Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_main.c
1 /*
2  * Copyright (c) 2004 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_main.c 1377 2004-12-23 19:57:12Z roland $
35  */
36
37 #include "ipoib.h"
38
39 #include <linux/module.h>
40
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/kernel.h>
44
45 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
46
47 #include <linux/ip.h>
48 #include <linux/in.h>
49
50 #include <net/dst.h>
51
52 MODULE_AUTHOR("Roland Dreier");
53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
54 MODULE_LICENSE("Dual BSD/GPL");
55
56 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
58
59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
63
64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
65 int ipoib_debug_level;
66
67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
69 #endif
70
71 struct ipoib_path_iter {
72         struct net_device *dev;
73         struct ipoib_path  path;
74 };
75
76 static const u8 ipv4_bcast_addr[] = {
77         0x00, 0xff, 0xff, 0xff,
78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
80 };
81
82 struct workqueue_struct *ipoib_workqueue;
83
84 struct ib_sa_client ipoib_sa_client;
85
86 static void ipoib_add_one(struct ib_device *device);
87 static void ipoib_remove_one(struct ib_device *device);
88
89 static struct ib_client ipoib_client = {
90         .name   = "ipoib",
91         .add    = ipoib_add_one,
92         .remove = ipoib_remove_one
93 };
94
95 int ipoib_open(struct net_device *dev)
96 {
97         struct ipoib_dev_priv *priv = netdev_priv(dev);
98
99         ipoib_dbg(priv, "bringing up interface\n");
100
101         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
102
103         if (ipoib_pkey_dev_delay_open(dev))
104                 return 0;
105
106         if (ipoib_ib_dev_open(dev))
107                 return -EINVAL;
108
109         if (ipoib_ib_dev_up(dev)) {
110                 ipoib_ib_dev_stop(dev);
111                 return -EINVAL;
112         }
113
114         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
115                 struct ipoib_dev_priv *cpriv;
116
117                 /* Bring up any child interfaces too */
118                 mutex_lock(&priv->vlan_mutex);
119                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
120                         int flags;
121
122                         flags = cpriv->dev->flags;
123                         if (flags & IFF_UP)
124                                 continue;
125
126                         dev_change_flags(cpriv->dev, flags | IFF_UP);
127                 }
128                 mutex_unlock(&priv->vlan_mutex);
129         }
130
131         netif_start_queue(dev);
132
133         return 0;
134 }
135
136 static int ipoib_stop(struct net_device *dev)
137 {
138         struct ipoib_dev_priv *priv = netdev_priv(dev);
139
140         ipoib_dbg(priv, "stopping interface\n");
141
142         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
143
144         netif_stop_queue(dev);
145
146         clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
147
148         /*
149          * Now flush workqueue to make sure a scheduled task doesn't
150          * bring our internal state back up.
151          */
152         flush_workqueue(ipoib_workqueue);
153
154         ipoib_ib_dev_down(dev, 1);
155         ipoib_ib_dev_stop(dev);
156
157         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
158                 struct ipoib_dev_priv *cpriv;
159
160                 /* Bring down any child interfaces too */
161                 mutex_lock(&priv->vlan_mutex);
162                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
163                         int flags;
164
165                         flags = cpriv->dev->flags;
166                         if (!(flags & IFF_UP))
167                                 continue;
168
169                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
170                 }
171                 mutex_unlock(&priv->vlan_mutex);
172         }
173
174         return 0;
175 }
176
177 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
178 {
179         struct ipoib_dev_priv *priv = netdev_priv(dev);
180
181         /* dev->mtu > 2K ==> connected mode */
182         if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
183                 if (new_mtu > priv->mcast_mtu)
184                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
185                                    priv->mcast_mtu);
186                 dev->mtu = new_mtu;
187                 return 0;
188         }
189
190         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
191                 return -EINVAL;
192         }
193
194         priv->admin_mtu = new_mtu;
195
196         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
197
198         return 0;
199 }
200
201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
202 {
203         struct ipoib_dev_priv *priv = netdev_priv(dev);
204         struct rb_node *n = priv->path_tree.rb_node;
205         struct ipoib_path *path;
206         int ret;
207
208         while (n) {
209                 path = rb_entry(n, struct ipoib_path, rb_node);
210
211                 ret = memcmp(gid, path->pathrec.dgid.raw,
212                              sizeof (union ib_gid));
213
214                 if (ret < 0)
215                         n = n->rb_left;
