1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/skbuff.h>
36 #include <linux/rtnetlink.h>
37 #include <linux/moduleparam.h>
38 #include <linux/ip.h>
39 #include <linux/in.h>
40 #include <linux/igmp.h>
41 #include <linux/inetdevice.h>
42 #include <linux/delay.h>
43 #include <linux/completion.h>
44 #include <linux/slab.h>
45 
46 #include <net/dst.h>
47 
48 #include "ipoib.h"
49 
50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
51 static int mcast_debug_level;
52 
53 module_param(mcast_debug_level, int, 0644);
54 MODULE_PARM_DESC(mcast_debug_level,
55 		 "Enable multicast debug tracing if > 0");
56 #endif
57 
58 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 /* join state that allows creating mcg with sendonly member request */
68 #define SENDONLY_FULLMEMBER_JOIN	8
69 
70 /*
71  * This should be called with the priv->lock held
72  */
73 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
74 					       struct ipoib_mcast *mcast,
75 					       bool delay)
76 {
77 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
78 		return;
79 
80 	/*
81 	 * We will be scheduling *something*, so cancel whatever is
82 	 * currently scheduled first
83 	 */
84 	cancel_delayed_work(&priv->mcast_task);
85 	if (mcast && delay) {
86 		/*
87 		 * We had a failure and want to schedule a retry later
88 		 */
89 		mcast->backoff *= 2;
90 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
91 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
92 		mcast->delay_until = jiffies + (mcast->backoff * HZ);
93 		/*
94 		 * Mark this mcast for its delay, but restart the
95 		 * task immediately.  The join task will make sure to
96 		 * clear out all entries without delays, and then
97 		 * schedule itself to run again when the earliest
98 		 * delay expires
99 		 */
100 		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
101 	} else if (delay) {
102 		/*
103 		 * Special case of retrying after a failure to
104 		 * allocate the broadcast multicast group, wait
105 		 * 1 second and try again
106 		 */
107 		queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
108 	} else
109 		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
110 }
111 
112 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
113 {
114 	struct net_device *dev = mcast->dev;
115 	int tx_dropped = 0;
116 
117 	ipoib_dbg_mcast(ipoib_priv(dev), "deleting multicast group %pI6\n",
118 			mcast->mcmember.mgid.raw);
119 
120 	/* remove all neigh connected to this mcast */
121 	ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
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 	netif_tx_lock_bh(dev);
132 	dev->stats.tx_dropped += tx_dropped;
133 	netif_tx_unlock_bh(dev);
134 
135 	kfree(mcast);
136 }
137 
138 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev)
139 {
140 	struct ipoib_mcast *mcast;
141 
142 	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
143 	if (!mcast)
144 		return NULL;
145 
146 	mcast->dev = dev;
147 	mcast->created = jiffies;
148 	mcast->delay_until = 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, void *mgid)
159 {
160 	struct ipoib_dev_priv *priv = ipoib_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, 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 = ipoib_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 = ipoib_priv(dev);
215 	struct rdma_netdev *rn = netdev_priv(dev);
216 	struct ipoib_ah *ah;
217 	struct rdma_ah_attr av;
218 	int ret;
219 	int set_qkey = 0;
220 	int mtu;
221 
222 	mcast->mcmember = *mcmember;
223 
224 	/* Set the multicast MTU and cached Q_Key before we attach if it's
225 	 * the broadcast group.
