[NET]: Make NAPI polling independent of struct net_device objects.
[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  * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
35  */
36
37 #include "ipoib.h"
38
39 #include <linux/module.h>
40
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/kernel.h>
44
45 #include <linux/if_arp.h>       /* For ARPHRD_xxx */
46
47 #include <linux/ip.h>
48 #include <linux/in.h>
49
50 #include <net/dst.h>
51
52 MODULE_AUTHOR("Roland Dreier");
53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
54 MODULE_LICENSE("Dual BSD/GPL");
55
56 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
58
59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
63
64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
65 int ipoib_debug_level;
66
67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
69 #endif
70
71 struct ipoib_path_iter {
72         struct net_device *dev;
73         struct ipoib_path  path;
74 };
75
76 static const u8 ipv4_bcast_addr[] = {
77         0x00, 0xff, 0xff, 0xff,
78         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
79         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
80 };
81
82 struct workqueue_struct *ipoib_workqueue;
83
84 struct ib_sa_client ipoib_sa_client;
85
86 static void ipoib_add_one(struct ib_device *device);
87 static void ipoib_remove_one(struct ib_device *device);
88
89 static struct ib_client ipoib_client = {
90         .name   = "ipoib",
91         .add    = ipoib_add_one,
92         .remove = ipoib_remove_one
93 };
94
95 int ipoib_open(struct net_device *dev)
96 {
97         struct ipoib_dev_priv *priv = netdev_priv(dev);
98
99         ipoib_dbg(priv, "bringing up interface\n");
100
101         napi_enable(&priv->napi);
102         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
103
104         if (ipoib_pkey_dev_delay_open(dev))
105                 return 0;
106
107         if (ipoib_ib_dev_open(dev)) {
108                 napi_disable(&priv->napi);
109                 return -EINVAL;
110         }
111
112         if (ipoib_ib_dev_up(dev)) {
113                 ipoib_ib_dev_stop(dev, 1);
114                 napi_disable(&priv->napi);
115                 return -EINVAL;
116         }
117
118         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
119                 struct ipoib_dev_priv *cpriv;
120
121                 /* Bring up any child interfaces too */
122                 mutex_lock(&priv->vlan_mutex);
123                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
124                         int flags;
125
126                         flags = cpriv->dev->flags;
127                         if (flags & IFF_UP)
128                                 continue;
129
130                         dev_change_flags(cpriv->dev, flags | IFF_UP);
131                 }
132                 mutex_unlock(&priv->vlan_mutex);
133         }
134
135         netif_start_queue(dev);
136
137         return 0;
138 }
139
140 static int ipoib_stop(struct net_device *dev)
141 {
142         struct ipoib_dev_priv *priv = netdev_priv(dev);
143
144         ipoib_dbg(priv, "stopping interface\n");
145
146         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
147         napi_disable(&priv->napi);
148
149         netif_stop_queue(dev);
150
151         clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
152
153         /*
154          * Now flush workqueue to make sure a scheduled task doesn't
155          * bring our internal state back up.
