e03b070d522210e332ebd3b3996015cdb932cd3e
[linux-2.6.git] / drivers / infiniband / ulp / ipoib / ipoib_multicast.c
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  * $Id: ipoib_multicast.c 1362 2004-12-18 15:56:29Z roland $
35  */
36
37 #include <linux/skbuff.h>
38 #include <linux/rtnetlink.h>
39 #include <linux/ip.h>
40 #include <linux/in.h>
41 #include <linux/igmp.h>
42 #include <linux/inetdevice.h>
43 #include <linux/delay.h>
44 #include <linux/completion.h>
45
46 #include "ipoib.h"
47
48 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
49 static int mcast_debug_level;
50
51 module_param(mcast_debug_level, int, 0644);
52 MODULE_PARM_DESC(mcast_debug_level,
53                  "Enable multicast debug tracing if > 0");
54 #endif
55
56 static DECLARE_MUTEX(mcast_mutex);
57
58 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
59 struct ipoib_mcast {
60         struct ib_sa_mcmember_rec mcmember;
61         struct ipoib_ah          *ah;
62
63         struct rb_node    rb_node;
64         struct list_head  list;
65         struct completion done;
66
67         int                 query_id;
68         struct ib_sa_query *query;
69
70         unsigned long created;
71         unsigned long backoff;
72
73         unsigned long flags;
74         unsigned char logcount;
75
76         struct list_head  neigh_list;
77
78         struct sk_buff_head pkt_queue;
79
80         struct net_device *dev;
81 };
82
83 struct ipoib_mcast_iter {
84         struct net_device *dev;
85         union ib_gid       mgid;
86         unsigned long      created;
87         unsigned int       queuelen;
88         unsigned int       complete;
89         unsigned int       send_only;
90 };
91
92 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
93 {
94         struct net_device *dev = mcast->dev;
95         struct ipoib_dev_priv *priv = netdev_priv(dev);
96         struct ipoib_neigh *neigh, *tmp;
97         unsigned long flags;
98         LIST_HEAD(ah_list);
99         struct ipoib_ah *ah, *tah;
100
101         ipoib_dbg_mcast(netdev_priv(dev),
102                         "deleting multicast group " IPOIB_GID_FMT "\n",
103                         IPOIB_GID_ARG(mcast->mcmember.mgid));
104
105         spin_lock_irqsave(&priv->lock, flags);
106
107         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
108                 if (neigh->ah)
109                         list_add_tail(&neigh->ah->list, &ah_list);
110                 *to_ipoib_neigh(neigh->neighbour) = NULL;
111                 neigh->neighbour->ops->destructor = NULL;
112                 kfree(neigh);
113         }
114
115         spin_unlock_irqrestore(&priv->lock, flags);
116
117         list_for_each_entry_safe(ah, tah, &ah_list, list)
118                 ipoib_put_ah(ah);
119
120         if (mcast->ah)
121                 ipoib_put_ah(mcast->ah);
122
123         while (!skb_queue_empty(&mcast->pkt_queue)) {
124                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
125
126                 skb->dev = dev;
127                 dev_kfree_skb_any(skb);
128         }
129
130         kfree(mcast);
131 }
132
133 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
134                                              int can_sleep)
135 {
136         struct ipoib_mcast *mcast;
137
138         mcast = kmalloc(sizeof (*mcast), can_sleep ? GFP_KERNEL : GFP_ATOMIC);
139         if (!mcast)
140                 return NULL;
141
142         memset(mcast, 0, sizeof (*mcast));
143
144         init_completion(&mcast->done);
145
146         mcast->dev = dev;
147         mcast->created = jiffies;
148         mcast->backoff = HZ;
149         mcast->logcount = 0;
150
151         INIT_LIST_HEAD(&mcast->list);
152         INIT_LIST_HEAD(&mcast->neigh_list);
153         skb_queue_head_init(&mcast->pkt_queue);
154
155         mcast->ah    = NULL;
156         mcast->query = NULL;
157
158         return mcast;
159 }
160
161 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
162 {
163         struct ipoib_dev_priv *priv = netdev_priv(dev);
164         struct rb_node *n = priv->multicast_tree.rb_node;
165
166         while (n) {
167                 struct ipoib_mcast *mcast;
168                 int ret;
169
170                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
171
172                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
173                              sizeof (union ib_gid));
174                 if (ret < 0)
175                         n = n->rb_left;
176                 else if (ret > 0)
177                         n = n->rb_right;
178                 else
179                         return mcast;
180         }
181
182         return NULL;
183 }
184
185 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
186 {
187         struct ipoib_dev_priv *priv = netdev_priv(dev);
188         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
189
190         while (*n) {
191                 struct ipoib_mcast *tmcast;
192                 int ret;
193
194                 pn = *n;
195                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
196
197                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
198                              sizeof (union ib_gid));
199                 if (ret < 0)
200                         n = &pn->rb_left;
201                 else if (ret > 0)
202                         n = &pn->rb_right;
203                 else
204                         return -EEXIST;
205         }
206
207         rb_link_node(&mcast->rb_node, pn, n);
208         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
209
210         return 0;
211 }
212
213 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
214                                    struct ib_sa_mcmember_rec *mcmember)
215 {
216         struct net_device *dev = mcast->dev;
217         struct ipoib_dev_priv *priv = netdev_priv(dev);
218         int ret;
219
220         mcast->mcmember = *mcmember;
221
222         /* Set the cached Q_Key before we attach if it's the broadcast group */
223         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
224                     sizeof (union ib_gid))) {
225                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
226                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
227         }
