652 		if (ret)
653 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
654 	}
655 
656 	return 0;
657 }
658 
659 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
660 {
661 	struct ipoib_dev_priv *priv = netdev_priv(dev);
662 	struct ipoib_mcast *mcast;
663 	unsigned long flags;
664 
665 	spin_lock_irqsave(&priv->lock, flags);
666 
667 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
668 	    !priv->broadcast					||
669 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
670 		++dev->stats.tx_dropped;
671 		dev_kfree_skb_any(skb);
672 		goto unlock;
673 	}
674 
675 	mcast = __ipoib_mcast_find(dev, mgid);
676 	if (!mcast) {
677 		/* Let's create a new send only group now */
678 		ipoib_dbg_mcast(priv, "setting up send only multicast group for "
679 				IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
680 
681 		mcast = ipoib_mcast_alloc(dev, 0);
682 		if (!mcast) {
683 			ipoib_warn(priv, "unable to allocate memory for "
684 				   "multicast structure\n");
685 			++dev->stats.tx_dropped;
686 			dev_kfree_skb_any(skb);
687 			goto out;
688 		}
689 
690 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
691 		memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
692 		__ipoib_mcast_add(dev, mcast);
693 		list_add_tail(&mcast->list, &priv->multicast_list);
694 	}
695 
696 	if (!mcast->ah) {
697 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
698 			skb_queue_tail(&mcast->pkt_queue, skb);
699 		else {
700 			++dev->stats.tx_dropped;
701 			dev_kfree_skb_any(skb);
702 		}
703 
704 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
705 			ipoib_dbg_mcast(priv, "no address vector, "
706 					"but multicast join already started\n");
707 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
708 			ipoib_mcast_sendonly_join(mcast);
709 
710 		/*
711 		 * If lookup completes between here and out:, don't
712 		 * want to send packet twice.
713 		 */
714 		mcast = NULL;
715 	}
716 
717 out:
718 	if (mcast && mcast->ah) {
719 		if (skb->dst		&&
720 		    skb->dst->neighbour &&
721 		    !*to_ipoib_neigh(skb->dst->neighbour)) {
722 			struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour,
723 									skb->dev);
724 
725 			if (neigh) {
726 				kref_get(&mcast->ah->ref);
727 				neigh->ah	= mcast->ah;
728 				list_add_tail(&neigh->list, &mcast->neigh_list);
729 			}
730 		}
731 
732 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
733 	}
734 
735 unlock:
736 	spin_unlock_irqrestore(&priv->lock, flags);
737 }
738 
739 void ipoib_mcast_dev_flush(struct net_device *dev)
740 {
741 	struct ipoib_dev_priv *priv = netdev_priv(dev);
742 	LIST_HEAD(remove_list);
743 	struct ipoib_mcast *mcast, *tmcast;
744 	unsigned long flags;
745 
746 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
747 
748 	spin_lock_irqsave(&priv->lock, flags);
749 
750 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
751 		list_del(&mcast->list);
752 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
753 		list_add_tail(&mcast->list, &remove_list);
754 	}
755 
756 	if (priv->broadcast) {
757 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
758 		list_add_tail(&priv->broadcast->list, &remove_list);
759 		priv->broadcast = NULL;
760 	}
761 
762 	spin_unlock_irqrestore(&priv->lock, flags);
763 
764 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
765 		ipoib_mcast_leave(dev, mcast);
766 		ipoib_mcast_free(mcast);
767 	}
768 }
769 
770 void ipoib_mcast_restart_task(struct work_struct *work)
771 {
772 	struct ipoib_dev_priv *priv =
773 		container_of(work, struct ipoib_dev_priv, restart_task);
774 	struct net_device *dev = priv->dev;
775 	struct dev_mc_list *mclist;
776 	struct ipoib_mcast *mcast, *tmcast;
777 	LIST_HEAD(remove_list);
778 	unsigned long flags;
779 	struct ib_sa_mcmember_rec rec;
780 
781 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
782 
783 	ipoib_mcast_stop_thread(dev, 0);
784 
785 	local_irq_save(flags);
786 	netif_addr_lock(dev);
787 	spin_lock(&priv->lock);
788 
789 	/*
790 	 * Unfortunately, the networking core only gives us a list of all of
791 	 * the multicast hardware addresses. We need to figure out which ones
792 	 * are new and which ones have been removed
793 	 */
794 
795 	/* Clear out the found flag */
796 	list_for_each_entry(mcast, &priv->multicast_list, list)
797 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