216                 else if (ret > 0)
217                         n = n->rb_right;
218                 else
219                         return path;
220         }
221
222         return NULL;
223 }
224
225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
226 {
227         struct ipoib_dev_priv *priv = netdev_priv(dev);
228         struct rb_node **n = &priv->path_tree.rb_node;
229         struct rb_node *pn = NULL;
230         struct ipoib_path *tpath;
231         int ret;
232
233         while (*n) {
234                 pn = *n;
235                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
236
237                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
238                              sizeof (union ib_gid));
239                 if (ret < 0)
240                         n = &pn->rb_left;
241                 else if (ret > 0)
242                         n = &pn->rb_right;
243                 else
244                         return -EEXIST;
245         }
246
247         rb_link_node(&path->rb_node, pn, n);
248         rb_insert_color(&path->rb_node, &priv->path_tree);
249
250         list_add_tail(&path->list, &priv->path_list);
251
252         return 0;
253 }
254
255 static void path_free(struct net_device *dev, struct ipoib_path *path)
256 {
257         struct ipoib_dev_priv *priv = netdev_priv(dev);
258         struct ipoib_neigh *neigh, *tn;
259         struct sk_buff *skb;
260         unsigned long flags;
261
262         while ((skb = __skb_dequeue(&path->queue)))
263                 dev_kfree_skb_irq(skb);
264
265         spin_lock_irqsave(&priv->lock, flags);
266
267         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
268                 /*
269                  * It's safe to call ipoib_put_ah() inside priv->lock
270                  * here, because we know that path->ah will always
271                  * hold one more reference, so ipoib_put_ah() will
272                  * never do more than decrement the ref count.
273                  */
274                 if (neigh->ah)
275                         ipoib_put_ah(neigh->ah);
276
277                 ipoib_neigh_free(dev, neigh);
278         }
279
280         spin_unlock_irqrestore(&priv->lock, flags);
281
282         if (path->ah)
283                 ipoib_put_ah(path->ah);
284
285         kfree(path);
286 }
287
288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
289
290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
291 {
292         struct ipoib_path_iter *iter;
293
294         iter = kmalloc(sizeof *iter, GFP_KERNEL);
295         if (!iter)
296                 return NULL;
297
298         iter->dev = dev;
299         memset(iter->path.pathrec.dgid.raw, 0, 16);
300
301         if (ipoib_path_iter_next(iter)) {
302                 kfree(iter);
303                 return NULL;
304         }
305
306         return iter;
307 }
308
309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
310 {
311         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
312         struct rb_node *n;
313         struct ipoib_path *path;
314         int ret = 1;
315
316         spin_lock_irq(&priv->lock);
317
318         n = rb_first(&priv->path_tree);
319
320         while (n) {
321                 path = rb_entry(n, struct ipoib_path, rb_node);
322
323                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
324                            sizeof (union ib_gid)) < 0) {
325                         iter->path = *path;
326                         ret = 0;
327                         break;
328                 }
329
330                 n = rb_next(n);
331         }
332
333         spin_unlock_irq(&priv->lock);
334
335         return ret;
336 }
337
338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
339                           struct ipoib_path *path)
340 {
341         *path = iter->path;
342 }
343
344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
345
346 void ipoib_flush_paths(struct net_device *dev)
347 {
348         struct ipoib_dev_priv *priv = netdev_priv(dev);
349         struct ipoib_path *path, *tp;
350         LIST_HEAD(remove_list);
351
352         spin_lock_irq(&priv->tx_lock);
353         spin_lock(&priv->lock);
354
355         list_splice(&priv->path_list, &remove_list);
356         INIT_LIST_HEAD(&priv->path_list);
357
358         list_for_each_entry(path, &remove_list, list)
359                 rb_erase(&path->rb_node, &priv->path_tree);
360
361         list_for_each_entry_safe(path, tp, &remove_list, list) {
362                 if (path->query)
363                         ib_sa_cancel_query(path->query_id, path->query);
364                 spin_unlock(&priv->lock);
365                 spin_unlock_irq(&priv->tx_lock);
366                 wait_for_completion(&path->done);
367                 path_free(dev, path);
368                 spin_lock_irq(&priv->tx_lock);
369                 spin_lock(&priv->lock);
370         }
371         spin_unlock(&priv->lock);
372         spin_unlock_irq(&priv->tx_lock);
373 }
374
375 static void path_rec_completion(int status,
376                                 struct ib_sa_path_rec *pathrec,
377                                 void *path_ptr)
378 {
379         struct ipoib_path *path = path_ptr;
380         struct net_device *dev = path->dev;
381         struct ipoib_dev_priv *priv = netdev_priv(dev);
382         struct ipoib_ah *ah = NULL;
383         struct ipoib_neigh *neigh, *tn;