228
229         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
230                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
231                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
232                                    " already attached\n",
233                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
234
235                         return 0;
236                 }
237
238                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
239                                          &mcast->mcmember.mgid);
240                 if (ret < 0) {
241                         ipoib_warn(priv, "couldn't attach QP to multicast group "
242                                    IPOIB_GID_FMT "\n",
243                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
244
245                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
246                         return ret;
247                 }
248         }
249
250         {
251                 struct ib_ah_attr av = {
252                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
253                         .port_num      = priv->port,
254                         .sl            = mcast->mcmember.sl,
255                         .ah_flags      = IB_AH_GRH,
256                         .grh           = {
257                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
258                                 .hop_limit     = mcast->mcmember.hop_limit,
259                                 .sgid_index    = 0,
260                                 .traffic_class = mcast->mcmember.traffic_class
261                         }
262                 };
263                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
264
265                 av.grh.dgid = mcast->mcmember.mgid;
266
267                 if (path_rate > 0 && priv->local_rate > path_rate)
268                         av.static_rate = (priv->local_rate - 1) / path_rate;
269
270                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
271                                 av.static_rate, priv->local_rate,
272                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
273
274                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
275                 if (!mcast->ah) {
276                         ipoib_warn(priv, "ib_address_create failed\n");
277                 } else {
278                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
279                                         " AV %p, LID 0x%04x, SL %d\n",
280                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
281                                         mcast->ah->ah,
282                                         be16_to_cpu(mcast->mcmember.mlid),
283                                         mcast->mcmember.sl);
284                 }
285         }
286
287         /* actually send any queued packets */
288         while (!skb_queue_empty(&mcast->pkt_queue)) {
289                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
290
291                 skb->dev = dev;
292
293                 if (!skb->dst || !skb->dst->neighbour) {
294                         /* put pseudoheader back on for next time */
295                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
296                 }
297
298                 if (dev_queue_xmit(skb))
299                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
300         }
301
302         return 0;
303 }
304
305 static void
306 ipoib_mcast_sendonly_join_complete(int status,
307                                    struct ib_sa_mcmember_rec *mcmember,
308                                    void *mcast_ptr)
309 {
310         struct ipoib_mcast *mcast = mcast_ptr;
311         struct net_device *dev = mcast->dev;
312
313         if (!status)
314                 ipoib_mcast_join_finish(mcast, mcmember);
315         else {
316                 if (mcast->logcount++ < 20)
317                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
318                                         IPOIB_GID_FMT ", status %d\n",
319                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
320
321                 /* Flush out any queued packets */
322                 while (!skb_queue_empty(&mcast->pkt_queue)) {
323                         struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
324
325                         skb->dev = dev;
326
327                         dev_kfree_skb_any(skb);
328                 }
329
330                 /* Clear the busy flag so we try again */
331                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
332         }
333
334         complete(&mcast->done);
335 }
336
337 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
338 {
339         struct net_device *dev = mcast->dev;
340         struct ipoib_dev_priv *priv = netdev_priv(dev);
341         struct ib_sa_mcmember_rec rec = {
342 #if 0                           /* Some SMs don't support send-only yet */
343                 .join_state = 4
344 #else
345                 .join_state = 1
346 #endif
347         };
348         int ret = 0;
349
350         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
351                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
352                 return -ENODEV;
353         }
354
355         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
356                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
357                 return -EBUSY;
358         }
359
360         rec.mgid     = mcast->mcmember.mgid;
361         rec.port_gid = priv->local_gid;
362         rec.pkey     = be16_to_cpu(priv->pkey);
363
364         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
365                                      IB_SA_MCMEMBER_REC_MGID            |
366                                      IB_SA_MCMEMBER_REC_PORT_GID        |
367                                      IB_SA_MCMEMBER_REC_PKEY            |
368                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
369                                      1000, GFP_ATOMIC,
370                                      ipoib_mcast_sendonly_join_complete,
371                                      mcast, &mcast->query);
372         if (ret < 0) {