798 
799 	/* Mark all of the entries that are found or don't exist */
800 	for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
801 		union ib_gid mgid;
802 
803 		memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
804 
805 		mcast = __ipoib_mcast_find(dev, &mgid);
806 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
807 			struct ipoib_mcast *nmcast;
808 
809 			/* ignore group which is directly joined by userspace */
810 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
811 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
812 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid "
813 						IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
814 				continue;
815 			}
816 
817 			/* Not found or send-only group, let's add a new entry */
818 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
819 					IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
820 
821 			nmcast = ipoib_mcast_alloc(dev, 0);
822 			if (!nmcast) {
823 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
824 				continue;
825 			}
826 
827 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
828 
829 			nmcast->mcmember.mgid = mgid;
830 
831 			if (mcast) {
832 				/* Destroy the send only entry */
833 				list_move_tail(&mcast->list, &remove_list);
834 
835 				rb_replace_node(&mcast->rb_node,
836 						&nmcast->rb_node,
837 						&priv->multicast_tree);
838 			} else
839 				__ipoib_mcast_add(dev, nmcast);
840 
841 			list_add_tail(&nmcast->list, &priv->multicast_list);
842 		}
843 
844 		if (mcast)
845 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
846 	}
847 
848 	/* Remove all of the entries don't exist anymore */
849 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
850 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
851 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
852 			ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
853 					IPOIB_GID_ARG(mcast->mcmember.mgid));
854 
855 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
856 
857 			/* Move to the remove list */
858 			list_move_tail(&mcast->list, &remove_list);
859 		}
860 	}
861 
862 	spin_unlock(&priv->lock);
863 	netif_addr_unlock(dev);
864 	local_irq_restore(flags);
865 
866 	/* We have to cancel outside of the spinlock */
867 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
868 		ipoib_mcast_leave(mcast->dev, mcast);
869 		ipoib_mcast_free(mcast);
870 	}
871 
872 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
873 		ipoib_mcast_start_thread(dev);
874 }
875 
876 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
877 
878 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
879 {
880 	struct ipoib_mcast_iter *iter;
881 
882 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
883 	if (!iter)
884 		return NULL;
885 
886 	iter->dev = dev;
887 	memset(iter->mgid.raw, 0, 16);
888 
889 	if (ipoib_mcast_iter_next(iter)) {
890 		kfree(iter);
891 		return NULL;
892 	}
893 
894 	return iter;
895 }
896 
897 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
898 {
899 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
900 	struct rb_node *n;
901 	struct ipoib_mcast *mcast;
902 	int ret = 1;
903 
904 	spin_lock_irq(&priv->lock);
905 
906 	n = rb_first(&priv->multicast_tree);
907 
908 	while (n) {
909 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
910 
911 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
912 			   sizeof (union ib_gid)) < 0) {
913 			iter->mgid      = mcast->mcmember.mgid;
914 			iter->created   = mcast->created;
915 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
916 			iter->complete  = !!mcast->ah;
917 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
918 
919 			ret = 0;
920 
921 			break;
922 		}
923 
924 		n = rb_next(n);
925 	}
926 
927 	spin_unlock_irq(&priv->lock);
928 
929 	return ret;
930 }
931 
932 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
933 			   union ib_gid *mgid,
934 			   unsigned long *created,
935 			   unsigned int *queuelen,
936 			   unsigned int *complete,
937 			   unsigned int *send_only)
938 {
939 	*mgid      = iter->mgid;
940 	*created   = iter->created;
941 	*queuelen  = iter->queuelen;
942 	*complete  = iter->complete;
943 	*send_only = iter->send_only;
944 }
945 
946 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
947