226 	 */
227 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
228 		    sizeof (union ib_gid))) {
229 		spin_lock_irq(&priv->lock);
230 		if (!priv->broadcast) {
231 			spin_unlock_irq(&priv->lock);
232 			return -EAGAIN;
233 		}
234 		/*update priv member according to the new mcast*/
235 		priv->broadcast->mcmember.qkey = mcmember->qkey;
236 		priv->broadcast->mcmember.mtu = mcmember->mtu;
237 		priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
238 		priv->broadcast->mcmember.rate = mcmember->rate;
239 		priv->broadcast->mcmember.sl = mcmember->sl;
240 		priv->broadcast->mcmember.flow_label = mcmember->flow_label;
241 		priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
242 		/* assume if the admin and the mcast are the same both can be changed */
243 		mtu = rdma_mtu_enum_to_int(priv->ca,  priv->port,
244 					   priv->broadcast->mcmember.mtu);
245 		if (priv->mcast_mtu == priv->admin_mtu)
246 			priv->admin_mtu = IPOIB_UD_MTU(mtu);
247 		priv->mcast_mtu = IPOIB_UD_MTU(mtu);
248 		rn->mtu = priv->mcast_mtu;
249 
250 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
251 		spin_unlock_irq(&priv->lock);
252 		priv->tx_wr.remote_qkey = priv->qkey;
253 		set_qkey = 1;
254 	}
255 
256 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
257 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
258 			ipoib_warn(priv, "multicast group %pI6 already attached\n",
259 				   mcast->mcmember.mgid.raw);
260 
261 			return 0;
262 		}
263 
264 		ret = rn->attach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
265 				       be16_to_cpu(mcast->mcmember.mlid),
266 				       set_qkey, priv->qkey);
267 		if (ret < 0) {
268 			ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
269 				   mcast->mcmember.mgid.raw);
270 
271 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
272 			return ret;
273 		}
274 	}
275 
276 	memset(&av, 0, sizeof(av));
277 	av.type = rdma_ah_find_type(priv->ca, priv->port);
278 	rdma_ah_set_dlid(&av, be16_to_cpu(mcast->mcmember.mlid));
279 	rdma_ah_set_port_num(&av, priv->port);
280 	rdma_ah_set_sl(&av, mcast->mcmember.sl);
281 	rdma_ah_set_static_rate(&av, mcast->mcmember.rate);
282 
283 	rdma_ah_set_grh(&av, &mcast->mcmember.mgid,
284 			be32_to_cpu(mcast->mcmember.flow_label),
285 			0, mcast->mcmember.hop_limit,
286 			mcast->mcmember.traffic_class);
287 
288 	ah = ipoib_create_ah(dev, priv->pd, &av);
289 	if (IS_ERR(ah)) {
290 		ipoib_warn(priv, "ib_address_create failed %ld\n",
291 			   -PTR_ERR(ah));
292 		/* use original error */
293 		return PTR_ERR(ah);
294 	}
295 	spin_lock_irq(&priv->lock);
296 	mcast->ah = ah;
297 	spin_unlock_irq(&priv->lock);
298 
299 	ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
300 			mcast->mcmember.mgid.raw,
301 			mcast->ah->ah,
302 			be16_to_cpu(mcast->mcmember.mlid),
303 			mcast->mcmember.sl);
304 
305 	/* actually send any queued packets */
306 	netif_tx_lock_bh(dev);
307 	while (!skb_queue_empty(&mcast->pkt_queue)) {
308 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
309 
310 		netif_tx_unlock_bh(dev);
311 
312 		skb->dev = dev;
313 
314 		ret = dev_queue_xmit(skb);
315 		if (ret)
316 			ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
317 				   __func__, ret);
318 		netif_tx_lock_bh(dev);
319 	}
320 	netif_tx_unlock_bh(dev);
321 
322 	return 0;
323 }
324 
325 void ipoib_mcast_carrier_on_task(struct work_struct *work)
326 {
327 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
328 						   carrier_on_task);
329 	struct ib_port_attr attr;
330 
331 	if (ib_query_port(priv->ca, priv->port, &attr) ||
332 	    attr.state != IB_PORT_ACTIVE) {
333 		ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
334 		return;
335 	}
336 	/*
337 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
338 	 * turning the carrier back on while a device is being
339 	 * removed.  However, ipoib_stop() will attempt to flush
340 	 * the workqueue while holding the rtnl lock, so loop
341 	 * on trylock until either we get the lock or we see
342 	 * FLAG_OPER_UP go away as that signals that we are bailing
343 	 * and can safely ignore the carrier on work.
