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