Merge branch 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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
35 #include "ipoib.h"
36
37 #include <linux/module.h>
38
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
43
44 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
45
46 #include <linux/ip.h>
47 #include <linux/in.h>
48
49 #include <net/dst.h>
50
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
53 MODULE_LICENSE("Dual BSD/GPL");
54
55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
57
58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
62
63 static int lro;
64 module_param(lro, bool, 0444);
65 MODULE_PARM_DESC(lro,  "Enable LRO (Large Receive Offload)");
66
67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
68 module_param(lro_max_aggr, int, 0644);
69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
70                 "(default = 64)");
71
72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
73 int ipoib_debug_level;
74
75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
77 #endif
78
79 struct ipoib_path_iter {
80         struct net_device *dev;
81         struct ipoib_path  path;
82 };
83
84 static const u8 ipv4_bcast_addr[] = {
85         0x00, 0xff, 0xff, 0xff,
86         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
87         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
88 };
89
90 struct workqueue_struct *ipoib_workqueue;
91
92 struct ib_sa_client ipoib_sa_client;
93
94 static void ipoib_add_one(struct ib_device *device);
95 static void ipoib_remove_one(struct ib_device *device);
96
97 static struct ib_client ipoib_client = {
98         .name   = "ipoib",
99         .add    = ipoib_add_one,
100         .remove = ipoib_remove_one
101 };
102
103 int ipoib_open(struct net_device *dev)
104 {
105         struct ipoib_dev_priv *priv = netdev_priv(dev);
106
107         ipoib_dbg(priv, "bringing up interface\n");
108
109         napi_enable(&priv->napi);
110         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
111
112         if (ipoib_pkey_dev_delay_open(dev))
113                 return 0;
114
115         if (ipoib_ib_dev_open(dev)) {
116                 napi_disable(&priv->napi);
117                 return -EINVAL;
118         }
119
120         if (ipoib_ib_dev_up(dev)) {
121                 ipoib_ib_dev_stop(dev, 1);
122                 napi_disable(&priv->napi);
123                 return -EINVAL;
124         }
125
126         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
127                 struct ipoib_dev_priv *cpriv;
128
129                 /* Bring up any child interfaces too */
130                 mutex_lock(&priv->vlan_mutex);
131                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
132                         int flags;
133
134                         flags = cpriv->dev->flags;
135                         if (flags & IFF_UP)
136                                 continue;
137
138                         dev_change_flags(cpriv->dev, flags | IFF_UP);
139                 }
140                 mutex_unlock(&priv->vlan_mutex);
141         }
142
143         netif_start_queue(dev);
144
145         return 0;
146 }
147
148 static int ipoib_stop(struct net_device *dev)
149 {
150         struct ipoib_dev_priv *priv = netdev_priv(dev);
151
152         ipoib_dbg(priv, "stopping interface\n");
153
154         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
155         napi_disable(&priv->napi);
156
157         netif_stop_queue(dev);
158
159         /*
160          * Now flush workqueue to make sure a scheduled task doesn't
161          * bring our internal state back up.
162          */
163         flush_workqueue(ipoib_workqueue);
164
165         ipoib_ib_dev_down(dev, 1);
166         ipoib_ib_dev_stop(dev, 1);
167
168         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
169                 struct ipoib_dev_priv *cpriv;
170
171                 /* Bring down any child interfaces too */
172                 mutex_lock(&priv->vlan_mutex);
173                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
174                         int flags;
175
176                         flags = cpriv->dev->flags;
177                         if (!(flags & IFF_UP))
178                                 continue;
179
180                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
181                 }
182                 mutex_unlock(&priv->vlan_mutex);
183         }
184
185         return 0;
186 }
187
188 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
189 {
190         struct ipoib_dev_priv *priv = netdev_priv(dev);
191
192         /* dev->mtu > 2K ==> connected mode */
193         if (ipoib_cm_admin_enabled(dev)) {
194                 if (new_mtu > ipoib_cm_max_mtu(dev))
195                         return -EINVAL;
196
197                 if (new_mtu > priv->mcast_mtu)
198                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
199                                    priv->mcast_mtu);
200
201                 dev->mtu = new_mtu;
202                 return 0;
203         }
204
205         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
206                 return -EINVAL;
207
208         priv->admin_mtu = new_mtu;
209
210         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
211
212         return 0;
213 }
214
215 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
216 {
217         struct ipoib_dev_priv *priv = netdev_priv(dev);
218         struct rb_node *n = priv->path_tree.rb_node;
219         struct ipoib_path *path;
220         int ret;
221
222         while (n) {
223                 path = rb_entry(n, struct ipoib_path, rb_node);
224
225                 ret = memcmp(gid, path->pathrec.dgid.raw,
226                              sizeof (union ib_gid));
227
228                 if (ret < 0)
229                         n = n->rb_left;
230                 else if (ret > 0)
231                         n = n->rb_right;
232                 else
233                         return path;
234         }
235
236         return NULL;
237 }
238
