Merge branch 'upstream' of git://ftp.linux-mips.org/pub/scm/upstream-linus
[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         path = kzalloc(sizeof *path, GFP_ATOMIC);
464         if (!path)
465                 return NULL;
466
467         path->dev = dev;
468
469         skb_queue_head_init(&path->queue);
470
471         INIT_LIST_HEAD(&path->neigh_list);
472
473         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
474         path->pathrec.sgid          = priv->local_gid;
475         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
476         path->pathrec.numb_path     = 1;
477         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
478
479         return path;
480 }
481
482 static int path_rec_start(struct net_device *dev,
483                           struct ipoib_path *path)
484 {
485         struct ipoib_dev_priv *priv = netdev_priv(dev);
486
487         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
488                   IPOIB_GID_ARG(path->pathrec.dgid));
489
490         init_completion(&path->done);
491
492         path->query_id =
493                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
494                                    &path->pathrec,
495                                    IB_SA_PATH_REC_DGID          |
496                                    IB_SA_PATH_REC_SGID          |
497                                    IB_SA_PATH_REC_NUMB_PATH     |
498                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
499                                    IB_SA_PATH_REC_PKEY,
500                                    1000, GFP_ATOMIC,
501                                    path_rec_completion,
502                                    path, &path->query);
503         if (path->query_id < 0) {
504                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
505                 path->query = NULL;
506                 return path->query_id;
507         }
508
509         return 0;
510 }
511
512 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
513 {
514         struct ipoib_dev_priv *priv = netdev_priv(dev);
515         struct ipoib_path *path;
516         struct ipoib_neigh *neigh;
517
518         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
519         if (!neigh) {
520                 ++dev->stats.tx_dropped;
521                 dev_kfree_skb_any(skb);
522                 return;
523         }
524
525         /*
526          * We can only be called from ipoib_start_xmit, so we're
527          * inside tx_lock -- no need to save/restore flags.
528          */
529         spin_lock(&priv->lock);
530
531         path = __path_find(dev, skb->dst->neighbour->ha + 4);
532         if (!path) {
533                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
534                 if (!path)
535                         goto err_path;
536
537                 __path_add(dev, path);
538         }
539
540         list_add_tail(&neigh->list, &path->neigh_list);
541
542         if (path->ah) {
543                 kref_get(&path->ah->ref);
544                 neigh->ah = path->ah;
545                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
546                        sizeof(union ib_gid));
547
548                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
549                         if (!ipoib_cm_get(neigh))
550                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
551                         if (!ipoib_cm_get(neigh)) {
552                                 list_del(&neigh->list);
553                                 if (neigh->ah)
554                                         ipoib_put_ah(neigh->ah);
555                                 ipoib_neigh_free(dev, neigh);
556                                 goto err_drop;
557                         }
558                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
559                                 __skb_queue_tail(&neigh->queue, skb);
560                         else {
561                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
562                                            skb_queue_len(&neigh->queue));
563                                 goto err_drop;
564                         }
565                 } else
566                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
567         } else {
568                 neigh->ah  = NULL;
569
570                 if (!path->query && path_rec_start(dev, path))
571                         goto err_list;
572
573                 __skb_queue_tail(&neigh->queue, skb);
574         }
575
576         spin_unlock(&priv->lock);
577         return;
578
579 err_list:
580         list_del(&neigh->list);
581
582 err_path:
583         ipoib_neigh_free(dev, neigh);
584 err_drop:
585         ++dev->stats.tx_dropped;
586         dev_kfree_skb_any(skb);
587
588         spin_unlock(&priv->lock);
589 }
590
591 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
592 {
593         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
594
595         /* Look up path record for unicasts */
596         if (skb->dst->neighbour->ha[4] != 0xff) {
597                 neigh_add_path(skb, dev);
598                 return;
599         }
600
601         /* Add in the P_Key for multicasts */
602         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
603         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
604         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
605 }
606
607 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
608                              struct ipoib_pseudoheader *phdr)
609 {
610         struct ipoib_dev_priv *priv = netdev_priv(dev);
611         struct ipoib_path *path;
612
613         /*
614          * We can only be called from ipoib_start_xmit, so we're
615          * inside tx_lock -- no need to save/restore flags.
