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