[PATCH] drivers: use list_move()
[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 <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 static DEFINE_MUTEX(mcast_mutex);
59
60 /* Used for all multicast joins (broadcast, IPv4 mcast and IPv6 mcast) */
61 struct ipoib_mcast {
62         struct ib_sa_mcmember_rec mcmember;
63         struct ipoib_ah          *ah;
64
65         struct rb_node    rb_node;
66         struct list_head  list;
67         struct completion done;
68
69         int                 query_id;
70         struct ib_sa_query *query;
71
72         unsigned long created;
73         unsigned long backoff;
74
75         unsigned long flags;
76         unsigned char logcount;
77
78         struct list_head  neigh_list;
79
80         struct sk_buff_head pkt_queue;
81
82         struct net_device *dev;
83 };
84
85 struct ipoib_mcast_iter {
86         struct net_device *dev;
87         union ib_gid       mgid;
88         unsigned long      created;
89         unsigned int       queuelen;
90         unsigned int       complete;
91         unsigned int       send_only;
92 };
93
94 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
95 {
96         struct net_device *dev = mcast->dev;
97         struct ipoib_dev_priv *priv = netdev_priv(dev);
98         struct ipoib_neigh *neigh, *tmp;
99         unsigned long flags;
100         int tx_dropped = 0;
101
102         ipoib_dbg_mcast(netdev_priv(dev),
103                         "deleting multicast group " IPOIB_GID_FMT "\n",
104                         IPOIB_GID_ARG(mcast->mcmember.mgid));
105
106         spin_lock_irqsave(&priv->lock, flags);
107
108         list_for_each_entry_safe(neigh, tmp, &mcast->neigh_list, list) {
109                 /*
110                  * It's safe to call ipoib_put_ah() inside priv->lock
111                  * here, because we know that mcast->ah will always
112                  * hold one more reference, so ipoib_put_ah() will
113                  * never do more than decrement the ref count.
114                  */
115                 if (neigh->ah)
116                         ipoib_put_ah(neigh->ah);
117                 ipoib_neigh_free(neigh);
118         }
119
120         spin_unlock_irqrestore(&priv->lock, flags);
121
122         if (mcast->ah)
123                 ipoib_put_ah(mcast->ah);
124
125         while (!skb_queue_empty(&mcast->pkt_queue)) {
126                 ++tx_dropped;
127                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
128         }
129
130         spin_lock_irqsave(&priv->tx_lock, flags);
131         priv->stats.tx_dropped += tx_dropped;
132         spin_unlock_irqrestore(&priv->tx_lock, flags);
133
134         kfree(mcast);
135 }
136
137 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev,
138                                              int can_sleep)
139 {
140         struct ipoib_mcast *mcast;
141
142         mcast = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
143         if (!mcast)
144                 return NULL;
145
146         mcast->dev = dev;
147         mcast->created = jiffies;
148         mcast->backoff = 1;
149
150         INIT_LIST_HEAD(&mcast->list);
151         INIT_LIST_HEAD(&mcast->neigh_list);
152         skb_queue_head_init(&mcast->pkt_queue);
153
154         return mcast;
155 }
156
157 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
158 {
159         struct ipoib_dev_priv *priv = netdev_priv(dev);
160         struct rb_node *n = priv->multicast_tree.rb_node;
161
162         while (n) {
163                 struct ipoib_mcast *mcast;
164                 int ret;
165
166                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
167
168                 ret = memcmp(mgid, mcast->mcmember.mgid.raw,
169                              sizeof (union ib_gid));
170                 if (ret < 0)
171                         n = n->rb_left;
172                 else if (ret > 0)
173                         n = n->rb_right;
174                 else
175                         return mcast;
176         }
177
178         return NULL;
179 }
180
181 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
182 {
183         struct ipoib_dev_priv *priv = netdev_priv(dev);
184         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
185
186         while (*n) {
187                 struct ipoib_mcast *tmcast;
188                 int ret;
189
190                 pn = *n;
191                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
192
193                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
194                              sizeof (union ib_gid));
195                 if (ret < 0)
196                         n = &pn->rb_left;
197                 else if (ret > 0)
198                         n = &pn->rb_right;
199                 else
200                         return -EEXIST;
201         }
202
203         rb_link_node(&mcast->rb_node, pn, n);
204         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
205
206         return 0;
207 }
208
209 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
210                                    struct ib_sa_mcmember_rec *mcmember)
211 {
212         struct net_device *dev = mcast->dev;
213         struct ipoib_dev_priv *priv = netdev_priv(dev);
214         struct ipoib_ah *ah;
215         int ret;
216
217         mcast->mcmember = *mcmember;
218
219         /* Set the cached Q_Key before we attach if it's the broadcast group */
220         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
221                     sizeof (union ib_gid))) {
222                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
223                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
224         }
225
226         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
227                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
228                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
229                                    " already attached\n",
230                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
231