616          */
617         spin_lock(&priv->lock);
618
619         path = __path_find(dev, phdr->hwaddr + 4);
620         if (!path) {
621                 path = path_rec_create(dev, phdr->hwaddr + 4);
622                 if (path) {
623                         /* put pseudoheader back on for next time */
624                         skb_push(skb, sizeof *phdr);
625                         __skb_queue_tail(&path->queue, skb);
626
627                         if (path_rec_start(dev, path)) {
628                                 spin_unlock(&priv->lock);
629                                 path_free(dev, path);
630                                 return;
631                         } else
632                                 __path_add(dev, path);
633                 } else {
634                         ++dev->stats.tx_dropped;
635                         dev_kfree_skb_any(skb);
636                 }
637
638                 spin_unlock(&priv->lock);
639                 return;
640         }
641
642         if (path->ah) {
643                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
644                           be16_to_cpu(path->pathrec.dlid));
645
646                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
647         } else if ((path->query || !path_rec_start(dev, path)) &&
648                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
649                 /* put pseudoheader back on for next time */
650                 skb_push(skb, sizeof *phdr);
651                 __skb_queue_tail(&path->queue, skb);
652         } else {
653                 ++dev->stats.tx_dropped;
654                 dev_kfree_skb_any(skb);
655         }
656
657         spin_unlock(&priv->lock);
658 }
659
660 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
661 {
662         struct ipoib_dev_priv *priv = netdev_priv(dev);
663         struct ipoib_neigh *neigh;
664         unsigned long flags;
665
666         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
667                 return NETDEV_TX_LOCKED;
668
669         /*
670          * Check if our queue is stopped.  Since we have the LLTX bit
671          * set, we can't rely on netif_stop_queue() preventing our
672          * xmit function from being called with a full queue.
673          */
674         if (unlikely(netif_queue_stopped(dev))) {
675                 spin_unlock_irqrestore(&priv->tx_lock, flags);
676                 return NETDEV_TX_BUSY;
677         }
678
679         if (likely(skb->dst && skb->dst->neighbour)) {
680                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
681                         ipoib_path_lookup(skb, dev);
682                         goto out;
683                 }
684
685                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
686
687                 if (ipoib_cm_get(neigh)) {
688                         if (ipoib_cm_up(neigh)) {
689                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
690                                 goto out;
691                         }
692                 } else if (neigh->ah) {
693                         if (unlikely((memcmp(&neigh->dgid.raw,
694                                             skb->dst->neighbour->ha + 4,
695                                             sizeof(union ib_gid))) ||
696                                          (neigh->dev != dev))) {
697                                 spin_lock(&priv->lock);
698                                 /*
699                                  * It's safe to call ipoib_put_ah() inside
700                                  * priv->lock here, because we know that
701                                  * path->ah will always hold one more reference,
702                                  * so ipoib_put_ah() will never do more than
703                                  * decrement the ref count.
