e9ca3cb57d5219960f9f67adf7eadb19dabc206e
[linux-2.6.git] / drivers / infiniband / ulp / ipoib / ipoib_main.c
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34
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         ipoib_ib_dev_down(dev, 0);
160         ipoib_ib_dev_stop(dev, 0);
161
162         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
163                 struct ipoib_dev_priv *cpriv;
164
165                 /* Bring down any child interfaces too */
166                 mutex_lock(&priv->vlan_mutex);
167                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
168                         int flags;
169
170                         flags = cpriv->dev->flags;
171                         if (!(flags & IFF_UP))
172                                 continue;
173
174                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
175                 }
176                 mutex_unlock(&priv->vlan_mutex);
177         }
178
179         return 0;
180 }
181
182 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
183 {
184         struct ipoib_dev_priv *priv = netdev_priv(dev);
185
186         /* dev->mtu > 2K ==> connected mode */
187         if (ipoib_cm_admin_enabled(dev)) {
188                 if (new_mtu > ipoib_cm_max_mtu(dev))
189                         return -EINVAL;
190
191                 if (new_mtu > priv->mcast_mtu)
192                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
193                                    priv->mcast_mtu);
194
195                 dev->mtu = new_mtu;
196                 return 0;
197         }
198
199         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
200                 return -EINVAL;
201
202         priv->admin_mtu = new_mtu;
203
204         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
205
206         return 0;
207 }
208
209 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
210 {
211         struct ipoib_dev_priv *priv = netdev_priv(dev);
212         struct rb_node *n = priv->path_tree.rb_node;
213         struct ipoib_path *path;
214         int ret;
215
216         while (n) {
217                 path = rb_entry(n, struct ipoib_path, rb_node);
218
219                 ret = memcmp(gid, path->pathrec.dgid.raw,
220                              sizeof (union ib_gid));
221
222                 if (ret < 0)
223                         n = n->rb_left;
224                 else if (ret > 0)
225                         n = n->rb_right;
226                 else
227                         return path;
228         }
229
230         return NULL;
231 }
232
233 static int __path_add(struct net_device *dev, struct ipoib_path *path)
234 {
235         struct ipoib_dev_priv *priv = netdev_priv(dev);
236         struct rb_node **n = &priv->path_tree.rb_node;
237         struct rb_node *pn = NULL;
238         struct ipoib_path *tpath;
239         int ret;
240
241         while (*n) {
242                 pn = *n;
243                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
244
245                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
246                              sizeof (union ib_gid));
247                 if (ret < 0)
248                         n = &pn->rb_left;
249                 else if (ret > 0)
250                         n = &pn->rb_right;
251                 else
252                         return -EEXIST;
253         }
254
255         rb_link_node(&path->rb_node, pn, n);
256         rb_insert_color(&path->rb_node, &priv->path_tree);
257
258         list_add_tail(&path->list, &priv->path_list);
259
260         return 0;
261 }
262
263 static void path_free(struct net_device *dev, struct ipoib_path *path)
264 {
265         struct ipoib_dev_priv *priv = netdev_priv(dev);
266         struct ipoib_neigh *neigh, *tn;
267         struct sk_buff *skb;
268         unsigned long flags;
269
270         while ((skb = __skb_dequeue(&path->queue)))
271                 dev_kfree_skb_irq(skb);
272
273         spin_lock_irqsave(&priv->lock, flags);
274
275         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
276                 /*
277                  * It's safe to call ipoib_put_ah() inside priv->lock
278                  * here, because we know that path->ah will always
279                  * hold one more reference, so ipoib_put_ah() will
280                  * never do more than decrement the ref count.
281                  */
282                 if (neigh->ah)
283                         ipoib_put_ah(neigh->ah);
284
285                 ipoib_neigh_free(dev, neigh);
286         }
287
288         spin_unlock_irqrestore(&priv->lock, flags);
289
290         if (path->ah)
291                 ipoib_put_ah(path->ah);
292
293         kfree(path);
294 }
295
296 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
297
298 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
299 {
300         struct ipoib_path_iter *iter;
301
302         iter = kmalloc(sizeof *iter, GFP_KERNEL);
303         if (!iter)
304                 return NULL;
305
306         iter->dev = dev;
307         memset(iter->path.pathrec.dgid.raw, 0, 16);
308
309         if (ipoib_path_iter_next(iter)) {
310                 kfree(iter);
311                 return NULL;
312         }
313
314         return iter;
315 }
316
317 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
318 {
319         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
320         struct rb_node *n;
321         struct ipoib_path *path;
322         int ret = 1;
323
324         spin_lock_irq(&priv->lock);
325
326         n = rb_first(&priv->path_tree);
327
328         while (n) {
329                 path = rb_entry(n, struct ipoib_path, rb_node);
330
331                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
332                            sizeof (union ib_gid)) < 0) {
333                         iter->path = *path;
334                         ret = 0;
335                         break;
336                 }
337
338                 n = rb_next(n);
339         }
340
341         spin_unlock_irq(&priv->lock);
342
343         return ret;
344 }
345
346 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
347                           struct ipoib_path *path)
348 {
349         *path = iter->path;
350 }
351
352 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
353
354 void ipoib_mark_paths_invalid(struct net_device *dev)
355 {
356         struct ipoib_dev_priv *priv = netdev_priv(dev);
357         struct ipoib_path *path, *tp;
358
359         spin_lock_irq(&priv->lock);
360
361         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
362                 ipoib_dbg(priv, "mark path LID 0x%04x GID " IPOIB_GID_FMT " invalid\n",
363                         be16_to_cpu(path->pathrec.dlid),
364                         IPOIB_GID_ARG(path->pathrec.dgid));
365                 path->valid =  0;
366         }
367
368         spin_unlock_irq(&priv->lock);
369 }
370
371 void ipoib_flush_paths(struct net_device *dev)
372 {
373         struct ipoib_dev_priv *priv = netdev_priv(dev);
374         struct ipoib_path *path, *tp;
375         LIST_HEAD(remove_list);
376
377         spin_lock_irq(&priv->tx_lock);
378         spin_lock(&priv->lock);
379
380         list_splice_init(&priv->path_list, &remove_list);
381
382         list_for_each_entry(path, &remove_list, list)
383                 rb_erase(&path->rb_node, &priv->path_tree);
384
385         list_for_each_entry_safe(path, tp, &remove_list, list) {
386                 if (path->query)
387                         ib_sa_cancel_query(path->query_id, path->query);
388                 spin_unlock(&priv->lock);
389                 spin_unlock_irq(&priv->tx_lock);
390                 wait_for_completion(&path->done);
391                 path_free(dev, path);
392                 spin_lock_irq(&priv->tx_lock);
393                 spin_lock(&priv->lock);
394         }
395         spin_unlock(&priv->lock);
396         spin_unlock_irq(&priv->tx_lock);
397 }
398
399 static void path_rec_completion(int status,
400                                 struct ib_sa_path_rec *pathrec,
401                                 void *path_ptr)
402 {
403         struct ipoib_path *path = path_ptr;
404         struct net_device *dev = path->dev;
405         struct ipoib_dev_priv *priv = netdev_priv(dev);
406         struct ipoib_ah *ah = NULL;
407         struct ipoib_ah *old_ah = NULL;
408         struct ipoib_neigh *neigh, *tn;
409         struct sk_buff_head skqueue;
410         struct sk_buff *skb;
411         unsigned long flags;
412
413         if (!status)
414                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
415                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
416         else
417                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
418                           status, IPOIB_GID_ARG(path->pathrec.dgid));
419
420         skb_queue_head_init(&skqueue);
421
422         if (!status) {
423                 struct ib_ah_attr av;
424
425                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
426                         ah = ipoib_create_ah(dev, priv->pd, &av);
427         }
428
429         spin_lock_irqsave(&priv->lock, flags);
430
431         if (ah) {
432                 path->pathrec = *pathrec;
433
434                 old_ah   = path->ah;
435                 path->ah = ah;
436
437                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
438                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
439
440                 while ((skb = __skb_dequeue(&path->queue)))
441                         __skb_queue_tail(&skqueue, skb);
442
443                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
444                         if (neigh->ah) {
445                                 WARN_ON(neigh->ah != old_ah);
446                                 /*
447                                  * Dropping the ah reference inside
448                                  * priv->lock is safe here, because we
449                                  * will hold one more reference from
450                                  * the original value of path->ah (ie
451                                  * old_ah).
452                                  */
453                                 ipoib_put_ah(neigh->ah);
454                         }
455                         kref_get(&path->ah->ref);
456                         neigh->ah = path->ah;
457                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
458                                sizeof(union ib_gid));
459
460                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
461                                 if (!ipoib_cm_get(neigh))
462                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
463                                                                                path,
464                                                                                neigh));
465                                 if (!ipoib_cm_get(neigh)) {
466                                         list_del(&neigh->list);
467                                         if (neigh->ah)
468                                                 ipoib_put_ah(neigh->ah);
469                                         ipoib_neigh_free(dev, neigh);
470                                         continue;
471                                 }
472                         }
473
474                         while ((skb = __skb_dequeue(&neigh->queue)))
475                                 __skb_queue_tail(&skqueue, skb);
476                 }
477                 path->valid = 1;
478         }
479
480         path->query = NULL;
481         complete(&path->done);
482
483         spin_unlock_irqrestore(&priv->lock, flags);
484
485         if (old_ah)
486                 ipoib_put_ah(old_ah);
487
488         while ((skb = __skb_dequeue(&skqueue))) {
489                 skb->dev = dev;
490                 if (dev_queue_xmit(skb))
491                         ipoib_warn(priv, "dev_queue_xmit failed "
492                                    "to requeue packet\n");
493         }
494 }
495
496 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
497 {
498         struct ipoib_dev_priv *priv = netdev_priv(dev);
499         struct ipoib_path *path;
500
501         if (!priv->broadcast)
502                 return NULL;
503
504         path = kzalloc(sizeof *path, GFP_ATOMIC);
505         if (!path)
506                 return NULL;
507
508         path->dev = dev;
509
510         skb_queue_head_init(&path->queue);
511
512         INIT_LIST_HEAD(&path->neigh_list);
513
514         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
515         path->pathrec.sgid          = priv->local_gid;
516         path->pathrec.pkey          = cpu_to_be16(priv->pkey);
517         path->pathrec.numb_path     = 1;
518         path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
519
520         return path;
521 }
522
523 static int path_rec_start(struct net_device *dev,
524                           struct ipoib_path *path)
525 {
526         struct ipoib_dev_priv *priv = netdev_priv(dev);
527
528         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
529                   IPOIB_GID_ARG(path->pathrec.dgid));
530
531         init_completion(&path->done);
532
533         path->query_id =
534                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
535                                    &path->pathrec,
536                                    IB_SA_PATH_REC_DGID          |
537                                    IB_SA_PATH_REC_SGID          |
538                                    IB_SA_PATH_REC_NUMB_PATH     |
539                                    IB_SA_PATH_REC_TRAFFIC_CLASS |
540                                    IB_SA_PATH_REC_PKEY,
541                                    1000, GFP_ATOMIC,
542                                    path_rec_completion,
543                                    path, &path->query);
544         if (path->query_id < 0) {
545                 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
546                 path->query = NULL;
547                 return path->query_id;
548         }
549
550         return 0;
551 }
552
553 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
554 {
555         struct ipoib_dev_priv *priv = netdev_priv(dev);
556         struct ipoib_path *path;
557         struct ipoib_neigh *neigh;
558
559         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
560         if (!neigh) {
561                 ++dev->stats.tx_dropped;
562                 dev_kfree_skb_any(skb);
563                 return;
564         }
565
566         /*
567          * We can only be called from ipoib_start_xmit, so we're
568          * inside tx_lock -- no need to save/restore flags.
569          */
570         spin_lock(&priv->lock);
571
572         path = __path_find(dev, skb->dst->neighbour->ha + 4);
573         if (!path) {
574                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
575                 if (!path)
576                         goto err_path;
577
578                 __path_add(dev, path);
579         }
580
581         list_add_tail(&neigh->list, &path->neigh_list);
582
583         if (path->ah) {
584                 kref_get(&path->ah->ref);
585                 neigh->ah = path->ah;
586                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
587                        sizeof(union ib_gid));
588
589                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
590                         if (!ipoib_cm_get(neigh))
591                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
592                         if (!ipoib_cm_get(neigh)) {
593                                 list_del(&neigh->list);
594                                 if (neigh->ah)
595                                         ipoib_put_ah(neigh->ah);
596                                 ipoib_neigh_free(dev, neigh);
597                                 goto err_drop;
598                         }
599                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
600                                 __skb_queue_tail(&neigh->queue, skb);
601                         else {
602                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
603                                            skb_queue_len(&neigh->queue));
604                                 goto err_drop;
605                         }
606                 } else
607                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
608         } else {
609                 neigh->ah  = NULL;
610
611                 if (!path->query && path_rec_start(dev, path))
612                         goto err_list;
613
614                 __skb_queue_tail(&neigh->queue, skb);
615         }
616
617         spin_unlock(&priv->lock);
618         return;
619
620 err_list:
621         list_del(&neigh->list);
622
623 err_path:
624         ipoib_neigh_free(dev, neigh);
625 err_drop:
626         ++dev->stats.tx_dropped;
627         dev_kfree_skb_any(skb);
628
629         spin_unlock(&priv->lock);
630 }
631
632 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
633 {
634         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
635
636         /* Look up path record for unicasts */
637         if (skb->dst->neighbour->ha[4] != 0xff) {
638                 neigh_add_path(skb, dev);
639                 return;
640         }
641
642         /* Add in the P_Key for multicasts */
643         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
644         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
645         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
646 }
647
648 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
649                              struct ipoib_pseudoheader *phdr)
650 {
651         struct ipoib_dev_priv *priv = netdev_priv(dev);
652         struct ipoib_path *path;
653
654         /*
655          * We can only be called from ipoib_start_xmit, so we're
656          * inside tx_lock -- no need to save/restore flags.
657          */
658         spin_lock(&priv->lock);
659
660         path = __path_find(dev, phdr->hwaddr + 4);
661         if (!path || !path->valid) {
662                 if (!path)
663                         path = path_rec_create(dev, phdr->hwaddr + 4);
664                 if (path) {
665                         /* put pseudoheader back on for next time */
666                         skb_push(skb, sizeof *phdr);
667                         __skb_queue_tail(&path->queue, skb);
668
669                         if (path_rec_start(dev, path)) {
670                                 spin_unlock(&priv->lock);
671                                 path_free(dev, path);
672                                 return;
673                         } else
674                                 __path_add(dev, path);
675                 } else {
676                         ++dev->stats.tx_dropped;
677                         dev_kfree_skb_any(skb);
678                 }
679
680                 spin_unlock(&priv->lock);
681                 return;
682         }
683
684         if (path->ah) {
685                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
686                           be16_to_cpu(path->pathrec.dlid));
687
688                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
689         } else if ((path->query || !path_rec_start(dev, path)) &&
690                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
691                 /* put pseudoheader back on for next time */
692                 skb_push(skb, sizeof *phdr);
693                 __skb_queue_tail(&path->queue, skb);
694         } else {
695                 ++dev->stats.tx_dropped;
696                 dev_kfree_skb_any(skb);
697         }
698
699         spin_unlock(&priv->lock);
700 }
701
702 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
703 {
704         struct ipoib_dev_priv *priv = netdev_priv(dev);
705         struct ipoib_neigh *neigh;
706         unsigned long flags;
707
708         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
709                 return NETDEV_TX_LOCKED;
710
711         if (likely(skb->dst && skb->dst->neighbour)) {
712                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
713                         ipoib_path_lookup(skb, dev);
714                         goto out;
715                 }
716
717                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
718
719                 if (neigh->ah)
720                         if (unlikely((memcmp(&neigh->dgid.raw,
721                                             skb->dst->neighbour->ha + 4,
722                                             sizeof(union ib_gid))) ||
723                                          (neigh->dev != dev))) {
724                                 spin_lock(&priv->lock);
725                                 /*
726                                  * It's safe to call ipoib_put_ah() inside
727                                  * priv->lock here, because we know that
728                                  * path->ah will always hold one more reference,
729                                  * so ipoib_put_ah() will never do more than
730                                  * decrement the ref count.
731                                  */
732                                 ipoib_put_ah(neigh->ah);
733                                 list_del(&neigh->list);
734                                 ipoib_neigh_free(dev, neigh);
735                                 spin_unlock(&priv->lock);
736                                 ipoib_path_lookup(skb, dev);
737                                 goto out;
738                         }
739
740                 if (ipoib_cm_get(neigh)) {
741                         if (ipoib_cm_up(neigh)) {
742                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
743                                 goto out;
744                         }
745                 } else if (neigh->ah) {
746                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
747                         goto out;
748                 }
749
750                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
751                         spin_lock(&priv->lock);
752                         __skb_queue_tail(&neigh->queue, skb);
753                         spin_unlock(&priv->lock);
754                 } else {
755                         ++dev->stats.tx_dropped;
756                         dev_kfree_skb_any(skb);
757                 }
758         } else {
759                 struct ipoib_pseudoheader *phdr =
760                         (struct ipoib_pseudoheader *) skb->data;
761                 skb_pull(skb, sizeof *phdr);
762
763                 if (phdr->hwaddr[4] == 0xff) {
764                         /* Add in the P_Key for multicast*/
765                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
766                         phdr->hwaddr[9] = priv->pkey & 0xff;
767
768                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
769                 } else {
770                         /* unicast GID -- should be ARP or RARP reply */
771
772                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
773                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
774                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
775                                            IPOIB_GID_FMT "\n",
776                                            skb->dst ? "neigh" : "dst",
777                                            be16_to_cpup((__be16 *) skb->data),
778                                            IPOIB_QPN(phdr->hwaddr),
779                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
780                                 dev_kfree_skb_any(skb);
781                                 ++dev->stats.tx_dropped;
782                                 goto out;
783                         }
784
785                         unicast_arp_send(skb, dev, phdr);
786                 }
787         }
788
789 out:
790         spin_unlock_irqrestore(&priv->tx_lock, flags);
791
792         return NETDEV_TX_OK;
793 }
794
795 static void ipoib_timeout(struct net_device *dev)
796 {
797         struct ipoib_dev_priv *priv = netdev_priv(dev);
798
799         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
800                    jiffies_to_msecs(jiffies - dev->trans_start));
801         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
802                    netif_queue_stopped(dev),
803                    priv->tx_head, priv->tx_tail);
804         /* XXX reset QP, etc. */
805 }
806
807 static int ipoib_hard_header(struct sk_buff *skb,
808                              struct net_device *dev,
809                              unsigned short type,
810                              const void *daddr, const void *saddr, unsigned len)
811 {
812         struct ipoib_header *header;
813
814         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
815
816         header->proto = htons(type);
817         header->reserved = 0;
818
819         /*
820          * If we don't have a neighbour structure, stuff the
821          * destination address onto the front of the skb so we can
822          * figure out where to send the packet later.
823          */
824         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
825                 struct ipoib_pseudoheader *phdr =
826                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
827                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
828         }
829
830         return 0;
831 }
832
833 static void ipoib_set_mcast_list(struct net_device *dev)
834 {
835         struct ipoib_dev_priv *priv = netdev_priv(dev);
836
837         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
838                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
839                 return;
840         }
841
842         queue_work(ipoib_workqueue, &priv->restart_task);
843 }
844
845 static void ipoib_neigh_cleanup(struct neighbour *n)
846 {
847         struct ipoib_neigh *neigh;
848         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
849         unsigned long flags;
850         struct ipoib_ah *ah = NULL;
851
852         neigh = *to_ipoib_neigh(n);
853         if (neigh)
854                 priv = netdev_priv(neigh->dev);
855         else
856                 return;
857         ipoib_dbg(priv,
858                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
859                   IPOIB_QPN(n->ha),
860                   IPOIB_GID_RAW_ARG(n->ha + 4));
861
862         spin_lock_irqsave(&priv->lock, flags);
863
864         if (neigh->ah)
865                 ah = neigh->ah;
866         list_del(&neigh->list);
867         ipoib_neigh_free(n->dev, neigh);
868
869         spin_unlock_irqrestore(&priv->lock, flags);
870
871         if (ah)
872                 ipoib_put_ah(ah);
873 }
874
875 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
876                                       struct net_device *dev)
877 {
878         struct ipoib_neigh *neigh;
879
880         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
881         if (!neigh)
882                 return NULL;
883
884         neigh->neighbour = neighbour;
885         neigh->dev = dev;
886         *to_ipoib_neigh(neighbour) = neigh;
887         skb_queue_head_init(&neigh->queue);
888         ipoib_cm_set(neigh, NULL);
889
890         return neigh;
891 }
892
893 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
894 {
895         struct sk_buff *skb;
896         *to_ipoib_neigh(neigh->neighbour) = NULL;
897         while ((skb = __skb_dequeue(&neigh->queue))) {
898                 ++dev->stats.tx_dropped;
899                 dev_kfree_skb_any(skb);
900         }
901         if (ipoib_cm_get(neigh))
902                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
903         kfree(neigh);
904 }
905
906 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
907 {
908         parms->neigh_cleanup = ipoib_neigh_cleanup;
909
910         return 0;
911 }
912
913 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
914 {
915         struct ipoib_dev_priv *priv = netdev_priv(dev);
916
917         /* Allocate RX/TX "rings" to hold queued skbs */
918         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
919                                 GFP_KERNEL);
920         if (!priv->rx_ring) {
921                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
922                        ca->name, ipoib_recvq_size);
923                 goto out;
924         }
925
926         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
927         if (!priv->tx_ring) {
928                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
929                        ca->name, ipoib_sendq_size);
930                 goto out_rx_ring_cleanup;
931         }
932         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
933
934         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
935
936         if (ipoib_ib_dev_init(dev, ca, port))
937                 goto out_tx_ring_cleanup;
938
939         return 0;
940
941 out_tx_ring_cleanup:
942         vfree(priv->tx_ring);
943
944 out_rx_ring_cleanup:
945         kfree(priv->rx_ring);
946
947 out:
948         return -ENOMEM;
949 }
950
951 void ipoib_dev_cleanup(struct net_device *dev)
952 {
953         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
954
955         ipoib_delete_debug_files(dev);
956
957         /* Delete any child interfaces first */
958         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
959                 unregister_netdev(cpriv->dev);
960                 ipoib_dev_cleanup(cpriv->dev);
961                 free_netdev(cpriv->dev);
962         }
963
964         ipoib_ib_dev_cleanup(dev);
965
966         kfree(priv->rx_ring);
967         vfree(priv->tx_ring);
968
969         priv->rx_ring = NULL;
970         priv->tx_ring = NULL;
971 }
972
973 static const struct header_ops ipoib_header_ops = {
974         .create = ipoib_hard_header,
975 };
976
977 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
978                        void **tcph, u64 *hdr_flags, void *priv)
979 {
980         unsigned int ip_len;
981         struct iphdr *iph;
982
983         if (unlikely(skb->protocol != htons(ETH_P_IP)))
984                 return -1;
985
986         /*
987          * In the future we may add an else clause that verifies the
988          * checksum and allows devices which do not calculate checksum
989          * to use LRO.
990          */
991         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
992                 return -1;
993
994         /* Check for non-TCP packet */
995         skb_reset_network_header(skb);
996         iph = ip_hdr(skb);
997         if (iph->protocol != IPPROTO_TCP)
998                 return -1;
999
1000         ip_len = ip_hdrlen(skb);
1001         skb_set_transport_header(skb, ip_len);
1002         *tcph = tcp_hdr(skb);
1003
1004         /* check if IP header and TCP header are complete */
1005         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1006                 return -1;
1007
1008         *hdr_flags = LRO_IPV4 | LRO_TCP;
1009         *iphdr = iph;
1010
1011         return 0;
1012 }
1013
1014 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1015 {
1016         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1017         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1018         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1019         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1020         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1021         priv->lro.lro_mgr.dev            = priv->dev;
1022         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1023 }
1024
1025 static void ipoib_setup(struct net_device *dev)
1026 {
1027         struct ipoib_dev_priv *priv = netdev_priv(dev);
1028
1029         dev->open                = ipoib_open;
1030         dev->stop                = ipoib_stop;
1031         dev->change_mtu          = ipoib_change_mtu;
1032         dev->hard_start_xmit     = ipoib_start_xmit;
1033         dev->tx_timeout          = ipoib_timeout;
1034         dev->header_ops          = &ipoib_header_ops;
1035         dev->set_multicast_list  = ipoib_set_mcast_list;
1036         dev->neigh_setup         = ipoib_neigh_setup_dev;
1037
1038         ipoib_set_ethtool_ops(dev);
1039
1040         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1041
1042         dev->watchdog_timeo      = HZ;
1043
1044         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1045
1046         /*
1047          * We add in INFINIBAND_ALEN to allow for the destination
1048          * address "pseudoheader" for skbs without neighbour struct.
1049          */
1050         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1051         dev->addr_len            = INFINIBAND_ALEN;
1052         dev->type                = ARPHRD_INFINIBAND;
1053         dev->tx_queue_len        = ipoib_sendq_size * 2;
1054         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1055                                     NETIF_F_LLTX                |
1056                                     NETIF_F_HIGHDMA);
1057
1058         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1059
1060         netif_carrier_off(dev);
1061
1062         priv->dev = dev;
1063
1064         ipoib_lro_setup(priv);
1065
1066         spin_lock_init(&priv->lock);
1067         spin_lock_init(&priv->tx_lock);
1068
1069         mutex_init(&priv->vlan_mutex);
1070
1071         INIT_LIST_HEAD(&priv->path_list);
1072         INIT_LIST_HEAD(&priv->child_intfs);
1073         INIT_LIST_HEAD(&priv->dead_ahs);
1074         INIT_LIST_HEAD(&priv->multicast_list);
1075
1076         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1077         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1078         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1079         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1080         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1081         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1082         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1083         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1084 }
1085
1086 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1087 {
1088         struct net_device *dev;
1089
1090         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1091                            ipoib_setup);
1092         if (!dev)
1093                 return NULL;
1094
1095         return netdev_priv(dev);
1096 }
1097
1098 static ssize_t show_pkey(struct device *dev,
1099                          struct device_attribute *attr, char *buf)
1100 {
1101         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1102
1103         return sprintf(buf, "0x%04x\n", priv->pkey);
1104 }
1105 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1106
1107 static ssize_t show_umcast(struct device *dev,
1108                            struct device_attribute *attr, char *buf)
1109 {
1110         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1111
1112         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1113 }
1114
1115 static ssize_t set_umcast(struct device *dev,
1116                           struct device_attribute *attr,
1117                           const char *buf, size_t count)
1118 {
1119         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1120         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1121
1122         if (umcast_val > 0) {
1123                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1124                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1125                                 "by userspace\n");
1126         } else
1127                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1128
1129         return count;
1130 }
1131 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1132
1133 int ipoib_add_umcast_attr(struct net_device *dev)
1134 {
1135         return device_create_file(&dev->dev, &dev_attr_umcast);
1136 }
1137
1138 static ssize_t create_child(struct device *dev,
1139                             struct device_attribute *attr,
1140                             const char *buf, size_t count)
1141 {
1142         int pkey;
1143         int ret;
1144
1145         if (sscanf(buf, "%i", &pkey) != 1)
1146                 return -EINVAL;
1147
1148         if (pkey < 0 || pkey > 0xffff)
1149                 return -EINVAL;
1150
1151         /*
1152          * Set the full membership bit, so that we join the right
1153          * broadcast group, etc.
1154          */
1155         pkey |= 0x8000;
1156
1157         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1158
1159         return ret ? ret : count;
1160 }
1161 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1162
1163 static ssize_t delete_child(struct device *dev,
1164                             struct device_attribute *attr,
1165                             const char *buf, size_t count)
1166 {
1167         int pkey;
1168         int ret;
1169
1170         if (sscanf(buf, "%i", &pkey) != 1)
1171                 return -EINVAL;
1172
1173         if (pkey < 0 || pkey > 0xffff)
1174                 return -EINVAL;
1175
1176         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1177
1178         return ret ? ret : count;
1179
1180 }
1181 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1182
1183 int ipoib_add_pkey_attr(struct net_device *dev)
1184 {
1185         return device_create_file(&dev->dev, &dev_attr_pkey);
1186 }
1187
1188 static struct net_device *ipoib_add_port(const char *format,
1189                                          struct ib_device *hca, u8 port)
1190 {
1191         struct ipoib_dev_priv *priv;
1192         struct ib_device_attr *device_attr;
1193         struct ib_port_attr attr;
1194         int result = -ENOMEM;
1195
1196         priv = ipoib_intf_alloc(format);
1197         if (!priv)
1198                 goto alloc_mem_failed;
1199
1200         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1201
1202         if (!ib_query_port(hca, port, &attr))
1203                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1204         else {
1205                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1206                        hca->name, port);
1207                 goto device_init_failed;
1208         }
1209
1210         /* MTU will be reset when mcast join happens */
1211         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1212         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1213
1214         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1215         if (result) {
1216                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1217                        hca->name, port, result);
1218                 goto device_init_failed;
1219         }
1220
1221         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1222         if (!device_attr) {
1223                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1224                        hca->name, sizeof *device_attr);
1225                 goto device_init_failed;
1226         }
1227
1228         result = ib_query_device(hca, device_attr);
1229         if (result) {
1230                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1231                        hca->name, result);
1232                 kfree(device_attr);
1233                 goto device_init_failed;
1234         }
1235         priv->hca_caps = device_attr->device_cap_flags;
1236
1237         kfree(device_attr);
1238
1239         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1240                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1241                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1242         }
1243
1244         if (lro)
1245                 priv->dev->features |= NETIF_F_LRO;
1246
1247         /*
1248          * Set the full membership bit, so that we join the right
1249          * broadcast group, etc.
1250          */
1251         priv->pkey |= 0x8000;
1252
1253         priv->dev->broadcast[8] = priv->pkey >> 8;
1254         priv->dev->broadcast[9] = priv->pkey & 0xff;
1255
1256         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1257         if (result) {
1258                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1259                        hca->name, port, result);
1260                 goto device_init_failed;
1261         } else
1262                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1263
1264         result = ipoib_dev_init(priv->dev, hca, port);
1265         if (result < 0) {
1266                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1267                        hca->name, port, result);
1268                 goto device_init_failed;
1269         }
1270
1271         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1272                               priv->ca, ipoib_event);
1273         result = ib_register_event_handler(&priv->event_handler);
1274         if (result < 0) {
1275                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1276                        "port %d (ret = %d)\n",
1277                        hca->name, port, result);
1278                 goto event_failed;
1279         }
1280
1281         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1282                 priv->dev->features |= NETIF_F_TSO;
1283
1284         result = register_netdev(priv->dev);
1285         if (result) {
1286                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1287                        hca->name, port, result);
1288                 goto register_failed;
1289         }
1290
1291         ipoib_create_debug_files(priv->dev);
1292
1293         if (ipoib_cm_add_mode_attr(priv->dev))
1294                 goto sysfs_failed;
1295         if (ipoib_add_pkey_attr(priv->dev))
1296                 goto sysfs_failed;
1297         if (ipoib_add_umcast_attr(priv->dev))
1298                 goto sysfs_failed;
1299         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1300                 goto sysfs_failed;
1301         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1302                 goto sysfs_failed;
1303
1304         return priv->dev;
1305
1306 sysfs_failed:
1307         ipoib_delete_debug_files(priv->dev);
1308         unregister_netdev(priv->dev);
1309
1310 register_failed:
1311         ib_unregister_event_handler(&priv->event_handler);
1312         flush_workqueue(ipoib_workqueue);
1313
1314 event_failed:
1315         ipoib_dev_cleanup(priv->dev);
1316
1317 device_init_failed:
1318         free_netdev(priv->dev);
1319
1320 alloc_mem_failed:
1321         return ERR_PTR(result);
1322 }
1323
1324 static void ipoib_add_one(struct ib_device *device)
1325 {
1326         struct list_head *dev_list;
1327         struct net_device *dev;
1328         struct ipoib_dev_priv *priv;
1329         int s, e, p;
1330
1331         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1332                 return;
1333
1334         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1335         if (!dev_list)
1336                 return;
1337
1338         INIT_LIST_HEAD(dev_list);
1339
1340         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1341                 s = 0;
1342                 e = 0;
1343         } else {
1344                 s = 1;
1345                 e = device->phys_port_cnt;
1346         }
1347
1348         for (p = s; p <= e; ++p) {
1349                 dev = ipoib_add_port("ib%d", device, p);
1350                 if (!IS_ERR(dev)) {
1351                         priv = netdev_priv(dev);
1352                         list_add_tail(&priv->list, dev_list);
1353                 }
1354         }
1355
1356         ib_set_client_data(device, &ipoib_client, dev_list);
1357 }
1358
1359 static void ipoib_remove_one(struct ib_device *device)
1360 {
1361         struct ipoib_dev_priv *priv, *tmp;
1362         struct list_head *dev_list;
1363
1364         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1365                 return;
1366
1367         dev_list = ib_get_client_data(device, &ipoib_client);
1368
1369         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1370                 ib_unregister_event_handler(&priv->event_handler);
1371
1372                 rtnl_lock();
1373                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1374                 rtnl_unlock();
1375
1376                 flush_workqueue(ipoib_workqueue);
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);