156          */
157         flush_workqueue(ipoib_workqueue);
158
159         ipoib_ib_dev_down(dev, 1);
160         ipoib_ib_dev_stop(dev, 1);
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) && new_mtu <= IPOIB_CM_MTU) {
188                 if (new_mtu > priv->mcast_mtu)
189                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
190                                    priv->mcast_mtu);
191                 dev->mtu = new_mtu;
192                 return 0;
193         }
194
195         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
196                 return -EINVAL;
197         }
198
199         priv->admin_mtu = new_mtu;
200
201         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
202
203         return 0;
204 }
205
206 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
207 {
208         struct ipoib_dev_priv *priv = netdev_priv(dev);
209         struct rb_node *n = priv->path_tree.rb_node;
210         struct ipoib_path *path;
211         int ret;
212
213         while (n) {
214                 path = rb_entry(n, struct ipoib_path, rb_node);
215
216                 ret = memcmp(gid, path->pathrec.dgid.raw,
217                              sizeof (union ib_gid));
218
219                 if (ret < 0)
220                         n = n->rb_left;
221                 else if (ret > 0)
222                         n = n->rb_right;
223                 else
224                         return path;
225         }
226
227         return NULL;
228 }
229
230 static int __path_add(struct net_device *dev, struct ipoib_path *path)
231 {
232         struct ipoib_dev_priv *priv = netdev_priv(dev);
233         struct rb_node **n = &priv->path_tree.rb_node;
234         struct rb_node *pn = NULL;
235         struct ipoib_path *tpath;
236         int ret;
237
238         while (*n) {
239                 pn = *n;
240                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
241
242                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
243                              sizeof (union ib_gid));
244                 if (ret < 0)
245                         n = &pn->rb_left;
246                 else if (ret > 0)
247                         n = &pn->rb_right;
248                 else
249                         return -EEXIST;
250         }
251
252         rb_link_node(&path->rb_node, pn, n);
253         rb_insert_color(&path->rb_node, &priv->path_tree);
254
255         list_add_tail(&path->list, &priv->path_list);
256
257         return 0;
258 }
259
260 static void path_free(struct net_device *dev, struct ipoib_path *path)
261 {
262         struct ipoib_dev_priv *priv = netdev_priv(dev);
263         struct ipoib_neigh *neigh, *tn;
264         struct sk_buff *skb;
265         unsigned long flags;
266
267         while ((skb = __skb_dequeue(&path->queue)))
268                 dev_kfree_skb_irq(skb);
269
270         spin_lock_irqsave(&priv->lock, flags);
271
272         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
273                 /*
274                  * It's safe to call ipoib_put_ah() inside priv->lock
275                  * here, because we know that path->ah will always
276                  * hold one more reference, so ipoib_put_ah() will
277                  * never do more than decrement the ref count.
278                  */
279                 if (neigh->ah)
280                         ipoib_put_ah(neigh->ah);
281
282                 ipoib_neigh_free(dev, neigh);
283         }
284
285         spin_unlock_irqrestore(&priv->lock, flags);
286
287         if (path->ah)
288                 ipoib_put_ah(path->ah);
289
290         kfree(path);
291 }
292
293 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
294
295 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
296 {
297         struct ipoib_path_iter *iter;
298
299         iter = kmalloc(sizeof *iter, GFP_KERNEL);
300         if (!iter)
301                 return NULL;
302
303         iter->dev = dev;
304         memset(iter->path.pathrec.dgid.raw, 0, 16);
305
306         if (ipoib_path_iter_next(iter)) {
307                 kfree(iter);
308                 return NULL;
309         }
310
311         return iter;
312 }
313
314 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
315 {
316         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
317         struct rb_node *n;
318         struct ipoib_path *path;
319         int ret = 1;
320
321         spin_lock_irq(&priv->lock);
322
323         n = rb_first(&priv->path_tree);
324
325         while (n) {
326                 path = rb_entry(n, struct ipoib_path, rb_node);
327
328                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
329                            sizeof (union ib_gid)) < 0) {
330                         iter->path = *path;
331                         ret = 0;
332                         break;
333                 }
334
335                 n = rb_next(n);
336         }
337
338         spin_unlock_irq(&priv->lock);
339
340         return ret;
341 }
342
343 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
344                           struct ipoib_path *path)
345 {
346         *path = iter->path;
347 }
348
349 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
350
351 void ipoib_flush_paths(struct net_device *dev)
352 {
353         struct ipoib_dev_priv *priv = netdev_priv(dev);
354         struct ipoib_path *path, *tp;
355         LIST_HEAD(remove_list);
356
357         spin_lock_irq(&priv->tx_lock);
358         spin_lock(&priv->lock);
359
360         list_splice(&priv->path_list, &remove_list);
361         INIT_LIST_HEAD(&priv->path_list);
362
363         list_for_each_entry(path, &remove_list, list)
364                 rb_erase(&path->rb_node, &priv->path_tree);
365
366         list_for_each_entry_safe(path, tp, &remove_list, list) {
367                 if (path->query)
368                         ib_sa_cancel_query(path->query_id, path->query);
369                 spin_unlock(&priv->lock);
370                 spin_unlock_irq(&priv->tx_lock);
371                 wait_for_completion(&path->done);
