ieee1394: merge from Linus
[linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_main.c
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
35  */
36
37 #include "ipoib.h"
38
39 #include <linux/module.h>
40
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/vmalloc.h>
44 #include <linux/kernel.h>
45
46 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
47
48 #include <linux/ip.h>
49 #include <linux/in.h>
50
51 #include <net/dst.h>
52
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
55 MODULE_LICENSE("Dual BSD/GPL");
56
57 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
58 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
59
60 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
61 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
62 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
63 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
64
65 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
66 int ipoib_debug_level;
67
68 module_param_named(debug_level, ipoib_debug_level, int, 0644);
69 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
70 #endif
71
72 struct ipoib_path_iter {
73         struct net_device *dev;
74         struct ipoib_path  path;
75 };
76
77 static const u8 ipv4_bcast_addr[] = {
78         0x00, 0xff, 0xff, 0xff,
79         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
80         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
81 };
82
83 struct workqueue_struct *ipoib_workqueue;
84
85 static void ipoib_add_one(struct ib_device *device);
86 static void ipoib_remove_one(struct ib_device *device);
87
88 static struct ib_client ipoib_client = {
89         .name   = "ipoib",
90         .add    = ipoib_add_one,
91         .remove = ipoib_remove_one
92 };
93
94 int ipoib_open(struct net_device *dev)
95 {
96         struct ipoib_dev_priv *priv = netdev_priv(dev);
97
98         ipoib_dbg(priv, "bringing up interface\n");
99
100         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
101
102         if (ipoib_pkey_dev_delay_open(dev))
103                 return 0;
104
105         if (ipoib_ib_dev_open(dev))
106                 return -EINVAL;
107
108         if (ipoib_ib_dev_up(dev)) {
109                 ipoib_ib_dev_stop(dev);
110                 return -EINVAL;
111         }
112
113         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
114                 struct ipoib_dev_priv *cpriv;
115
116                 /* Bring up any child interfaces too */
117                 mutex_lock(&priv->vlan_mutex);
118                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
119                         int flags;
120
121                         flags = cpriv->dev->flags;
122                         if (flags & IFF_UP)
123                                 continue;
124
125                         dev_change_flags(cpriv->dev, flags | IFF_UP);
126                 }
127                 mutex_unlock(&priv->vlan_mutex);
128         }
129
130         netif_start_queue(dev);
131
132         return 0;
133 }
134
135 static int ipoib_stop(struct net_device *dev)
136 {
137         struct ipoib_dev_priv *priv = netdev_priv(dev);
138
139         ipoib_dbg(priv, "stopping interface\n");
140
141         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
142
143         netif_stop_queue(dev);
144
145         /*
146          * Now flush workqueue to make sure a scheduled task doesn't
147          * bring our internal state back up.
148          */
149         flush_workqueue(ipoib_workqueue);
150
151         ipoib_ib_dev_down(dev, 1);
152         ipoib_ib_dev_stop(dev);
153
154         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
155                 struct ipoib_dev_priv *cpriv;
156
157                 /* Bring down any child interfaces too */
158                 mutex_lock(&priv->vlan_mutex);
159                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
160                         int flags;
161
162                         flags = cpriv->dev->flags;
163                         if (!(flags & IFF_UP))
164                                 continue;
165
166                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
167                 }
168                 mutex_unlock(&priv->vlan_mutex);
169         }
170
171         return 0;
172 }
173
174 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
175 {
176         struct ipoib_dev_priv *priv = netdev_priv(dev);
177
178         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
179                 return -EINVAL;
180
181         priv->admin_mtu = new_mtu;
182
183         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
184
185         return 0;
186 }
187
188 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
189 {
190         struct ipoib_dev_priv *priv = netdev_priv(dev);
191         struct rb_node *n = priv->path_tree.rb_node;
192         struct ipoib_path *path;
193         int ret;
194
195         while (n) {
196                 path = rb_entry(n, struct ipoib_path, rb_node);
197
198                 ret = memcmp(gid, path->pathrec.dgid.raw,
