IPoIB: In unicast_arp_send(), only free newly-created paths
[linux-3.10.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         if (!test_and_set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
110                 napi_enable(&priv->napi);
111
112         if (ipoib_pkey_dev_delay_open(dev))
113                 return 0;
114
115         if (ipoib_ib_dev_open(dev))
116                 goto err_disable;
117
118         if (ipoib_ib_dev_up(dev))
119                 goto err_stop;
120
121         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
122                 struct ipoib_dev_priv *cpriv;
123
124                 /* Bring up any child interfaces too */
125                 mutex_lock(&priv->vlan_mutex);
126                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
127                         int flags;
128
129                         flags = cpriv->dev->flags;
130                         if (flags & IFF_UP)
131                                 continue;
132
133                         dev_change_flags(cpriv->dev, flags | IFF_UP);
134                 }
135                 mutex_unlock(&priv->vlan_mutex);
136         }
137
138         netif_start_queue(dev);
139
140         return 0;
141
142 err_stop:
143         ipoib_ib_dev_stop(dev, 1);
144
145 err_disable:
146         napi_disable(&priv->napi);
147         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
148
149         return -EINVAL;
150 }
151
152 static int ipoib_stop(struct net_device *dev)
153 {
154         struct ipoib_dev_priv *priv = netdev_priv(dev);
155
156         ipoib_dbg(priv, "stopping interface\n");
157
158         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
159         napi_disable(&priv->napi);
160
161         netif_stop_queue(dev);
162
163         ipoib_ib_dev_down(dev, 0);
164         ipoib_ib_dev_stop(dev, 0);
165
166         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
167                 struct ipoib_dev_priv *cpriv;
168
169                 /* Bring down any child interfaces too */
170                 mutex_lock(&priv->vlan_mutex);
171                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
172                         int flags;
173
174                         flags = cpriv->dev->flags;
175                         if (!(flags & IFF_UP))
176                                 continue;
177
178                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
179                 }
180                 mutex_unlock(&priv->vlan_mutex);
181         }
182
183         return 0;
184 }
185
186 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
187 {
188         struct ipoib_dev_priv *priv = netdev_priv(dev);
189
190         /* dev->mtu > 2K ==> connected mode */
191         if (ipoib_cm_admin_enabled(dev)) {
192                 if (new_mtu > ipoib_cm_max_mtu(dev))
193                         return -EINVAL;
194
195                 if (new_mtu > priv->mcast_mtu)
196                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
197                                    priv->mcast_mtu);
198
199                 dev->mtu = new_mtu;
200                 return 0;
201         }
202
203         if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
204                 return -EINVAL;
205
206         priv->admin_mtu = new_mtu;
207
208         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
209
210         return 0;
211 }
212
213 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
214 {
215         struct ipoib_dev_priv *priv = netdev_priv(dev);
216         struct rb_node *n = priv->path_tree.rb_node;
217         struct ipoib_path *path;
218         int ret;
219
220         while (n) {
221                 path = rb_entry(n, struct ipoib_path, rb_node);
222
223                 ret = memcmp(gid, path->pathrec.dgid.raw,
224                              sizeof (union ib_gid));
225
226                 if (ret < 0)
227                         n = n->rb_left;
228                 else if (ret > 0)
229                         n = n->rb_right;
230                 else
231                         return path;
232         }
233
234         return NULL;
235 }
236
237 static int __path_add(struct net_device *dev, struct ipoib_path *path)
238 {
239         struct ipoib_dev_priv *priv = netdev_priv(dev);
240         struct rb_node **n = &priv->path_tree.rb_node;
241         struct rb_node *pn = NULL;
242         struct ipoib_path *tpath;
243         int ret;
244
245         while (*n) {
246                 pn = *n;
247                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
248
249                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
250                              sizeof (union ib_gid));
251                 if (ret < 0)
252                         n = &pn->rb_left;
253                 else if (ret > 0)
254                         n = &pn->rb_right;
255                 else
256                         return -EEXIST;
257         }
258
259         rb_link_node(&path->rb_node, pn, n);
260         rb_insert_color(&path->rb_node, &priv->path_tree);
261
262         list_add_tail(&path->list, &priv->path_list);
263
264         return 0;
265 }
266
267 static void path_free(struct net_device *dev, struct ipoib_path *path)
268 {
269         struct ipoib_dev_priv *priv = netdev_priv(dev);
270         struct ipoib_neigh *neigh, *tn;
271         struct sk_buff *skb;
272         unsigned long flags;
273
274         while ((skb = __skb_dequeue(&path->queue)))
275                 dev_kfree_skb_irq(skb);
276
277         spin_lock_irqsave(&priv->lock, flags);
278
279         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
280                 /*
281                  * It's safe to call ipoib_put_ah() inside priv->lock
282                  * here, because we know that path->ah will always
283                  * hold one more reference, so ipoib_put_ah() will
284                  * never do more than decrement the ref count.