372                 path_free(dev, path);
373                 spin_lock_irq(&priv->tx_lock);
374                 spin_lock(&priv->lock);
375         }
376         spin_unlock(&priv->lock);
377         spin_unlock_irq(&priv->tx_lock);
378 }
379
380 static void path_rec_completion(int status,
381                                 struct ib_sa_path_rec *pathrec,
382                                 void *path_ptr)
383 {
384         struct ipoib_path *path = path_ptr;
385         struct net_device *dev = path->dev;
386         struct ipoib_dev_priv *priv = netdev_priv(dev);
387         struct ipoib_ah *ah = NULL;
388         struct ipoib_neigh *neigh, *tn;
389         struct sk_buff_head skqueue;
390         struct sk_buff *skb;
391         unsigned long flags;
392
393         if (!status)
394                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
395                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
396         else
397                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
398                           status, IPOIB_GID_ARG(path->pathrec.dgid));
399
400         skb_queue_head_init(&skqueue);
401
402         if (!status) {
403                 struct ib_ah_attr av;
404
405                 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
406                         ah = ipoib_create_ah(dev, priv->pd, &av);
407         }
408
409         spin_lock_irqsave(&priv->lock, flags);
410
411         path->ah = ah;
412
413         if (ah) {
414                 path->pathrec = *pathrec;
415
416                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
417                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
418
419                 while ((skb = __skb_dequeue(&path->queue)))
420                         __skb_queue_tail(&skqueue, skb);
421
422                 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
423                         kref_get(&path->ah->ref);
424                         neigh->ah = path->ah;
425                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
426                                sizeof(union ib_gid));
427
428                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
429                                 if (!ipoib_cm_get(neigh))
430                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
431                                                                                path,
432                                                                                neigh));
433                                 if (!ipoib_cm_get(neigh)) {
434                                         list_del(&neigh->list);
435                                         if (neigh->ah)
436                                                 ipoib_put_ah(neigh->ah);
437                                         ipoib_neigh_free(dev, neigh);
438                                         continue;
439                                 }
440                         }
441
442                         while ((skb = __skb_dequeue(&neigh->queue)))
443                                 __skb_queue_tail(&skqueue, skb);
444                 }
445         }
446
447         path->query = NULL;
448         complete(&path->done);
449
450         spin_unlock_irqrestore(&priv->lock, flags);
451
452         while ((skb = __skb_dequeue(&skqueue))) {
453                 skb->dev = dev;
454                 if (dev_queue_xmit(skb))
455                         ipoib_warn(priv, "dev_queue_xmit failed "
456                                    "to requeue packet\n");
457         }
458 }
459
460 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
461 {
462         struct ipoib_dev_priv *priv = netdev_priv(dev);
463         struct ipoib_path *path;
464
465         path = kzalloc(sizeof *path, GFP_ATOMIC);
466         if (!path)
467                 return NULL;
468
469         path->dev = dev;
470
471         skb_queue_head_init(&path->queue);
472
473         INIT_LIST_HEAD(&path->neigh_list);
474
475         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
476         path->pathrec.sgid      = priv->local_gid;
477         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
478         path->pathrec.numb_path = 1;
479
480         return path;
481 }
482
483 static int path_rec_start(struct net_device *dev,
484                           struct ipoib_path *path)
485 {
486         struct ipoib_dev_priv *priv = netdev_priv(dev);
487
488         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
489                   IPOIB_GID_ARG(path->pathrec.dgid));
490
491         init_completion(&path->done);
492
493         path->query_id =
494                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
495                                    &path->pathrec,
496                                    IB_SA_PATH_REC_DGID          |
497                                    IB_SA_PATH_REC_SGID          |
498                                    IB_SA_PATH_REC_NUMB_PATH     |
499                                    IB_SA_PATH_REC_PKEY,
500                                    1000, GFP_ATOMIC,
501                                    path_rec_completion,
502                                    path, &path->query);
503         if (path->query_id < 0) {
504                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
505                 path->query = NULL;
506                 return path->query_id;
507         }
508
509         return 0;
510 }
511
512 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
513 {
514         struct ipoib_dev_priv *priv = netdev_priv(dev);
515         struct ipoib_path *path;
516         struct ipoib_neigh *neigh;
517
518         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
519         if (!neigh) {
520                 ++priv->stats.tx_dropped;
521                 dev_kfree_skb_any(skb);
522                 return;
523         }
524
525         /*
526          * We can only be called from ipoib_start_xmit, so we're
527          * inside tx_lock -- no need to save/restore flags.
