[IPoIB] add path record information in debugfs
[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 = kzalloc(sizeof *mcast, can_sleep ? GFP_KERNEL : GFP_ATOMIC);
139         if (!mcast)
140                 return NULL;
141
142         init_completion(&mcast->done);
143
144         mcast->dev = dev;
145         mcast->created = jiffies;
146         mcast->backoff = 1;
147         mcast->logcount = 0;
148
149         INIT_LIST_HEAD(&mcast->list);
150         INIT_LIST_HEAD(&mcast->neigh_list);
151         skb_queue_head_init(&mcast->pkt_queue);
152
153         mcast->ah    = NULL;
154         mcast->query = NULL;
155
156         return mcast;
157 }
158
159 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, union ib_gid *mgid)
160 {
161         struct ipoib_dev_priv *priv = netdev_priv(dev);
162         struct rb_node *n = priv->multicast_tree.rb_node;
163
164         while (n) {
165                 struct ipoib_mcast *mcast;
166                 int ret;
167
168                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
169
170                 ret = memcmp(mgid->raw, mcast->mcmember.mgid.raw,
171                              sizeof (union ib_gid));
172                 if (ret < 0)
173                         n = n->rb_left;
174                 else if (ret > 0)
175                         n = n->rb_right;
176                 else
177                         return mcast;
178         }
179
180         return NULL;
181 }
182
183 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
184 {
185         struct ipoib_dev_priv *priv = netdev_priv(dev);
186         struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
187
188         while (*n) {
189                 struct ipoib_mcast *tmcast;
190                 int ret;
191
192                 pn = *n;
193                 tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
194
195                 ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
196                              sizeof (union ib_gid));
197                 if (ret < 0)
198                         n = &pn->rb_left;
199                 else if (ret > 0)
200                         n = &pn->rb_right;
201                 else
202                         return -EEXIST;
203         }
204
205         rb_link_node(&mcast->rb_node, pn, n);
206         rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
207
208         return 0;
209 }
210
211 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
212                                    struct ib_sa_mcmember_rec *mcmember)
213 {
214         struct net_device *dev = mcast->dev;
215         struct ipoib_dev_priv *priv = netdev_priv(dev);
216         int ret;
217
218         mcast->mcmember = *mcmember;
219
220         /* Set the cached Q_Key before we attach if it's the broadcast group */
221         if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
222                     sizeof (union ib_gid))) {
223                 priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
224                 priv->tx_wr.wr.ud.remote_qkey = priv->qkey;
225         }
226
227         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
228                 if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
229                         ipoib_warn(priv, "multicast group " IPOIB_GID_FMT
230                                    " already attached\n",
231                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
232
233                         return 0;
234                 }
235
236                 ret = ipoib_mcast_attach(dev, be16_to_cpu(mcast->mcmember.mlid),
237                                          &mcast->mcmember.mgid);
238                 if (ret < 0) {
239                         ipoib_warn(priv, "couldn't attach QP to multicast group "
240                                    IPOIB_GID_FMT "\n",
241                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
242
243                         clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
244                         return ret;
245                 }
246         }
247
248         {
249                 struct ib_ah_attr av = {
250                         .dlid          = be16_to_cpu(mcast->mcmember.mlid),
251                         .port_num      = priv->port,
252                         .sl            = mcast->mcmember.sl,
253                         .ah_flags      = IB_AH_GRH,
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                 int path_rate = ib_sa_rate_enum_to_int(mcast->mcmember.rate);
262
263                 av.grh.dgid = mcast->mcmember.mgid;
264
265                 if (path_rate > 0 && priv->local_rate > path_rate)
