xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_multicast.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
3  *
4  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  */
36 
37 #include <sys/cdefs.h>
38 #include "ipoib.h"
39 
40 #include <linux/delay.h>
41 #include <linux/completion.h>
42 
43 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
44 static int mcast_debug_level = 1;
45 
46 module_param(mcast_debug_level, int, 0644);
47 MODULE_PARM_DESC(mcast_debug_level,
48 		 "Enable multicast debug tracing if > 0");
49 #endif
50 
51 static DEFINE_MUTEX(mcast_mutex);
52 
53 struct ipoib_mcast_iter {
54 	struct ipoib_dev_priv *priv;
55 	union ib_gid       mgid;
56 	unsigned long      created;
57 	unsigned int       queuelen;
58 	unsigned int       complete;
59 	unsigned int       send_only;
60 };
61 
62 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
63 {
64 	if_t dev = mcast->priv->dev;
65 	int tx_dropped = 0;
66 
67 	ipoib_dbg_mcast(mcast->priv, "deleting multicast group %16D\n",
68 			mcast->mcmember.mgid.raw, ":");
69 
70 	if (mcast->ah)
71 		ipoib_put_ah(mcast->ah);
72 
73 	tx_dropped = mcast->pkt_queue.ifq_len;
74 	_IF_DRAIN(&mcast->pkt_queue);	/* XXX Locking. */
75 
76 	if_inc_counter(dev, IFCOUNTER_OERRORS, tx_dropped);
77 
78 	kfree(mcast);
79 }
80 
81 static struct ipoib_mcast *ipoib_mcast_alloc(struct ipoib_dev_priv *priv,
82 					     int can_sleep)
83 {
84 	struct ipoib_mcast *mcast;
85 
86 	mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
87 	if (!mcast)
88 		return NULL;
89 
90 	mcast->priv = priv;
91 	mcast->created = jiffies;
92 	mcast->backoff = 1;
93 
94 	INIT_LIST_HEAD(&mcast->list);
95 	bzero(&mcast->pkt_queue, sizeof(mcast->pkt_queue));
96 
97 	return mcast;
98 }
99 
100 static struct ipoib_mcast *__ipoib_mcast_find(struct ipoib_dev_priv *priv,
101     void *mgid)
102 {
103 	struct rb_node *n = priv->multicast_tree.rb_node;
104 
105 	while (n) {
106 		struct ipoib_mcast *mcast;
107 		int ret;
108 
109 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
110 
111 		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
112 			     sizeof (union ib_gid));
113 		if (ret < 0)
114 			n = n->rb_left;
115 		else if (ret > 0)
116 			n = n->rb_right;
117 		else
118 			return mcast;
119 	}
120 
121 	return NULL;
122 }
123 
124 static int __ipoib_mcast_add(struct ipoib_dev_priv *priv,
125     struct ipoib_mcast *mcast)
126 {
127 	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
128 
129 	while (*n) {
130 		struct ipoib_mcast *tmcast;
131 		int ret;
132 
133 		pn = *n;
134 		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
135 
136 		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
137 			     sizeof (union ib_gid));
138 		if (ret < 0)
139 			n = &pn->rb_left;
140 		else if (ret > 0)
141 			n = &pn->rb_right;
142 		else
143 			return -EEXIST;
144 	}
145 
146 	rb_link_node(&mcast->rb_node, pn, n);
147 	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
148 
149 	return 0;
150 }
151 
152 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
153 				   struct ib_sa_mcmember_rec *mcmember)
154 {
155 	struct ipoib_dev_priv *priv = mcast->priv;
156 	if_t dev = priv->dev;
157 	struct ipoib_ah *ah;
158 	struct epoch_tracker et;
159 	int ret;
160 	int set_qkey = 0;
161 
162 	mcast->mcmember = *mcmember;
163 
164 	/* Set the cached Q_Key before we attach if it's the broadcast group */
165 	if (!memcmp(mcast->mcmember.mgid.raw, if_getbroadcastaddr(dev) + 4,
166 		    sizeof (union ib_gid))) {
167 		spin_lock_irq(&priv->lock);
168 		if (!priv->broadcast) {
169 			spin_unlock_irq(&priv->lock);
170 			return -EAGAIN;
171 		}
172 		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
173 		spin_unlock_irq(&priv->lock);
174 		priv->tx_wr.remote_qkey = priv->qkey;
175 		set_qkey = 1;
176 	}
177 
178 	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
179 		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