528          */
529         spin_lock(&priv->lock);
530
531         path = __path_find(dev, skb->dst->neighbour->ha + 4);
532         if (!path) {
533                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
534                 if (!path)
535                         goto err_path;
536
537                 __path_add(dev, path);
538         }
539
540         list_add_tail(&neigh->list, &path->neigh_list);
541
542         if (path->ah) {
543                 kref_get(&path->ah->ref);
544                 neigh->ah = path->ah;
545                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
546                        sizeof(union ib_gid));
547
548                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
549                         if (!ipoib_cm_get(neigh))
550                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
551                         if (!ipoib_cm_get(neigh)) {
552                                 list_del(&neigh->list);
553                                 if (neigh->ah)
554                                         ipoib_put_ah(neigh->ah);
555                                 ipoib_neigh_free(dev, neigh);
556                                 goto err_drop;
557                         }
558                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
559                                 __skb_queue_tail(&neigh->queue, skb);
560                         else {
561                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
562                                            skb_queue_len(&neigh->queue));
563                                 goto err_drop;
564                         }
565                 } else
566                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
567         } else {
568                 neigh->ah  = NULL;
569
570                 if (!path->query && path_rec_start(dev, path))
571                         goto err_list;
572
573                 __skb_queue_tail(&neigh->queue, skb);
574         }
575
576         spin_unlock(&priv->lock);
577         return;
578
579 err_list:
580         list_del(&neigh->list);
581
582 err_path:
583         ipoib_neigh_free(dev, neigh);
584 err_drop:
585         ++priv->stats.tx_dropped;
586         dev_kfree_skb_any(skb);
587
588         spin_unlock(&priv->lock);
589 }
590
591 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
592 {
593         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
594
595         /* Look up path record for unicasts */
596         if (skb->dst->neighbour->ha[4] != 0xff) {
597                 neigh_add_path(skb, dev);
598                 return;
599         }
600
601         /* Add in the P_Key for multicasts */
602         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
603         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
604         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
605 }
606
607 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
608                              struct ipoib_pseudoheader *phdr)
609 {
610         struct ipoib_dev_priv *priv = netdev_priv(dev);
611         struct ipoib_path *path;
612
613         /*
614          * We can only be called from ipoib_start_xmit, so we're
615          * inside tx_lock -- no need to save/restore flags.
616          */
617         spin_lock(&priv->lock);
618
619         path = __path_find(dev, phdr->hwaddr + 4);
620         if (!path) {
621                 path = path_rec_create(dev, phdr->hwaddr + 4);
622                 if (path) {
623                         /* put pseudoheader back on for next time */
624                         skb_push(skb, sizeof *phdr);
625                         __skb_queue_tail(&path->queue, skb);
626
627                         if (path_rec_start(dev, path)) {
628                                 spin_unlock(&priv->lock);
629                                 path_free(dev, path);
630                                 return;
631                         } else
632                                 __path_add(dev, path);
633                 } else {
634                         ++priv->stats.tx_dropped;
635                         dev_kfree_skb_any(skb);
636                 }
637
638                 spin_unlock(&priv->lock);
639                 return;
640         }
641
642         if (path->ah) {
643                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
644                           be16_to_cpu(path->pathrec.dlid));
645
646                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
647         } else if ((path->query || !path_rec_start(dev, path)) &&
648                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
649                 /* put pseudoheader back on for next time */
650                 skb_push(skb, sizeof *phdr);
651                 __skb_queue_tail(&path->queue, skb);
652         } else {
653                 ++priv->stats.tx_dropped;
654                 dev_kfree_skb_any(skb);
655         }
656
657         spin_unlock(&priv->lock);
658 }
659
660 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
661 {
662         struct ipoib_dev_priv *priv = netdev_priv(dev);
663         struct ipoib_neigh *neigh;
664         unsigned long flags;
665
666         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
667                 return NETDEV_TX_LOCKED;
668
669         /*
670          * Check if our queue is stopped.  Since we have the LLTX bit
671          * set, we can't rely on netif_stop_queue() preventing our
672          * xmit function from being called with a full queue.