373                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
374                            ret);
375         } else {
376                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
377                                 ", starting join\n",
378                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
379
380                 mcast->query_id = ret;
381         }
382
383         return ret;
384 }
385
386 static void ipoib_mcast_join_complete(int status,
387                                       struct ib_sa_mcmember_rec *mcmember,
388                                       void *mcast_ptr)
389 {
390         struct ipoib_mcast *mcast = mcast_ptr;
391         struct net_device *dev = mcast->dev;
392         struct ipoib_dev_priv *priv = netdev_priv(dev);
393
394         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
395                         " (status %d)\n",
396                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
397
398         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
399                 mcast->backoff = HZ;
400                 down(&mcast_mutex);
401                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
402                         queue_work(ipoib_workqueue, &priv->mcast_task);
403                 up(&mcast_mutex);
404                 complete(&mcast->done);
405                 return;
406         }
407
408         if (status == -EINTR) {
409                 complete(&mcast->done);
410                 return;
411         }
412
413         if (status && mcast->logcount++ < 20) {
414                 if (status == -ETIMEDOUT || status == -EINTR) {
415                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
416                                         ", status %d\n",
417                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
418                                         status);
419                 } else {
420                         ipoib_warn(priv, "multicast join failed for "
421                                    IPOIB_GID_FMT ", status %d\n",
422                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
423                                    status);
424                 }
425         }
426
427         mcast->backoff *= 2;
428         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
429                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
430
431         mcast->query = NULL;
432
433         down(&mcast_mutex);
434         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
435                 if (status == -ETIMEDOUT)
436                         queue_work(ipoib_workqueue, &priv->mcast_task);
437                 else
438                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
439                                            mcast->backoff * HZ);
440         } else
441                 complete(&mcast->done);
442         up(&mcast_mutex);
443
444         return;
445 }
446
447 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
448                              int create)
449 {
450         struct ipoib_dev_priv *priv = netdev_priv(dev);
451         struct ib_sa_mcmember_rec rec = {
452                 .join_state = 1
453         };
454         ib_sa_comp_mask comp_mask;
455         int ret = 0;
456
457         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
458                         IPOIB_GID_ARG(mcast->mcmember.mgid));
459
460         rec.mgid     = mcast->mcmember.mgid;
461         rec.port_gid = priv->local_gid;
462         rec.pkey     = be16_to_cpu(priv->pkey);
463
464         comp_mask =
465                 IB_SA_MCMEMBER_REC_MGID         |
466                 IB_SA_MCMEMBER_REC_PORT_GID     |
467                 IB_SA_MCMEMBER_REC_PKEY         |
468                 IB_SA_MCMEMBER_REC_JOIN_STATE;
469
470         if (create) {
471                 comp_mask |=
472                         IB_SA_MCMEMBER_REC_QKEY         |
473                         IB_SA_MCMEMBER_REC_SL           |
474                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
475                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
476
477                 rec.qkey          = priv->broadcast->mcmember.qkey;
478                 rec.sl            = priv->broadcast->mcmember.sl;
479                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
480                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
481         }
482
483         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
484                                      mcast->backoff * 1000, GFP_ATOMIC,
485                                      ipoib_mcast_join_complete,
486                                      mcast, &mcast->query);
487
488         if (ret < 0) {
489                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
490
491                 mcast->backoff *= 2;
492                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
493                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
494
495                 down(&mcast_mutex);
496                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
497                         queue_delayed_work(ipoib_workqueue,
498                                            &priv->mcast_task,
499                                            mcast->backoff);
500                 up(&mcast_mutex);
501         } else
502                 mcast->query_id = ret;
503 }
504
505 void ipoib_mcast_join_task(void *dev_ptr)
506 {
507         struct net_device *dev = dev_ptr;
508         struct ipoib_dev_priv *priv = netdev_priv(dev);
509
510         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
511                 return;
512
513         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
514                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
515         else
516                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
517
518         {
519                 struct ib_port_attr attr;
520
521                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
522                         priv->local_lid  = attr.lid;
523                         priv->local_rate = attr.active_speed *
524                                 ib_width_enum_to_int(attr.active_width);
525                 } else
