Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/drzeus/mmc
[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 #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(dev, 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
529                 if (!path->query && path_rec_start(dev, path))
530                         goto err_list;
531
532                 __skb_queue_tail(&neigh->queue, skb);
533         }
534
535         spin_unlock(&priv->lock);
536         return;
537
538 err_list:
539         list_del(&neigh->list);
540
541 err_path:
542         ipoib_neigh_free(dev, neigh);
543         ++priv->stats.tx_dropped;
544         dev_kfree_skb_any(skb);
545
546         spin_unlock(&priv->lock);
547 }
548
549 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
550 {
551         struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
552
553         /* Look up path record for unicasts */
554         if (skb->dst->neighbour->ha[4] != 0xff) {
555                 neigh_add_path(skb, dev);
556                 return;
557         }
558
559         /* Add in the P_Key for multicasts */
560         skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
561         skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
562         ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
563 }
564
565 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
566                              struct ipoib_pseudoheader *phdr)
567 {
568         struct ipoib_dev_priv *priv = netdev_priv(dev);
569         struct ipoib_path *path;
570
571         /*
572          * We can only be called from ipoib_start_xmit, so we're
573          * inside tx_lock -- no need to save/restore flags.
574          */
575         spin_lock(&priv->lock);
576
577         path = __path_find(dev, phdr->hwaddr + 4);
578         if (!path) {
579                 path = path_rec_create(dev, phdr->hwaddr + 4);
580                 if (path) {
581                         /* put pseudoheader back on for next time */
582                         skb_push(skb, sizeof *phdr);
583                         __skb_queue_tail(&path->queue, skb);
584
585                         if (path_rec_start(dev, path)) {
586                                 spin_unlock(&priv->lock);
587                                 path_free(dev, path);
588                                 return;
589                         } else
590                                 __path_add(dev, path);
591                 } else {
592                         ++priv->stats.tx_dropped;
593                         dev_kfree_skb_any(skb);
594                 }
595
596                 spin_unlock(&priv->lock);
597                 return;
598         }
599
600         if (path->ah) {
601                 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
602                           be16_to_cpu(path->pathrec.dlid));
603
604                 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
605         } else if ((path->query || !path_rec_start(dev, path)) &&
606                    skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
607                 /* put pseudoheader back on for next time */
608                 skb_push(skb, sizeof *phdr);
609                 __skb_queue_tail(&path->queue, skb);
610         } else {
611                 ++priv->stats.tx_dropped;
612                 dev_kfree_skb_any(skb);
613         }
614
615         spin_unlock(&priv->lock);
616 }
617
618 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
619 {
620         struct ipoib_dev_priv *priv = netdev_priv(dev);
621         struct ipoib_neigh *neigh;
622         unsigned long flags;
623
624         if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
625                 return NETDEV_TX_LOCKED;
626
627         /*
628          * Check if our queue is stopped.  Since we have the LLTX bit
629          * set, we can't rely on netif_stop_queue() preventing our
630          * xmit function from being called with a full queue.
631          */
632         if (unlikely(netif_queue_stopped(dev))) {
633                 spin_unlock_irqrestore(&priv->tx_lock, flags);
634                 return NETDEV_TX_BUSY;
635         }
636
637         if (likely(skb->dst && skb->dst->neighbour)) {
638                 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
639                         ipoib_path_lookup(skb, dev);
640                         goto out;
641                 }
642
643                 neigh = *to_ipoib_neigh(skb->dst->neighbour);
644
645                 if (likely(neigh->ah)) {
646                         if (unlikely(memcmp(&neigh->dgid.raw,
647                                             skb->dst->neighbour->ha + 4,
648                                             sizeof(union ib_gid)))) {
649                                 spin_lock(&priv->lock);
650                                 /*
651                                  * It's safe to call ipoib_put_ah() inside
652                                  * priv->lock here, because we know that
653                                  * path->ah will always hold one more reference,
654                                  * so ipoib_put_ah() will never do more than
655                                  * decrement the ref count.
656                                  */
657                                 ipoib_put_ah(neigh->ah);
658                                 list_del(&neigh->list);
659                                 ipoib_neigh_free(dev, neigh);
660                                 spin_unlock(&priv->lock);
661                                 ipoib_path_lookup(skb, dev);
662                                 goto out;
663                         }
664
665                         ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
666                         goto out;
667                 }
668
669                 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
670                         spin_lock(&priv->lock);
671                         __skb_queue_tail(&neigh->queue, skb);
672                         spin_unlock(&priv->lock);
673                 } else {
674                         ++priv->stats.tx_dropped;
675                         dev_kfree_skb_any(skb);
676                 }
677         } else {
678                 struct ipoib_pseudoheader *phdr =
679                         (struct ipoib_pseudoheader *) skb->data;
680                 skb_pull(skb, sizeof *phdr);
681
682                 if (phdr->hwaddr[4] == 0xff) {
683                         /* Add in the P_Key for multicast*/
684                         phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
685                         phdr->hwaddr[9] = priv->pkey & 0xff;
686
687                         ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
688                 } else {
689                         /* unicast GID -- should be ARP or RARP reply */
690
691                         if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
692                             (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
693                                 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
694                                            IPOIB_GID_FMT "\n",
695                                            skb->dst ? "neigh" : "dst",
696                                            be16_to_cpup((__be16 *) skb->data),
697                                            IPOIB_QPN(phdr->hwaddr),
698                                            IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
699                                 dev_kfree_skb_any(skb);
700                                 ++priv->stats.tx_dropped;
701                                 goto out;
702                         }
703
704                         unicast_arp_send(skb, dev, phdr);
705                 }
706         }
707
708 out:
709         spin_unlock_irqrestore(&priv->tx_lock, flags);
710
711         return NETDEV_TX_OK;
712 }
713
714 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
715 {
716         struct ipoib_dev_priv *priv = netdev_priv(dev);
717
718         return &priv->stats;
719 }
720
721 static void ipoib_timeout(struct net_device *dev)
722 {
723         struct ipoib_dev_priv *priv = netdev_priv(dev);
724
725         ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
726                    jiffies_to_msecs(jiffies - dev->trans_start));
727         ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
728                    netif_queue_stopped(dev),
729                    priv->tx_head, priv->tx_tail);
730         /* XXX reset QP, etc. */
731 }
732
733 static int ipoib_hard_header(struct sk_buff *skb,
734                              struct net_device *dev,
735                              unsigned short type,
736                              void *daddr, void *saddr, unsigned len)
737 {
738         struct ipoib_header *header;
739
740         header = (struct ipoib_header *) skb_push(skb, sizeof *header);
741
742         header->proto = htons(type);
743         header->reserved = 0;
744
745         /*
746          * If we don't have a neighbour structure, stuff the
747          * destination address onto the front of the skb so we can
748          * figure out where to send the packet later.
749          */
750         if ((!skb->dst || !skb->dst->neighbour) && daddr) {
751                 struct ipoib_pseudoheader *phdr =
752                         (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
753                 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
754         }
755
756         return 0;
757 }
758
759 static void ipoib_set_mcast_list(struct net_device *dev)
760 {
761         struct ipoib_dev_priv *priv = netdev_priv(dev);
762
763         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
764                 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
765                 return;
766         }
767
768         queue_work(ipoib_workqueue, &priv->restart_task);
769 }
770
771 static void ipoib_neigh_destructor(struct neighbour *n)
772 {
773         struct ipoib_neigh *neigh;
774         struct ipoib_dev_priv *priv = netdev_priv(n->dev);
775         unsigned long flags;
776         struct ipoib_ah *ah = NULL;
777
778         ipoib_dbg(priv,
779                   "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
780                   IPOIB_QPN(n->ha),
781                   IPOIB_GID_RAW_ARG(n->ha + 4));
782
783         spin_lock_irqsave(&priv->lock, flags);
784
785         neigh = *to_ipoib_neigh(n);
786         if (neigh) {
787                 if (neigh->ah)
788                         ah = neigh->ah;
789                 list_del(&neigh->list);
790                 ipoib_neigh_free(n->dev, neigh);
791         }
792
793         spin_unlock_irqrestore(&priv->lock, flags);
794
795         if (ah)
796                 ipoib_put_ah(ah);
797 }
798
799 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
800 {
801         struct ipoib_neigh *neigh;
802
803         neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
804         if (!neigh)
805                 return NULL;
806
807         neigh->neighbour = neighbour;
808         *to_ipoib_neigh(neighbour) = neigh;
809
810         return neigh;
811 }
812
813 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
814 {
815         struct ipoib_dev_priv *priv = netdev_priv(dev);
816         struct sk_buff *skb;
817         *to_ipoib_neigh(neigh->neighbour) = NULL;
818         while ((skb = __skb_dequeue(&neigh->queue))) {
819                 ++priv->stats.tx_dropped;
820                 dev_kfree_skb_any(skb);
821         }
822         kfree(neigh);
823 }
824
825 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
826 {
827         parms->neigh_destructor = ipoib_neigh_destructor;
828
829         return 0;
830 }
831
832 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
833 {
834         struct ipoib_dev_priv *priv = netdev_priv(dev);
835
836         /* Allocate RX/TX "rings" to hold queued skbs */
837         priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
838                                 GFP_KERNEL);
839         if (!priv->rx_ring) {
840                 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
841                        ca->name, ipoib_recvq_size);
842                 goto out;
843         }
844
845         priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
846                                 GFP_KERNEL);
847         if (!priv->tx_ring) {
848                 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
849                        ca->name, ipoib_sendq_size);
850                 goto out_rx_ring_cleanup;
851         }
852
853         /* priv->tx_head & tx_tail are already 0 */
854
855         if (ipoib_ib_dev_init(dev, ca, port))
856                 goto out_tx_ring_cleanup;
857
858         return 0;
859
860 out_tx_ring_cleanup:
861         kfree(priv->tx_ring);
862
863 out_rx_ring_cleanup:
864         kfree(priv->rx_ring);
865
866 out:
867         return -ENOMEM;
868 }
869
870 void ipoib_dev_cleanup(struct net_device *dev)
871 {
872         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
873
874         ipoib_delete_debug_files(dev);
875
876         /* Delete any child interfaces first */
877         list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
878                 unregister_netdev(cpriv->dev);
879                 ipoib_dev_cleanup(cpriv->dev);
880                 free_netdev(cpriv->dev);
881         }
882
883         ipoib_ib_dev_cleanup(dev);
884
885         kfree(priv->rx_ring);
886         kfree(priv->tx_ring);
887
888         priv->rx_ring = NULL;
889         priv->tx_ring = NULL;
890 }
891
892 static void ipoib_setup(struct net_device *dev)
893 {
894         struct ipoib_dev_priv *priv = netdev_priv(dev);
895
896         dev->open                = ipoib_open;
897         dev->stop                = ipoib_stop;
898         dev->change_mtu          = ipoib_change_mtu;
899         dev->hard_start_xmit     = ipoib_start_xmit;
900         dev->get_stats           = ipoib_get_stats;
901         dev->tx_timeout          = ipoib_timeout;
902         dev->hard_header         = ipoib_hard_header;
903         dev->set_multicast_list  = ipoib_set_mcast_list;
904         dev->neigh_setup         = ipoib_neigh_setup_dev;
905
906         dev->watchdog_timeo      = HZ;
907
908         dev->flags              |= IFF_BROADCAST | IFF_MULTICAST;
909
910         /*
911          * We add in INFINIBAND_ALEN to allow for the destination
912          * address "pseudoheader" for skbs without neighbour struct.
913          */
914         dev->hard_header_len     = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
915         dev->addr_len            = INFINIBAND_ALEN;
916         dev->type                = ARPHRD_INFINIBAND;
917         dev->tx_queue_len        = ipoib_sendq_size * 2;
918         dev->features            = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
919
920         /* MTU will be reset when mcast join happens */
921         dev->mtu                 = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
922         priv->mcast_mtu          = priv->admin_mtu = dev->mtu;
923
924         memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
925
926         netif_carrier_off(dev);
927
928         SET_MODULE_OWNER(dev);
929
930         priv->dev = dev;
931
932         spin_lock_init(&priv->lock);
933         spin_lock_init(&priv->tx_lock);
934
935         mutex_init(&priv->mcast_mutex);
936         mutex_init(&priv->vlan_mutex);
937
938         INIT_LIST_HEAD(&priv->path_list);
939         INIT_LIST_HEAD(&priv->child_intfs);
940         INIT_LIST_HEAD(&priv->dead_ahs);
941         INIT_LIST_HEAD(&priv->multicast_list);
942
943         INIT_DELAYED_WORK(&priv->pkey_task,    ipoib_pkey_poll);
944         INIT_DELAYED_WORK(&priv->mcast_task,   ipoib_mcast_join_task);
945         INIT_WORK(&priv->flush_task,   ipoib_ib_dev_flush);
946         INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
947         INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
948 }
949
950 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
951 {
952         struct net_device *dev;
953
954         dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
955                            ipoib_setup);
956         if (!dev)
957                 return NULL;
958
959         return netdev_priv(dev);
960 }
961
962 static ssize_t show_pkey(struct class_device *cdev, char *buf)
963 {
964         struct ipoib_dev_priv *priv =
965                 netdev_priv(container_of(cdev, struct net_device, class_dev));
966
967         return sprintf(buf, "0x%04x\n", priv->pkey);
968 }
969 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
970
971 static ssize_t create_child(struct class_device *cdev,
972                             const char *buf, size_t count)
973 {
974         int pkey;
975         int ret;
976
977         if (sscanf(buf, "%i", &pkey) != 1)
978                 return -EINVAL;
979
980         if (pkey < 0 || pkey > 0xffff)
981                 return -EINVAL;
982
983         /*
984          * Set the full membership bit, so that we join the right
985          * broadcast group, etc.
986          */
987         pkey |= 0x8000;
988
989         ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
990                              pkey);
991
992         return ret ? ret : count;
993 }
994 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
995
996 static ssize_t delete_child(struct class_device *cdev,
997                             const char *buf, size_t count)
998 {
999         int pkey;
1000         int ret;
1001
1002         if (sscanf(buf, "%i", &pkey) != 1)
1003                 return -EINVAL;
1004
1005         if (pkey < 0 || pkey > 0xffff)
1006                 return -EINVAL;
1007
1008         ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
1009                                 pkey);
1010
1011         return ret ? ret : count;
1012
1013 }
1014 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1015
1016 int ipoib_add_pkey_attr(struct net_device *dev)
1017 {
1018         return class_device_create_file(&dev->class_dev,
1019                                         &class_device_attr_pkey);
1020 }
1021
1022 static struct net_device *ipoib_add_port(const char *format,
1023                                          struct ib_device *hca, u8 port)
1024 {
1025         struct ipoib_dev_priv *priv;
1026         int result = -ENOMEM;
1027
1028         priv = ipoib_intf_alloc(format);
1029         if (!priv)
1030                 goto alloc_mem_failed;
1031
1032         SET_NETDEV_DEV(priv->dev, hca->dma_device);
1033
1034         result = ib_query_pkey(hca, port, 0, &priv->pkey);
1035         if (result) {
1036                 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1037                        hca->name, port, result);
1038                 goto alloc_mem_failed;
1039         }
1040
1041         /*
1042          * Set the full membership bit, so that we join the right
1043          * broadcast group, etc.
1044          */
1045         priv->pkey |= 0x8000;
1046
1047         priv->dev->broadcast[8] = priv->pkey >> 8;
1048         priv->dev->broadcast[9] = priv->pkey & 0xff;
1049
1050         result = ib_query_gid(hca, port, 0, &priv->local_gid);
1051         if (result) {
1052                 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1053                        hca->name, port, result);
1054                 goto alloc_mem_failed;
1055         } else
1056                 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1057
1058
1059         result = ipoib_dev_init(priv->dev, hca, port);
1060         if (result < 0) {
1061                 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1062                        hca->name, port, result);
1063                 goto device_init_failed;
1064         }
1065
1066         INIT_IB_EVENT_HANDLER(&priv->event_handler,
1067                               priv->ca, ipoib_event);
1068         result = ib_register_event_handler(&priv->event_handler);
1069         if (result < 0) {
1070                 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1071                        "port %d (ret = %d)\n",
1072                        hca->name, port, result);
1073                 goto event_failed;
1074         }
1075
1076         result = register_netdev(priv->dev);
1077         if (result) {
1078                 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1079                        hca->name, port, result);
1080                 goto register_failed;
1081         }
1082
1083         ipoib_create_debug_files(priv->dev);
1084
1085         if (ipoib_add_pkey_attr(priv->dev))
1086                 goto sysfs_failed;
1087         if (class_device_create_file(&priv->dev->class_dev,
1088                                      &class_device_attr_create_child))
1089                 goto sysfs_failed;
1090         if (class_device_create_file(&priv->dev->class_dev,
1091                                      &class_device_attr_delete_child))
1092                 goto sysfs_failed;
1093
1094         return priv->dev;
1095
1096 sysfs_failed:
1097         ipoib_delete_debug_files(priv->dev);
1098         unregister_netdev(priv->dev);
1099
1100 register_failed:
1101         ib_unregister_event_handler(&priv->event_handler);
1102         flush_scheduled_work();
1103
1104 event_failed:
1105         ipoib_dev_cleanup(priv->dev);
1106
1107 device_init_failed:
1108         free_netdev(priv->dev);
1109
1110 alloc_mem_failed:
1111         return ERR_PTR(result);
1112 }
1113
1114 static void ipoib_add_one(struct ib_device *device)
1115 {
1116         struct list_head *dev_list;
1117         struct net_device *dev;
1118         struct ipoib_dev_priv *priv;
1119         int s, e, p;
1120
1121         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1122                 return;
1123
1124         dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1125         if (!dev_list)
1126                 return;
1127
1128         INIT_LIST_HEAD(dev_list);
1129
1130         if (device->node_type == RDMA_NODE_IB_SWITCH) {
1131                 s = 0;
1132                 e = 0;
1133         } else {
1134                 s = 1;
1135                 e = device->phys_port_cnt;
1136         }
1137
1138         for (p = s; p <= e; ++p) {
1139                 dev = ipoib_add_port("ib%d", device, p);
1140                 if (!IS_ERR(dev)) {
1141                         priv = netdev_priv(dev);
1142                         list_add_tail(&priv->list, dev_list);
1143                 }
1144         }
1145
1146         ib_set_client_data(device, &ipoib_client, dev_list);
1147 }
1148
1149 static void ipoib_remove_one(struct ib_device *device)
1150 {
1151         struct ipoib_dev_priv *priv, *tmp;
1152         struct list_head *dev_list;
1153
1154         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1155                 return;
1156
1157         dev_list = ib_get_client_data(device, &ipoib_client);
1158
1159         list_for_each_entry_safe(priv, tmp, dev_list, list) {
1160                 ib_unregister_event_handler(&priv->event_handler);
1161                 flush_scheduled_work();
1162
1163                 unregister_netdev(priv->dev);
1164                 ipoib_dev_cleanup(priv->dev);
1165                 free_netdev(priv->dev);
1166         }
1167
1168         kfree(dev_list);
1169 }
1170
1171 static int __init ipoib_init_module(void)
1172 {
1173         int ret;
1174
1175         ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1176         ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1177         ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1178
1179         ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1180         ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1181         ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1182
1183         ret = ipoib_register_debugfs();
1184         if (ret)
1185                 return ret;
1186
1187         /*
1188          * We create our own workqueue mainly because we want to be
1189          * able to flush it when devices are being removed.  We can't
1190          * use schedule_work()/flush_scheduled_work() because both
1191          * unregister_netdev() and linkwatch_event take the rtnl lock,
1192          * so flush_scheduled_work() can deadlock during device
1193          * removal.
1194          */
1195         ipoib_workqueue = create_singlethread_workqueue("ipoib");
1196         if (!ipoib_workqueue) {
1197                 ret = -ENOMEM;
1198                 goto err_fs;
1199         }
1200
1201         ib_sa_register_client(&ipoib_sa_client);
1202
1203         ret = ib_register_client(&ipoib_client);
1204         if (ret)
1205                 goto err_sa;
1206
1207         return 0;
1208
1209 err_sa:
1210         ib_sa_unregister_client(&ipoib_sa_client);
1211         destroy_workqueue(ipoib_workqueue);
1212
1213 err_fs:
1214         ipoib_unregister_debugfs();
1215
1216         return ret;
1217 }
1218
1219 static void __exit ipoib_cleanup_module(void)
1220 {
1221         ib_unregister_client(&ipoib_client);
1222         ib_sa_unregister_client(&ipoib_sa_client);
1223         ipoib_unregister_debugfs();
1224         destroy_workqueue(ipoib_workqueue);
1225 }
1226
1227 module_init(ipoib_init_module);
1228 module_exit(ipoib_cleanup_module);