704                                  */
705                                 ipoib_put_ah(neigh->ah);
706                                 list_del(&neigh->list);
707                                 ipoib_neigh_free(dev, neigh);
708                                 spin_unlock(&priv->lock);
709                                 ipoib_path_lookup(skb, dev);
710                                 goto out;
711                         }
712
713                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
714                         goto out;
715                 }
716
717                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
718                         spin_lock(&priv->lock);
719                         __skb_queue_tail(&neigh->queue, skb);
720                         spin_unlock(&priv->lock);
721                 } else {
722                         ++dev->stats.tx_dropped;
723                         dev_kfree_skb_any(skb);
724                 }
725         } else {
726                 struct ipoib_pseudoheader *phdr =
727                         (struct ipoib_pseudoheader *) skb->data;
728                 skb_pull(skb, sizeof *phdr);
729
730                 if (phdr->hwaddr[4] == 0xff) {
731                         /* Add in the P_Key for multicast*/
732                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
733                         phdr->hwaddr[9] = priv->pkey & 0xff;
734
735                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
736                 } else {
737                         /* unicast GID -- should be ARP or RARP reply */
738
739                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
740                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
741                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
742                                            IPOIB_GID_FMT "\n",
743                                            skb->dst ? "neigh" : "dst",
744                                            be16_to_cpup((__be16 *) skb->data),
745                                            IPOIB_QPN(phdr->hwaddr),
746                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
747                                 dev_kfree_skb_any(skb);
748                                 ++dev->stats.tx_dropped;
749                                 goto out;
750                         }
751
752                         unicast_arp_send(skb, dev, phdr);
753                 }
754         }
755
756 out:
757         spin_unlock_irqrestore(&priv->tx_lock, flags);
758
759         return NETDEV_TX_OK;
760 }
761
762 static void ipoib_timeout(struct net_device *dev)
763 {
764         struct ipoib_dev_priv *priv = netdev_priv(dev);
765
766         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
767                    jiffies_to_msecs(jiffies - dev->trans_start));
768         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
769                    netif_queue_stopped(dev),
770                    priv->tx_head, priv->tx_tail);
771         /* XXX reset QP, etc. */
772 }
773
774 static int ipoib_hard_header(struct sk_buff *skb,
775                              struct net_device *dev,
776                              unsigned short type,
777                              const void *daddr, const void *saddr, unsigned len)
778 {
779         struct ipoib_header *header;
780
781         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
782
783         header->proto = htons(type);
784         header->reserved = 0;
785
786         /*
787          * If we don't have a neighbour structure, stuff the
788          * destination address onto the front of the skb so we can
789          * figure out where to send the packet later.
790          */
791         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
792                 struct ipoib_pseudoheader *phdr =
793                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
794                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
795         }
796
797         return 0;
798 }
799
800 static void ipoib_set_mcast_list(struct net_device *dev)
801 {
802         struct ipoib_dev_priv *priv = netdev_priv(dev);
803
804         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
805                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
806                 return;
807         }
808
809         queue_work(ipoib_workqueue, &priv->restart_task);
810 }
811
812 static void ipoib_neigh_cleanup(struct neighbour *n)
813 {
814         struct ipoib_neigh *neigh;
815         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
816         unsigned long flags;
817         struct ipoib_ah *ah = NULL;
818
819         neigh = *to_ipoib_neigh(n);
820         if (neigh) {
821                 priv = netdev_priv(neigh->dev);
822                 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
823                           n->dev->name);
824         } else
825                 return;
826         ipoib_dbg(priv,
827                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
828                   IPOIB_QPN(n->ha),
829                   IPOIB_GID_RAW_ARG(n->ha + 4));
830
831         spin_lock_irqsave(&priv->lock, flags);
832
833         if (neigh->ah)
834                 ah = neigh->ah;
835         list_del(&neigh->list);
836         ipoib_neigh_free(n->dev, neigh);
837