526                         ipoib_warn(priv, "ib_query_port failed\n");
527         }
528
529         if (!priv->broadcast) {
530                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
531                 if (!priv->broadcast) {
532                         ipoib_warn(priv, "failed to allocate broadcast group\n");
533                         down(&mcast_mutex);
534                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
535                                 queue_delayed_work(ipoib_workqueue,
536                                                    &priv->mcast_task, HZ);
537                         up(&mcast_mutex);
538                         return;
539                 }
540
541                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
542                        sizeof (union ib_gid));
543
544                 spin_lock_irq(&priv->lock);
545                 __ipoib_mcast_add(dev, priv->broadcast);
546                 spin_unlock_irq(&priv->lock);
547         }
548
549         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
550                 ipoib_mcast_join(dev, priv->broadcast, 0);
551                 return;
552         }
553
554         while (1) {
555                 struct ipoib_mcast *mcast = NULL;
556
557                 spin_lock_irq(&priv->lock);
558                 list_for_each_entry(mcast, &priv->multicast_list, list) {
559                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
560                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
561                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
562                                 /* Found the next unjoined group */
563                                 break;
564                         }
565                 }
566                 spin_unlock_irq(&priv->lock);
567
568                 if (&mcast->list == &priv->multicast_list) {
569                         /* All done */
570                         break;
571                 }
572
573                 ipoib_mcast_join(dev, mcast, 1);
574                 return;
575         }
576
577         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
578                 IPOIB_ENCAP_LEN;
579         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
580
581         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
582
583         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
584         netif_carrier_on(dev);
585 }
586
587 int ipoib_mcast_start_thread(struct net_device *dev)
588 {
589         struct ipoib_dev_priv *priv = netdev_priv(dev);
590
591         ipoib_dbg_mcast(priv, "starting multicast thread\n");
592
593         down(&mcast_mutex);
594         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
595                 queue_work(ipoib_workqueue, &priv->mcast_task);
596         up(&mcast_mutex);
597
598         return 0;
599 }
600
601 int ipoib_mcast_stop_thread(struct net_device *dev)
602 {
603         struct ipoib_dev_priv *priv = netdev_priv(dev);
604         struct ipoib_mcast *mcast;
605
606         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
607
608         down(&mcast_mutex);
609         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
610         cancel_delayed_work(&priv->mcast_task);
611         up(&mcast_mutex);
612
613         flush_workqueue(ipoib_workqueue);
614
615         if (priv->broadcast && priv->broadcast->query) {
616                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
617                 priv->broadcast->query = NULL;
618                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
619                 wait_for_completion(&priv->broadcast->done);
620         }
621
622         list_for_each_entry(mcast, &priv->multicast_list, list) {
623                 if (mcast->query) {
624                         ib_sa_cancel_query(mcast->query_id, mcast->query);
625                         mcast->query = NULL;
626                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
627                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
628                         wait_for_completion(&mcast->done);
629                 }
630         }
631
632         return 0;
633 }
634
635 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
636 {
637         struct ipoib_dev_priv *priv = netdev_priv(dev);
638         struct ib_sa_mcmember_rec rec = {
639                 .join_state = 1
640         };
641         int ret = 0;
642
643         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
644                 return 0;
645
646         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
647                         IPOIB_GID_ARG(mcast->mcmember.mgid));
648
649         rec.mgid     = mcast->mcmember.mgid;
650         rec.port_gid = priv->local_gid;
651         rec.pkey     = be16_to_cpu(priv->pkey);
652
653         /* Remove ourselves from the multicast group */
654         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
655                                  &mcast->mcmember.mgid);
656         if (ret)
657                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
658
659         /*
660          * Just make one shot at leaving and don't wait for a reply;
661          * if we fail, too bad.
662          */
663         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
664                                         IB_SA_MCMEMBER_REC_MGID         |
665                                         IB_SA_MCMEMBER_REC_PORT_GID     |
666                                         IB_SA_MCMEMBER_REC_PKEY         |
667                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
668                                         0, GFP_ATOMIC, NULL,
669                                         mcast, &mcast->query);
670         if (ret < 0)
671                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
672                            "for leave (result = %d)\n", ret);
673
674         return 0;
675 }
676
677 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
678                       struct sk_buff *skb)
679 {
680         struct ipoib_dev_priv *priv = netdev_priv(dev);
681         struct ipoib_mcast *mcast;
682
683         /*
684          * We can only be called from ipoib_start_xmit, so we're
685          * inside tx_lock -- no need to save/restore flags.
