85522daeb946a25a6c4cba91568fbeecf48508bc
[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 #define IPOIB_QPN(ha) (be32_to_cpup((__be32 *) ha) & 0xffffff)
53
54 MODULE_AUTHOR("Roland Dreier");
55 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
56 MODULE_LICENSE("Dual BSD/GPL");
57
58 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
59 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
60
61 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
62 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
63 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
64 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
65
66 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
67 int ipoib_debug_level;
68
69 module_param_named(debug_level, ipoib_debug_level, int, 0644);
70 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
71 #endif
72
73 struct ipoib_path_iter {
74         struct net_device *dev;
75         struct ipoib_path  path;
76 };
77
78 static const u8 ipv4_bcast_addr[] = {
79         0x00, 0xff, 0xff, 0xff,
80         0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
81         0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
82 };
83
84 struct workqueue_struct *ipoib_workqueue;
85
86 struct ib_sa_client ipoib_sa_client;
87
88 static void ipoib_add_one(struct ib_device *device);
89 static void ipoib_remove_one(struct ib_device *device);
90
91 static struct ib_client ipoib_client = {
92         .name   = "ipoib",
93         .add    = ipoib_add_one,
94         .remove = ipoib_remove_one
95 };
96
97 int ipoib_open(struct net_device *dev)
98 {
99         struct ipoib_dev_priv *priv = netdev_priv(dev);
100
101         ipoib_dbg(priv, "bringing up interface\n");
102
103         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
104
105         if (ipoib_pkey_dev_delay_open(dev))
106                 return 0;
107
108         if (ipoib_ib_dev_open(dev))
109                 return -EINVAL;
110
111         if (ipoib_ib_dev_up(dev)) {
112                 ipoib_ib_dev_stop(dev);
113                 return -EINVAL;
114         }
115
116         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
117                 struct ipoib_dev_priv *cpriv;
118
119                 /* Bring up any child interfaces too */
120                 mutex_lock(&priv->vlan_mutex);
121                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
122                         int flags;
123
124                         flags = cpriv->dev->flags;
125                         if (flags & IFF_UP)
126                                 continue;
127
128                         dev_change_flags(cpriv->dev, flags | IFF_UP);
129                 }
130                 mutex_unlock(&priv->vlan_mutex);
131         }
132
133         netif_start_queue(dev);
134
135         return 0;
136 }
137
138 static int ipoib_stop(struct net_device *dev)
139 {
140         struct ipoib_dev_priv *priv = netdev_priv(dev);
141
142         ipoib_dbg(priv, "stopping interface\n");
143
144         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
145
146         netif_stop_queue(dev);
147
148         /*
149          * Now flush workqueue to make sure a scheduled task doesn't
150          * bring our internal state back up.
151          */
152         flush_workqueue(ipoib_workqueue);
153
154         ipoib_ib_dev_down(dev, 1);
155         ipoib_ib_dev_stop(dev);
156
157         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
158                 struct ipoib_dev_priv *cpriv;
159
160                 /* Bring down any child interfaces too */
161                 mutex_lock(&priv->vlan_mutex);
162                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
163                         int flags;
164
165                         flags = cpriv->dev->flags;
166                         if (!(flags & IFF_UP))
167                                 continue;
168
169                         dev_change_flags(cpriv->dev, flags & ~IFF_UP);
170                 }
171                 mutex_unlock(&priv->vlan_mutex);
172         }
173
174         return 0;
175 }
176
177 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
178 {
179         struct ipoib_dev_priv *priv = netdev_priv(dev);
180
181         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
182                 return -EINVAL;
183
184         priv->admin_mtu = new_mtu;
185
186         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
187
188         return 0;
189 }
190
191 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
192 {
193         struct ipoib_dev_priv *priv = netdev_priv(dev);
194         struct rb_node *n = priv->path_tree.rb_node;
195         struct ipoib_path *path;
196         int ret;
197
198         while (n) {
199                 path = rb_entry(n, struct ipoib_path, rb_node);
200
201                 ret = memcmp(gid, path->pathrec.dgid.raw,
202                              sizeof (union ib_gid));
203
204                 if (ret < 0)
205                         n = n->rb_left;
206                 else if (ret > 0)
207                         n = n->rb_right;