344 	 */
345 	while (!rtnl_trylock()) {
346 		if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
347 			return;
348 		else
349 			msleep(20);
350 	}
351 	if (!ipoib_cm_admin_enabled(priv->dev))
352 		dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
353 	netif_carrier_on(priv->dev);
354 	rtnl_unlock();
355 }
356 
357 static int ipoib_mcast_join_complete(int status,
358 				     struct ib_sa_multicast *multicast)
359 {
360 	struct ipoib_mcast *mcast = multicast->context;
361 	struct net_device *dev = mcast->dev;
362 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
363 
364 	ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
365 			test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
366 			"sendonly " : "",
367 			mcast->mcmember.mgid.raw, status);
368 
369 	/* We trap for port events ourselves. */
370 	if (status == -ENETRESET) {
371 		status = 0;
372 		goto out;
373 	}
374 
375 	if (!status)
376 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
377 
378 	if (!status) {
379 		mcast->backoff = 1;
380 		mcast->delay_until = jiffies;
381 
382 		/*
383 		 * Defer carrier on work to priv->wq to avoid a
384 		 * deadlock on rtnl_lock here.  Requeue our multicast
385 		 * work too, which will end up happening right after
386 		 * our carrier on task work and will allow us to
387 		 * send out all of the non-broadcast joins
388 		 */
389 		if (mcast == priv->broadcast) {
390 			spin_lock_irq(&priv->lock);
391 			queue_work(priv->wq, &priv->carrier_on_task);
392 			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
393 			goto out_locked;
394 		}
395 	} else {
396 		bool silent_fail =
397 		    test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
398 		    status == -EINVAL;
399 
400 		if (mcast->logcount < 20) {
401 			if (status == -ETIMEDOUT || status == -EAGAIN ||
402 			    silent_fail) {
403 				ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
404 						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
405 						mcast->mcmember.mgid.raw, status);
406 			} else {
407 				ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
408 						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
409 					   mcast->mcmember.mgid.raw, status);
410 			}
411 
412 			if (!silent_fail)
413 				mcast->logcount++;
414 		}
415 
416 		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
417 		    mcast->backoff >= 2) {
418 			/*
419 			 * We only retry sendonly joins once before we drop
420 			 * the packet and quit trying to deal with the
421 			 * group.  However, we leave the group in the
422 			 * mcast list as an unjoined group.  If we want to
423 			 * try joining again, we simply queue up a packet
424 			 * and restart the join thread.  The empty queue
425 			 * is why the join thread ignores this group.
426 			 */
427 			mcast->backoff = 1;
428 			netif_tx_lock_bh(dev);
429 			while (!skb_queue_empty(&mcast->pkt_queue)) {
430 				++dev->stats.tx_dropped;
431 				dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
432 			}
433 			netif_tx_unlock_bh(dev);
434 		} else {
435 			spin_lock_irq(&priv->lock);
436 			/* Requeue this join task with a backoff delay */
437 			__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
438 			goto out_locked;
439 		}
440 	}
441 out:
442 	spin_lock_irq(&priv->lock);
443 out_locked:
444 	/*
445 	 * Make sure to set mcast->mc before we clear the busy flag to avoid
446 	 * racing with code that checks for BUSY before checking mcast->mc
447 	 */
448 	if (status)
449 		mcast->mc = NULL;
450 	else
451 		mcast->mc = multicast;
452 	clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
453 	spin_unlock_irq(&priv->lock);
454 	complete(&mcast->done);
455 
456 	return status;
457 }
458 
459 /*
460  * Caller must hold 'priv->lock'
461  */
462 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
463 {
464 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
465 	struct ib_sa_multicast *multicast;
466 	struct ib_sa_mcmember_rec rec = {
467 		.join_state = 1
468 	};
469 	ib_sa_comp_mask comp_mask;
470 	int ret = 0;
471 
472 	if (!priv->broadcast ||
473 	    !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
474 		return -EINVAL;
475 
476 	init_completion(&mcast->done);
477 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
478 
479 	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
480 
481 	rec.mgid     = mcast->mcmember.mgid;
482 	rec.port_gid = priv->local_gid;
483 	rec.pkey     = cpu_to_be16(priv->pkey);
484 
485 	comp_mask =
486 		IB_SA_MCMEMBER_REC_MGID		|
487 		IB_SA_MCMEMBER_REC_PORT_GID	|
488 		IB_SA_MCMEMBER_REC_PKEY		|
489 		IB_SA_MCMEMBER_REC_JOIN_STATE;
490 
491 	if (mcast != priv->broadcast) {
492 		/*
493 		 * RFC 4391:
494 		 *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
495 		 *  and HopLimit as those used in the broadcast-GID.  The rest
496 		 *  of attributes SHOULD follow the values used in the
497 		 *  broadcast-GID as well.
