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