?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'fix/pcm-hwptr' into for-linus
[linux/fpc-iii.git] / drivers / infiniband / ulp / ipoib / ipoib_main.c
blobe319d91f60a609be13d3e18c94df9477207178c6
1 /*
2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
11 *
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
14 * conditions are met:
15 *
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
18 * disclaimer.
19 *
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
24 *
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * SOFTWARE.
33 */
34
35 #include "ipoib.h"
36
37 #include <linux/module.h>
38
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
43
44 #include <linux/if_arp.h> /* For ARPHRD_xxx */
45
46 #include <linux/ip.h>
47 #include <linux/in.h>
48
49 #include <net/dst.h>
50
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
53 MODULE_LICENSE("Dual BSD/GPL");
54
55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
57
58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
62
63 static int lro;
64 module_param(lro, bool, 0444);
65 MODULE_PARM_DESC(lro, "Enable LRO (Large Receive Offload)");
66
67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
68 module_param(lro_max_aggr, int, 0644);
69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
70 "(default = 64)");
71
72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
73 int ipoib_debug_level;
74
75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
77 #endif
78
79 struct ipoib_path_iter {
80 struct net_device *dev;
81 struct ipoib_path path;
82 };
83
84 static const u8 ipv4_bcast_addr[] = {
85 0x00, 0xff, 0xff, 0xff,
86 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
87 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
88 };
89
90 struct workqueue_struct *ipoib_workqueue;
91
92 struct ib_sa_client ipoib_sa_client;
93
94 static void ipoib_add_one(struct ib_device *device);
95 static void ipoib_remove_one(struct ib_device *device);
96
97 static struct ib_client ipoib_client = {
98 .name = "ipoib",
99 .add = ipoib_add_one,
100 .remove = ipoib_remove_one
101 };
102
103 int ipoib_open(struct net_device *dev)
104 {
105 struct ipoib_dev_priv *priv = netdev_priv(dev);
106
107 ipoib_dbg(priv, "bringing up interface\n");
108
109 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
110
111 if (ipoib_pkey_dev_delay_open(dev))
112 return 0;
113
114 if (ipoib_ib_dev_open(dev))
115 goto err_disable;
116
117 if (ipoib_ib_dev_up(dev))
118 goto err_stop;
119
120 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
121 struct ipoib_dev_priv *cpriv;
122
123 /* Bring up any child interfaces too */
124 mutex_lock(&priv->vlan_mutex);
125 list_for_each_entry(cpriv, &priv->child_intfs, list) {
126 int flags;
127
128 flags = cpriv->dev->flags;
129 if (flags & IFF_UP)
130 continue;
131
132 dev_change_flags(cpriv->dev, flags | IFF_UP);
133 }
134 mutex_unlock(&priv->vlan_mutex);
135 }
136
137 netif_start_queue(dev);
138
139 return 0;
140
141 err_stop:
142 ipoib_ib_dev_stop(dev, 1);
143
144 err_disable:
145 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
146
147 return -EINVAL;
148 }
149
150 static int ipoib_stop(struct net_device *dev)
151 {
152 struct ipoib_dev_priv *priv = netdev_priv(dev);
153
154 ipoib_dbg(priv, "stopping interface\n");
155
156 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
157
158 netif_stop_queue(dev);
159
160 ipoib_ib_dev_down(dev, 0);
161 ipoib_ib_dev_stop(dev, 0);
162
163 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
164 struct ipoib_dev_priv *cpriv;
165
166 /* Bring down any child interfaces too */
167 mutex_lock(&priv->vlan_mutex);
168 list_for_each_entry(cpriv, &priv->child_intfs, list) {
169 int flags;
170
171 flags = cpriv->dev->flags;
172 if (!(flags & IFF_UP))
173 continue;
174
175 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
176 }
177 mutex_unlock(&priv->vlan_mutex);
178 }
179
180 return 0;
181 }
182
183 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
184 {
185 struct ipoib_dev_priv *priv = netdev_priv(dev);
186
187 /* dev->mtu > 2K ==> connected mode */
188 if (ipoib_cm_admin_enabled(dev)) {
189 if (new_mtu > ipoib_cm_max_mtu(dev))
190 return -EINVAL;
191
192 if (new_mtu > priv->mcast_mtu)
193 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
194 priv->mcast_mtu);
195
196 dev->mtu = new_mtu;
197 return 0;
198 }
199
200 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
201 return -EINVAL;
202
203 priv->admin_mtu = new_mtu;
204
205 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
206
207 return 0;
208 }
209
210 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
211 {
212 struct ipoib_dev_priv *priv = netdev_priv(dev);
213 struct rb_node *n = priv->path_tree.rb_node;
214 struct ipoib_path *path;
215 int ret;
216
217 while (n) {
218 path = rb_entry(n, struct ipoib_path, rb_node);
219
220 ret = memcmp(gid, path->pathrec.dgid.raw,
221 sizeof (union ib_gid));
222
223 if (ret < 0)
224 n = n->rb_left;
225 else if (ret > 0)
226 n = n->rb_right;
227 else
228 return path;
229 }
230
231 return NULL;
232 }
233
234 static int __path_add(struct net_device *dev, struct ipoib_path *path)
235 {
236 struct ipoib_dev_priv *priv = netdev_priv(dev);
237 struct rb_node **n = &priv->path_tree.rb_node;
238 struct rb_node *pn = NULL;
239 struct ipoib_path *tpath;
240 int ret;
241
242 while (*n) {
243 pn = *n;