232                         return 0;
233                 }
234
235                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
236                                          &mcast->mcmember.mgid);
237                 if (ret < 0) {
238                         ipoib_warn(priv, "couldn't attach QP to multicast group "
239                                    IPOIB_GID_FMT "\n",
240                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
241
242                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
243                         return ret;
244                 }
245         }
246
247         {
248                 struct ib_ah_attr av = {
249                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
250                         .port_num      = priv->port,
251                         .sl            = mcast->mcmember.sl,
252                         .ah_flags      = IB_AH_GRH,
253                         .static_rate   = mcast->mcmember.rate,
254                         .grh           = {
255                                 .flow_label    = be32_to_cpu(mcast->mcmember.flow_label),
256                                 .hop_limit     = mcast->mcmember.hop_limit,
257                                 .sgid_index    = 0,
258                                 .traffic_class = mcast->mcmember.traffic_class
259                         }
260                 };
261                 av.grh.dgid = mcast->mcmember.mgid;
262
263                 ah = ipoib_create_ah(dev, priv->pd, &av);
264                 if (!ah) {
265                         ipoib_warn(priv, "ib_address_create failed\n");
266                 } else {
267                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
268                                         " AV %p, LID 0x%04x, SL %d\n",
269                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
270                                         mcast->ah->ah,
271                                         be16_to_cpu(mcast->mcmember.mlid),
272                                         mcast->mcmember.sl);
273                 }
274
275                 spin_lock_irq(&priv->lock);
276                 mcast->ah = ah;
277                 spin_unlock_irq(&priv->lock);
278         }
279
280         /* actually send any queued packets */
281         spin_lock_irq(&priv->tx_lock);
282         while (!skb_queue_empty(&mcast->pkt_queue)) {
283                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
284                 spin_unlock_irq(&priv->tx_lock);
285
286                 skb->dev = dev;
287
288                 if (!skb->dst || !skb->dst->neighbour) {
289                         /* put pseudoheader back on for next time */
290                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
291                 }
292
293                 if (dev_queue_xmit(skb))
294                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
295                 spin_lock_irq(&priv->tx_lock);
296         }
297         spin_unlock_irq(&priv->tx_lock);
298
299         return 0;
300 }
301
302 static void
303 ipoib_mcast_sendonly_join_complete(int status,
304                                    struct ib_sa_mcmember_rec *mcmember,
305                                    void *mcast_ptr)
306 {
307         struct ipoib_mcast *mcast = mcast_ptr;
308         struct net_device *dev = mcast->dev;
309         struct ipoib_dev_priv *priv = netdev_priv(dev);
310
311         if (!status)
312                 ipoib_mcast_join_finish(mcast, mcmember);
313         else {
314                 if (mcast->logcount++ < 20)
315                         ipoib_dbg_mcast(netdev_priv(dev), "multicast join failed for "
316                                         IPOIB_GID_FMT ", status %d\n",
317                                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
318
319                 /* Flush out any queued packets */
320                 spin_lock_irq(&priv->tx_lock);
321                 while (!skb_queue_empty(&mcast->pkt_queue)) {
322                         ++priv->stats.tx_dropped;
323                         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
324                 }
325                 spin_unlock_irq(&priv->tx_lock);
326
327                 /* Clear the busy flag so we try again */
328                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
329         }
330
331         complete(&mcast->done);
332 }
333
334 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
335 {
336         struct net_device *dev = mcast->dev;
337         struct ipoib_dev_priv *priv = netdev_priv(dev);
338         struct ib_sa_mcmember_rec rec = {
339 #if 0                           /* Some SMs don't support send-only yet */
340                 .join_state = 4
341 #else
342                 .join_state = 1
343 #endif
344         };
345         int ret = 0;
346
347         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
348                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
349                 return -ENODEV;
350         }
351
352         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
353                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
354                 return -EBUSY;
355         }
356
357         rec.mgid     = mcast->mcmember.mgid;
358         rec.port_gid = priv->local_gid;
359         rec.pkey     = cpu_to_be16(priv->pkey);
360
361         init_completion(&mcast->done);
362
363         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
364                                      IB_SA_MCMEMBER_REC_MGID            |
365                                      IB_SA_MCMEMBER_REC_PORT_GID        |
366                                      IB_SA_MCMEMBER_REC_PKEY            |
367                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
368                                      1000, GFP_ATOMIC,
369                                      ipoib_mcast_sendonly_join_complete,
370                                      mcast, &mcast->query);
371         if (ret < 0) {
372                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