266                         av.static_rate = (priv->local_rate - 1) / path_rate;
267
268                 ipoib_dbg_mcast(priv, "static_rate %d for local port %dX, mcmember %dX\n",
269                                 av.static_rate, priv->local_rate,
270                                 ib_sa_rate_enum_to_int(mcast->mcmember.rate));
271
272                 mcast->ah = ipoib_create_ah(dev, priv->pd, &av);
273                 if (!mcast->ah) {
274                         ipoib_warn(priv, "ib_address_create failed\n");
275                 } else {
276                         ipoib_dbg_mcast(priv, "MGID " IPOIB_GID_FMT
277                                         " AV %p, LID 0x%04x, SL %d\n",
278                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
279                                         mcast->ah->ah,
280                                         be16_to_cpu(mcast->mcmember.mlid),
281                                         mcast->mcmember.sl);
282                 }
283         }
284
285         /* actually send any queued packets */
286         while (!skb_queue_empty(&mcast->pkt_queue)) {
287                 struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
288
289                 skb->dev = dev;
290
291                 if (!skb->dst || !skb->dst->neighbour) {
292                         /* put pseudoheader back on for next time */
293                         skb_push(skb, sizeof (struct ipoib_pseudoheader));
294                 }
295
296                 if (dev_queue_xmit(skb))
297                         ipoib_warn(priv, "dev_queue_xmit failed to requeue packet\n");
298         }
299
300         return 0;
301 }
302
303 static void
304 ipoib_mcast_sendonly_join_complete(int status,
305                                    struct ib_sa_mcmember_rec *mcmember,
306                                    void *mcast_ptr)
307 {
308         struct ipoib_mcast *mcast = mcast_ptr;
309         struct net_device *dev = mcast->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                 while (!skb_queue_empty(&mcast->pkt_queue)) {
321                         struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
322
323                         skb->dev = dev;
324
325                         dev_kfree_skb_any(skb);
326                 }
327
328                 /* Clear the busy flag so we try again */
329                 clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
330         }
331
332         complete(&mcast->done);
333 }
334
335 static int ipoib_mcast_sendonly_join(struct ipoib_mcast *mcast)
336 {
337         struct net_device *dev = mcast->dev;
338         struct ipoib_dev_priv *priv = netdev_priv(dev);
339         struct ib_sa_mcmember_rec rec = {
340 #if 0                           /* Some SMs don't support send-only yet */
341                 .join_state = 4
342 #else
343                 .join_state = 1
344 #endif
345         };
346         int ret = 0;
347
348         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
349                 ipoib_dbg_mcast(priv, "device shutting down, no multicast joins\n");
350                 return -ENODEV;
351         }
352
353         if (test_and_set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
354                 ipoib_dbg_mcast(priv, "multicast entry busy, skipping\n");
355                 return -EBUSY;
356         }
357
358         rec.mgid     = mcast->mcmember.mgid;
359         rec.port_gid = priv->local_gid;
360         rec.pkey     = cpu_to_be16(priv->pkey);
361
362         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec,
363                                      IB_SA_MCMEMBER_REC_MGID            |
364                                      IB_SA_MCMEMBER_REC_PORT_GID        |
365                                      IB_SA_MCMEMBER_REC_PKEY            |
366                                      IB_SA_MCMEMBER_REC_JOIN_STATE,
367                                      1000, GFP_ATOMIC,
368                                      ipoib_mcast_sendonly_join_complete,
369                                      mcast, &mcast->query);
370         if (ret < 0) {
371                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed (ret = %d)\n",
372                            ret);
373         } else {
374                 ipoib_dbg_mcast(priv, "no multicast record for " IPOIB_GID_FMT
375                                 ", starting join\n",
376                                 IPOIB_GID_ARG(mcast->mcmember.mgid));
377
378                 mcast->query_id = ret;
379         }
380
381         return ret;
382 }
383
384 static void ipoib_mcast_join_complete(int status,
385                                       struct ib_sa_mcmember_rec *mcmember,
386                                       void *mcast_ptr)
