Loading...
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 <linux/jhash.h>
50#include <net/arp.h>
51#include <net/addrconf.h>
52#include <linux/inetdevice.h>
53#include <rdma/ib_cache.h>
54
55MODULE_AUTHOR("Roland Dreier");
56MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
57MODULE_LICENSE("Dual BSD/GPL");
58
59int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
60int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
61
62module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
63MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
64module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
65MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
66
67#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
68int ipoib_debug_level;
69
70module_param_named(debug_level, ipoib_debug_level, int, 0644);
71MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
72#endif
73
74struct ipoib_path_iter {
75 struct net_device *dev;
76 struct ipoib_path path;
77};
78
79static const u8 ipv4_bcast_addr[] = {
80 0x00, 0xff, 0xff, 0xff,
81 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
82 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
83};
84
85struct workqueue_struct *ipoib_workqueue;
86
87struct ib_sa_client ipoib_sa_client;
88
89static int ipoib_add_one(struct ib_device *device);
90static void ipoib_remove_one(struct ib_device *device, void *client_data);
91static void ipoib_neigh_reclaim(struct rcu_head *rp);
92static struct net_device *ipoib_get_net_dev_by_params(
93 struct ib_device *dev, u32 port, u16 pkey,
94 const union ib_gid *gid, const struct sockaddr *addr,
95 void *client_data);
96static int ipoib_set_mac(struct net_device *dev, void *addr);
97static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
98 int cmd);
99
100static struct ib_client ipoib_client = {
101 .name = "ipoib",
102 .add = ipoib_add_one,
103 .remove = ipoib_remove_one,
104 .get_net_dev_by_params = ipoib_get_net_dev_by_params,
105};
106
107#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
108static int ipoib_netdev_event(struct notifier_block *this,
109 unsigned long event, void *ptr)
110{
111 struct netdev_notifier_info *ni = ptr;
112 struct net_device *dev = ni->dev;
113
114 if (dev->netdev_ops->ndo_open != ipoib_open)
115 return NOTIFY_DONE;
116
117 switch (event) {
118 case NETDEV_REGISTER:
119 ipoib_create_debug_files(dev);
120 break;
121 case NETDEV_CHANGENAME:
122 ipoib_delete_debug_files(dev);
123 ipoib_create_debug_files(dev);
124 break;
125 case NETDEV_UNREGISTER:
126 ipoib_delete_debug_files(dev);
127 break;
128 }
129
130 return NOTIFY_DONE;
131}
132#endif
133
134int ipoib_open(struct net_device *dev)
135{
136 struct ipoib_dev_priv *priv = ipoib_priv(dev);
137
138 ipoib_dbg(priv, "bringing up interface\n");
139
140 netif_carrier_off(dev);
141
142 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
143
144 if (ipoib_ib_dev_open(dev)) {
145 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
146 return 0;
147 goto err_disable;
148 }
149
150 ipoib_ib_dev_up(dev);
151
152 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
153 struct ipoib_dev_priv *cpriv;
154
155 /* Bring up any child interfaces too */
156 down_read(&priv->vlan_rwsem);
157 list_for_each_entry(cpriv, &priv->child_intfs, list) {
158 int flags;
159
160 flags = cpriv->dev->flags;
161 if (flags & IFF_UP)
162 continue;
163
164 dev_change_flags(cpriv->dev, flags | IFF_UP, NULL);
165 }
166 up_read(&priv->vlan_rwsem);
167 } else if (priv->parent) {
168 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
169
170 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &ppriv->flags))
171 ipoib_dbg(priv, "parent device %s is not up, so child device may be not functioning.\n",
172 ppriv->dev->name);
173 }
174 netif_start_queue(dev);
175
176 return 0;
177
178err_disable:
179 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
180
181 return -EINVAL;
182}
183
184static int ipoib_stop(struct net_device *dev)
185{
186 struct ipoib_dev_priv *priv = ipoib_priv(dev);
187
188 ipoib_dbg(priv, "stopping interface\n");
189
190 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
191
192 netif_stop_queue(dev);
193
194 ipoib_ib_dev_down(dev);
195 ipoib_ib_dev_stop(dev);
196
197 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
198 struct ipoib_dev_priv *cpriv;
199
200 /* Bring down any child interfaces too */
201 down_read(&priv->vlan_rwsem);
202 list_for_each_entry(cpriv, &priv->child_intfs, list) {
203 int flags;
204
205 flags = cpriv->dev->flags;
206 if (!(flags & IFF_UP))
207 continue;
208
209 dev_change_flags(cpriv->dev, flags & ~IFF_UP, NULL);
210 }
211 up_read(&priv->vlan_rwsem);
212 }
213
214 return 0;
215}
216
217static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
218{
219 struct ipoib_dev_priv *priv = ipoib_priv(dev);
220
221 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
222 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
223
224 return features;
225}
226
227static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
228{
229 struct ipoib_dev_priv *priv = ipoib_priv(dev);
230 int ret = 0;
231
232 /* dev->mtu > 2K ==> connected mode */
233 if (ipoib_cm_admin_enabled(dev)) {
234 if (new_mtu > ipoib_cm_max_mtu(dev))
235 return -EINVAL;
236
237 if (new_mtu > priv->mcast_mtu)
238 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
239 priv->mcast_mtu);
240
241 dev->mtu = new_mtu;
242 return 0;
243 }
244
245 if (new_mtu < (ETH_MIN_MTU + IPOIB_ENCAP_LEN) ||
246 new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
247 return -EINVAL;
248
249 priv->admin_mtu = new_mtu;
250
251 if (priv->mcast_mtu < priv->admin_mtu)
252 ipoib_dbg(priv, "MTU must be smaller than the underlying "
253 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
254
255 new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
256
257 if (priv->rn_ops->ndo_change_mtu) {
258 bool carrier_status = netif_carrier_ok(dev);
259
260 netif_carrier_off(dev);
261
262 /* notify lower level on the real mtu */
263 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
264
265 if (carrier_status)
266 netif_carrier_on(dev);
267 } else {
268 dev->mtu = new_mtu;
269 }
270
271 return ret;
272}
273
274static void ipoib_get_stats(struct net_device *dev,
275 struct rtnl_link_stats64 *stats)
276{
277 struct ipoib_dev_priv *priv = ipoib_priv(dev);
278
279 if (priv->rn_ops->ndo_get_stats64)
280 priv->rn_ops->ndo_get_stats64(dev, stats);
281 else
282 netdev_stats_to_stats64(stats, &dev->stats);
283}
284
285/* Called with an RCU read lock taken */
286static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
287 struct net_device *dev)
288{
289 struct net *net = dev_net(dev);
290 struct in_device *in_dev;
291 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
292 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
293 __be32 ret_addr;
294
295 switch (addr->sa_family) {
296 case AF_INET:
297 in_dev = in_dev_get(dev);
298 if (!in_dev)
299 return false;
300
301 ret_addr = inet_confirm_addr(net, in_dev, 0,
302 addr_in->sin_addr.s_addr,
303 RT_SCOPE_HOST);
304 in_dev_put(in_dev);
305 if (ret_addr)
306 return true;
307
308 break;
309 case AF_INET6:
310 if (IS_ENABLED(CONFIG_IPV6) &&
311 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
312 return true;
313
314 break;
315 }
316 return false;
317}
318
319/*
320 * Find the master net_device on top of the given net_device.
321 * @dev: base IPoIB net_device
322 *
323 * Returns the master net_device with a reference held, or the same net_device
324 * if no master exists.
325 */
326static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
327{
328 struct net_device *master;
329
330 rcu_read_lock();
331 master = netdev_master_upper_dev_get_rcu(dev);
332 if (master)
333 dev_hold(master);
334 rcu_read_unlock();
335
336 if (master)
337 return master;
338
339 dev_hold(dev);
340 return dev;
341}
342
343struct ipoib_walk_data {
344 const struct sockaddr *addr;
345 struct net_device *result;
346};
347
348static int ipoib_upper_walk(struct net_device *upper,
349 struct netdev_nested_priv *priv)
350{
351 struct ipoib_walk_data *data = (struct ipoib_walk_data *)priv->data;
352 int ret = 0;
353
354 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
355 dev_hold(upper);
356 data->result = upper;
357 ret = 1;
358 }
359
360 return ret;
361}
362
363/**
364 * ipoib_get_net_dev_match_addr - Find a net_device matching
365 * the given address, which is an upper device of the given net_device.
366 *
367 * @addr: IP address to look for.
368 * @dev: base IPoIB net_device
369 *
370 * If found, returns the net_device with a reference held. Otherwise return
371 * NULL.
372 */
373static struct net_device *ipoib_get_net_dev_match_addr(
374 const struct sockaddr *addr, struct net_device *dev)
375{
376 struct netdev_nested_priv priv;
377 struct ipoib_walk_data data = {
378 .addr = addr,
379 };
380
381 priv.data = (void *)&data;
382 rcu_read_lock();
383 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
384 dev_hold(dev);
385 data.result = dev;
386 goto out;
387 }
388
389 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &priv);
390out:
391 rcu_read_unlock();
392 return data.result;
393}
394
395/* returns the number of IPoIB netdevs on top a given ipoib device matching a
396 * pkey_index and address, if one exists.
397 *
398 * @found_net_dev: contains a matching net_device if the return value >= 1,
399 * with a reference held. */
400static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
401 const union ib_gid *gid,
402 u16 pkey_index,
403 const struct sockaddr *addr,
404 int nesting,
405 struct net_device **found_net_dev)
406{
407 struct ipoib_dev_priv *child_priv;
408 struct net_device *net_dev = NULL;
409 int matches = 0;
410
411 if (priv->pkey_index == pkey_index &&
412 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
413 if (!addr) {
414 net_dev = ipoib_get_master_net_dev(priv->dev);
415 } else {
416 /* Verify the net_device matches the IP address, as
417 * IPoIB child devices currently share a GID. */
418 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
419 }
420 if (net_dev) {
421 if (!*found_net_dev)
422 *found_net_dev = net_dev;
423 else
424 dev_put(net_dev);
425 ++matches;
426 }
427 }
428
429 /* Check child interfaces */
430 down_read_nested(&priv->vlan_rwsem, nesting);
431 list_for_each_entry(child_priv, &priv->child_intfs, list) {
432 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
433 pkey_index, addr,
434 nesting + 1,
435 found_net_dev);
436 if (matches > 1)
437 break;
438 }
439 up_read(&priv->vlan_rwsem);
440
441 return matches;
442}
443
444/* Returns the number of matching net_devs found (between 0 and 2). Also
445 * return the matching net_device in the @net_dev parameter, holding a
446 * reference to the net_device, if the number of matches >= 1 */
447static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u32 port,
448 u16 pkey_index,
449 const union ib_gid *gid,
450 const struct sockaddr *addr,
451 struct net_device **net_dev)
452{
453 struct ipoib_dev_priv *priv;
454 int matches = 0;
455
456 *net_dev = NULL;
457
458 list_for_each_entry(priv, dev_list, list) {
459 if (priv->port != port)
460 continue;
461
462 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
463 addr, 0, net_dev);
464 if (matches > 1)
465 break;
466 }
467
468 return matches;
469}
470
471static struct net_device *ipoib_get_net_dev_by_params(
472 struct ib_device *dev, u32 port, u16 pkey,
473 const union ib_gid *gid, const struct sockaddr *addr,
474 void *client_data)
475{
476 struct net_device *net_dev;
477 struct list_head *dev_list = client_data;
478 u16 pkey_index;
479 int matches;
480 int ret;
481
482 if (!rdma_protocol_ib(dev, port))
483 return NULL;
484
485 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
486 if (ret)
487 return NULL;
488
489 /* See if we can find a unique device matching the L2 parameters */
490 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
491 gid, NULL, &net_dev);
492
493 switch (matches) {
494 case 0:
495 return NULL;
496 case 1:
497 return net_dev;
498 }
499
500 dev_put(net_dev);
501
502 /* Couldn't find a unique device with L2 parameters only. Use L3
503 * address to uniquely match the net device */
504 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
505 gid, addr, &net_dev);
506 switch (matches) {
507 case 0:
508 return NULL;
509 default:
510 dev_warn_ratelimited(&dev->dev,
511 "duplicate IP address detected\n");
512 fallthrough;
513 case 1:
514 return net_dev;
515 }
516}
517
518int ipoib_set_mode(struct net_device *dev, const char *buf)
519{
520 struct ipoib_dev_priv *priv = ipoib_priv(dev);
521
522 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
523 !strcmp(buf, "connected\n")) ||
524 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
525 !strcmp(buf, "datagram\n"))) {
526 return 0;
527 }
528
529 /* flush paths if we switch modes so that connections are restarted */
530 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
531 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
532 ipoib_warn(priv, "enabling connected mode "
533 "will cause multicast packet drops\n");
534 netdev_update_features(dev);
535 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
536 netif_set_real_num_tx_queues(dev, 1);
537 rtnl_unlock();
538 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
539
540 ipoib_flush_paths(dev);
541 return (!rtnl_trylock()) ? -EBUSY : 0;
542 }
543
544 if (!strcmp(buf, "datagram\n")) {
545 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
546 netdev_update_features(dev);
547 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
548 netif_set_real_num_tx_queues(dev, dev->num_tx_queues);
549 rtnl_unlock();
550 ipoib_flush_paths(dev);
551 return (!rtnl_trylock()) ? -EBUSY : 0;
552 }
553
554 return -EINVAL;
555}
556
557struct ipoib_path *__path_find(struct net_device *dev, void *gid)
558{
559 struct ipoib_dev_priv *priv = ipoib_priv(dev);
560 struct rb_node *n = priv->path_tree.rb_node;
561 struct ipoib_path *path;
562 int ret;
563
564 while (n) {
565 path = rb_entry(n, struct ipoib_path, rb_node);
566
567 ret = memcmp(gid, path->pathrec.dgid.raw,
568 sizeof (union ib_gid));
569
570 if (ret < 0)
571 n = n->rb_left;
572 else if (ret > 0)
573 n = n->rb_right;
574 else
575 return path;
576 }
577
578 return NULL;
579}
580
581static int __path_add(struct net_device *dev, struct ipoib_path *path)
582{
583 struct ipoib_dev_priv *priv = ipoib_priv(dev);
584 struct rb_node **n = &priv->path_tree.rb_node;
585 struct rb_node *pn = NULL;
586 struct ipoib_path *tpath;
587 int ret;
588
589 while (*n) {
590 pn = *n;
591 tpath = rb_entry(pn, struct ipoib_path, rb_node);
592
593 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
594 sizeof (union ib_gid));
595 if (ret < 0)
596 n = &pn->rb_left;
597 else if (ret > 0)
598 n = &pn->rb_right;
599 else
600 return -EEXIST;
601 }
602
603 rb_link_node(&path->rb_node, pn, n);
604 rb_insert_color(&path->rb_node, &priv->path_tree);
605
606 list_add_tail(&path->list, &priv->path_list);
607
608 return 0;
609}
610
611static void path_free(struct net_device *dev, struct ipoib_path *path)
612{
613 struct sk_buff *skb;
614
615 while ((skb = __skb_dequeue(&path->queue)))
616 dev_kfree_skb_irq(skb);
617
618 ipoib_dbg(ipoib_priv(dev), "%s\n", __func__);
619
620 /* remove all neigh connected to this path */
621 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
622
623 if (path->ah)
624 ipoib_put_ah(path->ah);
625
626 kfree(path);
627}
628
629#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
630
631struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
632{
633 struct ipoib_path_iter *iter;
634
635 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
636 if (!iter)
637 return NULL;
638
639 iter->dev = dev;
640 memset(iter->path.pathrec.dgid.raw, 0, 16);
641
642 if (ipoib_path_iter_next(iter)) {
643 kfree(iter);
644 return NULL;
645 }
646
647 return iter;
648}
649
650int ipoib_path_iter_next(struct ipoib_path_iter *iter)
651{
652 struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
653 struct rb_node *n;
654 struct ipoib_path *path;
655 int ret = 1;
656
657 spin_lock_irq(&priv->lock);
658
659 n = rb_first(&priv->path_tree);
660
661 while (n) {
662 path = rb_entry(n, struct ipoib_path, rb_node);
663
664 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
665 sizeof (union ib_gid)) < 0) {
666 iter->path = *path;
667 ret = 0;
668 break;
669 }
670
671 n = rb_next(n);
672 }
673
674 spin_unlock_irq(&priv->lock);
675
676 return ret;
677}
678
679void ipoib_path_iter_read(struct ipoib_path_iter *iter,
680 struct ipoib_path *path)
681{
682 *path = iter->path;
683}
684
685#endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
686
687void ipoib_mark_paths_invalid(struct net_device *dev)
688{
689 struct ipoib_dev_priv *priv = ipoib_priv(dev);
690 struct ipoib_path *path, *tp;
691
692 spin_lock_irq(&priv->lock);
693
694 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
695 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
696 be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
697 path->pathrec.dgid.raw);
698 if (path->ah)
699 path->ah->valid = 0;
700 }
701
702 spin_unlock_irq(&priv->lock);
703}
704
705static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
706{
707 struct ipoib_pseudo_header *phdr;
708
709 phdr = skb_push(skb, sizeof(*phdr));
710 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
711}
712
713void ipoib_flush_paths(struct net_device *dev)
714{
715 struct ipoib_dev_priv *priv = ipoib_priv(dev);
716 struct ipoib_path *path, *tp;
717 LIST_HEAD(remove_list);
718 unsigned long flags;
719
720 netif_tx_lock_bh(dev);
721 spin_lock_irqsave(&priv->lock, flags);
722
723 list_splice_init(&priv->path_list, &remove_list);
724
725 list_for_each_entry(path, &remove_list, list)
726 rb_erase(&path->rb_node, &priv->path_tree);
727
728 list_for_each_entry_safe(path, tp, &remove_list, list) {
729 if (path->query)
730 ib_sa_cancel_query(path->query_id, path->query);
731 spin_unlock_irqrestore(&priv->lock, flags);
732 netif_tx_unlock_bh(dev);
733 wait_for_completion(&path->done);
734 path_free(dev, path);
735 netif_tx_lock_bh(dev);
736 spin_lock_irqsave(&priv->lock, flags);
737 }
738
739 spin_unlock_irqrestore(&priv->lock, flags);
740 netif_tx_unlock_bh(dev);
741}
742
743static void path_rec_completion(int status,
744 struct sa_path_rec *pathrec,
745 unsigned int num_prs, void *path_ptr)
746{
747 struct ipoib_path *path = path_ptr;
748 struct net_device *dev = path->dev;
749 struct ipoib_dev_priv *priv = ipoib_priv(dev);
750 struct ipoib_ah *ah = NULL;
751 struct ipoib_ah *old_ah = NULL;
752 struct ipoib_neigh *neigh, *tn;
753 struct sk_buff_head skqueue;
754 struct sk_buff *skb;
755 unsigned long flags;
756
757 if (!status)
758 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
759 be32_to_cpu(sa_path_get_dlid(pathrec)),
760 pathrec->dgid.raw);
761 else
762 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
763 status, path->pathrec.dgid.raw);
764
765 skb_queue_head_init(&skqueue);
766
767 if (!status) {
768 struct rdma_ah_attr av;
769
770 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
771 pathrec, &av, NULL)) {
772 ah = ipoib_create_ah(dev, priv->pd, &av);
773 rdma_destroy_ah_attr(&av);
774 }
775 }
776
777 spin_lock_irqsave(&priv->lock, flags);
778
779 if (!IS_ERR_OR_NULL(ah)) {
780 /*
781 * pathrec.dgid is used as the database key from the LLADDR,
782 * it must remain unchanged even if the SA returns a different
783 * GID to use in the AH.
