xref: /linux/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision 9ce7677cfd7cd871adb457c80bea3b581b839641)
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 "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 DECLARE_MUTEX(mcast_mutex);
57 
58 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
59 struct ipoib_mcast {
60 	struct ib_sa_mcmember_rec mcmember;
61 	struct ipoib_ah          *ah;
62 
63 	struct rb_node    rb_node;
64 	struct list_head  list;
65 	struct completion done;
66 
67 	int                 query_id;
68 	struct ib_sa_query *query;
69 
70 	unsigned long created;
71 	unsigned long backoff;
72 
73 	unsigned long flags;
74 	unsigned char logcount;
75 
76 	struct list_head  neigh_list;
77 
78 	struct sk_buff_head pkt_queue;
79 
80 	struct net_device *dev;
81 };
82 
83 struct ipoib_mcast_iter {
84 	struct net_device *dev;
85 	union ib_gid       mgid;
86 	unsigned long      created;
87 	unsigned int       queuelen;
88 	unsigned int       complete;
89 	unsigned int       send_only;
90 };
91 
92 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
93 {
94 	struct net_device *dev = mcast->dev;
95 	struct ipoib_dev_priv *priv = netdev_priv(dev);
96 	struct ipoib_neigh *neigh, *tmp;
97 	unsigned long flags;
98 	LIST_HEAD(ah_list);
99 	struct ipoib_ah *ah, *tah;
100 
101 	ipoib_dbg_mcast(netdev_priv(dev),
102 			"deleting multicast group " IPOIB_GID_FMT "\n",
103 			IPOIB_GID_ARG(mcast->mcmember.mgid));
104 
105 	spin_lock_irqsave(&priv->lock, flags);
106 
107 	list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
108 		if (neigh->ah)
109 			list_add_tail(&neigh->ah->list, &ah_list);
110 		*to_ipoib_neigh(neigh->neighbour) = NULL;
111 		neigh->neighbour->ops->destructor = NULL;
112 		kfree(neigh);
113 	}
114 
115 	spin_unlock_irqrestore(&priv->lock, flags);
116 
117 	list_for_each_entry_safe(ah, tah, &ah_list, list)
118 		ipoib_put_ah(ah);
119 
120 	if (mcast->ah)
121 		ipoib_put_ah(mcast->ah);
122 
123 	while (!skb_queue_empty(&mcast->pkt_queue))
124 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
125 
126 	kfree(mcast);
127 }
128 
129 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
130 					     int can_sleep)
131 {
132 	struct ipoib_mcast *mcast;
133 
134 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
135 	if (!mcast)
136 		return NULL;
137 
138 	mcast->dev = dev;
139 	mcast->created = jiffies;
140 	mcast->backoff = 1;
141 
142 	INIT_LIST_HEAD(&mcast->list);
143 	INIT_LIST_HEAD(&mcast->neigh_list);
144 	skb_queue_head_init(&mcast->pkt_queue);
145 
146 	return mcast;
147 }
148 
149 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
150 {
151 	struct ipoib_dev_priv *priv = netdev_priv(dev);
152 	struct rb_node *n = priv->multicast_tree.rb_node;
153 
154 	while (n) {
155 		struct ipoib_mcast *mcast;
156 		int ret;
157 
158 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
159 
160 		ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
161 			     sizeof (union ib_gid));
162 		if (ret < 0)
163 			n = n->rb_left;
164 		else if (ret > 0)
165 			n = n->rb_right;
166 		else
167 			return mcast;
168 	}
169 
170 	return NULL;
171 }
172 
173 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
174 {
175 	struct ipoib_dev_priv *priv = netdev_priv(dev);
176 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
177 
178 	while (*n) {
179 		struct ipoib_mcast *tmcast;
180 		int ret;
181 
182 		pn = *n;
183 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
184 
185 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
186 			     sizeof (union ib_gid));
187 		if (ret < 0)
188 			n = &pn->rb_left;
189 		else if (ret > 0)
190 			n = &pn->rb_right;
191 		else
192 			return -EEXIST;
193 	}
194 
195 	rb_link_node(&mcast->rb_node, pn, n);
196 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
197 
198 	return 0;
199 }
200 