387 {
388         struct ipoib_mcast *mcast = mcast_ptr;
389         struct net_device *dev = mcast->dev;
390         struct ipoib_dev_priv *priv = netdev_priv(dev);
391
392         ipoib_dbg_mcast(priv, "join completion for " IPOIB_GID_FMT
393                         " (status %d)\n",
394                         IPOIB_GID_ARG(mcast->mcmember.mgid), status);
395
396         if (!status && !ipoib_mcast_join_finish(mcast, mcmember)) {
397                 mcast->backoff = 1;
398                 down(&mcast_mutex);
399                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
400                         queue_work(ipoib_workqueue, &priv->mcast_task);
401                 up(&mcast_mutex);
402                 complete(&mcast->done);
403                 return;
404         }
405
406         if (status == -EINTR) {
407                 complete(&mcast->done);
408                 return;
409         }
410
411         if (status && mcast->logcount++ < 20) {
412                 if (status == -ETIMEDOUT || status == -EINTR) {
413                         ipoib_dbg_mcast(priv, "multicast join failed for " IPOIB_GID_FMT
414                                         ", status %d\n",
415                                         IPOIB_GID_ARG(mcast->mcmember.mgid),
416                                         status);
417                 } else {
418                         ipoib_warn(priv, "multicast join failed for "
419                                    IPOIB_GID_FMT ", status %d\n",
420                                    IPOIB_GID_ARG(mcast->mcmember.mgid),
421                                    status);
422                 }
423         }
424
425         mcast->backoff *= 2;
426         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
427                 mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
428
429         mcast->query = NULL;
430
431         down(&mcast_mutex);
432         if (test_bit(IPOIB_MCAST_RUN, &priv->flags)) {
433                 if (status == -ETIMEDOUT)
434                         queue_work(ipoib_workqueue, &priv->mcast_task);
435                 else
436                         queue_delayed_work(ipoib_workqueue, &priv->mcast_task,
437                                            mcast->backoff * HZ);
438         } else
439                 complete(&mcast->done);
440         up(&mcast_mutex);
441
442         return;
443 }
444
445 static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast,
446                              int create)
447 {
448         struct ipoib_dev_priv *priv = netdev_priv(dev);
449         struct ib_sa_mcmember_rec rec = {
450                 .join_state = 1
451         };
452         ib_sa_comp_mask comp_mask;
453         int ret = 0;
454
455         ipoib_dbg_mcast(priv, "joining MGID " IPOIB_GID_FMT "\n",
456                         IPOIB_GID_ARG(mcast->mcmember.mgid));
457
458         rec.mgid     = mcast->mcmember.mgid;
459         rec.port_gid = priv->local_gid;
460         rec.pkey     = cpu_to_be16(priv->pkey);
461
462         comp_mask =
463                 IB_SA_MCMEMBER_REC_MGID         |
464                 IB_SA_MCMEMBER_REC_PORT_GID     |
465                 IB_SA_MCMEMBER_REC_PKEY         |
466                 IB_SA_MCMEMBER_REC_JOIN_STATE;
467
468         if (create) {
469                 comp_mask |=
470                         IB_SA_MCMEMBER_REC_QKEY         |
471                         IB_SA_MCMEMBER_REC_SL           |
472                         IB_SA_MCMEMBER_REC_FLOW_LABEL   |
473                         IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
474
475                 rec.qkey          = priv->broadcast->mcmember.qkey;
476                 rec.sl            = priv->broadcast->mcmember.sl;
477                 rec.flow_label    = priv->broadcast->mcmember.flow_label;
478                 rec.traffic_class = priv->broadcast->mcmember.traffic_class;
479         }
480
481         ret = ib_sa_mcmember_rec_set(priv->ca, priv->port, &rec, comp_mask,
482                                      mcast->backoff * 1000, GFP_ATOMIC,
483                                      ipoib_mcast_join_complete,
484                                      mcast, &mcast->query);
485
486         if (ret < 0) {
487                 ipoib_warn(priv, "ib_sa_mcmember_rec_set failed, status %d\n", ret);
488
489                 mcast->backoff *= 2;
490                 if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
491                         mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
492
493                 down(&mcast_mutex);
494                 if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