373                            ret);
374         } else {
375                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
376                                 ", starting join\n",
377                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
378
379                 mcast->query_id = ret;
380         }
381
382         return ret;
383 }
384
385 static void ipoib_mcast_join_complete(int status,
386                                       struct ib_sa_mcmember_rec *mcmember,
387                                       void *mcast_ptr)
388 {
389         struct ipoib_mcast *mcast = mcast_ptr;
390         struct net_device *dev = mcast->dev;
391         struct ipoib_dev_priv *priv = netdev_priv(dev);
392
393         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
394                         " (status %d)\n",
395                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
396
397         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
398                 mcast->backoff = 1;
399                 mutex_lock(&mcast_mutex);
400                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
401                         queue_work(ipoib_workqueue, &priv->mcast_task);
402                 mutex_unlock(&mcast_mutex);
403                 complete(&mcast->done);
404                 return;
405         }
406
407         if (status == -EINTR) {
408                 complete(&mcast->done);
409                 return;
410         }
411
412         if (status && mcast->logcount++ < 20) {
413                 if (status == -ETIMEDOUT || status == -EINTR) {
414                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
415                                         ", status %d\n",
416                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
417                                         status);
418                 } else {
419                         ipoib_warn(priv, "multicast join failed for "
420                                    IPOIB_GID_FMT ", status %d\n",
421                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
422                                    status);
423                 }
424         }
425
426         mcast->backoff *= 2;
427         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
428                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
429
430         mutex_lock(&mcast_mutex);
431
432         spin_lock_irq(&priv->lock);
433         mcast->query = NULL;
434
435         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
436                 if (status == -ETIMEDOUT)
437                         queue_work(ipoib_workqueue, &priv->mcast_task);
438                 else
439                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
440                                            mcast->backoff * HZ);
441         } else
442                 complete(&mcast->done);
443         spin_unlock_irq(&priv->lock);
444         mutex_unlock(&mcast_mutex);
445
446         return;
447 }
448
449 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
450                              int create)
451 {
452         struct ipoib_dev_priv *priv = netdev_priv(dev);
453         struct ib_sa_mcmember_rec rec = {
454                 .join_state = 1
455         };
456         ib_sa_comp_mask comp_mask;
457         int ret = 0;
458
459         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
460                         IPOIB_GID_ARG(mcast->mcmember.mgid));
461
462         rec.mgid     = mcast->mcmember.mgid;
463         rec.port_gid = priv->local_gid;
464         rec.pkey     = cpu_to_be16(priv->pkey);
465
466         comp_mask =
467                 IB_SA_MCMEMBER_REC_MGID         |
468                 IB_SA_MCMEMBER_REC_PORT_GID     |
469                 IB_SA_MCMEMBER_REC_PKEY         |
470                 IB_SA_MCMEMBER_REC_JOIN_STATE;
471
472         if (create) {
473                 comp_mask |=
474                         IB_SA_MCMEMBER_REC_QKEY         |
475                         IB_SA_MCMEMBER_REC_SL           |
476                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
477                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
478
479                 rec.qkey          = priv->broadcast->mcmember.qkey;
480                 rec.sl            = priv->broadcast->mcmember.sl;
481                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
482                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
483         }
484
485         init_completion(&mcast->done);
486
487         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
488                                      mcast->backoff * 1000, GFP_ATOMIC,
489                                      ipoib_mcast_join_complete,
490                                      mcast, &mcast->query);
491
492         if (ret < 0) {
493                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
494
495                 mcast->backoff *= 2;
496                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
497                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
498
499                 mutex_lock(&mcast_mutex);
500                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
501                         queue_delayed_work(ipoib_workqueue,
502                                            &priv->mcast_task,
503                                            mcast->backoff * HZ);
504                 mutex_unlock(&mcast_mutex);
505         } else
506                 mcast->query_id = ret;
507 }
508
509 void ipoib_mcast_join_task(void *dev_ptr)
510 {
511         struct net_device *dev = dev_ptr;
512         struct ipoib_dev_priv *priv = netdev_priv(dev);
513
514         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
515                 return;
516
517         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
518                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
519         else
520                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
521