498 		 */
499 		comp_mask |=
500 			IB_SA_MCMEMBER_REC_QKEY			|
501 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
502 			IB_SA_MCMEMBER_REC_MTU			|
503 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
504 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
505 			IB_SA_MCMEMBER_REC_RATE			|
506 			IB_SA_MCMEMBER_REC_SL			|
507 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
508 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
509 
510 		rec.qkey	  = priv->broadcast->mcmember.qkey;
511 		rec.mtu_selector  = IB_SA_EQ;
512 		rec.mtu		  = priv->broadcast->mcmember.mtu;
513 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
514 		rec.rate_selector = IB_SA_EQ;
515 		rec.rate	  = priv->broadcast->mcmember.rate;
516 		rec.sl		  = priv->broadcast->mcmember.sl;
517 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
518 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
519 
520 		/*
521 		 * Send-only IB Multicast joins work at the core IB layer but
522 		 * require specific SM support.
523 		 * We can use such joins here only if the current SM supports that feature.
524 		 * However, if not, we emulate an Ethernet multicast send,
525 		 * which does not require a multicast subscription and will
526 		 * still send properly. The most appropriate thing to
527 		 * do is to create the group if it doesn't exist as that
528 		 * most closely emulates the behavior, from a user space
529 		 * application perspective, of Ethernet multicast operation.
530 		 */
531 		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
532 			rec.join_state = SENDONLY_FULLMEMBER_JOIN;
533 	}
534 	spin_unlock_irq(&priv->lock);
535 
536 	multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
537 					 &rec, comp_mask, GFP_KERNEL,
538 					 ipoib_mcast_join_complete, mcast);
539 	spin_lock_irq(&priv->lock);
540 	if (IS_ERR(multicast)) {
541 		ret = PTR_ERR(multicast);
542 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
543 		/* Requeue this join task with a backoff delay */
544 		__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
545 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
546 		spin_unlock_irq(&priv->lock);
547 		complete(&mcast->done);
548 		spin_lock_irq(&priv->lock);
549 	}
550 	return 0;
551 }
552 
553 void ipoib_mcast_join_task(struct work_struct *work)
554 {
555 	struct ipoib_dev_priv *priv =
556 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
557 	struct net_device *dev = priv->dev;
558 	struct ib_port_attr port_attr;
559 	unsigned long delay_until = 0;
560 	struct ipoib_mcast *mcast = NULL;
561 
562 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
563 		return;
564 
565 	if (ib_query_port(priv->ca, priv->port, &port_attr)) {
566 		ipoib_dbg(priv, "ib_query_port() failed\n");
567 		return;
568 	}
569 	if (port_attr.state != IB_PORT_ACTIVE) {
570 		ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
571 			  port_attr.state);
572 		return;
573 	}
574 	priv->local_lid = port_attr.lid;
575 	netif_addr_lock_bh(dev);
576 
577 	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
578 		netif_addr_unlock_bh(dev);
579 		return;
580 	}
581 	netif_addr_unlock_bh(dev);
582 
583 	spin_lock_irq(&priv->lock);
584 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
585 		goto out;
586 
587 	if (!priv->broadcast) {
588 		struct ipoib_mcast *broadcast;
589 
590 		broadcast = ipoib_mcast_alloc(dev);
591 		if (!broadcast) {
592 			ipoib_warn(priv, "failed to allocate broadcast group\n");
593 			/*
594 			 * Restart us after a 1 second delay to retry
595 			 * creating our broadcast group and attaching to
596 			 * it.  Until this succeeds, this ipoib dev is
597 			 * completely stalled (multicast wise).