285                  */
286                 if (neigh->ah)
287                         ipoib_put_ah(neigh->ah);
288
289                 ipoib_neigh_free(dev, neigh);
290         }
291
292         spin_unlock_irqrestore(&priv->lock, flags);
293
294         if (path->ah)
295                 ipoib_put_ah(path->ah);
296
297         kfree(path);
298 }
299
300 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
301
302 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
303 {
304         struct ipoib_path_iter *iter;
305
306         iter = kmalloc(sizeof *iter, GFP_KERNEL);
307         if (!iter)
308                 return NULL;
309
310         iter->dev = dev;
311         memset(iter->path.pathrec.dgid.raw, 0, 16);
312
313         if (ipoib_path_iter_next(iter)) {
314                 kfree(iter);
315                 return NULL;
316         }
317
318         return iter;
319 }
320
321 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
322 {
323         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
324         struct rb_node *n;
325         struct ipoib_path *path;
326         int ret = 1;
327
328         spin_lock_irq(&priv->lock);
329
330         n = rb_first(&priv->path_tree);
331
332         while (n) {
333                 path = rb_entry(n, struct ipoib_path, rb_node);
334
335                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
336                            sizeof (union ib_gid)) < 0) {
337                         iter->path = *path;
338                         ret = 0;
339                         break;
340                 }
341
342                 n = rb_next(n);
343         }
344
345         spin_unlock_irq(&priv->lock);
346
347         return ret;
348 }
349
350 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
351                           struct ipoib_path *path)
352 {
353         *path = iter->path;
354 }
355
356 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
357
358 void ipoib_mark_paths_invalid(struct net_device *dev)
359 {
360         struct ipoib_dev_priv *priv = netdev_priv(dev);
361         struct ipoib_path *path, *tp;
362
363         spin_lock_irq(&priv->lock);
364
365         list_for_each_entry_safe(path, tp, &priv->path_list, list) {
366                 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
367                         be16_to_cpu(path->pathrec.dlid),
368                         path->pathrec.dgid.raw);
369                 path->valid =  0;
370         }
371
372         spin_unlock_irq(&priv->lock);
373 }
374
375 void ipoib_flush_paths(struct net_device *dev)
376 {
377         struct ipoib_dev_priv *priv = netdev_priv(dev);
378         struct ipoib_path *path, *tp;
379         LIST_HEAD(remove_list);
380         unsigned long flags;
381
382         netif_tx_lock_bh(dev);
383         spin_lock_irqsave(&priv->lock, flags);
384
385         list_splice_init(&priv->path_list, &remove_list);
386
387         list_for_each_entry(path, &remove_list, list)
388                 rb_erase(&path->rb_node, &priv->path_tree);
389
390         list_for_each_entry_safe(path, tp, &remove_list, list) {
391                 if (path->query)
392                         ib_sa_cancel_query(path->query_id, path->query);
393                 spin_unlock_irqrestore(&priv->lock, flags);
394                 netif_tx_unlock_bh(dev);
395                 wait_for_completion(&path->done);
396                 path_free(dev, path);
397                 netif_tx_lock_bh(dev);
398                 spin_lock_irqsave(&priv->lock, flags);
399         }
400
401         spin_unlock_irqrestore(&priv->lock, flags);
402         netif_tx_unlock_bh(dev);
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 = NULL;
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 %pI6\n",
421                           be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
422         else
423                 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
424                           status, path->pathrec.dgid.raw);
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         if (ah) {
438                 path->pathrec = *pathrec;
439
440                 old_ah   = path->ah;
441                 path->ah = ah;
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 %pI6\n",
535                   path->pathrec.dgid.raw);
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: %d\n", path->query_id);
552                 path->query = NULL;
553                 complete(&path->done);
554                 return path->query_id;
555         }
556
557         return 0;
558 }
559
560 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
561 {
562         struct ipoib_dev_priv *priv = netdev_priv(dev);
563         struct ipoib_path *path;
564         struct ipoib_neigh *neigh;
565         unsigned long flags;
566
567         neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
568         if (!neigh) {
569                 ++dev->stats.tx_dropped;
570                 dev_kfree_skb_any(skb);
571                 return;
572         }
573
574         spin_lock_irqsave(&priv->lock, flags);
575
576         path = __path_find(dev, skb->dst->neighbour->ha + 4);