180 			ipoib_warn(priv, "multicast group %16D already attached\n",
181 				   mcast->mcmember.mgid.raw, ":");
182 
183 			return 0;
184 		}
185 
186 		ret = ipoib_mcast_attach(priv, be16_to_cpu(mcast->mcmember.mlid),
187 					 &mcast->mcmember.mgid, set_qkey);
188 		if (ret < 0) {
189 			ipoib_warn(priv, "couldn't attach QP to multicast group %16D\n",
190 				   mcast->mcmember.mgid.raw, ":");
191 
192 			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
193 			return ret;
194 		}
195 	}
196 
197 	{
198 		struct ib_ah_attr av = {
199 			.dlid	       = be16_to_cpu(mcast->mcmember.mlid),
200 			.port_num      = priv->port,
201 			.sl	       = mcast->mcmember.sl,
202 			.ah_flags      = IB_AH_GRH,
203 			.static_rate   = mcast->mcmember.rate,
204 			.grh	       = {
205 				.flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
206 				.hop_limit     = mcast->mcmember.hop_limit,
207 				.sgid_index    = 0,
208 				.traffic_class = mcast->mcmember.traffic_class
209 			}
210 		};
211 		av.grh.dgid = mcast->mcmember.mgid;
212 
213 		ah = ipoib_create_ah(priv, priv->pd, &av);
214 		if (!ah) {
215 			ipoib_warn(priv, "ib_address_create failed\n");
216 		} else {
217 			spin_lock_irq(&priv->lock);
218 			mcast->ah = ah;
219 			spin_unlock_irq(&priv->lock);
220 
221 			ipoib_dbg_mcast(priv, "MGID %16D AV %p, LID 0x%04x, SL %d\n",
222 					mcast->mcmember.mgid.raw, ":",
223 					mcast->ah->ah,
224 					be16_to_cpu(mcast->mcmember.mlid),
225 					mcast->mcmember.sl);
226 		}
227 	}
228 
229 	NET_EPOCH_ENTER(et);
230 
231 	/* actually send any queued packets */
232 	while (mcast->pkt_queue.ifq_len) {
233 		struct mbuf *mb;
234 		_IF_DEQUEUE(&mcast->pkt_queue, mb);
235 		mb->m_pkthdr.rcvif = dev;
236 
237 		if (if_transmit(dev, mb))
238 			ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
239 	}
240 
241 	NET_EPOCH_EXIT(et);
242 	return 0;
243 }
244 
245 static int
246 ipoib_mcast_sendonly_join_complete(int status,
247 				   struct ib_sa_multicast *multicast)
248 {
249 	struct ipoib_mcast *mcast = multicast->context;
250 	struct ipoib_dev_priv *priv = mcast->priv;
251 
252 	/* We trap for port events ourselves. */
253 	if (status == -ENETRESET)
254 		return 0;
255 
256 	if (!status)
257 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
258 
259 	if (status) {
260 		if (mcast->logcount++ < 20)
261 			ipoib_dbg_mcast(priv, "multicast join failed for %16D, status %d\n",
262 					mcast->mcmember.mgid.raw, ":", status);
263 
264 		/* Flush out any queued packets */
265 		if_inc_counter(priv->dev, IFCOUNTER_OERRORS, mcast->pkt_queue.ifq_len);
266 		_IF_DRAIN(&mcast->pkt_queue);
267 
268 		/* Clear the busy flag so we try again */
269 		status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY,
270 					    &mcast->flags);
271 	}
272 	return status;
273 }
274 
275 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
276 {
277 	struct ipoib_dev_priv *priv = mcast->priv;
278 	struct ib_sa_mcmember_rec rec = {
279 #if 0				/* Some SMs don't support send-only yet */
280 		.join_state = 4
281 #else
282 		.join_state = 1
283 #endif
284 	};
285 	int ret = 0;
286 
287 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
288 		ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
289 		return -ENODEV;
290 	}
291 
292 	if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
293 		ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
294 		return -EBUSY;
295 	}
296 
297 	rec.mgid     = mcast->mcmember.mgid;
298 	rec.port_gid = priv->local_gid;
299 	rec.pkey     = cpu_to_be16(priv->pkey);
300 
301 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca,
302 					 priv->port, &rec,
303 					 IB_SA_MCMEMBER_REC_MGID	|
304 					 IB_SA_MCMEMBER_REC_PORT_GID	|
305 					 IB_SA_MCMEMBER_REC_PKEY	|
306 					 IB_SA_MCMEMBER_REC_JOIN_STATE,
307 					 GFP_ATOMIC,
308 					 ipoib_mcast_sendonly_join_complete,
309 					 mcast);
310 	if (IS_ERR(mcast->mc)) {