208                 else
209                         return path;
210         }
211
212         return NULL;
213 }
214
215 static int __path_add(struct net_device *dev, struct ipoib_path *path)
216 {
217         struct ipoib_dev_priv *priv = netdev_priv(dev);
218         struct rb_node **n = &priv->path_tree.rb_node;
219         struct rb_node *pn = NULL;
220         struct ipoib_path *tpath;
221         int ret;
222
223         while (*n) {
224                 pn = *n;
225                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
226
227                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
228                              sizeof (union ib_gid));
229                 if (ret < 0)
230                         n = &pn->rb_left;
231                 else if (ret > 0)
232                         n = &pn->rb_right;
233                 else
234                         return -EEXIST;
235         }
236
237         rb_link_node(&path->rb_node, pn, n);
238         rb_insert_color(&path->rb_node, &priv->path_tree);
239
240         list_add_tail(&path->list, &priv->path_list);
241
242         return 0;
243 }
244
245 static void path_free(struct net_device *dev, struct ipoib_path *path)
246 {
247         struct ipoib_dev_priv *priv = netdev_priv(dev);
248         struct ipoib_neigh *neigh, *tn;
249         struct sk_buff *skb;
250         unsigned long flags;
251
252         while ((skb = __skb_dequeue(&path->queue)))
253                 dev_kfree_skb_irq(skb);
254
255         spin_lock_irqsave(&priv->lock, flags);
256
257         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
258                 /*
259                  * It's safe to call ipoib_put_ah() inside priv->lock
260                  * here, because we know that path->ah will always
261                  * hold one more reference, so ipoib_put_ah() will
262                  * never do more than decrement the ref count.
263                  */
264                 if (neigh->ah)
265                         ipoib_put_ah(neigh->ah);
266
267                 ipoib_neigh_free(neigh);
268         }
269
270         spin_unlock_irqrestore(&priv->lock, flags);
271
272         if (path->ah)
273                 ipoib_put_ah(path->ah);
274
275         kfree(path);
276 }
277
278 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
279
280 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
281 {
282         struct ipoib_path_iter *iter;
283
284         iter = kmalloc(sizeof *iter, GFP_KERNEL);
285         if (!iter)
286                 return NULL;
287
288         iter->dev = dev;
289         memset(iter->path.pathrec.dgid.raw, 0, 16);
290
291         if (ipoib_path_iter_next(iter)) {
292                 kfree(iter);
293                 return NULL;
294         }
295
296         return iter;
297 }
298
299 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
300 {
301         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
302         struct rb_node *n;
303         struct ipoib_path *path;
304         int ret = 1;
305
306         spin_lock_irq(&priv->lock);
307
308         n = rb_first(&priv->path_tree);
309
310         while (n) {
311                 path = rb_entry(n, struct ipoib_path, rb_node);
312
313                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
314                            sizeof (union ib_gid)) < 0) {
315                         iter->path = *path;
316                         ret = 0;
317                         break;
318                 }
319
320                 n = rb_next(n);
321         }
322
323         spin_unlock_irq(&priv->lock);
324
325         return ret;
326 }
327
328 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
329                           struct ipoib_path *path)
330 {
331         *path = iter->path;
332 }
333
334 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
335
336 void ipoib_flush_paths(struct net_device *dev)
337 {
338         struct ipoib_dev_priv *priv = netdev_priv(dev);
339         struct ipoib_path *path, *tp;
340         LIST_HEAD(remove_list);
341
342         spin_lock_irq(&priv->tx_lock);
343         spin_lock(&priv->lock);
344
345         list_splice(&priv->path_list, &remove_list);
346         INIT_LIST_HEAD(&priv->path_list);
347
348         list_for_each_entry(path, &remove_list, list)
349                 rb_erase(&path->rb_node, &priv->path_tree);
350
351         list_for_each_entry_safe(path, tp, &remove_list, list) {
352                 if (path->query)
353                         ib_sa_cancel_query(path->query_id, path->query);
354                 spin_unlock(&priv->lock);
355                 spin_unlock_irq(&priv->tx_lock);
356                 wait_for_completion(&path->done);
357                 path_free(dev, path);
358                 spin_lock_irq(&priv->tx_lock);
359                 spin_lock(&priv->lock);
360         }
361         spin_unlock(&priv->lock);
362         spin_unlock_irq(&priv->tx_lock);
363 }
364
365 static void path_rec_completion(int status,
366                                 struct ib_sa_path_rec *pathrec,
367                                 void *path_ptr)
368 {
369         struct ipoib_path *path = path_ptr;
370         struct net_device *dev = path->dev;
371         struct ipoib_dev_priv *priv = netdev_priv(dev);
372         struct ipoib_ah *ah = NULL;
373         struct ipoib_neigh *neigh;
374         struct sk_buff_head skqueue;
375         struct sk_buff *skb;
376         unsigned long flags;
377
378         if (pathrec)
379                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
380                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
381         else
382                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
383                           status, IPOIB_GID_ARG(path->pathrec.dgid));
384
385         skb_queue_head_init(&skqueue);
386
387         if (!status) {
388                 struct ib_ah_attr av = {
389                         .dlid          = be16_to_cpu(pathrec->dlid),
390                         .sl            = pathrec->sl,
391                         .port_num      = priv->port,
392                         .static_rate   = pathrec->rate
393                 };
394
395                 ah = ipoib_create_ah(dev, priv->pd, &av);
396         }
397
398         spin_lock_irqsave(&priv->lock, flags);
399
400         path->ah = ah;
401
402         if (ah) {
403                 path->pathrec = *pathrec;
404
405                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
406                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
407
408                 while ((skb = __skb_dequeue(&path->queue)))
409                         __skb_queue_tail(&skqueue, skb);
410
411                 list_for_each_entry(neigh, &path->neigh_list, list) {
412                         kref_get(&path->ah->ref);
413                         neigh->ah = path->ah;
414                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
415                                sizeof(union ib_gid));
416
417                         while ((skb = __skb_dequeue(&neigh->queue)))
418                                 __skb_queue_tail(&skqueue, skb);
419                 }
420         }
421
422         path->query = NULL;
423         complete(&path->done);
424
425         spin_unlock_irqrestore(&priv->lock, flags);
426
427         while ((skb = __skb_dequeue(&skqueue))) {
428                 skb->dev = dev;
429                 if (dev_queue_xmit(skb))
430                         ipoib_warn(priv, "dev_queue_xmit failed "
431                                    "to requeue packet\n");
432         }
433 }
434
435 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
436 {
437         struct ipoib_dev_priv *priv = netdev_priv(dev);
438         struct ipoib_path *path;
439
440         path = kzalloc(sizeof *path, GFP_ATOMIC);
441         if (!path)
442                 return NULL;
443
444         path->dev = dev;
445
446         skb_queue_head_init(&path->queue);
447
448         INIT_LIST_HEAD(&path->neigh_list);
449
450         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
451         path->pathrec.sgid      = priv->local_gid;
452         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
453         path->pathrec.numb_path = 1;
454
455         return path;
456 }
457
458 static int path_rec_start(struct net_device *dev,
459                           struct ipoib_path *path)
460 {
461         struct ipoib_dev_priv *priv = netdev_priv(dev);
462
463         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
464                   IPOIB_GID_ARG(path->pathrec.dgid));
465
466         init_completion(&path->done);
467
468         path->query_id =
469                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
470                                    &path->pathrec,
471                                    IB_SA_PATH_REC_DGID          |
472                                    IB_SA_PATH_REC_SGID          |
473                                    IB_SA_PATH_REC_NUMB_PATH     |
474                                    IB_SA_PATH_REC_PKEY,
475                                    1000, GFP_ATOMIC,
476                                    path_rec_completion,
477                                    path, &path->query);
478         if (path->query_id < 0) {
479                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
480                 path->query = NULL;
481                 return path->query_id;
482         }
483
484         return 0;
485 }
486
487 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
488 {
489         struct ipoib_dev_priv *priv = netdev_priv(dev);
490         struct ipoib_path *path;
491         struct ipoib_neigh *neigh;
492
493         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
494         if (!neigh) {
495                 ++priv->stats.tx_dropped;
496                 dev_kfree_skb_any(skb);
497                 return;
498         }
499
500         skb_queue_head_init(&neigh->queue);
501
502         /*
503          * We can only be called from ipoib_start_xmit, so we're
504          * inside tx_lock -- no need to save/restore flags.
505          */
506         spin_lock(&priv->lock);
507
508         path = __path_find(dev, skb->dst->neighbour->ha + 4);
509         if (!path) {
510                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
511                 if (!path)
512                         goto err_path;
513
514                 __path_add(dev, path);
515         }
516
517         list_add_tail(&neigh->list, &path->neigh_list);
518
519         if (path->ah) {
520                 kref_get(&path->ah->ref);
521                 neigh->ah = path->ah;
522                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
523                        sizeof(union ib_gid));
524
525                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
526         } else {
527                 neigh->ah  = NULL;
528                 __skb_queue_tail(&neigh->queue, skb);
529
530                 if (!path->query && path_rec_start(dev, path))
531                         goto err_list;
532         }
533
534         spin_unlock(&priv->lock);
535         return;
536
537 err_list:
538         list_del(&neigh->list);
539
540 err_path:
541         ipoib_neigh_free(neigh);
542         ++priv->stats.tx_dropped;
543         dev_kfree_skb_any(skb);
544
545         spin_unlock(&priv->lock);
546 }
547
548 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
549 {
550         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
551
552         /* Look up path record for unicasts */
553         if (skb->dst->neighbour->ha[4] != 0xff) {
554                 neigh_add_path(skb, dev);
555                 return;
556         }
557
558         /* Add in the P_Key for multicasts */
559         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
560         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
561         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
562 }
563
564 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
565                              struct ipoib_pseudoheader *phdr)
566 {
567         struct ipoib_dev_priv *priv = netdev_priv(dev);
568         struct ipoib_path *path;
569
570         /*
571          * We can only be called from ipoib_start_xmit, so we're
572          * inside tx_lock -- no need to save/restore flags.
573          */
574         spin_lock(&priv->lock);
575
576         path = __path_find(dev, phdr->hwaddr + 4);
577         if (!path) {
578                 path = path_rec_create(dev, phdr->hwaddr + 4);
579                 if (path) {
580                         /* put pseudoheader back on for next time */
581                         skb_push(skb, sizeof *phdr);
582                         __skb_queue_tail(&path->queue, skb);
583
584                         if (path_rec_start(dev, path)) {
585                                 spin_unlock(&priv->lock);
586                                 path_free(dev, path);
587                                 return;
588                         } else
589                                 __path_add(dev, path);
590                 } else {
591                         ++priv->stats.tx_dropped;
592                         dev_kfree_skb_any(skb);
593                 }
594
595                 spin_unlock(&priv->lock);
596                 return;
597         }
598
599         if (path->ah) {
600                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
601                           be16_to_cpu(path->pathrec.dlid));
602
603                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
604         } else if ((path->query || !path_rec_start(dev, path)) &&
605                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
606                 /* put pseudoheader back on for next time */
607                 skb_push(skb, sizeof *phdr);
608                 __skb_queue_tail(&path->queue, skb);
609         } else {
610                 ++priv->stats.tx_dropped;
611                 dev_kfree_skb_any(skb);
612         }
613
614         spin_unlock(&priv->lock);
615 }
616
617 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
618 {
619         struct ipoib_dev_priv *priv = netdev_priv(dev);
620         struct ipoib_neigh *neigh;
621         unsigned long flags;
622
623         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
624                 return NETDEV_TX_LOCKED;
625
626         /*
627          * Check if our queue is stopped.  Since we have the LLTX bit
628          * set, we can't rely on netif_stop_queue() preventing our
629          * xmit function from being called with a full queue.
630          */
631         if (unlikely(netif_queue_stopped(dev))) {
632                 spin_unlock_irqrestore(&priv->tx_lock, flags);
633                 return NETDEV_TX_BUSY;
634         }
635
636         if (likely(skb->dst && skb->dst->neighbour)) {
637                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
638                         ipoib_path_lookup(skb, dev);
639                         goto out;
640                 }
641
642                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
643
644                 if (likely(neigh->ah)) {
645                         if (unlikely(memcmp(&neigh->dgid.raw,
646                                             skb->dst->neighbour->ha + 4,
647                                             sizeof(union ib_gid)))) {
648                                 spin_lock(&priv->lock);
649                                 /*
650                                  * It's safe to call ipoib_put_ah() inside
651                                  * priv->lock here, because we know that
652                                  * path->ah will always hold one more reference,
653                                  * so ipoib_put_ah() will never do more than
654                                  * decrement the ref count.
655                                  */
656                                 ipoib_put_ah(neigh->ah);
657                                 list_del(&neigh->list);
658                                 ipoib_neigh_free(neigh);
659                                 spin_unlock(&priv->lock);
660                                 ipoib_path_lookup(skb, dev);
661                                 goto out;
662                         }
663
664                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
665                         goto out;
666                 }
667
668                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
669                         spin_lock(&priv->lock);
670                         __skb_queue_tail(&neigh->queue, skb);
671                         spin_unlock(&priv->lock);
672                 } else {
673                         ++priv->stats.tx_dropped;
674                         dev_kfree_skb_any(skb);
675                 }
676         } else {
677                 struct ipoib_pseudoheader *phdr =
678                         (struct ipoib_pseudoheader *) skb->data;
679                 skb_pull(skb, sizeof *phdr);
680
681                 if (phdr->hwaddr[4] == 0xff) {
682                         /* Add in the P_Key for multicast*/
683                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
684                         phdr->hwaddr[9] = priv->pkey & 0xff;
685
686                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
687                 } else {
688                         /* unicast GID -- should be ARP or RARP reply */
689
690                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
691                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
692                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
693                                            IPOIB_GID_FMT "\n",
694                                            skb->dst ? "neigh" : "dst",
695                                            be16_to_cpup((__be16 *) skb->data),
696                                            IPOIB_QPN(phdr->hwaddr),
697                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
698                                 dev_kfree_skb_any(skb);
699                                 ++priv->stats.tx_dropped;
700                                 goto out;
701                         }
702
703                         unicast_arp_send(skb, dev, phdr);
704                 }
705         }
706
707 out:
708         spin_unlock_irqrestore(&priv->tx_lock, flags);
709
710         return NETDEV_TX_OK;
711 }
712
713 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
714 {
715         struct ipoib_dev_priv *priv = netdev_priv(dev);
716
717         return &priv->stats;
718 }
719
720 static void ipoib_timeout(struct net_device *dev)
721 {
722         struct ipoib_dev_priv *priv = netdev_priv(dev);
723
724         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
725                    jiffies_to_msecs(jiffies - dev->trans_start));
726         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
727                    netif_queue_stopped(dev),
728                    priv->tx_head, priv->tx_tail);
729         /* XXX reset QP, etc. */
730 }
731
732 static int ipoib_hard_header(struct sk_buff *skb,
733                              struct net_device *dev,
734                              unsigned short type,
735                              void *daddr, void *saddr, unsigned len)
736 {
737         struct ipoib_header *header;
738
739         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
740
741         header->proto = htons(type);
742         header->reserved = 0;
743
744         /*
745          * If we don't have a neighbour structure, stuff the
746          * destination address onto the front of the skb so we can
747          * figure out where to send the packet later.
748          */
749         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
750                 struct ipoib_pseudoheader *phdr =
751                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
752                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
753         }
754
755         return 0;
756 }
757
758 static void ipoib_set_mcast_list(struct net_device *dev)
759 {
760         struct ipoib_dev_priv *priv = netdev_priv(dev);
761
762         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
763                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
764                 return;
765         }
766
767         queue_work(ipoib_workqueue, &priv->restart_task);
768 }
769
770 static void ipoib_neigh_destructor(struct neighbour *n)
771 {
772         struct ipoib_neigh *neigh;
773         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
774         unsigned long flags;
775         struct ipoib_ah *ah = NULL;
776
777         ipoib_dbg(priv,
778                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
779                   IPOIB_QPN(n->ha),
780                   IPOIB_GID_RAW_ARG(n->ha + 4));
781
782         spin_lock_irqsave(&priv->lock, flags);
783
784         neigh = *to_ipoib_neigh(n);
785         if (neigh) {
786                 if (neigh->ah)
787                         ah = neigh->ah;
788                 list_del(&neigh->list);
789                 ipoib_neigh_free(neigh);
790         }
791
792         spin_unlock_irqrestore(&priv->lock, flags);
793
794         if (ah)
795                 ipoib_put_ah(ah);
796 }
797
798 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
799 {
800         struct ipoib_neigh *neigh;
801
802         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
803         if (!neigh)
804                 return NULL;
805
806         neigh->neighbour = neighbour;
807         *to_ipoib_neigh(neighbour) = neigh;
808
809         return neigh;
810 }
811
812 void ipoib_neigh_free(struct ipoib_neigh *neigh)
813 {
814         *to_ipoib_neigh(neigh->neighbour) = NULL;
815         kfree(neigh);
816 }
817
818 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
819 {
820         parms->neigh_destructor = ipoib_neigh_destructor;
821
822         return 0;
823 }
824
825 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
826 {
827         struct ipoib_dev_priv *priv = netdev_priv(dev);
828
829         /* Allocate RX/TX "rings" to hold queued skbs */
830         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
831                                 GFP_KERNEL);
832         if (!priv->rx_ring) {
833                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
834                        ca->name, ipoib_recvq_size);
835                 goto out;
836         }
837
838         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
839                                 GFP_KERNEL);
840         if (!priv->tx_ring) {
841                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
842                        ca->name, ipoib_sendq_size);
843                 goto out_rx_ring_cleanup;
844         }
845
846         /* priv->tx_head & tx_tail are already 0 */
847
848         if (ipoib_ib_dev_init(dev, ca, port))
849                 goto out_tx_ring_cleanup;
850
851         return 0;
852
853 out_tx_ring_cleanup:
854         kfree(priv->tx_ring);
855
856 out_rx_ring_cleanup:
857         kfree(priv->rx_ring);
858
859 out:
860         return -ENOMEM;
861 }
862
863 void ipoib_dev_cleanup(struct net_device *dev)
864 {
865         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
866
867         ipoib_delete_debug_files(dev);
868
869         /* Delete any child interfaces first */
870         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
871                 unregister_netdev(cpriv->dev);
872                 ipoib_dev_cleanup(cpriv->dev);
873                 free_netdev(cpriv->dev);
874         }
875
876         ipoib_ib_dev_cleanup(dev);
877
878         kfree(priv->rx_ring);
879         kfree(priv->tx_ring);
880
881         priv->rx_ring = NULL;
882         priv->tx_ring = NULL;
883 }
884
885 static void ipoib_setup(struct net_device *dev)
886 {
887         struct ipoib_dev_priv *priv = netdev_priv(dev);
888
889         dev->open                = ipoib_open;
890         dev->stop                = ipoib_stop;
891         dev->change_mtu          = ipoib_change_mtu;
892         dev->hard_start_xmit     = ipoib_start_xmit;
893         dev->get_stats           = ipoib_get_stats;
894         dev->tx_timeout          = ipoib_timeout;
895         dev->hard_header         = ipoib_hard_header;
896         dev->set_multicast_list  = ipoib_set_mcast_list;
897         dev->neigh_setup         = ipoib_neigh_setup_dev;
898
899         dev->watchdog_timeo      = HZ;
900
901         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
902
903         /*
904          * We add in INFINIBAND_ALEN to allow for the destination
905          * address "pseudoheader" for skbs without neighbour struct.
906          */
907         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
908         dev->addr_len            = INFINIBAND_ALEN;
909         dev->type                = ARPHRD_INFINIBAND;
910         dev->tx_queue_len        = ipoib_sendq_size * 2;
911         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
912
913         /* MTU will be reset when mcast join happens */
914         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
915         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
916
917         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
918
919         netif_carrier_off(dev);
920
921         SET_MODULE_OWNER(dev);
922
923         priv->dev = dev;
924
925         spin_lock_init(&priv->lock);
926         spin_lock_init(&priv->tx_lock);
927
928         mutex_init(&priv->mcast_mutex);
929         mutex_init(&priv->vlan_mutex);
930
931         INIT_LIST_HEAD(&priv->path_list);
932         INIT_LIST_HEAD(&priv->child_intfs);
933         INIT_LIST_HEAD(&priv->dead_ahs);
934         INIT_LIST_HEAD(&priv->multicast_list);
935
936         INIT_WORK(&priv->pkey_task,    ipoib_pkey_poll,          priv->dev);
937         INIT_WORK(&priv->mcast_task,   ipoib_mcast_join_task,    priv->dev);
938         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush,       priv->dev);
939         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
940         INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah,            priv->dev);
941 }
942
943 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
944 {
945         struct net_device *dev;
946
947         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
948                            ipoib_setup);
949         if (!dev)
950                 return NULL;
951
952         return netdev_priv(dev);
953 }
954
955 static ssize_t show_pkey(struct class_device *cdev, char *buf)
956 {
957         struct ipoib_dev_priv *priv =
958                 netdev_priv(container_of(cdev, struct net_device, class_dev));
959
960         return sprintf(buf, "0x%04x\n", priv->pkey);
961 }
962 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
963
964 static ssize_t create_child(struct class_device *cdev,
965                             const char *buf, size_t count)
966 {
967         int pkey;
968         int ret;
969
970         if (sscanf(buf, "%i", &pkey) != 1)
971                 return -EINVAL;
972
973         if (pkey < 0 || pkey > 0xffff)
974                 return -EINVAL;
975
976         /*
977          * Set the full membership bit, so that we join the right
978          * broadcast group, etc.
979          */
980         pkey |= 0x8000;
981
982         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
983                              pkey);
984
985         return ret ? ret : count;
986 }
987 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
988
989 static ssize_t delete_child(struct class_device *cdev,
990                             const char *buf, size_t count)
991 {
992         int pkey;
993         int ret;
994
995         if (sscanf(buf, "%i", &pkey) != 1)
996                 return -EINVAL;
997
998         if (pkey < 0 || pkey > 0xffff)
999                 return -EINVAL;
1000
1001         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
1002                                 pkey);
1003
1004         return ret ? ret : count;
1005
1006 }
1007 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1008
1009 int ipoib_add_pkey_attr(struct net_device *dev)
1010 {
1011         return class_device_create_file(&dev->class_dev,
1012                                         &class_device_attr_pkey);
1013 }
1014
1015 static struct net_device *ipoib_add_port(const char *format,
1016                                          struct ib_device *hca, u8 port)
1017 {
1018         struct ipoib_dev_priv *priv;
1019         int result = -ENOMEM;
1020
1021         priv = ipoib_intf_alloc(format);
1022         if (!priv)
1023                 goto alloc_mem_failed;
1024
1025         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1026
1027         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1028         if (result) {
1029                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1030                        hca->name, port, result);
1031                 goto alloc_mem_failed;
1032         }
1033
1034         /*
1035          * Set the full membership bit, so that we join the right
1036          * broadcast group, etc.
1037          */
1038         priv->pkey |= 0x8000;
1039
1040         priv->dev->broadcast[8] = priv->pkey >> 8;
1041         priv->dev->broadcast[9] = priv->pkey & 0xff;
1042
1043         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1044         if (result) {
1045                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1046                        hca->name, port, result);
1047                 goto alloc_mem_failed;
1048         } else
1049                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1050
1051
1052         result = ipoib_dev_init(priv->dev, hca, port);
1053         if (result < 0) {
1054                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1055                        hca->name, port, result);
1056                 goto device_init_failed;
1057         }
1058
1059         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1060                               priv->ca, ipoib_event);
1061         result = ib_register_event_handler(&priv->event_handler);
1062         if (result < 0) {
1063                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1064                        "port %d (ret = %d)\n",
1065                        hca->name, port, result);
1066                 goto event_failed;
1067         }
1068
1069         result = register_netdev(priv->dev);
1070         if (result) {
1071                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1072                        hca->name, port, result);
1073                 goto register_failed;
1074         }
1075
1076         ipoib_create_debug_files(priv->dev);
1077
1078         if (ipoib_add_pkey_attr(priv->dev))
1079                 goto sysfs_failed;
1080         if (class_device_create_file(&priv->dev->class_dev,
1081                                      &class_device_attr_create_child))
1082                 goto sysfs_failed;
1083         if (class_device_create_file(&priv->dev->class_dev,
1084                                      &class_device_attr_delete_child))
1085                 goto sysfs_failed;
1086
1087         return priv->dev;
1088
1089 sysfs_failed:
1090         ipoib_delete_debug_files(priv->dev);
1091         unregister_netdev(priv->dev);
1092
1093 register_failed:
1094         ib_unregister_event_handler(&priv->event_handler);
1095         flush_scheduled_work();
1096
1097 event_failed:
1098         ipoib_dev_cleanup(priv->dev);
1099
1100 device_init_failed:
1101         free_netdev(priv->dev);
1102
1103 alloc_mem_failed:
1104         return ERR_PTR(result);
1105 }
1106
1107 static void ipoib_add_one(struct ib_device *device)
1108 {
1109         struct list_head *dev_list;
1110         struct net_device *dev;
1111         struct ipoib_dev_priv *priv;
1112         int s, e, p;
1113
1114         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1115                 return;
1116
1117         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1118         if (!dev_list)
1119                 return;
1120
1121         INIT_LIST_HEAD(dev_list);
1122
1123         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1124                 s = 0;
1125                 e = 0;
1126         } else {
1127                 s = 1;
1128                 e = device->phys_port_cnt;
1129         }
1130
1131         for (p = s; p <= e; ++p) {
1132                 dev = ipoib_add_port("ib%d", device, p);
1133                 if (!IS_ERR(dev)) {
1134                         priv = netdev_priv(dev);
1135                         list_add_tail(&priv->list, dev_list);
1136                 }
1137         }
1138
1139         ib_set_client_data(device, &ipoib_client, dev_list);
1140 }
1141
1142 static void ipoib_remove_one(struct ib_device *device)
1143 {
1144         struct ipoib_dev_priv *priv, *tmp;
1145         struct list_head *dev_list;
1146
1147         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1148                 return;
1149
1150         dev_list = ib_get_client_data(device, &ipoib_client);
1151
1152         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1153                 ib_unregister_event_handler(&priv->event_handler);
1154                 flush_scheduled_work();
1155
1156                 unregister_netdev(priv->dev);
1157                 ipoib_dev_cleanup(priv->dev);
1158                 free_netdev(priv->dev);
1159         }
1160
1161         kfree(dev_list);
1162 }
1163
1164 static int __init ipoib_init_module(void)
1165 {
1166         int ret;
1167
1168         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1169         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1170         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1171
1172         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1173         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1174         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1175
1176         ret = ipoib_register_debugfs();
1177         if (ret)
1178                 return ret;
1179
1180         /*
1181          * We create our own workqueue mainly because we want to be
1182          * able to flush it when devices are being removed.  We can't
1183          * use schedule_work()/flush_scheduled_work() because both
1184          * unregister_netdev() and linkwatch_event take the rtnl lock,
1185          * so flush_scheduled_work() can deadlock during device
1186          * removal.
1187          */
1188         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1189         if (!ipoib_workqueue) {
1190                 ret = -ENOMEM;
1191                 goto err_fs;
1192         }
1193
1194         ib_sa_register_client(&ipoib_sa_client);
1195
1196         ret = ib_register_client(&ipoib_client);
1197         if (ret)
1198                 goto err_sa;
1199
1200         return 0;
1201
1202 err_sa:
1203         ib_sa_unregister_client(&ipoib_sa_client);
1204         destroy_workqueue(ipoib_workqueue);
1205
1206 err_fs:
1207         ipoib_unregister_debugfs();
1208
1209         return ret;
1210 }
1211
1212 static void __exit ipoib_cleanup_module(void)
1213 {
1214         ib_unregister_client(&ipoib_client);
1215         ib_sa_unregister_client(&ipoib_sa_client);
1216         ipoib_unregister_debugfs();
1217         destroy_workqueue(ipoib_workqueue);
1218 }
1219
1220 module_init(ipoib_init_module);
1221 module_exit(ipoib_cleanup_module);