577         if (!path) {
578                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
579                 if (!path)
580                         goto err_path;
581
582                 __path_add(dev, path);
583         }
584
585         list_add_tail(&neigh->list, &path->neigh_list);
586
587         if (path->ah) {
588                 kref_get(&path->ah->ref);
589                 neigh->ah = path->ah;
590                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
591                        sizeof(union ib_gid));
592
593                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
594                         if (!ipoib_cm_get(neigh))
595                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
596                         if (!ipoib_cm_get(neigh)) {
597                                 list_del(&neigh->list);
598                                 if (neigh->ah)
599                                         ipoib_put_ah(neigh->ah);
600                                 ipoib_neigh_free(dev, neigh);
601                                 goto err_drop;
602                         }
603                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
604                                 __skb_queue_tail(&neigh->queue, skb);
605                         else {
606                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
607                                            skb_queue_len(&neigh->queue));
608                                 goto err_drop;
609                         }
610                 } else
611                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
612         } else {
613                 neigh->ah  = NULL;
614
615                 if (!path->query && path_rec_start(dev, path))
616                         goto err_list;
617
618                 __skb_queue_tail(&neigh->queue, skb);
619         }
620
621         spin_unlock_irqrestore(&priv->lock, flags);
622         return;
623
624 err_list:
625         list_del(&neigh->list);
626
627 err_path:
628         ipoib_neigh_free(dev, neigh);
629 err_drop:
630         ++dev->stats.tx_dropped;
631         dev_kfree_skb_any(skb);
632
633         spin_unlock_irqrestore(&priv->lock, flags);
634 }
635
636 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
637 {
638         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
639
640         /* Look up path record for unicasts */
641         if (skb->dst->neighbour->ha[4] != 0xff) {
642                 neigh_add_path(skb, dev);
643                 return;
644         }
645
646         /* Add in the P_Key for multicasts */
647         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
648         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
649         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
650 }
651
652 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
653                              struct ipoib_pseudoheader *phdr)
654 {
655         struct ipoib_dev_priv *priv = netdev_priv(dev);
656         struct ipoib_path *path;
657         unsigned long flags;
658
659         spin_lock_irqsave(&priv->lock, flags);
660
661         path = __path_find(dev, phdr->hwaddr + 4);
662         if (!path || !path->valid) {
663                 int new_path = 0;
664
665                 if (!path) {
666                         path = path_rec_create(dev, phdr->hwaddr + 4);
667                         new_path = 1;
668                 }
669                 if (path) {
670                         /* put pseudoheader back on for next time */
671                         skb_push(skb, sizeof *phdr);
672                         __skb_queue_tail(&path->queue, skb);
673
674                         if (!path->query && path_rec_start(dev, path)) {
675                                 spin_unlock_irqrestore(&priv->lock, flags);
676                                 if (new_path)
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_irqrestore(&priv->lock, flags);
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_irqrestore(&priv->lock, flags);
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 (likely(skb->dst && skb->dst->neighbour)) {
715                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
716                         ipoib_path_lookup(skb, dev);
717                         return NETDEV_TX_OK;
718                 }
719
720                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
721
722                 if (unlikely((memcmp(&neigh->dgid.raw,
723                                      skb->dst->neighbour->ha + 4,
724                                      sizeof(union ib_gid))) ||
725                              (neigh->dev != dev))) {
726                         spin_lock_irqsave(&priv->lock, flags);
727                         /*
728                          * It's safe to call ipoib_put_ah() inside
729                          * priv->lock here, because we know that
730                          * path->ah will always hold one more reference,
731                          * so ipoib_put_ah() will never do more than
732                          * decrement the ref count.