199                              sizeof (union ib_gid));
200
201                 if (ret < 0)
202                         n = n->rb_left;
203                 else if (ret > 0)
204                         n = n->rb_right;
205                 else
206                         return path;
207         }
208
209         return NULL;
210 }
211
212 static int __path_add(struct net_device *dev, struct ipoib_path *path)
213 {
214         struct ipoib_dev_priv *priv = netdev_priv(dev);
215         struct rb_node **n = &priv->path_tree.rb_node;
216         struct rb_node *pn = NULL;
217         struct ipoib_path *tpath;
218         int ret;
219
220         while (*n) {
221                 pn = *n;
222                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
223
224                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
225                              sizeof (union ib_gid));
226                 if (ret < 0)
227                         n = &pn->rb_left;
228                 else if (ret > 0)
229                         n = &pn->rb_right;
230                 else
231                         return -EEXIST;
232         }
233
234         rb_link_node(&path->rb_node, pn, n);
235         rb_insert_color(&path->rb_node, &priv->path_tree);
236
237         list_add_tail(&path->list, &priv->path_list);
238
239         return 0;
240 }
241
242 static void path_free(struct net_device *dev, struct ipoib_path *path)
243 {
244         struct ipoib_dev_priv *priv = netdev_priv(dev);
245         struct ipoib_neigh *neigh, *tn;
246         struct sk_buff *skb;
247         unsigned long flags;
248
249         while ((skb = __skb_dequeue(&path->queue)))
250                 dev_kfree_skb_irq(skb);
251
252         spin_lock_irqsave(&priv->lock, flags);
253
254         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
255                 /*
256                  * It's safe to call ipoib_put_ah() inside priv->lock
257                  * here, because we know that path->ah will always
258                  * hold one more reference, so ipoib_put_ah() will
259                  * never do more than decrement the ref count.
260                  */
261                 if (neigh->ah)
262                         ipoib_put_ah(neigh->ah);
263
264                 ipoib_neigh_free(neigh);
265         }
266
267         spin_unlock_irqrestore(&priv->lock, flags);
268
269         if (path->ah)
270                 ipoib_put_ah(path->ah);
271
272         kfree(path);
273 }
274
275 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
276
277 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
278 {
279         struct ipoib_path_iter *iter;
280
281         iter = kmalloc(sizeof *iter, GFP_KERNEL);
282         if (!iter)
283                 return NULL;
284
285         iter->dev = dev;
286         memset(iter->path.pathrec.dgid.raw, 0, 16);
287
288         if (ipoib_path_iter_next(iter)) {
289                 kfree(iter);
290                 return NULL;
291         }
292
293         return iter;
294 }
295
296 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
297 {
298         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
299         struct rb_node *n;
300         struct ipoib_path *path;
301         int ret = 1;
302
303         spin_lock_irq(&priv->lock);
304
305         n = rb_first(&priv->path_tree);
306
307         while (n) {
308                 path = rb_entry(n, struct ipoib_path, rb_node);
309
310                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
311                            sizeof (union ib_gid)) < 0) {
312                         iter->path = *path;
313                         ret = 0;
314                         break;
315                 }
316
317                 n = rb_next(n);
318         }
319
320         spin_unlock_irq(&priv->lock);
321
322         return ret;
323 }
324
325 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
326                           struct ipoib_path *path)
327 {
328         *path = iter->path;
329 }
330
331 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
332
333 void ipoib_flush_paths(struct net_device *dev)
334 {
335         struct ipoib_dev_priv *priv = netdev_priv(dev);
336         struct ipoib_path *path, *tp;
337         LIST_HEAD(remove_list);
338
339         spin_lock_irq(&priv->lock);
340
341         list_splice(&priv->path_list, &remove_list);
342         INIT_LIST_HEAD(&priv->path_list);
343
344         list_for_each_entry(path, &remove_list, list)
345                 rb_erase(&path->rb_node, &priv->path_tree);
346
347         list_for_each_entry_safe(path, tp, &remove_list, list) {
348                 if (path->query)
349                         ib_sa_cancel_query(path->query_id, path->query);
350                 spin_unlock_irq(&priv->lock);
351                 wait_for_completion(&path->done);
352                 path_free(dev, path);
353                 spin_lock_irq(&priv->lock);
354         }
355         spin_unlock_irq(&priv->lock);
356 }
357
358 static void path_rec_completion(int status,
359                                 struct ib_sa_path_rec *pathrec,