384         struct sk_buff_head skqueue;
385         struct sk_buff *skb;
386         unsigned long flags;
387
388         if (!status)
389                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
390                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
391         else
392                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
393                           status, IPOIB_GID_ARG(path->pathrec.dgid));
394
395         skb_queue_head_init(&skqueue);
396
397         if (!status) {
398                 struct ib_ah_attr av;
399
400                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
401                         ah = ipoib_create_ah(dev, priv->pd, &av);
402         }
403
404         spin_lock_irqsave(&priv->lock, flags);
405
406         path->ah = ah;
407
408         if (ah) {
409                 path->pathrec = *pathrec;
410
411                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
412                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
413
414                 while ((skb = __skb_dequeue(&path->queue)))
415                         __skb_queue_tail(&skqueue, skb);
416
417                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
418                         kref_get(&path->ah->ref);
419                         neigh->ah = path->ah;
420                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
421                                sizeof(union ib_gid));
422
423                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
424                                 if (!ipoib_cm_get(neigh))
425                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
426                                                                                path,
427                                                                                neigh));
428                                 if (!ipoib_cm_get(neigh)) {
429                                         list_del(&neigh->list);
430                                         if (neigh->ah)
431                                                 ipoib_put_ah(neigh->ah);
432                                         ipoib_neigh_free(dev, neigh);
433                                         continue;
434                                 }
435                         }
436
437                         while ((skb = __skb_dequeue(&neigh->queue)))
438                                 __skb_queue_tail(&skqueue, skb);
439                 }
440         }
441
442         path->query = NULL;
443         complete(&path->done);
444
445         spin_unlock_irqrestore(&priv->lock, flags);
446
447         while ((skb = __skb_dequeue(&skqueue))) {
448                 skb->dev = dev;
449                 if (dev_queue_xmit(skb))
450                         ipoib_warn(priv, "dev_queue_xmit failed "
451                                    "to requeue packet\n");
452         }
453 }
454
455 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
456 {
457         struct ipoib_dev_priv *priv = netdev_priv(dev);
458         struct ipoib_path *path;
459
460         path = kzalloc(sizeof *path, GFP_ATOMIC);
461         if (!path)
462                 return NULL;
463
464         path->dev = dev;
465
466         skb_queue_head_init(&path->queue);
467
468         INIT_LIST_HEAD(&path->neigh_list);
469
470         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
471         path->pathrec.sgid      = priv->local_gid;
472         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
473         path->pathrec.numb_path = 1;
474
475         return path;
476 }
477
478 static int path_rec_start(struct net_device *dev,
479                           struct ipoib_path *path)
480 {
481         struct ipoib_dev_priv *priv = netdev_priv(dev);
482
483         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
484                   IPOIB_GID_ARG(path->pathrec.dgid));
485
486         init_completion(&path->done);
487
488         path->query_id =
489                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
490                                    &path->pathrec,
491                                    IB_SA_PATH_REC_DGID          |
492                                    IB_SA_PATH_REC_SGID          |
493                                    IB_SA_PATH_REC_NUMB_PATH     |
494                                    IB_SA_PATH_REC_PKEY,
495                                    1000, GFP_ATOMIC,
496                                    path_rec_completion,
497                                    path, &path->query);
498         if (path->query_id < 0) {
499                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
500                 path->query = NULL;
501                 return path->query_id;
502         }
503
504         return 0;
505 }
506
507 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
508 {
509         struct ipoib_dev_priv *priv = netdev_priv(dev);
510         struct ipoib_path *path;
511         struct ipoib_neigh *neigh;
512
513         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
514         if (!neigh) {
515                 ++priv->stats.tx_dropped;
516                 dev_kfree_skb_any(skb);
517                 return;
518         }
519
520         /*
521          * We can only be called from ipoib_start_xmit, so we're
522          * inside tx_lock -- no need to save/restore flags.