733                          */
734                         if (neigh->ah)
735                                 ipoib_put_ah(neigh->ah);
736                         list_del(&neigh->list);
737                         ipoib_neigh_free(dev, neigh);
738                         spin_unlock_irqrestore(&priv->lock, flags);
739                         ipoib_path_lookup(skb, dev);
740                         return NETDEV_TX_OK;
741                 }
742
743                 if (ipoib_cm_get(neigh)) {
744                         if (ipoib_cm_up(neigh)) {
745                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
746                                 return NETDEV_TX_OK;
747                         }
748                 } else if (neigh->ah) {
749                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
750                         return NETDEV_TX_OK;
751                 }
752
753                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
754                         spin_lock_irqsave(&priv->lock, flags);
755                         __skb_queue_tail(&neigh->queue, skb);
756                         spin_unlock_irqrestore(&priv->lock, flags);
757                 } else {
758                         ++dev->stats.tx_dropped;
759                         dev_kfree_skb_any(skb);
760                 }
761         } else {
762                 struct ipoib_pseudoheader *phdr =
763                         (struct ipoib_pseudoheader *) skb->data;
764                 skb_pull(skb, sizeof *phdr);
765
766                 if (phdr->hwaddr[4] == 0xff) {
767                         /* Add in the P_Key for multicast*/
768                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
769                         phdr->hwaddr[9] = priv->pkey & 0xff;
770
771                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
772                 } else {
773                         /* unicast GID -- should be ARP or RARP reply */
774
775                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
776                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
777                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
778                                            skb->dst ? "neigh" : "dst",
779                                            be16_to_cpup((__be16 *) skb->data),
780                                            IPOIB_QPN(phdr->hwaddr),
781                                            phdr->hwaddr + 4);
782                                 dev_kfree_skb_any(skb);
783                                 ++dev->stats.tx_dropped;
784                                 return NETDEV_TX_OK;
785                         }
786
787                         unicast_arp_send(skb, dev, phdr);
788                 }
789         }
790
791         return NETDEV_TX_OK;
792 }
793
794 static void ipoib_timeout(struct net_device *dev)
795 {
796         struct ipoib_dev_priv *priv = netdev_priv(dev);
797
798         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
799                    jiffies_to_msecs(jiffies - dev->trans_start));
800         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
801                    netif_queue_stopped(dev),
802                    priv->tx_head, priv->tx_tail);
803         /* XXX reset QP, etc. */
804 }
805
806 static int ipoib_hard_header(struct sk_buff *skb,
807                              struct net_device *dev,
808                              unsigned short type,
809                              const void *daddr, const void *saddr, unsigned len)
810 {
811         struct ipoib_header *header;
812
813         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
814
815         header->proto = htons(type);
816         header->reserved = 0;
817
818         /*
819          * If we don't have a neighbour structure, stuff the
820          * destination address onto the front of the skb so we can
821          * figure out where to send the packet later.
