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 	 * Check if can send sendonly MCG's with sendonly-fullmember join state.
338 	 * It done here after the successfully join to the broadcast group,
339 	 * because the broadcast group must always be joined first and is always
340 	 * re-joined if the SM changes substantially.
341 	 */
342 	priv->sm_fullmember_sendonly_support =
343 		ib_sa_sendonly_fullmem_support(&ipoib_sa_client,
344 					       priv->ca, priv->port);
345 	/*
346 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
347 	 * turning the carrier back on while a device is being
348 	 * removed.  However, ipoib_stop() will attempt to flush
349 	 * the workqueue while holding the rtnl lock, so loop
350 	 * on trylock until either we get the lock or we see
351 	 * FLAG_OPER_UP go away as that signals that we are bailing
352 	 * and can safely ignore the carrier on work.
353 	 */
354 	while (!rtnl_trylock()) {
355 		if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
356 			return;
357 		else
358 			msleep(20);
359 	}
360 	if (!ipoib_cm_admin_enabled(priv->dev))
361 		dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
362 	netif_carrier_on(priv->dev);
363 	rtnl_unlock();
364 }
365 
366 static int ipoib_mcast_join_complete(int status,
367 				     struct ib_sa_multicast *multicast)
368 {
369 	struct ipoib_mcast *mcast = multicast->context;
370 	struct net_device *dev = mcast->dev;
371 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
372 
373 	ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
374 			test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
375 			"sendonly " : "",
376 			mcast->mcmember.mgid.raw, status);
377 
378 	/* We trap for port events ourselves. */
379 	if (status == -ENETRESET) {
380 		status = 0;
381 		goto out;
382 	}
383 
384 	if (!status)
385 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
386 
387 	if (!status) {
388 		mcast->backoff = 1;
389 		mcast->delay_until = jiffies;
390 
391 		/*
392 		 * Defer carrier on work to priv->wq to avoid a
393 		 * deadlock on rtnl_lock here.  Requeue our multicast
394 		 * work too, which will end up happening right after
395 		 * our carrier on task work and will allow us to
396 		 * send out all of the non-broadcast joins
397 		 */
398 		if (mcast == priv->broadcast) {
399 			spin_lock_irq(&priv->lock);
400 			queue_work(priv->wq, &priv->carrier_on_task);
401 			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
402 			goto out_locked;
403 		}
404 	} else {
405 		bool silent_fail =
406 		    test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
407 		    status == -EINVAL;
408 
409 		if (mcast->logcount < 20) {
410 			if (status == -ETIMEDOUT || status == -EAGAIN ||
411 			    silent_fail) {
412 				ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
413 						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
414 						mcast->mcmember.mgid.raw, status);
415 			} else {
416 				ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
417 						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
418 					   mcast->mcmember.mgid.raw, status);
419 			}
420 
421 			if (!silent_fail)
422 				mcast->logcount++;
423 		}
424 
425 		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
426 		    mcast->backoff >= 2) {
427 			/*
428 			 * We only retry sendonly joins once before we drop
429 			 * the packet and quit trying to deal with the
430 			 * group.  However, we leave the group in the
431 			 * mcast list as an unjoined group.  If we want to
432 			 * try joining again, we simply queue up a packet
433 			 * and restart the join thread.  The empty queue
434 			 * is why the join thread ignores this group.
435 			 */
436 			mcast->backoff = 1;
437 			netif_tx_lock_bh(dev);
438 			while (!skb_queue_empty(&mcast->pkt_queue)) {
439 				++dev->stats.tx_dropped;
440 				dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
441 			}
442 			netif_tx_unlock_bh(dev);
443 		} else {
444 			spin_lock_irq(&priv->lock);
445 			/* Requeue this join task with a backoff delay */
446 			__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
447 			goto out_locked;
448 		}
449 	}
450 out:
451 	spin_lock_irq(&priv->lock);
452 out_locked:
453 	/*
454 	 * Make sure to set mcast->mc before we clear the busy flag to avoid
455 	 * racing with code that checks for BUSY before checking mcast->mc
456 	 */
457 	if (status)
458 		mcast->mc = NULL;
459 	else
460 		mcast->mc = multicast;
461 	clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
462 	spin_unlock_irq(&priv->lock);
463 	complete(&mcast->done);
464 
465 	return status;
466 }
467 
468 /*
469  * Caller must hold 'priv->lock'
470  */
471 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
472 {
473 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
474 	struct ib_sa_multicast *multicast;
475 	struct ib_sa_mcmember_rec rec = {
476 		.join_state = 1
477 	};
478 	ib_sa_comp_mask comp_mask;
479 	int ret = 0;
480 
481 	if (!priv->broadcast ||
482 	    !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
483 		return -EINVAL;
484 
485 	init_completion(&mcast->done);
486 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
487 
488 	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
489 
490 	rec.mgid     = mcast->mcmember.mgid;
491 	rec.port_gid = priv->local_gid;
492 	rec.pkey     = cpu_to_be16(priv->pkey);
493 
494 	comp_mask =
495 		IB_SA_MCMEMBER_REC_MGID		|
496 		IB_SA_MCMEMBER_REC_PORT_GID	|
497 		IB_SA_MCMEMBER_REC_PKEY		|
498 		IB_SA_MCMEMBER_REC_JOIN_STATE;
499 
500 	if (mcast != priv->broadcast) {
501 		/*
502 		 * RFC 4391:
503 		 *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
504 		 *  and HopLimit as those used in the broadcast-GID.  The rest
505 		 *  of attributes SHOULD follow the values used in the
506 		 *  broadcast-GID as well.
507 		 */
508 		comp_mask |=
509 			IB_SA_MCMEMBER_REC_QKEY			|
510 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
511 			IB_SA_MCMEMBER_REC_MTU			|
512 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
513 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
514 			IB_SA_MCMEMBER_REC_RATE			|
515 			IB_SA_MCMEMBER_REC_SL			|
516 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
517 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
518 
519 		rec.qkey	  = priv->broadcast->mcmember.qkey;
520 		rec.mtu_selector  = IB_SA_EQ;
521 		rec.mtu		  = priv->broadcast->mcmember.mtu;
522 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
523 		rec.rate_selector = IB_SA_EQ;
524 		rec.rate	  = priv->broadcast->mcmember.rate;
525 		rec.sl		  = priv->broadcast->mcmember.sl;
526 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
527 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
528 
529 		/*
530 		 * Send-only IB Multicast joins work at the core IB layer but
531 		 * require specific SM support.
532 		 * We can use such joins here only if the current SM supports that feature.
533 		 * However, if not, we emulate an Ethernet multicast send,
534 		 * which does not require a multicast subscription and will
535 		 * still send properly. The most appropriate thing to
536 		 * do is to create the group if it doesn't exist as that
537 		 * most closely emulates the behavior, from a user space
538 		 * application perspective, of Ethernet multicast operation.
539 		 */
540 		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
541 		    priv->sm_fullmember_sendonly_support)
542 			/* SM supports sendonly-fullmember, otherwise fallback to full-member */
543 			rec.join_state = SENDONLY_FULLMEMBER_JOIN;
544 	}
545 	spin_unlock_irq(&priv->lock);
546 
547 	multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
548 					 &rec, comp_mask, GFP_KERNEL,
549 					 ipoib_mcast_join_complete, mcast);
550 	spin_lock_irq(&priv->lock);
551 	if (IS_ERR(multicast)) {
552 		ret = PTR_ERR(multicast);
553 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
554 		/* Requeue this join task with a backoff delay */
555 		__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
556 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
557 		spin_unlock_irq(&priv->lock);
558 		complete(&mcast->done);
559 		spin_lock_irq(&priv->lock);
560 	}
561 	return 0;
562 }
563 
564 void ipoib_mcast_join_task(struct work_struct *work)
565 {
566 	struct ipoib_dev_priv *priv =
567 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
568 	struct net_device *dev = priv->dev;
569 	struct ib_port_attr port_attr;
570 	unsigned long delay_until = 0;
571 	struct ipoib_mcast *mcast = NULL;
572 
573 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
574 		return;
575 
576 	if (ib_query_port(priv->ca, priv->port, &port_attr)) {
577 		ipoib_dbg(priv, "ib_query_port() failed\n");
578 		return;
579 	}
580 	if (port_attr.state != IB_PORT_ACTIVE) {
581 		ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
582 			  port_attr.state);
583 		return;
584 	}
585 	priv->local_lid = port_attr.lid;
586 	netif_addr_lock_bh(dev);
587 
588 	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
589 		netif_addr_unlock_bh(dev);
590 		return;
591 	}
592 	netif_addr_unlock_bh(dev);
593 
594 	spin_lock_irq(&priv->lock);
595 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
596 		goto out;
597 
598 	if (!priv->broadcast) {
599 		struct ipoib_mcast *broadcast;
600 
601 		broadcast = ipoib_mcast_alloc(dev);
602 		if (!broadcast) {
603 			ipoib_warn(priv, "failed to allocate broadcast group\n");
604 			/*
605 			 * Restart us after a 1 second delay to retry
606 			 * creating our broadcast group and attaching to
607 			 * it.  Until this succeeds, this ipoib dev is
608 			 * completely stalled (multicast wise).
609 			 */
610 			__ipoib_mcast_schedule_join_thread(priv, NULL, 1);
611 			goto out;
612 		}
613 
614 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
615 		       sizeof (union ib_gid));
616 		priv->broadcast = broadcast;
617 
618 		__ipoib_mcast_add(dev, priv->broadcast);
619 	}
620 
621 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
622 		if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
623 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
624 			mcast = priv->broadcast;
625 			if (mcast->backoff > 1 &&
626 			    time_before(jiffies, mcast->delay_until)) {
627 				delay_until = mcast->delay_until;
628 				mcast = NULL;
629 			}
630 		}
631 		goto out;
632 	}
633 
634 	/*
635 	 * We'll never get here until the broadcast group is both allocated
636 	 * and attached
637 	 */
638 	list_for_each_entry(mcast, &priv->multicast_list, list) {
639 		if (IS_ERR_OR_NULL(mcast->mc) &&
640 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
641 		    (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
642 		     !skb_queue_empty(&mcast->pkt_queue))) {
643 			if (mcast->backoff == 1 ||
644 			    time_after_eq(jiffies, mcast->delay_until)) {
645 				/* Found the next unjoined group */
646 				if (ipoib_mcast_join(dev, mcast)) {
647 					spin_unlock_irq(&priv->lock);
648 					return;
649 				}
650 			} else if (!delay_until ||
651 				 time_before(mcast->delay_until, delay_until))
652 				delay_until = mcast->delay_until;
653 		}
654 	}
655 
656 	mcast = NULL;
657 	ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
658 
659 out:
660 	if (delay_until) {
661 		cancel_delayed_work(&priv->mcast_task);
662 		queue_delayed_work(priv->wq, &priv->mcast_task,
663 				   delay_until - jiffies);
664 	}
665 	if (mcast)
666 		ipoib_mcast_join(dev, mcast);
667 
668 	spin_unlock_irq(&priv->lock);
669 }
670 
671 void ipoib_mcast_start_thread(struct net_device *dev)
672 {
673 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
674 	unsigned long flags;
675 
676 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
677 
678 	spin_lock_irqsave(&priv->lock, flags);
679 	__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
680 	spin_unlock_irqrestore(&priv->lock, flags);
681 }
682 
683 void ipoib_mcast_stop_thread(struct net_device *dev)
684 {
685 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
686 
687 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
688 
689 	cancel_delayed_work_sync(&priv->mcast_task);
690 }
691 
692 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
693 {
694 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
695 	struct rdma_netdev *rn = netdev_priv(dev);
696 	int ret = 0;
697 
698 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
699 		ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
700 
701 	if (!IS_ERR_OR_NULL(mcast->mc))
702 		ib_sa_free_multicast(mcast->mc);
703 
704 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
705 		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
706 				mcast->mcmember.mgid.raw);
707 
708 		/* Remove ourselves from the multicast group */
709 		ret = rn->detach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
710 				       be16_to_cpu(mcast->mcmember.mlid));
711 		if (ret)
712 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
713 	} else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
714 		ipoib_dbg(priv, "leaving with no mcmember but not a "
715 			  "SENDONLY join\n");
716 
717 	return 0;
718 }
719 
720 /*
721  * Check if the multicast group is sendonly. If so remove it from the maps
722  * and add to the remove list
723  */
724 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
725 				struct list_head *remove_list)
726 {
727 	/* Is this multicast ? */
728 	if (*mgid == 0xff) {
729 		struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
730 
731 		if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
732 			list_del(&mcast->list);
733 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
734 			list_add_tail(&mcast->list, remove_list);
735 		}
736 	}
737 }
738 
739 void ipoib_mcast_remove_list(struct list_head *remove_list)
740 {
741 	struct ipoib_mcast *mcast, *tmcast;
742 
743 	/*
744 	 * make sure the in-flight joins have finished before we attempt
745 	 * to leave
746 	 */
747 	list_for_each_entry_safe(mcast, tmcast, remove_list, list)
748 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
749 			wait_for_completion(&mcast->done);
750 
751 	list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
752 		ipoib_mcast_leave(mcast->dev, mcast);
753 		ipoib_mcast_free(mcast);
754 	}
755 }
756 
757 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
758 {
759 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
760 	struct rdma_netdev *rn = netdev_priv(dev);
761 	struct ipoib_mcast *mcast;
762 	unsigned long flags;
763 	void *mgid = daddr + 4;
764 
765 	spin_lock_irqsave(&priv->lock, flags);
766 
767 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
768 	    !priv->broadcast					||
769 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
770 		++dev->stats.tx_dropped;
771 		dev_kfree_skb_any(skb);
772 		goto unlock;
773 	}
774 
775 	mcast = __ipoib_mcast_find(dev, mgid);
776 	if (!mcast || !mcast->ah) {
777 		if (!mcast) {
778 			/* Let's create a new send only group now */
779 			ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
780 					mgid);
781 
782 			mcast = ipoib_mcast_alloc(dev);
783 			if (!mcast) {
784 				ipoib_warn(priv, "unable to allocate memory "
785 					   "for multicast structure\n");
786 				++dev->stats.tx_dropped;
787 				dev_kfree_skb_any(skb);
788 				goto unlock;
789 			}
790 
791 			set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
792 			memcpy(mcast->mcmember.mgid.raw, mgid,
793 			       sizeof (union ib_gid));
794 			__ipoib_mcast_add(dev, mcast);
795 			list_add_tail(&mcast->list, &priv->multicast_list);
796 		}
797 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
798 			/* put pseudoheader back on for next time */
799 			skb_push(skb, sizeof(struct ipoib_pseudo_header));
800 			skb_queue_tail(&mcast->pkt_queue, skb);
801 		} else {
802 			++dev->stats.tx_dropped;
803 			dev_kfree_skb_any(skb);
804 		}
805 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
806 			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
807 		}
808 	} else {
809 		struct ipoib_neigh *neigh;
810 
811 		spin_unlock_irqrestore(&priv->lock, flags);
812 		neigh = ipoib_neigh_get(dev, daddr);
813 		spin_lock_irqsave(&priv->lock, flags);
814 		if (!neigh) {
815 			neigh = ipoib_neigh_alloc(daddr, dev);
816 			/* Make sure that the neigh will be added only
817 			 * once to mcast list.
818 			 */
819 			if (neigh && list_empty(&neigh->list)) {
820 				kref_get(&mcast->ah->ref);
821 				neigh->ah	= mcast->ah;
822 				neigh->ah->valid = 1;
823 				list_add_tail(&neigh->list, &mcast->neigh_list);
824 			}
825 		}
826 		spin_unlock_irqrestore(&priv->lock, flags);
827 		mcast->ah->last_send = rn->send(dev, skb, mcast->ah->ah,
828 						IB_MULTICAST_QPN);
829 		if (neigh)
830 			ipoib_neigh_put(neigh);
831 		return;
832 	}
833 
834 unlock:
835 	spin_unlock_irqrestore(&priv->lock, flags);
836 }
837 
838 void ipoib_mcast_dev_flush(struct net_device *dev)
839 {
840 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
841 	LIST_HEAD(remove_list);
842 	struct ipoib_mcast *mcast, *tmcast;
843 	unsigned long flags;
844 
845 	mutex_lock(&priv->mcast_mutex);
846 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
847 
848 	spin_lock_irqsave(&priv->lock, flags);
849 
850 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
851 		list_del(&mcast->list);
852 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
853 		list_add_tail(&mcast->list, &remove_list);
854 	}
855 
856 	if (priv->broadcast) {
857 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
858 		list_add_tail(&priv->broadcast->list, &remove_list);
859 		priv->broadcast = NULL;
860 	}
861 
862 	spin_unlock_irqrestore(&priv->lock, flags);
863 
864 	ipoib_mcast_remove_list(&remove_list);
865 	mutex_unlock(&priv->mcast_mutex);
866 }
867 
868 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
869 {
870 	/* reserved QPN, prefix, scope */
871 	if (memcmp(addr, broadcast, 6))
872 		return 0;
873 	/* signature lower, pkey */
874 	if (memcmp(addr + 7, broadcast + 7, 3))
875 		return 0;
876 	return 1;
877 }
878 
879 void ipoib_mcast_restart_task(struct work_struct *work)
880 {
881 	struct ipoib_dev_priv *priv =
882 		container_of(work, struct ipoib_dev_priv, restart_task);
883 	struct net_device *dev = priv->dev;
884 	struct netdev_hw_addr *ha;
885 	struct ipoib_mcast *mcast, *tmcast;
886 	LIST_HEAD(remove_list);
887 	struct ib_sa_mcmember_rec rec;
888 
889 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
890 		/*
891 		 * shortcut...on shutdown flush is called next, just
892 		 * let it do all the work
893 		 */
894 		return;
895 
896 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
897 
898 	netif_addr_lock_bh(dev);
899 	spin_lock_irq(&priv->lock);
900 
901 	/*
902 	 * Unfortunately, the networking core only gives us a list of all of
903 	 * the multicast hardware addresses. We need to figure out which ones
904 	 * are new and which ones have been removed
905 	 */
906 
907 	/* Clear out the found flag */
908 	list_for_each_entry(mcast, &priv->multicast_list, list)
909 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
910 
911 	/* Mark all of the entries that are found or don't exist */
912 	netdev_for_each_mc_addr(ha, dev) {
913 		union ib_gid mgid;
914 
915 		if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
916 			continue;
917 
918 		memcpy(mgid.raw, ha->addr + 4, sizeof(mgid));
919 
920 		mcast = __ipoib_mcast_find(dev, &mgid);
921 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
922 			struct ipoib_mcast *nmcast;
923 
924 			/* ignore group which is directly joined by userspace */
925 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
926 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
927 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
928 						mgid.raw);
929 				continue;
930 			}
931 
932 			/* Not found or send-only group, let's add a new entry */
933 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
934 					mgid.raw);
935 
936 			nmcast = ipoib_mcast_alloc(dev);
937 			if (!nmcast) {
938 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
939 				continue;
940 			}
941 
942 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
943 
944 			nmcast->mcmember.mgid = mgid;
945 
946 			if (mcast) {
947 				/* Destroy the send only entry */
948 				list_move_tail(&mcast->list, &remove_list);
949 
950 				rb_replace_node(&mcast->rb_node,
951 						&nmcast->rb_node,
952 						&priv->multicast_tree);
953 			} else
954 				__ipoib_mcast_add(dev, nmcast);
955 
956 			list_add_tail(&nmcast->list, &priv->multicast_list);
957 		}
958 
959 		if (mcast)
960 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
961 	}
962 
963 	/* Remove all of the entries don't exist anymore */
964 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
965 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
966 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
967 			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
968 					mcast->mcmember.mgid.raw);
969 
970 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
971 
972 			/* Move to the remove list */
973 			list_move_tail(&mcast->list, &remove_list);
974 		}
975 	}
976 
977 	spin_unlock_irq(&priv->lock);
978 	netif_addr_unlock_bh(dev);
979 
980 	ipoib_mcast_remove_list(&remove_list);
981 
982 	/*
983 	 * Double check that we are still up
984 	 */
985 	if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
986 		spin_lock_irq(&priv->lock);
987 		__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
988 		spin_unlock_irq(&priv->lock);
989 	}
990 }
991 
992 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
993 
994 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
995 {
996 	struct ipoib_mcast_iter *iter;
997 
998 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
999 	if (!iter)
1000 		return NULL;
1001 
1002 	iter->dev = dev;
1003 	memset(iter->mgid.raw, 0, 16);
1004 
1005 	if (ipoib_mcast_iter_next(iter)) {
1006 		kfree(iter);
1007 		return NULL;
1008 	}
1009 
1010 	return iter;
1011 }
1012 
1013 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1014 {
1015 	struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
1016 	struct rb_node *n;
1017 	struct ipoib_mcast *mcast;
1018 	int ret = 1;
1019 
1020 	spin_lock_irq(&priv->lock);
1021 
1022 	n = rb_first(&priv->multicast_tree);
1023 
1024 	while (n) {
1025 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1026 
1027 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1028 			   sizeof (union ib_gid)) < 0) {
1029 			iter->mgid      = mcast->mcmember.mgid;
1030 			iter->created   = mcast->created;
1031 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1032 			iter->complete  = !!mcast->ah;
1033 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1034 
1035 			ret = 0;
1036 
1037 			break;
1038 		}
1039 
1040 		n = rb_next(n);
1041 	}
1042 
1043 	spin_unlock_irq(&priv->lock);
1044 
1045 	return ret;
1046 }
1047 
1048 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1049 			   union ib_gid *mgid,
1050 			   unsigned long *created,
1051 			   unsigned int *queuelen,
1052 			   unsigned int *complete,
1053 			   unsigned int *send_only)
1054 {
1055 	*mgid      = iter->mgid;
1056 	*created   = iter->created;
1057 	*queuelen  = iter->queuelen;
1058 	*complete  = iter->complete;
1059 	*send_only = iter->send_only;
1060 }
1061 
1062 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
1063