311 		ret = PTR_ERR(mcast->mc);
312 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
313 		ipoib_warn(priv, "ib_sa_join_multicast failed (ret = %d)\n",
314 			   ret);
315 	} else {
316 		ipoib_dbg_mcast(priv, "no multicast record for %16D, starting join\n",
317 				mcast->mcmember.mgid.raw, ":");
318 	}
319 
320 	return ret;
321 }
322 
323 void ipoib_mcast_carrier_on_task(struct work_struct *work)
324 {
325 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
326 						   carrier_on_task);
327 	struct ib_port_attr attr;
328 
329 	/*
330 	 * Take rtnl_lock to avoid racing with ipoib_stop() and
331 	 * turning the carrier back on while a device is being
332 	 * removed.
333 	 */
334 	if (ib_query_port(priv->ca, priv->port, &attr) ||
335 	    attr.state != IB_PORT_ACTIVE) {
336 		ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
337 		return;
338 	}
339 	if_link_state_change(priv->dev, LINK_STATE_UP);
340 }
341 
342 static int ipoib_mcast_join_complete(int status,
343 				     struct ib_sa_multicast *multicast)
344 {
345 	struct ipoib_mcast *mcast = multicast->context;
346 	struct ipoib_dev_priv *priv = mcast->priv;
347 
348 	ipoib_dbg_mcast(priv, "join completion for %16D (status %d)\n",
349 			mcast->mcmember.mgid.raw, ":", status);
350 
351 	/* We trap for port events ourselves. */
352 	if (status == -ENETRESET)
353 		return 0;
354 
355 	if (!status)
356 		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
357 
358 	if (!status) {
359 		mcast->backoff = 1;
360 		mutex_lock(&mcast_mutex);
361 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
362 			queue_delayed_work(ipoib_workqueue,
363 					   &priv->mcast_task, 0);
364 		mutex_unlock(&mcast_mutex);
365 
366 		/*
367 		 * Defer carrier on work to ipoib_workqueue to avoid a
368 		 * deadlock on rtnl_lock here.
369 		 */
370 		if (mcast == priv->broadcast)
371 			queue_work(ipoib_workqueue, &priv->carrier_on_task);
372 
373 		return 0;
374 	}
375 
376 	if (mcast->logcount++ < 20) {
377 		if (status == -ETIMEDOUT || status == -EAGAIN) {
378 			ipoib_dbg_mcast(priv, "multicast join failed for %16D, status %d\n",
379 					mcast->mcmember.mgid.raw, ":", status);
380 		} else {
381 			ipoib_warn(priv, "multicast join failed for %16D, status %d\n",
382 				   mcast->mcmember.mgid.raw, ":", status);
383 		}
384 	}
385 
386 	mcast->backoff *= 2;
387 	if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
388 		mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
389 
390 	/* Clear the busy flag so we try again */
391 	status = test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
392 
393 	mutex_lock(&mcast_mutex);
394 	spin_lock_irq(&priv->lock);
395 	if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
396 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
397 				   mcast->backoff * HZ);
398 	spin_unlock_irq(&priv->lock);
399 	mutex_unlock(&mcast_mutex);
400 
401 	return status;
402 }
403 
404 static void ipoib_mcast_join(struct ipoib_dev_priv *priv,
405     struct ipoib_mcast *mcast, int create)
406 {
407 	struct ib_sa_mcmember_rec rec = {
408 		.join_state = 1
409 	};
410 	ib_sa_comp_mask comp_mask;
411 	int ret = 0;
412 
413 	ipoib_dbg_mcast(priv, "joining MGID %16D\n",
414 	    mcast->mcmember.mgid.raw, ":");
415 
416 	rec.mgid     = mcast->mcmember.mgid;
417 	rec.port_gid = priv->local_gid;
418 	rec.pkey     = cpu_to_be16(priv->pkey);
419 
420 	comp_mask =
421 		IB_SA_MCMEMBER_REC_MGID		|
422 		IB_SA_MCMEMBER_REC_PORT_GID	|
423 		IB_SA_MCMEMBER_REC_PKEY		|
424 		IB_SA_MCMEMBER_REC_JOIN_STATE;
425 
426 	if (create) {
427 		comp_mask |=
428 			IB_SA_MCMEMBER_REC_QKEY			|
429 			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
430 			IB_SA_MCMEMBER_REC_MTU			|
431 			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
432 			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
433 			IB_SA_MCMEMBER_REC_RATE			|
434 			IB_SA_MCMEMBER_REC_SL			|