822          */
823         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
824                 struct ipoib_pseudoheader *phdr =
825                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
826                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
827         }
828
829         return 0;
830 }
831
832 static void ipoib_set_mcast_list(struct net_device *dev)
833 {
834         struct ipoib_dev_priv *priv = netdev_priv(dev);
835
836         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
837                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
838                 return;
839         }
840
841         queue_work(ipoib_workqueue, &priv->restart_task);
842 }
843
844 static void ipoib_neigh_cleanup(struct neighbour *n)
845 {
846         struct ipoib_neigh *neigh;
847         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
848         unsigned long flags;
849         struct ipoib_ah *ah = NULL;
850
851         neigh = *to_ipoib_neigh(n);
852         if (neigh)
853                 priv = netdev_priv(neigh->dev);
854         else
855                 return;
856         ipoib_dbg(priv,
857                   "neigh_cleanup for %06x %pI6\n",
858                   IPOIB_QPN(n->ha),
859                   n->ha + 4);
860
861         spin_lock_irqsave(&priv->lock, flags);
862
863         if (neigh->ah)
864                 ah = neigh->ah;
865         list_del(&neigh->list);
866         ipoib_neigh_free(n->dev, neigh);
867
868         spin_unlock_irqrestore(&priv->lock, flags);
869
870         if (ah)
871                 ipoib_put_ah(ah);
872 }
873
874 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
875                                       struct net_device *dev)
876 {
877         struct ipoib_neigh *neigh;
878
879         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
880         if (!neigh)
881                 return NULL;
882
883         neigh->neighbour = neighbour;
884         neigh->dev = dev;
885         *to_ipoib_neigh(neighbour) = neigh;
886         skb_queue_head_init(&neigh->queue);
887         ipoib_cm_set(neigh, NULL);
888
889         return neigh;
890 }
891
892 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
893 {
894         struct sk_buff *skb;
895         *to_ipoib_neigh(neigh->neighbour) = NULL;
896         while ((skb = __skb_dequeue(&neigh->queue))) {
897                 ++dev->stats.tx_dropped;
898                 dev_kfree_skb_any(skb);
899         }
900         if (ipoib_cm_get(neigh))
901                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
902         kfree(neigh);
903 }
904
905 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
906 {
907         parms->neigh_cleanup = ipoib_neigh_cleanup;
908
909         return 0;
910 }
911
912 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
913 {
914         struct ipoib_dev_priv *priv = netdev_priv(dev);
915
916         /* Allocate RX/TX "rings" to hold queued skbs */
917         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
918                                 GFP_KERNEL);
919         if (!priv->rx_ring) {
920                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
921                        ca->name, ipoib_recvq_size);
922                 goto out;
923         }
924
925         priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
926         if (!priv->tx_ring) {
927                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
928                        ca->name, ipoib_sendq_size);
929                 goto out_rx_ring_cleanup;
930         }
931         memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
932
933         /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
934
935         if (ipoib_ib_dev_init(dev, ca, port))
936                 goto out_tx_ring_cleanup;
937
938         return 0;
939
940 out_tx_ring_cleanup:
941         vfree(priv->tx_ring);
942
943 out_rx_ring_cleanup:
944         kfree(priv->rx_ring);
945
946 out:
947         return -ENOMEM;
948 }
949
950 void ipoib_dev_cleanup(struct net_device *dev)
951 {
952         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
953
954         ipoib_delete_debug_files(dev);
955
956         /* Delete any child interfaces first */
957         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
958                 unregister_netdev(cpriv->dev);
959                 ipoib_dev_cleanup(cpriv->dev);
960                 free_netdev(cpriv->dev);
961         }
962
963         ipoib_ib_dev_cleanup(dev);
964
965         kfree(priv->rx_ring);
966         vfree(priv->tx_ring);
967
968         priv->rx_ring = NULL;
969         priv->tx_ring = NULL;
970 }
971
972 static const struct header_ops ipoib_header_ops = {
973         .create = ipoib_hard_header,
974 };
975
976 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
977                        void **tcph, u64 *hdr_flags, void *priv)
978 {
979         unsigned int ip_len;
980         struct iphdr *iph;
981
982         if (unlikely(skb->protocol != htons(ETH_P_IP)))
983                 return -1;
984
985         /*
986          * In the future we may add an else clause that verifies the
987          * checksum and allows devices which do not calculate checksum
988          * to use LRO.