784 */
785 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
786 sizeof(union ib_gid))) {
787 ipoib_dbg(
788 priv,
789 "%s got PathRec for gid %pI6 while asked for %pI6\n",
790 dev->name, pathrec->dgid.raw,
791 path->pathrec.dgid.raw);
792 memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
793 sizeof(union ib_gid));
794 }
795
796 path->pathrec = *pathrec;
797
798 old_ah = path->ah;
799 path->ah = ah;
800
801 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
802 ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
803 pathrec->sl);
804
805 while ((skb = __skb_dequeue(&path->queue)))
806 __skb_queue_tail(&skqueue, skb);
807
808 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
809 if (neigh->ah) {
810 WARN_ON(neigh->ah != old_ah);
811 /*
812 * Dropping the ah reference inside
813 * priv->lock is safe here, because we
814 * will hold one more reference from
815 * the original value of path->ah (ie
816 * old_ah).
817 */
818 ipoib_put_ah(neigh->ah);
819 }
820 kref_get(&path->ah->ref);
821 neigh->ah = path->ah;
822
823 if (ipoib_cm_enabled(dev, neigh->daddr)) {
824 if (!ipoib_cm_get(neigh))
825 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
826 path,
827 neigh));
828 if (!ipoib_cm_get(neigh)) {
829 ipoib_neigh_free(neigh);
830 continue;
831 }
832 }
833
834 while ((skb = __skb_dequeue(&neigh->queue)))
835 __skb_queue_tail(&skqueue, skb);
836 }
837 path->ah->valid = 1;
838 }
839
840 path->query = NULL;
841 complete(&path->done);
842
843 spin_unlock_irqrestore(&priv->lock, flags);
844
845 if (IS_ERR_OR_NULL(ah))
846 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
847
848 if (old_ah)
849 ipoib_put_ah(old_ah);
850
851 while ((skb = __skb_dequeue(&skqueue))) {
852 int ret;
853 skb->dev = dev;
854 ret = dev_queue_xmit(skb);
855 if (ret)
856 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
857 __func__, ret);
858 }
859}
860
861static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
862 void *gid)
863{
864 path->dev = priv->dev;
865
866 if (rdma_cap_opa_ah(priv->ca, priv->port))
867 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
868 else
869 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
870
871 memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
872 path->pathrec.sgid = priv->local_gid;
873 path->pathrec.pkey = cpu_to_be16(priv->pkey);
874 path->pathrec.numb_path = 1;
875 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
876}
877
878static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
879{
880 struct ipoib_dev_priv *priv = ipoib_priv(dev);
881 struct ipoib_path *path;
882
883 if (!priv->broadcast)
884 return NULL;
885
886 path = kzalloc(sizeof(*path), GFP_ATOMIC);
887 if (!path)
888 return NULL;
889
890 skb_queue_head_init(&path->queue);
891
892 INIT_LIST_HEAD(&path->neigh_list);
893
894 init_path_rec(priv, path, gid);
895
896 return path;
897}
898
899static int path_rec_start(struct net_device *dev,
900 struct ipoib_path *path)
901{
902 struct ipoib_dev_priv *priv = ipoib_priv(dev);
903
904 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
905 path->pathrec.dgid.raw);
906
907 init_completion(&path->done);
908
909 path->query_id =
910 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
911 &path->pathrec,
912 IB_SA_PATH_REC_DGID |
913 IB_SA_PATH_REC_SGID |
914 IB_SA_PATH_REC_NUMB_PATH |
915 IB_SA_PATH_REC_TRAFFIC_CLASS |
916 IB_SA_PATH_REC_PKEY,
917 1000, GFP_ATOMIC,
918 path_rec_completion,
919 path, &path->query);
920 if (path->query_id < 0) {
921 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
922 path->query = NULL;
923 complete(&path->done);
924 return path->query_id;
925 }
926
927 return 0;
928}
929
930static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
931 struct net_device *dev)
932{
933 struct ipoib_dev_priv *priv = ipoib_priv(dev);
934 struct ipoib_path *path;
935 unsigned long flags;
936
937 spin_lock_irqsave(&priv->lock, flags);
938
939 path = __path_find(dev, daddr + 4);
940 if (!path)
941 goto out;
942 if (!path->query)
943 path_rec_start(dev, path);
944out:
945 spin_unlock_irqrestore(&priv->lock, flags);
946}
947
948static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
949 struct net_device *dev)
950{
951 struct ipoib_dev_priv *priv = ipoib_priv(dev);
952 struct rdma_netdev *rn = netdev_priv(dev);
953 struct ipoib_path *path;
954 struct ipoib_neigh *neigh;
955 unsigned long flags;
956
957 spin_lock_irqsave(&priv->lock, flags);
958 neigh = ipoib_neigh_alloc(daddr, dev);
959 if (!neigh) {
960 spin_unlock_irqrestore(&priv->lock, flags);
961 ++dev->stats.tx_dropped;
962 dev_kfree_skb_any(skb);
963 return NULL;
964 }
965
966 /* To avoid race condition, make sure that the
967 * neigh will be added only once.
968 */
969 if (unlikely(!list_empty(&neigh->list))) {
970 spin_unlock_irqrestore(&priv->lock, flags);
971 return neigh;
972 }
973
974 path = __path_find(dev, daddr + 4);
975 if (!path) {
976 path = path_rec_create(dev, daddr + 4);
977 if (!path)
978 goto err_path;
979
980 __path_add(dev, path);
981 }
982
983 list_add_tail(&neigh->list, &path->neigh_list);
984
985 if (path->ah && path->ah->valid) {
986 kref_get(&path->ah->ref);
987 neigh->ah = path->ah;
988
989 if (ipoib_cm_enabled(dev, neigh->daddr)) {
990 if (!ipoib_cm_get(neigh))
991 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
992 if (!ipoib_cm_get(neigh)) {
993 ipoib_neigh_free(neigh);
994 goto err_drop;
995 }
996 if (skb_queue_len(&neigh->queue) <
997 IPOIB_MAX_PATH_REC_QUEUE) {
998 push_pseudo_header(skb, neigh->daddr);
999 __skb_queue_tail(&neigh->queue, skb);
1000 } else {
1001 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
1002 skb_queue_len(&neigh->queue));
1003 goto err_drop;
1004 }
1005 } else {
1006 spin_unlock_irqrestore(&priv->lock, flags);
1007 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1008 IPOIB_QPN(daddr));
1009 ipoib_neigh_put(neigh);
1010 return NULL;
1011 }
1012 } else {
1013 neigh->ah = NULL;
1014
1015 if (!path->query && path_rec_start(dev, path))
1016 goto err_path;
1017 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1018 push_pseudo_header(skb, neigh->daddr);
1019 __skb_queue_tail(&neigh->queue, skb);
1020 } else {
1021 goto err_drop;
1022 }
1023 }
1024
1025 spin_unlock_irqrestore(&priv->lock, flags);
1026 ipoib_neigh_put(neigh);
1027 return NULL;
1028
1029err_path:
1030 ipoib_neigh_free(neigh);
1031err_drop:
1032 ++dev->stats.tx_dropped;
1033 dev_kfree_skb_any(skb);
1034
1035 spin_unlock_irqrestore(&priv->lock, flags);
1036 ipoib_neigh_put(neigh);
1037
1038 return NULL;
1039}
1040
1041static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1042 struct ipoib_pseudo_header *phdr)
1043{
1044 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1045 struct rdma_netdev *rn = netdev_priv(dev);
1046 struct ipoib_path *path;
1047 unsigned long flags;
1048
1049 spin_lock_irqsave(&priv->lock, flags);
1050
1051 /* no broadcast means that all paths are (going to be) not valid */
1052 if (!priv->broadcast)
1053 goto drop_and_unlock;
1054
1055 path = __path_find(dev, phdr->hwaddr + 4);
1056 if (!path || !path->ah || !path->ah->valid) {
1057 if (!path) {
1058 path = path_rec_create(dev, phdr->hwaddr + 4);
1059 if (!path)
1060 goto drop_and_unlock;
1061 __path_add(dev, path);
1062 } else {
1063 /*
1064 * make sure there are no changes in the existing
1065 * path record
1066 */
1067 init_path_rec(priv, path, phdr->hwaddr + 4);
1068 }
1069 if (!path->query && path_rec_start(dev, path)) {
1070 goto drop_and_unlock;
1071 }
1072
1073 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1074 push_pseudo_header(skb, phdr->hwaddr);
1075 __skb_queue_tail(&path->queue, skb);
1076 goto unlock;
1077 } else {
1078 goto drop_and_unlock;
1079 }
1080 }
1081
1082 spin_unlock_irqrestore(&priv->lock, flags);
1083 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1084 be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1085 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1086 IPOIB_QPN(phdr->hwaddr));
1087 return;
1088
1089drop_and_unlock:
1090 ++dev->stats.tx_dropped;
1091 dev_kfree_skb_any(skb);
1092unlock:
1093 spin_unlock_irqrestore(&priv->lock, flags);
1094}
1095
1096static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1097{
1098 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1099 struct rdma_netdev *rn = netdev_priv(dev);
1100 struct ipoib_neigh *neigh;
1101 struct ipoib_pseudo_header *phdr;
1102 struct ipoib_header *header;
1103 unsigned long flags;
1104
1105 phdr = (struct ipoib_pseudo_header *) skb->data;
1106 skb_pull(skb, sizeof(*phdr));
1107 header = (struct ipoib_header *) skb->data;
1108
1109 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1110 /* multicast, arrange "if" according to probability */
1111 if ((header->proto != htons(ETH_P_IP)) &&
1112 (header->proto != htons(ETH_P_IPV6)) &&
1113 (header->proto != htons(ETH_P_ARP)) &&
1114 (header->proto != htons(ETH_P_RARP)) &&
1115 (header->proto != htons(ETH_P_TIPC))) {
1116 /* ethertype not supported by IPoIB */
1117 ++dev->stats.tx_dropped;
1118 dev_kfree_skb_any(skb);
1119 return NETDEV_TX_OK;
1120 }
1121 /* Add in the P_Key for multicast*/
1122 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1123 phdr->hwaddr[9] = priv->pkey & 0xff;
1124
1125 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1126 if (likely(neigh))
1127 goto send_using_neigh;
1128 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1129 return NETDEV_TX_OK;
1130 }
1131
1132 /* unicast, arrange "switch" according to probability */
1133 switch (header->proto) {
1134 case htons(ETH_P_IP):
1135 case htons(ETH_P_IPV6):
1136 case htons(ETH_P_TIPC):
1137 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1138 if (unlikely(!neigh)) {
1139 neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1140 if (likely(!neigh))
1141 return NETDEV_TX_OK;
1142 }
1143 break;
1144 case htons(ETH_P_ARP):
1145 case htons(ETH_P_RARP):
1146 /* for unicast ARP and RARP should always perform path find */
1147 unicast_arp_send(skb, dev, phdr);
1148 return NETDEV_TX_OK;
1149 default:
1150 /* ethertype not supported by IPoIB */
1151 ++dev->stats.tx_dropped;
1152 dev_kfree_skb_any(skb);
1153 return NETDEV_TX_OK;
1154 }
1155
1156send_using_neigh:
1157 /* note we now hold a ref to neigh */
1158 if (ipoib_cm_get(neigh)) {
1159 if (ipoib_cm_up(neigh)) {
1160 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1161 goto unref;
1162 }
1163 } else if (neigh->ah && neigh->ah->valid) {
1164 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1165 IPOIB_QPN(phdr->hwaddr));
1166 goto unref;
1167 } else if (neigh->ah) {
1168 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1169 }
1170
1171 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1172 push_pseudo_header(skb, phdr->hwaddr);
1173 spin_lock_irqsave(&priv->lock, flags);
1174 __skb_queue_tail(&neigh->queue, skb);
1175 spin_unlock_irqrestore(&priv->lock, flags);
1176 } else {
1177 ++dev->stats.tx_dropped;
1178 dev_kfree_skb_any(skb);
1179 }
1180
1181unref:
1182 ipoib_neigh_put(neigh);
1183
1184 return NETDEV_TX_OK;
1185}
1186
1187static void ipoib_timeout(struct net_device *dev, unsigned int txqueue)
1188{
1189 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1190 struct rdma_netdev *rn = netdev_priv(dev);
1191
1192 if (rn->tx_timeout) {
1193 rn->tx_timeout(dev, txqueue);
1194 return;
1195 }
1196 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1197 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1198 ipoib_warn(priv,
1199 "queue stopped %d, tx_head %u, tx_tail %u, global_tx_head %u, global_tx_tail %u\n",
1200 netif_queue_stopped(dev), priv->tx_head, priv->tx_tail,
1201 priv->global_tx_head, priv->global_tx_tail);
1202
1203
1204 schedule_work(&priv->tx_timeout_work);
1205}
1206
1207void ipoib_ib_tx_timeout_work(struct work_struct *work)
1208{
1209 struct ipoib_dev_priv *priv = container_of(work,
1210 struct ipoib_dev_priv,
1211 tx_timeout_work);
1212 int err;
1213
1214 rtnl_lock();
1215
1216 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1217 goto unlock;
1218
1219 ipoib_stop(priv->dev);
1220 err = ipoib_open(priv->dev);
1221 if (err) {
1222 ipoib_warn(priv, "ipoib_open failed recovering from a tx_timeout, err(%d).\n",
1223 err);
1224 goto unlock;
1225 }
1226
1227 netif_tx_wake_all_queues(priv->dev);
1228unlock:
1229 rtnl_unlock();
1230
1231}
1232
1233static int ipoib_hard_header(struct sk_buff *skb,
1234 struct net_device *dev,
1235 unsigned short type,
1236 const void *daddr,
1237 const void *saddr,
1238 unsigned int len)
1239{
1240 struct ipoib_header *header;
1241
1242 header = skb_push(skb, sizeof(*header));
1243
1244 header->proto = htons(type);
1245 header->reserved = 0;
1246
1247 /*
1248 * we don't rely on dst_entry structure, always stuff the
1249 * destination address into skb hard header so we can figure out where
1250 * to send the packet later.