686          */
687         spin_lock(&priv->lock);
688
689         mcast = __ipoib_mcast_find(dev, mgid);
690         if (!mcast) {
691                 /* Let's create a new send only group now */
692                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
693                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
694
695                 mcast = ipoib_mcast_alloc(dev, 0);
696                 if (!mcast) {
697                         ipoib_warn(priv, "unable to allocate memory for "
698                                    "multicast structure\n");
699                         dev_kfree_skb_any(skb);
700                         goto out;
701                 }
702
703                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
704                 mcast->mcmember.mgid = *mgid;
705                 __ipoib_mcast_add(dev, mcast);
706                 list_add_tail(&mcast->list, &priv->multicast_list);
707         }
708
709         if (!mcast->ah) {
710                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
711                         skb_queue_tail(&mcast->pkt_queue, skb);
712                 else
713                         dev_kfree_skb_any(skb);
714
715                 if (mcast->query)
716                         ipoib_dbg_mcast(priv, "no address vector, "
717                                         "but multicast join already started\n");
718                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
719                         ipoib_mcast_sendonly_join(mcast);
720
721                 /*
722                  * If lookup completes between here and out:, don't
723                  * want to send packet twice.
724                  */
725                 mcast = NULL;
726         }
727
728 out:
729         if (mcast && mcast->ah) {
730                 if (skb->dst            &&
731                     skb->dst->neighbour &&
732                     !*to_ipoib_neigh(skb->dst->neighbour)) {
733                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
734
735                         if (neigh) {
736                                 kref_get(&mcast->ah->ref);
737                                 neigh->ah       = mcast->ah;
738                                 neigh->neighbour = skb->dst->neighbour;
739                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
740                                 list_add_tail(&neigh->list, &mcast->neigh_list);
741                         }
742                 }
743
744                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
745         }
746
747         spin_unlock(&priv->lock);
748 }
749
750 void ipoib_mcast_dev_flush(struct net_device *dev)
751 {
752         struct ipoib_dev_priv *priv = netdev_priv(dev);
753         LIST_HEAD(remove_list);
754         struct ipoib_mcast *mcast, *tmcast, *nmcast;
755         unsigned long flags;
756
757         ipoib_dbg_mcast(priv, "flushing multicast list\n");
758
759         spin_lock_irqsave(&priv->lock, flags);
760         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
761                 nmcast = ipoib_mcast_alloc(dev, 0);
762                 if (nmcast) {
763                         nmcast->flags =
764                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
765
766                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
767
768                         /* Add the new group in before the to-be-destroyed group */
769                         list_add_tail(&nmcast->list, &mcast->list);
770                         list_del_init(&mcast->list);
771
772                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
773                                         &priv->multicast_tree);
774
775                         list_add_tail(&mcast->list, &remove_list);
776                 } else {
777                         ipoib_warn(priv, "could not reallocate multicast group "
778                                    IPOIB_GID_FMT "\n",
779                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
780                 }
781         }
782
783         if (priv->broadcast) {
784                 nmcast = ipoib_mcast_alloc(dev, 0);
785                 if (nmcast) {
786                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
787
788                         rb_replace_node(&priv->broadcast->rb_node,
789                                         &nmcast->rb_node,
790                                         &priv->multicast_tree);
791
792                         list_add_tail(&priv->broadcast->list, &remove_list);
793                 }
794
795                 priv->broadcast = nmcast;
796         }
797
798         spin_unlock_irqrestore(&priv->lock, flags);
799
800         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
801                 ipoib_mcast_leave(dev, mcast);
802                 ipoib_mcast_free(mcast);
803         }
804 }
805
806 void ipoib_mcast_dev_down(struct net_device *dev)
807 {
808         struct ipoib_dev_priv *priv = netdev_priv(dev);
809         unsigned long flags;
810
811         /* Delete broadcast since it will be recreated */
812         if (priv->broadcast) {
813                 ipoib_dbg_mcast(priv, "deleting broadcast group\n");
814
815                 spin_lock_irqsave(&priv->lock, flags);
816                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
817                 spin_unlock_irqrestore(&priv->lock, flags);
818                 ipoib_mcast_leave(dev, priv->broadcast);
819                 ipoib_mcast_free(priv->broadcast);
820                 priv->broadcast = NULL;
821         }
822 }
823
824 void ipoib_mcast_restart_task(void *dev_ptr)
825 {
826         struct net_device *dev = dev_ptr;
827         struct ipoib_dev_priv *priv = netdev_priv(dev);
828         struct dev_mc_list *mclist;
829         struct ipoib_mcast *mcast, *tmcast;
830         LIST_HEAD(remove_list);
831         unsigned long flags;
832
833         ipoib_dbg_mcast(priv, "restarting multicast task\n");
834