989          */
990         if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
991                 return -1;
992
993         /* Check for non-TCP packet */
994         skb_reset_network_header(skb);
995         iph = ip_hdr(skb);
996         if (iph->protocol != IPPROTO_TCP)
997                 return -1;
998
999         ip_len = ip_hdrlen(skb);
1000         skb_set_transport_header(skb, ip_len);
1001         *tcph = tcp_hdr(skb);
1002
1003         /* check if IP header and TCP header are complete */
1004         if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1005                 return -1;
1006
1007         *hdr_flags = LRO_IPV4 | LRO_TCP;
1008         *iphdr = iph;
1009
1010         return 0;
1011 }
1012
1013 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1014 {
1015         priv->lro.lro_mgr.max_aggr       = lro_max_aggr;
1016         priv->lro.lro_mgr.max_desc       = IPOIB_MAX_LRO_DESCRIPTORS;
1017         priv->lro.lro_mgr.lro_arr        = priv->lro.lro_desc;
1018         priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1019         priv->lro.lro_mgr.features       = LRO_F_NAPI;
1020         priv->lro.lro_mgr.dev            = priv->dev;
1021         priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1022 }
1023
1024 static void ipoib_setup(struct net_device *dev)
1025 {
1026         struct ipoib_dev_priv *priv = netdev_priv(dev);
1027
1028         dev->open                = ipoib_open;
1029         dev->stop                = ipoib_stop;
1030         dev->change_mtu          = ipoib_change_mtu;
1031         dev->hard_start_xmit     = ipoib_start_xmit;
1032         dev->tx_timeout          = ipoib_timeout;
1033         dev->header_ops          = &ipoib_header_ops;
1034         dev->set_multicast_list  = ipoib_set_mcast_list;
1035         dev->neigh_setup         = ipoib_neigh_setup_dev;
1036
1037         ipoib_set_ethtool_ops(dev);
1038
1039         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1040
1041         dev->watchdog_timeo      = HZ;
1042
1043         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
1044
1045         /*
1046          * We add in INFINIBAND_ALEN to allow for the destination
1047          * address "pseudoheader" for skbs without neighbour struct.
1048          */
1049         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1050         dev->addr_len            = INFINIBAND_ALEN;
1051         dev->type                = ARPHRD_INFINIBAND;
1052         dev->tx_queue_len        = ipoib_sendq_size * 2;
1053         dev->features            = (NETIF_F_VLAN_CHALLENGED     |
1054                                     NETIF_F_HIGHDMA);
1055
1056         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1057
1058         netif_carrier_off(dev);
1059
1060         priv->dev = dev;
1061
1062         ipoib_lro_setup(priv);
1063
1064         spin_lock_init(&priv->lock);
1065
1066         mutex_init(&priv->vlan_mutex);
1067
1068         INIT_LIST_HEAD(&priv->path_list);
1069         INIT_LIST_HEAD(&priv->child_intfs);
1070         INIT_LIST_HEAD(&priv->dead_ahs);
1071         INIT_LIST_HEAD(&priv->multicast_list);
1072
1073         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1074         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
1075         INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1076         INIT_WORK(&priv->flush_light,   ipoib_ib_dev_flush_light);
1077         INIT_WORK(&priv->flush_normal,   ipoib_ib_dev_flush_normal);
1078         INIT_WORK(&priv->flush_heavy,   ipoib_ib_dev_flush_heavy);
1079         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1080         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1081 }
1082
1083 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1084 {
1085         struct net_device *dev;
1086
1087         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1088                            ipoib_setup);
1089         if (!dev)
1090                 return NULL;
1091
1092         return netdev_priv(dev);
1093 }
1094
1095 static ssize_t show_pkey(struct device *dev,
1096                          struct device_attribute *attr, char *buf)
1097 {
1098         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1099
1100         return sprintf(buf, "0x%04x\n", priv->pkey);
1101 }
1102 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1103
1104 static ssize_t show_umcast(struct device *dev,
1105                            struct device_attribute *attr, char *buf)
1106 {
1107         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1108
1109         return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1110 }
1111
1112 static ssize_t set_umcast(struct device *dev,
1113                           struct device_attribute *attr,
1114                           const char *buf, size_t count)
1115 {
1116         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1117         unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1118
1119         if (umcast_val > 0) {
1120                 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1121                 ipoib_warn(priv, "ignoring multicast groups joined directly "
1122                                 "by userspace\n");
1123         } else
1124                 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1125
1126         return count;
1127 }
1128 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1129
1130 int ipoib_add_umcast_attr(struct net_device *dev)
1131 {
1132         return device_create_file(&dev->dev, &dev_attr_umcast);
1133 }
1134
1135 static ssize_t create_child(struct device *dev,
1136                             struct device_attribute *attr,
1137                             const char *buf, size_t count)
1138 {
1139         int pkey;
1140         int ret;
1141
1142         if (sscanf(buf, "%i", &pkey) != 1)
1143                 return -EINVAL;
1144
1145         if (pkey < 0 || pkey > 0xffff)
1146                 return -EINVAL;
1147
1148         /*
1149          * Set the full membership bit, so that we join the right
1150          * broadcast group, etc.
