xref: /linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision de2fe5e07d58424bc286fff3fd3c1b0bf933cd58)
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 ipoib_ah          *ah;
64 
65 	struct rb_node    rb_node;
66 	struct list_head  list;
67 	struct completion done;
68 
69 	int                 query_id;
70 	struct ib_sa_query *query;
71 
72 	unsigned long created;
73 	unsigned long backoff;
74 
75 	unsigned long flags;
76 	unsigned char logcount;
77 
78 	struct list_head  neigh_list;
79 
80 	struct sk_buff_head pkt_queue;
81 
82 	struct net_device *dev;
83 };
84 
85 struct ipoib_mcast_iter {
86 	struct net_device *dev;
87 	union ib_gid       mgid;
88 	unsigned long      created;
89 	unsigned int       queuelen;
90 	unsigned int       complete;
91 	unsigned int       send_only;
92 };
93 
94 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
95 {
96 	struct net_device *dev = mcast->dev;
97 	struct ipoib_dev_priv *priv = netdev_priv(dev);
98 	struct ipoib_neigh *neigh, *tmp;
99 	unsigned long flags;
100 	int tx_dropped = 0;
101 
102 	ipoib_dbg_mcast(netdev_priv(dev),
103 			"deleting multicast group " IPOIB_GID_FMT "\n",
104 			IPOIB_GID_ARG(mcast->mcmember.mgid));
105 
106 	spin_lock_irqsave(&priv->lock, flags);
107 
108 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
109 		/*
110 		 * It's safe to call ipoib_put_ah() inside priv->lock
111 		 * here, because we know that mcast->ah will always
112 		 * hold one more reference, so ipoib_put_ah() will
113 		 * never do more than decrement the ref count.
114 		 */
115 		if (neigh->ah)
116 			ipoib_put_ah(neigh->ah);
117 		*to_ipoib_neigh(neigh->neighbour) = NULL;
118 		kfree(neigh);
119 	}
120 
121 	spin_unlock_irqrestore(&priv->lock, flags);
122 
123 	if (mcast->ah)
124 		ipoib_put_ah(mcast->ah);
125 
126 	while (!skb_queue_empty(&mcast->pkt_queue)) {
127 		++tx_dropped;
128 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
129 	}
130 
131 	spin_lock_irqsave(&priv->tx_lock, flags);
132 	priv->stats.tx_dropped += tx_dropped;
133 	spin_unlock_irqrestore(&priv->tx_lock, flags);
134 
135 	kfree(mcast);
136 }
137 
138 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
139 					     int can_sleep)
140 {
141 	struct ipoib_mcast *mcast;
142 
143 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
144 	if (!mcast)
145 		return NULL;
146 
147 	mcast->dev = dev;
148 	mcast->created = jiffies;
149 	mcast->backoff = 1;
150 
151 	INIT_LIST_HEAD(&mcast->list);
152 	INIT_LIST_HEAD(&mcast->neigh_list);
153 	skb_queue_head_init(&mcast->pkt_queue);
154 
155 	return mcast;
156 }
157 
158 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
159 {
160 	struct ipoib_dev_priv *priv = netdev_priv(dev);
161 	struct rb_node *n = priv->multicast_tree.rb_node;
162 
163 	while (n) {
164 		struct ipoib_mcast *mcast;
165 		int ret;
166 
167 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
168 
169 		ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
170 			     sizeof (union ib_gid));
171 		if (ret < 0)
172 			n = n->rb_left;
173 		else if (ret > 0)
174 			n = n->rb_right;
175 		else
176 			return mcast;
177 	}
178 
179 	return NULL;
180 }
181 
182 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
183 {
184 	struct ipoib_dev_priv *priv = netdev_priv(dev);
185 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
186 
187 	while (*n) {
188 		struct ipoib_mcast *tmcast;
189 		int ret;
190 
191 		pn = *n;
192 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
193 
194 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
195 			     sizeof (union ib_gid));
196 		if (ret < 0)
197 			n = &pn->rb_left;
198 		else if (ret > 0)
199 			n = &pn->rb_right;
200 		else
201 			return -EEXIST;
202 	}
203 
204 	rb_link_node(&mcast->rb_node, pn, n);
205 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
206 
207 	return 0;
208 }
209 
210 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
211 				   struct ib_sa_mcmember_rec *mcmember)
212 {
213 	struct net_device *dev = mcast->dev;
214 	struct ipoib_dev_priv *priv = netdev_priv(dev);
215 	struct ipoib_ah *ah;
216 	int ret;
217 
218 	mcast->mcmember = *mcmember;
219 
220 	/* Set the cached Q_Key before we attach if it's the broadcast group */
