Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[sfrench/cifs-2.6.git] / 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         napi_enable(&priv->napi);
102         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
103
104         if (ipoib_pkey_dev_delay_open(dev))
105                 return 0;
106
107         if (ipoib_ib_dev_open(dev)) {
108                 napi_disable(&priv->napi);
109                 return -EINVAL;
110         }
111
112         if (ipoib_ib_dev_up(dev)) {
113                 ipoib_ib_dev_stop(dev, 1);
114                 napi_disable(&priv->napi);
115                 return -EINVAL;
116         }
117
118         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
119                 struct ipoib_dev_priv *cpriv;
120
121                 /* Bring up any child interfaces too */
122                 mutex_lock(&priv->vlan_mutex);
123                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
124                         int flags;
125
126                         flags = cpriv->dev->flags;
127                         if (flags & IFF_UP)
128                                 continue;
129
130                         dev_change_flags(cpriv->dev, flags | IFF_UP);
131                 }
132                 mutex_unlock(&priv->vlan_mutex);
133         }
134
135         netif_start_queue(dev);
136
137         return 0;
138 }
139
140 static int ipoib_stop(struct net_device *dev)
141 {
142         struct ipoib_dev_priv *priv = netdev_priv(dev);
143
144         ipoib_dbg(priv, "stopping interface\n");
145
146         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
147         napi_disable(&priv->napi);
148
149         netif_stop_queue(dev);
150
151         /*
152          * Now flush workqueue to make sure a scheduled task doesn't
153          * bring our internal state back up.
154          */
155         flush_workqueue(ipoib_workqueue);
156
157         ipoib_ib_dev_down(dev, 1);
158         ipoib_ib_dev_stop(dev, 1);
159
160         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
161                 struct ipoib_dev_priv *cpriv;
162
163                 /* Bring down any child interfaces too */
164                 mutex_lock(&priv->vlan_mutex);
165                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
166                         int flags;
167
168                         flags = cpriv->dev->flags;
169                         if (!(flags & IFF_UP))
170                                 continue;
171
172                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
173                 }
174                 mutex_unlock(&priv->vlan_mutex);
175         }
176
177         return 0;
178 }
179
180 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
181 {
182         struct ipoib_dev_priv *priv = netdev_priv(dev);
183
184         /* dev->mtu > 2K ==> connected mode */
185         if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
186                 if (new_mtu > priv->mcast_mtu)
187                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
188                                    priv->mcast_mtu);
189                 dev->mtu = new_mtu;
190                 return 0;
191         }
192
193         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
194                 return -EINVAL;
195         }
196
197         priv->admin_mtu = new_mtu;
198
199         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
200
201         return 0;
202 }
203
204 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
205 {
206         struct ipoib_dev_priv *priv = netdev_priv(dev);
207         struct rb_node *n = priv->path_tree.rb_node;
208         struct ipoib_path *path;
209         int ret;
210
211         while (n) {
212                 path = rb_entry(n, struct ipoib_path, rb_node);
213
214                 ret = memcmp(gid, path->pathrec.dgid.raw,
215                              sizeof (union ib_gid));
216
217                 if (ret < 0)
218                         n = n->rb_left;
219                 else if (ret > 0)
220                         n = n->rb_right;
221                 else
222                         return path;
223         }
224
225         return NULL;
226 }
227
228 static int __path_add(struct net_device *dev, struct ipoib_path *path)
229 {
230         struct ipoib_dev_priv *priv = netdev_priv(dev);
231         struct rb_node **n = &priv->path_tree.rb_node;
232         struct rb_node *pn = NULL;
233         struct ipoib_path *tpath;
234         int ret;
235
236         while (*n) {
237                 pn = *n;
238                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
239
240                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
241                              sizeof (union ib_gid));
242                 if (ret < 0)
243                         n = &pn->rb_left;
244                 else if (ret > 0)
245                         n = &pn->rb_right;
246                 else
247                         return -EEXIST;
248         }
249
250         rb_link_node(&path->rb_node, pn, n);
251         rb_insert_color(&path->rb_node, &priv->path_tree);
252
253         list_add_tail(&path->list, &priv->path_list);
254
255         return 0;
256 }
257
258 static void path_free(struct net_device *dev, struct ipoib_path *path)
259 {
260         struct ipoib_dev_priv *priv = netdev_priv(dev);
261         struct ipoib_neigh *neigh, *tn;
262         struct sk_buff *skb;
263         unsigned long flags;
264
265         while ((skb = __skb_dequeue(&path->queue)))
266                 dev_kfree_skb_irq(skb);
267
268         spin_lock_irqsave(&priv->lock, flags);
269
270         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
271                 /*
272                  * It's safe to call ipoib_put_ah() inside priv->lock
273                  * here, because we know that path->ah will always
274                  * hold one more reference, so ipoib_put_ah() will
275                  * never do more than decrement the ref count.