1151          */
1152         pkey |= 0x8000;
1153
1154         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1155
1156         return ret ? ret : count;
1157 }
1158 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1159
1160 static ssize_t delete_child(struct device *dev,
1161                             struct device_attribute *attr,
1162                             const char *buf, size_t count)
1163 {
1164         int pkey;
1165         int ret;
1166
1167         if (sscanf(buf, "%i", &pkey) != 1)
1168                 return -EINVAL;
1169
1170         if (pkey < 0 || pkey > 0xffff)
1171                 return -EINVAL;
1172
1173         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1174
1175         return ret ? ret : count;
1176
1177 }
1178 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1179
1180 int ipoib_add_pkey_attr(struct net_device *dev)
1181 {
1182         return device_create_file(&dev->dev, &dev_attr_pkey);
1183 }
1184
1185 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1186 {
1187         struct ib_device_attr *device_attr;
1188         int result = -ENOMEM;
1189
1190         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1191         if (!device_attr) {
1192                 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1193                        hca->name, sizeof *device_attr);
1194                 return result;
1195         }
1196
1197         result = ib_query_device(hca, device_attr);
1198         if (result) {
1199                 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1200                        hca->name, result);
1201                 kfree(device_attr);
1202                 return result;
1203         }
1204         priv->hca_caps = device_attr->device_cap_flags;
1205
1206         kfree(device_attr);
1207
1208         if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1209                 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1210                 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1211         }
1212
1213         if (lro)
1214                 priv->dev->features |= NETIF_F_LRO;
1215
1216         if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1217                 priv->dev->features |= NETIF_F_TSO;
1218
1219         return 0;
1220 }
1221
1222
1223 static struct net_device *ipoib_add_port(const char *format,
1224                                          struct ib_device *hca, u8 port)
1225 {
1226         struct ipoib_dev_priv *priv;
1227         struct ib_port_attr attr;
1228         int result = -ENOMEM;
1229
1230         priv = ipoib_intf_alloc(format);
1231         if (!priv)
1232                 goto alloc_mem_failed;
1233
1234         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1235
1236         if (!ib_query_port(hca, port, &attr))
1237                 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1238         else {
1239                 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1240                        hca->name, port);
1241                 goto device_init_failed;
1242         }
1243
1244         /* MTU will be reset when mcast join happens */
1245         priv->dev->mtu  = IPOIB_UD_MTU(priv->max_ib_mtu);
1246         priv->mcast_mtu  = priv->admin_mtu = priv->dev->mtu;
1247
1248         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1249         if (result) {
1250                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1251                        hca->name, port, result);
1252                 goto device_init_failed;
1253         }
1254
1255         if (ipoib_set_dev_features(priv, hca))
1256                 goto device_init_failed;
1257
1258         /*
1259          * Set the full membership bit, so that we join the right
1260          * broadcast group, etc.
1261          */
1262         priv->pkey |= 0x8000;
1263
1264         priv->dev->broadcast[8] = priv->pkey >> 8;
1265         priv->dev->broadcast[9] = priv->pkey & 0xff;
1266
1267         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1268         if (result) {
1269                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1270                        hca->name, port, result);
1271                 goto device_init_failed;
1272         } else
1273                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1274
1275         result = ipoib_dev_init(priv->dev, hca, port);
1276         if (result < 0) {
1277                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1278                        hca->name, port, result);
1279                 goto device_init_failed;
1280         }
1281
1282         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1283                               priv->ca, ipoib_event);
1284         result = ib_register_event_handler(&priv->event_handler);
1285         if (result < 0) {
1286                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1287                        "port %d (ret = %d)\n",
1288                        hca->name, port, result);
1289                 goto event_failed;
1290         }
1291
1292         result = register_netdev(priv->dev);
1293         if (result) {
1294                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1295                        hca->name, port, result);
1296                 goto register_failed;
1297         }
1298
1299         ipoib_create_debug_files(priv->dev);
1300