360                                 void *path_ptr)
361 {
362         struct ipoib_path *path = path_ptr;
363         struct net_device *dev = path->dev;
364         struct ipoib_dev_priv *priv = netdev_priv(dev);
365         struct ipoib_ah *ah = NULL;
366         struct ipoib_neigh *neigh;
367         struct sk_buff_head skqueue;
368         struct sk_buff *skb;
369         unsigned long flags;
370
371         if (pathrec)
372                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
373                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
374         else
375                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
376                           status, IPOIB_GID_ARG(path->pathrec.dgid));
377
378         skb_queue_head_init(&skqueue);
379
380         if (!status) {
381                 struct ib_ah_attr av = {
382                         .dlid          = be16_to_cpu(pathrec->dlid),
383                         .sl            = pathrec->sl,
384                         .port_num      = priv->port,
385                         .static_rate   = pathrec->rate
386                 };
387
388                 ah = ipoib_create_ah(dev, priv->pd, &av);
389         }
390
391         spin_lock_irqsave(&priv->lock, flags);
392
393         path->ah = ah;
394
395         if (ah) {
396                 path->pathrec = *pathrec;
397
398                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
399                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
400
401                 while ((skb = __skb_dequeue(&path->queue)))
402                         __skb_queue_tail(&skqueue, skb);
403
404                 list_for_each_entry(neigh, &path->neigh_list, list) {
405                         kref_get(&path->ah->ref);
406                         neigh->ah = path->ah;
407                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
408                                sizeof(union ib_gid));
409
410                         while ((skb = __skb_dequeue(&neigh->queue)))
411                                 __skb_queue_tail(&skqueue, skb);
412                 }
413         }
414
415         path->query = NULL;
416         complete(&path->done);
417
418         spin_unlock_irqrestore(&priv->lock, flags);
419
420         while ((skb = __skb_dequeue(&skqueue))) {
421                 skb->dev = dev;
422                 if (dev_queue_xmit(skb))
423                         ipoib_warn(priv, "dev_queue_xmit failed "
424                                    "to requeue packet\n");
425         }
426 }
427
428 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
429 {
430         struct ipoib_dev_priv *priv = netdev_priv(dev);
431         struct ipoib_path *path;
432
433         path = kzalloc(sizeof *path, GFP_ATOMIC);
434         if (!path)
435                 return NULL;
436
437         path->dev = dev;
438
439         skb_queue_head_init(&path->queue);
440
441         INIT_LIST_HEAD(&path->neigh_list);
442
443         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
444         path->pathrec.sgid      = priv->local_gid;
445         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
446         path->pathrec.numb_path = 1;
447
448         return path;
449 }
450
451 static int path_rec_start(struct net_device *dev,
452                           struct ipoib_path *path)
453 {
454         struct ipoib_dev_priv *priv = netdev_priv(dev);
455
456         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
457                   IPOIB_GID_ARG(path->pathrec.dgid));
458
459         init_completion(&path->done);
460
461         path->query_id =
462                 ib_sa_path_rec_get(priv->ca, priv->port,
463                                    &path->pathrec,
464                                    IB_SA_PATH_REC_DGID          |
465                                    IB_SA_PATH_REC_SGID          |
466                                    IB_SA_PATH_REC_NUMB_PATH     |
467                                    IB_SA_PATH_REC_PKEY,
468                                    1000, GFP_ATOMIC,
469                                    path_rec_completion,
470                                    path, &path->query);
471         if (path->query_id < 0) {
472                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
473                 path->query = NULL;
474                 return path->query_id;
475         }
476
477         return 0;
478 }
479
480 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
481 {
482         struct ipoib_dev_priv *priv = netdev_priv(dev);
483         struct ipoib_path *path;
484         struct ipoib_neigh *neigh;
485
486         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
487         if (!neigh) {
488                 ++priv->stats.tx_dropped;
489                 dev_kfree_skb_any(skb);
490                 return;
491         }
492
493         skb_queue_head_init(&neigh->queue);
494
495         /*
496          * We can only be called from ipoib_start_xmit, so we're
497          * inside tx_lock -- no need to save/restore flags.
498          */
499         spin_lock(&priv->lock);
500
501         path = __path_find(dev, skb->dst->neighbour->ha + 4);
502         if (!path) {
503                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
504                 if (!path)
505                         goto err_path;
506
507                 __path_add(dev, path);
508         }
509
510         list_add_tail(&neigh->list, &path->neigh_list);
511
512         if (path->ah) {
513                 kref_get(&path->ah->ref);
514                 neigh->ah = path->ah;
515                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
516                        sizeof(union ib_gid));
517
518                 ipoib_send(dev, skb, path->ah,
519                            be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
520         } else {
521                 neigh->ah  = NULL;
522                 __skb_queue_tail(&neigh->queue, skb);
523
524                 if (!path->query && path_rec_start(dev, path))
525                         goto err_list;
526         }
527
528         spin_unlock(&priv->lock);
529         return;
530
531 err_list:
532         list_del(&neigh->list);
533
534 err_path:
535         ipoib_neigh_free(neigh);
536         ++priv->stats.tx_dropped;
537         dev_kfree_skb_any(skb);
538
539         spin_unlock(&priv->lock);
540 }
541
542 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
543 {
544         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
545
546         /* Look up path record for unicasts */
547         if (skb->dst->neighbour->ha[4] != 0xff) {
548                 neigh_add_path(skb, dev);
549                 return;
550         }
551
552         /* Add in the P_Key for multicasts */
553         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
554         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
555         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
556 }
557
558 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
559                              struct ipoib_pseudoheader *phdr)
560 {
561         struct ipoib_dev_priv *priv = netdev_priv(dev);
562         struct ipoib_path *path;
563
564         /*
565          * We can only be called from ipoib_start_xmit, so we're
566          * inside tx_lock -- no need to save/restore flags.
567          */
568         spin_lock(&priv->lock);
569
570         path = __path_find(dev, phdr->hwaddr + 4);
571         if (!path) {
572                 path = path_rec_create(dev, phdr->hwaddr + 4);
573                 if (path) {
574                         /* put pseudoheader back on for next time */
575                         skb_push(skb, sizeof *phdr);
576                         __skb_queue_tail(&path->queue, skb);
577
578                         if (path_rec_start(dev, path)) {
579                                 spin_unlock(&priv->lock);
580                                 path_free(dev, path);
581                                 return;
582                         } else
583                                 __path_add(dev, path);
584                 } else {
585                         ++priv->stats.tx_dropped;
586                         dev_kfree_skb_any(skb);
587                 }
588
589                 spin_unlock(&priv->lock);
590                 return;
591         }
592
593         if (path->ah) {
594                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
595                           be16_to_cpu(path->pathrec.dlid));
596
597                 ipoib_send(dev, skb, path->ah,
598                            be32_to_cpup((__be32 *) phdr->hwaddr));
599         } else if ((path->query || !path_rec_start(dev, path)) &&
600                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
601                 /* put pseudoheader back on for next time */
602                 skb_push(skb, sizeof *phdr);
603                 __skb_queue_tail(&path->queue, skb);
604         } else {
605                 ++priv->stats.tx_dropped;
606                 dev_kfree_skb_any(skb);
607         }
608
609         spin_unlock(&priv->lock);
610 }
611
612 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
613 {
614         struct ipoib_dev_priv *priv = netdev_priv(dev);
615         struct ipoib_neigh *neigh;
616         unsigned long flags;
617
618         if (!spin_trylock_irqsave(&priv->tx_lock, flags))
619                 return NETDEV_TX_LOCKED;
620
621         /*
622          * Check if our queue is stopped.  Since we have the LLTX bit
623          * set, we can't rely on netif_stop_queue() preventing our
624          * xmit function from being called with a full queue.
625          */
626         if (unlikely(netif_queue_stopped(dev))) {
627                 spin_unlock_irqrestore(&priv->tx_lock, flags);
628                 return NETDEV_TX_BUSY;
629         }
630
631         if (skb->dst && skb->dst->neighbour) {
632                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
633                         ipoib_path_lookup(skb, dev);
634                         goto out;
635                 }
636
637                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
638
639                 if (likely(neigh->ah)) {
640                         if (unlikely(memcmp(&neigh->dgid.raw,
641                                             skb->dst->neighbour->ha + 4,
642                                             sizeof(union ib_gid)))) {
643                                 spin_lock(&priv->lock);
644                                 /*
645                                  * It's safe to call ipoib_put_ah() inside
646                                  * priv->lock here, because we know that
647                                  * path->ah will always hold one more reference,
648                                  * so ipoib_put_ah() will never do more than
649                                  * decrement the ref count.
650                                  */
651                                 ipoib_put_ah(neigh->ah);
652                                 list_del(&neigh->list);
653                                 ipoib_neigh_free(neigh);
654                                 spin_unlock(&priv->lock);
655                                 ipoib_path_lookup(skb, dev);
656                                 goto out;
657                         }
658
659                         ipoib_send(dev, skb, neigh->ah,
660                                    be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
661                         goto out;
662                 }
663
664                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
665                         spin_lock(&priv->lock);
666                         __skb_queue_tail(&neigh->queue, skb);
667                         spin_unlock(&priv->lock);
668                 } else {
669                         ++priv->stats.tx_dropped;
670                         dev_kfree_skb_any(skb);
671                 }
672         } else {
673                 struct ipoib_pseudoheader *phdr =
674                         (struct ipoib_pseudoheader *) skb->data;
675                 skb_pull(skb, sizeof *phdr);
676
677                 if (phdr->hwaddr[4] == 0xff) {
678                         /* Add in the P_Key for multicast*/
679                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
680                         phdr->hwaddr[9] = priv->pkey & 0xff;
681
682                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
683                 } else {
684                         /* unicast GID -- should be ARP or RARP reply */
685
686                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
687                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
688                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
689                                            IPOIB_GID_FMT "\n",
690                                            skb->dst ? "neigh" : "dst",
691                                            be16_to_cpup((__be16 *) skb->data),
692                                            be32_to_cpup((__be32 *) phdr->hwaddr),
693                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
694                                 dev_kfree_skb_any(skb);
695                                 ++priv->stats.tx_dropped;
696                                 goto out;
697                         }
698
699                         unicast_arp_send(skb, dev, phdr);
700                 }
701         }
702
703 out:
704         spin_unlock_irqrestore(&priv->tx_lock, flags);
705
706         return NETDEV_TX_OK;
707 }
708
709 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
710 {
711         struct ipoib_dev_priv *priv = netdev_priv(dev);
712
713         return &priv->stats;
714 }
715
716 static void ipoib_timeout(struct net_device *dev)
717 {
718         struct ipoib_dev_priv *priv = netdev_priv(dev);
719
720         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
721                    jiffies_to_msecs(jiffies - dev->trans_start));
722         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
723                    netif_queue_stopped(dev),
724                    priv->tx_head, priv->tx_tail);
725         /* XXX reset QP, etc. */
726 }
727
728 static int ipoib_hard_header(struct sk_buff *skb,
729                              struct net_device *dev,
730                              unsigned short type,
731                              void *daddr, void *saddr, unsigned len)
732 {
733         struct ipoib_header *header;
734
735         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
736
737         header->proto = htons(type);
738         header->reserved = 0;
739
740         /*
741          * If we don't have a neighbour structure, stuff the
742          * destination address onto the front of the skb so we can
743          * figure out where to send the packet later.
744          */
745         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
746                 struct ipoib_pseudoheader *phdr =
747                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
748                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
749         }
750
751         return 0;
752 }
753
754 static void ipoib_set_mcast_list(struct net_device *dev)
755 {
756         struct ipoib_dev_priv *priv = netdev_priv(dev);
757
758         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
759                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
760                 return;
761         }
762
763         queue_work(ipoib_workqueue, &priv->restart_task);
764 }
765
766 static void ipoib_neigh_destructor(struct neighbour *n)
767 {
768         struct ipoib_neigh *neigh;
769         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
770         unsigned long flags;
771         struct ipoib_ah *ah = NULL;
772
773         ipoib_dbg(priv,
774                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
775                   be32_to_cpup((__be32 *) n->ha),
776                   IPOIB_GID_RAW_ARG(n->ha + 4));
777
778         spin_lock_irqsave(&priv->lock, flags);
779
780         neigh = *to_ipoib_neigh(n);
781         if (neigh) {
782                 if (neigh->ah)
783                         ah = neigh->ah;
784                 list_del(&neigh->list);
785                 ipoib_neigh_free(neigh);
786         }
787
788         spin_unlock_irqrestore(&priv->lock, flags);
789
790         if (ah)
791                 ipoib_put_ah(ah);
792 }
793
794 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
795 {
796         struct ipoib_neigh *neigh;
797
798         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
799         if (!neigh)
800                 return NULL;
801
802         neigh->neighbour = neighbour;
803         *to_ipoib_neigh(neighbour) = neigh;
804
805         return neigh;
806 }
807
808 void ipoib_neigh_free(struct ipoib_neigh *neigh)
809 {
810         *to_ipoib_neigh(neigh->neighbour) = NULL;
811         kfree(neigh);
812 }
813
814 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
815 {
816         parms->neigh_destructor = ipoib_neigh_destructor;
817
818         return 0;
819 }
820
821 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
822 {
823         struct ipoib_dev_priv *priv = netdev_priv(dev);
824
825         /* Allocate RX/TX "rings" to hold queued skbs */
826         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
827                                 GFP_KERNEL);
828         if (!priv->rx_ring) {
829                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
830                        ca->name, ipoib_recvq_size);
831                 goto out;
832         }
833
834         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
835                                 GFP_KERNEL);
836         if (!priv->tx_ring) {
837                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
838                        ca->name, ipoib_sendq_size);
839                 goto out_rx_ring_cleanup;
840         }
841
842         /* priv->tx_head & tx_tail are already 0 */
843
844         if (ipoib_ib_dev_init(dev, ca, port))
845                 goto out_tx_ring_cleanup;
846
847         return 0;
848
849 out_tx_ring_cleanup:
850         kfree(priv->tx_ring);
851
852 out_rx_ring_cleanup:
853         kfree(priv->rx_ring);
854
855 out:
856         return -ENOMEM;
857 }
858
859 void ipoib_dev_cleanup(struct net_device *dev)
860 {
861         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
862
863         ipoib_delete_debug_files(dev);
864
865         /* Delete any child interfaces first */
866         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
867                 unregister_netdev(cpriv->dev);
868                 ipoib_dev_cleanup(cpriv->dev);
869                 free_netdev(cpriv->dev);
870         }
871
872         ipoib_ib_dev_cleanup(dev);
873
874         kfree(priv->rx_ring);
875         kfree(priv->tx_ring);
876
877         priv->rx_ring = NULL;
878         priv->tx_ring = NULL;
879 }
880
881 static void ipoib_setup(struct net_device *dev)
882 {
883         struct ipoib_dev_priv *priv = netdev_priv(dev);
884
885         dev->open                = ipoib_open;
886         dev->stop                = ipoib_stop;
887         dev->change_mtu          = ipoib_change_mtu;
888         dev->hard_start_xmit     = ipoib_start_xmit;
889         dev->get_stats           = ipoib_get_stats;
890         dev->tx_timeout          = ipoib_timeout;
891         dev->hard_header         = ipoib_hard_header;
892         dev->set_multicast_list  = ipoib_set_mcast_list;
893         dev->neigh_setup         = ipoib_neigh_setup_dev;
894
895         dev->watchdog_timeo      = HZ;
896
897         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
898
899         /*
900          * We add in INFINIBAND_ALEN to allow for the destination
901          * address "pseudoheader" for skbs without neighbour struct.
902          */
903         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
904         dev->addr_len            = INFINIBAND_ALEN;
905         dev->type                = ARPHRD_INFINIBAND;
906         dev->tx_queue_len        = ipoib_sendq_size * 2;
907         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
908
909         /* MTU will be reset when mcast join happens */
910         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
911         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
912
913         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
914
915         netif_carrier_off(dev);
916
917         SET_MODULE_OWNER(dev);
918
919         priv->dev = dev;
920
921         spin_lock_init(&priv->lock);
922         spin_lock_init(&priv->tx_lock);
923
924         mutex_init(&priv->mcast_mutex);
925         mutex_init(&priv->vlan_mutex);
926
927         INIT_LIST_HEAD(&priv->path_list);
928         INIT_LIST_HEAD(&priv->child_intfs);
929         INIT_LIST_HEAD(&priv->dead_ahs);
930         INIT_LIST_HEAD(&priv->multicast_list);
931
932         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
933         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
934         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
935         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
936         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
937 }
938
939 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
940 {
941         struct net_device *dev;
942
943         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
944                            ipoib_setup);
945         if (!dev)
946                 return NULL;
947
948         return netdev_priv(dev);
949 }
950
951 static ssize_t show_pkey(struct class_device *cdev, char *buf)
952 {
953         struct ipoib_dev_priv *priv =
954                 netdev_priv(container_of(cdev, struct net_device, class_dev));
955
956         return sprintf(buf, "0x%04x\n", priv->pkey);
957 }
958 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
959
960 static ssize_t create_child(struct class_device *cdev,
961                             const char *buf, size_t count)
962 {
963         int pkey;
964         int ret;
965
966         if (sscanf(buf, "%i", &pkey) != 1)
967                 return -EINVAL;
968
969         if (pkey < 0 || pkey > 0xffff)
970                 return -EINVAL;
971
972         /*
973          * Set the full membership bit, so that we join the right
974          * broadcast group, etc.
975          */
976         pkey |= 0x8000;
977
978         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
979                              pkey);
980
981         return ret ? ret : count;
982 }
983 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
984
985 static ssize_t delete_child(struct class_device *cdev,
986                             const char *buf, size_t count)
987 {
988         int pkey;
989         int ret;
990
991         if (sscanf(buf, "%i", &pkey) != 1)
992                 return -EINVAL;
993
994         if (pkey < 0 || pkey > 0xffff)
995                 return -EINVAL;
996
997         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
998                                 pkey);
999
1000         return ret ? ret : count;
1001
1002 }
1003 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1004
1005 int ipoib_add_pkey_attr(struct net_device *dev)
1006 {
1007         return class_device_create_file(&dev->class_dev,
1008                                         &class_device_attr_pkey);
1009 }
1010
1011 static struct net_device *ipoib_add_port(const char *format,
1012                                          struct ib_device *hca, u8 port)
1013 {
1014         struct ipoib_dev_priv *priv;
1015         int result = -ENOMEM;
1016
1017         priv = ipoib_intf_alloc(format);
1018         if (!priv)
1019                 goto alloc_mem_failed;
1020
1021         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1022
1023         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1024         if (result) {
1025                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1026                        hca->name, port, result);
1027                 goto alloc_mem_failed;
1028         }
1029
1030         /*
1031          * Set the full membership bit, so that we join the right
1032          * broadcast group, etc.
1033          */
1034         priv->pkey |= 0x8000;
1035
1036         priv->dev->broadcast[8] = priv->pkey >> 8;
1037         priv->dev->broadcast[9] = priv->pkey & 0xff;
1038
1039         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1040         if (result) {
1041                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1042                        hca->name, port, result);
1043                 goto alloc_mem_failed;
1044         } else
1045                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1046
1047
1048         result = ipoib_dev_init(priv->dev, hca, port);
1049         if (result < 0) {
1050                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1051                        hca->name, port, result);
1052                 goto device_init_failed;
1053         }
1054
1055         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1056                               priv->ca, ipoib_event);
1057         result = ib_register_event_handler(&priv->event_handler);
1058         if (result < 0) {
1059                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1060                        "port %d (ret = %d)\n",
1061                        hca->name, port, result);
1062                 goto event_failed;
1063         }
1064
1065         result = register_netdev(priv->dev);
1066         if (result) {
1067                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1068                        hca->name, port, result);
1069                 goto register_failed;
1070         }
1071
1072         ipoib_create_debug_files(priv->dev);
1073
1074         if (ipoib_add_pkey_attr(priv->dev))
1075                 goto sysfs_failed;
1076         if (class_device_create_file(&priv->dev->class_dev,
1077                                      &class_device_attr_create_child))
1078                 goto sysfs_failed;
1079         if (class_device_create_file(&priv->dev->class_dev,
1080                                      &class_device_attr_delete_child))
1081                 goto sysfs_failed;
1082
1083         return priv->dev;
1084
1085 sysfs_failed:
1086         ipoib_delete_debug_files(priv->dev);
1087         unregister_netdev(priv->dev);
1088
1089 register_failed:
1090         ib_unregister_event_handler(&priv->event_handler);
1091         flush_scheduled_work();
1092
1093 event_failed:
1094         ipoib_dev_cleanup(priv->dev);
1095
1096 device_init_failed:
1097         free_netdev(priv->dev);
1098
1099 alloc_mem_failed:
1100         return ERR_PTR(result);
1101 }
1102
1103 static void ipoib_add_one(struct ib_device *device)
1104 {
1105         struct list_head *dev_list;
1106         struct net_device *dev;
1107         struct ipoib_dev_priv *priv;
1108         int s, e, p;
1109
1110         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1111         if (!dev_list)
1112                 return;
1113
1114         INIT_LIST_HEAD(dev_list);
1115
1116         if (device->node_type == IB_NODE_SWITCH) {
1117                 s = 0;
1118                 e = 0;
1119         } else {
1120                 s = 1;
1121                 e = device->phys_port_cnt;
1122         }
1123
1124         for (p = s; p <= e; ++p) {
1125                 dev = ipoib_add_port("ib%d", device, p);
1126                 if (!IS_ERR(dev)) {
1127                         priv = netdev_priv(dev);
1128                         list_add_tail(&priv->list, dev_list);
1129                 }
1130         }
1131
1132         ib_set_client_data(device, &ipoib_client, dev_list);
1133 }
1134
1135 static void ipoib_remove_one(struct ib_device *device)
1136 {
1137         struct ipoib_dev_priv *priv, *tmp;
1138         struct list_head *dev_list;
1139
1140         dev_list = ib_get_client_data(device, &ipoib_client);
1141
1142         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1143                 ib_unregister_event_handler(&priv->event_handler);
1144                 flush_scheduled_work();
1145
1146                 unregister_netdev(priv->dev);
1147                 ipoib_dev_cleanup(priv->dev);
1148                 free_netdev(priv->dev);
1149         }
1150
1151         kfree(dev_list);
1152 }
1153
1154 static int __init ipoib_init_module(void)
1155 {
1156         int ret;
1157
1158         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1159         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1160         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1161
1162         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1163         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1164         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1165
1166         ret = ipoib_register_debugfs();
1167         if (ret)
1168                 return ret;
1169
1170         /*
1171          * We create our own workqueue mainly because we want to be
1172          * able to flush it when devices are being removed.  We can't
1173          * use schedule_work()/flush_scheduled_work() because both
1174          * unregister_netdev() and linkwatch_event take the rtnl lock,
1175          * so flush_scheduled_work() can deadlock during device
1176          * removal.
1177          */
1178         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1179         if (!ipoib_workqueue) {
1180                 ret = -ENOMEM;
1181                 goto err_fs;
1182         }
1183
1184         ret = ib_register_client(&ipoib_client);
1185         if (ret)
1186                 goto err_wq;
1187
1188         return 0;
1189
1190 err_wq:
1191         destroy_workqueue(ipoib_workqueue);
1192
1193 err_fs:
1194         ipoib_unregister_debugfs();
1195
1196         return ret;
1197 }
1198
1199 static void __exit ipoib_cleanup_module(void)
1200 {
1201         ib_unregister_client(&ipoib_client);
1202         ipoib_unregister_debugfs();
1203         destroy_workqueue(ipoib_workqueue);
1204 }
1205
1206 module_init(ipoib_init_module);
1207 module_exit(ipoib_cleanup_module);