495                         queue_delayed_work(ipoib_workqueue,
496                                            &priv->mcast_task,
497                                            mcast->backoff * HZ);
498                 up(&mcast_mutex);
499         } else
500                 mcast->query_id = ret;
501 }
502
503 void ipoib_mcast_join_task(void *dev_ptr)
504 {
505         struct net_device *dev = dev_ptr;
506         struct ipoib_dev_priv *priv = netdev_priv(dev);
507
508         if (!test_bit(IPOIB_MCAST_RUN, &priv->flags))
509                 return;
510
511         if (ib_query_gid(priv->ca, priv->port, 0, &priv->local_gid))
512                 ipoib_warn(priv, "ib_gid_entry_get() failed\n");
513         else
514                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
515
516         {
517                 struct ib_port_attr attr;
518
519                 if (!ib_query_port(priv->ca, priv->port, &attr)) {
520                         priv->local_lid  = attr.lid;
521                         priv->local_rate = attr.active_speed *
522                                 ib_width_enum_to_int(attr.active_width);
523                 } else
524                         ipoib_warn(priv, "ib_query_port failed\n");
525         }
526
527         if (!priv->broadcast) {
528                 priv->broadcast = ipoib_mcast_alloc(dev, 1);
529                 if (!priv->broadcast) {
530                         ipoib_warn(priv, "failed to allocate broadcast group\n");
531                         down(&mcast_mutex);
532                         if (test_bit(IPOIB_MCAST_RUN, &priv->flags))
533                                 queue_delayed_work(ipoib_workqueue,
534                                                    &priv->mcast_task, HZ);
535                         up(&mcast_mutex);
536                         return;
537                 }
538
539                 memcpy(priv->broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
540                        sizeof (union ib_gid));
541
542                 spin_lock_irq(&priv->lock);
543                 __ipoib_mcast_add(dev, priv->broadcast);
544                 spin_unlock_irq(&priv->lock);
545         }
546
547         if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
548                 ipoib_mcast_join(dev, priv->broadcast, 0);
549                 return;
550         }
551
552         while (1) {
553                 struct ipoib_mcast *mcast = NULL;
554
555                 spin_lock_irq(&priv->lock);
556                 list_for_each_entry(mcast, &priv->multicast_list, list) {
557                         if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)
558                             && !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)
559                             && !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
560                                 /* Found the next unjoined group */
561                                 break;
562                         }
563                 }
564                 spin_unlock_irq(&priv->lock);
565
566                 if (&mcast->list == &priv->multicast_list) {
567                         /* All done */
568                         break;
569                 }
570
571                 ipoib_mcast_join(dev, mcast, 1);
572                 return;
573         }
574
575         priv->mcast_mtu = ib_mtu_enum_to_int(priv->broadcast->mcmember.mtu) -
576                 IPOIB_ENCAP_LEN;
577         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
578
579         ipoib_dbg_mcast(priv, "successfully joined all multicast groups\n");
580
581         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
582         netif_carrier_on(dev);
583 }
584
585 int ipoib_mcast_start_thread(struct net_device *dev)
586 {
587         struct ipoib_dev_priv *priv = netdev_priv(dev);
588
589         ipoib_dbg_mcast(priv, "starting multicast thread\n");
590
591         down(&mcast_mutex);
592         if (!test_and_set_bit(IPOIB_MCAST_RUN, &priv->flags))
593                 queue_work(ipoib_workqueue, &priv->mcast_task);
594         up(&mcast_mutex);
595
596         return 0;
597 }
598
599 int ipoib_mcast_stop_thread(struct net_device *dev, int flush)
600 {
601         struct ipoib_dev_priv *priv = netdev_priv(dev);
602         struct ipoib_mcast *mcast;
603
604         ipoib_dbg_mcast(priv, "stopping multicast thread\n");
605
606         down(&mcast_mutex);
607         clear_bit(IPOIB_MCAST_RUN, &priv->flags);
608         cancel_delayed_work(&priv->mcast_task);
609         up(&mcast_mutex);
610
611         if (flush)
612                 flush_workqueue(ipoib_workqueue);
613
614         if (priv->broadcast && priv->broadcast->query) {