522         {
523                 struct ib_port_attr attr;
524
525                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
526                         priv->local_lid  = attr.lid;
527                         priv->local_rate = attr.active_speed *
528                                 ib_width_enum_to_int(attr.active_width);
529                 } else
530                         ipoib_warn(priv, "ib_query_port failed\n");
531         }
532
533         if (!priv->broadcast) {
534                 struct ipoib_mcast *broadcast;
535
536                 broadcast = ipoib_mcast_alloc(dev, 1);
537                 if (!broadcast) {
538                         ipoib_warn(priv, "failed to allocate broadcast group\n");
539                         mutex_lock(&mcast_mutex);
540                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
541                                 queue_delayed_work(ipoib_workqueue,
542                                                    &priv->mcast_task, HZ);
543                         mutex_unlock(&mcast_mutex);
544                         return;
545                 }
546
547                 spin_lock_irq(&priv->lock);
548                 memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
549                        sizeof (union ib_gid));
550                 priv->broadcast = broadcast;
551
552                 __ipoib_mcast_add(dev, priv->broadcast);
553                 spin_unlock_irq(&priv->lock);
554         }
555
556         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
557                 ipoib_mcast_join(dev, priv->broadcast, 0);
558                 return;
559         }
560
561         while (1) {
562                 struct ipoib_mcast *mcast = NULL;
563
564                 spin_lock_irq(&priv->lock);
565                 list_for_each_entry(mcast, &priv->multicast_list, list) {
566                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
567                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
568                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
569                                 /* Found the next unjoined group */
570                                 break;
571                         }
572                 }
573                 spin_unlock_irq(&priv->lock);
574
575                 if (&mcast->list == &priv->multicast_list) {
576                         /* All done */
577                         break;
578                 }
579
580                 ipoib_mcast_join(dev, mcast, 1);
581                 return;
582         }
583
584         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
585                 IPOIB_ENCAP_LEN;
586         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
587
588         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
589
590         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
591         netif_carrier_on(dev);
592 }
593
594 int ipoib_mcast_start_thread(struct net_device *dev)
595 {
596         struct ipoib_dev_priv *priv = netdev_priv(dev);
597
598         ipoib_dbg_mcast(priv, "starting multicast thread\n");
599
600         mutex_lock(&mcast_mutex);
601         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
602                 queue_work(ipoib_workqueue, &priv->mcast_task);
603         mutex_unlock(&mcast_mutex);
604
605         spin_lock_irq(&priv->lock);
606         set_bit(IPOIB_MCAST_STARTED, &priv->flags);
607         spin_unlock_irq(&priv->lock);
608
609         return 0;
610 }
611
612 static void wait_for_mcast_join(struct ipoib_dev_priv *priv,
613                                 struct ipoib_mcast *mcast)
614 {
615         spin_lock_irq(&priv->lock);
616         if (mcast && mcast->query) {
617                 ib_sa_cancel_query(mcast->query_id, mcast->query);
618                 mcast->query = NULL;
619                 spin_unlock_irq(&priv->lock);
620                 ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
621                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
622                 wait_for_completion(&mcast->done);
623         }
624         else
625                 spin_unlock_irq(&priv->lock);
626 }
627
628 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
629 {
630         struct ipoib_dev_priv *priv = netdev_priv(dev);
631         struct ipoib_mcast *mcast;
632
633         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
634
635         spin_lock_irq(&priv->lock);
636         clear_bit(IPOIB_MCAST_STARTED, &priv->flags);
637         spin_unlock_irq(&priv->lock);
638
639         mutex_lock(&mcast_mutex);
640         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
641         cancel_delayed_work(&priv->mcast_task);
642         mutex_unlock(&mcast_mutex);
643
644         if (flush)
645                 flush_workqueue(ipoib_workqueue);
646
647         wait_for_mcast_join(priv, priv->broadcast);
648
649         list_for_each_entry(mcast, &priv->multicast_list, list)
650                 wait_for_mcast_join(priv, mcast);
651
652         return 0;
653 }
654
655 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
656 {
657         struct ipoib_dev_priv *priv = netdev_priv(dev);
658         struct ib_sa_mcmember_rec rec = {
659                 .join_state = 1
660         };
661         int ret = 0;
662
663         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
664                 return 0;
665
666         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
667                         IPOIB_GID_ARG(mcast->mcmember.mgid));
668
669         rec.mgid     = mcast->mcmember.mgid;
670         rec.port_gid = priv->local_gid;
671         rec.pkey     = cpu_to_be16(priv->pkey);
672
673         /* Remove ourselves from the multicast group */
674         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
675                                  &mcast->mcmember.mgid);
676         if (ret)
677                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
678
679         /*
680          * Just make one shot at leaving and don't wait for a reply;
681          * if we fail, too bad.
682          */
683         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
684                                         IB_SA_MCMEMBER_REC_MGID         |
685                                         IB_SA_MCMEMBER_REC_PORT_GID     |
686                                         IB_SA_MCMEMBER_REC_PKEY         |
687                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
688                                         0, GFP_ATOMIC, NULL,
689                                         mcast, &mcast->query);
690         if (ret < 0)
691                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
692                            "for leave (result = %d)\n", ret);
693
694         return 0;
695 }
696
697 void ipoib_mcast_send(struct net_device *dev, void *mgid, struct sk_buff *skb)
698 {
699         struct ipoib_dev_priv *priv = netdev_priv(dev);
700         struct ipoib_mcast *mcast;
701
702         /*
703          * We can only be called from ipoib_start_xmit, so we're
704          * inside tx_lock -- no need to save/restore flags.
705          */
706         spin_lock(&priv->lock);
707
708         if (!test_bit(IPOIB_MCAST_STARTED, &priv->flags)        ||
709             !priv->broadcast                                    ||
710             !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
711                 ++priv->stats.tx_dropped;
712                 dev_kfree_skb_any(skb);
713                 goto unlock;
714         }
715
716         mcast = __ipoib_mcast_find(dev, mgid);
717         if (!mcast) {
718                 /* Let's create a new send only group now */
719                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
720                                 IPOIB_GID_FMT "\n", IPOIB_GID_RAW_ARG(mgid));
721
722                 mcast = ipoib_mcast_alloc(dev, 0);
723                 if (!mcast) {
724                         ipoib_warn(priv, "unable to allocate memory for "
725                                    "multicast structure\n");
726                         ++priv->stats.tx_dropped;
727                         dev_kfree_skb_any(skb);
728                         goto out;
729                 }
730
731                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
732                 memcpy(mcast->mcmember.mgid.raw, mgid, sizeof (union ib_gid));
733                 __ipoib_mcast_add(dev, mcast);
734                 list_add_tail(&mcast->list, &priv->multicast_list);
735         }
736
737         if (!mcast->ah) {
738                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
739                         skb_queue_tail(&mcast->pkt_queue, skb);
740                 else {
741                         ++priv->stats.tx_dropped;
742                         dev_kfree_skb_any(skb);
743                 }
744
745                 if (mcast->query)
746                         ipoib_dbg_mcast(priv, "no address vector, "
747                                         "but multicast join already started\n");
748                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
749                         ipoib_mcast_sendonly_join(mcast);
750
751                 /*
752                  * If lookup completes between here and out:, don't
753                  * want to send packet twice.
754                  */
755                 mcast = NULL;
756         }
757
758 out:
759         if (mcast && mcast->ah) {
760                 if (skb->dst            &&
761                     skb->dst->neighbour &&
762                     !*to_ipoib_neigh(skb->dst->neighbour)) {
763                         struct ipoib_neigh *neigh = ipoib_neigh_alloc(skb->dst->neighbour);
764
765                         if (neigh) {
766                                 kref_get(&mcast->ah->ref);
767                                 neigh->ah       = mcast->ah;
768                                 list_add_tail(&neigh->list, &mcast->neigh_list);
769                         }
770                 }
771
772                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
773         }
774
775 unlock:
776         spin_unlock(&priv->lock);
777 }
778
779 void ipoib_mcast_dev_flush(struct net_device *dev)
780 {
781         struct ipoib_dev_priv *priv = netdev_priv(dev);
782         LIST_HEAD(remove_list);
783         struct ipoib_mcast *mcast, *tmcast;
784         unsigned long flags;
785
786         ipoib_dbg_mcast(priv, "flushing multicast list\n");
787
788         spin_lock_irqsave(&priv->lock, flags);
789
790         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
791                 list_del(&mcast->list);
792                 rb_erase(&mcast->rb_node, &priv->multicast_tree);
793                 list_add_tail(&mcast->list, &remove_list);
794         }
795
796         if (priv->broadcast) {
797                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
798                 list_add_tail(&priv->broadcast->list, &remove_list);
799                 priv->broadcast = NULL;
800         }
801
802         spin_unlock_irqrestore(&priv->lock, flags);
803
804         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
805                 ipoib_mcast_leave(dev, mcast);
806                 ipoib_mcast_free(mcast);
807         }
808 }
809
810 void ipoib_mcast_restart_task(void *dev_ptr)
811 {
812         struct net_device *dev = dev_ptr;
813         struct ipoib_dev_priv *priv = netdev_priv(dev);
814         struct dev_mc_list *mclist;
815         struct ipoib_mcast *mcast, *tmcast;
816         LIST_HEAD(remove_list);
817         unsigned long flags;
818
819         ipoib_dbg_mcast(priv, "restarting multicast task\n");
820
821         ipoib_mcast_stop_thread(dev, 0);
822
823         local_irq_save(flags);
824         netif_tx_lock(dev);
825         spin_lock(&priv->lock);
826
827         /*
828          * Unfortunately, the networking core only gives us a list of all of