838         spin_unlock_irqrestore(&priv->lock, flags);
839
840         if (ah)
841                 ipoib_put_ah(ah);
842 }
843
844 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
845                                       struct net_device *dev)
846 {
847         struct ipoib_neigh *neigh;
848
849         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
850         if (!neigh)
851                 return NULL;
852
853         neigh->neighbour = neighbour;
854         neigh->dev = dev;
855         *to_ipoib_neigh(neighbour) = neigh;
856         skb_queue_head_init(&neigh->queue);
857         ipoib_cm_set(neigh, NULL);
858
859         return neigh;
860 }
861
862 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
863 {
864         struct sk_buff *skb;
865         *to_ipoib_neigh(neigh->neighbour) = NULL;
866         while ((skb = __skb_dequeue(&neigh->queue))) {
867                 ++dev->stats.tx_dropped;
868                 dev_kfree_skb_any(skb);
869         }
870         if (ipoib_cm_get(neigh))
871                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
872         kfree(neigh);
873 }
874
875 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
876 {
877         parms->neigh_cleanup = ipoib_neigh_cleanup;
878
879         return 0;
880 }
881
882 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
883 {
884         struct ipoib_dev_priv *priv = netdev_priv(dev);
885
886         /* Allocate RX/TX "rings" to hold queued skbs */
887         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
888                                 GFP_KERNEL);
889         if (!priv->rx_ring) {
890                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
891                        ca->name, ipoib_recvq_size);
892                 goto out;
893         }
894
895         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
896                                 GFP_KERNEL);
897         if (!priv->tx_ring) {
898                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
899                        ca->name, ipoib_sendq_size);
900                 goto out_rx_ring_cleanup;
901         }
902
903         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
904
905         if (ipoib_ib_dev_init(dev, ca, port))
906                 goto out_tx_ring_cleanup;
907
908         return 0;
909
910 out_tx_ring_cleanup:
911         kfree(priv->tx_ring);
912
913 out_rx_ring_cleanup:
914         kfree(priv->rx_ring);
915
916 out:
917         return -ENOMEM;
918 }
919
920 void ipoib_dev_cleanup(struct net_device *dev)
921 {
922         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
923
924         ipoib_delete_debug_files(dev);
925
926         /* Delete any child interfaces first */
927         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
928                 unregister_netdev(cpriv->dev);
929                 ipoib_dev_cleanup(cpriv->dev);
930                 free_netdev(cpriv->dev);
931         }
932
933         ipoib_ib_dev_cleanup(dev);
934
935         kfree(priv->rx_ring);
936         kfree(priv->tx_ring);
937
938         priv->rx_ring = NULL;
939         priv->tx_ring = NULL;
940 }
941
942 static const struct header_ops ipoib_header_ops = {
943         .create = ipoib_hard_header,
944 };
945
946 static void ipoib_setup(struct net_device *dev)
947 {
948         struct ipoib_dev_priv *priv = netdev_priv(dev);
949
950         dev->open                = ipoib_open;
951         dev->stop                = ipoib_stop;
952         dev->change_mtu          = ipoib_change_mtu;
953         dev->hard_start_xmit     = ipoib_start_xmit;
954         dev->tx_timeout          = ipoib_timeout;
955         dev->header_ops          = &ipoib_header_ops;
956         dev->set_multicast_list  = ipoib_set_mcast_list;
957         dev->neigh_setup         = ipoib_neigh_setup_dev;
958
959         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
960
961         dev->watchdog_timeo      = HZ;
962
963         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
964
965         /*
966          * We add in INFINIBAND_ALEN to allow for the destination
967          * address "pseudoheader" for skbs without neighbour struct.
968          */
969         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
970         dev->addr_len            = INFINIBAND_ALEN;
971         dev->type                = ARPHRD_INFINIBAND;
972         dev->tx_queue_len        = ipoib_sendq_size * 2;
973         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
974
975         /* MTU will be reset when mcast join happens */
976         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
977         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
978
979         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
980
981         netif_carrier_off(dev);
982
983         priv->dev = dev;
984
985         spin_lock_init(&priv->lock);
986         spin_lock_init(&priv->tx_lock);
987
988         mutex_init(&priv->mcast_mutex);
989         mutex_init(&priv->vlan_mutex);
990
991         INIT_LIST_HEAD(&priv->path_list);
992         INIT_LIST_HEAD(&priv->child_intfs);
993         INIT_LIST_HEAD(&priv->dead_ahs);
994         INIT_LIST_HEAD(&priv->multicast_list);
995
996         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
997         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
998         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