523          */
524         spin_lock(&priv->lock);
525
526         path = __path_find(dev, skb->dst->neighbour->ha + 4);
527         if (!path) {
528                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
529                 if (!path)
530                         goto err_path;
531
532                 __path_add(dev, path);
533         }
534
535         list_add_tail(&neigh->list, &path->neigh_list);
536
537         if (path->ah) {
538                 kref_get(&path->ah->ref);
539                 neigh->ah = path->ah;
540                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
541                        sizeof(union ib_gid));
542
543                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
544                         if (!ipoib_cm_get(neigh))
545                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
546                         if (!ipoib_cm_get(neigh)) {
547                                 list_del(&neigh->list);
548                                 if (neigh->ah)
549                                         ipoib_put_ah(neigh->ah);
550                                 ipoib_neigh_free(dev, neigh);
551                                 goto err_drop;
552                         }
553                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
554                                 __skb_queue_tail(&neigh->queue, skb);
555                         else {
556                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
557                                            skb_queue_len(&neigh->queue));
558                                 goto err_drop;
559                         }
560                 } else
561                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
562         } else {
563                 neigh->ah  = NULL;
564
565                 if (!path->query && path_rec_start(dev, path))
566                         goto err_list;
567
568                 __skb_queue_tail(&neigh->queue, skb);
569         }
570
571         spin_unlock(&priv->lock);
572         return;
573
574 err_list:
575         list_del(&neigh->list);
576
577 err_path:
578         ipoib_neigh_free(dev, neigh);
579 err_drop:
580         ++priv->stats.tx_dropped;
581         dev_kfree_skb_any(skb);
582
583         spin_unlock(&priv->lock);
584 }
585
586 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
587 {
588         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
589
590         /* Look up path record for unicasts */
591         if (skb->dst->neighbour->ha[4] != 0xff) {
592                 neigh_add_path(skb, dev);
593                 return;
594         }
595
596         /* Add in the P_Key for multicasts */
597         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
598         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
599         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
600 }
601
602 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
603                              struct ipoib_pseudoheader *phdr)
604 {
605         struct ipoib_dev_priv *priv = netdev_priv(dev);
606         struct ipoib_path *path;
607
608         /*
609          * We can only be called from ipoib_start_xmit, so we're
610          * inside tx_lock -- no need to save/restore flags.
611          */
612         spin_lock(&priv->lock);
613
614         path = __path_find(dev, phdr->hwaddr + 4);
615         if (!path) {
616                 path = path_rec_create(dev, phdr->hwaddr + 4);
617                 if (path) {
618                         /* put pseudoheader back on for next time */
619                         skb_push(skb, sizeof *phdr);
620                         __skb_queue_tail(&path->queue, skb);
621
622                         if (path_rec_start(dev, path)) {
623                                 spin_unlock(&priv->lock);
624                                 path_free(dev, path);
625                                 return;
626                         } else
627                                 __path_add(dev, path);
628                 } else {
629                         ++priv->stats.tx_dropped;
630                         dev_kfree_skb_any(skb);
631                 }
632
633                 spin_unlock(&priv->lock);
634                 return;
635         }
636
637         if (path->ah) {
638                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
639                           be16_to_cpu(path->pathrec.dlid));
640
641                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
642         } else if ((path->query || !path_rec_start(dev, path)) &&
643                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
644                 /* put pseudoheader back on for next time */
645                 skb_push(skb, sizeof *phdr);
646                 __skb_queue_tail(&path->queue, skb);
647         } else {
648                 ++priv->stats.tx_dropped;
649                 dev_kfree_skb_any(skb);
650         }
651
652         spin_unlock(&priv->lock);
653 }
654
655 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
656 {
657         struct ipoib_dev_priv *priv = netdev_priv(dev);
658         struct ipoib_neigh *neigh;
659         unsigned long flags;
660
661         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
662                 return NETDEV_TX_LOCKED;
663
664         /*
665          * Check if our queue is stopped.  Since we have the LLTX bit
666          * set, we can't rely on netif_stop_queue() preventing our
667          * xmit function from being called with a full queue.
668          */
669         if (unlikely(netif_queue_stopped(dev))) {
670                 spin_unlock_irqrestore(&priv->tx_lock, flags);
671                 return NETDEV_TX_BUSY;
672         }
673
674         if (likely(skb->dst && skb->dst->neighbour)) {
675                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
676                         ipoib_path_lookup(skb, dev);
677                         goto out;
678                 }
679
680                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
681
682                 if (ipoib_cm_get(neigh)) {
683                         if (ipoib_cm_up(neigh)) {
684                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
685                                 goto out;
686                         }
687                 } else if (neigh->ah) {
688                         if (unlikely(memcmp(&neigh->dgid.raw,
689                                             skb->dst->neighbour->ha + 4,
690                                             sizeof(union ib_gid)))) {
691                                 spin_lock(&priv->lock);
692                                 /*
693                                  * It's safe to call ipoib_put_ah() inside
694                                  * priv->lock here, because we know that
695                                  * path->ah will always hold one more reference,
696                                  * so ipoib_put_ah() will never do more than
697                                  * decrement the ref count.
698                                  */
699                                 ipoib_put_ah(neigh->ah);
700                                 list_del(&neigh->list);
701                                 ipoib_neigh_free(dev, neigh);
702                                 spin_unlock(&priv->lock);
703                                 ipoib_path_lookup(skb, dev);
704                                 goto out;
705                         }
706
707                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
708                         goto out;
709                 }
710
711                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
712                         spin_lock(&priv->lock);
713                         __skb_queue_tail(&neigh->queue, skb);
714                         spin_unlock(&priv->lock);
715                 } else {
716                         ++priv->stats.tx_dropped;
717                         dev_kfree_skb_any(skb);
718                 }
719         } else {
720                 struct ipoib_pseudoheader *phdr =
721                         (struct ipoib_pseudoheader *) skb->data;
722                 skb_pull(skb, sizeof *phdr);
723
724                 if (phdr->hwaddr[4] == 0xff) {
725                         /* Add in the P_Key for multicast*/
726                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
727                         phdr->hwaddr[9] = priv->pkey & 0xff;
728
729                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
730                 } else {
731                         /* unicast GID -- should be ARP or RARP reply */
732
733                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
734                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
735                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
736                                            IPOIB_GID_FMT "\n",
737                                            skb->dst ? "neigh" : "dst",
738                                            be16_to_cpup((__be16 *) skb->data),
739                                            IPOIB_QPN(phdr->hwaddr),
740                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
741                                 dev_kfree_skb_any(skb);
742                                 ++priv->stats.tx_dropped;
743                                 goto out;
744                         }
745
746                         unicast_arp_send(skb, dev, phdr);
747                 }
748         }
749
750 out:
751         spin_unlock_irqrestore(&priv->tx_lock, flags);
752
753         return NETDEV_TX_OK;
754 }
755
756 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
757 {
758         struct ipoib_dev_priv *priv = netdev_priv(dev);
759
760         return &priv->stats;
761 }
762
763 static void ipoib_timeout(struct net_device *dev)
764 {
765         struct ipoib_dev_priv *priv = netdev_priv(dev);
766
767         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
768                    jiffies_to_msecs(jiffies - dev->trans_start));
769         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
770                    netif_queue_stopped(dev),
771                    priv->tx_head, priv->tx_tail);
772         /* XXX reset QP, etc. */
773 }
774
775 static int ipoib_hard_header(struct sk_buff *skb,
776                              struct net_device *dev,
777                              unsigned short type,
778                              void *daddr, void *saddr, unsigned len)
779 {
780         struct ipoib_header *header;
781
782         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
783
784         header->proto = htons(type);
785         header->reserved = 0;
786
787         /*
788          * If we don't have a neighbour structure, stuff the
789          * destination address onto the front of the skb so we can
790          * figure out where to send the packet later.
791          */
792         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
793                 struct ipoib_pseudoheader *phdr =
794                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
795                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
796         }
797
798         return 0;
799 }
800
801 static void ipoib_set_mcast_list(struct net_device *dev)
802 {
803         struct ipoib_dev_priv *priv = netdev_priv(dev);
804
805         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
806                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
807                 return;
808         }
809
810         queue_work(ipoib_workqueue, &priv->restart_task);
811 }
812
813 static void ipoib_neigh_cleanup(struct neighbour *n)
814 {
815         struct ipoib_neigh *neigh;
816         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
817         unsigned long flags;
818         struct ipoib_ah *ah = NULL;
819
820         ipoib_dbg(priv,
821                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
822                   IPOIB_QPN(n->ha),
823                   IPOIB_GID_RAW_ARG(n->ha + 4));
824
825         spin_lock_irqsave(&priv->lock, flags);
826
827         neigh = *to_ipoib_neigh(n);
828         if (neigh) {
829                 if (neigh->ah)
830                         ah = neigh->ah;
831                 list_del(&neigh->list);
832                 ipoib_neigh_free(n->dev, neigh);
833         }
834
835         spin_unlock_irqrestore(&priv->lock, flags);
836
837         if (ah)
838                 ipoib_put_ah(ah);
839 }
840
841 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
842 {
843         struct ipoib_neigh *neigh;
844
845         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
846         if (!neigh)
847                 return NULL;
848
849         neigh->neighbour = neighbour;
850         *to_ipoib_neigh(neighbour) = neigh;
851         skb_queue_head_init(&neigh->queue);
852         ipoib_cm_set(neigh, NULL);
853
854         return neigh;
855 }
856
857 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
858 {
859         struct ipoib_dev_priv *priv = netdev_priv(dev);
860         struct sk_buff *skb;
861         *to_ipoib_neigh(neigh->neighbour) = NULL;
862         while ((skb = __skb_dequeue(&neigh->queue))) {
863                 ++priv->stats.tx_dropped;
864                 dev_kfree_skb_any(skb);
865         }
866         if (ipoib_cm_get(neigh))
867                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
868         kfree(neigh);
869 }
870
871 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
872 {
873         parms->neigh_cleanup = ipoib_neigh_cleanup;
874
875         return 0;
876 }
877
878 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
879 {
880         struct ipoib_dev_priv *priv = netdev_priv(dev);
881
882         /* Allocate RX/TX "rings" to hold queued skbs */
883         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
884                                 GFP_KERNEL);
885         if (!priv->rx_ring) {
886                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
887                        ca->name, ipoib_recvq_size);
888                 goto out;
889         }
890
891         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
892                                 GFP_KERNEL);
893         if (!priv->tx_ring) {
894                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
895                        ca->name, ipoib_sendq_size);
896                 goto out_rx_ring_cleanup;
897         }
898
899         /* priv->tx_head & tx_tail are already 0 */
900
901         if (ipoib_ib_dev_init(dev, ca, port))
902                 goto out_tx_ring_cleanup;
903
904         return 0;
905
906 out_tx_ring_cleanup:
907         kfree(priv->tx_ring);
908
909 out_rx_ring_cleanup:
910         kfree(priv->rx_ring);
911
912 out:
913         return -ENOMEM;
914 }
915
916 void ipoib_dev_cleanup(struct net_device *dev)
917 {
918         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
919
920         ipoib_delete_debug_files(dev);
921
922         /* Delete any child interfaces first */
923         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
924                 unregister_netdev(cpriv->dev);
925                 ipoib_dev_cleanup(cpriv->dev);
926                 free_netdev(cpriv->dev);
927         }
928
929         ipoib_ib_dev_cleanup(dev);
930
931         kfree(priv->rx_ring);
932         kfree(priv->tx_ring);
933
934         priv->rx_ring = NULL;
935         priv->tx_ring = NULL;
936 }
937
938 static void ipoib_setup(struct net_device *dev)
939 {
940         struct ipoib_dev_priv *priv = netdev_priv(dev);
941
942         dev->open                = ipoib_open;
943         dev->stop                = ipoib_stop;
944         dev->change_mtu          = ipoib_change_mtu;
945         dev->hard_start_xmit     = ipoib_start_xmit;
946         dev->get_stats           = ipoib_get_stats;
947         dev->tx_timeout          = ipoib_timeout;
948         dev->hard_header         = ipoib_hard_header;
949         dev->set_multicast_list  = ipoib_set_mcast_list;
950         dev->neigh_setup         = ipoib_neigh_setup_dev;
951
952         dev->watchdog_timeo      = HZ;
953
954         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
955
956         /*
957          * We add in INFINIBAND_ALEN to allow for the destination
958          * address "pseudoheader" for skbs without neighbour struct.
959          */
960         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
961         dev->addr_len            = INFINIBAND_ALEN;
962         dev->type                = ARPHRD_INFINIBAND;
963         dev->tx_queue_len        = ipoib_sendq_size * 2;
964         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
965
966         /* MTU will be reset when mcast join happens */
967         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
968         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
969
970         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
971
972         netif_carrier_off(dev);
973
974         SET_MODULE_OWNER(dev);
975
976         priv->dev = dev;
977
978         spin_lock_init(&priv->lock);
979         spin_lock_init(&priv->tx_lock);
980
981         mutex_init(&priv->mcast_mutex);
982         mutex_init(&priv->vlan_mutex);
983
984         INIT_LIST_HEAD(&priv->path_list);
985         INIT_LIST_HEAD(&priv->child_intfs);
986         INIT_LIST_HEAD(&priv->dead_ahs);
987         INIT_LIST_HEAD(&priv->multicast_list);
988
989         INIT_DELAYED_WORK(&priv->pkey_task,    ipoib_pkey_poll);
990         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
991         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
992         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
993         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
994 }
995
996 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
997 {
998         struct net_device *dev;
999
1000         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1001                            ipoib_setup);
1002         if (!dev)
1003                 return NULL;
1004
1005         return netdev_priv(dev);
1006 }
1007
1008 static ssize_t show_pkey(struct device *dev,
1009                          struct device_attribute *attr, char *buf)
1010 {
1011         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1012
1013         return sprintf(buf, "0x%04x\n", priv->pkey);
1014 }
1015 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1016
1017 static ssize_t create_child(struct device *dev,
1018                             struct device_attribute *attr,
1019                             const char *buf, size_t count)
1020 {
1021         int pkey;
1022         int ret;
1023
1024         if (sscanf(buf, "%i", &pkey) != 1)
1025                 return -EINVAL;
1026
1027         if (pkey < 0 || pkey > 0xffff)
1028                 return -EINVAL;
1029
1030         /*
1031          * Set the full membership bit, so that we join the right
1032          * broadcast group, etc.
1033          */
1034         pkey |= 0x8000;
1035
1036         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1037
1038         return ret ? ret : count;
1039 }
1040 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1041
1042 static ssize_t delete_child(struct device *dev,
1043                             struct device_attribute *attr,
1044                             const char *buf, size_t count)
1045 {
1046         int pkey;
1047         int ret;
1048
1049         if (sscanf(buf, "%i", &pkey) != 1)
1050                 return -EINVAL;
1051
1052         if (pkey < 0 || pkey > 0xffff)
1053                 return -EINVAL;
1054
1055         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1056
1057         return ret ? ret : count;
1058
1059 }
1060 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1061
1062 int ipoib_add_pkey_attr(struct net_device *dev)
1063 {
1064         return device_create_file(&dev->dev, &dev_attr_pkey);
1065 }
1066
1067 static struct net_device *ipoib_add_port(const char *format,
1068                                          struct ib_device *hca, u8 port)
1069 {
1070         struct ipoib_dev_priv *priv;
1071         int result = -ENOMEM;
1072
1073         priv = ipoib_intf_alloc(format);
1074         if (!priv)
1075                 goto alloc_mem_failed;
1076
1077         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1078
1079         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1080         if (result) {
1081                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1082                        hca->name, port, result);
1083                 goto alloc_mem_failed;
1084         }
1085
1086         /*
1087          * Set the full membership bit, so that we join the right
1088          * broadcast group, etc.