1251 */
1252 push_pseudo_header(skb, daddr);
1253
1254 return IPOIB_HARD_LEN;
1255}
1256
1257static void ipoib_set_mcast_list(struct net_device *dev)
1258{
1259 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1260
1261 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1262 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1263 return;
1264 }
1265
1266 queue_work(priv->wq, &priv->restart_task);
1267}
1268
1269static int ipoib_get_iflink(const struct net_device *dev)
1270{
1271 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1272
1273 /* parent interface */
1274 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1275 return dev->ifindex;
1276
1277 /* child/vlan interface */
1278 return priv->parent->ifindex;
1279}
1280
1281static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1282{
1283 /*
1284 * Use only the address parts that contributes to spreading
1285 * The subnet prefix is not used as one can not connect to
1286 * same remote port (GUID) using the same remote QPN via two
1287 * different subnets.
1288 */
1289 /* qpn octets[1:4) & port GUID octets[12:20) */
1290 u32 *d32 = (u32 *) daddr;
1291 u32 hv;
1292
1293 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1294 return hv & htbl->mask;
1295}
1296
1297struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1298{
1299 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1300 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1301 struct ipoib_neigh_hash *htbl;
1302 struct ipoib_neigh *neigh = NULL;
1303 u32 hash_val;
1304
1305 rcu_read_lock_bh();
1306
1307 htbl = rcu_dereference_bh(ntbl->htbl);
1308
1309 if (!htbl)
1310 goto out_unlock;
1311
1312 hash_val = ipoib_addr_hash(htbl, daddr);
1313 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1314 neigh != NULL;
1315 neigh = rcu_dereference_bh(neigh->hnext)) {
1316 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1317 /* found, take one ref on behalf of the caller */
1318 if (!refcount_inc_not_zero(&neigh->refcnt)) {
1319 /* deleted */
1320 neigh = NULL;
1321 goto out_unlock;
1322 }
1323
1324 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1325 neigh->alive = jiffies;
1326 goto out_unlock;
1327 }
1328 }
1329
1330out_unlock:
1331 rcu_read_unlock_bh();
1332 return neigh;
1333}
1334
1335static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1336{
1337 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1338 struct ipoib_neigh_hash *htbl;
1339 unsigned long neigh_obsolete;
1340 unsigned long dt;
1341 unsigned long flags;
1342 int i;
1343 LIST_HEAD(remove_list);
1344
1345 spin_lock_irqsave(&priv->lock, flags);
1346
1347 htbl = rcu_dereference_protected(ntbl->htbl,
1348 lockdep_is_held(&priv->lock));
1349
1350 if (!htbl)
1351 goto out_unlock;
1352
1353 /* neigh is obsolete if it was idle for two GC periods */
1354 dt = 2 * arp_tbl.gc_interval;
1355 neigh_obsolete = jiffies - dt;
1356
1357 for (i = 0; i < htbl->size; i++) {
1358 struct ipoib_neigh *neigh;
1359 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1360
1361 while ((neigh = rcu_dereference_protected(*np,
1362 lockdep_is_held(&priv->lock))) != NULL) {
1363 /* was the neigh idle for two GC periods */
1364 if (time_after(neigh_obsolete, neigh->alive)) {
1365
1366 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1367
1368 rcu_assign_pointer(*np,
1369 rcu_dereference_protected(neigh->hnext,
1370 lockdep_is_held(&priv->lock)));
1371 /* remove from path/mc list */
1372 list_del_init(&neigh->list);
1373 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1374 } else {
1375 np = &neigh->hnext;
1376 }
1377
1378 }
1379 }
1380
1381out_unlock:
1382 spin_unlock_irqrestore(&priv->lock, flags);
1383 ipoib_mcast_remove_list(&remove_list);
1384}
1385
1386static void ipoib_reap_neigh(struct work_struct *work)
1387{
1388 struct ipoib_dev_priv *priv =
1389 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1390
1391 __ipoib_reap_neigh(priv);
1392
1393 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1394 arp_tbl.gc_interval);
1395}
1396
1397
1398static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1399 struct net_device *dev)
1400{
1401 struct ipoib_neigh *neigh;
1402
1403 neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1404 if (!neigh)
1405 return NULL;
1406
1407 neigh->dev = dev;
1408 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1409 skb_queue_head_init(&neigh->queue);
1410 INIT_LIST_HEAD(&neigh->list);
1411 ipoib_cm_set(neigh, NULL);
1412 /* one ref on behalf of the caller */
1413 refcount_set(&neigh->refcnt, 1);
1414
1415 return neigh;
1416}
1417
1418struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1419 struct net_device *dev)
1420{
1421 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1422 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1423 struct ipoib_neigh_hash *htbl;
1424 struct ipoib_neigh *neigh;
1425 u32 hash_val;
1426
1427 htbl = rcu_dereference_protected(ntbl->htbl,
1428 lockdep_is_held(&priv->lock));
1429 if (!htbl) {
1430 neigh = NULL;
1431 goto out_unlock;
1432 }
1433
1434 /* need to add a new neigh, but maybe some other thread succeeded?
1435 * recalc hash, maybe hash resize took place so we do a search
1436 */
1437 hash_val = ipoib_addr_hash(htbl, daddr);
1438 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1439 lockdep_is_held(&priv->lock));
1440 neigh != NULL;
1441 neigh = rcu_dereference_protected(neigh->hnext,
1442 lockdep_is_held(&priv->lock))) {
1443 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1444 /* found, take one ref on behalf of the caller */
1445 if (!refcount_inc_not_zero(&neigh->refcnt)) {
1446 /* deleted */
1447 neigh = NULL;
1448 break;
1449 }
1450 neigh->alive = jiffies;
1451 goto out_unlock;
1452 }
1453 }
1454
1455 neigh = ipoib_neigh_ctor(daddr, dev);
1456 if (!neigh)
1457 goto out_unlock;
1458
1459 /* one ref on behalf of the hash table */
1460 refcount_inc(&neigh->refcnt);
1461 neigh->alive = jiffies;
1462 /* put in hash */
1463 rcu_assign_pointer(neigh->hnext,
1464 rcu_dereference_protected(htbl->buckets[hash_val],
1465 lockdep_is_held(&priv->lock)));
1466 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1467 atomic_inc(&ntbl->entries);
1468
1469out_unlock:
1470
1471 return neigh;
1472}
1473
1474void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1475{
1476 /* neigh reference count was dropprd to zero */
1477 struct net_device *dev = neigh->dev;
1478 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1479 struct sk_buff *skb;
1480 if (neigh->ah)
1481 ipoib_put_ah(neigh->ah);
1482 while ((skb = __skb_dequeue(&neigh->queue))) {
1483 ++dev->stats.tx_dropped;
1484 dev_kfree_skb_any(skb);
1485 }
1486 if (ipoib_cm_get(neigh))
1487 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1488 ipoib_dbg(ipoib_priv(dev),
1489 "neigh free for %06x %pI6\n",
1490 IPOIB_QPN(neigh->daddr),
1491 neigh->daddr + 4);
1492 kfree(neigh);
1493 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1494 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1495 complete(&priv->ntbl.flushed);
1496 }
1497}
1498
1499static void ipoib_neigh_reclaim(struct rcu_head *rp)
1500{
1501 /* Called as a result of removal from hash table */
1502 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1503 /* note TX context may hold another ref */
1504 ipoib_neigh_put(neigh);
1505}
1506
1507void ipoib_neigh_free(struct ipoib_neigh *neigh)
1508{
1509 struct net_device *dev = neigh->dev;
1510 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1511 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1512 struct ipoib_neigh_hash *htbl;
1513 struct ipoib_neigh __rcu **np;
1514 struct ipoib_neigh *n;
1515 u32 hash_val;
1516
1517 htbl = rcu_dereference_protected(ntbl->htbl,
1518 lockdep_is_held(&priv->lock));
1519 if (!htbl)
1520 return;
1521
1522 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1523 np = &htbl->buckets[hash_val];
1524 for (n = rcu_dereference_protected(*np,
1525 lockdep_is_held(&priv->lock));
1526 n != NULL;
1527 n = rcu_dereference_protected(*np,
1528 lockdep_is_held(&priv->lock))) {
1529 if (n == neigh) {
1530 /* found */
1531 rcu_assign_pointer(*np,
1532 rcu_dereference_protected(neigh->hnext,
1533 lockdep_is_held(&priv->lock)));
1534 /* remove from parent list */
1535 list_del_init(&neigh->list);
1536 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1537 return;
1538 } else {
1539 np = &n->hnext;
1540 }
1541 }
1542}
1543
1544static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1545{
1546 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1547 struct ipoib_neigh_hash *htbl;
1548 struct ipoib_neigh __rcu **buckets;
1549 u32 size;
1550
1551 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1552 ntbl->htbl = NULL;
1553 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1554 if (!htbl)
1555 return -ENOMEM;
1556 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1557 buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1558 if (!buckets) {
1559 kfree(htbl);
1560 return -ENOMEM;
1561 }
1562 htbl->size = size;
1563 htbl->mask = (size - 1);
1564 htbl->buckets = buckets;
1565 RCU_INIT_POINTER(ntbl->htbl, htbl);
1566 htbl->ntbl = ntbl;
1567 atomic_set(&ntbl->entries, 0);
1568
1569 /* start garbage collection */
1570 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1571 arp_tbl.gc_interval);
1572
1573 return 0;
1574}
1575
1576static void neigh_hash_free_rcu(struct rcu_head *head)
1577{
1578 struct ipoib_neigh_hash *htbl = container_of(head,
1579 struct ipoib_neigh_hash,
1580 rcu);
1581 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1582 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1583
1584 kvfree(buckets);
1585 kfree(htbl);
1586 complete(&ntbl->deleted);
1587}
1588
1589void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1590{
1591 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1592 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1593 struct ipoib_neigh_hash *htbl;
1594 unsigned long flags;
1595 int i;
1596
1597 /* remove all neigh connected to a given path or mcast */
1598 spin_lock_irqsave(&priv->lock, flags);
1599
1600 htbl = rcu_dereference_protected(ntbl->htbl,
1601 lockdep_is_held(&priv->lock));
1602
1603 if (!htbl)
1604 goto out_unlock;
1605
1606 for (i = 0; i < htbl->size; i++) {
1607 struct ipoib_neigh *neigh;
1608 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1609
1610 while ((neigh = rcu_dereference_protected(*np,
1611 lockdep_is_held(&priv->lock))) != NULL) {
1612 /* delete neighs belong to this parent */
1613 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1614 rcu_assign_pointer(*np,
1615 rcu_dereference_protected(neigh->hnext,
1616 lockdep_is_held(&priv->lock)));
1617 /* remove from parent list */
1618 list_del_init(&neigh->list);
1619 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1620 } else {
1621 np = &neigh->hnext;
1622 }
1623
1624 }
1625 }
1626out_unlock:
1627 spin_unlock_irqrestore(&priv->lock, flags);
1628}
1629
1630static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1631{
1632 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1633 struct ipoib_neigh_hash *htbl;
1634 unsigned long flags;
1635 int i, wait_flushed = 0;
1636
1637 init_completion(&priv->ntbl.flushed);
1638 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1639
1640 spin_lock_irqsave(&priv->lock, flags);
1641
1642 htbl = rcu_dereference_protected(ntbl->htbl,
1643 lockdep_is_held(&priv->lock));
1644 if (!htbl)
1645 goto out_unlock;
1646
1647 wait_flushed = atomic_read(&priv->ntbl.entries);
1648 if (!wait_flushed)
1649 goto free_htbl;
1650
1651 for (i = 0; i < htbl->size; i++) {
1652 struct ipoib_neigh *neigh;
1653 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1654
1655 while ((neigh = rcu_dereference_protected(*np,
1656 lockdep_is_held(&priv->lock))) != NULL) {
1657 rcu_assign_pointer(*np,
1658 rcu_dereference_protected(neigh->hnext,
1659 lockdep_is_held(&priv->lock)));
1660 /* remove from path/mc list */
1661 list_del_init(&neigh->list);
1662 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1663 }
1664 }
1665
1666free_htbl:
1667 rcu_assign_pointer(ntbl->htbl, NULL);
1668 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1669
1670out_unlock:
1671 spin_unlock_irqrestore(&priv->lock, flags);
1672 if (wait_flushed)
1673 wait_for_completion(&priv->ntbl.flushed);
1674}
1675
1676static void ipoib_neigh_hash_uninit(struct net_device *dev)
1677{
1678 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1679
1680 ipoib_dbg(priv, "%s\n", __func__);
1681 init_completion(&priv->ntbl.deleted);
1682
1683 cancel_delayed_work_sync(&priv->neigh_reap_task);
1684
1685 ipoib_flush_neighs(priv);
1686
1687 wait_for_completion(&priv->ntbl.deleted);
1688}
1689
1690static void ipoib_napi_add(struct net_device *dev)
1691{
1692 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1693
1694 netif_napi_add_weight(dev, &priv->recv_napi, ipoib_rx_poll,
1695 IPOIB_NUM_WC);
1696 netif_napi_add_weight(dev, &priv->send_napi, ipoib_tx_poll,
1697 MAX_SEND_CQE);
1698}
1699
1700static void ipoib_napi_del(struct net_device *dev)
1701{
1702 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1703
1704 netif_napi_del(&priv->recv_napi);
1705 netif_napi_del(&priv->send_napi);
1706}
1707
1708static void ipoib_dev_uninit_default(struct net_device *dev)
1709{
1710 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1711
1712 ipoib_transport_dev_cleanup(dev);
1713
1714 ipoib_napi_del(dev);
1715
1716 ipoib_cm_dev_cleanup(dev);
1717
1718 kfree(priv->rx_ring);
1719 vfree(priv->tx_ring);
1720
1721 priv->rx_ring = NULL;
1722 priv->tx_ring = NULL;
1723}
1724
1725static int ipoib_dev_init_default(struct net_device *dev)
1726{
1727 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1728 u8 addr_mod[3];
1729
1730 ipoib_napi_add(dev);
1731
1732 /* Allocate RX/TX "rings" to hold queued skbs */
1733 priv->rx_ring = kcalloc(ipoib_recvq_size,
1734 sizeof(*priv->rx_ring),
1735 GFP_KERNEL);
1736 if (!priv->rx_ring)
1737 goto out;
1738
1739 priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1740 sizeof(*priv->tx_ring)));
1741 if (!priv->tx_ring) {
1742 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1743 priv->ca->name, ipoib_sendq_size);
1744 goto out_rx_ring_cleanup;
1745 }
1746
1747 /* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
1748
1749 if (ipoib_transport_dev_init(dev, priv->ca)) {
1750 pr_warn("%s: ipoib_transport_dev_init failed\n",
1751 priv->ca->name);
1752 goto out_tx_ring_cleanup;
1753 }
1754
1755 /* after qp created set dev address */
1756 addr_mod[0] = (priv->qp->qp_num >> 16) & 0xff;
1757 addr_mod[1] = (priv->qp->qp_num >> 8) & 0xff;
1758 addr_mod[2] = (priv->qp->qp_num) & 0xff;
1759 dev_addr_mod(priv->dev, 1, addr_mod, sizeof(addr_mod));
1760
1761 return 0;
1762
1763out_tx_ring_cleanup:
1764 vfree(priv->tx_ring);
1765
1766out_rx_ring_cleanup:
1767 kfree(priv->rx_ring);
1768
1769out:
1770 ipoib_napi_del(dev);
1771 return -ENOMEM;
1772}
1773
1774static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1775 int cmd)
1776{
1777 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1778
1779 if (!priv->rn_ops->ndo_eth_ioctl)
1780 return -EOPNOTSUPP;
1781
1782 return priv->rn_ops->ndo_eth_ioctl(dev, ifr, cmd);
1783}
1784
1785static int ipoib_dev_init(struct net_device *dev)
1786{
1787 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1788 int ret = -ENOMEM;
1789
1790 priv->qp = NULL;
1791
1792 /*
1793 * the various IPoIB tasks assume they will never race against
1794 * themselves, so always use a single thread workqueue
1795 */
1796 priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1797 if (!priv->wq) {
1798 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1799 goto out;
1800 }
1801
1802 /* create pd, which used both for control and datapath*/
1803 priv->pd = ib_alloc_pd(priv->ca, 0);
1804 if (IS_ERR(priv->pd)) {
1805 pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1806 goto clean_wq;
1807 }
1808
1809 ret = priv->rn_ops->ndo_init(dev);
1810 if (ret) {
1811 pr_warn("%s failed to init HW resource\n", dev->name);
1812 goto out_free_pd;
1813 }
1814
1815 ret = ipoib_neigh_hash_init(priv);
1816 if (ret) {
1817 pr_warn("%s failed to init neigh hash\n", dev->name);
1818 goto out_dev_uninit;
1819 }
1820
1821 if (dev->flags & IFF_UP) {
1822 if (ipoib_ib_dev_open(dev)) {
1823 pr_warn("%s failed to open device\n", dev->name);
1824 ret = -ENODEV;
1825 goto out_hash_uninit;
1826 }
1827 }
1828
1829 return 0;
1830
1831out_hash_uninit:
1832 ipoib_neigh_hash_uninit(dev);
1833
1834out_dev_uninit:
1835 ipoib_ib_dev_cleanup(dev);
1836
1837out_free_pd:
1838 if (priv->pd) {
1839 ib_dealloc_pd(priv->pd);
1840 priv->pd = NULL;
1841 }
1842
1843clean_wq:
1844 if (priv->wq) {
1845 destroy_workqueue(priv->wq);
1846 priv->wq = NULL;
1847 }
1848
1849out:
1850 return ret;
1851}
1852
1853/*
1854 * This must be called before doing an unregister_netdev on a parent device to
1855 * shutdown the IB event handler.
1856 */
1857static void ipoib_parent_unregister_pre(struct net_device *ndev)
1858{
1859 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1860
1861 /*
1862 * ipoib_set_mac checks netif_running before pushing work, clearing
1863 * running ensures the it will not add more work.
1864 */
1865 rtnl_lock();
1866 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP, NULL);
1867 rtnl_unlock();
1868
1869 /* ipoib_event() cannot be running once this returns */
1870 ib_unregister_event_handler(&priv->event_handler);
1871
1872 /*
1873 * Work on the queue grabs the rtnl lock, so this cannot be done while
1874 * also holding it.
1875 */
1876 flush_workqueue(ipoib_workqueue);
1877}
1878
1879static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1880{
1881 priv->hca_caps = priv->ca->attrs.device_cap_flags;
1882 priv->kernel_caps = priv->ca->attrs.kernel_cap_flags;
1883
1884 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1885 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1886
1887 if (priv->kernel_caps & IBK_UD_TSO)
1888 priv->dev->hw_features |= NETIF_F_TSO;
1889
1890 priv->dev->features |= priv->dev->hw_features;
1891 }
1892}
1893
1894static int ipoib_parent_init(struct net_device *ndev)
1895{
1896 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1897 struct ib_port_attr attr;
1898 int result;
1899
1900 result = ib_query_port(priv->ca, priv->port, &attr);
1901 if (result) {
1902 pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1903 priv->port);
1904 return result;
1905 }
1906 priv->max_ib_mtu = rdma_mtu_from_attr(priv->ca, priv->port, &attr);
1907
1908 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1909 if (result) {
1910 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1911 priv->ca->name, priv->port, result);
1912 return result;
1913 }
1914
1915 result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1916 if (result) {
1917 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1918 priv->ca->name, priv->port, result);
1919 return result;
1920 }
1921 dev_addr_mod(priv->dev, 4, priv->local_gid.raw, sizeof(union ib_gid));
1922
1923 SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1924 priv->dev->dev_port = priv->port - 1;
1925 /* Let's set this one too for backwards compatibility. */
1926 priv->dev->dev_id = priv->port - 1;
1927
1928 return 0;
1929}
1930
1931static void ipoib_child_init(struct net_device *ndev)
1932{
1933 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1934 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1935
1936 priv->max_ib_mtu = ppriv->max_ib_mtu;
1937 set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1938 if (memchr_inv(priv->dev->dev_addr, 0, INFINIBAND_ALEN))
1939 memcpy(&priv->local_gid, priv->dev->dev_addr + 4,
1940 sizeof(priv->local_gid));
1941 else {
1942 __dev_addr_set(priv->dev, ppriv->dev->dev_addr,
1943 INFINIBAND_ALEN);
1944 memcpy(&priv->local_gid, &ppriv->local_gid,
1945 sizeof(priv->local_gid));
1946 }
1947}
1948
1949static int ipoib_ndo_init(struct net_device *ndev)
1950{
1951 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1952 int rc;
1953 struct rdma_netdev *rn = netdev_priv(ndev);
1954
1955 if (priv->parent) {
1956 ipoib_child_init(ndev);
1957 } else {
1958 rc = ipoib_parent_init(ndev);
1959 if (rc)
1960 return rc;
1961 }
1962
1963 /* MTU will be reset when mcast join happens */
1964 ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1965 priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
1966 rn->mtu = priv->mcast_mtu;
1967 ndev->max_mtu = IPOIB_CM_MTU;
1968
1969 ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
1970
1971 /*
1972 * Set the full membership bit, so that we join the right
1973 * broadcast group, etc.
1974 */
1975 priv->pkey |= 0x8000;
1976
1977 ndev->broadcast[8] = priv->pkey >> 8;
1978 ndev->broadcast[9] = priv->pkey & 0xff;
1979 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1980
1981 ipoib_set_dev_features(priv);
1982
1983 rc = ipoib_dev_init(ndev);
1984 if (rc) {
1985 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1986 priv->ca->name, priv->dev->name, priv->port, rc);
1987 return rc;
1988 }
1989
1990 if (priv->parent) {
1991 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1992
1993 dev_hold(priv->parent);
1994
1995 down_write(&ppriv->vlan_rwsem);
1996 list_add_tail(&priv->list, &ppriv->child_intfs);
1997 up_write(&ppriv->vlan_rwsem);
1998 }
1999
2000 return 0;
2001}
2002
2003static void ipoib_ndo_uninit(struct net_device *dev)
2004{
2005 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2006
2007 ASSERT_RTNL();
2008
2009 /*
2010 * ipoib_remove_one guarantees the children are removed before the
2011 * parent, and that is the only place where a parent can be removed.
2012 */
2013 WARN_ON(!list_empty(&priv->child_intfs));
2014
2015 if (priv->parent) {
2016 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
2017
2018 down_write(&ppriv->vlan_rwsem);
2019 list_del(&priv->list);
2020 up_write(&ppriv->vlan_rwsem);
2021 }
2022
2023 ipoib_neigh_hash_uninit(dev);
2024
2025 ipoib_ib_dev_cleanup(dev);
2026
2027 /* no more works over the priv->wq */
2028 if (priv->wq) {
2029 /* See ipoib_mcast_carrier_on_task() */
2030 WARN_ON(test_bit(IPOIB_FLAG_OPER_UP, &priv->flags));
2031 destroy_workqueue(priv->wq);
2032 priv->wq = NULL;
2033 }
2034
2035 dev_put(priv->parent);
2036}
2037
2038static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
2039{
2040 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2041
2042 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
2043}
2044
2045static int ipoib_get_vf_config(struct net_device *dev, int vf,
2046 struct ifla_vf_info *ivf)
2047{
2048 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2049 int err;
2050
2051 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2052 if (err)
2053 return err;
2054
2055 ivf->vf = vf;
2056 memcpy(ivf->mac, dev->dev_addr, dev->addr_len);
2057
2058 return 0;
2059}
2060
2061static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2062{
2063 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2064
2065 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2066 return -EINVAL;
2067
2068 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2069}
2070
2071static int ipoib_get_vf_guid(struct net_device *dev, int vf,
2072 struct ifla_vf_guid *node_guid,
2073 struct ifla_vf_guid *port_guid)
2074{
2075 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2076
2077 return ib_get_vf_guid(priv->ca, vf, priv->port, node_guid, port_guid);
2078}
2079
2080static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2081 struct ifla_vf_stats *vf_stats)
2082{
2083 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2084
2085 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2086}
2087
2088static const struct header_ops ipoib_header_ops = {
2089 .create = ipoib_hard_header,
2090};
2091
2092static const struct net_device_ops ipoib_netdev_ops_pf = {
2093 .ndo_init = ipoib_ndo_init,
2094 .ndo_uninit = ipoib_ndo_uninit,
2095 .ndo_open = ipoib_open,
2096 .ndo_stop = ipoib_stop,
2097 .ndo_change_mtu = ipoib_change_mtu,
2098 .ndo_fix_features = ipoib_fix_features,
2099 .ndo_start_xmit = ipoib_start_xmit,
2100 .ndo_tx_timeout = ipoib_timeout,
2101 .ndo_set_rx_mode = ipoib_set_mcast_list,
2102 .ndo_get_iflink = ipoib_get_iflink,
2103 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
2104 .ndo_get_vf_config = ipoib_get_vf_config,
2105 .ndo_get_vf_stats = ipoib_get_vf_stats,
2106 .ndo_get_vf_guid = ipoib_get_vf_guid,
2107 .ndo_set_vf_guid = ipoib_set_vf_guid,
2108 .ndo_set_mac_address = ipoib_set_mac,
2109 .ndo_get_stats64 = ipoib_get_stats,
2110 .ndo_eth_ioctl = ipoib_ioctl,
2111};
2112
2113static const struct net_device_ops ipoib_netdev_ops_vf = {
2114 .ndo_init = ipoib_ndo_init,
2115 .ndo_uninit = ipoib_ndo_uninit,
2116 .ndo_open = ipoib_open,
2117 .ndo_stop = ipoib_stop,
2118 .ndo_change_mtu = ipoib_change_mtu,
2119 .ndo_fix_features = ipoib_fix_features,
2120 .ndo_start_xmit = ipoib_start_xmit,
2121 .ndo_tx_timeout = ipoib_timeout,
2122 .ndo_set_rx_mode = ipoib_set_mcast_list,
2123 .ndo_get_iflink = ipoib_get_iflink,
2124 .ndo_get_stats64 = ipoib_get_stats,
2125 .ndo_eth_ioctl = ipoib_ioctl,
2126};
2127
2128static const struct net_device_ops ipoib_netdev_default_pf = {
2129 .ndo_init = ipoib_dev_init_default,
2130 .ndo_uninit = ipoib_dev_uninit_default,
2131 .ndo_open = ipoib_ib_dev_open_default,
2132 .ndo_stop = ipoib_ib_dev_stop_default,
2133};
2134
2135void ipoib_setup_common(struct net_device *dev)
2136{
2137 dev->header_ops = &ipoib_header_ops;
2138 dev->netdev_ops = &ipoib_netdev_default_pf;
2139
2140 ipoib_set_ethtool_ops(dev);
2141
2142 dev->watchdog_timeo = 10 * HZ;
2143
2144 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
2145
2146 dev->hard_header_len = IPOIB_HARD_LEN;
2147 dev->addr_len = INFINIBAND_ALEN;
2148 dev->type = ARPHRD_INFINIBAND;
2149 dev->tx_queue_len = ipoib_sendq_size * 2;
2150 dev->features = (NETIF_F_VLAN_CHALLENGED |
2151 NETIF_F_HIGHDMA);
2152 netif_keep_dst(dev);
2153
2154 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2155
2156 /*
2157 * unregister_netdev always frees the netdev, we use this mode
2158 * consistently to unify all the various unregister paths, including
2159 * those connected to rtnl_link_ops which require it.
2160 */
2161 dev->needs_free_netdev = true;
2162}
2163
2164static void ipoib_build_priv(struct net_device *dev)
2165{
2166 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2167
2168 priv->dev = dev;
2169 spin_lock_init(&priv->lock);
2170 init_rwsem(&priv->vlan_rwsem);
2171 mutex_init(&priv->mcast_mutex);
2172
2173 INIT_LIST_HEAD(&priv->path_list);
2174 INIT_LIST_HEAD(&priv->child_intfs);
2175 INIT_LIST_HEAD(&priv->dead_ahs);
2176 INIT_LIST_HEAD(&priv->multicast_list);
2177
2178 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
2179 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2180 INIT_WORK(&priv->reschedule_napi_work, ipoib_napi_schedule_work);
2181 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
2182 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
2183 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
2184 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2185 INIT_WORK(&priv->tx_timeout_work, ipoib_ib_tx_timeout_work);
2186 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2187 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2188}
2189
2190static struct net_device *ipoib_alloc_netdev(struct ib_device *hca, u32 port,
2191 const char *name)
2192{
2193 struct net_device *dev;
2194
2195 dev = rdma_alloc_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2196 NET_NAME_UNKNOWN, ipoib_setup_common);
2197 if (!IS_ERR(dev) || PTR_ERR(dev) != -EOPNOTSUPP)
2198 return dev;
2199
2200 dev = alloc_netdev(sizeof(struct rdma_netdev), name, NET_NAME_UNKNOWN,
2201 ipoib_setup_common);
2202 if (!dev)
2203 return ERR_PTR(-ENOMEM);
2204 return dev;
2205}
2206
2207int ipoib_intf_init(struct ib_device *hca, u32 port, const char *name,
2208 struct net_device *dev)
2209{
2210 struct rdma_netdev *rn = netdev_priv(dev);
2211 struct ipoib_dev_priv *priv;
2212 int rc;
2213
2214 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2215 if (!priv)
2216 return -ENOMEM;
2217
2218 priv->ca = hca;
2219 priv->port = port;
2220
2221 rc = rdma_init_netdev(hca, port, RDMA_NETDEV_IPOIB, name,
2222 NET_NAME_UNKNOWN, ipoib_setup_common, dev);
2223 if (rc) {
2224 if (rc != -EOPNOTSUPP)
2225 goto out;
2226
2227 rn->send = ipoib_send;
2228 rn->attach_mcast = ipoib_mcast_attach;
2229 rn->detach_mcast = ipoib_mcast_detach;
2230 rn->hca = hca;
2231
2232 rc = netif_set_real_num_tx_queues(dev, 1);
2233 if (rc)
2234 goto out;
2235
2236 rc = netif_set_real_num_rx_queues(dev, 1);
2237 if (rc)
2238 goto out;
2239 }
2240
2241 priv->rn_ops = dev->netdev_ops;
2242
2243 if (hca->attrs.kernel_cap_flags & IBK_VIRTUAL_FUNCTION)
2244 dev->netdev_ops = &ipoib_netdev_ops_vf;
2245 else
2246 dev->netdev_ops = &ipoib_netdev_ops_pf;
2247
2248 rn->clnt_priv = priv;
2249 /*
2250 * Only the child register_netdev flows can handle priv_destructor
2251 * being set, so we force it to NULL here and handle manually until it
2252 * is safe to turn on.
2253 */
2254 priv->next_priv_destructor = dev->priv_destructor;
2255 dev->priv_destructor = NULL;
2256
2257 ipoib_build_priv(dev);
2258
2259 return 0;
2260
2261out:
2262 kfree(priv);
2263 return rc;
2264}
2265
2266struct net_device *ipoib_intf_alloc(struct ib_device *hca, u32 port,
2267 const char *name)
2268{
2269 struct net_device *dev;
2270 int rc;
2271
2272 dev = ipoib_alloc_netdev(hca, port, name);
2273 if (IS_ERR(dev))
2274 return dev;
2275
2276 rc = ipoib_intf_init(hca, port, name, dev);
2277 if (rc) {
2278 free_netdev(dev);
2279 return ERR_PTR(rc);
2280 }
2281
2282 /*
2283 * Upon success the caller must ensure ipoib_intf_free is called or
2284 * register_netdevice succeed'd and priv_destructor is set to
2285 * ipoib_intf_free.
2286 */
2287 return dev;
2288}
2289
2290void ipoib_intf_free(struct net_device *dev)
2291{
2292 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2293 struct rdma_netdev *rn = netdev_priv(dev);
2294
2295 dev->priv_destructor = priv->next_priv_destructor;
2296 if (dev->priv_destructor)
2297 dev->priv_destructor(dev);
2298
2299 /*
2300 * There are some error flows around register_netdev failing that may
2301 * attempt to call priv_destructor twice, prevent that from happening.
2302 */
2303 dev->priv_destructor = NULL;
2304
2305 /* unregister/destroy is very complicated. Make bugs more obvious. */
2306 rn->clnt_priv = NULL;
2307
2308 kfree(priv);
2309}
2310
2311static ssize_t pkey_show(struct device *dev, struct device_attribute *attr,
2312 char *buf)
2313{
2314 struct net_device *ndev = to_net_dev(dev);
2315 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2316
2317 return sysfs_emit(buf, "0x%04x\n", priv->pkey);
2318}
2319static DEVICE_ATTR_RO(pkey);
2320
2321static ssize_t umcast_show(struct device *dev, struct device_attribute *attr,
2322 char *buf)
2323{
2324 struct net_device *ndev = to_net_dev(dev);
2325 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2326
2327 return sysfs_emit(buf, "%d\n",
2328 test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2329}
2330
2331void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2332{
2333 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2334
2335 if (umcast_val > 0) {
2336 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2337 ipoib_warn(priv, "ignoring multicast groups joined directly "
2338 "by userspace\n");
2339 } else
2340 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2341}
2342
2343static ssize_t umcast_store(struct device *dev, struct device_attribute *attr,
2344 const char *buf, size_t count)
2345{
2346 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2347
2348 ipoib_set_umcast(to_net_dev(dev), umcast_val);
2349
2350 return count;
2351}
2352static DEVICE_ATTR_RW(umcast);
2353
2354int ipoib_add_umcast_attr(struct net_device *dev)
2355{
2356 return device_create_file(&dev->dev, &dev_attr_umcast);
2357}
2358
2359static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2360{
2361 struct ipoib_dev_priv *child_priv;
2362 struct net_device *netdev = priv->dev;
2363
2364 netif_addr_lock_bh(netdev);
2365
2366 memcpy(&priv->local_gid.global.interface_id,
2367 &gid->global.interface_id,
2368 sizeof(gid->global.interface_id));
2369 dev_addr_mod(netdev, 4, (u8 *)&priv->local_gid, sizeof(priv->local_gid));
2370 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2371
2372 netif_addr_unlock_bh(netdev);
2373
2374 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2375 down_read(&priv->vlan_rwsem);
2376 list_for_each_entry(child_priv, &priv->child_intfs, list)
2377 set_base_guid(child_priv, gid);
2378 up_read(&priv->vlan_rwsem);
2379 }
2380}
2381
2382static int ipoib_check_lladdr(struct net_device *dev,
2383 struct sockaddr_storage *ss)
2384{
2385 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2386 int ret = 0;
2387
2388 netif_addr_lock_bh(dev);
2389
2390 /* Make sure the QPN, reserved and subnet prefix match the current
2391 * lladdr, it also makes sure the lladdr is unicast.
2392 */
2393 if (memcmp(dev->dev_addr, ss->__data,
2394 4 + sizeof(gid->global.subnet_prefix)) ||
2395 gid->global.interface_id == 0)
2396 ret = -EINVAL;
2397
2398 netif_addr_unlock_bh(dev);
2399
2400 return ret;
2401}
2402
2403static int ipoib_set_mac(struct net_device *dev, void *addr)
2404{
2405 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2406 struct sockaddr_storage *ss = addr;
2407 int ret;
2408
2409 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2410 return -EBUSY;
2411
2412 ret = ipoib_check_lladdr(dev, ss);
2413 if (ret)
2414 return ret;
2415
2416 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2417
2418 queue_work(ipoib_workqueue, &priv->flush_light);
2419
2420 return 0;
2421}
2422
2423static ssize_t create_child_store(struct device *dev,
2424 struct device_attribute *attr,
2425 const char *buf, size_t count)
2426{
2427 int pkey;
2428 int ret;
2429
2430 if (sscanf(buf, "%i", &pkey) != 1)
2431 return -EINVAL;
2432
2433 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2434 return -EINVAL;
2435
2436 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2437
2438 return ret ? ret : count;
2439}
2440static DEVICE_ATTR_WO(create_child);
2441
2442static ssize_t delete_child_store(struct device *dev,
2443 struct device_attribute *attr,
2444 const char *buf, size_t count)
2445{
2446 int pkey;
2447 int ret;
2448
2449 if (sscanf(buf, "%i", &pkey) != 1)
2450 return -EINVAL;
2451
2452 if (pkey < 0 || pkey > 0xffff)
2453 return -EINVAL;
2454
2455 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2456
2457 return ret ? ret : count;
2458
2459}
2460static DEVICE_ATTR_WO(delete_child);
2461
2462int ipoib_add_pkey_attr(struct net_device *dev)
2463{
2464 return device_create_file(&dev->dev, &dev_attr_pkey);
2465}
2466
2467/*
2468 * We erroneously exposed the iface's port number in the dev_id
2469 * sysfs field long after dev_port was introduced for that purpose[1],
2470 * and we need to stop everyone from relying on that.
2471 * Let's overload the shower routine for the dev_id file here
2472 * to gently bring the issue up.
2473 *
2474 * [1] https://www.spinics.net/lists/netdev/msg272123.html
2475 */
2476static ssize_t dev_id_show(struct device *dev,
2477 struct device_attribute *attr, char *buf)
2478{
2479 struct net_device *ndev = to_net_dev(dev);
2480
2481 /*
2482 * ndev->dev_port will be equal to 0 in old kernel prior to commit
2483 * 9b8b2a323008 ("IB/ipoib: Use dev_port to expose network interface
2484 * port numbers") Zero was chosen as special case for user space
2485 * applications to fallback and query dev_id to check if it has
2486 * different value or not.
2487 *
2488 * Don't print warning in such scenario.
2489 *
2490 * https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c#L358
2491 */
2492 if (ndev->dev_port && ndev->dev_id == ndev->dev_port)
2493 netdev_info_once(ndev,
2494 "\"%s\" wants to know my dev_id. Should it look at dev_port instead? See Documentation/ABI/testing/sysfs-class-net for more info.\n",
2495 current->comm);
2496
2497 return sysfs_emit(buf, "%#x\n", ndev->dev_id);
2498}
2499static DEVICE_ATTR_RO(dev_id);
2500
2501static int ipoib_intercept_dev_id_attr(struct net_device *dev)
2502{
2503 device_remove_file(&dev->dev, &dev_attr_dev_id);
2504 return device_create_file(&dev->dev, &dev_attr_dev_id);
2505}
2506
2507static struct net_device *ipoib_add_port(const char *format,
2508 struct ib_device *hca, u32 port)
2509{
2510 struct rtnl_link_ops *ops = ipoib_get_link_ops();
2511 struct rdma_netdev_alloc_params params;
2512 struct ipoib_dev_priv *priv;
2513 struct net_device *ndev;
2514 int result;
2515
2516 ndev = ipoib_intf_alloc(hca, port, format);
2517 if (IS_ERR(ndev)) {
2518 pr_warn("%s, %d: ipoib_intf_alloc failed %ld\n", hca->name, port,
2519 PTR_ERR(ndev));
2520 return ndev;
2521 }
2522 priv = ipoib_priv(ndev);
2523
2524 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2525 priv->ca, ipoib_event);
2526 ib_register_event_handler(&priv->event_handler);
2527
2528 /* call event handler to ensure pkey in sync */
2529 queue_work(ipoib_workqueue, &priv->flush_heavy);
2530
2531 ndev->rtnl_link_ops = ipoib_get_link_ops();
2532
2533 result = register_netdev(ndev);
2534 if (result) {
2535 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2536 hca->name, port, result);
2537
2538 ipoib_parent_unregister_pre(ndev);
2539 ipoib_intf_free(ndev);
2540 free_netdev(ndev);
2541
2542 return ERR_PTR(result);
2543 }
2544
2545 if (hca->ops.rdma_netdev_get_params) {
2546 int rc = hca->ops.rdma_netdev_get_params(hca, port,
2547 RDMA_NETDEV_IPOIB,
2548 ¶ms);
2549
2550 if (!rc && ops->priv_size < params.sizeof_priv)
2551 ops->priv_size = params.sizeof_priv;
2552 }
2553 /*
2554 * We cannot set priv_destructor before register_netdev because we
2555 * need priv to be always valid during the error flow to execute
2556 * ipoib_parent_unregister_pre(). Instead handle it manually and only
2557 * enter priv_destructor mode once we are completely registered.
2558 */
2559 ndev->priv_destructor = ipoib_intf_free;
2560
2561 if (ipoib_intercept_dev_id_attr(ndev))
2562 goto sysfs_failed;
2563 if (ipoib_cm_add_mode_attr(ndev))
2564 goto sysfs_failed;
2565 if (ipoib_add_pkey_attr(ndev))
2566 goto sysfs_failed;
2567 if (ipoib_add_umcast_attr(ndev))
2568 goto sysfs_failed;
2569 if (device_create_file(&ndev->dev, &dev_attr_create_child))
2570 goto sysfs_failed;
2571 if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2572 goto sysfs_failed;
2573
2574 return ndev;
2575
2576sysfs_failed:
2577 ipoib_parent_unregister_pre(ndev);
2578 unregister_netdev(ndev);
2579 return ERR_PTR(-ENOMEM);
2580}
2581
2582static int ipoib_add_one(struct ib_device *device)
2583{
2584 struct list_head *dev_list;
2585 struct net_device *dev;
2586 struct ipoib_dev_priv *priv;
2587 unsigned int p;
2588 int count = 0;
2589
2590 dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2591 if (!dev_list)
2592 return -ENOMEM;
2593
2594 INIT_LIST_HEAD(dev_list);
2595
2596 rdma_for_each_port (device, p) {
2597 if (!rdma_protocol_ib(device, p))
2598 continue;
2599 dev = ipoib_add_port("ib%d", device, p);
2600 if (!IS_ERR(dev)) {
2601 priv = ipoib_priv(dev);
2602 list_add_tail(&priv->list, dev_list);
2603 count++;
2604 }
2605 }
2606
2607 if (!count) {
2608 kfree(dev_list);
2609 return -EOPNOTSUPP;
2610 }
2611
2612 ib_set_client_data(device, &ipoib_client, dev_list);
2613 return 0;
2614}
2615
2616static void ipoib_remove_one(struct ib_device *device, void *client_data)
2617{
2618 struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2619 struct list_head *dev_list = client_data;
2620
2621 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2622 LIST_HEAD(head);
2623 ipoib_parent_unregister_pre(priv->dev);
2624
2625 rtnl_lock();
2626
2627 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2628 list)
2629 unregister_netdevice_queue(cpriv->dev, &head);
2630 unregister_netdevice_queue(priv->dev, &head);
2631 unregister_netdevice_many(&head);
2632
2633 rtnl_unlock();
2634 }
2635
2636 kfree(dev_list);
2637}
2638
2639#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2640static struct notifier_block ipoib_netdev_notifier = {
2641 .notifier_call = ipoib_netdev_event,
2642};
2643#endif
2644
2645static int __init ipoib_init_module(void)
2646{
2647 int ret;
2648
2649 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2650 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2651 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2652
2653 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2654 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2655 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2656#ifdef CONFIG_INFINIBAND_IPOIB_CM
2657 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2658 ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2659#endif
2660
2661 /*
2662 * When copying small received packets, we only copy from the
2663 * linear data part of the SKB, so we rely on this condition.
2664 */
2665 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2666
2667 ipoib_register_debugfs();
2668
2669 /*
2670 * We create a global workqueue here that is used for all flush
2671 * operations. However, if you attempt to flush a workqueue
2672 * from a task on that same workqueue, it deadlocks the system.
2673 * We want to be able to flush the tasks associated with a
2674 * specific net device, so we also create a workqueue for each
2675 * netdevice. We queue up the tasks for that device only on
2676 * its private workqueue, and we only queue up flush events
2677 * on our global flush workqueue. This avoids the deadlocks.
2678 */
2679 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2680 if (!ipoib_workqueue) {
2681 ret = -ENOMEM;
2682 goto err_fs;
2683 }
2684
2685 ib_sa_register_client(&ipoib_sa_client);
2686
2687 ret = ib_register_client(&ipoib_client);
2688 if (ret)
2689 goto err_sa;
2690
2691 ret = ipoib_netlink_init();
2692 if (ret)
2693 goto err_client;
2694
2695#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2696 register_netdevice_notifier(&ipoib_netdev_notifier);
2697#endif
2698 return 0;
2699
2700err_client:
2701 ib_unregister_client(&ipoib_client);
2702
2703err_sa:
2704 ib_sa_unregister_client(&ipoib_sa_client);
2705 destroy_workqueue(ipoib_workqueue);
2706
2707err_fs:
2708 ipoib_unregister_debugfs();
2709
2710 return ret;
2711}
2712
2713static void __exit ipoib_cleanup_module(void)
2714{
2715#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2716 unregister_netdevice_notifier(&ipoib_netdev_notifier);
2717#endif
2718 ipoib_netlink_fini();
2719 ib_unregister_client(&ipoib_client);
2720 ib_sa_unregister_client(&ipoib_sa_client);
2721 ipoib_unregister_debugfs();
2722 destroy_workqueue(ipoib_workqueue);
2723}
2724
2725module_init(ipoib_init_module);
2726module_exit(ipoib_cleanup_module);
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 <linux/jhash.h>
50#include <net/arp.h>
51#include <net/addrconf.h>
52#include <linux/inetdevice.h>
53#include <rdma/ib_cache.h>
54#include <linux/pci.h>
55
56#define DRV_VERSION "1.0.0"
57
58const char ipoib_driver_version[] = DRV_VERSION;
59
60MODULE_AUTHOR("Roland Dreier");
61MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
62MODULE_LICENSE("Dual BSD/GPL");
63MODULE_VERSION(DRV_VERSION);
64
65int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
66int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
67
68module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
69MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
70module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
71MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
72
73#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
74int ipoib_debug_level;
75
76module_param_named(debug_level, ipoib_debug_level, int, 0644);
77MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
78#endif
79
80struct ipoib_path_iter {
81 struct net_device *dev;
82 struct ipoib_path path;
83};
84
85static const u8 ipv4_bcast_addr[] = {
86 0x00, 0xff, 0xff, 0xff,
87 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
88 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
89};
90
91struct workqueue_struct *ipoib_workqueue;
92
93struct ib_sa_client ipoib_sa_client;
94
95static void ipoib_add_one(struct ib_device *device);
96static void ipoib_remove_one(struct ib_device *device, void *client_data);
97static void ipoib_neigh_reclaim(struct rcu_head *rp);
98static struct net_device *ipoib_get_net_dev_by_params(
99 struct ib_device *dev, u8 port, u16 pkey,
100 const union ib_gid *gid, const struct sockaddr *addr,
101 void *client_data);
102
103static struct ib_client ipoib_client = {
104 .name = "ipoib",
105 .add = ipoib_add_one,
106 .remove = ipoib_remove_one,
107 .get_net_dev_by_params = ipoib_get_net_dev_by_params,
108};
109
110int ipoib_open(struct net_device *dev)
111{
112 struct ipoib_dev_priv *priv = netdev_priv(dev);
113
114 ipoib_dbg(priv, "bringing up interface\n");
115
116 netif_carrier_off(dev);
117
118 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
119
120 if (ipoib_ib_dev_open(dev)) {
121 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
122 return 0;
123 goto err_disable;
124 }
125
126 if (ipoib_ib_dev_up(dev))
127 goto err_stop;
128
129 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
130 struct ipoib_dev_priv *cpriv;
131
132 /* Bring up any child interfaces too */
133 down_read(&priv->vlan_rwsem);
134 list_for_each_entry(cpriv, &priv->child_intfs, list) {
135 int flags;
136
137 flags = cpriv->dev->flags;
138 if (flags & IFF_UP)
139 continue;
140
141 dev_change_flags(cpriv->dev, flags | IFF_UP);
142 }
143 up_read(&priv->vlan_rwsem);
144 }
145
146 netif_start_queue(dev);
147
148 return 0;
149
150err_stop:
151 ipoib_ib_dev_stop(dev);
152
153err_disable:
154 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
155
156 return -EINVAL;
157}
158
159static int ipoib_stop(struct net_device *dev)
160{
161 struct ipoib_dev_priv *priv = netdev_priv(dev);
162
163 ipoib_dbg(priv, "stopping interface\n");
164
165 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
166
167 netif_stop_queue(dev);
168
169 ipoib_ib_dev_down(dev);
170 ipoib_ib_dev_stop(dev);
171
172 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
173 struct ipoib_dev_priv *cpriv;
174
175 /* Bring down any child interfaces too */
176 down_read(&priv->vlan_rwsem);
177 list_for_each_entry(cpriv, &priv->child_intfs, list) {
178 int flags;
179
180 flags = cpriv->dev->flags;
181 if (!(flags & IFF_UP))
182 continue;
183
184 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
185 }
186 up_read(&priv->vlan_rwsem);
187 }
188
189 return 0;
190}
191
192static void ipoib_uninit(struct net_device *dev)
193{
194 ipoib_dev_cleanup(dev);
195}
196
197static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
198{
199 struct ipoib_dev_priv *priv = netdev_priv(dev);
200
201 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
202 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
203
204 return features;
205}
206
207static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
208{
209 struct ipoib_dev_priv *priv = netdev_priv(dev);
210
211 /* dev->mtu > 2K ==> connected mode */
212 if (ipoib_cm_admin_enabled(dev)) {
213 if (new_mtu > ipoib_cm_max_mtu(dev))
214 return -EINVAL;
215
216 if (new_mtu > priv->mcast_mtu)
217 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
218 priv->mcast_mtu);
219
220 dev->mtu = new_mtu;
221 return 0;
222 }
223
224 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
225 return -EINVAL;
226
227 priv->admin_mtu = new_mtu;
228
229 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
230
231 return 0;
232}
233
234/* Called with an RCU read lock taken */
235static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
236 struct net_device *dev)
237{
238 struct net *net = dev_net(dev);
239 struct in_device *in_dev;
240 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
241 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
242 __be32 ret_addr;
243
244 switch (addr->sa_family) {
245 case AF_INET:
246 in_dev = in_dev_get(dev);
247 if (!in_dev)
248 return false;
249
250 ret_addr = inet_confirm_addr(net, in_dev, 0,
251 addr_in->sin_addr.s_addr,
252 RT_SCOPE_HOST);
253 in_dev_put(in_dev);
254 if (ret_addr)
255 return true;
256
257 break;
258 case AF_INET6:
259 if (IS_ENABLED(CONFIG_IPV6) &&
260 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
261 return true;
262
263 break;
264 }
265 return false;
266}
267
268/**
269 * Find the master net_device on top of the given net_device.
270 * @dev: base IPoIB net_device
271 *
272 * Returns the master net_device with a reference held, or the same net_device
273 * if no master exists.
274 */
275static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
276{
277 struct net_device *master;
278
279 rcu_read_lock();
280 master = netdev_master_upper_dev_get_rcu(dev);
281 if (master)
282 dev_hold(master);
283 rcu_read_unlock();
284
285 if (master)
286 return master;
287
288 dev_hold(dev);
289 return dev;
290}
291
292/**
293 * Find a net_device matching the given address, which is an upper device of
294 * the given net_device.