999         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
1000         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1001         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1002 }
1003
1004 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1005 {
1006         struct net_device *dev;
1007
1008         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1009                            ipoib_setup);
1010         if (!dev)
1011                 return NULL;
1012
1013         return netdev_priv(dev);
1014 }
1015
1016 static ssize_t show_pkey(struct device *dev,
1017                          struct device_attribute *attr, char *buf)
1018 {
1019         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1020
1021         return sprintf(buf, "0x%04x\n", priv->pkey);
1022 }
1023 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1024
1025 static ssize_t show_umcast(struct device *dev,
1026                            struct device_attribute *attr, char *buf)
1027 {
1028         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1029
1030         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1031 }
1032
1033 static ssize_t set_umcast(struct device *dev,
1034                           struct device_attribute *attr,
1035                           const char *buf, size_t count)
1036 {
1037         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1038         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1039
1040         if (umcast_val > 0) {
1041                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1042                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1043                                 "by userspace\n");
1044         } else
1045                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1046
1047         return count;
1048 }
1049 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1050
1051 int ipoib_add_umcast_attr(struct net_device *dev)
1052 {
1053         return device_create_file(&dev->dev, &dev_attr_umcast);
1054 }
1055
1056 static ssize_t create_child(struct device *dev,
1057                             struct device_attribute *attr,
1058                             const char *buf, size_t count)
1059 {
1060         int pkey;
1061         int ret;
1062
1063         if (sscanf(buf, "%i", &pkey) != 1)
1064                 return -EINVAL;
1065
1066         if (pkey < 0 || pkey > 0xffff)
1067                 return -EINVAL;
1068
1069         /*
1070          * Set the full membership bit, so that we join the right
1071          * broadcast group, etc.
1072          */
1073         pkey |= 0x8000;
1074
1075         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1076
1077         return ret ? ret : count;
1078 }
1079 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1080
1081 static ssize_t delete_child(struct device *dev,
1082                             struct device_attribute *attr,
1083                             const char *buf, size_t count)
1084 {
1085         int pkey;
1086         int ret;
1087
1088         if (sscanf(buf, "%i", &pkey) != 1)
1089                 return -EINVAL;
1090
1091         if (pkey < 0 || pkey > 0xffff)
1092                 return -EINVAL;
1093
1094         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1095
1096         return ret ? ret : count;
1097
1098 }
1099 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1100
1101 int ipoib_add_pkey_attr(struct net_device *dev)
1102 {
1103         return device_create_file(&dev->dev, &dev_attr_pkey);
1104 }
1105
1106 static struct net_device *ipoib_add_port(const char *format,
1107                                          struct ib_device *hca, u8 port)
1108 {
1109         struct ipoib_dev_priv *priv;
1110         int result = -ENOMEM;
1111
1112         priv = ipoib_intf_alloc(format);
1113         if (!priv)
1114                 goto alloc_mem_failed;
1115
1116         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1117
1118         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1119         if (result) {
1120                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1121                        hca->name, port, result);
1122                 goto device_init_failed;
1123         }
1124
1125         /*
1126          * Set the full membership bit, so that we join the right
1127          * broadcast group, etc.
1128          */
1129         priv->pkey |= 0x8000;
1130
1131         priv->dev->broadcast[8] = priv->pkey >> 8;
1132         priv->dev->broadcast[9] = priv->pkey & 0xff;
1133
1134         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1135         if (result) {
1136                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1137                        hca->name, port, result);
1138                 goto device_init_failed;
1139         } else
1140                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1141
1142
1143         result = ipoib_dev_init(priv->dev, hca, port);
1144         if (result < 0) {
1145                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1146                        hca->name, port, result);
1147                 goto device_init_failed;
1148         }
1149
1150         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1151                               priv->ca, ipoib_event);
1152         result = ib_register_event_handler(&priv->event_handler);