673          */
674         if (unlikely(netif_queue_stopped(dev))) {
675                 spin_unlock_irqrestore(&priv->tx_lock, flags);
676                 return NETDEV_TX_BUSY;
677         }
678
679         if (likely(skb->dst && skb->dst->neighbour)) {
680                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
681                         ipoib_path_lookup(skb, dev);
682                         goto out;
683                 }
684
685                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
686
687                 if (ipoib_cm_get(neigh)) {
688                         if (ipoib_cm_up(neigh)) {
689                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
690                                 goto out;
691                         }
692                 } else if (neigh->ah) {
693                         if (unlikely(memcmp(&neigh->dgid.raw,
694                                             skb->dst->neighbour->ha + 4,
695                                             sizeof(union ib_gid)))) {
696                                 spin_lock(&priv->lock);
697                                 /*
698                                  * It's safe to call ipoib_put_ah() inside
699                                  * priv->lock here, because we know that
700                                  * path->ah will always hold one more reference,
701                                  * so ipoib_put_ah() will never do more than
702                                  * decrement the ref count.
703                                  */
704                                 ipoib_put_ah(neigh->ah);
705                                 list_del(&neigh->list);
706                                 ipoib_neigh_free(dev, neigh);
707                                 spin_unlock(&priv->lock);
708                                 ipoib_path_lookup(skb, dev);
709                                 goto out;
710                         }
711
712                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
713                         goto out;
714                 }
715
716                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
717                         spin_lock(&priv->lock);
718                         __skb_queue_tail(&neigh->queue, skb);
719                         spin_unlock(&priv->lock);
720                 } else {
721                         ++priv->stats.tx_dropped;
722                         dev_kfree_skb_any(skb);
723                 }
724         } else {
725                 struct ipoib_pseudoheader *phdr =
726                         (struct ipoib_pseudoheader *) skb->data;
727                 skb_pull(skb, sizeof *phdr);
728
729                 if (phdr->hwaddr[4] == 0xff) {
730                         /* Add in the P_Key for multicast*/
731                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
732                         phdr->hwaddr[9] = priv->pkey & 0xff;
733
734                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
735                 } else {
736                         /* unicast GID -- should be ARP or RARP reply */
737
738                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
739                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
740                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
741                                            IPOIB_GID_FMT "\n",
742                                            skb->dst ? "neigh" : "dst",
743                                            be16_to_cpup((__be16 *) skb->data),
744                                            IPOIB_QPN(phdr->hwaddr),
745                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
746                                 dev_kfree_skb_any(skb);
747                                 ++priv->stats.tx_dropped;
748                                 goto out;
749                         }
750
751                         unicast_arp_send(skb, dev, phdr);
752                 }
753         }
754
755 out:
756         spin_unlock_irqrestore(&priv->tx_lock, flags);
757
758         return NETDEV_TX_OK;
759 }
760
761 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
762 {
763         struct ipoib_dev_priv *priv = netdev_priv(dev);
764
765         return &priv->stats;
766 }
767
768 static void ipoib_timeout(struct net_device *dev)
769 {
770         struct ipoib_dev_priv *priv = netdev_priv(dev);
771
772         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
773                    jiffies_to_msecs(jiffies - dev->trans_start));
774         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
775                    netif_queue_stopped(dev),
776                    priv->tx_head, priv->tx_tail);
777         /* XXX reset QP, etc. */
778 }
779
780 static int ipoib_hard_header(struct sk_buff *skb,
781                              struct net_device *dev,
782                              unsigned short type,
783                              void *daddr, void *saddr, unsigned len)
784 {
785         struct ipoib_header *header;
786
787         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
788
789         header->proto = htons(type);
790         header->reserved = 0;
791
792         /*
793          * If we don't have a neighbour structure, stuff the
794          * destination address onto the front of the skb so we can
795          * figure out where to send the packet later.