435 			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
436 			IB_SA_MCMEMBER_REC_HOP_LIMIT;
437 
438 		rec.qkey	  = priv->broadcast->mcmember.qkey;
439 		rec.mtu_selector  = IB_SA_EQ;
440 		rec.mtu		  = priv->broadcast->mcmember.mtu;
441 		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
442 		rec.rate_selector = IB_SA_EQ;
443 		rec.rate	  = priv->broadcast->mcmember.rate;
444 		rec.sl		  = priv->broadcast->mcmember.sl;
445 		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
446 		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
447 	}
448 
449 	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
450 	mcast->mc = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
451 					 &rec, comp_mask, GFP_KERNEL,
452 					 ipoib_mcast_join_complete, mcast);
453 	if (IS_ERR(mcast->mc)) {
454 		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
455 		ret = PTR_ERR(mcast->mc);
456 		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
457 
458 		mcast->backoff *= 2;
459 		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
460 			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
461 
462 		mutex_lock(&mcast_mutex);
463 		if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
464 			queue_delayed_work(ipoib_workqueue,
465 					   &priv->mcast_task,
466 					   mcast->backoff * HZ);
467 		mutex_unlock(&mcast_mutex);
468 	}
469 }
470 
471 void ipoib_mcast_join_task(struct work_struct *work)
472 {
473 	struct ipoib_dev_priv *priv =
474 		container_of(work, struct ipoib_dev_priv, mcast_task.work);
475 	if_t dev = priv->dev;
476 	struct ib_port_attr attr;
477 
478 	ipoib_dbg_mcast(priv, "Running join task. flags 0x%lX\n", priv->flags);
479 
480 	if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
481 		return;
482 
483 	if (ib_query_port(priv->ca, priv->port, &attr) ||
484             attr.state != IB_PORT_ACTIVE) {
485 		ipoib_dbg(priv, "%s: port state is not ACTIVE (state = %d) suspend task.\n",
486                           __func__, attr.state);
487 		return;
488 	}
489 
490 	if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid, NULL))
491 		ipoib_warn(priv, "ib_query_gid() failed\n");
492 	else
493 		memcpy(if_getlladdr(dev) + 4, priv->local_gid.raw, sizeof (union ib_gid));
494 
495 	{
496 		struct ib_port_attr attr;
497 
498 		if (!ib_query_port(priv->ca, priv->port, &attr))
499 			priv->local_lid = attr.lid;
500 		else
501 			ipoib_warn(priv, "ib_query_port failed\n");
502 	}
503 
504 	if (!priv->broadcast) {
505 		struct ipoib_mcast *broadcast;
506 
507 		if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
508 			return;
509 
510 		broadcast = ipoib_mcast_alloc(priv, 1);
511 		if (!broadcast) {
512 			ipoib_warn(priv, "failed to allocate broadcast group\n");
513 			mutex_lock(&mcast_mutex);
514 			if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
515 				queue_delayed_work(ipoib_workqueue,
516 						   &priv->mcast_task, HZ);
517 			mutex_unlock(&mcast_mutex);
518 			return;
519 		}
520 
521 		spin_lock_irq(&priv->lock);
522 		memcpy(broadcast->mcmember.mgid.raw, if_getbroadcastaddr(dev) + 4,
523 		       sizeof (union ib_gid));
524 		priv->broadcast = broadcast;
525 
526 		__ipoib_mcast_add(priv, priv->broadcast);
527 		spin_unlock_irq(&priv->lock);
528 	}
529 
530 	if (priv->broadcast &&
531 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
532 		if (priv->broadcast &&
533 		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags))
534 			ipoib_mcast_join(priv, priv->broadcast, 0);
535 		return;
536 	}
537 
538 	while (1) {
539 		struct ipoib_mcast *mcast = NULL;
540 
541 		spin_lock_irq(&priv->lock);
542 		list_for_each_entry(mcast, &priv->multicast_list, list) {
543 			if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
544 			    && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
545 			    && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
546 				/* Found the next unjoined group */