239 static int __path_add(struct net_device *dev, struct ipoib_path *path)
240 {
241         struct ipoib_dev_priv *priv = netdev_priv(dev);
242         struct rb_node **n = &priv->path_tree.rb_node;
243         struct rb_node *pn = NULL;
244         struct ipoib_path *tpath;
245         int ret;
246
247         while (*n) {
248                 pn = *n;
249                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
250
251                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
252                              sizeof (union ib_gid));
253                 if (ret < 0)
254                         n = &pn->rb_left;
255                 else if (ret > 0)
256                         n = &pn->rb_right;
257                 else
258                         return -EEXIST;
259         }
260
261         rb_link_node(&path->rb_node, pn, n);
262         rb_insert_color(&path->rb_node, &priv->path_tree);
263
264         list_add_tail(&path->list, &priv->path_list);
265
266         return 0;
267 }
268
269 static void path_free(struct net_device *dev, struct ipoib_path *path)
270 {
271         struct ipoib_dev_priv *priv = netdev_priv(dev);
272         struct ipoib_neigh *neigh, *tn;
273         struct sk_buff *skb;
274         unsigned long flags;
275
276         while ((skb = __skb_dequeue(&path->queue)))
277                 dev_kfree_skb_irq(skb);
278
279         spin_lock_irqsave(&priv->lock, flags);
280
281         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
282                 /*
283                  * It's safe to call ipoib_put_ah() inside priv->lock
284                  * here, because we know that path->ah will always
285                  * hold one more reference, so ipoib_put_ah() will
286                  * never do more than decrement the ref count.
287                  */
288                 if (neigh->ah)
289                         ipoib_put_ah(neigh->ah);
290
291                 ipoib_neigh_free(dev, neigh);
292         }
293
294         spin_unlock_irqrestore(&priv->lock, flags);
295
296         if (path->ah)
297                 ipoib_put_ah(path->ah);
298
299         kfree(path);
300 }
301
302 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
303
304 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
305 {
306         struct ipoib_path_iter *iter;
307
308         iter = kmalloc(sizeof *iter, GFP_KERNEL);
309         if (!iter)
310                 return NULL;
311
312         iter->dev = dev;
313         memset(iter->path.pathrec.dgid.raw, 0, 16);
314
315         if (ipoib_path_iter_next(iter)) {
316                 kfree(iter);
317                 return NULL;
318         }
319
320         return iter;
321 }
322
323 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
324 {
325         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
326         struct rb_node *n;
327         struct ipoib_path *path;
328         int ret = 1;
329
330         spin_lock_irq(&priv->lock);
331
332         n = rb_first(&priv->path_tree);
333
334         while (n) {
335                 path = rb_entry(n, struct ipoib_path, rb_node);
336
337                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
338                            sizeof (union ib_gid)) < 0) {
339                         iter->path = *path;
340                         ret = 0;
341                         break;
342                 }
343
344                 n = rb_next(n);
345         }
346
347         spin_unlock_irq(&priv->lock);
348
349         return ret;
350 }
351
352 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
353                           struct ipoib_path *path)
354 {
355         *path = iter->path;
356 }
357
358 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
359
360 void ipoib_mark_paths_invalid(struct net_device *dev)
361 {
362         struct ipoib_dev_priv *priv = netdev_priv(dev);
363         struct ipoib_path *path, *tp;
364
365         spin_lock_irq(&priv->lock);
366
367         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
368                 ipoib_dbg(priv, "mark path LID 0x%04x GID " IPOIB_GID_FMT " invalid\n",
369                         be16_to_cpu(path->pathrec.dlid),
370                         IPOIB_GID_ARG(path->pathrec.dgid));
371                 path->valid =  0;
372         }
373
374         spin_unlock_irq(&priv->lock);
375 }
376
377 void ipoib_flush_paths(struct net_device *dev)
378 {
379         struct ipoib_dev_priv *priv = netdev_priv(dev);
380         struct ipoib_path *path, *tp;
381         LIST_HEAD(remove_list);
382
383         spin_lock_irq(&priv->tx_lock);
384         spin_lock(&priv->lock);
385
386         list_splice_init(&priv->path_list, &remove_list);
387
388         list_for_each_entry(path, &remove_list, list)
389                 rb_erase(&path->rb_node, &priv->path_tree);
390
391         list_for_each_entry_safe(path, tp, &remove_list, list) {
392                 if (path->query)
393                         ib_sa_cancel_query(path->query_id, path->query);
394                 spin_unlock(&priv->lock);
395                 spin_unlock_irq(&priv->tx_lock);
396                 wait_for_completion(&path->done);
397                 path_free(dev, path);
398                 spin_lock_irq(&priv->tx_lock);
399                 spin_lock(&priv->lock);
400         }
401         spin_unlock(&priv->lock);
402         spin_unlock_irq(&priv->tx_lock);
403 }
404
405 static void path_rec_completion(int status,
406                                 struct ib_sa_path_rec *pathrec,
407                                 void *path_ptr)
408 {
409         struct ipoib_path *path = path_ptr;
410         struct net_device *dev = path->dev;
411         struct ipoib_dev_priv *priv = netdev_priv(dev);
412         struct ipoib_ah *ah = NULL;
413         struct ipoib_ah *old_ah;
414         struct ipoib_neigh *neigh, *tn;
415         struct sk_buff_head skqueue;
416         struct sk_buff *skb;
417         unsigned long flags;
418
419         if (!status)
420                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
421                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