276                  */
277                 if (neigh->ah)
278                         ipoib_put_ah(neigh->ah);
279
280                 ipoib_neigh_free(dev, neigh);
281         }
282
283         spin_unlock_irqrestore(&priv->lock, flags);
284
285         if (path->ah)
286                 ipoib_put_ah(path->ah);
287
288         kfree(path);
289 }
290
291 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
292
293 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
294 {
295         struct ipoib_path_iter *iter;
296
297         iter = kmalloc(sizeof *iter, GFP_KERNEL);
298         if (!iter)
299                 return NULL;
300
301         iter->dev = dev;
302         memset(iter->path.pathrec.dgid.raw, 0, 16);
303
304         if (ipoib_path_iter_next(iter)) {
305                 kfree(iter);
306                 return NULL;
307         }
308
309         return iter;
310 }
311
312 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
313 {
314         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
315         struct rb_node *n;
316         struct ipoib_path *path;
317         int ret = 1;
318
319         spin_lock_irq(&priv->lock);
320
321         n = rb_first(&priv->path_tree);
322
323         while (n) {
324                 path = rb_entry(n, struct ipoib_path, rb_node);
325
326                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
327                            sizeof (union ib_gid)) < 0) {
328                         iter->path = *path;
329                         ret = 0;
330                         break;
331                 }
332
333                 n = rb_next(n);
334         }
335
336         spin_unlock_irq(&priv->lock);
337
338         return ret;
339 }
340
341 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
342                           struct ipoib_path *path)
343 {
344         *path = iter->path;
345 }
346
347 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
348
349 void ipoib_flush_paths(struct net_device *dev)
350 {
351         struct ipoib_dev_priv *priv = netdev_priv(dev);
352         struct ipoib_path *path, *tp;
353         LIST_HEAD(remove_list);
354
355         spin_lock_irq(&priv->tx_lock);
356         spin_lock(&priv->lock);
357
358         list_splice(&priv->path_list, &remove_list);
359         INIT_LIST_HEAD(&priv->path_list);
360
361         list_for_each_entry(path, &remove_list, list)
362                 rb_erase(&path->rb_node, &priv->path_tree);
363
364         list_for_each_entry_safe(path, tp, &remove_list, list) {
365                 if (path->query)
366                         ib_sa_cancel_query(path->query_id, path->query);
367                 spin_unlock(&priv->lock);
368                 spin_unlock_irq(&priv->tx_lock);
369                 wait_for_completion(&path->done);
370                 path_free(dev, path);
371                 spin_lock_irq(&priv->tx_lock);
372                 spin_lock(&priv->lock);
373         }
374         spin_unlock(&priv->lock);
375         spin_unlock_irq(&priv->tx_lock);
376 }
377
378 static void path_rec_completion(int status,
379                                 struct ib_sa_path_rec *pathrec,
380                                 void *path_ptr)
381 {
382         struct ipoib_path *path = path_ptr;
383         struct net_device *dev = path->dev;
384         struct ipoib_dev_priv *priv = netdev_priv(dev);
385         struct ipoib_ah *ah = NULL;
386         struct ipoib_neigh *neigh, *tn;
387         struct sk_buff_head skqueue;
388         struct sk_buff *skb;
389         unsigned long flags;
390
391         if (!status)
392                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
393                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
394         else
395                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
396                           status, IPOIB_GID_ARG(path->pathrec.dgid));
397
398         skb_queue_head_init(&skqueue);
399
400         if (!status) {
401                 struct ib_ah_attr av;
402
403                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
404                         ah = ipoib_create_ah(dev, priv->pd, &av);
405         }
406
407         spin_lock_irqsave(&priv->lock, flags);
408
409         path->ah = ah;
410
411         if (ah) {
412                 path->pathrec = *pathrec;
413
414                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
415                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
416
417                 while ((skb = __skb_dequeue(&path->queue)))