835         ipoib_mcast_stop_thread(dev);
836
837         spin_lock_irqsave(&priv->lock, flags);
838
839         /*
840          * Unfortunately, the networking core only gives us a list of all of
841          * the multicast hardware addresses. We need to figure out which ones
842          * are new and which ones have been removed
843          */
844
845         /* Clear out the found flag */
846         list_for_each_entry(mcast, &priv->multicast_list, list)
847                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
848
849         /* Mark all of the entries that are found or don't exist */
850         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
851                 union ib_gid mgid;
852
853                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
854
855                 /* Add in the P_Key */
856                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
857                 mgid.raw[5] = priv->pkey & 0xff;
858
859                 mcast = __ipoib_mcast_find(dev, &mgid);
860                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
861                         struct ipoib_mcast *nmcast;
862
863                         /* Not found or send-only group, let's add a new entry */
864                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
865                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
866
867                         nmcast = ipoib_mcast_alloc(dev, 0);
868                         if (!nmcast) {
869                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
870                                 continue;
871                         }
872
873                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
874
875                         nmcast->mcmember.mgid = mgid;
876
877                         if (mcast) {
878                                 /* Destroy the send only entry */
879                                 list_del(&mcast->list);
880                                 list_add_tail(&mcast->list, &remove_list);
881
882                                 rb_replace_node(&mcast->rb_node,
883                                                 &nmcast->rb_node,
884                                                 &priv->multicast_tree);
885                         } else
886                                 __ipoib_mcast_add(dev, nmcast);
887
888                         list_add_tail(&nmcast->list, &priv->multicast_list);
889                 }
890
891                 if (mcast)
892                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
893         }
894
895         /* Remove all of the entries don't exist anymore */
896         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
897                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
898                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
899                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
900                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
901
902                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
903
904                         /* Move to the remove list */
905                         list_del(&mcast->list);
906                         list_add_tail(&mcast->list, &remove_list);
907                 }
908         }
909         spin_unlock_irqrestore(&priv->lock, flags);
910
911         /* We have to cancel outside of the spinlock */
912         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
913                 ipoib_mcast_leave(mcast->dev, mcast);
914                 ipoib_mcast_free(mcast);
915         }
916
917         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
918                 ipoib_mcast_start_thread(dev);
919 }
920
921 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
922 {
923         struct ipoib_mcast_iter *iter;
924
925         iter = kmalloc(sizeof *iter, GFP_KERNEL);
926         if (!iter)
927                 return NULL;
928
929         iter->dev = dev;
930         memset(iter->mgid.raw, 0, sizeof iter->mgid);
931
932         if (ipoib_mcast_iter_next(iter)) {
933                 ipoib_mcast_iter_free(iter);
934                 return NULL;
935         }
936
937         return iter;
938 }
939
940 void ipoib_mcast_iter_free(struct ipoib_mcast_iter *iter)
941 {
942         kfree(iter);
943 }
944
945 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
946 {
947         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
948         struct rb_node *n;
949         struct ipoib_mcast *mcast;
950         int ret = 1;
951
952         spin_lock_irq(&priv->lock);
953
954         n = rb_first(&priv->multicast_tree);
955
956         while (n) {
957                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
958
959                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
960                            sizeof (union ib_gid)) < 0) {
961                         iter->mgid      = mcast->mcmember.mgid;
962                         iter->created   = mcast->created;
963                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
964                         iter->complete  = !!mcast->ah;
965                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
966
967                         ret = 0;
968
969                         break;
970                 }
971
972                 n = rb_next(n);
973         }
974
975         spin_unlock_irq(&priv->lock);
976
977         return ret;
978 }
979
980 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
981                            union ib_gid *mgid,
982                            unsigned long *created,
983                            unsigned int *queuelen,
984                            unsigned int *complete,
985                            unsigned int *send_only)
986 {
987         *mgid      = iter->mgid;
988         *created   = iter->created;
989         *queuelen  = iter->queuelen;
990         *complete  = iter->complete;
991         *send_only = iter->send_only;
992 }