221 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
222 		    sizeof (union ib_gid))) {
223 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
224 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
225 	}
226 
227 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
228 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
229 			ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
230 				   " already attached\n",
231 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
232 
233 			return 0;
234 		}
235 
236 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
237 					 &mcast->mcmember.mgid);
238 		if (ret < 0) {
239 			ipoib_warn(priv, "couldn't attach QP to multicast group "
240 				   IPOIB_GID_FMT "\n",
241 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
242 
243 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
244 			return ret;
245 		}
246 	}
247 
248 	{
249 		struct ib_ah_attr av = {
250 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
251 			.port_num      = priv->port,
252 			.sl	       = mcast->mcmember.sl,
253 			.ah_flags      = IB_AH_GRH,
254 			.grh	       = {
255 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
256 				.hop_limit     = mcast->mcmember.hop_limit,
257 				.sgid_index    = 0,
258 				.traffic_class = mcast->mcmember.traffic_class
259 			}
260 		};
261 		int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
262 
263 		av.grh.dgid = mcast->mcmember.mgid;
264 
265 		if (path_rate > 0 && priv->local_rate > path_rate)
266 			av.static_rate = (priv->local_rate - 1) / path_rate;
267 
268 		ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
269 				av.static_rate, priv->local_rate,
270 				ib_sa_rate_enum_to_int(mcast->mcmember.rate));
271 
272 		ah = ipoib_create_ah(dev, priv->pd, &av);
273 		if (!ah) {
274 			ipoib_warn(priv, "ib_address_create failed\n");
275 		} else {
276 			ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
277 					" AV %p, LID 0x%04x, SL %d\n",
278 					IPOIB_GID_ARG(mcast->mcmember.mgid),
279 					mcast->ah->ah,
280 					be16_to_cpu(mcast->mcmember.mlid),
281 					mcast->mcmember.sl);
282 		}
283 
284 		spin_lock_irq(&priv->lock);
285 		mcast->ah = ah;
286 		spin_unlock_irq(&priv->lock);
287 	}
288 
289 	/* actually send any queued packets */
290 	spin_lock_irq(&priv->tx_lock);
291 	while (!skb_queue_empty(&mcast->pkt_queue)) {
292 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
293 		spin_unlock_irq(&priv->tx_lock);
294 
295 		skb->dev = dev;
296 
297 		if (!skb->dst || !skb->dst->neighbour) {
298 			/* put pseudoheader back on for next time */
299 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
300 		}
301 
302 		if (dev_queue_xmit(skb))
303 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
304 		spin_lock_irq(&priv->tx_lock);
305 	}
306 	spin_unlock_irq(&priv->tx_lock);
307 
308 	return 0;
309 }
310 
311 static void
312 ipoib_mcast_sendonly_join_complete(int status,
313 				   struct ib_sa_mcmember_rec *mcmember,
314 				   void *mcast_ptr)
315 {
316 	struct ipoib_mcast *mcast = mcast_ptr;
317 	struct net_device *dev = mcast->dev;
318 	struct ipoib_dev_priv *priv = netdev_priv(dev);
319 
320 	if (!status)
321 		ipoib_mcast_join_finish(mcast, mcmember);
322 	else {
323 		if (mcast->logcount++ < 20)
324 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
325 					IPOIB_GID_FMT ", status %d\n",
326 					IPOIB_GID_ARG(mcast->mcmember.mgid), status);
327 
328 		/* Flush out any queued packets */
329 		spin_lock_irq(&priv->tx_lock);
330 		while (!skb_queue_empty(&mcast->pkt_queue)) {
331 			++priv->stats.tx_dropped;
332 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
333 		}
334 		spin_unlock_irq(&priv->tx_lock);
335 
336 		/* Clear the busy flag so we try again */
337 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
338 	}
339 
340 	complete(&mcast->done);
341 }
342 
343 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
344 {
345 	struct net_device *dev = mcast->dev;
346 	struct ipoib_dev_priv *priv = netdev_priv(dev);
347 	struct ib_sa_mcmember_rec rec = {
348 #if 0				/* Some SMs don't support send-only yet */
349 		.join_state = 4
350 #else
351 		.join_state = 1
352 #endif
353 	};
354 	int ret = 0;
355 
356 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
357 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
358 		return -ENODEV;
359 	}
360 
361 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
362 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
363 		return -EBUSY;
364 	}
365 
366 	rec.mgid     = mcast->mcmember.mgid;
367 	rec.port_gid = priv->local_gid;
368 	rec.pkey     = cpu_to_be16(priv->pkey);
369 
370 	init_completion(&mcast->done);
371 
372 	ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
373 				     IB_SA_MCMEMBER_REC_MGID		|
374 				     IB_SA_MCMEMBER_REC_PORT_GID	|
375 				     IB_SA_MCMEMBER_REC_PKEY		|
376 				     IB_SA_MCMEMBER_REC_JOIN_STATE,
377 				     1000, GFP_ATOMIC,
378 				     ipoib_mcast_sendonly_join_complete,
379 				     mcast, &mcast->query);
380 	if (ret < 0) {
381 		ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
382 			   ret);
383 	} else {
384 		ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
385 				", starting join\n",
386 				IPOIB_GID_ARG(mcast->mcmember.mgid));
387 
388 		mcast->query_id = ret;
389 	}
390 
391 	return ret;
392 }
393 
394 static void ipoib_mcast_join_complete(int status,
395 				      struct ib_sa_mcmember_rec *mcmember,
396 				      void *mcast_ptr)
397 {
398 	struct ipoib_mcast *mcast = mcast_ptr;
399 	struct net_device *dev = mcast->dev;
400 	struct ipoib_dev_priv *priv = netdev_priv(dev);
401 
402 	ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
403 			" (status %d)\n",
404 			IPOIB_GID_ARG(mcast->mcmember.mgid), status);
405 
406 	if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
407 		mcast->backoff = 1;
408 		mutex_lock(&mcast_mutex);
409 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
410 			queue_work(ipoib_workqueue, &priv->mcast_task);
411 		mutex_unlock(&mcast_mutex);
412 		complete(&mcast->done);
413 		return;
414 	}
415 
416 	if (status == -EINTR) {
417 		complete(&mcast->done);
418 		return;
419 	}
420 
421 	if (status && mcast->logcount++ < 20) {
422 		if (status == -ETIMEDOUT || status == -EINTR) {
423 			ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
424 					", status %d\n",
425 					IPOIB_GID_ARG(mcast->mcmember.mgid),
426 					status);
427 		} else {
428 			ipoib_warn(priv, "multicast join failed for "
429 				   IPOIB_GID_FMT ", status %d\n",
430 				   IPOIB_GID_ARG(mcast->mcmember.mgid),
431 				   status);
432 		}
433 	}
434 
435 	mcast->backoff *= 2;
436 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
437 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
438 
439 	mutex_lock(&mcast_mutex);
440 
441 	spin_lock_irq(&priv->lock);
442 	mcast->query = NULL;
443 
444 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
445 		if (status == -ETIMEDOUT)
446 			queue_work(ipoib_workqueue, &priv->mcast_task);
447 		else
448 			queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
449 					   mcast->backoff * HZ);
450 	} else
451 		complete(&mcast->done);
452 	spin_unlock_irq(&priv->lock);
453 	mutex_unlock(&mcast_mutex);
454 
455 	return;
456 }
457 
458 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
459 			     int create)
460 {
461 	struct ipoib_dev_priv *priv = netdev_priv(dev);
462 	struct ib_sa_mcmember_rec rec = {
463 		.join_state = 1
464 	};
465 	ib_sa_comp_mask comp_mask;
466 	int ret = 0;
467 
468 	ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
469 			IPOIB_GID_ARG(mcast->mcmember.mgid));
470 
471 	rec.mgid     = mcast->mcmember.mgid;
472 	rec.port_gid = priv->local_gid;
473 	rec.pkey     = cpu_to_be16(priv->pkey);
474 
475 	comp_mask =
476 		IB_SA_MCMEMBER_REC_MGID		|
477 		IB_SA_MCMEMBER_REC_PORT_GID	|
478 		IB_SA_MCMEMBER_REC_PKEY		|
479 		IB_SA_MCMEMBER_REC_JOIN_STATE;
480 
481 	if (create) {
482 		comp_mask |=
483 			IB_SA_MCMEMBER_REC_QKEY		|
484 			IB_SA_MCMEMBER_REC_SL		|
485 			IB_SA_MCMEMBER_REC_FLOW_LABEL	|
486 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
487 
488 		rec.qkey	  = priv->broadcast->mcmember.qkey;
489 		rec.sl		  = priv->broadcast->mcmember.sl;
490 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
491 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
492 	}
493 
494 	init_completion(&mcast->done);
495 
496 	ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
497 				     mcast->backoff * 1000, GFP_ATOMIC,
498 				     ipoib_mcast_join_complete,
499 				     mcast, &mcast->query);
500 
501 	if (ret < 0) {
502 		ipoib_warn(priv, "ib_sa_mcmember_rec_set 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 	} else
515 		mcast->query_id = ret;
516 }
517 
518 void ipoib_mcast_join_task(void *dev_ptr)
519 {
520 	struct net_device *dev = dev_ptr;
521 	struct ipoib_dev_priv *priv = netdev_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 			priv->local_rate = attr.active_speed *
537 				ib_width_enum_to_int(attr.active_width);
538 		} else
539 			ipoib_warn(priv, "ib_query_port failed\n");
540 	}
541 
542 	if (!priv->broadcast) {
543 		struct ipoib_mcast *broadcast;
544 
545 		broadcast = ipoib_mcast_alloc(dev, 1);
546 		if (!broadcast) {
547 			ipoib_warn(priv, "failed to allocate broadcast group\n");
548 			mutex_lock(&mcast_mutex);
549 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
550 				queue_delayed_work(ipoib_workqueue,
551 						   &priv->mcast_task, HZ);
552 			mutex_unlock(&mcast_mutex);
553 			return;
554 		}
555 
556 		spin_lock_irq(&priv->lock);
557 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
558 		       sizeof (union ib_gid));
559 		priv->broadcast = broadcast;
560 
561 		__ipoib_mcast_add(dev, priv->broadcast);
562 		spin_unlock_irq(&priv->lock);
563 	}
564 
565 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &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 = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
594 		IPOIB_ENCAP_LEN;
595 	dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
596 
597 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
598 
599 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
600 	netif_carrier_on(dev);
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_work(ipoib_workqueue, &priv->mcast_task);
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 	struct ipoib_mcast *mcast;
625 
626 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
627 
628 	spin_lock_irq(&priv->lock);
629 	clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
630 	spin_unlock_irq(&priv->lock);
631 
632 	mutex_lock(&mcast_mutex);
633 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
634 	cancel_delayed_work(&priv->mcast_task);
635 	mutex_unlock(&mcast_mutex);
636 
637 	if (flush)
638 		flush_workqueue(ipoib_workqueue);
639 
640 	spin_lock_irq(&priv->lock);
641 	if (priv->broadcast && priv->broadcast->query) {
642 		ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
643 		priv->broadcast->query = NULL;
644 		spin_unlock_irq(&priv->lock);
645 		ipoib_dbg_mcast(priv, "waiting for bcast\n");
646 		wait_for_completion(&priv->broadcast->done);
647 	} else
648 		spin_unlock_irq(&priv->lock);
649 
650 	list_for_each_entry(mcast, &priv->multicast_list, list) {
651 		spin_lock_irq(&priv->lock);
652 		if (mcast->query) {
653 			ib_sa_cancel_query(mcast->query_id, mcast->query);
654 			mcast->query = NULL;
655 			spin_unlock_irq(&priv->lock);
656 			ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
657 					IPOIB_GID_ARG(mcast->mcmember.mgid));
658 			wait_for_completion(&mcast->done);
659 		} else
660 			spin_unlock_irq(&priv->lock);
661 	}
662 
663 	return 0;
664 }
665 
666 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
667 {
668 	struct ipoib_dev_priv *priv = netdev_priv(dev);
669 	struct ib_sa_mcmember_rec rec = {
670 		.join_state = 1
671 	};
672 	int ret = 0;
673 
674 	if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
675 		return 0;
676 
677 	ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
678 			IPOIB_GID_ARG(mcast->mcmember.mgid));
679 
680 	rec.mgid     = mcast->mcmember.mgid;
681 	rec.port_gid = priv->local_gid;
682 	rec.pkey     = cpu_to_be16(priv->pkey);
683 
684 	/* Remove ourselves from the multicast group */
685 	ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
686 				 &mcast->mcmember.mgid);
687 	if (ret)
688 		ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
689 
690 	/*
691 	 * Just make one shot at leaving and don't wait for a reply;
692 	 * if we fail, too bad.
693 	 */
694 	ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
695 					IB_SA_MCMEMBER_REC_MGID		|
696 					IB_SA_MCMEMBER_REC_PORT_GID	|
697 					IB_SA_MCMEMBER_REC_PKEY		|
698 					IB_SA_MCMEMBER_REC_JOIN_STATE,
699 					0, GFP_ATOMIC, NULL,
700 					mcast, &mcast->query);
701 	if (ret < 0)
702 		ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
703 			   "for leave (result = %d)\n", ret);
704 
705 	return 0;
706 }
707 
708 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
709 		      struct sk_buff *skb)
710 {
711 	struct ipoib_dev_priv *priv = netdev_priv(dev);
712 	struct ipoib_mcast *mcast;
713 
714 	/*
715 	 * We can only be called from ipoib_start_xmit, so we're
716 	 * inside tx_lock -- no need to save/restore flags.
717 	 */
718 	spin_lock(&priv->lock);
719 
720 	if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)	||
721 	    !priv->broadcast					||
722 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
723 		++priv->stats.tx_dropped;
724 		dev_kfree_skb_any(skb);
725 		goto unlock;
726 	}
727 
728 	mcast = __ipoib_mcast_find(dev, mgid);
729 	if (!mcast) {
730 		/* Let's create a new send only group now */
731 		ipoib_dbg_mcast(priv, "setting up send only multicast group for "
732 				IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
733 
734 		mcast = ipoib_mcast_alloc(dev, 0);
735 		if (!mcast) {
736 			ipoib_warn(priv, "unable to allocate memory for "
737 				   "multicast structure\n");
738 			++priv->stats.tx_dropped;
739 			dev_kfree_skb_any(skb);
740 			goto out;
741 		}
742 
743 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
744 		mcast->mcmember.mgid = *mgid;
745 		__ipoib_mcast_add(dev, mcast);
746 		list_add_tail(&mcast->list, &priv->multicast_list);
747 	}
748 
749 	if (!mcast->ah) {
750 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
751 			skb_queue_tail(&mcast->pkt_queue, skb);
752 		else {
753 			++priv->stats.tx_dropped;
754 			dev_kfree_skb_any(skb);
755 		}
756 
757 		if (mcast->query)
758 			ipoib_dbg_mcast(priv, "no address vector, "
759 					"but multicast join already started\n");
760 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
761 			ipoib_mcast_sendonly_join(mcast);
762 
763 		/*
764 		 * If lookup completes between here and out:, don't
765 		 * want to send packet twice.
766 		 */
767 		mcast = NULL;
768 	}
769 
770 out:
771 	if (mcast && mcast->ah) {
772 		if (skb->dst            &&
773 		    skb->dst->neighbour &&
774 		    !*to_ipoib_neigh(skb->dst->neighbour)) {
775 			struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
776 
777 			if (neigh) {
778 				kref_get(&mcast->ah->ref);
779 				neigh->ah  	= mcast->ah;
780 				neigh->neighbour = skb->dst->neighbour;
781 				*to_ipoib_neigh(skb->dst->neighbour) = neigh;
782 				list_add_tail(&neigh->list, &mcast->neigh_list);
783 			}
784 		}
785 
786 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
787 	}
788 
789 unlock:
790 	spin_unlock(&priv->lock);
791 }
792 
793 void ipoib_mcast_dev_flush(struct net_device *dev)
794 {
795 	struct ipoib_dev_priv *priv = netdev_priv(dev);
796 	LIST_HEAD(remove_list);
797 	struct ipoib_mcast *mcast, *tmcast;
798 	unsigned long flags;
799 
800 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
801 
802 	spin_lock_irqsave(&priv->lock, flags);
803 
804 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
805 		list_del(&mcast->list);
806 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
807 		list_add_tail(&mcast->list, &remove_list);
808 	}
809 
810 	if (priv->broadcast) {
811  		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
812 		list_add_tail(&priv->broadcast->list, &remove_list);
813 		priv->broadcast = NULL;
814 	}
815 
816 	spin_unlock_irqrestore(&priv->lock, flags);
817 
818 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
819 		ipoib_mcast_leave(dev, mcast);
820 		ipoib_mcast_free(mcast);
821 	}
822 }
823 
824 void ipoib_mcast_restart_task(void *dev_ptr)
825 {
826 	struct net_device *dev = dev_ptr;
827 	struct ipoib_dev_priv *priv = netdev_priv(dev);
828 	struct dev_mc_list *mclist;
829 	struct ipoib_mcast *mcast, *tmcast;
830 	LIST_HEAD(remove_list);
831 	unsigned long flags;
832 
833 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
834 
835 	ipoib_mcast_stop_thread(dev, 0);
836 
837 	spin_lock_irqsave(&dev->xmit_lock, flags);
838 	spin_lock(&priv->lock);
839 
840 	/*
841 	 * Unfortunately, the networking core only gives us a list of all of
842 	 * the multicast hardware addresses. We need to figure out which ones
843 	 * are new and which ones have been removed
844 	 */
845 
846 	/* Clear out the found flag */
847 	list_for_each_entry(mcast, &priv->multicast_list, list)
848 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
849 
850 	/* Mark all of the entries that are found or don't exist */
851 	for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
852 		union ib_gid mgid;
853 
854 		memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
855 
856 		/* Add in the P_Key */
857 		mgid.raw[4] = (priv->pkey >> 8) & 0xff;
858 		mgid.raw[5] = priv->pkey & 0xff;
859 
860 		mcast = __ipoib_mcast_find(dev, &mgid);
861 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
862 			struct ipoib_mcast *nmcast;
863 
864 			/* Not found or send-only group, let's add a new entry */
865 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
866 					IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
867 
868 			nmcast = ipoib_mcast_alloc(dev, 0);
869 			if (!nmcast) {
870 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
871 				continue;
872 			}
873 
874 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
875 
876 			nmcast->mcmember.mgid = mgid;
877 
878 			if (mcast) {
879 				/* Destroy the send only entry */
880 				list_del(&mcast->list);
881 				list_add_tail(&mcast->list, &remove_list);
882 
883 				rb_replace_node(&mcast->rb_node,
884 						&nmcast->rb_node,
885 						&priv->multicast_tree);
886 			} else
887 				__ipoib_mcast_add(dev, nmcast);
888 
889 			list_add_tail(&nmcast->list, &priv->multicast_list);
890 		}
891 
892 		if (mcast)
893 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
894 	}
895 
896 	/* Remove all of the entries don't exist anymore */
897 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
898 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
899 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
900 			ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
901 					IPOIB_GID_ARG(mcast->mcmember.mgid));
902 
903 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
904 
905 			/* Move to the remove list */
906 			list_del(&mcast->list);
907 			list_add_tail(&mcast->list, &remove_list);
908 		}
909 	}
910 
911 	spin_unlock(&priv->lock);
912 	spin_unlock_irqrestore(&dev->xmit_lock, flags);
913 
914 	/* We have to cancel outside of the spinlock */
915 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
916 		ipoib_mcast_leave(mcast->dev, mcast);
917 		ipoib_mcast_free(mcast);
918 	}
919 
920 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
921 		ipoib_mcast_start_thread(dev);
922 }
923 
924 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
925 
926 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
927 {
928 	struct ipoib_mcast_iter *iter;
929 
930 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
931 	if (!iter)
932 		return NULL;
933 
934 	iter->dev = dev;
935 	memset(iter->mgid.raw, 0, 16);
936 
937 	if (ipoib_mcast_iter_next(iter)) {
938 		kfree(iter);
939 		return NULL;
940 	}
941 
942 	return iter;
943 }
944 
945 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
946 {
947 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
948 	struct rb_node *n;
949 	struct ipoib_mcast *mcast;
950 	int ret = 1;
951 
952 	spin_lock_irq(&priv->lock);
953 
954 	n = rb_first(&priv->multicast_tree);
955 
956 	while (n) {
957 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
958 
959 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
960 			   sizeof (union ib_gid)) < 0) {
961 			iter->mgid      = mcast->mcmember.mgid;
962 			iter->created   = mcast->created;
963 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
964 			iter->complete  = !!mcast->ah;
965 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
966 
967 			ret = 0;
968 
969 			break;
970 		}
971 
972 		n = rb_next(n);
973 	}
974 
975 	spin_unlock_irq(&priv->lock);
976 
977 	return ret;
978 }
979 
980 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
981 			   union ib_gid *mgid,
982 			   unsigned long *created,
983 			   unsigned int *queuelen,
984 			   unsigned int *complete,
985 			   unsigned int *send_only)
986 {
987 	*mgid      = iter->mgid;
988 	*created   = iter->created;
989 	*queuelen  = iter->queuelen;
990 	*complete  = iter->complete;
991 	*send_only = iter->send_only;
992 }
993 
994 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
995