615                 ib_sa_cancel_query(priv->broadcast->query_id, priv->broadcast->query);
616                 priv->broadcast->query = NULL;
617                 ipoib_dbg_mcast(priv, "waiting for bcast\n");
618                 wait_for_completion(&priv->broadcast->done);
619         }
620
621         list_for_each_entry(mcast, &priv->multicast_list, list) {
622                 if (mcast->query) {
623                         ib_sa_cancel_query(mcast->query_id, mcast->query);
624                         mcast->query = NULL;
625                         ipoib_dbg_mcast(priv, "waiting for MGID " IPOIB_GID_FMT "\n",
626                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
627                         wait_for_completion(&mcast->done);
628                 }
629         }
630
631         return 0;
632 }
633
634 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
635 {
636         struct ipoib_dev_priv *priv = netdev_priv(dev);
637         struct ib_sa_mcmember_rec rec = {
638                 .join_state = 1
639         };
640         int ret = 0;
641
642         if (!test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags))
643                 return 0;
644
645         ipoib_dbg_mcast(priv, "leaving MGID " IPOIB_GID_FMT "\n",
646                         IPOIB_GID_ARG(mcast->mcmember.mgid));
647
648         rec.mgid     = mcast->mcmember.mgid;
649         rec.port_gid = priv->local_gid;
650         rec.pkey     = cpu_to_be16(priv->pkey);
651
652         /* Remove ourselves from the multicast group */
653         ret = ipoib_mcast_detach(dev, be16_to_cpu(mcast->mcmember.mlid),
654                                  &mcast->mcmember.mgid);
655         if (ret)
656                 ipoib_warn(priv, "ipoib_mcast_detach failed (result = %d)\n", ret);
657
658         /*
659          * Just make one shot at leaving and don't wait for a reply;
660          * if we fail, too bad.
661          */
662         ret = ib_sa_mcmember_rec_delete(priv->ca, priv->port, &rec,
663                                         IB_SA_MCMEMBER_REC_MGID         |
664                                         IB_SA_MCMEMBER_REC_PORT_GID     |
665                                         IB_SA_MCMEMBER_REC_PKEY         |
666                                         IB_SA_MCMEMBER_REC_JOIN_STATE,
667                                         0, GFP_ATOMIC, NULL,
668                                         mcast, &mcast->query);
669         if (ret < 0)
670                 ipoib_warn(priv, "ib_sa_mcmember_rec_delete failed "
671                            "for leave (result = %d)\n", ret);
672
673         return 0;
674 }
675
676 void ipoib_mcast_send(struct net_device *dev, union ib_gid *mgid,
677                       struct sk_buff *skb)
678 {
679         struct ipoib_dev_priv *priv = netdev_priv(dev);
680         struct ipoib_mcast *mcast;
681
682         /*
683          * We can only be called from ipoib_start_xmit, so we're
684          * inside tx_lock -- no need to save/restore flags.
685          */
686         spin_lock(&priv->lock);
687
688         mcast = __ipoib_mcast_find(dev, mgid);
689         if (!mcast) {
690                 /* Let's create a new send only group now */
691                 ipoib_dbg_mcast(priv, "setting up send only multicast group for "
692                                 IPOIB_GID_FMT "\n", IPOIB_GID_ARG(*mgid));
693
694                 mcast = ipoib_mcast_alloc(dev, 0);
695                 if (!mcast) {
696                         ipoib_warn(priv, "unable to allocate memory for "
697                                    "multicast structure\n");
698                         dev_kfree_skb_any(skb);
699                         goto out;
700                 }
701
702                 set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
703                 mcast->mcmember.mgid = *mgid;
704                 __ipoib_mcast_add(dev, mcast);
705                 list_add_tail(&mcast->list, &priv->multicast_list);
706         }
707
708         if (!mcast->ah) {
709                 if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE)
710                         skb_queue_tail(&mcast->pkt_queue, skb);
711                 else
712                         dev_kfree_skb_any(skb);
713
714                 if (mcast->query)
715                         ipoib_dbg_mcast(priv, "no address vector, "
716                                         "but multicast join already started\n");
717                 else if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
718                         ipoib_mcast_sendonly_join(mcast);
719
720                 /*
721                  * If lookup completes between here and out:, don't
722                  * want to send packet twice.
723                  */
724                 mcast = NULL;
725         }
726
727 out:
728         if (mcast && mcast->ah) {
729                 if (skb->dst            &&
730                     skb->dst->neighbour &&
731                     !*to_ipoib_neigh(skb->dst->neighbour)) {
732                         struct ipoib_neigh *neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
733
734                         if (neigh) {
735                                 kref_get(&mcast->ah->ref);
736                                 neigh->ah       = mcast->ah;
737                                 neigh->neighbour = skb->dst->neighbour;
738                                 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
739                                 list_add_tail(&neigh->list, &mcast->neigh_list);
740                         }
741                 }
742
743                 ipoib_send(dev, skb, mcast->ah, IB_MULTICAST_QPN);
744         }
745
746         spin_unlock(&priv->lock);
747 }
748
749 void ipoib_mcast_dev_flush(struct net_device *dev)
750 {
751         struct ipoib_dev_priv *priv = netdev_priv(dev);
752         LIST_HEAD(remove_list);
753         struct ipoib_mcast *mcast, *tmcast, *nmcast;
754         unsigned long flags;
755
756         ipoib_dbg_mcast(priv, "flushing multicast list\n");
757
758         spin_lock_irqsave(&priv->lock, flags);
759         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
760                 nmcast = ipoib_mcast_alloc(dev, 0);
761                 if (nmcast) {
762                         nmcast->flags =
763                                 mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY);
764
765                         nmcast->mcmember.mgid = mcast->mcmember.mgid;
766
767                         /* Add the new group in before the to-be-destroyed group */
768                         list_add_tail(&nmcast->list, &mcast->list);
769                         list_del_init(&mcast->list);
770
771                         rb_replace_node(&mcast->rb_node, &nmcast->rb_node,
772                                         &priv->multicast_tree);
773
774                         list_add_tail(&mcast->list, &remove_list);
775                 } else {
776                         ipoib_warn(priv, "could not reallocate multicast group "
777                                    IPOIB_GID_FMT "\n",
778                                    IPOIB_GID_ARG(mcast->mcmember.mgid));
779                 }
780         }
781
782         if (priv->broadcast) {
783                 nmcast = ipoib_mcast_alloc(dev, 0);
784                 if (nmcast) {
785                         nmcast->mcmember.mgid = priv->broadcast->mcmember.mgid;
786
787                         rb_replace_node(&priv->broadcast->rb_node,
788                                         &nmcast->rb_node,
789                                         &priv->multicast_tree);
790
791                         list_add_tail(&priv->broadcast->list, &remove_list);
792                 }
793
794                 priv->broadcast = nmcast;
795         }
796
797         spin_unlock_irqrestore(&priv->lock, flags);
798
799         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
800                 ipoib_mcast_leave(dev, mcast);
801                 ipoib_mcast_free(mcast);
802         }
803 }
804
805 void ipoib_mcast_dev_down(struct net_device *dev)
806 {
807         struct ipoib_dev_priv *priv = netdev_priv(dev);
808         unsigned long flags;
809
810         /* Delete broadcast since it will be recreated */
811         if (priv->broadcast) {
812                 ipoib_dbg_mcast(priv, "deleting broadcast group\n");
813
814                 spin_lock_irqsave(&priv->lock, flags);
815                 rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
816                 spin_unlock_irqrestore(&priv->lock, flags);
817                 ipoib_mcast_leave(dev, priv->broadcast);
818                 ipoib_mcast_free(priv->broadcast);
819                 priv->broadcast = NULL;
820         }
821 }
822
823 void ipoib_mcast_restart_task(void *dev_ptr)
824 {
825         struct net_device *dev = dev_ptr;
826         struct ipoib_dev_priv *priv = netdev_priv(dev);
827         struct dev_mc_list *mclist;
828         struct ipoib_mcast *mcast, *tmcast;
829         LIST_HEAD(remove_list);
830         unsigned long flags;
831
832         ipoib_dbg_mcast(priv, "restarting multicast task\n");
833
834         ipoib_mcast_stop_thread(dev, 0);
835
836         spin_lock_irqsave(&priv->lock, flags);
837
838         /*
839          * Unfortunately, the networking core only gives us a list of all of
840          * the multicast hardware addresses. We need to figure out which ones
841          * are new and which ones have been removed
842          */
843
844         /* Clear out the found flag */
845         list_for_each_entry(mcast, &priv->multicast_list, list)
846                 clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
847
848         /* Mark all of the entries that are found or don't exist */
849         for (mclist = dev->mc_list; mclist; mclist = mclist->next) {
850                 union ib_gid mgid;
851
852                 memcpy(mgid.raw, mclist->dmi_addr + 4, sizeof mgid);
853
854                 /* Add in the P_Key */