598 			 */
599 			__ipoib_mcast_schedule_join_thread(priv, NULL, 1);
600 			goto out;
601 		}
602 
603 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
604 		       sizeof (union ib_gid));
605 		priv->broadcast = broadcast;
606 
607 		__ipoib_mcast_add(dev, priv->broadcast);
608 	}
609 
610 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
611 		if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
612 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
613 			mcast = priv->broadcast;
614 			if (mcast->backoff > 1 &&
615 			    time_before(jiffies, mcast->delay_until)) {
616 				delay_until = mcast->delay_until;
617 				mcast = NULL;
618 			}
619 		}
620 		goto out;
621 	}
622 
623 	/*
624 	 * We'll never get here until the broadcast group is both allocated
625 	 * and attached
626 	 */
627 	list_for_each_entry(mcast, &priv->multicast_list, list) {
628 		if (IS_ERR_OR_NULL(mcast->mc) &&
629 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
630 		    (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
631 		     !skb_queue_empty(&mcast->pkt_queue))) {
632 			if (mcast->backoff == 1 ||
633 			    time_after_eq(jiffies, mcast->delay_until)) {
634 				/* Found the next unjoined group */
635 				if (ipoib_mcast_join(dev, mcast)) {
636 					spin_unlock_irq(&priv->lock);
637 					return;
638 				}
639 			} else if (!delay_until ||
640 				 time_before(mcast->delay_until, delay_until))
641 				delay_until = mcast->delay_until;
642 		}
643 	}
644 
645 	mcast = NULL;
646 	ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
647 
648 out:
649 	if (delay_until) {
650 		cancel_delayed_work(&priv->mcast_task);
651 		queue_delayed_work(priv->wq, &priv->mcast_task,
652 				   delay_until - jiffies);
653 	}
654 	if (mcast)
655 		ipoib_mcast_join(dev, mcast);
656 
657 	spin_unlock_irq(&priv->lock);
658 }
659 
660 void ipoib_mcast_start_thread(struct net_device *dev)
661 {
662 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
663 	unsigned long flags;
664 
665 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
666 
667 	spin_lock_irqsave(&priv->lock, flags);
668 	__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
669 	spin_unlock_irqrestore(&priv->lock, flags);
670 }
671 
672 void ipoib_mcast_stop_thread(struct net_device *dev)
673 {
674 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
675 
676 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
677 
678 	cancel_delayed_work_sync(&priv->mcast_task);
679 }
680 
681 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
682 {
683 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
684 	struct rdma_netdev *rn = netdev_priv(dev);
685 	int ret = 0;
686 
687 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
688 		ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
689 
690 	if (!IS_ERR_OR_NULL(mcast->mc))
691 		ib_sa_free_multicast(mcast->mc);
692 
693 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
694 		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
695 				mcast->mcmember.mgid.raw);
696 
697 		/* Remove ourselves from the multicast group */
698 		ret = rn->detach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
699 				       be16_to_cpu(mcast->mcmember.mlid));
700 		if (ret)
701 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
702 	} else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
703 		ipoib_dbg(priv, "leaving with no mcmember but not a "
704 			  "SENDONLY join\n");
705 
706 	return 0;
707 }
708 
709 /*
710  * Check if the multicast group is sendonly. If so remove it from the maps
711  * and add to the remove list
712  */
713 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
714 				struct list_head *remove_list)
715 {
716 	/* Is this multicast ? */
717 	if (*mgid == 0xff) {
718 		struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
719 
720 		if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
721 			list_del(&mcast->list);
722 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
723 			list_add_tail(&mcast->list, remove_list);
724 		}
725 	}
726 }
727 
728 void ipoib_mcast_remove_list(struct list_head *remove_list)
729 {
730 	struct ipoib_mcast *mcast, *tmcast;
731 
732 	/*
733 	 * make sure the in-flight joins have finished before we attempt
734 	 * to leave
735 	 */
736 	list_for_each_entry_safe(mcast, tmcast, remove_list, list)
737 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
738 			wait_for_completion(&mcast->done);
739 
740 	list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
741 		ipoib_mcast_leave(mcast->dev, mcast);
742 		ipoib_mcast_free(mcast);
743 	}
744 }
745 
746 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
747 {
748 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
749 	struct rdma_netdev *rn = netdev_priv(dev);
750 	struct ipoib_mcast *mcast;
751 	unsigned long flags;
752 	void *mgid = daddr + 4;
753 
754 	spin_lock_irqsave(&priv->lock, flags);
755 
756 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
757 	    !priv->broadcast					||
758 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
759 		++dev->stats.tx_dropped;
760 		dev_kfree_skb_any(skb);
761 		goto unlock;
762 	}
763 
764 	mcast = __ipoib_mcast_find(dev, mgid);
765 	if (!mcast || !mcast->ah) {
766 		if (!mcast) {
767 			/* Let's create a new send only group now */
768 			ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
769 					mgid);
770 
771 			mcast = ipoib_mcast_alloc(dev);
772 			if (!mcast) {
773 				ipoib_warn(priv, "unable to allocate memory "
774 					   "for multicast structure\n");
775 				++dev->stats.tx_dropped;
776 				dev_kfree_skb_any(skb);
777 				goto unlock;
778 			}
779 
780 			set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
781 			memcpy(mcast->mcmember.mgid.raw, mgid,
782 			       sizeof (union ib_gid));
783 			__ipoib_mcast_add(dev, mcast);
784 			list_add_tail(&mcast->list, &priv->multicast_list);
785 		}
786 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
787 			/* put pseudoheader back on for next time */
788 			skb_push(skb, sizeof(struct ipoib_pseudo_header));
789 			skb_queue_tail(&mcast->pkt_queue, skb);
790 		} else {
791 			++dev->stats.tx_dropped;
792 			dev_kfree_skb_any(skb);
793 		}
794 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
795 			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
796 		}
797 	} else {
798 		struct ipoib_neigh *neigh;
799 
800 		spin_unlock_irqrestore(&priv->lock, flags);
801 		neigh = ipoib_neigh_get(dev, daddr);
802 		spin_lock_irqsave(&priv->lock, flags);
803 		if (!neigh) {
804 			neigh = ipoib_neigh_alloc(daddr, dev);
805 			/* Make sure that the neigh will be added only
806 			 * once to mcast list.
807 			 */
808 			if (neigh && list_empty(&neigh->list)) {
809 				kref_get(&mcast->ah->ref);
810 				neigh->ah	= mcast->ah;
811 				neigh->ah->valid = 1;
812 				list_add_tail(&neigh->list, &mcast->neigh_list);
813 			}
814 		}
815 		spin_unlock_irqrestore(&priv->lock, flags);
816 		mcast->ah->last_send = rn->send(dev, skb, mcast->ah->ah,
817 						IB_MULTICAST_QPN);
818 		if (neigh)
819 			ipoib_neigh_put(neigh);
820 		return;
821 	}
822 
823 unlock:
824 	spin_unlock_irqrestore(&priv->lock, flags);
825 }
826 
827 void ipoib_mcast_dev_flush(struct net_device *dev)
828 {
829 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
830 	LIST_HEAD(remove_list);
831 	struct ipoib_mcast *mcast, *tmcast;
832 	unsigned long flags;
833 
834 	mutex_lock(&priv->mcast_mutex);
835 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
836 
837 	spin_lock_irqsave(&priv->lock, flags);
838 
839 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
840 		list_del(&mcast->list);
841 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
842 		list_add_tail(&mcast->list, &remove_list);
843 	}
844 
845 	if (priv->broadcast) {
846 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
847 		list_add_tail(&priv->broadcast->list, &remove_list);
848 		priv->broadcast = NULL;
849 	}
850 
851 	spin_unlock_irqrestore(&priv->lock, flags);
852 
853 	ipoib_mcast_remove_list(&remove_list);
854 	mutex_unlock(&priv->mcast_mutex);
855 }
856 
857 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
858 {
859 	/* reserved QPN, prefix, scope */
860 	if (memcmp(addr, broadcast, 6))
861 		return 0;
862 	/* signature lower, pkey */
863 	if (memcmp(addr + 7, broadcast + 7, 3))
864 		return 0;
865 	return 1;
866 }
867 
868 void ipoib_mcast_restart_task(struct work_struct *work)
869 {
870 	struct ipoib_dev_priv *priv =
871 		container_of(work, struct ipoib_dev_priv, restart_task);
872 	struct net_device *dev = priv->dev;
873 	struct netdev_hw_addr *ha;
874 	struct ipoib_mcast *mcast, *tmcast;
875 	LIST_HEAD(remove_list);
876 	struct ib_sa_mcmember_rec rec;
877 
878 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
879 		/*
880 		 * shortcut...on shutdown flush is called next, just
881 		 * let it do all the work
882 		 */
883 		return;
884 
885 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
886 
887 	netif_addr_lock_bh(dev);
888 	spin_lock_irq(&priv->lock);
889 
890 	/*
891 	 * Unfortunately, the networking core only gives us a list of all of
892 	 * the multicast hardware addresses. We need to figure out which ones
893 	 * are new and which ones have been removed
894 	 */
895 
896 	/* Clear out the found flag */
897 	list_for_each_entry(mcast, &priv->multicast_list, list)
898 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
899 
900 	/* Mark all of the entries that are found or don't exist */
901 	netdev_for_each_mc_addr(ha, dev) {
902 		union ib_gid mgid;
903 
904 		if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
905 			continue;
906 
907 		memcpy(mgid.raw, ha->addr + 4, sizeof(mgid));
908 
909 		mcast = __ipoib_mcast_find(dev, &mgid);
910 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
911 			struct ipoib_mcast *nmcast;
912 
913 			/* ignore group which is directly joined by userspace */
914 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
915 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
916 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
917 						mgid.raw);
918 				continue;
919 			}
920 
921 			/* Not found or send-only group, let's add a new entry */
922 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
923 					mgid.raw);
924 
925 			nmcast = ipoib_mcast_alloc(dev);
926 			if (!nmcast) {
927 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
928 				continue;
929 			}
930 
931 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
932 
933 			nmcast->mcmember.mgid = mgid;
934 
935 			if (mcast) {
936 				/* Destroy the send only entry */
937 				list_move_tail(&mcast->list, &remove_list);
938 
939 				rb_replace_node(&mcast->rb_node,
940 						&nmcast->rb_node,
941 						&priv->multicast_tree);
942 			} else
943 				__ipoib_mcast_add(dev, nmcast);
944 
945 			list_add_tail(&nmcast->list, &priv->multicast_list);
946 		}
947 
948 		if (mcast)
949 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
950 	}
951 
952 	/* Remove all of the entries don't exist anymore */
953 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
954 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
955 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
956 			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
957 					mcast->mcmember.mgid.raw);
958 
959 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
960 
961 			/* Move to the remove list */
962 			list_move_tail(&mcast->list, &remove_list);
963 		}
964 	}
965 
966 	spin_unlock_irq(&priv->lock);
967 	netif_addr_unlock_bh(dev);
968 
969 	ipoib_mcast_remove_list(&remove_list);
970 
971 	/*
972 	 * Double check that we are still up
973 	 */
974 	if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
975 		spin_lock_irq(&priv->lock);
976 		__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
977 		spin_unlock_irq(&priv->lock);
978 	}
979 }
980 
981 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
982 
983 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
984 {
985 	struct ipoib_mcast_iter *iter;
986 
987 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
988 	if (!iter)
989 		return NULL;
990 
991 	iter->dev = dev;
992 	memset(iter->mgid.raw, 0, 16);
993 
994 	if (ipoib_mcast_iter_next(iter)) {
995 		kfree(iter);
996 		return NULL;
997 	}
998 
999 	return iter;
1000 }
1001 
1002 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1003 {
1004 	struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
1005 	struct rb_node *n;
1006 	struct ipoib_mcast *mcast;
1007 	int ret = 1;
1008 
1009 	spin_lock_irq(&priv->lock);
1010 
1011 	n = rb_first(&priv->multicast_tree);
1012 
1013 	while (n) {
1014 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1015 
1016 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1017 			   sizeof (union ib_gid)) < 0) {
1018 			iter->mgid      = mcast->mcmember.mgid;
1019 			iter->created   = mcast->created;
1020 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1021 			iter->complete  = !!mcast->ah;
1022 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1023 
1024 			ret = 0;
1025 
1026 			break;
1027 		}
1028 
1029 		n = rb_next(n);
1030 	}
1031 
1032 	spin_unlock_irq(&priv->lock);
1033 
1034 	return ret;
1035 }
1036 
1037 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1038 			   union ib_gid *mgid,
1039 			   unsigned long *created,
1040 			   unsigned int *queuelen,
1041 			   unsigned int *complete,
1042 			   unsigned int *send_only)
1043 {
1044 	*mgid      = iter->mgid;
1045 	*created   = iter->created;
1046 	*queuelen  = iter->queuelen;
1047 	*complete  = iter->complete;
1048 	*send_only = iter->send_only;
1049 }
1050 
1051 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
1052