201 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
202 				   struct ib_sa_mcmember_rec *mcmember)
203 {
204 	struct net_device *dev = mcast->dev;
205 	struct ipoib_dev_priv *priv = netdev_priv(dev);
206 	int ret;
207 
208 	mcast->mcmember = *mcmember;
209 
210 	/* Set the cached Q_Key before we attach if it's the broadcast group */
211 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
212 		    sizeof (union ib_gid))) {
213 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
214 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
215 	}
216 
217 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
218 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
219 			ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
220 				   " already attached\n",
221 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
222 
223 			return 0;
224 		}
225 
226 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
227 					 &mcast->mcmember.mgid);
228 		if (ret < 0) {
229 			ipoib_warn(priv, "couldn't attach QP to multicast group "
230 				   IPOIB_GID_FMT "\n",
231 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
232 
233 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
234 			return ret;
235 		}
236 	}
237 
238 	{
239 		struct ib_ah_attr av = {
240 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
241 			.port_num      = priv->port,
242 			.sl	       = mcast->mcmember.sl,
243 			.ah_flags      = IB_AH_GRH,
244 			.grh	       = {
245 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
246 				.hop_limit     = mcast->mcmember.hop_limit,
247 				.sgid_index    = 0,
248 				.traffic_class = mcast->mcmember.traffic_class
249 			}
250 		};
251 		int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
252 
253 		av.grh.dgid = mcast->mcmember.mgid;
254 
255 		if (path_rate > 0 && priv->local_rate > path_rate)
256 			av.static_rate = (priv->local_rate - 1) / path_rate;
257 
258 		ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
259 				av.static_rate, priv->local_rate,
260 				ib_sa_rate_enum_to_int(mcast->mcmember.rate));
261 
262 		mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
263 		if (!mcast->ah) {
264 			ipoib_warn(priv, "ib_address_create failed\n");
265 		} else {
266 			ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
267 					" AV %p, LID 0x%04x, SL %d\n",
268 					IPOIB_GID_ARG(mcast->mcmember.mgid),
269 					mcast->ah->ah,
270 					be16_to_cpu(mcast->mcmember.mlid),
271 					mcast->mcmember.sl);
272 		}
273 	}
274 
275 	/* actually send any queued packets */
276 	while (!skb_queue_empty(&mcast->pkt_queue)) {
277 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
278 
279 		skb->dev = dev;
280 
281 		if (!skb->dst || !skb->dst->neighbour) {
282 			/* put pseudoheader back on for next time */
283 			skb_push(skb, sizeof (struct ipoib_pseudoheader));
284 		}
285 
286 		if (dev_queue_xmit(skb))
287 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
288 	}
289 
290 	return 0;
291 }
292 
293 static void
294 ipoib_mcast_sendonly_join_complete(int status,
295 				   struct ib_sa_mcmember_rec *mcmember,
296 				   void *mcast_ptr)
297 {
298 	struct ipoib_mcast *mcast = mcast_ptr;
299 	struct net_device *dev = mcast->dev;
300 
301 	if (!status)
302 		ipoib_mcast_join_finish(mcast, mcmember);
303 	else {
304 		if (mcast->logcount++ < 20)
305 			ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
306 					IPOIB_GID_FMT ", status %d\n",
307 					IPOIB_GID_ARG(mcast->mcmember.mgid), status);
308 
309 		/* Flush out any queued packets */
310 		while (!skb_queue_empty(&mcast->pkt_queue))
311 			dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
312 
313 		/* Clear the busy flag so we try again */
314 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
315 	}
316 
317 	complete(&mcast->done);
318 }
319 
320 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
321 {
322 	struct net_device *dev = mcast->dev;
323 	struct ipoib_dev_priv *priv = netdev_priv(dev);
324 	struct ib_sa_mcmember_rec rec = {
325 #if 0				/* Some SMs don't support send-only yet */
326 		.join_state = 4
327 #else
328 		.join_state = 1
329 #endif
330 	};
331 	int ret = 0;
332 
333 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
334 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
335 		return -ENODEV;
336 	}
337 
338 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
339 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
340 		return -EBUSY;
341 	}
342 
343 	rec.mgid     = mcast->mcmember.mgid;
344 	rec.port_gid = priv->local_gid;
345 	rec.pkey     = cpu_to_be16(priv->pkey);
346 
347 	init_completion(&mcast->done);
348 
349 	ret = ib_sa_mcmember_rec_set(priv->ca, 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 				     1000, GFP_ATOMIC,
355 				     ipoib_mcast_sendonly_join_complete,
356 				     mcast, &mcast->query);
357 	if (ret < 0) {
358 		ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
359 			   ret);
360 	} else {
361 		ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
362 				", starting join\n",
363 				IPOIB_GID_ARG(mcast->mcmember.mgid));
364 
365 		mcast->query_id = ret;
366 	}
367 
368 	return ret;
369 }
370 
371 static void ipoib_mcast_join_complete(int status,
372 				      struct ib_sa_mcmember_rec *mcmember,
373 				      void *mcast_ptr)
374 {
375 	struct ipoib_mcast *mcast = mcast_ptr;
376 	struct net_device *dev = mcast->dev;
377 	struct ipoib_dev_priv *priv = netdev_priv(dev);
378 
379 	ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
380 			" (status %d)\n",
381 			IPOIB_GID_ARG(mcast->mcmember.mgid), status);
382 
383 	if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
384 		mcast->backoff = 1;
385 		down(&mcast_mutex);
386 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
387 			queue_work(ipoib_workqueue, &priv->mcast_task);
388 		up(&mcast_mutex);
389 		complete(&mcast->done);
390 		return;
391 	}
392 
393 	if (status == -EINTR) {
394 		complete(&mcast->done);
395 		return;
396 	}
397 
398 	if (status && mcast->logcount++ < 20) {
399 		if (status == -ETIMEDOUT || status == -EINTR) {
400 			ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
401 					", status %d\n",
402 					IPOIB_GID_ARG(mcast->mcmember.mgid),
403 					status);
404 		} else {
405 			ipoib_warn(priv, "multicast join failed for "
406 				   IPOIB_GID_FMT ", status %d\n",
407 				   IPOIB_GID_ARG(mcast->mcmember.mgid),
408 				   status);
409 		}
410 	}
411 
412 	mcast->backoff *= 2;
413 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
414 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
415 
416 	mcast->query = NULL;
417 
418 	down(&mcast_mutex);
419 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
420 		if (status == -ETIMEDOUT)
421 			queue_work(ipoib_workqueue, &priv->mcast_task);
422 		else
423 			queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
424 					   mcast->backoff * HZ);
425 	} else
426 		complete(&mcast->done);
427 	up(&mcast_mutex);
428 
429 	return;
430 }
431 
432 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
433 			     int create)
434 {
435 	struct ipoib_dev_priv *priv = netdev_priv(dev);
436 	struct ib_sa_mcmember_rec rec = {
437 		.join_state = 1
438 	};
439 	ib_sa_comp_mask comp_mask;
440 	int ret = 0;
441 
442 	ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
443 			IPOIB_GID_ARG(mcast->mcmember.mgid));
444 
445 	rec.mgid     = mcast->mcmember.mgid;
446 	rec.port_gid = priv->local_gid;
447 	rec.pkey     = cpu_to_be16(priv->pkey);
448 
449 	comp_mask =
450 		IB_SA_MCMEMBER_REC_MGID		|
451 		IB_SA_MCMEMBER_REC_PORT_GID	|
452 		IB_SA_MCMEMBER_REC_PKEY		|
453 		IB_SA_MCMEMBER_REC_JOIN_STATE;
454 
455 	if (create) {
456 		comp_mask |=
457 			IB_SA_MCMEMBER_REC_QKEY		|
458 			IB_SA_MCMEMBER_REC_SL		|
459 			IB_SA_MCMEMBER_REC_FLOW_LABEL	|
460 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
461 
462 		rec.qkey	  = priv->broadcast->mcmember.qkey;
463 		rec.sl		  = priv->broadcast->mcmember.sl;
464 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
465 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
466 	}
467 
468 	init_completion(&mcast->done);
469 
470 	ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
471 				     mcast->backoff * 1000, GFP_ATOMIC,
472 				     ipoib_mcast_join_complete,
473 				     mcast, &mcast->query);
474 
475 	if (ret < 0) {
476 		ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
477 
478 		mcast->backoff *= 2;
479 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
480 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
481 
482 		down(&mcast_mutex);
483 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
484 			queue_delayed_work(ipoib_workqueue,
485 					   &priv->mcast_task,
486 					   mcast->backoff * HZ);
487 		up(&mcast_mutex);
488 	} else
489 		mcast->query_id = ret;
490 }
491 
492 void ipoib_mcast_join_task(void *dev_ptr)
493 {
494 	struct net_device *dev = dev_ptr;
495 	struct ipoib_dev_priv *priv = netdev_priv(dev);
496 
497 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
498 		return;
499 
500 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
501 		ipoib_warn(priv, "ib_gid_entry_get() failed\n");
502 	else
503 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
504 
505 	{
506 		struct ib_port_attr attr;
507 
508 		if (!ib_query_port(priv->ca, priv->port, &attr)) {
509 			priv->local_lid  = attr.lid;
510 			priv->local_rate = attr.active_speed *
511 				ib_width_enum_to_int(attr.active_width);
512 		} else
513 			ipoib_warn(priv, "ib_query_port failed\n");
514 	}
515 
516 	if (!priv->broadcast) {
517 		priv->broadcast = ipoib_mcast_alloc(dev, 1);
518 		if (!priv->broadcast) {
519 			ipoib_warn(priv, "failed to allocate broadcast group\n");
520 			down(&mcast_mutex);
521 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
522 				queue_delayed_work(ipoib_workqueue,
523 						   &priv->mcast_task, HZ);
524 			up(&mcast_mutex);
525 			return;
526 		}
527 
528 		memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
529 		       sizeof (union ib_gid));
530 
531 		spin_lock_irq(&priv->lock);
532 		__ipoib_mcast_add(dev, priv->broadcast);
533 		spin_unlock_irq(&priv->lock);
534 	}
535 
536 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
537 		ipoib_mcast_join(dev, priv->broadcast, 0);
538 		return;
539 	}
540 
541 	while (1) {
542 		struct ipoib_mcast *mcast = NULL;
543 
544 		spin_lock_irq(&priv->lock);
545 		list_for_each_entry(mcast, &priv->multicast_list, list) {
546 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
547 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
548 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
549 				/* Found the next unjoined group */
550 				break;
551 			}
552 		}
553 		spin_unlock_irq(&priv->lock);
554 
555 		if (&mcast->list == &priv->multicast_list) {
556 			/* All done */
557 			break;
558 		}
559 
560 		ipoib_mcast_join(dev, mcast, 1);
561 		return;
562 	}
563 
564 	priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
565 		IPOIB_ENCAP_LEN;
566 	dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
567 
568 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
569 
570 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
571 	netif_carrier_on(dev);
572 }
573 
574 int ipoib_mcast_start_thread(struct net_device *dev)
575 {
576 	struct ipoib_dev_priv *priv = netdev_priv(dev);
577 
578 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
579 
580 	down(&mcast_mutex);
581 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
582 		queue_work(ipoib_workqueue, &priv->mcast_task);
583 	up(&mcast_mutex);
584 
585 	return 0;
586 }
587 
588 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
589 {
590 	struct ipoib_dev_priv *priv = netdev_priv(dev);
591 	struct ipoib_mcast *mcast;
592 
593 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
594 
595 	down(&mcast_mutex);
596 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
597 	cancel_delayed_work(&priv->mcast_task);
598 	up(&mcast_mutex);
599 
600 	if (flush)
601 		flush_workqueue(ipoib_workqueue);
602 
603 	if (priv->broadcast && priv->broadcast->query) {
604 		ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
605 		priv->broadcast->query = NULL;
606 		ipoib_dbg_mcast(priv, "waiting for bcast\n");
607 		wait_for_completion(&priv->broadcast->done);
608 	}
609 
610 	list_for_each_entry(mcast, &priv->multicast_list, list) {
611 		if (mcast->query) {
612 			ib_sa_cancel_query(mcast->query_id, mcast->query);
613 			mcast->query = NULL;
614 			ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
615 					IPOIB_GID_ARG(mcast->mcmember.mgid));
616 			wait_for_completion(&mcast->done);
617 		}
618 	}
619 
620 	return 0;
621 }
622 
623 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
624 {
625 	struct ipoib_dev_priv *priv = netdev_priv(dev);
626 	struct ib_sa_mcmember_rec rec = {
627 		.join_state = 1
628 	};
629 	int ret = 0;
630 
631 	if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
632 		return 0;
633 
634 	ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
635 			IPOIB_GID_ARG(mcast->mcmember.mgid));
636 
637 	rec.mgid     = mcast->mcmember.mgid;
638 	rec.port_gid = priv->local_gid;
639 	rec.pkey     = cpu_to_be16(priv->pkey);
640 
641 	/* Remove ourselves from the multicast group */
642 	ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
643 				 &mcast->mcmember.mgid);
644 	if (ret)
645 		ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
646 
647 	/*
648 	 * Just make one shot at leaving and don't wait for a reply;
649 	 * if we fail, too bad.
650 	 */
651 	ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
652 					IB_SA_MCMEMBER_REC_MGID		|
653 					IB_SA_MCMEMBER_REC_PORT_GID	|
654 					IB_SA_MCMEMBER_REC_PKEY		|
655 					IB_SA_MCMEMBER_REC_JOIN_STATE,
656 					0, GFP_ATOMIC, NULL,
657 					mcast, &mcast->query);
658 	if (ret < 0)
659 		ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
660 			   "for leave (result = %d)\n", ret);
661 
662 	return 0;
663 }
664 
665 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
666 		      struct sk_buff *skb)
667 {
668 	struct ipoib_dev_priv *priv = netdev_priv(dev);
669 	struct ipoib_mcast *mcast;
670 
671 	/*
672 	 * We can only be called from ipoib_start_xmit, so we're
673 	 * inside tx_lock -- no need to save/restore flags.
674 	 */
675 	spin_lock(&priv->lock);
676 
677 	mcast = __ipoib_mcast_find(dev, mgid);
678 	if (!mcast) {
679 		/* Let's create a new send only group now */
680 		ipoib_dbg_mcast(priv, "setting up send only multicast group for "
681 				IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
682 
683 		mcast = ipoib_mcast_alloc(dev, 0);
684 		if (!mcast) {
685 			ipoib_warn(priv, "unable to allocate memory for "
686 				   "multicast structure\n");
687 			dev_kfree_skb_any(skb);
688 			goto out;
689 		}
690 
691 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
692 		mcast->mcmember.mgid = *mgid;
693 		__ipoib_mcast_add(dev, mcast);
694 		list_add_tail(&mcast->list, &priv->multicast_list);
695 	}
696 
697 	if (!mcast->ah) {
698 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
699 			skb_queue_tail(&mcast->pkt_queue, skb);
700 		else
701 			dev_kfree_skb_any(skb);
702 
703 		if (mcast->query)
704 			ipoib_dbg_mcast(priv, "no address vector, "
705 					"but multicast join already started\n");
706 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
707 			ipoib_mcast_sendonly_join(mcast);
708 
709 		/*
710 		 * If lookup completes between here and out:, don't
711 		 * want to send packet twice.