1153         if (result < 0) {
1154                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1155                        "port %d (ret = %d)\n",
1156                        hca->name, port, result);
1157                 goto event_failed;
1158         }
1159
1160         result = register_netdev(priv->dev);
1161         if (result) {
1162                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1163                        hca->name, port, result);
1164                 goto register_failed;
1165         }
1166
1167         ipoib_create_debug_files(priv->dev);
1168
1169         if (ipoib_cm_add_mode_attr(priv->dev))
1170                 goto sysfs_failed;
1171         if (ipoib_add_pkey_attr(priv->dev))
1172                 goto sysfs_failed;
1173         if (ipoib_add_umcast_attr(priv->dev))
1174                 goto sysfs_failed;
1175         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1176                 goto sysfs_failed;
1177         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1178                 goto sysfs_failed;
1179
1180         return priv->dev;
1181
1182 sysfs_failed:
1183         ipoib_delete_debug_files(priv->dev);
1184         unregister_netdev(priv->dev);
1185
1186 register_failed:
1187         ib_unregister_event_handler(&priv->event_handler);
1188         flush_scheduled_work();
1189
1190 event_failed:
1191         ipoib_dev_cleanup(priv->dev);
1192
1193 device_init_failed:
1194         free_netdev(priv->dev);
1195
1196 alloc_mem_failed:
1197         return ERR_PTR(result);
1198 }
1199
1200 static void ipoib_add_one(struct ib_device *device)
1201 {
1202         struct list_head *dev_list;
1203         struct net_device *dev;
1204         struct ipoib_dev_priv *priv;
1205         int s, e, p;
1206
1207         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1208                 return;
1209
1210         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1211         if (!dev_list)
1212                 return;
1213
1214         INIT_LIST_HEAD(dev_list);
1215
1216         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1217                 s = 0;
1218                 e = 0;
1219         } else {
1220                 s = 1;
1221                 e = device->phys_port_cnt;
1222         }
1223
1224         for (p = s; p <= e; ++p) {
1225                 dev = ipoib_add_port("ib%d", device, p);
1226                 if (!IS_ERR(dev)) {
1227                         priv = netdev_priv(dev);
1228                         list_add_tail(&priv->list, dev_list);
1229                 }
1230         }
1231
1232         ib_set_client_data(device, &ipoib_client, dev_list);
1233 }
1234
1235 static void ipoib_remove_one(struct ib_device *device)
1236 {
1237         struct ipoib_dev_priv *priv, *tmp;
1238         struct list_head *dev_list;
1239
1240         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1241                 return;
1242
1243         dev_list = ib_get_client_data(device, &ipoib_client);
1244
1245         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1246                 ib_unregister_event_handler(&priv->event_handler);
1247                 flush_scheduled_work();
1248
1249                 unregister_netdev(priv->dev);
1250                 ipoib_dev_cleanup(priv->dev);
1251                 free_netdev(priv->dev);
1252         }
1253
1254         kfree(dev_list);
1255 }
1256
1257 static int __init ipoib_init_module(void)
1258 {
1259         int ret;
1260
1261         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1262         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1263         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1264
1265         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1266         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1267         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1268
1269         ret = ipoib_register_debugfs();
1270         if (ret)
1271                 return ret;
1272
1273         /*
1274          * We create our own workqueue mainly because we want to be
1275          * able to flush it when devices are being removed.  We can't
1276          * use schedule_work()/flush_scheduled_work() because both
1277          * unregister_netdev() and linkwatch_event take the rtnl lock,
1278          * so flush_scheduled_work() can deadlock during device
1279          * removal.
1280          */
1281         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1282         if (!ipoib_workqueue) {
1283                 ret = -ENOMEM;
1284                 goto err_fs;
1285         }
1286
1287         ib_sa_register_client(&ipoib_sa_client);
1288
1289         ret = ib_register_client(&ipoib_client);
1290         if (ret)
1291                 goto err_sa;
1292
1293         return 0;
1294
1295 err_sa:
1296         ib_sa_unregister_client(&ipoib_sa_client);
1297         destroy_workqueue(ipoib_workqueue);
1298
1299 err_fs:
1300         ipoib_unregister_debugfs();
1301
1302         return ret;
1303 }
1304
1305 static void __exit ipoib_cleanup_module(void)
1306 {
1307         ib_unregister_client(&ipoib_client);
1308         ib_sa_unregister_client(&ipoib_sa_client);
1309         ipoib_unregister_debugfs();
1310         destroy_workqueue(ipoib_workqueue);
1311 }
1312
1313 module_init(ipoib_init_module);
1314 module_exit(ipoib_cleanup_module);