418                         __skb_queue_tail(&skqueue, skb);
419
420                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
421                         kref_get(&path->ah->ref);
422                         neigh->ah = path->ah;
423                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
424                                sizeof(union ib_gid));
425
426                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
427                                 if (!ipoib_cm_get(neigh))
428                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
429                                                                                path,
430                                                                                neigh));
431                                 if (!ipoib_cm_get(neigh)) {
432                                         list_del(&neigh->list);
433                                         if (neigh->ah)
434                                                 ipoib_put_ah(neigh->ah);
435                                         ipoib_neigh_free(dev, neigh);
436                                         continue;
437                                 }
438                         }
439
440                         while ((skb = __skb_dequeue(&neigh->queue)))
441                                 __skb_queue_tail(&skqueue, skb);
442                 }
443         }
444
445         path->query = NULL;
446         complete(&path->done);
447
448         spin_unlock_irqrestore(&priv->lock, flags);
449
450         while ((skb = __skb_dequeue(&skqueue))) {
451                 skb->dev = dev;
452                 if (dev_queue_xmit(skb))
453                         ipoib_warn(priv, "dev_queue_xmit failed "
454                                    "to requeue packet\n");
455         }
456 }
457
458 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
459 {
460         struct ipoib_dev_priv *priv = netdev_priv(dev);
461         struct ipoib_path *path;
462
463         if (!priv->broadcast)
464                 return NULL;
465
466         path = kzalloc(sizeof *path, GFP_ATOMIC);
467         if (!path)
468                 return NULL;
469
470         path->dev = dev;
471
472         skb_queue_head_init(&path->queue);
473
474         INIT_LIST_HEAD(&path->neigh_list);
475
476         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
477         path->pathrec.sgid          = priv->local_gid;
478         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
479         path->pathrec.numb_path     = 1;
480         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
481
482         return path;
483 }
484
485 static int path_rec_start(struct net_device *dev,
486                           struct ipoib_path *path)
487 {
488         struct ipoib_dev_priv *priv = netdev_priv(dev);
489
490         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
491                   IPOIB_GID_ARG(path->pathrec.dgid));
492
493         init_completion(&path->done);
494
495         path->query_id =
496                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
497                                    &path->pathrec,
498                                    IB_SA_PATH_REC_DGID          |
499                                    IB_SA_PATH_REC_SGID          |
500                                    IB_SA_PATH_REC_NUMB_PATH     |
501                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
502                                    IB_SA_PATH_REC_PKEY,
503                                    1000, GFP_ATOMIC,
504                                    path_rec_completion,
505                                    path, &path->query);
506         if (path->query_id < 0) {
507                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
508                 path->query = NULL;
509                 return path->query_id;
510         }
511
512         return 0;
513 }
514
515 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
516 {
517         struct ipoib_dev_priv *priv = netdev_priv(dev);
518         struct ipoib_path *path;
519         struct ipoib_neigh *neigh;
520
521         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
522         if (!neigh) {
523                 ++dev->stats.tx_dropped;
524                 dev_kfree_skb_any(skb);
525                 return;
526         }
527
528         /*
529          * We can only be called from ipoib_start_xmit, so we're
530          * inside tx_lock -- no need to save/restore flags.