547 				break;
548 			}
549 		}
550 		spin_unlock_irq(&priv->lock);
551 
552 		if (&mcast->list == &priv->multicast_list) {
553 			/* All done */
554 			break;
555 		}
556 
557 		ipoib_mcast_join(priv, mcast, 1);
558 		return;
559 	}
560 
561 	spin_lock_irq(&priv->lock);
562 	if (priv->broadcast)
563 		priv->mcast_mtu = IPOIB_UD_MTU(ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu));
564 	else
565 		priv->mcast_mtu = priv->admin_mtu;
566 	spin_unlock_irq(&priv->lock);
567 
568 	if (!ipoib_cm_admin_enabled(priv))
569 		ipoib_change_mtu(priv, min(priv->mcast_mtu, priv->admin_mtu),
570 		    true);
571 
572 	ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
573 
574 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
575 }
576 
577 int ipoib_mcast_start_thread(struct ipoib_dev_priv *priv)
578 {
579 	ipoib_dbg_mcast(priv, "starting multicast thread flags 0x%lX\n",
580 	    priv->flags);
581 
582 	mutex_lock(&mcast_mutex);
583 	if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
584 		queue_delayed_work(ipoib_workqueue, &priv->mcast_task, 0);
585 	mutex_unlock(&mcast_mutex);
586 
587 	return 0;
588 }
589 
590 int ipoib_mcast_stop_thread(struct ipoib_dev_priv *priv, int flush)
591 {
592 
593 	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
594 
595 	mutex_lock(&mcast_mutex);
596 	clear_bit(IPOIB_MCAST_RUN, &priv->flags);
597 	cancel_delayed_work(&priv->mcast_task);
598 	mutex_unlock(&mcast_mutex);
599 
600 	if (flush)
601 		flush_workqueue(ipoib_workqueue);
602 
603 	return 0;
604 }
605 
606 static int ipoib_mcast_leave(struct ipoib_dev_priv *priv, struct ipoib_mcast *mcast)
607 {
608 	int ret = 0;
609 
610 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
611 		ib_sa_free_multicast(mcast->mc);
612 
613 	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
614 		ipoib_dbg_mcast(priv, "leaving MGID %16D\n",
615 				mcast->mcmember.mgid.raw, ":");
616 
617 		/* Remove ourselves from the multicast group */
618 		ret = ib_detach_mcast(priv->qp, &mcast->mcmember.mgid,
619 				      be16_to_cpu(mcast->mcmember.mlid));
620 		if (ret)
621 			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
622 	}
623 
624 	return 0;
625 }
626 
627 void
628 ipoib_mcast_send(struct ipoib_dev_priv *priv, void *mgid, struct mbuf *mb)
629 {
630 	if_t dev = priv->dev;
631 	struct ipoib_mcast *mcast;
632 
633 	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
634 	    !priv->broadcast					||
635 	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
636 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
637 		m_freem(mb);
638 		return;
639 	}
640 
641 	mcast = __ipoib_mcast_find(priv, mgid);
642 	if (!mcast) {
643 		/* Let's create a new send only group now */
644 		ipoib_dbg_mcast(priv, "setting up send only multicast group for %16D\n",
645 				mgid, ":");
646 
647 		mcast = ipoib_mcast_alloc(priv, 0);
648 		if (!mcast) {
649 			ipoib_warn(priv, "unable to allocate memory for "
650 				   "multicast structure\n");
651 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
652 			m_freem(mb);
653 			goto out;
654 		}
655 
656 		set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
657 		memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
658 		__ipoib_mcast_add(priv, mcast);
659 		list_add_tail(&mcast->list, &priv->multicast_list);
660 	}
661 
662 	if (!mcast->ah) {
663 		if (mcast->pkt_queue.ifq_len < IPOIB_MAX_MCAST_QUEUE) {
664 			_IF_ENQUEUE(&mcast->pkt_queue, mb);
665 		} else {
666 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
667 			m_freem(mb);
668 		}
669 
670 		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
671 			ipoib_dbg_mcast(priv, "no address vector, "
672 					"but multicast join already started\n");
673 		else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
674 			ipoib_mcast_sendonly_join(mcast);
675 
676 		/*
677 		 * If lookup completes between here and out:, don't
678 		 * want to send packet twice.