796          */
797         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
798                 struct ipoib_pseudoheader *phdr =
799                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
800                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
801         }
802
803         return 0;
804 }
805
806 static void ipoib_set_mcast_list(struct net_device *dev)
807 {
808         struct ipoib_dev_priv *priv = netdev_priv(dev);
809
810         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
811                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
812                 return;
813         }
814
815         queue_work(ipoib_workqueue, &priv->restart_task);
816 }
817
818 static void ipoib_neigh_cleanup(struct neighbour *n)
819 {
820         struct ipoib_neigh *neigh;
821         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
822         unsigned long flags;
823         struct ipoib_ah *ah = NULL;
824
825         ipoib_dbg(priv,
826                   "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
827                   IPOIB_QPN(n->ha),
828                   IPOIB_GID_RAW_ARG(n->ha + 4));
829
830         spin_lock_irqsave(&priv->lock, flags);
831
832         neigh = *to_ipoib_neigh(n);
833         if (neigh) {
834                 if (neigh->ah)
835                         ah = neigh->ah;
836                 list_del(&neigh->list);
837                 ipoib_neigh_free(n->dev, neigh);
838         }
839
840         spin_unlock_irqrestore(&priv->lock, flags);
841
842         if (ah)
843                 ipoib_put_ah(ah);
844 }
845
846 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
847 {
848         struct ipoib_neigh *neigh;
849
850         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
851         if (!neigh)
852                 return NULL;
853
854         neigh->neighbour = neighbour;
855         *to_ipoib_neigh(neighbour) = neigh;
856         skb_queue_head_init(&neigh->queue);
857         ipoib_cm_set(neigh, NULL);
858
859         return neigh;
860 }
861
862 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
863 {
864         struct ipoib_dev_priv *priv = netdev_priv(dev);
865         struct sk_buff *skb;
866         *to_ipoib_neigh(neigh->neighbour) = NULL;
867         while ((skb = __skb_dequeue(&neigh->queue))) {
868                 ++priv->stats.tx_dropped;
869                 dev_kfree_skb_any(skb);
870         }
871         if (ipoib_cm_get(neigh))
872                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
873         kfree(neigh);
874 }
875
876 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
877 {
878         parms->neigh_cleanup = ipoib_neigh_cleanup;
879
880         return 0;
881 }
882
883 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
884 {
885         struct ipoib_dev_priv *priv = netdev_priv(dev);
886
887         /* Allocate RX/TX "rings" to hold queued skbs */
888         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
889                                 GFP_KERNEL);
890         if (!priv->rx_ring) {
891                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
892                        ca->name, ipoib_recvq_size);
893                 goto out;
894         }
895
896         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
897                                 GFP_KERNEL);
898         if (!priv->tx_ring) {
899                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
900                        ca->name, ipoib_sendq_size);
901                 goto out_rx_ring_cleanup;
902         }
903
904         /* priv->tx_head & tx_tail are already 0 */
905
906         if (ipoib_ib_dev_init(dev, ca, port))
907                 goto out_tx_ring_cleanup;
908
909         return 0;
910
911 out_tx_ring_cleanup:
912         kfree(priv->tx_ring);
913
914 out_rx_ring_cleanup:
915         kfree(priv->rx_ring);
916
917 out:
918         return -ENOMEM;
919 }
920
921 void ipoib_dev_cleanup(struct net_device *dev)
922 {
923         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
924
925         ipoib_delete_debug_files(dev);
926
927         /* Delete any child interfaces first */
928         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
929                 unregister_netdev(cpriv->dev);
930                 ipoib_dev_cleanup(cpriv->dev);
931                 free_netdev(cpriv->dev);
932         }
933
934         ipoib_ib_dev_cleanup(dev);
935
936         kfree(priv->rx_ring);
937         kfree(priv->tx_ring);
938
939         priv->rx_ring = NULL;
940         priv->tx_ring = NULL;
941 }
942
943 static void ipoib_setup(struct net_device *dev)
944 {
945         struct ipoib_dev_priv *priv = netdev_priv(dev);
946
947         dev->open                = ipoib_open;
948         dev->stop                = ipoib_stop;
949         dev->change_mtu          = ipoib_change_mtu;
950         dev->hard_start_xmit     = ipoib_start_xmit;
951         dev->get_stats           = ipoib_get_stats;
952         dev->tx_timeout          = ipoib_timeout;
953         dev->hard_header         = ipoib_hard_header;
954         dev->set_multicast_list  = ipoib_set_mcast_list;
955         dev->neigh_setup         = ipoib_neigh_setup_dev;
956
957         netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
958
959         dev->watchdog_timeo      = HZ;
960
961         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
962
963         /*
964          * We add in INFINIBAND_ALEN to allow for the destination
965          * address "pseudoheader" for skbs without neighbour struct.