712 		 */
713 		mcast = NULL;
714 	}
715 
716 out:
717 	if (mcast && mcast->ah) {
718 		if (skb->dst            &&
719 		    skb->dst->neighbour &&
720 		    !*to_ipoib_neigh(skb->dst->neighbour)) {
721 			struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
722 
723 			if (neigh) {
724 				kref_get(&mcast->ah->ref);
725 				neigh->ah  	= mcast->ah;
726 				neigh->neighbour = skb->dst->neighbour;
727 				*to_ipoib_neigh(skb->dst->neighbour) = neigh;
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 	spin_unlock(&priv->lock);
736 }
737 
738 void ipoib_mcast_dev_flush(struct net_device *dev)
739 {
740 	struct ipoib_dev_priv *priv = netdev_priv(dev);
741 	LIST_HEAD(remove_list);
742 	struct ipoib_mcast *mcast, *tmcast, *nmcast;
743 	unsigned long flags;
744 
745 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
746 
747 	spin_lock_irqsave(&priv->lock, flags);
748 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
749 		nmcast = ipoib_mcast_alloc(dev, 0);
750 		if (nmcast) {
751 			nmcast->flags =
752 				mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
753 
754 			nmcast->mcmember.mgid = mcast->mcmember.mgid;
755 
756 			/* Add the new group in before the to-be-destroyed group */
757 			list_add_tail(&nmcast->list, &mcast->list);
758 			list_del_init(&mcast->list);
759 
760 			rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
761 					&priv->multicast_tree);
762 
763 			list_add_tail(&mcast->list, &remove_list);
764 		} else {
765 			ipoib_warn(priv, "could not reallocate multicast group "
766 				   IPOIB_GID_FMT "\n",
767 				   IPOIB_GID_ARG(mcast->mcmember.mgid));
768 		}
769 	}
770 
771 	if (priv->broadcast) {
772 		nmcast = ipoib_mcast_alloc(dev, 0);
773 		if (nmcast) {
774 			nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
775 
776 			rb_replace_node(&priv->broadcast->rb_node,
777 					&nmcast->rb_node,
778 					&priv->multicast_tree);
779 
780 			list_add_tail(&priv->broadcast->list, &remove_list);
781 		}
782 
783 		priv->broadcast = nmcast;
784 	}
785 
786 	spin_unlock_irqrestore(&priv->lock, flags);
787 
788 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
789 		ipoib_mcast_leave(dev, mcast);
790 		ipoib_mcast_free(mcast);
791 	}
792 }
793 
794 void ipoib_mcast_dev_down(struct net_device *dev)
795 {
796 	struct ipoib_dev_priv *priv = netdev_priv(dev);
797 	unsigned long flags;
798 
799 	/* Delete broadcast since it will be recreated */
800 	if (priv->broadcast) {
801 		ipoib_dbg_mcast(priv, "deleting broadcast group\n");
802 
803 		spin_lock_irqsave(&priv->lock, flags);
804 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
805 		spin_unlock_irqrestore(&priv->lock, flags);
806 		ipoib_mcast_leave(dev, priv->broadcast);
807 		ipoib_mcast_free(priv->broadcast);
808 		priv->broadcast = NULL;
809 	}
810 }
811 
812 void ipoib_mcast_restart_task(void *dev_ptr)
813 {
814 	struct net_device *dev = dev_ptr;
815 	struct ipoib_dev_priv *priv = netdev_priv(dev);
816 	struct dev_mc_list *mclist;
817 	struct ipoib_mcast *mcast, *tmcast;
818 	LIST_HEAD(remove_list);
819 	unsigned long flags;
820 
821 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
822 
823 	ipoib_mcast_stop_thread(dev, 0);
824 
825 	spin_lock_irqsave(&priv->lock, flags);
826 
827 	/*
828 	 * Unfortunately, the networking core only gives us a list of all of
829 	 * the multicast hardware addresses. We need to figure out which ones
830 	 * are new and which ones have been removed
831 	 */
832 
833 	/* Clear out the found flag */
834 	list_for_each_entry(mcast, &priv->multicast_list, list)
835 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
836 
837 	/* Mark all of the entries that are found or don't exist */
838 	for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
839 		union ib_gid mgid;
840 
841 		memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
842 