531          */
532         spin_lock(&priv->lock);
533
534         path = __path_find(dev, skb->dst->neighbour->ha + 4);
535         if (!path) {
536                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
537                 if (!path)
538                         goto err_path;
539
540                 __path_add(dev, path);
541         }
542
543         list_add_tail(&neigh->list, &path->neigh_list);
544
545         if (path->ah) {
546                 kref_get(&path->ah->ref);
547                 neigh->ah = path->ah;
548                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
549                        sizeof(union ib_gid));
550
551                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
552                         if (!ipoib_cm_get(neigh))
553                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
554                         if (!ipoib_cm_get(neigh)) {
555                                 list_del(&neigh->list);
556                                 if (neigh->ah)
557                                         ipoib_put_ah(neigh->ah);
558                                 ipoib_neigh_free(dev, neigh);
559                                 goto err_drop;
560                         }
561                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
562                                 __skb_queue_tail(&neigh->queue, skb);
563                         else {
564                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
565                                            skb_queue_len(&neigh->queue));
566                                 goto err_drop;
567                         }
568                 } else
569                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
570         } else {
571                 neigh->ah  = NULL;
572
573                 if (!path->query && path_rec_start(dev, path))
574                         goto err_list;
575
576                 __skb_queue_tail(&neigh->queue, skb);
577         }
578
579         spin_unlock(&priv->lock);
580         return;
581
582 err_list:
583         list_del(&neigh->list);
584
585 err_path:
586         ipoib_neigh_free(dev, neigh);
587 err_drop:
588         ++dev->stats.tx_dropped;
589         dev_kfree_skb_any(skb);
590
591         spin_unlock(&priv->lock);
592 }
593
594 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
595 {
596         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
597
598         /* Look up path record for unicasts */
599         if (skb->dst->neighbour->ha[4] != 0xff) {
600                 neigh_add_path(skb, dev);
601                 return;
602         }
603
604         /* Add in the P_Key for multicasts */
605         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
606         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
607         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
608 }
609
610 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
611                              struct ipoib_pseudoheader *phdr)
612 {
613         struct ipoib_dev_priv *priv = netdev_priv(dev);
614         struct ipoib_path *path;
615
616         /*
617          * We can only be called from ipoib_start_xmit, so we're
618          * inside tx_lock -- no need to save/restore flags.
619          */
620         spin_lock(&priv->lock);
621
622         path = __path_find(dev, phdr->hwaddr + 4);
623         if (!path) {
624                 path = path_rec_create(dev, phdr->hwaddr + 4);
625                 if (path) {
626                         /* put pseudoheader back on for next time */
627                         skb_push(skb, sizeof *phdr);
628                         __skb_queue_tail(&path->queue, skb);
629
630                         if (path_rec_start(dev, path)) {
631                                 spin_unlock(&priv->lock);
632                                 path_free(dev, path);
633                                 return;
634                         } else
635                                 __path_add(dev, path);
636                 } else {
637                         ++dev->stats.tx_dropped;
638                         dev_kfree_skb_any(skb);
639                 }
640
641                 spin_unlock(&priv->lock);
642                 return;
643         }
644
645         if (path->ah) {
646                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
647                           be16_to_cpu(path->pathrec.dlid));
648
649                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
650         } else if ((path->query || !path_rec_start(dev, path)) &&
651                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
652                 /* put pseudoheader back on for next time */
653                 skb_push(skb, sizeof *phdr);
654                 __skb_queue_tail(&path->queue, skb);
655         } else {
656                 ++dev->stats.tx_dropped;
657                 dev_kfree_skb_any(skb);
658         }
659
660         spin_unlock(&priv->lock);
661 }
662
663 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
664 {
665         struct ipoib_dev_priv *priv = netdev_priv(dev);
666         struct ipoib_neigh *neigh;
667         unsigned long flags;
668
669         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
670                 return NETDEV_TX_LOCKED;
671
672         /*
673          * Check if our queue is stopped.  Since we have the LLTX bit
674          * set, we can't rely on netif_stop_queue() preventing our
675          * xmit function from being called with a full queue.
676          */
677         if (unlikely(netif_queue_stopped(dev))) {
678                 spin_unlock_irqrestore(&priv->tx_lock, flags);
679                 return NETDEV_TX_BUSY;
680         }
681
682         if (likely(skb->dst && skb->dst->neighbour)) {
683                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
684                         ipoib_path_lookup(skb, dev);
685                         goto out;
686                 }
687
688                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
689
690                 if (ipoib_cm_get(neigh)) {
691                         if (ipoib_cm_up(neigh)) {
692                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
693                                 goto out;
694                         }
695                 } else if (neigh->ah) {
696                         if (unlikely((memcmp(&neigh->dgid.raw,
697                                             skb->dst->neighbour->ha + 4,
698                                             sizeof(union ib_gid))) ||
699                                          (neigh->dev != dev))) {
700                                 spin_lock(&priv->lock);
701                                 /*
702                                  * It's safe to call ipoib_put_ah() inside
703                                  * priv->lock here, because we know that
704                                  * path->ah will always hold one more reference,
705                                  * so ipoib_put_ah() will never do more than
706                                  * decrement the ref count.
707                                  */
708                                 ipoib_put_ah(neigh->ah);
709                                 list_del(&neigh->list);
710                                 ipoib_neigh_free(dev, neigh);
711                                 spin_unlock(&priv->lock);
712                                 ipoib_path_lookup(skb, dev);
713                                 goto out;
714                         }
715
716                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
717                         goto out;
718                 }
719
720                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
721                         spin_lock(&priv->lock);
722                         __skb_queue_tail(&neigh->queue, skb);
723                         spin_unlock(&priv->lock);
724                 } else {
725                         ++dev->stats.tx_dropped;
726                         dev_kfree_skb_any(skb);
727                 }
728         } else {
729                 struct ipoib_pseudoheader *phdr =
730                         (struct ipoib_pseudoheader *) skb->data;
731                 skb_pull(skb, sizeof *phdr);
732
733                 if (phdr->hwaddr[4] == 0xff) {
734                         /* Add in the P_Key for multicast*/
735                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
736                         phdr->hwaddr[9] = priv->pkey & 0xff;
737
738                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
739                 } else {
740                         /* unicast GID -- should be ARP or RARP reply */
741
742                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
743                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
744                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
745                                            IPOIB_GID_FMT "\n",
746                                            skb->dst ? "neigh" : "dst",
747                                            be16_to_cpup((__be16 *) skb->data),
748                                            IPOIB_QPN(phdr->hwaddr),
749                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
750                                 dev_kfree_skb_any(skb);
751                                 ++dev->stats.tx_dropped;
752                                 goto out;
753                         }
754
755                         unicast_arp_send(skb, dev, phdr);
756                 }
757         }
758
759 out:
760         spin_unlock_irqrestore(&priv->tx_lock, flags);
761
762         return NETDEV_TX_OK;
763 }
764
765 static void ipoib_timeout(struct net_device *dev)
766 {
767         struct ipoib_dev_priv *priv = netdev_priv(dev);
768
769         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
770                    jiffies_to_msecs(jiffies - dev->trans_start));
771         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
772                    netif_queue_stopped(dev),
773                    priv->tx_head, priv->tx_tail);
774         /* XXX reset QP, etc. */
775 }
776
777 static int ipoib_hard_header(struct sk_buff *skb,
778                              struct net_device *dev,
779                              unsigned short type,
780                              const void *daddr, const void *saddr, unsigned len)
781 {
782         struct ipoib_header *header;
783
784         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
785
786         header->proto = htons(type);
787         header->reserved = 0;
788
789         /*
790          * If we don't have a neighbour structure, stuff the
791          * destination address onto the front of the skb so we can
792          * figure out where to send the packet later.
793          */
794         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
795                 struct ipoib_pseudoheader *phdr =
796                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
797                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
798         }
799
800         return 0;
801 }
802
803 static void ipoib_set_mcast_list(struct net_device *dev)
804 {
805         struct ipoib_dev_priv *priv = netdev_priv(dev);
806
807         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
808                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
809                 return;
810         }
811
812         queue_work(ipoib_workqueue, &priv->restart_task);
813 }
814
815 static void ipoib_neigh_cleanup(struct neighbour *n)
816 {
817         struct ipoib_neigh *neigh;
818         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
819         unsigned long flags;
820         struct ipoib_ah *ah = NULL;
821
822         neigh = *to_ipoib_neigh(n);
823         if (neigh) {
824                 priv = netdev_priv(neigh->dev);
825                 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
826                           n->dev->name);
827         } else
828                 return;
829         ipoib_dbg(priv,
830                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
831                   IPOIB_QPN(n->ha),
832                   IPOIB_GID_RAW_ARG(n->ha + 4));
833
834         spin_lock_irqsave(&priv->lock, flags);
835
836         if (neigh->ah)
837                 ah = neigh->ah;
838         list_del(&neigh->list);
839         ipoib_neigh_free(n->dev, neigh);
840
841         spin_unlock_irqrestore(&priv->lock, flags);
842
843         if (ah)
844                 ipoib_put_ah(ah);
845 }
846
847 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
848                                       struct net_device *dev)
849 {
850         struct ipoib_neigh *neigh;
851
852         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
853         if (!neigh)
854                 return NULL;
855
856         neigh->neighbour = neighbour;
857         neigh->dev = dev;
858         *to_ipoib_neigh(neighbour) = neigh;
859         skb_queue_head_init(&neigh->queue);
860         ipoib_cm_set(neigh, NULL);
861
862         return neigh;
863 }
864
865 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
866 {
867         struct sk_buff *skb;
868         *to_ipoib_neigh(neigh->neighbour) = NULL;
869         while ((skb = __skb_dequeue(&neigh->queue))) {
870                 ++dev->stats.tx_dropped;
871                 dev_kfree_skb_any(skb);
872         }
873         if (ipoib_cm_get(neigh))
874                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
875         kfree(neigh);
876 }
877
878 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
879 {
880         parms->neigh_cleanup = ipoib_neigh_cleanup;
881
882         return 0;
883 }
884
885 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
886 {
887         struct ipoib_dev_priv *priv = netdev_priv(dev);
888
889         /* Allocate RX/TX "rings" to hold queued skbs */
890         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
891                                 GFP_KERNEL);
892         if (!priv->rx_ring) {
893                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
894                        ca->name, ipoib_recvq_size);
895                 goto out;
896         }
897
898         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
899                                 GFP_KERNEL);
900         if (!priv->tx_ring) {
901                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
902                        ca->name, ipoib_sendq_size);
903                 goto out_rx_ring_cleanup;
904         }
905
906         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
907
908         if (ipoib_ib_dev_init(dev, ca, port))
909                 goto out_tx_ring_cleanup;
910
911         return 0;
912
913 out_tx_ring_cleanup:
914         kfree(priv->tx_ring);
915
916 out_rx_ring_cleanup:
917         kfree(priv->rx_ring);
918
919 out:
920         return -ENOMEM;
921 }
922
923 void ipoib_dev_cleanup(struct net_device *dev)
924 {
925         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
926
927         ipoib_delete_debug_files(dev);
928
929         /* Delete any child interfaces first */
930         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
931                 unregister_netdev(cpriv->dev);
932                 ipoib_dev_cleanup(cpriv->dev);
933                 free_netdev(cpriv->dev);
934         }
935
936         ipoib_ib_dev_cleanup(dev);
937
938         kfree(priv->rx_ring);
939         kfree(priv->tx_ring);
940
941         priv->rx_ring = NULL;
942         priv->tx_ring = NULL;
943 }
944
945 static const struct header_ops ipoib_header_ops = {
946         .create = ipoib_hard_header,
947 };
948
949 static void ipoib_setup(struct net_device *dev)
950 {
951         struct ipoib_dev_priv *priv = netdev_priv(dev);
952
953         dev->open                = ipoib_open;
954         dev->stop                = ipoib_stop;
955         dev->change_mtu          = ipoib_change_mtu;
956         dev->hard_start_xmit     = ipoib_start_xmit;
957         dev->tx_timeout          = ipoib_timeout;
958         dev->header_ops          = &ipoib_header_ops;
959         dev->set_multicast_list  = ipoib_set_mcast_list;
960         dev->neigh_setup         = ipoib_neigh_setup_dev;
961
962         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
963
964         dev->watchdog_timeo      = HZ;
965
966         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
967
968         /*
969          * We add in INFINIBAND_ALEN to allow for the destination
970          * address "pseudoheader" for skbs without neighbour struct.
971          */
972         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
973         dev->addr_len            = INFINIBAND_ALEN;
974         dev->type                = ARPHRD_INFINIBAND;
975         dev->tx_queue_len        = ipoib_sendq_size * 2;
976         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
977
978         /* MTU will be reset when mcast join happens */
979         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
980         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
981
982         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
983
984         netif_carrier_off(dev);
985
986         priv->dev = dev;
987
988         spin_lock_init(&priv->lock);
989         spin_lock_init(&priv->tx_lock);
990
991         mutex_init(&priv->mcast_mutex);
992         mutex_init(&priv->vlan_mutex);
993
994         INIT_LIST_HEAD(&priv->path_list);
995         INIT_LIST_HEAD(&priv->child_intfs);
996         INIT_LIST_HEAD(&priv->dead_ahs);
997         INIT_LIST_HEAD(&priv->multicast_list);
998
999         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1000         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
1001         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1002         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
1003         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1004         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1005 }
1006
1007 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1008 {
1009         struct net_device *dev;
1010
1011         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1012                            ipoib_setup);
1013         if (!dev)
1014                 return NULL;
1015
1016         return netdev_priv(dev);
1017 }
1018
1019 static ssize_t show_pkey(struct device *dev,
1020                          struct device_attribute *attr, char *buf)
1021 {
1022         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1023
1024         return sprintf(buf, "0x%04x\n", priv->pkey);
1025 }
1026 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1027
1028 static ssize_t show_umcast(struct device *dev,
1029                            struct device_attribute *attr, char *buf)
1030 {
1031         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1032
1033         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1034 }
1035
1036 static ssize_t set_umcast(struct device *dev,
1037                           struct device_attribute *attr,
1038                           const char *buf, size_t count)
1039 {
1040         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1041         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1042
1043         if (umcast_val > 0) {
1044                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1045                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1046                                 "by userspace\n");
1047         } else
1048                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1049
1050         return count;
1051 }
1052 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1053
1054 int ipoib_add_umcast_attr(struct net_device *dev)
1055 {
1056         return device_create_file(&dev->dev, &dev_attr_umcast);
1057 }
1058
1059 static ssize_t create_child(struct device *dev,
1060                             struct device_attribute *attr,
1061                             const char *buf, size_t count)
1062 {
1063         int pkey;
1064         int ret;
1065
1066         if (sscanf(buf, "%i", &pkey) != 1)
1067                 return -EINVAL;
1068
1069         if (pkey < 0 || pkey > 0xffff)
1070                 return -EINVAL;
1071
1072         /*
1073          * Set the full membership bit, so that we join the right
1074          * broadcast group, etc.
1075          */
1076         pkey |= 0x8000;
1077
1078         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1079
1080         return ret ? ret : count;
1081 }
1082 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1083
1084 static ssize_t delete_child(struct device *dev,
1085                             struct device_attribute *attr,
1086                             const char *buf, size_t count)
1087 {
1088         int pkey;
1089         int ret;
1090
1091         if (sscanf(buf, "%i", &pkey) != 1)
1092                 return -EINVAL;
1093
1094         if (pkey < 0 || pkey > 0xffff)
1095                 return -EINVAL;
1096
1097         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1098
1099         return ret ? ret : count;
1100
1101 }
1102 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1103
1104 int ipoib_add_pkey_attr(struct net_device *dev)
1105 {
1106         return device_create_file(&dev->dev, &dev_attr_pkey);
1107 }
1108
1109 static struct net_device *ipoib_add_port(const char *format,
1110                                          struct ib_device *hca, u8 port)
1111 {
1112         struct ipoib_dev_priv *priv;
1113         int result = -ENOMEM;
1114
1115         priv = ipoib_intf_alloc(format);
1116         if (!priv)
1117                 goto alloc_mem_failed;
1118
1119         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1120
1121         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1122         if (result) {
1123                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1124                        hca->name, port, result);
1125                 goto device_init_failed;
1126         }
1127
1128         /*
1129          * Set the full membership bit, so that we join the right
1130          * broadcast group, etc.