855                 mgid.raw[4] = (priv->pkey >> 8) & 0xff;
856                 mgid.raw[5] = priv->pkey & 0xff;
857
858                 mcast = __ipoib_mcast_find(dev, &mgid);
859                 if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
860                         struct ipoib_mcast *nmcast;
861
862                         /* Not found or send-only group, let's add a new entry */
863                         ipoib_dbg_mcast(priv, "adding multicast entry for mgid "
864                                         IPOIB_GID_FMT "\n", IPOIB_GID_ARG(mgid));
865
866                         nmcast = ipoib_mcast_alloc(dev, 0);
867                         if (!nmcast) {
868                                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
869                                 continue;
870                         }
871
872                         set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
873
874                         nmcast->mcmember.mgid = mgid;
875
876                         if (mcast) {
877                                 /* Destroy the send only entry */
878                                 list_del(&mcast->list);
879                                 list_add_tail(&mcast->list, &remove_list);
880
881                                 rb_replace_node(&mcast->rb_node,
882                                                 &nmcast->rb_node,
883                                                 &priv->multicast_tree);
884                         } else
885                                 __ipoib_mcast_add(dev, nmcast);
886
887                         list_add_tail(&nmcast->list, &priv->multicast_list);
888                 }
889
890                 if (mcast)
891                         set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
892         }
893
894         /* Remove all of the entries don't exist anymore */
895         list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
896                 if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
897                     !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
898                         ipoib_dbg_mcast(priv, "deleting multicast group " IPOIB_GID_FMT "\n",
899                                         IPOIB_GID_ARG(mcast->mcmember.mgid));
900
901                         rb_erase(&mcast->rb_node, &priv->multicast_tree);
902
903                         /* Move to the remove list */
904                         list_del(&mcast->list);
905                         list_add_tail(&mcast->list, &remove_list);
906                 }
907         }
908         spin_unlock_irqrestore(&priv->lock, flags);
909
910         /* We have to cancel outside of the spinlock */
911         list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
912                 ipoib_mcast_leave(mcast->dev, mcast);
913                 ipoib_mcast_free(mcast);
914         }
915
916         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
917                 ipoib_mcast_start_thread(dev);
918 }
919
920 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
921
922 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
923 {
924         struct ipoib_mcast_iter *iter;
925
926         iter = kmalloc(sizeof *iter, GFP_KERNEL);
927         if (!iter)
928                 return NULL;
929
930         iter->dev = dev;
931         memset(iter->mgid.raw, 0, 16);
932
933         if (ipoib_mcast_iter_next(iter)) {
934                 kfree(iter);
935                 return NULL;
936         }
937
938         return iter;
939 }
940
941 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
942 {
943         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
944         struct rb_node *n;
945         struct ipoib_mcast *mcast;
946         int ret = 1;
947
948         spin_lock_irq(&priv->lock);
949
950         n = rb_first(&priv->multicast_tree);
951
952         while (n) {
953                 mcast = rb_entry(n, struct ipoib_mcast, rb_node);
954
955                 if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
956                            sizeof (union ib_gid)) < 0) {
957                         iter->mgid      = mcast->mcmember.mgid;
958                         iter->created   = mcast->created;
959                         iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
960                         iter->complete  = !!mcast->ah;
961                         iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
962
963                         ret = 0;
964
965                         break;
966                 }
967
968                 n = rb_next(n);
969         }
970
971         spin_unlock_irq(&priv->lock);
972
973         return ret;
974 }
975
976 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
977                            union ib_gid *mgid,
978                            unsigned long *created,
979                            unsigned int *queuelen,
980                            unsigned int *complete,
981                            unsigned int *send_only)
982 {
983         *mgid      = iter->mgid;
984         *created   = iter->created;
985         *queuelen  = iter->queuelen;
986         *complete  = iter->complete;
987         *send_only = iter->send_only;
988 }
989
990 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */