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 /*
68  * This should be called with the priv->lock held
69  */
70 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
71 					       struct ipoib_mcast *mcast,
72 					       bool delay)
73 {
74 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
75 		return;
76 
77 	/*
78 	 * We will be scheduling *something*, so cancel whatever is
79 	 * currently scheduled first
80 	 */
81 	cancel_delayed_work(&priv->mcast_task);
82 	if (mcast && delay) {
83 		/*
84 		 * We had a failure and want to schedule a retry later
85 		 */
86 		mcast->backoff *= 2;
87 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
88 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
89 		mcast->delay_until = jiffies + (mcast->backoff * HZ);
90 		/*
91 		 * Mark this mcast for its delay, but restart the
92 		 * task immediately.  The join task will make sure to
93 		 * clear out all entries without delays, and then
94 		 * schedule itself to run again when the earliest
95 		 * delay expires
96 		 */
97 		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
98 	} else if (delay) {
99 		/*
100 		 * Special case of retrying after a failure to
101 		 * allocate the broadcast multicast group, wait
102 		 * 1 second and try again
103 		 */
104 		queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
105 	} else
106 		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
107 }
108 
109 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
110 {
111 	struct net_device *dev = mcast->dev;
112 	int tx_dropped = 0;
113 
114 	ipoib_dbg_mcast(netdev_priv(dev), "deleting multicast group %pI6\n",
115 			mcast->mcmember.mgid.raw);
116 
117 	/* remove all neigh connected to this mcast */
118 	ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
119 
120 	if (mcast->ah)
121 		ipoib_put_ah(mcast->ah);
122 
123 	while (!skb_queue_empty(&mcast->pkt_queue)) {
124 		++tx_dropped;
125 		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
126 	}
127 
128 	netif_tx_lock_bh(dev);
129 	dev->stats.tx_dropped += tx_dropped;
130 	netif_tx_unlock_bh(dev);
131 
132 	kfree(mcast);
133 }
134 
135 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
136 					     int can_sleep)
137 {
138 	struct ipoib_mcast *mcast;
139 
140 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
141 	if (!mcast)
142 		return NULL;
143 
144 	mcast->dev = dev;
145 	mcast->created = jiffies;
146 	mcast->delay_until = jiffies;
147 	mcast->backoff = 1;
148 
149 	INIT_LIST_HEAD(&mcast->list);
150 	INIT_LIST_HEAD(&mcast->neigh_list);
151 	skb_queue_head_init(&mcast->pkt_queue);
152 
153 	return mcast;
154 }
155 
156 struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
157 {
158 	struct ipoib_dev_priv *priv = netdev_priv(dev);
159 	struct rb_node *n = priv->multicast_tree.rb_node;
160 
161 	while (n) {
162 		struct ipoib_mcast *mcast;
163 		int ret;
164 
165 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
166 
167 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
168 			     sizeof (union ib_gid));
169 		if (ret < 0)
170 			n = n->rb_left;
171 		else if (ret > 0)
172 			n = n->rb_right;
173 		else
174 			return mcast;
175 	}
176 
177 	return NULL;
178 }
179 
180 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
181 {
182 	struct ipoib_dev_priv *priv = netdev_priv(dev);
183 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
184 
185 	while (*n) {
186 		struct ipoib_mcast *tmcast;
187 		int ret;
188 
189 		pn = *n;
190 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
191 
192 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
193 			     sizeof (union ib_gid));
194 		if (ret < 0)
195 			n = &pn->rb_left;
196 		else if (ret > 0)
197 			n = &pn->rb_right;
198 		else
199 			return -EEXIST;
200 	}
201 
202 	rb_link_node(&mcast->rb_node, pn, n);
203 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
204 
205 	return 0;
206 }
207 
208 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
209 				   struct ib_sa_mcmember_rec *mcmember)
210 {
211 	struct net_device *dev = mcast->dev;
212 	struct ipoib_dev_priv *priv = netdev_priv(dev);
213 	struct ipoib_ah *ah;
214 	int ret;
215 	int set_qkey = 0;
216 
217 	mcast->mcmember = *mcmember;
218 
219 	/* Set the multicast MTU and cached Q_Key before we attach if it's
220 	 * the broadcast group.
221 	 */
222 	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
223 		    sizeof (union ib_gid))) {
224 		spin_lock_irq(&priv->lock);
225 		if (!priv->broadcast) {
226 			spin_unlock_irq(&priv->lock);
227 			return -EAGAIN;
228 		}
229 		/*update priv member according to the new mcast*/
230 		priv->broadcast->mcmember.qkey = mcmember->qkey;
231 		priv->broadcast->mcmember.mtu = mcmember->mtu;
232 		priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
233 		priv->broadcast->mcmember.rate = mcmember->rate;
234 		priv->broadcast->mcmember.sl = mcmember->sl;
235 		priv->broadcast->mcmember.flow_label = mcmember->flow_label;
236 		priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
237 		/* assume if the admin and the mcast are the same both can be changed */
238 		if (priv->mcast_mtu == priv->admin_mtu)
239 			priv->admin_mtu =
240 			priv->mcast_mtu =
241 			IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
242 		else
243 			priv->mcast_mtu =
244 			IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
245 
246 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
247 		spin_unlock_irq(&priv->lock);
248 		priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
249 		set_qkey = 1;
250 	}
251 
252 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
253 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
254 			ipoib_warn(priv, "multicast group %pI6 already attached\n",
255 				   mcast->mcmember.mgid.raw);
256 
257 			return 0;
258 		}
259 
260 		ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
261 					 &mcast->mcmember.mgid, set_qkey);
262 		if (ret < 0) {
263 			ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
264 				   mcast->mcmember.mgid.raw);
265 
266 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
267 			return ret;
268 		}
269 	}
270 
271 	{
272 		struct ib_ah_attr av = {
273 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
274 			.port_num      = priv->port,
275 			.sl	       = mcast->mcmember.sl,
276 			.ah_flags      = IB_AH_GRH,
277 			.static_rate   = mcast->mcmember.rate,
278 			.grh	       = {
279 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
280 				.hop_limit     = mcast->mcmember.hop_limit,
281 				.sgid_index    = 0,
282 				.traffic_class = mcast->mcmember.traffic_class
283 			}
284 		};
285 		av.grh.dgid = mcast->mcmember.mgid;
286 
287 		ah = ipoib_create_ah(dev, priv->pd, &av);
288 		if (IS_ERR(ah)) {
289 			ipoib_warn(priv, "ib_address_create failed %ld\n",
290 				-PTR_ERR(ah));
291 			/* use original error */
292 			return PTR_ERR(ah);
293 		} else {
294 			spin_lock_irq(&priv->lock);
295 			mcast->ah = ah;
296 			spin_unlock_irq(&priv->lock);
297 
298 			ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
299 					mcast->mcmember.mgid.raw,
300 					mcast->ah->ah,
301 					be16_to_cpu(mcast->mcmember.mlid),
302 					mcast->mcmember.sl);
303 		}
304 	}
305 
306 	/* actually send any queued packets */
307 	netif_tx_lock_bh(dev);
308 	while (!skb_queue_empty(&mcast->pkt_queue)) {
309 		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
310 
311 		netif_tx_unlock_bh(dev);
312 
313 		skb->dev = dev;
314 		if (dev_queue_xmit(skb))
315 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
316 
317 		netif_tx_lock_bh(dev);
318 	}
319 	netif_tx_unlock_bh(dev);
320 
321 	return 0;
322 }
323 
324 void ipoib_mcast_carrier_on_task(struct work_struct *work)
325 {
326 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
327 						   carrier_on_task);
328 	struct ib_port_attr attr;
329 
330 	if (ib_query_port(priv->ca, priv->port, &attr) ||
331 	    attr.state != IB_PORT_ACTIVE) {
332 		ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
333 		return;
334 	}
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 = netdev_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 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
460 {
461 	struct ipoib_dev_priv *priv = netdev_priv(dev);
462 	struct ib_sa_multicast *multicast;
463 	struct ib_sa_mcmember_rec rec = {
464 		.join_state = 1
465 	};
466 	ib_sa_comp_mask comp_mask;
467 	int ret = 0;
468 
469 	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
470 
471 	rec.mgid     = mcast->mcmember.mgid;
472 	rec.port_gid = priv->local_gid;
473 	rec.pkey     = cpu_to_be16(priv->pkey);
474 
475 	comp_mask =
476 		IB_SA_MCMEMBER_REC_MGID		|
477 		IB_SA_MCMEMBER_REC_PORT_GID	|
478 		IB_SA_MCMEMBER_REC_PKEY		|
479 		IB_SA_MCMEMBER_REC_JOIN_STATE;
480 
481 	if (mcast != priv->broadcast) {
482 		/*
483 		 * RFC 4391:
484 		 *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
485 		 *  and HopLimit as those used in the broadcast-GID.  The rest
486 		 *  of attributes SHOULD follow the values used in the
487 		 *  broadcast-GID as well.
488 		 */
489 		comp_mask |=
490 			IB_SA_MCMEMBER_REC_QKEY			|
491 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
492 			IB_SA_MCMEMBER_REC_MTU			|
493 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
494 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
495 			IB_SA_MCMEMBER_REC_RATE			|
496 			IB_SA_MCMEMBER_REC_SL			|
497 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
498 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
499 
500 		rec.qkey	  = priv->broadcast->mcmember.qkey;
501 		rec.mtu_selector  = IB_SA_EQ;
502 		rec.mtu		  = priv->broadcast->mcmember.mtu;
503 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
504 		rec.rate_selector = IB_SA_EQ;
505 		rec.rate	  = priv->broadcast->mcmember.rate;
506 		rec.sl		  = priv->broadcast->mcmember.sl;
507 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
508 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
509 
510 		/*
511 		 * Send-only IB Multicast joins do not work at the core
512 		 * IB layer yet, so we can't use them here.  However,
513 		 * we are emulating an Ethernet multicast send, which
514 		 * does not require a multicast subscription and will
515 		 * still send properly.  The most appropriate thing to
516 		 * do is to create the group if it doesn't exist as that
517 		 * most closely emulates the behavior, from a user space
518 		 * application perspecitive, of Ethernet multicast
519 		 * operation.  For now, we do a full join, maybe later
520 		 * when the core IB layers support send only joins we
521 		 * will use them.
522 		 */
523 #if 0
524 		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
525 			rec.join_state = 4;
526 #endif
527 	}
528 
529 	multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
530 					 &rec, comp_mask, GFP_KERNEL,
531 					 ipoib_mcast_join_complete, mcast);
532 	if (IS_ERR(multicast)) {
533 		ret = PTR_ERR(multicast);
534 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
535 		spin_lock_irq(&priv->lock);
536 		/* Requeue this join task with a backoff delay */
537 		__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
538 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
539 		spin_unlock_irq(&priv->lock);
540 		complete(&mcast->done);
541 	}
542 }
543 
544 void ipoib_mcast_join_task(struct work_struct *work)
545 {
546 	struct ipoib_dev_priv *priv =
547 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
548 	struct net_device *dev = priv->dev;
549 	struct ib_port_attr port_attr;
550 	unsigned long delay_until = 0;
551 	struct ipoib_mcast *mcast = NULL;
552 
553 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
554 		return;
555 
556 	if (ib_query_port(priv->ca, priv->port, &port_attr) ||
557 	    port_attr.state != IB_PORT_ACTIVE) {
558 		ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
559 			  port_attr.state);
560 		return;
561 	}
562 	priv->local_lid = port_attr.lid;
563 
564 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
565 		ipoib_warn(priv, "ib_query_gid() failed\n");
566 	else
567 		memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
568 
569 	spin_lock_irq(&priv->lock);
570 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
571 		goto out;
572 
573 	if (!priv->broadcast) {
574 		struct ipoib_mcast *broadcast;
575 
576 		broadcast = ipoib_mcast_alloc(dev, 0);
577 		if (!broadcast) {
578 			ipoib_warn(priv, "failed to allocate broadcast group\n");
579 			/*
580 			 * Restart us after a 1 second delay to retry
581 			 * creating our broadcast group and attaching to
582 			 * it.  Until this succeeds, this ipoib dev is
583 			 * completely stalled (multicast wise).
584 			 */
585 			__ipoib_mcast_schedule_join_thread(priv, NULL, 1);
586 			goto out;
587 		}
588 
589 		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
590 		       sizeof (union ib_gid));
591 		priv->broadcast = broadcast;
592 
593 		__ipoib_mcast_add(dev, priv->broadcast);
594 	}
595 
596 	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
597 		if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
598 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
599 			mcast = priv->broadcast;
600 			if (mcast->backoff > 1 &&
601 			    time_before(jiffies, mcast->delay_until)) {
602 				delay_until = mcast->delay_until;
603 				mcast = NULL;
604 			}
605 		}
606 		goto out;
607 	}
608 
609 	/*
610 	 * We'll never get here until the broadcast group is both allocated
611 	 * and attached
612 	 */
613 	list_for_each_entry(mcast, &priv->multicast_list, list) {
614 		if (IS_ERR_OR_NULL(mcast->mc) &&
615 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
616 		    (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
617 		     !skb_queue_empty(&mcast->pkt_queue))) {
618 			if (mcast->backoff == 1 ||
619 			    time_after_eq(jiffies, mcast->delay_until)) {
620 				/* Found the next unjoined group */
621 				init_completion(&mcast->done);
622 				set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
623 				spin_unlock_irq(&priv->lock);
624 				ipoib_mcast_join(dev, mcast);
625 				spin_lock_irq(&priv->lock);
626 			} else if (!delay_until ||
627 				 time_before(mcast->delay_until, delay_until))
628 				delay_until = mcast->delay_until;
629 		}
630 	}
631 
632 	mcast = NULL;
633 	ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
634 
635 out:
636 	if (delay_until) {
637 		cancel_delayed_work(&priv->mcast_task);
638 		queue_delayed_work(priv->wq, &priv->mcast_task,
639 				   delay_until - jiffies);
640 	}
641 	if (mcast) {
642 		init_completion(&mcast->done);
643 		set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
644 	}
645 	spin_unlock_irq(&priv->lock);
646 	if (mcast)
647 		ipoib_mcast_join(dev, mcast);
648 }
649 
650 int ipoib_mcast_start_thread(struct net_device *dev)
651 {
652 	struct ipoib_dev_priv *priv = netdev_priv(dev);
653 	unsigned long flags;
654 
655 	ipoib_dbg_mcast(priv, "starting multicast thread\n");
656 
657 	spin_lock_irqsave(&priv->lock, flags);
658 	__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
659 	spin_unlock_irqrestore(&priv->lock, flags);
660 
661 	return 0;
662 }
663 
664 int ipoib_mcast_stop_thread(struct net_device *dev)
665 {
666 	struct ipoib_dev_priv *priv = netdev_priv(dev);
667 	unsigned long flags;
668 
669 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
670 
671 	spin_lock_irqsave(&priv->lock, flags);
672 	cancel_delayed_work(&priv->mcast_task);
673 	spin_unlock_irqrestore(&priv->lock, flags);
674 
675 	flush_workqueue(priv->wq);
676 
677 	return 0;
678 }
679 
680 int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
681 {
682 	struct ipoib_dev_priv *priv = netdev_priv(dev);
683 	int ret = 0;
684 
685 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
686 		ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
687 
688 	if (!IS_ERR_OR_NULL(mcast->mc))
689 		ib_sa_free_multicast(mcast->mc);
690 
691 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
692 		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
693 				mcast->mcmember.mgid.raw);
694 
695 		/* Remove ourselves from the multicast group */
696 		ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
697 				      be16_to_cpu(mcast->mcmember.mlid));
698 		if (ret)
699 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
700 	} else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
701 		ipoib_dbg(priv, "leaving with no mcmember but not a "
702 			  "SENDONLY join\n");
703 
704 	return 0;
705 }
706 
707 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
708 {
709 	struct ipoib_dev_priv *priv = netdev_priv(dev);
710 	struct ipoib_mcast *mcast;
711 	unsigned long flags;
712 	void *mgid = daddr + 4;
713 
714 	spin_lock_irqsave(&priv->lock, flags);
715 
716 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
717 	    !priv->broadcast					||
718 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
719 		++dev->stats.tx_dropped;
720 		dev_kfree_skb_any(skb);
721 		goto unlock;
722 	}
723 
724 	mcast = __ipoib_mcast_find(dev, mgid);
725 	if (!mcast || !mcast->ah) {
726 		if (!mcast) {
727 			/* Let's create a new send only group now */
728 			ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
729 					mgid);
730 
731 			mcast = ipoib_mcast_alloc(dev, 0);
732 			if (!mcast) {
733 				ipoib_warn(priv, "unable to allocate memory "
734 					   "for multicast structure\n");
735 				++dev->stats.tx_dropped;
736 				dev_kfree_skb_any(skb);
737 				goto unlock;
738 			}
739 
740 			set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
741 			memcpy(mcast->mcmember.mgid.raw, mgid,
742 			       sizeof (union ib_gid));
743 			__ipoib_mcast_add(dev, mcast);
744 			list_add_tail(&mcast->list, &priv->multicast_list);
745 		}
746 		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
747 			skb_queue_tail(&mcast->pkt_queue, skb);
748 		else {
749 			++dev->stats.tx_dropped;
750 			dev_kfree_skb_any(skb);
751 		}
752 		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
753 			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
754 		}
755 	} else {
756 		struct ipoib_neigh *neigh;
757 
758 		spin_unlock_irqrestore(&priv->lock, flags);
759 		neigh = ipoib_neigh_get(dev, daddr);
760 		spin_lock_irqsave(&priv->lock, flags);
761 		if (!neigh) {
762 			neigh = ipoib_neigh_alloc(daddr, dev);
763 			if (neigh) {
764 				kref_get(&mcast->ah->ref);
765 				neigh->ah	= mcast->ah;
766 				list_add_tail(&neigh->list, &mcast->neigh_list);
767 			}
768 		}
769 		spin_unlock_irqrestore(&priv->lock, flags);
770 		ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
771 		if (neigh)
772 			ipoib_neigh_put(neigh);
773 		return;
774 	}
775 
776 unlock:
777 	spin_unlock_irqrestore(&priv->lock, flags);
778 }
779 
780 void ipoib_mcast_dev_flush(struct net_device *dev)
781 {
782 	struct ipoib_dev_priv *priv = netdev_priv(dev);
783 	LIST_HEAD(remove_list);
784 	struct ipoib_mcast *mcast, *tmcast;
785 	unsigned long flags;
786 
787 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
788 
789 	spin_lock_irqsave(&priv->lock, flags);
790 
791 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
792 		list_del(&mcast->list);
793 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
794 		list_add_tail(&mcast->list, &remove_list);
795 	}
796 
797 	if (priv->broadcast) {
798 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
799 		list_add_tail(&priv->broadcast->list, &remove_list);
800 		priv->broadcast = NULL;
801 	}
802 
803 	spin_unlock_irqrestore(&priv->lock, flags);
804 
805 	/*
806 	 * make sure the in-flight joins have finished before we attempt
807 	 * to leave
808 	 */
809 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
810 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
811 			wait_for_completion(&mcast->done);
812 
813 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
814 		ipoib_mcast_leave(dev, mcast);
815 		ipoib_mcast_free(mcast);
816 	}
817 }
818 
819 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
820 {
821 	/* reserved QPN, prefix, scope */
822 	if (memcmp(addr, broadcast, 6))
823 		return 0;
824 	/* signature lower, pkey */
825 	if (memcmp(addr + 7, broadcast + 7, 3))
826 		return 0;
827 	return 1;
828 }
829 
830 void ipoib_mcast_restart_task(struct work_struct *work)
831 {
832 	struct ipoib_dev_priv *priv =
833 		container_of(work, struct ipoib_dev_priv, restart_task);
834 	struct net_device *dev = priv->dev;
835 	struct netdev_hw_addr *ha;
836 	struct ipoib_mcast *mcast, *tmcast;
837 	LIST_HEAD(remove_list);
838 	unsigned long flags;
839 	struct ib_sa_mcmember_rec rec;
840 
841 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
842 		/*
843 		 * shortcut...on shutdown flush is called next, just
844 		 * let it do all the work
845 		 */
846 		return;
847 
848 	ipoib_dbg_mcast(priv, "restarting multicast task\n");
849 
850 	local_irq_save(flags);
851 	netif_addr_lock(dev);
852 	spin_lock(&priv->lock);
853 
854 	/*
855 	 * Unfortunately, the networking core only gives us a list of all of
856 	 * the multicast hardware addresses. We need to figure out which ones
857 	 * are new and which ones have been removed
858 	 */
859 
860 	/* Clear out the found flag */
861 	list_for_each_entry(mcast, &priv->multicast_list, list)
862 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
863 
864 	/* Mark all of the entries that are found or don't exist */
865 	netdev_for_each_mc_addr(ha, dev) {
866 		union ib_gid mgid;
867 
868 		if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
869 			continue;
870 
871 		memcpy(mgid.raw, ha->addr + 4, sizeof mgid);
872 
873 		mcast = __ipoib_mcast_find(dev, &mgid);
874 		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
875 			struct ipoib_mcast *nmcast;
876 
877 			/* ignore group which is directly joined by userspace */
878 			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
879 			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
880 				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
881 						mgid.raw);
882 				continue;
883 			}
884 
885 			/* Not found or send-only group, let's add a new entry */
886 			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
887 					mgid.raw);
888 
889 			nmcast = ipoib_mcast_alloc(dev, 0);
890 			if (!nmcast) {
891 				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
892 				continue;
893 			}
894 
895 			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
896 
897 			nmcast->mcmember.mgid = mgid;
898 
899 			if (mcast) {
900 				/* Destroy the send only entry */
901 				list_move_tail(&mcast->list, &remove_list);
902 
903 				rb_replace_node(&mcast->rb_node,
904 						&nmcast->rb_node,
905 						&priv->multicast_tree);
906 			} else
907 				__ipoib_mcast_add(dev, nmcast);
908 
909 			list_add_tail(&nmcast->list, &priv->multicast_list);
910 		}
911 
912 		if (mcast)
913 			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
914 	}
915 
916 	/* Remove all of the entries don't exist anymore */
917 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
918 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
919 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
920 			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
921 					mcast->mcmember.mgid.raw);
922 
923 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
924 
925 			/* Move to the remove list */
926 			list_move_tail(&mcast->list, &remove_list);
927 		}
928 	}
929 
930 	spin_unlock(&priv->lock);
931 	netif_addr_unlock(dev);
932 	local_irq_restore(flags);
933 
934 	/*
935 	 * make sure the in-flight joins have finished before we attempt
936 	 * to leave
937 	 */
938 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
939 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
940 			wait_for_completion(&mcast->done);
941 
942 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
943 		ipoib_mcast_leave(mcast->dev, mcast);
944 		ipoib_mcast_free(mcast);
945 	}
946 
947 	/*
948 	 * Double check that we are still up
949 	 */
950 	if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
951 		spin_lock_irqsave(&priv->lock, flags);
952 		__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
953 		spin_unlock_irqrestore(&priv->lock, flags);
954 	}
955 }
956 
957 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
958 
959 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
960 {
961 	struct ipoib_mcast_iter *iter;
962 
963 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
964 	if (!iter)
965 		return NULL;
966 
967 	iter->dev = dev;
968 	memset(iter->mgid.raw, 0, 16);
969 
970 	if (ipoib_mcast_iter_next(iter)) {
971 		kfree(iter);
972 		return NULL;
973 	}
974 
975 	return iter;
976 }
977 
978 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
979 {
980 	struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
981 	struct rb_node *n;
982 	struct ipoib_mcast *mcast;
983 	int ret = 1;
984 
985 	spin_lock_irq(&priv->lock);
986 
987 	n = rb_first(&priv->multicast_tree);
988 
989 	while (n) {
990 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
991 
992 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
993 			   sizeof (union ib_gid)) < 0) {
994 			iter->mgid      = mcast->mcmember.mgid;
995 			iter->created   = mcast->created;
996 			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
997 			iter->complete  = !!mcast->ah;
998 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
999 
1000 			ret = 0;
1001 
1002 			break;
1003 		}
1004 
1005 		n = rb_next(n);
1006 	}
1007 
1008 	spin_unlock_irq(&priv->lock);
1009 
1010 	return ret;
1011 }
1012 
1013 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1014 			   union ib_gid *mgid,
1015 			   unsigned long *created,
1016 			   unsigned int *queuelen,
1017 			   unsigned int *complete,
1018 			   unsigned int *send_only)
1019 {
1020 	*mgid      = iter->mgid;
1021 	*created   = iter->created;
1022 	*queuelen  = iter->queuelen;
1023 	*complete  = iter->complete;
1024 	*send_only = iter->send_only;
1025 }
1026 
1027 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
1028