1131          */
1132         priv->pkey |= 0x8000;
1133
1134         priv->dev->broadcast[8] = priv->pkey >> 8;
1135         priv->dev->broadcast[9] = priv->pkey & 0xff;
1136
1137         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1138         if (result) {
1139                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1140                        hca->name, port, result);
1141                 goto device_init_failed;
1142         } else
1143                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1144
1145
1146         result = ipoib_dev_init(priv->dev, hca, port);
1147         if (result < 0) {
1148                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1149                        hca->name, port, result);
1150                 goto device_init_failed;
1151         }
1152
1153         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1154                               priv->ca, ipoib_event);
1155         result = ib_register_event_handler(&priv->event_handler);
1156         if (result < 0) {
1157                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1158                        "port %d (ret = %d)\n",
1159                        hca->name, port, result);
1160                 goto event_failed;
1161         }
1162
1163         result = register_netdev(priv->dev);
1164         if (result) {
1165                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1166                        hca->name, port, result);
1167                 goto register_failed;
1168         }
1169
1170         ipoib_create_debug_files(priv->dev);
1171
1172         if (ipoib_cm_add_mode_attr(priv->dev))
1173                 goto sysfs_failed;
1174         if (ipoib_add_pkey_attr(priv->dev))
1175                 goto sysfs_failed;
1176         if (ipoib_add_umcast_attr(priv->dev))
1177                 goto sysfs_failed;
1178         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1179                 goto sysfs_failed;
1180         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1181                 goto sysfs_failed;
1182
1183         return priv->dev;
1184
1185 sysfs_failed:
1186         ipoib_delete_debug_files(priv->dev);
1187         unregister_netdev(priv->dev);
1188
1189 register_failed:
1190         ib_unregister_event_handler(&priv->event_handler);
1191         flush_scheduled_work();
1192
1193 event_failed:
1194         ipoib_dev_cleanup(priv->dev);
1195
1196 device_init_failed:
1197         free_netdev(priv->dev);
1198
1199 alloc_mem_failed:
1200         return ERR_PTR(result);
1201 }
1202
1203 static void ipoib_add_one(struct ib_device *device)
1204 {
1205         struct list_head *dev_list;
1206         struct net_device *dev;
1207         struct ipoib_dev_priv *priv;
1208         int s, e, p;
1209
1210         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1211                 return;
1212
1213         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1214         if (!dev_list)
1215                 return;
1216
1217         INIT_LIST_HEAD(dev_list);
1218
1219         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1220                 s = 0;
1221                 e = 0;
1222         } else {
1223                 s = 1;
1224                 e = device->phys_port_cnt;
1225         }
1226
1227         for (p = s; p <= e; ++p) {
1228                 dev = ipoib_add_port("ib%d", device, p);
1229                 if (!IS_ERR(dev)) {
1230                         priv = netdev_priv(dev);
1231                         list_add_tail(&priv->list, dev_list);
1232                 }
1233         }
1234
1235         ib_set_client_data(device, &ipoib_client, dev_list);
1236 }
1237
1238 static void ipoib_remove_one(struct ib_device *device)
1239 {
1240         struct ipoib_dev_priv *priv, *tmp;
1241         struct list_head *dev_list;
1242
1243         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1244                 return;
1245
1246         dev_list = ib_get_client_data(device, &ipoib_client);
1247
1248         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1249                 ib_unregister_event_handler(&priv->event_handler);
1250                 flush_scheduled_work();
1251
1252                 unregister_netdev(priv->dev);
1253                 ipoib_dev_cleanup(priv->dev);
1254                 free_netdev(priv->dev);
1255         }
1256
1257         kfree(dev_list);
1258 }
1259
1260 static int __init ipoib_init_module(void)
1261 {
1262         int ret;
1263
1264         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1265         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1266         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1267
1268         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1269         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1270         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1271
1272         ret = ipoib_register_debugfs();
1273         if (ret)
1274                 return ret;
1275
1276         /*
1277          * We create our own workqueue mainly because we want to be
1278          * able to flush it when devices are being removed.  We can't
1279          * use schedule_work()/flush_scheduled_work() because both
1280          * unregister_netdev() and linkwatch_event take the rtnl lock,
1281          * so flush_scheduled_work() can deadlock during device
1282          * removal.
1283          */
1284         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1285         if (!ipoib_workqueue) {
1286                 ret = -ENOMEM;
1287                 goto err_fs;
1288         }
1289
1290         ib_sa_register_client(&ipoib_sa_client);
1291
1292         ret = ib_register_client(&ipoib_client);
1293         if (ret)
1294                 goto err_sa;
1295
1296         return 0;
1297
1298 err_sa:
1299         ib_sa_unregister_client(&ipoib_sa_client);
1300         destroy_workqueue(ipoib_workqueue);
1301
1302 err_fs:
1303         ipoib_unregister_debugfs();
1304
1305         return ret;
1306 }
1307
1308 static void __exit ipoib_cleanup_module(void)
1309 {
1310         ib_unregister_client(&ipoib_client);
1311         ib_sa_unregister_client(&ipoib_sa_client);
1312         ipoib_unregister_debugfs();
1313         destroy_workqueue(ipoib_workqueue);
1314 }
1315
1316 module_init(ipoib_init_module);
1317 module_exit(ipoib_cleanup_module);