244 tpath = rb_entry(pn, struct ipoib_path, rb_node);
245
246 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
247 sizeof (union ib_gid));
248 if (ret < 0)
249 n = &pn->rb_left;
250 else if (ret > 0)
251 n = &pn->rb_right;
252 else
253 return -EEXIST;
254 }
255
256 rb_link_node(&path->rb_node, pn, n);
257 rb_insert_color(&path->rb_node, &priv->path_tree);
258
259 list_add_tail(&path->list, &priv->path_list);
260
261 return 0;
262 }
263
264 static void path_free(struct net_device *dev, struct ipoib_path *path)
265 {
266 struct ipoib_dev_priv *priv = netdev_priv(dev);
267 struct ipoib_neigh *neigh, *tn;
268 struct sk_buff *skb;
269 unsigned long flags;
270
271 while ((skb = __skb_dequeue(&path->queue)))
272 dev_kfree_skb_irq(skb);
273
274 spin_lock_irqsave(&priv->lock, flags);
275
276 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
277 /*
278 * It's safe to call ipoib_put_ah() inside priv->lock
279 * here, because we know that path->ah will always
280 * hold one more reference, so ipoib_put_ah() will
281 * never do more than decrement the ref count.
282 */
283 if (neigh->ah)
284 ipoib_put_ah(neigh->ah);
285
286 ipoib_neigh_free(dev, neigh);
287 }
288
289 spin_unlock_irqrestore(&priv->lock, flags);
290
291 if (path->ah)
292 ipoib_put_ah(path->ah);
293
294 kfree(path);
295 }
296
297 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
298
299 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
300 {
301 struct ipoib_path_iter *iter;
302
303 iter = kmalloc(sizeof *iter, GFP_KERNEL);
304 if (!iter)
305 return NULL;
306
307 iter->dev = dev;
308 memset(iter->path.pathrec.dgid.raw, 0, 16);
309
310 if (ipoib_path_iter_next(iter)) {
311 kfree(iter);
312 return NULL;
313 }
314
315 return iter;
316 }
317
318 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
319 {
320 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
321 struct rb_node *n;
322 struct ipoib_path *path;
323 int ret = 1;
324
325 spin_lock_irq(&priv->lock);
326
327 n = rb_first(&priv->path_tree);
328
329 while (n) {
330 path = rb_entry(n, struct ipoib_path, rb_node);
331
332 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
333 sizeof (union ib_gid)) < 0) {
334 iter->path = *path;
335 ret = 0;
336 break;
337 }
338
339 n = rb_next(n);
340 }
341
342 spin_unlock_irq(&priv->lock);
343
344 return ret;
345 }
346
347 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
348 struct ipoib_path *path)
349 {
350 *path = iter->path;
351 }
352
353 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
354
355 void ipoib_mark_paths_invalid(struct net_device *dev)
356 {
357 struct ipoib_dev_priv *priv = netdev_priv(dev);
358 struct ipoib_path *path, *tp;
359
360 spin_lock_irq(&priv->lock);
361
362 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
363 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
364 be16_to_cpu(path->pathrec.dlid),
365 path->pathrec.dgid.raw);
366 path->valid = 0;
367 }
368
369 spin_unlock_irq(&priv->lock);
370 }
371
372 void ipoib_flush_paths(struct net_device *dev)
373 {
374 struct ipoib_dev_priv *priv = netdev_priv(dev);
375 struct ipoib_path *path, *tp;
376 LIST_HEAD(remove_list);
377 unsigned long flags;
378
379 netif_tx_lock_bh(dev);
380 spin_lock_irqsave(&priv->lock, flags);
381
382 list_splice_init(&priv->path_list, &remove_list);
383
384 list_for_each_entry(path, &remove_list, list)
385 rb_erase(&path->rb_node, &priv->path_tree);
386
387 list_for_each_entry_safe(path, tp, &remove_list, list) {
388 if (path->query)
389 ib_sa_cancel_query(path->query_id, path->query);
390 spin_unlock_irqrestore(&priv->lock, flags);
391 netif_tx_unlock_bh(dev);
392 wait_for_completion(&path->done);
393 path_free(dev, path);
394 netif_tx_lock_bh(dev);
395 spin_lock_irqsave(&priv->lock, flags);
396 }
397
398 spin_unlock_irqrestore(&priv->lock, flags);
399 netif_tx_unlock_bh(dev);
400 }
401
402 static void path_rec_completion(int status,
403 struct ib_sa_path_rec *pathrec,
404 void *path_ptr)
405 {
406 struct ipoib_path *path = path_ptr;
407 struct net_device *dev = path->dev;
408 struct ipoib_dev_priv *priv = netdev_priv(dev);
409 struct ipoib_ah *ah = NULL;
410 struct ipoib_ah *old_ah = NULL;
411 struct ipoib_neigh *neigh, *tn;
412 struct sk_buff_head skqueue;
413 struct sk_buff *skb;
414 unsigned long flags;
415
416 if (!status)
417 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
418 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
419 else
420 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
421 status, path->pathrec.dgid.raw);
422
423 skb_queue_head_init(&skqueue);
424
425 if (!status) {
426 struct ib_ah_attr av;
427
428 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
429 ah = ipoib_create_ah(dev, priv->pd, &av);
430 }
431
432 spin_lock_irqsave(&priv->lock, flags);
433
434 if (ah) {
435 path->pathrec = *pathrec;
436
437 old_ah = path->ah;
438 path->ah = ah;
439
440 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
441 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
442
443 while ((skb = __skb_dequeue(&path->queue)))
444 __skb_queue_tail(&skqueue, skb);
445
446 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
447 if (neigh->ah) {
448 WARN_ON(neigh->ah != old_ah);
449 /*
450 * Dropping the ah reference inside
451 * priv->lock is safe here, because we
452 * will hold one more reference from
453 * the original value of path->ah (ie
454 * old_ah).
455 */
456 ipoib_put_ah(neigh->ah);
457 }
458 kref_get(&path->ah->ref);
459 neigh->ah = path->ah;
460 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
461 sizeof(union ib_gid));
462
463 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
464 if (!ipoib_cm_get(neigh))
465 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
466 path,
467 neigh));
468 if (!ipoib_cm_get(neigh)) {
469 list_del(&neigh->list);
470 if (neigh->ah)
471 ipoib_put_ah(neigh->ah);
472 ipoib_neigh_free(dev, neigh);
473 continue;
474 }
475 }
476
477 while ((skb = __skb_dequeue(&neigh->queue)))
478 __skb_queue_tail(&skqueue, skb);
479 }
480 path->valid = 1;
481 }
482
483 path->query = NULL;
484 complete(&path->done);
485
486 spin_unlock_irqrestore(&priv->lock, flags);
487
488 if (old_ah)
489 ipoib_put_ah(old_ah);
490
491 while ((skb = __skb_dequeue(&skqueue))) {
492 skb->dev = dev;
493 if (dev_queue_xmit(skb))
494 ipoib_warn(priv, "dev_queue_xmit failed "
495 "to requeue packet\n");
496 }
497 }
498
499 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
500 {
501 struct ipoib_dev_priv *priv = netdev_priv(dev);
502 struct ipoib_path *path;
503
504 if (!priv->broadcast)
505 return NULL;
506
507 path = kzalloc(sizeof *path, GFP_ATOMIC);
508 if (!path)
509 return NULL;
510
511 path->dev = dev;
512
513 skb_queue_head_init(&path->queue);
514
515 INIT_LIST_HEAD(&path->neigh_list);
516
517 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
518 path->pathrec.sgid = priv->local_gid;
519 path->pathrec.pkey = cpu_to_be16(priv->pkey);
520 path->pathrec.numb_path = 1;
521 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
522
523 return path;
524 }
525
526 static int path_rec_start(struct net_device *dev,
527 struct ipoib_path *path)
528 {
529 struct ipoib_dev_priv *priv = netdev_priv(dev);
530
531 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
532 path->pathrec.dgid.raw);
533
534 init_completion(&path->done);
535
536 path->query_id =
537 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
538 &path->pathrec,
539 IB_SA_PATH_REC_DGID |
540 IB_SA_PATH_REC_SGID |
541 IB_SA_PATH_REC_NUMB_PATH |
542 IB_SA_PATH_REC_TRAFFIC_CLASS |
543 IB_SA_PATH_REC_PKEY,
544 1000, GFP_ATOMIC,
545 path_rec_completion,
546 path, &path->query);
547 if (path->query_id < 0) {
548 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
549 path->query = NULL;
550 complete(&path->done);
551 return path->query_id;
552 }
553
554 return 0;
555 }
556
557 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
558 {
559 struct ipoib_dev_priv *priv = netdev_priv(dev);
560 struct ipoib_path *path;
561 struct ipoib_neigh *neigh;
562 unsigned long flags;
563
564 neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour, skb->dev);
565 if (!neigh) {
566 ++dev->stats.tx_dropped;
567 dev_kfree_skb_any(skb);
568 return;
569 }
570
571 spin_lock_irqsave(&priv->lock, flags);
572
573 path = __path_find(dev, skb_dst(skb)->neighbour->ha + 4);
574 if (!path) {
575 path = path_rec_create(dev, skb_dst(skb)->neighbour->ha + 4);
576 if (!path)
577 goto err_path;
578
579 __path_add(dev, path);
580 }
581
582 list_add_tail(&neigh->list, &path->neigh_list);
583
584 if (path->ah) {
585 kref_get(&path->ah->ref);
586 neigh->ah = path->ah;
587 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
588 sizeof(union ib_gid));
589
590 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
591 if (!ipoib_cm_get(neigh))
592 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
593 if (!ipoib_cm_get(neigh)) {
594 list_del(&neigh->list);
595 if (neigh->ah)
596 ipoib_put_ah(neigh->ah);
597 ipoib_neigh_free(dev, neigh);
598 goto err_drop;
599 }
600 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
601 __skb_queue_tail(&neigh->queue, skb);
602 else {
603 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
604 skb_queue_len(&neigh->queue));
605 goto err_drop;
606 }
607 } else
608 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
609 } else {
610 neigh->ah = NULL;
611
612 if (!path->query && path_rec_start(dev, path))
613 goto err_list;
614
615 __skb_queue_tail(&neigh->queue, skb);
616 }
617
618 spin_unlock_irqrestore(&priv->lock, flags);
619 return;
620
621 err_list:
622 list_del(&neigh->list);
623
624 err_path:
625 ipoib_neigh_free(dev, neigh);
626 err_drop:
627 ++dev->stats.tx_dropped;
628 dev_kfree_skb_any(skb);
629
630 spin_unlock_irqrestore(&priv->lock, flags);
631 }
632
633 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
634 {
635 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
636
637 /* Look up path record for unicasts */
638 if (skb_dst(skb)->neighbour->ha[4] != 0xff) {
639 neigh_add_path(skb, dev);
640 return;
641 }
642
643 /* Add in the P_Key for multicasts */
644 skb_dst(skb)->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
645 skb_dst(skb)->neighbour->ha[9] = priv->pkey & 0xff;
646 ipoib_mcast_send(dev, skb_dst(skb)->neighbour->ha + 4, skb);
647 }
648
649 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
650 struct ipoib_pseudoheader *phdr)
651 {
652 struct ipoib_dev_priv *priv = netdev_priv(dev);
653 struct ipoib_path *path;
654 unsigned long flags;
655
656 spin_lock_irqsave(&priv->lock, flags);
657
658 path = __path_find(dev, phdr->hwaddr + 4);
659 if (!path || !path->valid) {
660 int new_path = 0;
661
662 if (!path) {
663 path = path_rec_create(dev, phdr->hwaddr + 4);
664 new_path = 1;
665 }
666 if (path) {
667 /* put pseudoheader back on for next time */
668 skb_push(skb, sizeof *phdr);
669 __skb_queue_tail(&path->queue, skb);
670
671 if (!path->query && path_rec_start(dev, path)) {
672 spin_unlock_irqrestore(&priv->lock, flags);
673 if (new_path)
674 path_free(dev, path);
675 return;
676 } else
677 __path_add(dev, path);
678 } else {
679 ++dev->stats.tx_dropped;
680 dev_kfree_skb_any(skb);
681 }
682
683 spin_unlock_irqrestore(&priv->lock, flags);
684 return;
685 }
686
687 if (path->ah) {
688 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
689 be16_to_cpu(path->pathrec.dlid));
690
691 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
692 } else if ((path->query || !path_rec_start(dev, path)) &&
693 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
694 /* put pseudoheader back on for next time */
695 skb_push(skb, sizeof *phdr);
696 __skb_queue_tail(&path->queue, skb);
697 } else {
698 ++dev->stats.tx_dropped;
699 dev_kfree_skb_any(skb);
700 }
701
702 spin_unlock_irqrestore(&priv->lock, flags);
703 }
704
705 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
706 {
707 struct ipoib_dev_priv *priv = netdev_priv(dev);
708 struct ipoib_neigh *neigh;
709 unsigned long flags;
710
711 if (likely(skb_dst(skb) && skb_dst(skb)->neighbour)) {
712 if (unlikely(!*to_ipoib_neigh(skb_dst(skb)->neighbour))) {
713 ipoib_path_lookup(skb, dev);
714 return NETDEV_TX_OK;
715 }
716
717 neigh = *to_ipoib_neigh(skb_dst(skb)->neighbour);
718
719 if (unlikely((memcmp(&neigh->dgid.raw,
720 skb_dst(skb)->neighbour->ha + 4,
721 sizeof(union ib_gid))) ||
722 (neigh->dev != dev))) {
723 spin_lock_irqsave(&priv->lock, flags);
724 /*
725 * It's safe to call ipoib_put_ah() inside
726 * priv->lock here, because we know that
727 * path->ah will always hold one more reference,
728 * so ipoib_put_ah() will never do more than
729 * decrement the ref count.
730 */
731 if (neigh->ah)
732 ipoib_put_ah(neigh->ah);
733 list_del(&neigh->list);
734 ipoib_neigh_free(dev, neigh);
735 spin_unlock_irqrestore(&priv->lock, flags);
736 ipoib_path_lookup(skb, dev);
737 return NETDEV_TX_OK;
738 }
739
740 if (ipoib_cm_get(neigh)) {
741 if (ipoib_cm_up(neigh)) {
742 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
743 return NETDEV_TX_OK;
744 }
745 } else if (neigh->ah) {
746 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
747 return NETDEV_TX_OK;
748 }
749
750 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
751 spin_lock_irqsave(&priv->lock, flags);
752 __skb_queue_tail(&neigh->queue, skb);
753 spin_unlock_irqrestore(&priv->lock, flags);
754 } else {
755 ++dev->stats.tx_dropped;
756 dev_kfree_skb_any(skb);
757 }
758 } else {
759 struct ipoib_pseudoheader *phdr =
760 (struct ipoib_pseudoheader *) skb->data;
761 skb_pull(skb, sizeof *phdr);
762
763 if (phdr->hwaddr[4] == 0xff) {
764 /* Add in the P_Key for multicast*/
765 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
766 phdr->hwaddr[9] = priv->pkey & 0xff;
767
768 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
769 } else {
770 /* unicast GID -- should be ARP or RARP reply */
771
772 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
773 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
774 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
775 skb_dst(skb) ? "neigh" : "dst",
776 be16_to_cpup((__be16 *) skb->data),
777 IPOIB_QPN(phdr->hwaddr),
778 phdr->hwaddr + 4);
779 dev_kfree_skb_any(skb);
780 ++dev->stats.tx_dropped;
781 return NETDEV_TX_OK;
782 }
783
784 unicast_arp_send(skb, dev, phdr);
785 }
786 }
787
788 return NETDEV_TX_OK;
789 }
790
791 static void ipoib_timeout(struct net_device *dev)
792 {
793 struct ipoib_dev_priv *priv = netdev_priv(dev);
794
795 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
796 jiffies_to_msecs(jiffies - dev->trans_start));
797 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
798 netif_queue_stopped(dev),
799 priv->tx_head, priv->tx_tail);
800 /* XXX reset QP, etc. */
801 }
802
803 static int ipoib_hard_header(struct sk_buff *skb,
804 struct net_device *dev,
805 unsigned short type,
806 const void *daddr, const void *saddr, unsigned len)
807 {
808 struct ipoib_header *header;
809
810 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
811
812 header->proto = htons(type);
813 header->reserved = 0;
814
815 /*
816 * If we don't have a neighbour structure, stuff the
817 * destination address onto the front of the skb so we can
818 * figure out where to send the packet later.
819 */
820 if ((!skb_dst(skb) || !skb_dst(skb)->neighbour) && daddr) {
821 struct ipoib_pseudoheader *phdr =
822 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
823 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
824 }
825
826 return 0;
827 }
828
829 static void ipoib_set_mcast_list(struct net_device *dev)
830 {
831 struct ipoib_dev_priv *priv = netdev_priv(dev);
832
833 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
834 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
835 return;
836 }
837
838 queue_work(ipoib_workqueue, &priv->restart_task);
839 }
840
841 static void ipoib_neigh_cleanup(struct neighbour *n)
842 {
843 struct ipoib_neigh *neigh;
844 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
845 unsigned long flags;
846 struct ipoib_ah *ah = NULL;
847
848 neigh = *to_ipoib_neigh(n);
849 if (neigh)
850 priv = netdev_priv(neigh->dev);
851 else
852 return;
853 ipoib_dbg(priv,
854 "neigh_cleanup for %06x %pI6\n",
855 IPOIB_QPN(n->ha),
856 n->ha + 4);
857
858 spin_lock_irqsave(&priv->lock, flags);
859
860 if (neigh->ah)
861 ah = neigh->ah;
862 list_del(&neigh->list);
863 ipoib_neigh_free(n->dev, neigh);
864
865 spin_unlock_irqrestore(&priv->lock, flags);
866
867 if (ah)
868 ipoib_put_ah(ah);
869 }
870
871 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
872 struct net_device *dev)
873 {
874 struct ipoib_neigh *neigh;
875
876 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
877 if (!neigh)
878 return NULL;
879
880 neigh->neighbour = neighbour;
881 neigh->dev = dev;
882 *to_ipoib_neigh(neighbour) = neigh;
883 skb_queue_head_init(&neigh->queue);
884 ipoib_cm_set(neigh, NULL);
885
886 return neigh;
887 }
888
889 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
890 {
891 struct sk_buff *skb;
892 *to_ipoib_neigh(neigh->neighbour) = NULL;
893 while ((skb = __skb_dequeue(&neigh->queue))) {
894 ++dev->stats.tx_dropped;
895 dev_kfree_skb_any(skb);
896 }
897 if (ipoib_cm_get(neigh))
898 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
899 kfree(neigh);
900 }
901
902 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
903 {
904 parms->neigh_cleanup = ipoib_neigh_cleanup;
905
906 return 0;
907 }
908
909 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
910 {
911 struct ipoib_dev_priv *priv = netdev_priv(dev);
912
913 /* Allocate RX/TX "rings" to hold queued skbs */
914 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
915 GFP_KERNEL);
916 if (!priv->rx_ring) {
917 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
918 ca->name, ipoib_recvq_size);
919 goto out;
920 }
921
922 priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
923 if (!priv->tx_ring) {
924 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
925 ca->name, ipoib_sendq_size);
926 goto out_rx_ring_cleanup;
927 }
928 memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
929
930 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
931
932 if (ipoib_ib_dev_init(dev, ca, port))
933 goto out_tx_ring_cleanup;
934
935 return 0;
936
937 out_tx_ring_cleanup:
938 vfree(priv->tx_ring);
939
940 out_rx_ring_cleanup:
941 kfree(priv->rx_ring);
942
943 out:
944 return -ENOMEM;
945 }
946
947 void ipoib_dev_cleanup(struct net_device *dev)
948 {
949 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
950
951 ipoib_delete_debug_files(dev);
952
953 /* Delete any child interfaces first */
954 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
955 unregister_netdev(cpriv->dev);
956 ipoib_dev_cleanup(cpriv->dev);
957 free_netdev(cpriv->dev);
958 }
959
960 ipoib_ib_dev_cleanup(dev);
961
962 kfree(priv->rx_ring);
963 vfree(priv->tx_ring);
964
965 priv->rx_ring = NULL;
966 priv->tx_ring = NULL;
967 }
968
969 static const struct header_ops ipoib_header_ops = {
970 .create = ipoib_hard_header,
971 };
972
973 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
974 void **tcph, u64 *hdr_flags, void *priv)
975 {
976 unsigned int ip_len;
977 struct iphdr *iph;
978
979 if (unlikely(skb->protocol != htons(ETH_P_IP)))
980 return -1;
981
982 /*
983 * In the future we may add an else clause that verifies the
984 * checksum and allows devices which do not calculate checksum
985 * to use LRO.
986 */
987 if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
988 return -1;
989
990 /* Check for non-TCP packet */
991 skb_reset_network_header(skb);
992 iph = ip_hdr(skb);
993 if (iph->protocol != IPPROTO_TCP)
994 return -1;
995
996 ip_len = ip_hdrlen(skb);
997 skb_set_transport_header(skb, ip_len);
998 *tcph = tcp_hdr(skb);
999
1000 /* check if IP header and TCP header are complete */
1001 if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
1002 return -1;
1003
1004 *hdr_flags = LRO_IPV4 | LRO_TCP;
1005 *iphdr = iph;
1006
1007 return 0;
1008 }
1009
1010 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1011 {
1012 priv->lro.lro_mgr.max_aggr = lro_max_aggr;
1013 priv->lro.lro_mgr.max_desc = IPOIB_MAX_LRO_DESCRIPTORS;
1014 priv->lro.lro_mgr.lro_arr = priv->lro.lro_desc;
1015 priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1016 priv->lro.lro_mgr.features = LRO_F_NAPI;
1017 priv->lro.lro_mgr.dev = priv->dev;
1018 priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1019 }
1020
1021 static const struct net_device_ops ipoib_netdev_ops = {
1022 .ndo_open = ipoib_open,
1023 .ndo_stop = ipoib_stop,
1024 .ndo_change_mtu = ipoib_change_mtu,
1025 .ndo_start_xmit = ipoib_start_xmit,
1026 .ndo_tx_timeout = ipoib_timeout,
1027 .ndo_set_multicast_list = ipoib_set_mcast_list,
1028 .ndo_neigh_setup = ipoib_neigh_setup_dev,
1029 };
1030
1031 static void ipoib_setup(struct net_device *dev)
1032 {
1033 struct ipoib_dev_priv *priv = netdev_priv(dev);
1034
1035 dev->netdev_ops = &ipoib_netdev_ops;
1036 dev->header_ops = &ipoib_header_ops;
1037
1038 ipoib_set_ethtool_ops(dev);
1039
1040 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1041
1042 dev->watchdog_timeo = HZ;
1043
1044 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1045
1046 /*
1047 * We add in INFINIBAND_ALEN to allow for the destination
1048 * address "pseudoheader" for skbs without neighbour struct.
1049 */
1050 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1051 dev->addr_len = INFINIBAND_ALEN;
1052 dev->type = ARPHRD_INFINIBAND;
1053 dev->tx_queue_len = ipoib_sendq_size * 2;
1054 dev->features = (NETIF_F_VLAN_CHALLENGED |
1055 NETIF_F_HIGHDMA);
1056 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
1057
1058 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1059
1060 netif_carrier_off(dev);
1061
1062 priv->dev = dev;
1063
1064 ipoib_lro_setup(priv);
1065
1066 spin_lock_init(&priv->lock);
1067
1068 mutex_init(&priv->vlan_mutex);
1069
1070 INIT_LIST_HEAD(&priv->path_list);
1071 INIT_LIST_HEAD(&priv->child_intfs);
1072 INIT_LIST_HEAD(&priv->dead_ahs);
1073 INIT_LIST_HEAD(&priv->multicast_list);
1074
1075 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1076 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1077 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1078 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1079 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1080 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1081 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1082 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1083 }
1084
1085 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1086 {
1087 struct net_device *dev;
1088
1089 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1090 ipoib_setup);
1091 if (!dev)
1092 return NULL;
1093
1094 return netdev_priv(dev);
1095 }
1096
1097 static ssize_t show_pkey(struct device *dev,
1098 struct device_attribute *attr, char *buf)
1099 {
1100 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1101
1102 return sprintf(buf, "0x%04x\n", priv->pkey);
1103 }
1104 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1105
1106 static ssize_t show_umcast(struct device *dev,
1107 struct device_attribute *attr, char *buf)
1108 {
1109 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1110
1111 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1112 }
1113
1114 static ssize_t set_umcast(struct device *dev,
1115 struct device_attribute *attr,
1116 const char *buf, size_t count)
1117 {
1118 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1119 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1120
1121 if (umcast_val > 0) {
1122 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1123 ipoib_warn(priv, "ignoring multicast groups joined directly "
1124 "by userspace\n");
1125 } else
1126 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1127
1128 return count;
1129 }
1130 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1131
1132 int ipoib_add_umcast_attr(struct net_device *dev)
1133 {
1134 return device_create_file(&dev->dev, &dev_attr_umcast);
1135 }
1136
1137 static ssize_t create_child(struct device *dev,
1138 struct device_attribute *attr,
1139 const char *buf, size_t count)
1140 {
1141 int pkey;
1142 int ret;
1143
1144 if (sscanf(buf, "%i", &pkey) != 1)
1145 return -EINVAL;
1146
1147 if (pkey < 0 || pkey > 0xffff)
1148 return -EINVAL;
1149
1150 /*
1151 * Set the full membership bit, so that we join the right
1152 * broadcast group, etc.
1153 */
1154 pkey |= 0x8000;
1155
1156 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1157
1158 return ret ? ret : count;
1159 }
1160 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1161
1162 static ssize_t delete_child(struct device *dev,
1163 struct device_attribute *attr,
1164 const char *buf, size_t count)
1165 {
1166 int pkey;
1167 int ret;
1168
1169 if (sscanf(buf, "%i", &pkey) != 1)
1170 return -EINVAL;
1171
1172 if (pkey < 0 || pkey > 0xffff)
1173 return -EINVAL;
1174
1175 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1176
1177 return ret ? ret : count;
1178
1179 }
1180 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1181
1182 int ipoib_add_pkey_attr(struct net_device *dev)
1183 {
1184 return device_create_file(&dev->dev, &dev_attr_pkey);
1185 }
1186
1187 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1188 {
1189 struct ib_device_attr *device_attr;
1190 int result = -ENOMEM;
1191
1192 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1193 if (!device_attr) {
1194 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1195 hca->name, sizeof *device_attr);
1196 return result;
1197 }
1198
1199 result = ib_query_device(hca, device_attr);
1200 if (result) {
1201 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1202 hca->name, result);
1203 kfree(device_attr);
1204 return result;
1205 }
1206 priv->hca_caps = device_attr->device_cap_flags;
1207
1208 kfree(device_attr);
1209
1210 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1211 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1212 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1213 }
1214
1215 if (lro)
1216 priv->dev->features |= NETIF_F_LRO;
1217
1218 if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1219 priv->dev->features |= NETIF_F_TSO;
1220
1221 return 0;
1222 }
1223
1224
1225 static struct net_device *ipoib_add_port(const char *format,
1226 struct ib_device *hca, u8 port)
1227 {
1228 struct ipoib_dev_priv *priv;
1229 struct ib_port_attr attr;
1230 int result = -ENOMEM;
1231
1232 priv = ipoib_intf_alloc(format);
1233 if (!priv)
1234 goto alloc_mem_failed;
1235
1236 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1237
1238 if (!ib_query_port(hca, port, &attr))
1239 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1240 else {
1241 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1242 hca->name, port);
1243 goto device_init_failed;
1244 }
1245
1246 /* MTU will be reset when mcast join happens */
1247 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1248 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1249
1250 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1251 if (result) {
1252 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1253 hca->name, port, result);
1254 goto device_init_failed;
1255 }
1256
1257 if (ipoib_set_dev_features(priv, hca))
1258 goto device_init_failed;
1259
1260 /*
1261 * Set the full membership bit, so that we join the right
1262 * broadcast group, etc.
1263 */
1264 priv->pkey |= 0x8000;
1265
1266 priv->dev->broadcast[8] = priv->pkey >> 8;
1267 priv->dev->broadcast[9] = priv->pkey & 0xff;
1268
1269 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1270 if (result) {
1271 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1272 hca->name, port, result);
1273 goto device_init_failed;
1274 } else
1275 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1276
1277 result = ipoib_dev_init(priv->dev, hca, port);
1278 if (result < 0) {
1279 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1280 hca->name, port, result);
1281 goto device_init_failed;
1282 }
1283
1284 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1285 priv->ca, ipoib_event);
1286 result = ib_register_event_handler(&priv->event_handler);
1287 if (result < 0) {
1288 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1289 "port %d (ret = %d)\n",
1290 hca->name, port, result);
1291 goto event_failed;
1292 }
1293
1294 result = register_netdev(priv->dev);
1295 if (result) {
1296 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1297 hca->name, port, result);
1298 goto register_failed;
1299 }
1300
1301 ipoib_create_debug_files(priv->dev);
1302
1303 if (ipoib_cm_add_mode_attr(priv->dev))
1304 goto sysfs_failed;
1305 if (ipoib_add_pkey_attr(priv->dev))
1306 goto sysfs_failed;
1307 if (ipoib_add_umcast_attr(priv->dev))
1308 goto sysfs_failed;
1309 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1310 goto sysfs_failed;
1311 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1312 goto sysfs_failed;
1313
1314 return priv->dev;
1315
1316 sysfs_failed:
1317 ipoib_delete_debug_files(priv->dev);
1318 unregister_netdev(priv->dev);
1319
1320 register_failed:
1321 ib_unregister_event_handler(&priv->event_handler);
1322 flush_workqueue(ipoib_workqueue);
1323
1324 event_failed:
1325 ipoib_dev_cleanup(priv->dev);
1326
1327 device_init_failed:
1328 free_netdev(priv->dev);
1329
1330 alloc_mem_failed:
1331 return ERR_PTR(result);
1332 }
1333
1334 static void ipoib_add_one(struct ib_device *device)
1335 {
1336 struct list_head *dev_list;
1337 struct net_device *dev;
1338 struct ipoib_dev_priv *priv;
1339 int s, e, p;
1340
1341 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1342 return;
1343
1344 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1345 if (!dev_list)
1346 return;
1347
1348 INIT_LIST_HEAD(dev_list);
1349
1350 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1351 s = 0;
1352 e = 0;
1353 } else {
1354 s = 1;
1355 e = device->phys_port_cnt;
1356 }
1357
1358 for (p = s; p <= e; ++p) {
1359 dev = ipoib_add_port("ib%d", device, p);
1360 if (!IS_ERR(dev)) {
1361 priv = netdev_priv(dev);
1362 list_add_tail(&priv->list, dev_list);
1363 }
1364 }
1365
1366 ib_set_client_data(device, &ipoib_client, dev_list);
1367 }
1368
1369 static void ipoib_remove_one(struct ib_device *device)
1370 {
1371 struct ipoib_dev_priv *priv, *tmp;
1372 struct list_head *dev_list;
1373
1374 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1375 return;
1376
1377 dev_list = ib_get_client_data(device, &ipoib_client);
1378
1379 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1380 ib_unregister_event_handler(&priv->event_handler);
1381
1382 rtnl_lock();
1383 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1384 rtnl_unlock();
1385
1386 flush_workqueue(ipoib_workqueue);
1387
1388 unregister_netdev(priv->dev);
1389 ipoib_dev_cleanup(priv->dev);
1390 free_netdev(priv->dev);
1391 }
1392
1393 kfree(dev_list);
1394 }
1395
1396 static int __init ipoib_init_module(void)
1397 {
1398 int ret;
1399
1400 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1401 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1402 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1403
1404 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1405 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1406 ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1407 IPOIB_MIN_QUEUE_SIZE));
1408 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1409 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1410 #endif
1411
1412 /*
1413 * When copying small received packets, we only copy from the
1414 * linear data part of the SKB, so we rely on this condition.
1415 */
1416 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1417
1418 ret = ipoib_register_debugfs();
1419 if (ret)
1420 return ret;
1421
1422 /*
1423 * We create our own workqueue mainly because we want to be
1424 * able to flush it when devices are being removed. We can't
1425 * use schedule_work()/flush_scheduled_work() because both
1426 * unregister_netdev() and linkwatch_event take the rtnl lock,
1427 * so flush_scheduled_work() can deadlock during device
1428 * removal.
1429 */
1430 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1431 if (!ipoib_workqueue) {
1432 ret = -ENOMEM;
1433 goto err_fs;
1434 }
1435
1436 ib_sa_register_client(&ipoib_sa_client);
1437
1438 ret = ib_register_client(&ipoib_client);
1439 if (ret)
1440 goto err_sa;
1441
1442 return 0;
1443
1444 err_sa:
1445 ib_sa_unregister_client(&ipoib_sa_client);
1446 destroy_workqueue(ipoib_workqueue);
1447
1448 err_fs:
1449 ipoib_unregister_debugfs();
1450
1451 return ret;
1452 }
1453
1454 static void __exit ipoib_cleanup_module(void)
1455 {
1456 ib_unregister_client(&ipoib_client);
1457 ib_sa_unregister_client(&ipoib_sa_client);
1458 ipoib_unregister_debugfs();
1459 destroy_workqueue(ipoib_workqueue);
1460 }
1461
1462 module_init(ipoib_init_module);
1463 module_exit(ipoib_cleanup_module);
Linux with added FPC-III support