1301         if (ipoib_cm_add_mode_attr(priv->dev))
1302                 goto sysfs_failed;
1303         if (ipoib_add_pkey_attr(priv->dev))
1304                 goto sysfs_failed;
1305         if (ipoib_add_umcast_attr(priv->dev))
1306                 goto sysfs_failed;
1307         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1308                 goto sysfs_failed;
1309         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1310                 goto sysfs_failed;
1311
1312         return priv->dev;
1313
1314 sysfs_failed:
1315         ipoib_delete_debug_files(priv->dev);
1316         unregister_netdev(priv->dev);
1317
1318 register_failed:
1319         ib_unregister_event_handler(&priv->event_handler);
1320         flush_workqueue(ipoib_workqueue);
1321
1322 event_failed:
1323         ipoib_dev_cleanup(priv->dev);
1324
1325 device_init_failed:
1326         free_netdev(priv->dev);
1327
1328 alloc_mem_failed:
1329         return ERR_PTR(result);
1330 }
1331
1332 static void ipoib_add_one(struct ib_device *device)
1333 {
1334         struct list_head *dev_list;
1335         struct net_device *dev;
1336         struct ipoib_dev_priv *priv;
1337         int s, e, p;
1338
1339         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1340                 return;
1341
1342         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1343         if (!dev_list)
1344                 return;
1345
1346         INIT_LIST_HEAD(dev_list);
1347
1348         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1349                 s = 0;
1350                 e = 0;
1351         } else {
1352                 s = 1;
1353                 e = device->phys_port_cnt;
1354         }
1355
1356         for (p = s; p <= e; ++p) {
1357                 dev = ipoib_add_port("ib%d", device, p);
1358                 if (!IS_ERR(dev)) {
1359                         priv = netdev_priv(dev);
1360                         list_add_tail(&priv->list, dev_list);
1361                 }
1362         }
1363
1364         ib_set_client_data(device, &ipoib_client, dev_list);
1365 }
1366
1367 static void ipoib_remove_one(struct ib_device *device)
1368 {
1369         struct ipoib_dev_priv *priv, *tmp;
1370         struct list_head *dev_list;
1371
1372         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1373                 return;
1374
1375         dev_list = ib_get_client_data(device, &ipoib_client);
1376
1377         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1378                 ib_unregister_event_handler(&priv->event_handler);
1379
1380                 rtnl_lock();
1381                 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1382                 rtnl_unlock();
1383
1384                 flush_workqueue(ipoib_workqueue);
1385
1386                 unregister_netdev(priv->dev);
1387                 ipoib_dev_cleanup(priv->dev);
1388                 free_netdev(priv->dev);
1389         }
1390
1391         kfree(dev_list);
1392 }
1393
1394 static int __init ipoib_init_module(void)
1395 {
1396         int ret;
1397
1398         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1399         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1400         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1401
1402         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1403         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1404         ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1405                                                      IPOIB_MIN_QUEUE_SIZE));
1406 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1407         ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1408 #endif
1409
1410         /*
1411          * When copying small received packets, we only copy from the
1412          * linear data part of the SKB, so we rely on this condition.
1413          */
1414         BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1415
1416         ret = ipoib_register_debugfs();
1417         if (ret)
1418                 return ret;
1419
1420         /*
1421          * We create our own workqueue mainly because we want to be
1422          * able to flush it when devices are being removed.  We can't
1423          * use schedule_work()/flush_scheduled_work() because both
1424          * unregister_netdev() and linkwatch_event take the rtnl lock,
1425          * so flush_scheduled_work() can deadlock during device
1426          * removal.
1427          */
1428         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1429         if (!ipoib_workqueue) {
1430                 ret = -ENOMEM;
1431                 goto err_fs;
1432         }
1433
1434         ib_sa_register_client(&ipoib_sa_client);
1435
1436         ret = ib_register_client(&ipoib_client);
1437         if (ret)
1438                 goto err_sa;
1439
1440         return 0;
1441
1442 err_sa:
1443         ib_sa_unregister_client(&ipoib_sa_client);
1444         destroy_workqueue(ipoib_workqueue);
1445
1446 err_fs:
1447         ipoib_unregister_debugfs();
1448
1449         return ret;
1450 }
1451
1452 static void __exit ipoib_cleanup_module(void)
1453 {
1454         ib_unregister_client(&ipoib_client);
1455         ib_sa_unregister_client(&ipoib_sa_client);
1456         ipoib_unregister_debugfs();
1457         destroy_workqueue(ipoib_workqueue);
1458 }
1459
1460 module_init(ipoib_init_module);
1461 module_exit(ipoib_cleanup_module);