1089          */
1090         priv->pkey |= 0x8000;
1091
1092         priv->dev->broadcast[8] = priv->pkey >> 8;
1093         priv->dev->broadcast[9] = priv->pkey & 0xff;
1094
1095         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1096         if (result) {
1097                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1098                        hca->name, port, result);
1099                 goto alloc_mem_failed;
1100         } else
1101                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1102
1103
1104         result = ipoib_dev_init(priv->dev, hca, port);
1105         if (result < 0) {
1106                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1107                        hca->name, port, result);
1108                 goto device_init_failed;
1109         }
1110
1111         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1112                               priv->ca, ipoib_event);
1113         result = ib_register_event_handler(&priv->event_handler);
1114         if (result < 0) {
1115                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1116                        "port %d (ret = %d)\n",
1117                        hca->name, port, result);
1118                 goto event_failed;
1119         }
1120
1121         result = register_netdev(priv->dev);
1122         if (result) {
1123                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1124                        hca->name, port, result);
1125                 goto register_failed;
1126         }
1127
1128         ipoib_create_debug_files(priv->dev);
1129
1130         if (ipoib_cm_add_mode_attr(priv->dev))
1131                 goto sysfs_failed;
1132         if (ipoib_add_pkey_attr(priv->dev))
1133                 goto sysfs_failed;
1134         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1135                 goto sysfs_failed;
1136         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1137                 goto sysfs_failed;
1138
1139         return priv->dev;
1140
1141 sysfs_failed:
1142         ipoib_delete_debug_files(priv->dev);
1143         unregister_netdev(priv->dev);
1144
1145 register_failed:
1146         ib_unregister_event_handler(&priv->event_handler);
1147         flush_scheduled_work();
1148
1149 event_failed:
1150         ipoib_dev_cleanup(priv->dev);
1151
1152 device_init_failed:
1153         free_netdev(priv->dev);
1154
1155 alloc_mem_failed:
1156         return ERR_PTR(result);
1157 }
1158
1159 static void ipoib_add_one(struct ib_device *device)
1160 {
1161         struct list_head *dev_list;
1162         struct net_device *dev;
1163         struct ipoib_dev_priv *priv;
1164         int s, e, p;
1165
1166         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1167                 return;
1168
1169         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1170         if (!dev_list)
1171                 return;
1172
1173         INIT_LIST_HEAD(dev_list);
1174
1175         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1176                 s = 0;
1177                 e = 0;
1178         } else {
1179                 s = 1;
1180                 e = device->phys_port_cnt;
1181         }
1182
1183         for (p = s; p <= e; ++p) {
1184                 dev = ipoib_add_port("ib%d", device, p);
1185                 if (!IS_ERR(dev)) {
1186                         priv = netdev_priv(dev);
1187                         list_add_tail(&priv->list, dev_list);
1188                 }
1189         }
1190
1191         ib_set_client_data(device, &ipoib_client, dev_list);
1192 }
1193
1194 static void ipoib_remove_one(struct ib_device *device)
1195 {
1196         struct ipoib_dev_priv *priv, *tmp;
1197         struct list_head *dev_list;
1198
1199         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1200                 return;
1201
1202         dev_list = ib_get_client_data(device, &ipoib_client);
1203
1204         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1205                 ib_unregister_event_handler(&priv->event_handler);
1206                 flush_scheduled_work();
1207
1208                 unregister_netdev(priv->dev);
1209                 ipoib_dev_cleanup(priv->dev);
1210                 free_netdev(priv->dev);
1211         }
1212
1213         kfree(dev_list);
1214 }
1215
1216 static int __init ipoib_init_module(void)
1217 {
1218         int ret;
1219
1220         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1221         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1222         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1223
1224         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1225         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1226         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1227
1228         ret = ipoib_register_debugfs();
1229         if (ret)
1230                 return ret;
1231
1232         /*
1233          * We create our own workqueue mainly because we want to be
1234          * able to flush it when devices are being removed.  We can't
1235          * use schedule_work()/flush_scheduled_work() because both
1236          * unregister_netdev() and linkwatch_event take the rtnl lock,
1237          * so flush_scheduled_work() can deadlock during device
1238          * removal.
1239          */
1240         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1241         if (!ipoib_workqueue) {
1242                 ret = -ENOMEM;
1243                 goto err_fs;
1244         }
1245
1246         ib_sa_register_client(&ipoib_sa_client);
1247
1248         ret = ib_register_client(&ipoib_client);
1249         if (ret)
1250                 goto err_sa;
1251
1252         return 0;
1253
1254 err_sa:
1255         ib_sa_unregister_client(&ipoib_sa_client);
1256         destroy_workqueue(ipoib_workqueue);
1257
1258 err_fs:
1259         ipoib_unregister_debugfs();
1260
1261         return ret;
1262 }
1263
1264 static void __exit ipoib_cleanup_module(void)
1265 {
1266         ib_unregister_client(&ipoib_client);
1267         ib_sa_unregister_client(&ipoib_sa_client);
1268         ipoib_unregister_debugfs();
1269         destroy_workqueue(ipoib_workqueue);
1270 }
1271
1272 module_init(ipoib_init_module);
1273 module_exit(ipoib_cleanup_module);