966          */
967         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
968         dev->addr_len            = INFINIBAND_ALEN;
969         dev->type                = ARPHRD_INFINIBAND;
970         dev->tx_queue_len        = ipoib_sendq_size * 2;
971         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
972
973         /* MTU will be reset when mcast join happens */
974         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
975         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
976
977         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
978
979         netif_carrier_off(dev);
980
981         SET_MODULE_OWNER(dev);
982
983         priv->dev = dev;
984
985         spin_lock_init(&priv->lock);
986         spin_lock_init(&priv->tx_lock);
987
988         mutex_init(&priv->mcast_mutex);
989         mutex_init(&priv->vlan_mutex);
990
991         INIT_LIST_HEAD(&priv->path_list);
992         INIT_LIST_HEAD(&priv->child_intfs);
993         INIT_LIST_HEAD(&priv->dead_ahs);
994         INIT_LIST_HEAD(&priv->multicast_list);
995
996         INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
997         INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
998         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
999         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
1000         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1001         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1002 }
1003
1004 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1005 {
1006         struct net_device *dev;
1007
1008         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1009                            ipoib_setup);
1010         if (!dev)
1011                 return NULL;
1012
1013         return netdev_priv(dev);
1014 }
1015
1016 static ssize_t show_pkey(struct device *dev,
1017                          struct device_attribute *attr, char *buf)
1018 {
1019         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1020
1021         return sprintf(buf, "0x%04x\n", priv->pkey);
1022 }
1023 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1024
1025 static ssize_t create_child(struct device *dev,
1026                             struct device_attribute *attr,
1027                             const char *buf, size_t count)
1028 {
1029         int pkey;
1030         int ret;
1031
1032         if (sscanf(buf, "%i", &pkey) != 1)
1033                 return -EINVAL;
1034
1035         if (pkey < 0 || pkey > 0xffff)
1036                 return -EINVAL;
1037
1038         /*
1039          * Set the full membership bit, so that we join the right
1040          * broadcast group, etc.
1041          */
1042         pkey |= 0x8000;
1043
1044         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1045
1046         return ret ? ret : count;
1047 }
1048 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1049
1050 static ssize_t delete_child(struct device *dev,
1051                             struct device_attribute *attr,
1052                             const char *buf, size_t count)
1053 {
1054         int pkey;
1055         int ret;
1056
1057         if (sscanf(buf, "%i", &pkey) != 1)
1058                 return -EINVAL;
1059
1060         if (pkey < 0 || pkey > 0xffff)
1061                 return -EINVAL;
1062
1063         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1064
1065         return ret ? ret : count;
1066
1067 }
1068 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1069
1070 int ipoib_add_pkey_attr(struct net_device *dev)
1071 {
1072         return device_create_file(&dev->dev, &dev_attr_pkey);
1073 }
1074
1075 static struct net_device *ipoib_add_port(const char *format,
1076                                          struct ib_device *hca, u8 port)
1077 {
1078         struct ipoib_dev_priv *priv;
1079         int result = -ENOMEM;
1080
1081         priv = ipoib_intf_alloc(format);
1082         if (!priv)
1083                 goto alloc_mem_failed;
1084
1085         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1086
1087         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1088         if (result) {
1089                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1090                        hca->name, port, result);
1091                 goto alloc_mem_failed;
1092         }
1093
1094         /*
1095          * Set the full membership bit, so that we join the right
1096          * broadcast group, etc.