679 		 */
680 		mcast = NULL;
681 	}
682 
683 out:
684 	if (mcast && mcast->ah)
685 		ipoib_send(priv, mb, mcast->ah, IB_MULTICAST_QPN);
686 }
687 
688 void ipoib_mcast_dev_flush(struct ipoib_dev_priv *priv)
689 {
690 	LIST_HEAD(remove_list);
691 	struct ipoib_mcast *mcast, *tmcast;
692 	unsigned long flags;
693 
694 	ipoib_dbg_mcast(priv, "flushing multicast list\n");
695 
696 	spin_lock_irqsave(&priv->lock, flags);
697 
698 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
699 		list_del(&mcast->list);
700 		rb_erase(&mcast->rb_node, &priv->multicast_tree);
701 		list_add_tail(&mcast->list, &remove_list);
702 	}
703 
704 	if (priv->broadcast) {
705 		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
706 		list_add_tail(&priv->broadcast->list, &remove_list);
707 		priv->broadcast = NULL;
708 	}
709 
710 	spin_unlock_irqrestore(&priv->lock, flags);
711 
712 	list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
713 		ipoib_mcast_leave(priv, mcast);
714 		ipoib_mcast_free(mcast);
715 	}
716 }
717 
718 static int ipoib_mcast_addr_is_valid(const u8 *addr, unsigned int addrlen,
719 				     const u8 *broadcast)
720 {
721 	if (addrlen != INFINIBAND_ALEN)
722 		return 0;
723 	/* reserved QPN, prefix, scope */
724 	if (memcmp(addr, broadcast, 6))
725 		return 0;
726 	/* signature lower, pkey */
727 	if (memcmp(addr + 7, broadcast + 7, 3))
728 		return 0;
729 	return 1;
730 }
731 
732 void ipoib_mcast_restart_task(struct work_struct *work)
733 {
734 	struct ipoib_dev_priv *priv =
735 		container_of(work, struct ipoib_dev_priv, restart_task);
736 	ipoib_mcast_restart(priv);
737 }
738 
739 struct ipoib_mcast_ctx {
740 	struct ipoib_dev_priv *priv;
741 	struct list_head remove_list;
742 };
743 
744 static u_int
745 ipoib_process_maddr(void *arg, struct sockaddr_dl *sdl, u_int cnt)
746 {
747 	struct ipoib_mcast_ctx *ctx = arg;
748 	struct ipoib_dev_priv *priv = ctx->priv;
749 	struct ipoib_mcast *mcast;
750 	struct ib_sa_mcmember_rec rec;
751 	union ib_gid mgid;
752 	uint8_t *addr;
753 	int addrlen;
754 
755 	addr = LLADDR(sdl);
756 	addrlen = sdl->sdl_alen;
757 	if (!ipoib_mcast_addr_is_valid(addr, addrlen,
758 	    if_getbroadcastaddr(priv->dev)))
759 		return (0);
760 
761 	memcpy(mgid.raw, addr + 4, sizeof mgid);
762 
763 	mcast = __ipoib_mcast_find(priv, &mgid);
764 	if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
765 		struct ipoib_mcast *nmcast;
766 
767 		/* ignore group which is directly joined by userspace */
768 		if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
769 		    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
770 			ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %16D\n",
771 					mgid.raw, ":");
772 			return (0);
773 		}
774 
775 		/* Not found or send-only group, let's add a new entry */
776 		ipoib_dbg_mcast(priv, "adding multicast entry for mgid %16D\n",
777 				mgid.raw, ":");
778 
779 		nmcast = ipoib_mcast_alloc(priv, 0);
780 		if (!nmcast) {
781 			ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
782 			return (0);
783 		}
784 
785 		set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
786 
787 		nmcast->mcmember.mgid = mgid;
788 
789 		if (mcast) {
790 			/* Destroy the send only entry */
791 			list_move_tail(&mcast->list, &ctx->remove_list);
792 
793 			rb_replace_node(&mcast->rb_node,
794 					&nmcast->rb_node,
795 					&priv->multicast_tree);
796 		} else
797 			__ipoib_mcast_add(priv, nmcast);
798 
799 		list_add_tail(&nmcast->list, &priv->multicast_list);
800 	}
801 
802 	if (mcast)
803 		set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
804 
805 	return (1);