829          * the multicast hardware addresses. We need to figure out which ones
830          * are new and which ones have been removed
831          */
832
833         /* Clear out the found flag */
834         list_for_each_entry(mcast, &priv->multicast_list, list)
835                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
836
837         /* Mark all of the entries that are found or don't exist */
838         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
839                 union ib_gid mgid;
840
841                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
842
843                 /* Add in the P_Key */
844                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
845                 mgid.raw[5] = priv->pkey & 0xff;
846
847                 mcast = __ipoib_mcast_find(dev, &mgid);
848                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
849                         struct ipoib_mcast *nmcast;
850
851                         /* Not found or send-only group, let's add a new entry */
852                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
853                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
854
855                         nmcast = ipoib_mcast_alloc(dev, 0);
856                         if (!nmcast) {
857                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
858                                 continue;
859                         }
860
861                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
862
863                         nmcast->mcmember.mgid = mgid;
864
865                         if (mcast) {
866                                 /* Destroy the send only entry */
867                                 list_move_tail(&mcast->list, &remove_list);
868
869                                 rb_replace_node(&mcast->rb_node,
870                                                 &nmcast->rb_node,
871                                                 &priv->multicast_tree);
872                         } else
873                                 __ipoib_mcast_add(dev, nmcast);
874
875                         list_add_tail(&nmcast->list, &priv->multicast_list);
876                 }
877
878                 if (mcast)
879                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
880         }
881
882         /* Remove all of the entries don't exist anymore */
883         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
884                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
885                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
886                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
887                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
888
889                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
890
891                         /* Move to the remove list */
892                         list_move_tail(&mcast->list, &remove_list);
893                 }
894         }
895
896         spin_unlock(&priv->lock);
897         netif_tx_unlock(dev);
898         local_irq_restore(flags);
899
900         /* We have to cancel outside of the spinlock */
901         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
902                 wait_for_mcast_join(priv, mcast);
903                 ipoib_mcast_leave(mcast->dev, mcast);
904                 ipoib_mcast_free(mcast);
905         }
906
907         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
908                 ipoib_mcast_start_thread(dev);
909 }
910
911 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
912
913 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
914 {
915         struct ipoib_mcast_iter *iter;
916
917         iter = kmalloc(sizeof *iter, GFP_KERNEL);
918         if (!iter)
919                 return NULL;
920
921         iter->dev = dev;
922         memset(iter->mgid.raw, 0, 16);
923
924         if (ipoib_mcast_iter_next(iter)) {
925                 kfree(iter);
926                 return NULL;
927         }
928
929         return iter;
930 }
931
932 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
933 {
934         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
935         struct rb_node *n;
936         struct ipoib_mcast *mcast;
937         int ret = 1;
938
939         spin_lock_irq(&priv->lock);
940
941         n = rb_first(&priv->multicast_tree);
942
943         while (n) {
944                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
945
946                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
947                            sizeof (union ib_gid)) < 0) {
948                         iter->mgid      = mcast->mcmember.mgid;
949                         iter->created   = mcast->created;
950                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
951                         iter->complete  = !!mcast->ah;
952                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
953
954                         ret = 0;
955
956                         break;
957                 }
958
959                 n = rb_next(n);
960         }
961
962         spin_unlock_irq(&priv->lock);
963
964         return ret;
965 }
966
967 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
968                            union ib_gid *mgid,
969                            unsigned long *created,
970                            unsigned int *queuelen,
971                            unsigned int *complete,
972                            unsigned int *send_only)
973 {
974         *mgid      = iter->mgid;
975         *created   = iter->created;
976         *queuelen  = iter->queuelen;
977         *complete  = iter->complete;
978         *send_only = iter->send_only;
979 }
980
981 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */