IPoIB: Connected mode experimental support
[safe/jmp/linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_main.c
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $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         set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
102
103         if (ipoib_pkey_dev_delay_open(dev))
104                 return 0;
105
106         if (ipoib_ib_dev_open(dev))
107                 return -EINVAL;
108
109         if (ipoib_ib_dev_up(dev)) {
110                 ipoib_ib_dev_stop(dev);
111                 return -EINVAL;
112         }
113
114         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
115                 struct ipoib_dev_priv *cpriv;
116
117                 /* Bring up any child interfaces too */
118                 mutex_lock(&priv->vlan_mutex);
119                 list_for_each_entry(cpriv, &priv->child_intfs, list) {
120                         int flags;
121
122                         flags = cpriv->dev->flags;
123                         if (flags & IFF_UP)
124                                 continue;
125
126                         dev_change_flags(cpriv->dev, flags | IFF_UP);
127                 }
128                 mutex_unlock(&priv->vlan_mutex);
129         }
130
131         netif_start_queue(dev);
132
133         return 0;
134 }
135
136 static int ipoib_stop(struct net_device *dev)
137 {
138         struct ipoib_dev_priv *priv = netdev_priv(dev);
139
140         ipoib_dbg(priv, "stopping interface\n");
141
142         clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
143
144         netif_stop_queue(dev);
145
146         clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
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         /* dev->mtu > 2K ==> connected mode */
182         if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
183                 if (new_mtu > priv->mcast_mtu)
184                         ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
185                                    priv->mcast_mtu);
186                 dev->mtu = new_mtu;
187                 return 0;
188         }
189
190         if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
191                 return -EINVAL;
192         }
193
194         priv->admin_mtu = new_mtu;
195
196         dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
197
198         return 0;
199 }
200
201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
202 {
203         struct ipoib_dev_priv *priv = netdev_priv(dev);
204         struct rb_node *n = priv->path_tree.rb_node;
205         struct ipoib_path *path;
206         int ret;
207
208         while (n) {
209                 path = rb_entry(n, struct ipoib_path, rb_node);
210
211                 ret = memcmp(gid, path->pathrec.dgid.raw,
212                              sizeof (union ib_gid));
213
214                 if (ret < 0)
215                         n = n->rb_left;
216                 else if (ret > 0)
217                         n = n->rb_right;
218                 else
219                         return path;
220         }
221
222         return NULL;
223 }
224
225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
226 {
227         struct ipoib_dev_priv *priv = netdev_priv(dev);
228         struct rb_node **n = &priv->path_tree.rb_node;
229         struct rb_node *pn = NULL;
230         struct ipoib_path *tpath;
231         int ret;
232
233         while (*n) {
234                 pn = *n;
235                 tpath = rb_entry(pn, struct ipoib_path, rb_node);
236
237                 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
238                              sizeof (union ib_gid));
239                 if (ret < 0)
240                         n = &pn->rb_left;
241                 else if (ret > 0)
242                         n = &pn->rb_right;
243                 else
244                         return -EEXIST;
245         }
246
247         rb_link_node(&path->rb_node, pn, n);
248         rb_insert_color(&path->rb_node, &priv->path_tree);
249
250         list_add_tail(&path->list, &priv->path_list);
251
252         return 0;
253 }
254
255 static void path_free(struct net_device *dev, struct ipoib_path *path)
256 {
257         struct ipoib_dev_priv *priv = netdev_priv(dev);
258         struct ipoib_neigh *neigh, *tn;
259         struct sk_buff *skb;
260         unsigned long flags;
261
262         while ((skb = __skb_dequeue(&path->queue)))
263                 dev_kfree_skb_irq(skb);
264
265         spin_lock_irqsave(&priv->lock, flags);
266
267         list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
268                 /*
269                  * It's safe to call ipoib_put_ah() inside priv->lock
270                  * here, because we know that path->ah will always
271                  * hold one more reference, so ipoib_put_ah() will
272                  * never do more than decrement the ref count.
273                  */
274                 if (neigh->ah)
275                         ipoib_put_ah(neigh->ah);
276
277                 ipoib_neigh_free(dev, neigh);
278         }
279
280         spin_unlock_irqrestore(&priv->lock, flags);
281
282         if (path->ah)
283                 ipoib_put_ah(path->ah);
284
285         kfree(path);
286 }
287
288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
289
290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
291 {
292         struct ipoib_path_iter *iter;
293
294         iter = kmalloc(sizeof *iter, GFP_KERNEL);
295         if (!iter)
296                 return NULL;
297
298         iter->dev = dev;
299         memset(iter->path.pathrec.dgid.raw, 0, 16);
300
301         if (ipoib_path_iter_next(iter)) {
302                 kfree(iter);
303                 return NULL;
304         }
305
306         return iter;
307 }
308
309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
310 {
311         struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
312         struct rb_node *n;
313         struct ipoib_path *path;
314         int ret = 1;
315
316         spin_lock_irq(&priv->lock);
317
318         n = rb_first(&priv->path_tree);
319
320         while (n) {
321                 path = rb_entry(n, struct ipoib_path, rb_node);
322
323                 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
324                            sizeof (union ib_gid)) < 0) {
325                         iter->path = *path;
326                         ret = 0;
327                         break;
328                 }
329
330                 n = rb_next(n);
331         }
332
333         spin_unlock_irq(&priv->lock);
334
335         return ret;
336 }
337
338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
339                           struct ipoib_path *path)
340 {
341         *path = iter->path;
342 }
343
344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
345
346 void ipoib_flush_paths(struct net_device *dev)
347 {
348         struct ipoib_dev_priv *priv = netdev_priv(dev);
349         struct ipoib_path *path, *tp;
350         LIST_HEAD(remove_list);
351
352         spin_lock_irq(&priv->tx_lock);
353         spin_lock(&priv->lock);
354
355         list_splice(&priv->path_list, &remove_list);
356         INIT_LIST_HEAD(&priv->path_list);
357
358         list_for_each_entry(path, &remove_list, list)
359                 rb_erase(&path->rb_node, &priv->path_tree);
360
361         list_for_each_entry_safe(path, tp, &remove_list, list) {
362                 if (path->query)
363                         ib_sa_cancel_query(path->query_id, path->query);
364                 spin_unlock(&priv->lock);
365                 spin_unlock_irq(&priv->tx_lock);
366                 wait_for_completion(&path->done);
367                 path_free(dev, path);
368                 spin_lock_irq(&priv->tx_lock);
369                 spin_lock(&priv->lock);
370         }
371         spin_unlock(&priv->lock);
372         spin_unlock_irq(&priv->tx_lock);
373 }
374
375 static void path_rec_completion(int status,
376                                 struct ib_sa_path_rec *pathrec,
377                                 void *path_ptr)
378 {
379         struct ipoib_path *path = path_ptr;
380         struct net_device *dev = path->dev;
381         struct ipoib_dev_priv *priv = netdev_priv(dev);
382         struct ipoib_ah *ah = NULL;
383         struct ipoib_neigh *neigh;
384         struct sk_buff_head skqueue;
385         struct sk_buff *skb;
386         unsigned long flags;
387
388         if (pathrec)
389                 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
390                           be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
391         else
392                 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
393                           status, IPOIB_GID_ARG(path->pathrec.dgid));
394
395         skb_queue_head_init(&skqueue);
396
397         if (!status) {
398                 struct ib_ah_attr av = {
399                         .dlid          = be16_to_cpu(pathrec->dlid),
400                         .sl            = pathrec->sl,
401                         .port_num      = priv->port,
402                         .static_rate   = pathrec->rate
403                 };
404
405                 ah = ipoib_create_ah(dev, priv->pd, &av);
406         }
407
408         spin_lock_irqsave(&priv->lock, flags);
409
410         path->ah = ah;
411
412         if (ah) {
413                 path->pathrec = *pathrec;
414
415                 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
416                           ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
417
418                 while ((skb = __skb_dequeue(&path->queue)))
419                         __skb_queue_tail(&skqueue, skb);
420
421                 list_for_each_entry(neigh, &path->neigh_list, list) {
422                         kref_get(&path->ah->ref);
423                         neigh->ah = path->ah;
424                         memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
425                                sizeof(union ib_gid));
426
427                         if (ipoib_cm_enabled(dev, neigh->neighbour)) {
428                                 if (!ipoib_cm_get(neigh))
429                                         ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
430                                                                                path,
431                                                                                neigh));
432                                 if (!ipoib_cm_get(neigh)) {
433                                         list_del(&neigh->list);
434                                         if (neigh->ah)
435                                                 ipoib_put_ah(neigh->ah);
436                                         ipoib_neigh_free(dev, neigh);
437                                         continue;
438                                 }
439                         }
440
441                         while ((skb = __skb_dequeue(&neigh->queue)))
442                                 __skb_queue_tail(&skqueue, skb);
443                 }
444         }
445
446         path->query = NULL;
447         complete(&path->done);
448
449         spin_unlock_irqrestore(&priv->lock, flags);
450
451         while ((skb = __skb_dequeue(&skqueue))) {
452                 skb->dev = dev;
453                 if (dev_queue_xmit(skb))
454                         ipoib_warn(priv, "dev_queue_xmit failed "
455                                    "to requeue packet\n");
456         }
457 }
458
459 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
460 {
461         struct ipoib_dev_priv *priv = netdev_priv(dev);
462         struct ipoib_path *path;
463
464         path = kzalloc(sizeof *path, GFP_ATOMIC);
465         if (!path)
466                 return NULL;
467
468         path->dev = dev;
469
470         skb_queue_head_init(&path->queue);
471
472         INIT_LIST_HEAD(&path->neigh_list);
473
474         memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
475         path->pathrec.sgid      = priv->local_gid;
476         path->pathrec.pkey      = cpu_to_be16(priv->pkey);
477         path->pathrec.numb_path = 1;
478
479         return path;
480 }
481
482 static int path_rec_start(struct net_device *dev,
483                           struct ipoib_path *path)
484 {
485         struct ipoib_dev_priv *priv = netdev_priv(dev);
486
487         ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
488                   IPOIB_GID_ARG(path->pathrec.dgid));
489
490         init_completion(&path->done);
491
492         path->query_id =
493                 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
494                                    &path->pathrec,
495                                    IB_SA_PATH_REC_DGID          |
496                                    IB_SA_PATH_REC_SGID          |
497                                    IB_SA_PATH_REC_NUMB_PATH     |
498                                    IB_SA_PATH_REC_PKEY,
499                                    1000, GFP_ATOMIC,
500                                    path_rec_completion,
501                                    path, &path->query);
502         if (path->query_id < 0) {
503                 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
504                 path->query = NULL;
505                 return path->query_id;
506         }
507
508         return 0;
509 }
510
511 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
512 {
513         struct ipoib_dev_priv *priv = netdev_priv(dev);
514         struct ipoib_path *path;
515         struct ipoib_neigh *neigh;
516
517         neigh = ipoib_neigh_alloc(skb->dst->neighbour);
518         if (!neigh) {
519                 ++priv->stats.tx_dropped;
520                 dev_kfree_skb_any(skb);
521                 return;
522         }
523
524         /*
525          * We can only be called from ipoib_start_xmit, so we're
526          * inside tx_lock -- no need to save/restore flags.
527          */
528         spin_lock(&priv->lock);
529
530         path = __path_find(dev, skb->dst->neighbour->ha + 4);
531         if (!path) {
532                 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
533                 if (!path)
534                         goto err_path;
535
536                 __path_add(dev, path);
537         }
538
539         list_add_tail(&neigh->list, &path->neigh_list);
540
541         if (path->ah) {
542                 kref_get(&path->ah->ref);
543                 neigh->ah = path->ah;
544                 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
545                        sizeof(union ib_gid));
546
547                 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
548                         if (!ipoib_cm_get(neigh))
549                                 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
550                         if (!ipoib_cm_get(neigh)) {
551                                 list_del(&neigh->list);
552                                 if (neigh->ah)
553                                         ipoib_put_ah(neigh->ah);
554                                 ipoib_neigh_free(dev, neigh);
555                                 goto err_drop;
556                         }
557                         if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
558                                 __skb_queue_tail(&neigh->queue, skb);
559                         else {
560                                 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
561                                            skb_queue_len(&neigh->queue));
562                                 goto err_drop;
563                         }
564                 } else
565                         ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
566         } else {
567                 neigh->ah  = NULL;
568
569                 if (!path->query && path_rec_start(dev, path))
570                         goto err_list;
571
572                 __skb_queue_tail(&neigh->queue, skb);
573         }
574
575         spin_unlock(&priv->lock);
576         return;
577
578 err_list:
579         list_del(&neigh->list);
580
581 err_path:
582         ipoib_neigh_free(dev, neigh);
583 err_drop:
584         ++priv->stats.tx_dropped;
585         dev_kfree_skb_any(skb);
586
587         spin_unlock(&priv->lock);
588 }
589
590 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
591 {
592         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
593
594         /* Look up path record for unicasts */
595         if (skb->dst->neighbour->ha[4] != 0xff) {
596                 neigh_add_path(skb, dev);
597                 return;
598         }
599
600         /* Add in the P_Key for multicasts */
601         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
602         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
603         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
604 }
605
606 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
607                              struct ipoib_pseudoheader *phdr)
608 {
609         struct ipoib_dev_priv *priv = netdev_priv(dev);
610         struct ipoib_path *path;
611
612         /*
613          * We can only be called from ipoib_start_xmit, so we're
614          * inside tx_lock -- no need to save/restore flags.
615          */
616         spin_lock(&priv->lock);
617
618         path = __path_find(dev, phdr->hwaddr + 4);
619         if (!path) {
620                 path = path_rec_create(dev, phdr->hwaddr + 4);
621                 if (path) {
622                         /* put pseudoheader back on for next time */
623                         skb_push(skb, sizeof *phdr);
624                         __skb_queue_tail(&path->queue, skb);
625
626                         if (path_rec_start(dev, path)) {
627                                 spin_unlock(&priv->lock);
628                                 path_free(dev, path);
629                                 return;
630                         } else
631                                 __path_add(dev, path);
632                 } else {
633                         ++priv->stats.tx_dropped;
634                         dev_kfree_skb_any(skb);
635                 }
636
637                 spin_unlock(&priv->lock);
638                 return;
639         }
640
641         if (path->ah) {
642                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
643                           be16_to_cpu(path->pathrec.dlid));
644
645                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
646         } else if ((path->query || !path_rec_start(dev, path)) &&
647                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
648                 /* put pseudoheader back on for next time */
649                 skb_push(skb, sizeof *phdr);
650                 __skb_queue_tail(&path->queue, skb);
651         } else {
652                 ++priv->stats.tx_dropped;
653                 dev_kfree_skb_any(skb);
654         }
655
656         spin_unlock(&priv->lock);
657 }
658
659 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
660 {
661         struct ipoib_dev_priv *priv = netdev_priv(dev);
662         struct ipoib_neigh *neigh;
663         unsigned long flags;
664
665         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
666                 return NETDEV_TX_LOCKED;
667
668         /*
669          * Check if our queue is stopped.  Since we have the LLTX bit
670          * set, we can't rely on netif_stop_queue() preventing our
671          * xmit function from being called with a full queue.
672          */
673         if (unlikely(netif_queue_stopped(dev))) {
674                 spin_unlock_irqrestore(&priv->tx_lock, flags);
675                 return NETDEV_TX_BUSY;
676         }
677
678         if (likely(skb->dst && skb->dst->neighbour)) {
679                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
680                         ipoib_path_lookup(skb, dev);
681                         goto out;
682                 }
683
684                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
685
686                 if (ipoib_cm_get(neigh)) {
687                         if (ipoib_cm_up(neigh)) {
688                                 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
689                                 goto out;
690                         }
691                 } else if (neigh->ah) {
692                         if (unlikely(memcmp(&neigh->dgid.raw,
693                                             skb->dst->neighbour->ha + 4,
694                                             sizeof(union ib_gid)))) {
695                                 spin_lock(&priv->lock);
696                                 /*
697                                  * It's safe to call ipoib_put_ah() inside
698                                  * priv->lock here, because we know that
699                                  * path->ah will always hold one more reference,
700                                  * so ipoib_put_ah() will never do more than
701                                  * decrement the ref count.
702                                  */
703                                 ipoib_put_ah(neigh->ah);
704                                 list_del(&neigh->list);
705                                 ipoib_neigh_free(dev, neigh);
706                                 spin_unlock(&priv->lock);
707                                 ipoib_path_lookup(skb, dev);
708                                 goto out;
709                         }
710
711                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
712                         goto out;
713                 }
714
715                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
716                         spin_lock(&priv->lock);
717                         __skb_queue_tail(&neigh->queue, skb);
718                         spin_unlock(&priv->lock);
719                 } else {
720                         ++priv->stats.tx_dropped;
721                         dev_kfree_skb_any(skb);
722                 }
723         } else {
724                 struct ipoib_pseudoheader *phdr =
725                         (struct ipoib_pseudoheader *) skb->data;
726                 skb_pull(skb, sizeof *phdr);
727
728                 if (phdr->hwaddr[4] == 0xff) {
729                         /* Add in the P_Key for multicast*/
730                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
731                         phdr->hwaddr[9] = priv->pkey & 0xff;
732
733                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
734                 } else {
735                         /* unicast GID -- should be ARP or RARP reply */
736
737                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
738                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
739                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
740                                            IPOIB_GID_FMT "\n",
741                                            skb->dst ? "neigh" : "dst",
742                                            be16_to_cpup((__be16 *) skb->data),
743                                            IPOIB_QPN(phdr->hwaddr),
744                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
745                                 dev_kfree_skb_any(skb);
746                                 ++priv->stats.tx_dropped;
747                                 goto out;
748                         }
749
750                         unicast_arp_send(skb, dev, phdr);
751                 }
752         }
753
754 out:
755         spin_unlock_irqrestore(&priv->tx_lock, flags);
756
757         return NETDEV_TX_OK;
758 }
759
760 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
761 {
762         struct ipoib_dev_priv *priv = netdev_priv(dev);
763
764         return &priv->stats;
765 }
766
767 static void ipoib_timeout(struct net_device *dev)
768 {
769         struct ipoib_dev_priv *priv = netdev_priv(dev);
770
771         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
772                    jiffies_to_msecs(jiffies - dev->trans_start));
773         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
774                    netif_queue_stopped(dev),
775                    priv->tx_head, priv->tx_tail);
776         /* XXX reset QP, etc. */
777 }
778
779 static int ipoib_hard_header(struct sk_buff *skb,
780                              struct net_device *dev,
781                              unsigned short type,
782                              void *daddr, void *saddr, unsigned len)
783 {
784         struct ipoib_header *header;
785
786         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
787
788         header->proto = htons(type);
789         header->reserved = 0;
790
791         /*
792          * If we don't have a neighbour structure, stuff the
793          * destination address onto the front of the skb so we can
794          * figure out where to send the packet later.
795          */
796         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
797                 struct ipoib_pseudoheader *phdr =
798                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
799                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
800         }
801
802         return 0;
803 }
804
805 static void ipoib_set_mcast_list(struct net_device *dev)
806 {
807         struct ipoib_dev_priv *priv = netdev_priv(dev);
808
809         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
810                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
811                 return;
812         }
813
814         queue_work(ipoib_workqueue, &priv->restart_task);
815 }
816
817 static void ipoib_neigh_destructor(struct neighbour *n)
818 {
819         struct ipoib_neigh *neigh;
820         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
821         unsigned long flags;
822         struct ipoib_ah *ah = NULL;
823
824         ipoib_dbg(priv,
825                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
826                   IPOIB_QPN(n->ha),
827                   IPOIB_GID_RAW_ARG(n->ha + 4));
828
829         spin_lock_irqsave(&priv->lock, flags);
830
831         neigh = *to_ipoib_neigh(n);
832         if (neigh) {
833                 if (neigh->ah)
834                         ah = neigh->ah;
835                 list_del(&neigh->list);
836                 ipoib_neigh_free(n->dev, neigh);
837         }
838
839         spin_unlock_irqrestore(&priv->lock, flags);
840
841         if (ah)
842                 ipoib_put_ah(ah);
843 }
844
845 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
846 {
847         struct ipoib_neigh *neigh;
848
849         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
850         if (!neigh)
851                 return NULL;
852
853         neigh->neighbour = neighbour;
854         *to_ipoib_neigh(neighbour) = neigh;
855         skb_queue_head_init(&neigh->queue);
856         ipoib_cm_set(neigh, NULL);
857
858         return neigh;
859 }
860
861 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
862 {
863         struct ipoib_dev_priv *priv = netdev_priv(dev);
864         struct sk_buff *skb;
865         *to_ipoib_neigh(neigh->neighbour) = NULL;
866         while ((skb = __skb_dequeue(&neigh->queue))) {
867                 ++priv->stats.tx_dropped;
868                 dev_kfree_skb_any(skb);
869         }
870         if (ipoib_cm_get(neigh))
871                 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
872         kfree(neigh);
873 }
874
875 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
876 {
877         parms->neigh_destructor = ipoib_neigh_destructor;
878
879         return 0;
880 }
881
882 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
883 {
884         struct ipoib_dev_priv *priv = netdev_priv(dev);
885
886         /* Allocate RX/TX "rings" to hold queued skbs */
887         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
888                                 GFP_KERNEL);
889         if (!priv->rx_ring) {
890                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
891                        ca->name, ipoib_recvq_size);
892                 goto out;
893         }
894
895         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
896                                 GFP_KERNEL);
897         if (!priv->tx_ring) {
898                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
899                        ca->name, ipoib_sendq_size);
900                 goto out_rx_ring_cleanup;
901         }
902
903         /* priv->tx_head & tx_tail are already 0 */
904
905         if (ipoib_ib_dev_init(dev, ca, port))
906                 goto out_tx_ring_cleanup;
907
908         return 0;
909
910 out_tx_ring_cleanup:
911         kfree(priv->tx_ring);
912
913 out_rx_ring_cleanup:
914         kfree(priv->rx_ring);
915
916 out:
917         return -ENOMEM;
918 }
919
920 void ipoib_dev_cleanup(struct net_device *dev)
921 {
922         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
923
924         ipoib_delete_debug_files(dev);
925
926         /* Delete any child interfaces first */
927         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
928                 unregister_netdev(cpriv->dev);
929                 ipoib_dev_cleanup(cpriv->dev);
930                 free_netdev(cpriv->dev);
931         }
932
933         ipoib_ib_dev_cleanup(dev);
934
935         kfree(priv->rx_ring);
936         kfree(priv->tx_ring);
937
938         priv->rx_ring = NULL;
939         priv->tx_ring = NULL;
940 }
941
942 static void ipoib_setup(struct net_device *dev)
943 {
944         struct ipoib_dev_priv *priv = netdev_priv(dev);
945
946         dev->open                = ipoib_open;
947         dev->stop                = ipoib_stop;
948         dev->change_mtu          = ipoib_change_mtu;
949         dev->hard_start_xmit     = ipoib_start_xmit;
950         dev->get_stats           = ipoib_get_stats;
951         dev->tx_timeout          = ipoib_timeout;
952         dev->hard_header         = ipoib_hard_header;
953         dev->set_multicast_list  = ipoib_set_mcast_list;
954         dev->neigh_setup         = ipoib_neigh_setup_dev;
955
956         dev->watchdog_timeo      = HZ;
957
958         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
959
960         /*
961          * We add in INFINIBAND_ALEN to allow for the destination
962          * address "pseudoheader" for skbs without neighbour struct.
963          */
964         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
965         dev->addr_len            = INFINIBAND_ALEN;
966         dev->type                = ARPHRD_INFINIBAND;
967         dev->tx_queue_len        = ipoib_sendq_size * 2;
968         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
969
970         /* MTU will be reset when mcast join happens */
971         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
972         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
973
974         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
975
976         netif_carrier_off(dev);
977
978         SET_MODULE_OWNER(dev);
979
980         priv->dev = dev;
981
982         spin_lock_init(&priv->lock);
983         spin_lock_init(&priv->tx_lock);
984
985         mutex_init(&priv->mcast_mutex);
986         mutex_init(&priv->vlan_mutex);
987
988         INIT_LIST_HEAD(&priv->path_list);
989         INIT_LIST_HEAD(&priv->child_intfs);
990         INIT_LIST_HEAD(&priv->dead_ahs);
991         INIT_LIST_HEAD(&priv->multicast_list);
992
993         INIT_DELAYED_WORK(&priv->pkey_task,    ipoib_pkey_poll);
994         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
995         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
996         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
997         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
998 }
999
1000 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1001 {
1002         struct net_device *dev;
1003
1004         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1005                            ipoib_setup);
1006         if (!dev)
1007                 return NULL;
1008
1009         return netdev_priv(dev);
1010 }
1011
1012 static ssize_t show_pkey(struct device *dev,
1013                          struct device_attribute *attr, char *buf)
1014 {
1015         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1016
1017         return sprintf(buf, "0x%04x\n", priv->pkey);
1018 }
1019 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1020
1021 static ssize_t create_child(struct device *dev,
1022                             struct device_attribute *attr,
1023                             const char *buf, size_t count)
1024 {
1025         int pkey;
1026         int ret;
1027
1028         if (sscanf(buf, "%i", &pkey) != 1)
1029                 return -EINVAL;
1030
1031         if (pkey < 0 || pkey > 0xffff)
1032                 return -EINVAL;
1033
1034         /*
1035          * Set the full membership bit, so that we join the right
1036          * broadcast group, etc.
1037          */
1038         pkey |= 0x8000;
1039
1040         ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1041
1042         return ret ? ret : count;
1043 }
1044 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1045
1046 static ssize_t delete_child(struct device *dev,
1047                             struct device_attribute *attr,
1048                             const char *buf, size_t count)
1049 {
1050         int pkey;
1051         int ret;
1052
1053         if (sscanf(buf, "%i", &pkey) != 1)
1054                 return -EINVAL;
1055
1056         if (pkey < 0 || pkey > 0xffff)
1057                 return -EINVAL;
1058
1059         ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1060
1061         return ret ? ret : count;
1062
1063 }
1064 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1065
1066 int ipoib_add_pkey_attr(struct net_device *dev)
1067 {
1068         return device_create_file(&dev->dev, &dev_attr_pkey);
1069 }
1070
1071 static struct net_device *ipoib_add_port(const char *format,
1072                                          struct ib_device *hca, u8 port)
1073 {
1074         struct ipoib_dev_priv *priv;
1075         int result = -ENOMEM;
1076
1077         priv = ipoib_intf_alloc(format);
1078         if (!priv)
1079                 goto alloc_mem_failed;
1080
1081         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1082
1083         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1084         if (result) {
1085                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1086                        hca->name, port, result);
1087                 goto alloc_mem_failed;
1088         }
1089
1090         /*
1091          * Set the full membership bit, so that we join the right
1092          * broadcast group, etc.
1093          */
1094         priv->pkey |= 0x8000;
1095
1096         priv->dev->broadcast[8] = priv->pkey >> 8;
1097         priv->dev->broadcast[9] = priv->pkey & 0xff;
1098
1099         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1100         if (result) {
1101                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1102                        hca->name, port, result);
1103                 goto alloc_mem_failed;
1104         } else
1105                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1106
1107
1108         result = ipoib_dev_init(priv->dev, hca, port);
1109         if (result < 0) {
1110                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1111                        hca->name, port, result);
1112                 goto device_init_failed;
1113         }
1114
1115         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1116                               priv->ca, ipoib_event);
1117         result = ib_register_event_handler(&priv->event_handler);
1118         if (result < 0) {
1119                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1120                        "port %d (ret = %d)\n",
1121                        hca->name, port, result);
1122                 goto event_failed;
1123         }
1124
1125         result = register_netdev(priv->dev);
1126         if (result) {
1127                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1128                        hca->name, port, result);
1129                 goto register_failed;
1130         }
1131
1132         ipoib_create_debug_files(priv->dev);
1133
1134         if (ipoib_cm_add_mode_attr(priv->dev))
1135                 goto sysfs_failed;
1136         if (ipoib_add_pkey_attr(priv->dev))
1137                 goto sysfs_failed;
1138         if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1139                 goto sysfs_failed;
1140         if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1141                 goto sysfs_failed;
1142
1143         return priv->dev;
1144
1145 sysfs_failed:
1146         ipoib_delete_debug_files(priv->dev);
1147         unregister_netdev(priv->dev);
1148
1149 register_failed:
1150         ib_unregister_event_handler(&priv->event_handler);
1151         flush_scheduled_work();
1152
1153 event_failed:
1154         ipoib_dev_cleanup(priv->dev);
1155
1156 device_init_failed:
1157         free_netdev(priv->dev);
1158
1159 alloc_mem_failed:
1160         return ERR_PTR(result);
1161 }
1162
1163 static void ipoib_add_one(struct ib_device *device)
1164 {
1165         struct list_head *dev_list;
1166         struct net_device *dev;
1167         struct ipoib_dev_priv *priv;
1168         int s, e, p;
1169
1170         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1171                 return;
1172
1173         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1174         if (!dev_list)
1175                 return;
1176
1177         INIT_LIST_HEAD(dev_list);
1178
1179         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1180                 s = 0;
1181                 e = 0;
1182         } else {
1183                 s = 1;
1184                 e = device->phys_port_cnt;
1185         }
1186
1187         for (p = s; p <= e; ++p) {
1188                 dev = ipoib_add_port("ib%d", device, p);
1189                 if (!IS_ERR(dev)) {
1190                         priv = netdev_priv(dev);
1191                         list_add_tail(&priv->list, dev_list);
1192                 }
1193         }
1194
1195         ib_set_client_data(device, &ipoib_client, dev_list);
1196 }
1197
1198 static void ipoib_remove_one(struct ib_device *device)
1199 {
1200         struct ipoib_dev_priv *priv, *tmp;
1201         struct list_head *dev_list;
1202
1203         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1204                 return;
1205
1206         dev_list = ib_get_client_data(device, &ipoib_client);
1207
1208         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1209                 ib_unregister_event_handler(&priv->event_handler);
1210                 flush_scheduled_work();
1211
1212                 unregister_netdev(priv->dev);
1213                 ipoib_dev_cleanup(priv->dev);
1214                 free_netdev(priv->dev);
1215         }
1216
1217         kfree(dev_list);
1218 }
1219
1220 static int __init ipoib_init_module(void)
1221 {
1222         int ret;
1223
1224         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1225         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1226         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1227
1228         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1229         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1230         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1231
1232         ret = ipoib_register_debugfs();
1233         if (ret)
1234                 return ret;
1235
1236         /*
1237          * We create our own workqueue mainly because we want to be
1238          * able to flush it when devices are being removed.  We can't
1239          * use schedule_work()/flush_scheduled_work() because both
1240          * unregister_netdev() and linkwatch_event take the rtnl lock,
1241          * so flush_scheduled_work() can deadlock during device
1242          * removal.
1243          */
1244         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1245         if (!ipoib_workqueue) {
1246                 ret = -ENOMEM;
1247                 goto err_fs;
1248         }
1249
1250         ib_sa_register_client(&ipoib_sa_client);
1251
1252         ret = ib_register_client(&ipoib_client);
1253         if (ret)
1254                 goto err_sa;
1255
1256         return 0;
1257
1258 err_sa:
1259         ib_sa_unregister_client(&ipoib_sa_client);
1260         destroy_workqueue(ipoib_workqueue);
1261
1262 err_fs:
1263         ipoib_unregister_debugfs();
1264
1265         return ret;
1266 }
1267
1268 static void __exit ipoib_cleanup_module(void)
1269 {
1270         ib_unregister_client(&ipoib_client);
1271         ib_sa_unregister_client(&ipoib_sa_client);
1272         ipoib_unregister_debugfs();
1273         destroy_workqueue(ipoib_workqueue);
1274 }
1275
1276 module_init(ipoib_init_module);
1277 module_exit(ipoib_cleanup_module);