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