422         else
423                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
424                           status, IPOIB_GID_ARG(path->pathrec.dgid));
425
426         skb_queue_head_init(&skqueue);
427
428         if (!status) {
429                 struct ib_ah_attr av;
430
431                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
432                         ah = ipoib_create_ah(dev, priv->pd, &av);
433         }
434
435         spin_lock_irqsave(&priv->lock, flags);
436
437         old_ah   = path->ah;
438         path->ah = ah;
439
440         if (ah) {
441                 path->pathrec = *pathrec;
442
443                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
444                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
445
446                 while ((skb = __skb_dequeue(&path->queue)))
447                         __skb_queue_tail(&skqueue, skb);
448
449                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
450                         if (neigh->ah) {
451                                 WARN_ON(neigh->ah != old_ah);
452                                 /*
453                                  * Dropping the ah reference inside
454                                  * priv->lock is safe here, because we
455                                  * will hold one more reference from
456                                  * the original value of path->ah (ie
457                                  * old_ah).
458                                  */
459                                 ipoib_put_ah(neigh->ah);
460                         }
461                         kref_get(&path->ah->ref);
462                         neigh->ah = path->ah;
463                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
464                                sizeof(union ib_gid));
465
466                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
467                                 if (!ipoib_cm_get(neigh))
468                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
469                                                                                path,
470                                                                                neigh));
471                                 if (!ipoib_cm_get(neigh)) {
472                                         list_del(&neigh->list);
473                                         if (neigh->ah)
474                                                 ipoib_put_ah(neigh->ah);
475                                         ipoib_neigh_free(dev, neigh);
476                                         continue;
477                                 }
478                         }
479
480                         while ((skb = __skb_dequeue(&neigh->queue)))
481                                 __skb_queue_tail(&skqueue, skb);
482                 }
483                 path->valid = 1;
484         }
485
486         path->query = NULL;
487         complete(&path->done);
488
489         spin_unlock_irqrestore(&priv->lock, flags);
490
491         if (old_ah)
492                 ipoib_put_ah(old_ah);
493
494         while ((skb = __skb_dequeue(&skqueue))) {
495                 skb->dev = dev;
496                 if (dev_queue_xmit(skb))
497                         ipoib_warn(priv, "dev_queue_xmit failed "
498                                    "to requeue packet\n");
499         }
500 }
501
502 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
503 {
504         struct ipoib_dev_priv *priv = netdev_priv(dev);
505         struct ipoib_path *path;
506
507         if (!priv->broadcast)
508                 return NULL;
509
510         path = kzalloc(sizeof *path, GFP_ATOMIC);
511         if (!path)
512                 return NULL;
513
514         path->dev = dev;
515
516         skb_queue_head_init(&path->queue);
517
518         INIT_LIST_HEAD(&path->neigh_list);
519
520         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
521         path->pathrec.sgid          = priv->local_gid;
522         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
523         path->pathrec.numb_path     = 1;
524         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
525
526         return path;
527 }
528
529 static int path_rec_start(struct net_device *dev,
530                           struct ipoib_path *path)
531 {
532         struct ipoib_dev_priv *priv = netdev_priv(dev);
533
534         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
535                   IPOIB_GID_ARG(path->pathrec.dgid));
536
537         init_completion(&path->done);
538
539         path->query_id =
540                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
541                                    &path->pathrec,
542                                    IB_SA_PATH_REC_DGID          |
543                                    IB_SA_PATH_REC_SGID          |
544                                    IB_SA_PATH_REC_NUMB_PATH     |
545                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
546                                    IB_SA_PATH_REC_PKEY,
547                                    1000, GFP_ATOMIC,
548                                    path_rec_completion,
549                                    path, &path->query);
550         if (path->query_id < 0) {
551                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
552                 path->query = NULL;
553                 return path->query_id;
554         }
555
556         return 0;
557 }
558
559 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
560 {
561         struct ipoib_dev_priv *priv = netdev_priv(dev);
562         struct ipoib_path *path;
563         struct ipoib_neigh *neigh;
564
565         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
566         if (!neigh) {
567                 ++dev->stats.tx_dropped;
568                 dev_kfree_skb_any(skb);
569                 return;
570         }
571
572         /*
573          * We can only be called from ipoib_start_xmit, so we're
574          * inside tx_lock -- no need to save/restore flags.