806 }
807 
808 void ipoib_mcast_restart(struct ipoib_dev_priv *priv)
809 {
810 	struct ipoib_mcast_ctx ctx = { priv,
811 	    { &ctx.remove_list, &ctx.remove_list }};
812 	if_t dev = priv->dev;
813 	struct ipoib_mcast *mcast, *tmcast;
814 
815 	ipoib_dbg_mcast(priv, "restarting multicast task flags 0x%lX\n",
816 	    priv->flags);
817 
818 	ipoib_mcast_stop_thread(priv, 0);
819 
820 	spin_lock(&priv->lock);
821 
822 	/*
823 	 * Unfortunately, the networking core only gives us a list of all of
824 	 * the multicast hardware addresses. We need to figure out which ones
825 	 * are new and which ones have been removed
826 	 */
827 
828 	/* Clear out the found flag */
829 	list_for_each_entry(mcast, &priv->multicast_list, list)
830 		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
831 
832 	/* Mark all of the entries that are found or don't exist */
833 	ctx.priv = priv;
834 	if_foreach_llmaddr(dev, ipoib_process_maddr, &ctx);
835 
836 	/* Remove all of the entries don't exist anymore */
837 	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
838 		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
839 		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
840 			ipoib_dbg_mcast(priv, "deleting multicast group %16D\n",
841 					mcast->mcmember.mgid.raw, ":");
842 
843 			rb_erase(&mcast->rb_node, &priv->multicast_tree);
844 
845 			/* Move to the remove list */
846 			list_move_tail(&mcast->list, &ctx.remove_list);
847 		}
848 	}
849 
850 	spin_unlock(&priv->lock);
851 
852 	/* We have to cancel outside of the spinlock */
853 	list_for_each_entry_safe(mcast, tmcast, &ctx.remove_list, list) {
854 		ipoib_mcast_leave(mcast->priv, mcast);
855 		ipoib_mcast_free(mcast);
856 	}
857 
858 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
859 		ipoib_mcast_start_thread(priv);
860 }
861 
862 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
863 
864 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct ipoib_dev_priv *priv)
865 {
866 	struct ipoib_mcast_iter *iter;
867 
868 	iter = kmalloc(sizeof *iter, GFP_KERNEL);
869 	if (!iter)
870 		return NULL;
871 
872 	iter->priv = priv;
873 	memset(iter->mgid.raw, 0, 16);
874 
875 	if (ipoib_mcast_iter_next(iter)) {
876 		kfree(iter);
877 		return NULL;
878 	}
879 
880 	return iter;
881 }
882 
883 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
884 {
885 	struct ipoib_dev_priv *priv = iter->priv;
886 	struct rb_node *n;
887 	struct ipoib_mcast *mcast;
888 	int ret = 1;
889 
890 	spin_lock_irq(&priv->lock);
891 
892 	n = rb_first(&priv->multicast_tree);
893 
894 	while (n) {
895 		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
896 
897 		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
898 			   sizeof (union ib_gid)) < 0) {
899 			iter->mgid      = mcast->mcmember.mgid;
900 			iter->created   = mcast->created;
901 			iter->queuelen  = mcast->pkt_queue.ifq_len;
902 			iter->complete  = !!mcast->ah;
903 			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
904 
905 			ret = 0;
906 
907 			break;
908 		}
909 
910 		n = rb_next(n);
911 	}
912 
913 	spin_unlock_irq(&priv->lock);
914 
915 	return ret;
916 }
917 
918 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
919 			   union ib_gid *mgid,
920 			   unsigned long *created,
921 			   unsigned int *queuelen,
922 			   unsigned int *complete,
923 			   unsigned int *send_only)
924 {
925 	*mgid      = iter->mgid;
926 	*created   = iter->created;
927 	*queuelen  = iter->queuelen;
928 	*complete  = iter->complete;
929 	*send_only = iter->send_only;
930 }
931 
932 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
933