843 		/* Add in the P_Key */
844 		mgid.raw[4] = (priv->pkey >> 8) & 0xff;
845 		mgid.raw[5] = priv->pkey & 0xff;
846 
847 		mcast = __ipoib_mcast_find(dev, &mgid);
848 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
849 			struct ipoib_mcast *nmcast;
850 
851 			/* Not found or send-only group, let's add a new entry */
852 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
853 					IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
854 
855 			nmcast = ipoib_mcast_alloc(dev, 0);
856 			if (!nmcast) {
857 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
858 				continue;
859 			}
860 
861 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
862 
863 			nmcast->mcmember.mgid = mgid;
864 
865 			if (mcast) {
866 				/* Destroy the send only entry */
867 				list_del(&mcast->list);
868 				list_add_tail(&mcast->list, &remove_list);
869 
870 				rb_replace_node(&mcast->rb_node,
871 						&nmcast->rb_node,
872 						&priv->multicast_tree);
873 			} else
874 				__ipoib_mcast_add(dev, nmcast);
875 
876 			list_add_tail(&nmcast->list, &priv->multicast_list);
877 		}
878 
879 		if (mcast)
880 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
881 	}
882 
883 	/* Remove all of the entries don't exist anymore */
884 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
885 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
886 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
887 			ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
888 					IPOIB_GID_ARG(mcast->mcmember.mgid));
889 
890 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
891 
892 			/* Move to the remove list */
893 			list_del(&mcast->list);
894 			list_add_tail(&mcast->list, &remove_list);
895 		}
896 	}
897 	spin_unlock_irqrestore(&priv->lock, flags);
898 
899 	/* We have to cancel outside of the spinlock */
900 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
901 		ipoib_mcast_leave(mcast->dev, mcast);
902 		ipoib_mcast_free(mcast);
903 	}
904 
905 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
906 		ipoib_mcast_start_thread(dev);
907 }
908 
909 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
910 
911 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
912 {
913 	struct ipoib_mcast_iter *iter;
914 
915 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
916 	if (!iter)
917 		return NULL;
918 
919 	iter->dev = dev;
920 	memset(iter->mgid.raw, 0, 16);
921 
922 	if (ipoib_mcast_iter_next(iter)) {
923 		kfree(iter);
924 		return NULL;
925 	}
926 
927 	return iter;
928 }
929 
930 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
931 {
932 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
933 	struct rb_node *n;
934 	struct ipoib_mcast *mcast;
935 	int ret = 1;
936 
937 	spin_lock_irq(&priv->lock);
938 
939 	n = rb_first(&priv->multicast_tree);
940 
941 	while (n) {
942 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
943 
944 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
945 			   sizeof (union ib_gid)) < 0) {
946 			iter->mgid      = mcast->mcmember.mgid;
947 			iter->created   = mcast->created;
948 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
949 			iter->complete  = !!mcast->ah;
950 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
951 
952 			ret = 0;
953 
954 			break;
955 		}
956 
957 		n = rb_next(n);
958 	}
959 
960 	spin_unlock_irq(&priv->lock);
961 
962 	return ret;
963 }
964 
965 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
966 			   union ib_gid *mgid,
967 			   unsigned long *created,
968 			   unsigned int *queuelen,
969 			   unsigned int *complete,
970 			   unsigned int *send_only)
971 {
972 	*mgid      = iter->mgid;
973 	*created   = iter->created;
974 	*queuelen  = iter->queuelen;
975 	*complete  = iter->complete;
976 	*send_only = iter->send_only;
977 }
978 
979 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
980