575          */
576         spin_lock(&priv->lock);
577
578         path = __path_find(dev, skb->dst->neighbour->ha + 4);
579         if (!path) {
580                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
581                 if (!path)
582                         goto err_path;
583
584                 __path_add(dev, path);
585         }
586
587         list_add_tail(&neigh->list, &path->neigh_list);
588
589         if (path->ah) {
590                 kref_get(&path->ah->ref);
591                 neigh->ah = path->ah;
592                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
593                        sizeof(union ib_gid));
594
595                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
596                         if (!ipoib_cm_get(neigh))
597                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
598                         if (!ipoib_cm_get(neigh)) {
599                                 list_del(&neigh->list);
600                                 if (neigh->ah)
601                                         ipoib_put_ah(neigh->ah);
602                                 ipoib_neigh_free(dev, neigh);
603                                 goto err_drop;
604                         }
605                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
606                                 __skb_queue_tail(&neigh->queue, skb);
607                         else {
608                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
609                                            skb_queue_len(&neigh->queue));
610                                 goto err_drop;
611                         }
612                 } else
613                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
614         } else {
615                 neigh->ah  = NULL;
616
617                 if (!path->query && path_rec_start(dev, path))
618                         goto err_list;
619
620                 __skb_queue_tail(&neigh->queue, skb);
621         }
622
623         spin_unlock(&priv->lock);
624         return;
625
626 err_list:
627         list_del(&neigh->list);
628
629 err_path:
630         ipoib_neigh_free(dev, neigh);
631 err_drop:
632         ++dev->stats.tx_dropped;
633         dev_kfree_skb_any(skb);
634
635         spin_unlock(&priv->lock);
636 }
637
638 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
639 {
640         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
641
642         /* Look up path record for unicasts */
643         if (skb->dst->neighbour->ha[4] != 0xff) {
644                 neigh_add_path(skb, dev);
645                 return;
646         }
647
648         /* Add in the P_Key for multicasts */
649         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
650         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
651         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
652 }
653
654 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
655                              struct ipoib_pseudoheader *phdr)
656 {
657         struct ipoib_dev_priv *priv = netdev_priv(dev);
658         struct ipoib_path *path;
659
660         /*
661          * We can only be called from ipoib_start_xmit, so we're
662          * inside tx_lock -- no need to save/restore flags.
663          */
664         spin_lock(&priv->lock);
665
666         path = __path_find(dev, phdr->hwaddr + 4);
667         if (!path || !path->valid) {
668                 if (!path)
669                         path = path_rec_create(dev, phdr->hwaddr + 4);
670                 if (path) {
671                         /* put pseudoheader back on for next time */
672                         skb_push(skb, sizeof *phdr);
673                         __skb_queue_tail(&path->queue, skb);
674
675                         if (path_rec_start(dev, path)) {
676                                 spin_unlock(&priv->lock);
677                                 path_free(dev, path);
678                                 return;
679                         } else
680                                 __path_add(dev, path);
681                 } else {
682                         ++dev->stats.tx_dropped;
683                         dev_kfree_skb_any(skb);
684                 }
685
686                 spin_unlock(&priv->lock);
687                 return;
688         }
689
690         if (path->ah) {
691                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
692                           be16_to_cpu(path->pathrec.dlid));
693
694                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
695         } else if ((path->query || !path_rec_start(dev, path)) &&
696                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
697                 /* put pseudoheader back on for next time */
698                 skb_push(skb, sizeof *phdr);
699                 __skb_queue_tail(&path->queue, skb);
700         } else {
701                 ++dev->stats.tx_dropped;
702                 dev_kfree_skb_any(skb);
703         }
704
705         spin_unlock(&priv->lock);
706 }
707
708 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
709 {
710         struct ipoib_dev_priv *priv = netdev_priv(dev);
711         struct ipoib_neigh *neigh;
712         unsigned long flags;
713
714         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
715                 return NETDEV_TX_LOCKED;
716
717         if (likely(skb->dst && skb->dst->neighbour)) {
718                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
719                         ipoib_path_lookup(skb, dev);
720                         goto out;
721                 }
722
723                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
724
725                 if (neigh->ah)
726                         if (unlikely((memcmp(&neigh->dgid.raw,
727                                             skb->dst->neighbour->ha + 4,
728                                             sizeof(union ib_gid))) ||
729                                          (neigh->dev != dev))) {
730                                 spin_lock(&priv->lock);
731                                 /*
732                                  * It's safe to call ipoib_put_ah() inside
733                                  * priv->lock here, because we know that
734                                  * path->ah will always hold one more reference,
735                                  * so ipoib_put_ah() will never do more than
736                                  * decrement the ref count.