295 * @addr: IP address to look for.
296 * @dev: base IPoIB net_device
297 *
298 * If found, returns the net_device with a reference held. Otherwise return
299 * NULL.
300 */
301static struct net_device *ipoib_get_net_dev_match_addr(
302 const struct sockaddr *addr, struct net_device *dev)
303{
304 struct net_device *upper,
305 *result = NULL;
306 struct list_head *iter;
307
308 rcu_read_lock();
309 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
310 dev_hold(dev);
311 result = dev;
312 goto out;
313 }
314
315 netdev_for_each_all_upper_dev_rcu(dev, upper, iter) {
316 if (ipoib_is_dev_match_addr_rcu(addr, upper)) {
317 dev_hold(upper);
318 result = upper;
319 break;
320 }
321 }
322out:
323 rcu_read_unlock();
324 return result;
325}
326
327/* returns the number of IPoIB netdevs on top a given ipoib device matching a
328 * pkey_index and address, if one exists.
329 *
330 * @found_net_dev: contains a matching net_device if the return value >= 1,
331 * with a reference held. */
332static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
333 const union ib_gid *gid,
334 u16 pkey_index,
335 const struct sockaddr *addr,
336 int nesting,
337 struct net_device **found_net_dev)
338{
339 struct ipoib_dev_priv *child_priv;
340 struct net_device *net_dev = NULL;
341 int matches = 0;
342
343 if (priv->pkey_index == pkey_index &&
344 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
345 if (!addr) {
346 net_dev = ipoib_get_master_net_dev(priv->dev);
347 } else {
348 /* Verify the net_device matches the IP address, as
349 * IPoIB child devices currently share a GID. */
350 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
351 }
352 if (net_dev) {
353 if (!*found_net_dev)
354 *found_net_dev = net_dev;
355 else
356 dev_put(net_dev);
357 ++matches;
358 }
359 }
360
361 /* Check child interfaces */
362 down_read_nested(&priv->vlan_rwsem, nesting);
363 list_for_each_entry(child_priv, &priv->child_intfs, list) {
364 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
365 pkey_index, addr,
366 nesting + 1,
367 found_net_dev);
368 if (matches > 1)
369 break;
370 }
371 up_read(&priv->vlan_rwsem);
372
373 return matches;
374}
375
376/* Returns the number of matching net_devs found (between 0 and 2). Also
377 * return the matching net_device in the @net_dev parameter, holding a
378 * reference to the net_device, if the number of matches >= 1 */
379static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
380 u16 pkey_index,
381 const union ib_gid *gid,
382 const struct sockaddr *addr,
383 struct net_device **net_dev)
384{
385 struct ipoib_dev_priv *priv;
386 int matches = 0;
387
388 *net_dev = NULL;
389
390 list_for_each_entry(priv, dev_list, list) {
391 if (priv->port != port)
392 continue;
393
394 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
395 addr, 0, net_dev);
396 if (matches > 1)
397 break;
398 }
399
400 return matches;
401}
402
403static struct net_device *ipoib_get_net_dev_by_params(
404 struct ib_device *dev, u8 port, u16 pkey,
405 const union ib_gid *gid, const struct sockaddr *addr,
406 void *client_data)
407{
408 struct net_device *net_dev;
409 struct list_head *dev_list = client_data;
410 u16 pkey_index;
411 int matches;
412 int ret;
413
414 if (!rdma_protocol_ib(dev, port))
415 return NULL;
416
417 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
418 if (ret)
419 return NULL;
420
421 if (!dev_list)
422 return NULL;
423
424 /* See if we can find a unique device matching the L2 parameters */
425 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
426 gid, NULL, &net_dev);
427
428 switch (matches) {
429 case 0:
430 return NULL;
431 case 1:
432 return net_dev;
433 }
434
435 dev_put(net_dev);
436
437 /* Couldn't find a unique device with L2 parameters only. Use L3
438 * address to uniquely match the net device */
439 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
440 gid, addr, &net_dev);
441 switch (matches) {
442 case 0:
443 return NULL;
444 default:
445 dev_warn_ratelimited(&dev->dev,
446 "duplicate IP address detected\n");
447 /* Fall through */
448 case 1:
449 return net_dev;
450 }
451}
452
453int ipoib_set_mode(struct net_device *dev, const char *buf)
454{
455 struct ipoib_dev_priv *priv = netdev_priv(dev);
456
457 /* flush paths if we switch modes so that connections are restarted */
458 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
459 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
460 ipoib_warn(priv, "enabling connected mode "
461 "will cause multicast packet drops\n");
462 netdev_update_features(dev);
463 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
464 rtnl_unlock();
465 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
466
467 ipoib_flush_paths(dev);
468 rtnl_lock();
469 return 0;
470 }
471
472 if (!strcmp(buf, "datagram\n")) {
473 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
474 netdev_update_features(dev);
475 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
476 rtnl_unlock();
477 ipoib_flush_paths(dev);
478 rtnl_lock();
479 return 0;
480 }
481
482 return -EINVAL;
483}
484
485static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
486{
487 struct ipoib_dev_priv *priv = netdev_priv(dev);
488 struct rb_node *n = priv->path_tree.rb_node;
489 struct ipoib_path *path;
490 int ret;
491
492 while (n) {
493 path = rb_entry(n, struct ipoib_path, rb_node);
494
495 ret = memcmp(gid, path->pathrec.dgid.raw,
496 sizeof (union ib_gid));
497
498 if (ret < 0)
499 n = n->rb_left;
500 else if (ret > 0)
501 n = n->rb_right;
502 else
503 return path;
504 }
505
506 return NULL;
507}
508
509static int __path_add(struct net_device *dev, struct ipoib_path *path)
510{
511 struct ipoib_dev_priv *priv = netdev_priv(dev);
512 struct rb_node **n = &priv->path_tree.rb_node;
513 struct rb_node *pn = NULL;
514 struct ipoib_path *tpath;
515 int ret;
516
517 while (*n) {
518 pn = *n;
519 tpath = rb_entry(pn, struct ipoib_path, rb_node);
520
521 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
522 sizeof (union ib_gid));
523 if (ret < 0)
524 n = &pn->rb_left;
525 else if (ret > 0)
526 n = &pn->rb_right;
527 else
528 return -EEXIST;
529 }
530
531 rb_link_node(&path->rb_node, pn, n);
532 rb_insert_color(&path->rb_node, &priv->path_tree);
533
534 list_add_tail(&path->list, &priv->path_list);
535
536 return 0;
537}
538
539static void path_free(struct net_device *dev, struct ipoib_path *path)
540{
541 struct sk_buff *skb;
542
543 while ((skb = __skb_dequeue(&path->queue)))
544 dev_kfree_skb_irq(skb);
545
546 ipoib_dbg(netdev_priv(dev), "path_free\n");
547
548 /* remove all neigh connected to this path */
549 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
550
551 if (path->ah)
552 ipoib_put_ah(path->ah);
553
554 kfree(path);
555}
556
557#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
558
559struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
560{
561 struct ipoib_path_iter *iter;
562
563 iter = kmalloc(sizeof *iter, GFP_KERNEL);
564 if (!iter)
565 return NULL;
566
567 iter->dev = dev;
568 memset(iter->path.pathrec.dgid.raw, 0, 16);
569
570 if (ipoib_path_iter_next(iter)) {
571 kfree(iter);
572 return NULL;
573 }
574
575 return iter;
576}
577
578int ipoib_path_iter_next(struct ipoib_path_iter *iter)
579{
580 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
581 struct rb_node *n;
582 struct ipoib_path *path;
583 int ret = 1;
584
585 spin_lock_irq(&priv->lock);
586
587 n = rb_first(&priv->path_tree);
588
589 while (n) {
590 path = rb_entry(n, struct ipoib_path, rb_node);
591
592 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
593 sizeof (union ib_gid)) < 0) {
594 iter->path = *path;
595 ret = 0;
596 break;
597 }
598
599 n = rb_next(n);
600 }
601
602 spin_unlock_irq(&priv->lock);
603
604 return ret;
605}
606
607void ipoib_path_iter_read(struct ipoib_path_iter *iter,
608 struct ipoib_path *path)
609{
610 *path = iter->path;
611}
612
613#endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
614
615void ipoib_mark_paths_invalid(struct net_device *dev)
616{
617 struct ipoib_dev_priv *priv = netdev_priv(dev);
618 struct ipoib_path *path, *tp;
619
620 spin_lock_irq(&priv->lock);
621
622 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
623 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
624 be16_to_cpu(path->pathrec.dlid),
625 path->pathrec.dgid.raw);
626 path->valid = 0;
627 }
628
629 spin_unlock_irq(&priv->lock);
630}
631
632void ipoib_flush_paths(struct net_device *dev)
633{
634 struct ipoib_dev_priv *priv = netdev_priv(dev);
635 struct ipoib_path *path, *tp;
636 LIST_HEAD(remove_list);
637 unsigned long flags;
638
639 netif_tx_lock_bh(dev);
640 spin_lock_irqsave(&priv->lock, flags);
641
642 list_splice_init(&priv->path_list, &remove_list);
643
644 list_for_each_entry(path, &remove_list, list)
645 rb_erase(&path->rb_node, &priv->path_tree);
646
647 list_for_each_entry_safe(path, tp, &remove_list, list) {
648 if (path->query)
649 ib_sa_cancel_query(path->query_id, path->query);
650 spin_unlock_irqrestore(&priv->lock, flags);
651 netif_tx_unlock_bh(dev);
652 wait_for_completion(&path->done);
653 path_free(dev, path);
654 netif_tx_lock_bh(dev);
655 spin_lock_irqsave(&priv->lock, flags);
656 }
657
658 spin_unlock_irqrestore(&priv->lock, flags);
659 netif_tx_unlock_bh(dev);
660}
661
662static void path_rec_completion(int status,
663 struct ib_sa_path_rec *pathrec,
664 void *path_ptr)
665{
666 struct ipoib_path *path = path_ptr;
667 struct net_device *dev = path->dev;
668 struct ipoib_dev_priv *priv = netdev_priv(dev);
669 struct ipoib_ah *ah = NULL;
670 struct ipoib_ah *old_ah = NULL;
671 struct ipoib_neigh *neigh, *tn;
672 struct sk_buff_head skqueue;
673 struct sk_buff *skb;
674 unsigned long flags;
675
676 if (!status)
677 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
678 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
679 else
680 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
681 status, path->pathrec.dgid.raw);
682
683 skb_queue_head_init(&skqueue);
684
685 if (!status) {
686 struct ib_ah_attr av;
687
688 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
689 ah = ipoib_create_ah(dev, priv->pd, &av);
690 }
691
692 spin_lock_irqsave(&priv->lock, flags);
693
694 if (!IS_ERR_OR_NULL(ah)) {
695 path->pathrec = *pathrec;
696
697 old_ah = path->ah;
698 path->ah = ah;
699
700 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
701 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
702
703 while ((skb = __skb_dequeue(&path->queue)))
704 __skb_queue_tail(&skqueue, skb);
705
706 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
707 if (neigh->ah) {
708 WARN_ON(neigh->ah != old_ah);
709 /*
710 * Dropping the ah reference inside
711 * priv->lock is safe here, because we
712 * will hold one more reference from
713 * the original value of path->ah (ie
714 * old_ah).
715 */
716 ipoib_put_ah(neigh->ah);
717 }
718 kref_get(&path->ah->ref);
719 neigh->ah = path->ah;
720
721 if (ipoib_cm_enabled(dev, neigh->daddr)) {
722 if (!ipoib_cm_get(neigh))
723 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
724 path,
725 neigh));
726 if (!ipoib_cm_get(neigh)) {
727 ipoib_neigh_free(neigh);
728 continue;
729 }
730 }
731
732 while ((skb = __skb_dequeue(&neigh->queue)))
733 __skb_queue_tail(&skqueue, skb);
734 }
735 path->valid = 1;
736 }
737
738 path->query = NULL;
739 complete(&path->done);
740
741 spin_unlock_irqrestore(&priv->lock, flags);
742
743 if (IS_ERR_OR_NULL(ah))
744 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
745
746 if (old_ah)
747 ipoib_put_ah(old_ah);
748
749 while ((skb = __skb_dequeue(&skqueue))) {
750 skb->dev = dev;
751 if (dev_queue_xmit(skb))
752 ipoib_warn(priv, "dev_queue_xmit failed "
753 "to requeue packet\n");
754 }
755}
756
757static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
758{
759 struct ipoib_dev_priv *priv = netdev_priv(dev);
760 struct ipoib_path *path;
761
762 if (!priv->broadcast)
763 return NULL;
764
765 path = kzalloc(sizeof *path, GFP_ATOMIC);
766 if (!path)
767 return NULL;
768
769 path->dev = dev;
770
771 skb_queue_head_init(&path->queue);
772
773 INIT_LIST_HEAD(&path->neigh_list);
774
775 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
776 path->pathrec.sgid = priv->local_gid;
777 path->pathrec.pkey = cpu_to_be16(priv->pkey);
778 path->pathrec.numb_path = 1;
779 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
780
781 return path;
782}
783
784static int path_rec_start(struct net_device *dev,
785 struct ipoib_path *path)
786{
787 struct ipoib_dev_priv *priv = netdev_priv(dev);
788
789 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
790 path->pathrec.dgid.raw);
791
792 init_completion(&path->done);
793
794 path->query_id =
795 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
796 &path->pathrec,
797 IB_SA_PATH_REC_DGID |
798 IB_SA_PATH_REC_SGID |
799 IB_SA_PATH_REC_NUMB_PATH |
800 IB_SA_PATH_REC_TRAFFIC_CLASS |
801 IB_SA_PATH_REC_PKEY,
802 1000, GFP_ATOMIC,
803 path_rec_completion,
804 path, &path->query);
805 if (path->query_id < 0) {
806 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
807 path->query = NULL;
808 complete(&path->done);
809 return path->query_id;
810 }
811
812 return 0;
813}
814
815static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
816 struct net_device *dev)
817{
818 struct ipoib_dev_priv *priv = netdev_priv(dev);
819 struct ipoib_path *path;
820 struct ipoib_neigh *neigh;
821 unsigned long flags;
822
823 spin_lock_irqsave(&priv->lock, flags);
824 neigh = ipoib_neigh_alloc(daddr, dev);
825 if (!neigh) {
826 spin_unlock_irqrestore(&priv->lock, flags);
827 ++dev->stats.tx_dropped;
828 dev_kfree_skb_any(skb);
829 return;
830 }
831
832 path = __path_find(dev, daddr + 4);
833 if (!path) {
834 path = path_rec_create(dev, daddr + 4);
835 if (!path)
836 goto err_path;
837
838 __path_add(dev, path);
839 }
840
841 list_add_tail(&neigh->list, &path->neigh_list);
842
843 if (path->ah) {
844 kref_get(&path->ah->ref);
845 neigh->ah = path->ah;
846
847 if (ipoib_cm_enabled(dev, neigh->daddr)) {
848 if (!ipoib_cm_get(neigh))
849 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
850 if (!ipoib_cm_get(neigh)) {
851 ipoib_neigh_free(neigh);
852 goto err_drop;
853 }
854 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
855 __skb_queue_tail(&neigh->queue, skb);
856 else {
857 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
858 skb_queue_len(&neigh->queue));
859 goto err_drop;
860 }
861 } else {
862 spin_unlock_irqrestore(&priv->lock, flags);
863 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
864 ipoib_neigh_put(neigh);
865 return;
866 }
867 } else {
868 neigh->ah = NULL;
869
870 if (!path->query && path_rec_start(dev, path))
871 goto err_path;
872 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
873 __skb_queue_tail(&neigh->queue, skb);
874 else
875 goto err_drop;
876 }
877
878 spin_unlock_irqrestore(&priv->lock, flags);
879 ipoib_neigh_put(neigh);
880 return;
881
882err_path:
883 ipoib_neigh_free(neigh);
884err_drop:
885 ++dev->stats.tx_dropped;
886 dev_kfree_skb_any(skb);
887
888 spin_unlock_irqrestore(&priv->lock, flags);
889 ipoib_neigh_put(neigh);
890}
891
892static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
893 struct ipoib_cb *cb)
894{
895 struct ipoib_dev_priv *priv = netdev_priv(dev);
896 struct ipoib_path *path;
897 unsigned long flags;
898
899 spin_lock_irqsave(&priv->lock, flags);
900
901 path = __path_find(dev, cb->hwaddr + 4);
902 if (!path || !path->valid) {
903 int new_path = 0;
904
905 if (!path) {
906 path = path_rec_create(dev, cb->hwaddr + 4);
907 new_path = 1;
908 }
909 if (path) {
910 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
911 __skb_queue_tail(&path->queue, skb);
912 } else {
913 ++dev->stats.tx_dropped;
914 dev_kfree_skb_any(skb);
915 }
916
917 if (!path->query && path_rec_start(dev, path)) {
918 spin_unlock_irqrestore(&priv->lock, flags);
919 if (new_path)
920 path_free(dev, path);
921 return;
922 } else
923 __path_add(dev, path);
924 } else {
925 ++dev->stats.tx_dropped;
926 dev_kfree_skb_any(skb);
927 }
928
929 spin_unlock_irqrestore(&priv->lock, flags);
930 return;
931 }
932
933 if (path->ah) {
934 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
935 be16_to_cpu(path->pathrec.dlid));
936
937 spin_unlock_irqrestore(&priv->lock, flags);
938 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
939 return;
940 } else if ((path->query || !path_rec_start(dev, path)) &&
941 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
942 __skb_queue_tail(&path->queue, skb);
943 } else {
944 ++dev->stats.tx_dropped;
945 dev_kfree_skb_any(skb);
946 }
947
948 spin_unlock_irqrestore(&priv->lock, flags);
949}
950
951static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
952{
953 struct ipoib_dev_priv *priv = netdev_priv(dev);
954 struct ipoib_neigh *neigh;
955 struct ipoib_cb *cb = ipoib_skb_cb(skb);
956 struct ipoib_header *header;
957 unsigned long flags;
958
959 header = (struct ipoib_header *) skb->data;
960
961 if (unlikely(cb->hwaddr[4] == 0xff)) {
962 /* multicast, arrange "if" according to probability */
963 if ((header->proto != htons(ETH_P_IP)) &&
964 (header->proto != htons(ETH_P_IPV6)) &&
965 (header->proto != htons(ETH_P_ARP)) &&
966 (header->proto != htons(ETH_P_RARP)) &&
967 (header->proto != htons(ETH_P_TIPC))) {
968 /* ethertype not supported by IPoIB */
969 ++dev->stats.tx_dropped;
970 dev_kfree_skb_any(skb);
971 return NETDEV_TX_OK;
972 }
973 /* Add in the P_Key for multicast*/
974 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
975 cb->hwaddr[9] = priv->pkey & 0xff;
976
977 neigh = ipoib_neigh_get(dev, cb->hwaddr);
978 if (likely(neigh))
979 goto send_using_neigh;
980 ipoib_mcast_send(dev, cb->hwaddr, skb);
981 return NETDEV_TX_OK;
982 }
983
984 /* unicast, arrange "switch" according to probability */
985 switch (header->proto) {
986 case htons(ETH_P_IP):
987 case htons(ETH_P_IPV6):
988 case htons(ETH_P_TIPC):
989 neigh = ipoib_neigh_get(dev, cb->hwaddr);
990 if (unlikely(!neigh)) {
991 neigh_add_path(skb, cb->hwaddr, dev);
992 return NETDEV_TX_OK;
993 }
994 break;
995 case htons(ETH_P_ARP):
996 case htons(ETH_P_RARP):
997 /* for unicast ARP and RARP should always perform path find */
998 unicast_arp_send(skb, dev, cb);
999 return NETDEV_TX_OK;
1000 default:
1001 /* ethertype not supported by IPoIB */
1002 ++dev->stats.tx_dropped;
1003 dev_kfree_skb_any(skb);
1004 return NETDEV_TX_OK;
1005 }
1006
1007send_using_neigh:
1008 /* note we now hold a ref to neigh */
1009 if (ipoib_cm_get(neigh)) {
1010 if (ipoib_cm_up(neigh)) {
1011 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1012 goto unref;
1013 }
1014 } else if (neigh->ah) {
1015 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
1016 goto unref;
1017 }
1018
1019 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1020 spin_lock_irqsave(&priv->lock, flags);
1021 __skb_queue_tail(&neigh->queue, skb);
1022 spin_unlock_irqrestore(&priv->lock, flags);
1023 } else {
1024 ++dev->stats.tx_dropped;
1025 dev_kfree_skb_any(skb);
1026 }
1027
1028unref:
1029 ipoib_neigh_put(neigh);
1030
1031 return NETDEV_TX_OK;
1032}
1033
1034static void ipoib_timeout(struct net_device *dev)
1035{
1036 struct ipoib_dev_priv *priv = netdev_priv(dev);
1037
1038 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1039 jiffies_to_msecs(jiffies - dev->trans_start));
1040 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1041 netif_queue_stopped(dev),
1042 priv->tx_head, priv->tx_tail);
1043 /* XXX reset QP, etc. */
1044}
1045
1046static int ipoib_hard_header(struct sk_buff *skb,
1047 struct net_device *dev,
1048 unsigned short type,
1049 const void *daddr, const void *saddr, unsigned len)
1050{
1051 struct ipoib_header *header;
1052 struct ipoib_cb *cb = ipoib_skb_cb(skb);
1053
1054 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1055
1056 header->proto = htons(type);
1057 header->reserved = 0;
1058
1059 /*
1060 * we don't rely on dst_entry structure, always stuff the
1061 * destination address into skb->cb so we can figure out where
1062 * to send the packet later.
1063 */
1064 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
1065
1066 return sizeof *header;
1067}
1068
1069static void ipoib_set_mcast_list(struct net_device *dev)
1070{
1071 struct ipoib_dev_priv *priv = netdev_priv(dev);
1072
1073 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1074 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1075 return;
1076 }
1077
1078 queue_work(priv->wq, &priv->restart_task);
1079}
1080
1081static int ipoib_get_iflink(const struct net_device *dev)
1082{
1083 struct ipoib_dev_priv *priv = netdev_priv(dev);
1084
1085 /* parent interface */
1086 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1087 return dev->ifindex;
1088
1089 /* child/vlan interface */
1090 return priv->parent->ifindex;
1091}
1092
1093static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1094{
1095 /*
1096 * Use only the address parts that contributes to spreading
1097 * The subnet prefix is not used as one can not connect to
1098 * same remote port (GUID) using the same remote QPN via two
1099 * different subnets.
1100 */
1101 /* qpn octets[1:4) & port GUID octets[12:20) */
1102 u32 *d32 = (u32 *) daddr;
1103 u32 hv;
1104
1105 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1106 return hv & htbl->mask;
1107}
1108
1109struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1110{
1111 struct ipoib_dev_priv *priv = netdev_priv(dev);
1112 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1113 struct ipoib_neigh_hash *htbl;
1114 struct ipoib_neigh *neigh = NULL;
1115 u32 hash_val;
1116
1117 rcu_read_lock_bh();
1118
1119 htbl = rcu_dereference_bh(ntbl->htbl);
1120
1121 if (!htbl)
1122 goto out_unlock;
1123
1124 hash_val = ipoib_addr_hash(htbl, daddr);
1125 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1126 neigh != NULL;
1127 neigh = rcu_dereference_bh(neigh->hnext)) {
1128 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1129 /* found, take one ref on behalf of the caller */
1130 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1131 /* deleted */
1132 neigh = NULL;
1133 goto out_unlock;
1134 }
1135 neigh->alive = jiffies;
1136 goto out_unlock;
1137 }
1138 }
1139
1140out_unlock:
1141 rcu_read_unlock_bh();
1142 return neigh;
1143}
1144
1145static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1146{
1147 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1148 struct ipoib_neigh_hash *htbl;
1149 unsigned long neigh_obsolete;
1150 unsigned long dt;
1151 unsigned long flags;
1152 int i;
1153 LIST_HEAD(remove_list);
1154
1155 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1156 return;
1157
1158 spin_lock_irqsave(&priv->lock, flags);
1159
1160 htbl = rcu_dereference_protected(ntbl->htbl,
1161 lockdep_is_held(&priv->lock));
1162
1163 if (!htbl)
1164 goto out_unlock;
1165
1166 /* neigh is obsolete if it was idle for two GC periods */
1167 dt = 2 * arp_tbl.gc_interval;
1168 neigh_obsolete = jiffies - dt;
1169 /* handle possible race condition */
1170 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1171 goto out_unlock;
1172
1173 for (i = 0; i < htbl->size; i++) {
1174 struct ipoib_neigh *neigh;
1175 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1176
1177 while ((neigh = rcu_dereference_protected(*np,
1178 lockdep_is_held(&priv->lock))) != NULL) {
1179 /* was the neigh idle for two GC periods */
1180 if (time_after(neigh_obsolete, neigh->alive)) {
1181
1182 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1183
1184 rcu_assign_pointer(*np,
1185 rcu_dereference_protected(neigh->hnext,
1186 lockdep_is_held(&priv->lock)));
1187 /* remove from path/mc list */
1188 list_del(&neigh->list);
1189 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1190 } else {
1191 np = &neigh->hnext;
1192 }
1193
1194 }
1195 }
1196
1197out_unlock:
1198 spin_unlock_irqrestore(&priv->lock, flags);
1199 ipoib_mcast_remove_list(&remove_list);
1200}
1201
1202static void ipoib_reap_neigh(struct work_struct *work)
1203{
1204 struct ipoib_dev_priv *priv =
1205 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1206
1207 __ipoib_reap_neigh(priv);
1208
1209 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1210 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1211 arp_tbl.gc_interval);
1212}
1213
1214
1215static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1216 struct net_device *dev)
1217{
1218 struct ipoib_neigh *neigh;
1219
1220 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1221 if (!neigh)
1222 return NULL;
1223
1224 neigh->dev = dev;
1225 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1226 skb_queue_head_init(&neigh->queue);
1227 INIT_LIST_HEAD(&neigh->list);
1228 ipoib_cm_set(neigh, NULL);
1229 /* one ref on behalf of the caller */
1230 atomic_set(&neigh->refcnt, 1);
1231
1232 return neigh;
1233}
1234
1235struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1236 struct net_device *dev)
1237{
1238 struct ipoib_dev_priv *priv = netdev_priv(dev);
1239 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1240 struct ipoib_neigh_hash *htbl;
1241 struct ipoib_neigh *neigh;
1242 u32 hash_val;
1243
1244 htbl = rcu_dereference_protected(ntbl->htbl,
1245 lockdep_is_held(&priv->lock));
1246 if (!htbl) {
1247 neigh = NULL;
1248 goto out_unlock;
1249 }
1250
1251 /* need to add a new neigh, but maybe some other thread succeeded?
1252 * recalc hash, maybe hash resize took place so we do a search
1253 */
1254 hash_val = ipoib_addr_hash(htbl, daddr);
1255 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1256 lockdep_is_held(&priv->lock));
1257 neigh != NULL;
1258 neigh = rcu_dereference_protected(neigh->hnext,
1259 lockdep_is_held(&priv->lock))) {
1260 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1261 /* found, take one ref on behalf of the caller */
1262 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1263 /* deleted */
1264 neigh = NULL;
1265 break;
1266 }
1267 neigh->alive = jiffies;
1268 goto out_unlock;
1269 }
1270 }
1271
1272 neigh = ipoib_neigh_ctor(daddr, dev);
1273 if (!neigh)
1274 goto out_unlock;
1275
1276 /* one ref on behalf of the hash table */
1277 atomic_inc(&neigh->refcnt);
1278 neigh->alive = jiffies;
1279 /* put in hash */
1280 rcu_assign_pointer(neigh->hnext,
1281 rcu_dereference_protected(htbl->buckets[hash_val],
1282 lockdep_is_held(&priv->lock)));
1283 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1284 atomic_inc(&ntbl->entries);
1285
1286out_unlock:
1287
1288 return neigh;
1289}
1290
1291void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1292{
1293 /* neigh reference count was dropprd to zero */
1294 struct net_device *dev = neigh->dev;
1295 struct ipoib_dev_priv *priv = netdev_priv(dev);
1296 struct sk_buff *skb;
1297 if (neigh->ah)
1298 ipoib_put_ah(neigh->ah);
1299 while ((skb = __skb_dequeue(&neigh->queue))) {
1300 ++dev->stats.tx_dropped;
1301 dev_kfree_skb_any(skb);
1302 }
1303 if (ipoib_cm_get(neigh))
1304 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1305 ipoib_dbg(netdev_priv(dev),
1306 "neigh free for %06x %pI6\n",
1307 IPOIB_QPN(neigh->daddr),
1308 neigh->daddr + 4);
1309 kfree(neigh);
1310 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1311 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1312 complete(&priv->ntbl.flushed);
1313 }
1314}
1315
1316static void ipoib_neigh_reclaim(struct rcu_head *rp)
1317{
1318 /* Called as a result of removal from hash table */
1319 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1320 /* note TX context may hold another ref */
1321 ipoib_neigh_put(neigh);
1322}
1323
1324void ipoib_neigh_free(struct ipoib_neigh *neigh)
1325{
1326 struct net_device *dev = neigh->dev;
1327 struct ipoib_dev_priv *priv = netdev_priv(dev);
1328 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1329 struct ipoib_neigh_hash *htbl;
1330 struct ipoib_neigh __rcu **np;
1331 struct ipoib_neigh *n;
1332 u32 hash_val;
1333
1334 htbl = rcu_dereference_protected(ntbl->htbl,
1335 lockdep_is_held(&priv->lock));
1336 if (!htbl)
1337 return;
1338
1339 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1340 np = &htbl->buckets[hash_val];
1341 for (n = rcu_dereference_protected(*np,
1342 lockdep_is_held(&priv->lock));
1343 n != NULL;
1344 n = rcu_dereference_protected(*np,
1345 lockdep_is_held(&priv->lock))) {
1346 if (n == neigh) {
1347 /* found */
1348 rcu_assign_pointer(*np,
1349 rcu_dereference_protected(neigh->hnext,
1350 lockdep_is_held(&priv->lock)));
1351 /* remove from parent list */
1352 list_del(&neigh->list);
1353 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1354 return;
1355 } else {
1356 np = &n->hnext;
1357 }
1358 }
1359}
1360
1361static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1362{
1363 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1364 struct ipoib_neigh_hash *htbl;
1365 struct ipoib_neigh __rcu **buckets;
1366 u32 size;
1367
1368 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1369 ntbl->htbl = NULL;
1370 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1371 if (!htbl)
1372 return -ENOMEM;
1373 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1374 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1375 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1376 if (!buckets) {
1377 kfree(htbl);
1378 return -ENOMEM;
1379 }
1380 htbl->size = size;
1381 htbl->mask = (size - 1);
1382 htbl->buckets = buckets;
1383 RCU_INIT_POINTER(ntbl->htbl, htbl);
1384 htbl->ntbl = ntbl;
1385 atomic_set(&ntbl->entries, 0);
1386
1387 /* start garbage collection */
1388 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1389 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1390 arp_tbl.gc_interval);
1391
1392 return 0;
1393}
1394
1395static void neigh_hash_free_rcu(struct rcu_head *head)
1396{
1397 struct ipoib_neigh_hash *htbl = container_of(head,
1398 struct ipoib_neigh_hash,
1399 rcu);
1400 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1401 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1402
1403 kfree(buckets);
1404 kfree(htbl);
1405 complete(&ntbl->deleted);
1406}
1407
1408void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1409{
1410 struct ipoib_dev_priv *priv = netdev_priv(dev);
1411 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1412 struct ipoib_neigh_hash *htbl;
1413 unsigned long flags;
1414 int i;
1415
1416 /* remove all neigh connected to a given path or mcast */
1417 spin_lock_irqsave(&priv->lock, flags);
1418
1419 htbl = rcu_dereference_protected(ntbl->htbl,
1420 lockdep_is_held(&priv->lock));
1421
1422 if (!htbl)
1423 goto out_unlock;
1424
1425 for (i = 0; i < htbl->size; i++) {
1426 struct ipoib_neigh *neigh;
1427 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1428
1429 while ((neigh = rcu_dereference_protected(*np,
1430 lockdep_is_held(&priv->lock))) != NULL) {
1431 /* delete neighs belong to this parent */
1432 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1433 rcu_assign_pointer(*np,
1434 rcu_dereference_protected(neigh->hnext,
1435 lockdep_is_held(&priv->lock)));
1436 /* remove from parent list */
1437 list_del(&neigh->list);
1438 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1439 } else {
1440 np = &neigh->hnext;
1441 }
1442
1443 }
1444 }
1445out_unlock:
1446 spin_unlock_irqrestore(&priv->lock, flags);
1447}
1448
1449static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1450{
1451 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1452 struct ipoib_neigh_hash *htbl;
1453 unsigned long flags;
1454 int i, wait_flushed = 0;
1455
1456 init_completion(&priv->ntbl.flushed);
1457
1458 spin_lock_irqsave(&priv->lock, flags);
1459
1460 htbl = rcu_dereference_protected(ntbl->htbl,
1461 lockdep_is_held(&priv->lock));
1462 if (!htbl)
1463 goto out_unlock;
1464
1465 wait_flushed = atomic_read(&priv->ntbl.entries);
1466 if (!wait_flushed)
1467 goto free_htbl;
1468
1469 for (i = 0; i < htbl->size; i++) {
1470 struct ipoib_neigh *neigh;
1471 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1472
1473 while ((neigh = rcu_dereference_protected(*np,
1474 lockdep_is_held(&priv->lock))) != NULL) {
1475 rcu_assign_pointer(*np,
1476 rcu_dereference_protected(neigh->hnext,
1477 lockdep_is_held(&priv->lock)));
1478 /* remove from path/mc list */
1479 list_del(&neigh->list);
1480 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1481 }
1482 }
1483
1484free_htbl:
1485 rcu_assign_pointer(ntbl->htbl, NULL);
1486 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1487
1488out_unlock:
1489 spin_unlock_irqrestore(&priv->lock, flags);
1490 if (wait_flushed)
1491 wait_for_completion(&priv->ntbl.flushed);
1492}
1493
1494static void ipoib_neigh_hash_uninit(struct net_device *dev)
1495{
1496 struct ipoib_dev_priv *priv = netdev_priv(dev);
1497 int stopped;
1498
1499 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1500 init_completion(&priv->ntbl.deleted);
1501 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1502
1503 /* Stop GC if called at init fail need to cancel work */
1504 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1505 if (!stopped)
1506 cancel_delayed_work(&priv->neigh_reap_task);
1507
1508 ipoib_flush_neighs(priv);
1509
1510 wait_for_completion(&priv->ntbl.deleted);
1511}
1512
1513
1514int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1515{
1516 struct ipoib_dev_priv *priv = netdev_priv(dev);
1517
1518 /* Allocate RX/TX "rings" to hold queued skbs */
1519 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1520 GFP_KERNEL);
1521 if (!priv->rx_ring) {
1522 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1523 ca->name, ipoib_recvq_size);
1524 goto out;
1525 }
1526
1527 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1528 if (!priv->tx_ring) {
1529 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1530 ca->name, ipoib_sendq_size);
1531 goto out_rx_ring_cleanup;
1532 }
1533
1534 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1535
1536 if (ipoib_ib_dev_init(dev, ca, port))
1537 goto out_tx_ring_cleanup;
1538
1539 /*
1540 * Must be after ipoib_ib_dev_init so we can allocate a per
1541 * device wq there and use it here
1542 */
1543 if (ipoib_neigh_hash_init(priv) < 0)
1544 goto out_dev_uninit;
1545
1546 return 0;
1547
1548out_dev_uninit:
1549 ipoib_ib_dev_cleanup(dev);
1550
1551out_tx_ring_cleanup:
1552 vfree(priv->tx_ring);
1553
1554out_rx_ring_cleanup:
1555 kfree(priv->rx_ring);
1556
1557out:
1558 return -ENOMEM;
1559}
1560
1561void ipoib_dev_cleanup(struct net_device *dev)
1562{
1563 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1564 LIST_HEAD(head);
1565
1566 ASSERT_RTNL();
1567
1568 ipoib_delete_debug_files(dev);
1569
1570 /* Delete any child interfaces first */
1571 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1572 /* Stop GC on child */
1573 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1574 cancel_delayed_work(&cpriv->neigh_reap_task);
1575 unregister_netdevice_queue(cpriv->dev, &head);
1576 }
1577 unregister_netdevice_many(&head);
1578
1579 /*
1580 * Must be before ipoib_ib_dev_cleanup or we delete an in use
1581 * work queue
1582 */
1583 ipoib_neigh_hash_uninit(dev);
1584
1585 ipoib_ib_dev_cleanup(dev);
1586
1587 kfree(priv->rx_ring);
1588 vfree(priv->tx_ring);
1589
1590 priv->rx_ring = NULL;
1591 priv->tx_ring = NULL;
1592}
1593
1594static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1595{
1596 struct ipoib_dev_priv *priv = netdev_priv(dev);
1597
1598 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1599}
1600
1601static int ipoib_get_vf_config(struct net_device *dev, int vf,
1602 struct ifla_vf_info *ivf)
1603{
1604 struct ipoib_dev_priv *priv = netdev_priv(dev);
1605 int err;
1606
1607 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1608 if (err)
1609 return err;
1610
1611 ivf->vf = vf;
1612
1613 return 0;
1614}
1615
1616static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1617{
1618 struct ipoib_dev_priv *priv = netdev_priv(dev);
1619
1620 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1621 return -EINVAL;
1622
1623 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1624}
1625
1626static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1627 struct ifla_vf_stats *vf_stats)
1628{
1629 struct ipoib_dev_priv *priv = netdev_priv(dev);
1630
1631 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1632}
1633
1634static const struct header_ops ipoib_header_ops = {
1635 .create = ipoib_hard_header,
1636};
1637
1638static const struct net_device_ops ipoib_netdev_ops_pf = {
1639 .ndo_uninit = ipoib_uninit,
1640 .ndo_open = ipoib_open,
1641 .ndo_stop = ipoib_stop,
1642 .ndo_change_mtu = ipoib_change_mtu,
1643 .ndo_fix_features = ipoib_fix_features,
1644 .ndo_start_xmit = ipoib_start_xmit,
1645 .ndo_tx_timeout = ipoib_timeout,
1646 .ndo_set_rx_mode = ipoib_set_mcast_list,
1647 .ndo_get_iflink = ipoib_get_iflink,
1648 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
1649 .ndo_get_vf_config = ipoib_get_vf_config,
1650 .ndo_get_vf_stats = ipoib_get_vf_stats,
1651 .ndo_set_vf_guid = ipoib_set_vf_guid,
1652};
1653
1654static const struct net_device_ops ipoib_netdev_ops_vf = {
1655 .ndo_uninit = ipoib_uninit,
1656 .ndo_open = ipoib_open,
1657 .ndo_stop = ipoib_stop,
1658 .ndo_change_mtu = ipoib_change_mtu,
1659 .ndo_fix_features = ipoib_fix_features,
1660 .ndo_start_xmit = ipoib_start_xmit,
1661 .ndo_tx_timeout = ipoib_timeout,
1662 .ndo_set_rx_mode = ipoib_set_mcast_list,
1663 .ndo_get_iflink = ipoib_get_iflink,
1664};
1665
1666void ipoib_setup(struct net_device *dev)
1667{
1668 struct ipoib_dev_priv *priv = netdev_priv(dev);
1669
1670 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1671 dev->netdev_ops = &ipoib_netdev_ops_vf;
1672 else
1673 dev->netdev_ops = &ipoib_netdev_ops_pf;
1674
1675 dev->header_ops = &ipoib_header_ops;
1676
1677 ipoib_set_ethtool_ops(dev);
1678
1679 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1680
1681 dev->watchdog_timeo = HZ;
1682
1683 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1684
1685 dev->hard_header_len = IPOIB_ENCAP_LEN;
1686 dev->addr_len = INFINIBAND_ALEN;
1687 dev->type = ARPHRD_INFINIBAND;
1688 dev->tx_queue_len = ipoib_sendq_size * 2;
1689 dev->features = (NETIF_F_VLAN_CHALLENGED |
1690 NETIF_F_HIGHDMA);
1691 netif_keep_dst(dev);
1692
1693 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1694
1695 priv->dev = dev;
1696
1697 spin_lock_init(&priv->lock);
1698
1699 init_rwsem(&priv->vlan_rwsem);
1700
1701 INIT_LIST_HEAD(&priv->path_list);
1702 INIT_LIST_HEAD(&priv->child_intfs);
1703 INIT_LIST_HEAD(&priv->dead_ahs);
1704 INIT_LIST_HEAD(&priv->multicast_list);
1705
1706 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1707 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1708 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1709 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1710 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1711 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1712 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1713 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1714}
1715
1716struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1717{
1718 struct net_device *dev;
1719
1720 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1721 NET_NAME_UNKNOWN, ipoib_setup);
1722 if (!dev)
1723 return NULL;
1724
1725 return netdev_priv(dev);
1726}
1727
1728static ssize_t show_pkey(struct device *dev,
1729 struct device_attribute *attr, char *buf)
1730{
1731 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1732
1733 return sprintf(buf, "0x%04x\n", priv->pkey);
1734}
1735static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1736
1737static ssize_t show_umcast(struct device *dev,
1738 struct device_attribute *attr, char *buf)
1739{
1740 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1741
1742 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1743}
1744
1745void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1746{
1747 struct ipoib_dev_priv *priv = netdev_priv(ndev);
1748
1749 if (umcast_val > 0) {
1750 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1751 ipoib_warn(priv, "ignoring multicast groups joined directly "
1752 "by userspace\n");
1753 } else
1754 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1755}
1756
1757static ssize_t set_umcast(struct device *dev,
1758 struct device_attribute *attr,
1759 const char *buf, size_t count)
1760{
1761 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1762
1763 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1764
1765 return count;
1766}
1767static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1768
1769int ipoib_add_umcast_attr(struct net_device *dev)
1770{
1771 return device_create_file(&dev->dev, &dev_attr_umcast);
1772}
1773
1774static ssize_t create_child(struct device *dev,
1775 struct device_attribute *attr,
1776 const char *buf, size_t count)
1777{
1778 int pkey;
1779 int ret;
1780
1781 if (sscanf(buf, "%i", &pkey) != 1)
1782 return -EINVAL;
1783
1784 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1785 return -EINVAL;
1786
1787 /*
1788 * Set the full membership bit, so that we join the right
1789 * broadcast group, etc.
1790 */
1791 pkey |= 0x8000;
1792
1793 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1794
1795 return ret ? ret : count;
1796}
1797static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1798
1799static ssize_t delete_child(struct device *dev,
1800 struct device_attribute *attr,
1801 const char *buf, size_t count)
1802{
1803 int pkey;
1804 int ret;
1805
1806 if (sscanf(buf, "%i", &pkey) != 1)
1807 return -EINVAL;
1808
1809 if (pkey < 0 || pkey > 0xffff)
1810 return -EINVAL;
1811
1812 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1813
1814 return ret ? ret : count;
1815
1816}
1817static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1818
1819int ipoib_add_pkey_attr(struct net_device *dev)
1820{
1821 return device_create_file(&dev->dev, &dev_attr_pkey);
1822}
1823
1824int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1825{
1826 priv->hca_caps = hca->attrs.device_cap_flags;
1827
1828 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1829 priv->dev->hw_features = NETIF_F_SG |
1830 NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1831
1832 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1833 priv->dev->hw_features |= NETIF_F_TSO;
1834
1835 priv->dev->features |= priv->dev->hw_features;
1836 }
1837
1838 return 0;
1839}
1840
1841static struct net_device *ipoib_add_port(const char *format,
1842 struct ib_device *hca, u8 port)
1843{
1844 struct ipoib_dev_priv *priv;
1845 struct ib_port_attr attr;
1846 int result = -ENOMEM;
1847
1848 priv = ipoib_intf_alloc(format);
1849 if (!priv)
1850 goto alloc_mem_failed;
1851
1852 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1853 priv->dev->dev_id = port - 1;
1854
1855 result = ib_query_port(hca, port, &attr);
1856 if (!result)
1857 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1858 else {
1859 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1860 hca->name, port);
1861 goto device_init_failed;
1862 }
1863
1864 /* MTU will be reset when mcast join happens */
1865 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1866 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1867
1868 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1869
1870 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1871 if (result) {
1872 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1873 hca->name, port, result);
1874 goto device_init_failed;
1875 }
1876
1877 result = ipoib_set_dev_features(priv, hca);
1878 if (result)
1879 goto device_init_failed;
1880
1881 /*
1882 * Set the full membership bit, so that we join the right
1883 * broadcast group, etc.
1884 */
1885 priv->pkey |= 0x8000;
1886
1887 priv->dev->broadcast[8] = priv->pkey >> 8;
1888 priv->dev->broadcast[9] = priv->pkey & 0xff;
1889
1890 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
1891 if (result) {
1892 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1893 hca->name, port, result);
1894 goto device_init_failed;
1895 } else
1896 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1897
1898 result = ipoib_dev_init(priv->dev, hca, port);
1899 if (result < 0) {
1900 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1901 hca->name, port, result);
1902 goto device_init_failed;
1903 }
1904
1905 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1906 priv->ca, ipoib_event);
1907 result = ib_register_event_handler(&priv->event_handler);
1908 if (result < 0) {
1909 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1910 "port %d (ret = %d)\n",
1911 hca->name, port, result);
1912 goto event_failed;
1913 }
1914
1915 result = register_netdev(priv->dev);
1916 if (result) {
1917 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1918 hca->name, port, result);
1919 goto register_failed;
1920 }
1921
1922 ipoib_create_debug_files(priv->dev);
1923
1924 if (ipoib_cm_add_mode_attr(priv->dev))
1925 goto sysfs_failed;
1926 if (ipoib_add_pkey_attr(priv->dev))
1927 goto sysfs_failed;
1928 if (ipoib_add_umcast_attr(priv->dev))
1929 goto sysfs_failed;
1930 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1931 goto sysfs_failed;
1932 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1933 goto sysfs_failed;
1934
1935 return priv->dev;
1936
1937sysfs_failed:
1938 ipoib_delete_debug_files(priv->dev);
1939 unregister_netdev(priv->dev);
1940
1941register_failed:
1942 ib_unregister_event_handler(&priv->event_handler);
1943 flush_workqueue(ipoib_workqueue);
1944 /* Stop GC if started before flush */
1945 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1946 cancel_delayed_work(&priv->neigh_reap_task);
1947 flush_workqueue(priv->wq);
1948
1949event_failed:
1950 ipoib_dev_cleanup(priv->dev);
1951
1952device_init_failed:
1953 free_netdev(priv->dev);
1954
1955alloc_mem_failed:
1956 return ERR_PTR(result);
1957}
1958
1959static void ipoib_add_one(struct ib_device *device)
1960{
1961 struct list_head *dev_list;
1962 struct net_device *dev;
1963 struct ipoib_dev_priv *priv;
1964 int p;
1965 int count = 0;
1966
1967 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1968 if (!dev_list)
1969 return;
1970
1971 INIT_LIST_HEAD(dev_list);
1972
1973 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
1974 if (!rdma_protocol_ib(device, p))
1975 continue;
1976 dev = ipoib_add_port("ib%d", device, p);
1977 if (!IS_ERR(dev)) {
1978 priv = netdev_priv(dev);
1979 list_add_tail(&priv->list, dev_list);
1980 count++;
1981 }
1982 }
1983
1984 if (!count) {
1985 kfree(dev_list);
1986 return;
1987 }
1988
1989 ib_set_client_data(device, &ipoib_client, dev_list);
1990}
1991
1992static void ipoib_remove_one(struct ib_device *device, void *client_data)
1993{
1994 struct ipoib_dev_priv *priv, *tmp;
1995 struct list_head *dev_list = client_data;
1996
1997 if (!dev_list)
1998 return;
1999
2000 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2001 ib_unregister_event_handler(&priv->event_handler);
2002 flush_workqueue(ipoib_workqueue);
2003
2004 rtnl_lock();
2005 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2006 rtnl_unlock();
2007
2008 /* Stop GC */
2009 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2010 cancel_delayed_work(&priv->neigh_reap_task);
2011 flush_workqueue(priv->wq);
2012
2013 unregister_netdev(priv->dev);
2014 free_netdev(priv->dev);
2015 }
2016
2017 kfree(dev_list);
2018}
2019
2020static int __init ipoib_init_module(void)
2021{
2022 int ret;
2023
2024 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2025 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2026 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2027
2028 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2029 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2030 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2031#ifdef CONFIG_INFINIBAND_IPOIB_CM
2032 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2033#endif
2034
2035 /*
2036 * When copying small received packets, we only copy from the
2037 * linear data part of the SKB, so we rely on this condition.
2038 */
2039 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2040
2041 ret = ipoib_register_debugfs();
2042 if (ret)
2043 return ret;
2044
2045 /*
2046 * We create a global workqueue here that is used for all flush
2047 * operations. However, if you attempt to flush a workqueue
2048 * from a task on that same workqueue, it deadlocks the system.
2049 * We want to be able to flush the tasks associated with a
2050 * specific net device, so we also create a workqueue for each
2051 * netdevice. We queue up the tasks for that device only on
2052 * its private workqueue, and we only queue up flush events
2053 * on our global flush workqueue. This avoids the deadlocks.
2054 */
2055 ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
2056 if (!ipoib_workqueue) {
2057 ret = -ENOMEM;
2058 goto err_fs;
2059 }
2060
2061 ib_sa_register_client(&ipoib_sa_client);
2062
2063 ret = ib_register_client(&ipoib_client);
2064 if (ret)
2065 goto err_sa;
2066
2067 ret = ipoib_netlink_init();
2068 if (ret)
2069 goto err_client;
2070
2071 return 0;
2072
2073err_client:
2074 ib_unregister_client(&ipoib_client);
2075
2076err_sa:
2077 ib_sa_unregister_client(&ipoib_sa_client);
2078 destroy_workqueue(ipoib_workqueue);
2079
2080err_fs:
2081 ipoib_unregister_debugfs();
2082
2083 return ret;
2084}
2085
2086static void __exit ipoib_cleanup_module(void)
2087{
2088 ipoib_netlink_fini();
2089 ib_unregister_client(&ipoib_client);
2090 ib_sa_unregister_client(&ipoib_sa_client);
2091 ipoib_unregister_debugfs();
2092 destroy_workqueue(ipoib_workqueue);
2093}
2094
2095module_init(ipoib_init_module);
2096module_exit(ipoib_cleanup_module);