1097          */
1098         priv->pkey |= 0x8000;
1099
1100         priv->dev->broadcast[8] = priv->pkey >> 8;
1101         priv->dev->broadcast[9] = priv->pkey & 0xff;
1102
1103         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1104         if (result) {
1105                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1106                        hca->name, port, result);
1107                 goto alloc_mem_failed;
1108         } else
1109                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1110
1111
1112         result = ipoib_dev_init(priv->dev, hca, port);
1113         if (result < 0) {
1114                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1115                        hca->name, port, result);
1116                 goto device_init_failed;
1117         }
1118
1119         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1120                               priv->ca, ipoib_event);
1121         result = ib_register_event_handler(&priv->event_handler);
1122         if (result < 0) {
1123                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1124                        "port %d (ret = %d)\n",
1125                        hca->name, port, result);
1126                 goto event_failed;
1127         }
1128
1129         result = register_netdev(priv->dev);
1130         if (result) {
1131                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1132                        hca->name, port, result);
1133                 goto register_failed;
1134         }
1135
1136         ipoib_create_debug_files(priv->dev);
1137
1138         if (ipoib_cm_add_mode_attr(priv->dev))
1139                 goto sysfs_failed;
1140         if (ipoib_add_pkey_attr(priv->dev))
1141                 goto sysfs_failed;
1142         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1143                 goto sysfs_failed;
1144         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1145                 goto sysfs_failed;
1146
1147         return priv->dev;
1148
1149 sysfs_failed:
1150         ipoib_delete_debug_files(priv->dev);
1151         unregister_netdev(priv->dev);
1152
1153 register_failed:
1154         ib_unregister_event_handler(&priv->event_handler);
1155         flush_scheduled_work();
1156
1157 event_failed:
1158         ipoib_dev_cleanup(priv->dev);
1159
1160 device_init_failed:
1161         free_netdev(priv->dev);
1162
1163 alloc_mem_failed:
1164         return ERR_PTR(result);
1165 }
1166
1167 static void ipoib_add_one(struct ib_device *device)
1168 {
1169         struct list_head *dev_list;
1170         struct net_device *dev;
1171         struct ipoib_dev_priv *priv;
1172         int s, e, p;
1173
1174         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1175                 return;
1176
1177         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1178         if (!dev_list)
1179                 return;
1180
1181         INIT_LIST_HEAD(dev_list);
1182
1183         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1184                 s = 0;
1185                 e = 0;
1186         } else {
1187                 s = 1;
1188                 e = device->phys_port_cnt;
1189         }
1190
1191         for (p = s; p <= e; ++p) {
1192                 dev = ipoib_add_port("ib%d", device, p);
1193                 if (!IS_ERR(dev)) {
1194                         priv = netdev_priv(dev);
1195                         list_add_tail(&priv->list, dev_list);
1196                 }
1197         }
1198
1199         ib_set_client_data(device, &ipoib_client, dev_list);
1200 }
1201
1202 static void ipoib_remove_one(struct ib_device *device)
1203 {
1204         struct ipoib_dev_priv *priv, *tmp;
1205         struct list_head *dev_list;
1206
1207         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1208                 return;
1209
1210         dev_list = ib_get_client_data(device, &ipoib_client);
1211
1212         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1213                 ib_unregister_event_handler(&priv->event_handler);
1214                 flush_scheduled_work();
1215
1216                 unregister_netdev(priv->dev);
1217                 ipoib_dev_cleanup(priv->dev);
1218                 free_netdev(priv->dev);
1219         }
1220
1221         kfree(dev_list);
1222 }
1223
1224 static int __init ipoib_init_module(void)
1225 {
1226         int ret;
1227
1228         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1229         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1230         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1231
1232         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1233         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1234         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1235
1236         ret = ipoib_register_debugfs();
1237         if (ret)
1238                 return ret;
1239
1240         /*
1241          * We create our own workqueue mainly because we want to be
1242          * able to flush it when devices are being removed.  We can't
1243          * use schedule_work()/flush_scheduled_work() because both
1244          * unregister_netdev() and linkwatch_event take the rtnl lock,
1245          * so flush_scheduled_work() can deadlock during device
1246          * removal.
1247          */
1248         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1249         if (!ipoib_workqueue) {
1250                 ret = -ENOMEM;
1251                 goto err_fs;
1252         }
1253
1254         ib_sa_register_client(&ipoib_sa_client);
1255
1256         ret = ib_register_client(&ipoib_client);
1257         if (ret)
1258                 goto err_sa;
1259
1260         return 0;
1261
1262 err_sa:
1263         ib_sa_unregister_client(&ipoib_sa_client);
1264         destroy_workqueue(ipoib_workqueue);
1265
1266 err_fs:
1267         ipoib_unregister_debugfs();
1268
1269         return ret;
1270 }
1271
1272 static void __exit ipoib_cleanup_module(void)
1273 {
1274         ib_unregister_client(&ipoib_client);
1275         ib_sa_unregister_client(&ipoib_sa_client);
1276         ipoib_unregister_debugfs();
1277         destroy_workqueue(ipoib_workqueue);
1278 }
1279
1280 module_init(ipoib_init_module);
1281 module_exit(ipoib_cleanup_module);