737                                  */
738                                 ipoib_put_ah(neigh->ah);
739                                 list_del(&neigh->list);
740                                 ipoib_neigh_free(dev, neigh);
741                                 spin_unlock(&priv->lock);
742                                 ipoib_path_lookup(skb, dev);
743                                 goto out;
744                         }
745
746                 if (ipoib_cm_get(neigh)) {
747                         if (ipoib_cm_up(neigh)) {
748                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
749                                 goto out;
750                         }
751                 } else if (neigh->ah) {
752                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
753                         goto out;
754                 }
755
756                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
757                         spin_lock(&priv->lock);
758                         __skb_queue_tail(&neigh->queue, skb);
759                         spin_unlock(&priv->lock);
760                 } else {
761                         ++dev->stats.tx_dropped;
762                         dev_kfree_skb_any(skb);
763                 }
764         } else {
765                 struct ipoib_pseudoheader *phdr =
766                         (struct ipoib_pseudoheader *) skb->data;
767                 skb_pull(skb, sizeof *phdr);
768
769                 if (phdr->hwaddr[4] == 0xff) {
770                         /* Add in the P_Key for multicast*/
771                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
772                         phdr->hwaddr[9] = priv->pkey & 0xff;
773
774                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
775                 } else {
776                         /* unicast GID -- should be ARP or RARP reply */
777
778                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
779                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
780                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
781                                            IPOIB_GID_FMT "\n",
782                                            skb->dst ? "neigh" : "dst",
783                                            be16_to_cpup((__be16 *) skb->data),
784                                            IPOIB_QPN(phdr->hwaddr),
785                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
786                                 dev_kfree_skb_any(skb);
787                                 ++dev->stats.tx_dropped;
788                                 goto out;
789                         }
790
791                         unicast_arp_send(skb, dev, phdr);
792                 }
793         }
794
795 out:
796         spin_unlock_irqrestore(&priv->tx_lock, flags);
797
798         return NETDEV_TX_OK;
799 }
800
801 static void ipoib_timeout(struct net_device *dev)
802 {
803         struct ipoib_dev_priv *priv = netdev_priv(dev);
804
805         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
806                    jiffies_to_msecs(jiffies - dev->trans_start));
807         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
808                    netif_queue_stopped(dev),
809                    priv->tx_head, priv->tx_tail);
810         /* XXX reset QP, etc. */
811 }
812
813 static int ipoib_hard_header(struct sk_buff *skb,
814                              struct net_device *dev,
815                              unsigned short type,
816                              const void *daddr, const void *saddr, unsigned len)
817 {
818         struct ipoib_header *header;
819
820         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
821
822         header->proto = htons(type);
823         header->reserved = 0;
824
825         /*
826          * If we don't have a neighbour structure, stuff the
827          * destination address onto the front of the skb so we can
828          * figure out where to send the packet later.
829          */
830         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
831                 struct ipoib_pseudoheader *phdr =
832                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
833                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
834         }
835
836         return 0;
837 }
838
839 static void ipoib_set_mcast_list(struct net_device *dev)
840 {
841         struct ipoib_dev_priv *priv = netdev_priv(dev);
842
843         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
844                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
845                 return;
846         }
847
848         queue_work(ipoib_workqueue, &priv->restart_task);
849 }
850
851 static void ipoib_neigh_cleanup(struct neighbour *n)
852 {
853         struct ipoib_neigh *neigh;
854         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
855         unsigned long flags;
856         struct ipoib_ah *ah = NULL;
857
858         neigh = *to_ipoib_neigh(n);
859         if (neigh)
860                 priv = netdev_priv(neigh->dev);
861         else
862                 return;
863         ipoib_dbg(priv,
864                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
865                   IPOIB_QPN(n->ha),
866                   IPOIB_GID_RAW_ARG(n->ha + 4));
867
868         spin_lock_irqsave(&priv->lock, flags);
869
870         if (neigh->ah)
871                 ah = neigh->ah;
872         list_del(&neigh->list);
873         ipoib_neigh_free(n->dev, neigh);
874
875         spin_unlock_irqrestore(&priv->lock, flags);
876
877         if (ah)
878                 ipoib_put_ah(ah);
879 }
880
881 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
882                                       struct net_device *dev)
883 {
884         struct ipoib_neigh *neigh;
885
886         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
887         if (!neigh)
888                 return NULL;
889
890         neigh->neighbour = neighbour;
891         neigh->dev = dev;
892         *to_ipoib_neigh(neighbour) = neigh;
893         skb_queue_head_init(&neigh->queue);
894         ipoib_cm_set(neigh, NULL);
895
896         return neigh;
897 }
898
899 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
900 {
901         struct sk_buff *skb;
902         *to_ipoib_neigh(neigh->neighbour) = NULL;
903         while ((skb = __skb_dequeue(&neigh->queue))) {
904                 ++dev->stats.tx_dropped;
905                 dev_kfree_skb_any(skb);
906         }
907         if (ipoib_cm_get(neigh))
908                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
909         kfree(neigh);
910 }
911
912 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
913 {
914         parms->neigh_cleanup = ipoib_neigh_cleanup;
915
916         return 0;
917 }
918
919 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
920 {
921         struct ipoib_dev_priv *priv = netdev_priv(dev);
922
923         /* Allocate RX/TX "rings" to hold queued skbs */
924         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
925                                 GFP_KERNEL);
926         if (!priv->rx_ring) {
927                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
928                        ca->name, ipoib_recvq_size);
929                 goto out;
930         }
931
932         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
933         if (!priv->tx_ring) {
934                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
935                        ca->name, ipoib_sendq_size);
936                 goto out_rx_ring_cleanup;
937         }
938         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
939
940         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
941
942         if (ipoib_ib_dev_init(dev, ca, port))
943                 goto out_tx_ring_cleanup;
944
945         return 0;
946
947 out_tx_ring_cleanup:
948         vfree(priv->tx_ring);
949
950 out_rx_ring_cleanup:
951         kfree(priv->rx_ring);
952
953 out:
954         return -ENOMEM;
955 }
956
957 void ipoib_dev_cleanup(struct net_device *dev)
958 {
959         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
960
961         ipoib_delete_debug_files(dev);
962
963         /* Delete any child interfaces first */
964         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
965                 unregister_netdev(cpriv->dev);
966                 ipoib_dev_cleanup(cpriv->dev);
967                 free_netdev(cpriv->dev);
968         }
969
970         ipoib_ib_dev_cleanup(dev);
971
972         kfree(priv->rx_ring);
973         vfree(priv->tx_ring);
974
975         priv->rx_ring = NULL;
976         priv->tx_ring = NULL;
977 }
978
979 static const struct header_ops ipoib_header_ops = {
980         .create = ipoib_hard_header,
981 };
982
983 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
984                        void **tcph, u64 *hdr_flags, void *priv)
985 {
986         unsigned int ip_len;
987         struct iphdr *iph;
988
989         if (unlikely(skb->protocol != htons(ETH_P_IP)))
990                 return -1;
991
992         /*
993          * In the future we may add an else clause that verifies the
994          * checksum and allows devices which do not calculate checksum
995          * to use LRO.
996          */
997         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
998                 return -1;
999
1000         /* Check for non-TCP packet */
1001         skb_reset_network_header(skb);
1002         iph = ip_hdr(skb);
1003         if (iph->protocol != IPPROTO_TCP)
1004                 return -1;
1005
1006         ip_len = ip_hdrlen(skb);
1007         skb_set_transport_header(skb, ip_len);
1008         *tcph = tcp_hdr(skb);
1009
1010         /* check if IP header and TCP header are complete */
1011         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1012                 return -1;
1013
1014         *hdr_flags = LRO_IPV4 | LRO_TCP;
1015         *iphdr = iph;
1016
1017         return 0;
1018 }
1019
1020 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1021 {
1022         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1023         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1024         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1025         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1026         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1027         priv->lro.lro_mgr.dev            = priv->dev;
1028         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1029 }
1030
1031 static void ipoib_setup(struct net_device *dev)
1032 {
1033         struct ipoib_dev_priv *priv = netdev_priv(dev);
1034
1035         dev->open                = ipoib_open;
1036         dev->stop                = ipoib_stop;
1037         dev->change_mtu          = ipoib_change_mtu;
1038         dev->hard_start_xmit     = ipoib_start_xmit;
1039         dev->tx_timeout          = ipoib_timeout;
1040         dev->header_ops          = &ipoib_header_ops;
1041         dev->set_multicast_list  = ipoib_set_mcast_list;
1042         dev->neigh_setup         = ipoib_neigh_setup_dev;
1043
1044         ipoib_set_ethtool_ops(dev);
1045
1046         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1047
1048         dev->watchdog_timeo      = HZ;
1049
1050         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1051
1052         /*
1053          * We add in INFINIBAND_ALEN to allow for the destination
1054          * address "pseudoheader" for skbs without neighbour struct.
1055          */
1056         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1057         dev->addr_len            = INFINIBAND_ALEN;
1058         dev->type                = ARPHRD_INFINIBAND;
1059         dev->tx_queue_len        = ipoib_sendq_size * 2;
1060         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1061                                     NETIF_F_LLTX                |
1062                                     NETIF_F_HIGHDMA);
1063
1064         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1065
1066         netif_carrier_off(dev);
1067
1068         priv->dev = dev;
1069
1070         ipoib_lro_setup(priv);
1071
1072         spin_lock_init(&priv->lock);
1073         spin_lock_init(&priv->tx_lock);
1074
1075         mutex_init(&priv->vlan_mutex);
1076
1077         INIT_LIST_HEAD(&priv->path_list);
1078         INIT_LIST_HEAD(&priv->child_intfs);
1079         INIT_LIST_HEAD(&priv->dead_ahs);
1080         INIT_LIST_HEAD(&priv->multicast_list);
1081
1082         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1083         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1084         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1085         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1086         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1087         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1088         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1089 }
1090
1091 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1092 {
1093         struct net_device *dev;
1094
1095         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1096                            ipoib_setup);
1097         if (!dev)
1098                 return NULL;
1099
1100         return netdev_priv(dev);
1101 }
1102
1103 static ssize_t show_pkey(struct device *dev,
1104                          struct device_attribute *attr, char *buf)
1105 {
1106         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1107
1108         return sprintf(buf, "0x%04x\n", priv->pkey);
1109 }
1110 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1111
1112 static ssize_t show_umcast(struct device *dev,
1113                            struct device_attribute *attr, char *buf)
1114 {
1115         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1116
1117         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1118 }
1119
1120 static ssize_t set_umcast(struct device *dev,
1121                           struct device_attribute *attr,
1122                           const char *buf, size_t count)
1123 {
1124         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1125         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1126
1127         if (umcast_val > 0) {
1128                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1129                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1130                                 "by userspace\n");
1131         } else
1132                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1133
1134         return count;
1135 }
1136 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1137
1138 int ipoib_add_umcast_attr(struct net_device *dev)
1139 {
1140         return device_create_file(&dev->dev, &dev_attr_umcast);
1141 }
1142
1143 static ssize_t create_child(struct device *dev,
1144                             struct device_attribute *attr,
1145                             const char *buf, size_t count)
1146 {
1147         int pkey;
1148         int ret;
1149
1150         if (sscanf(buf, "%i", &pkey) != 1)
1151                 return -EINVAL;
1152
1153         if (pkey < 0 || pkey > 0xffff)
1154                 return -EINVAL;
1155
1156         /*
1157          * Set the full membership bit, so that we join the right
1158          * broadcast group, etc.
1159          */
1160         pkey |= 0x8000;
1161
1162         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1163
1164         return ret ? ret : count;
1165 }
1166 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1167
1168 static ssize_t delete_child(struct device *dev,
1169                             struct device_attribute *attr,
1170                             const char *buf, size_t count)
1171 {
1172         int pkey;
1173         int ret;
1174
1175         if (sscanf(buf, "%i", &pkey) != 1)
1176                 return -EINVAL;
1177
1178         if (pkey < 0 || pkey > 0xffff)
1179                 return -EINVAL;
1180
1181         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1182
1183         return ret ? ret : count;
1184
1185 }
1186 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1187
1188 int ipoib_add_pkey_attr(struct net_device *dev)
1189 {
1190         return device_create_file(&dev->dev, &dev_attr_pkey);
1191 }
1192
1193 static struct net_device *ipoib_add_port(const char *format,
1194                                          struct ib_device *hca, u8 port)
1195 {
1196         struct ipoib_dev_priv *priv;
1197         struct ib_device_attr *device_attr;
1198         struct ib_port_attr attr;
1199         int result = -ENOMEM;
1200
1201         priv = ipoib_intf_alloc(format);
1202         if (!priv)
1203                 goto alloc_mem_failed;
1204
1205         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1206
1207         if (!ib_query_port(hca, port, &attr))
1208                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1209         else {
1210                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1211                        hca->name, port);
1212                 goto device_init_failed;
1213         }
1214
1215         /* MTU will be reset when mcast join happens */
1216         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1217         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1218
1219         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1220         if (result) {
1221                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1222                        hca->name, port, result);
1223                 goto device_init_failed;
1224         }
1225
1226         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1227         if (!device_attr) {
1228                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1229                        hca->name, sizeof *device_attr);
1230                 goto device_init_failed;
1231         }
1232
1233         result = ib_query_device(hca, device_attr);
1234         if (result) {
1235                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1236                        hca->name, result);
1237                 kfree(device_attr);
1238                 goto device_init_failed;
1239         }
1240         priv->hca_caps = device_attr->device_cap_flags;
1241
1242         kfree(device_attr);
1243
1244         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1245                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1246                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1247         }
1248
1249         if (lro)
1250                 priv->dev->features |= NETIF_F_LRO;
1251
1252         /*
1253          * Set the full membership bit, so that we join the right
1254          * broadcast group, etc.
1255          */
1256         priv->pkey |= 0x8000;
1257
1258         priv->dev->broadcast[8] = priv->pkey >> 8;
1259         priv->dev->broadcast[9] = priv->pkey & 0xff;
1260
1261         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1262         if (result) {
1263                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1264                        hca->name, port, result);
1265                 goto device_init_failed;
1266         } else
1267                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1268
1269         result = ipoib_dev_init(priv->dev, hca, port);
1270         if (result < 0) {
1271                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1272                        hca->name, port, result);
1273                 goto device_init_failed;
1274         }
1275
1276         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1277                               priv->ca, ipoib_event);
1278         result = ib_register_event_handler(&priv->event_handler);
1279         if (result < 0) {
1280                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1281                        "port %d (ret = %d)\n",
1282                        hca->name, port, result);
1283                 goto event_failed;
1284         }
1285
1286         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1287                 priv->dev->features |= NETIF_F_TSO;
1288
1289         result = register_netdev(priv->dev);
1290         if (result) {
1291                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1292                        hca->name, port, result);
1293                 goto register_failed;
1294         }
1295
1296         ipoib_create_debug_files(priv->dev);
1297
1298         if (ipoib_cm_add_mode_attr(priv->dev))
1299                 goto sysfs_failed;
1300         if (ipoib_add_pkey_attr(priv->dev))
1301                 goto sysfs_failed;
1302         if (ipoib_add_umcast_attr(priv->dev))
1303                 goto sysfs_failed;
1304         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1305                 goto sysfs_failed;
1306         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1307                 goto sysfs_failed;
1308
1309         return priv->dev;
1310
1311 sysfs_failed:
1312         ipoib_delete_debug_files(priv->dev);
1313         unregister_netdev(priv->dev);
1314
1315 register_failed:
1316         ib_unregister_event_handler(&priv->event_handler);
1317         flush_scheduled_work();
1318
1319 event_failed:
1320         ipoib_dev_cleanup(priv->dev);
1321
1322 device_init_failed:
1323         free_netdev(priv->dev);
1324
1325 alloc_mem_failed:
1326         return ERR_PTR(result);
1327 }
1328
1329 static void ipoib_add_one(struct ib_device *device)
1330 {
1331         struct list_head *dev_list;
1332         struct net_device *dev;
1333         struct ipoib_dev_priv *priv;
1334         int s, e, p;
1335
1336         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1337                 return;
1338
1339         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1340         if (!dev_list)
1341                 return;
1342
1343         INIT_LIST_HEAD(dev_list);
1344
1345         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1346                 s = 0;
1347                 e = 0;
1348         } else {
1349                 s = 1;
1350                 e = device->phys_port_cnt;
1351         }
1352
1353         for (p = s; p <= e; ++p) {
1354                 dev = ipoib_add_port("ib%d", device, p);
1355                 if (!IS_ERR(dev)) {
1356                         priv = netdev_priv(dev);
1357                         list_add_tail(&priv->list, dev_list);
1358                 }
1359         }
1360
1361         ib_set_client_data(device, &ipoib_client, dev_list);
1362 }
1363
1364 static void ipoib_remove_one(struct ib_device *device)
1365 {
1366         struct ipoib_dev_priv *priv, *tmp;
1367         struct list_head *dev_list;
1368
1369         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1370                 return;
1371
1372         dev_list = ib_get_client_data(device, &ipoib_client);
1373
1374         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1375                 ib_unregister_event_handler(&priv->event_handler);
1376                 flush_scheduled_work();
1377
1378                 unregister_netdev(priv->dev);
1379                 ipoib_dev_cleanup(priv->dev);
1380                 free_netdev(priv->dev);
1381         }
1382
1383         kfree(dev_list);
1384 }
1385
1386 static int __init ipoib_init_module(void)
1387 {
1388         int ret;
1389
1390         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1391         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1392         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1393
1394         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1395         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1396         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1397                                                      IPOIB_MIN_QUEUE_SIZE));
1398 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1399         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1400 #endif
1401
1402         /*
1403          * When copying small received packets, we only copy from the
1404          * linear data part of the SKB, so we rely on this condition.
1405          */
1406         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1407
1408         ret = ipoib_register_debugfs();
1409         if (ret)
1410                 return ret;
1411
1412         /*
1413          * We create our own workqueue mainly because we want to be
1414          * able to flush it when devices are being removed.  We can't
1415          * use schedule_work()/flush_scheduled_work() because both
1416          * unregister_netdev() and linkwatch_event take the rtnl lock,
1417          * so flush_scheduled_work() can deadlock during device
1418          * removal.
1419          */
1420         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1421         if (!ipoib_workqueue) {
1422                 ret = -ENOMEM;
1423                 goto err_fs;
1424         }
1425
1426         ib_sa_register_client(&ipoib_sa_client);
1427
1428         ret = ib_register_client(&ipoib_client);
1429         if (ret)
1430                 goto err_sa;
1431
1432         return 0;
1433
1434 err_sa:
1435         ib_sa_unregister_client(&ipoib_sa_client);
1436         destroy_workqueue(ipoib_workqueue);
1437
1438 err_fs:
1439         ipoib_unregister_debugfs();
1440
1441         return ret;
1442 }
1443
1444 static void __exit ipoib_cleanup_module(void)
1445 {
1446         ib_unregister_client(&ipoib_client);
1447         ib_sa_unregister_client(&ipoib_sa_client);
1448         ipoib_unregister_debugfs();
1449         destroy_workqueue(ipoib_workqueue);
1450 }
1451
1452 module_init(ipoib_init_module);
1453 module_exit(ipoib_cleanup_module);