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	MODULE_AUTHOR("Roland Dreier");
56	MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
57	MODULE_LICENSE("Dual BSD/GPL");
58
59	int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
60	int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
61
62	module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
63	MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
64	module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
65	MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
66
67	#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
68	int ipoib_debug_level;
69
70	module_param_named(debug_level, ipoib_debug_level, int, 0644);
71	MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
72	#endif
73
74	struct ipoib_path_iter {
75	struct net_device *dev;
76	struct ipoib_path path;
77	};
78
79	static 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
85	struct workqueue_struct *ipoib_workqueue;
86
87	struct ib_sa_client ipoib_sa_client;
88
89	static int ipoib_add_one(struct ib_device *device);
90	static void ipoib_remove_one(struct ib_device *device, void *client_data);
91	static void ipoib_neigh_reclaim(struct rcu_head *rp);
92	static 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);
96	static int ipoib_set_mac(struct net_device *dev, void *addr);
97	static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
98	int cmd);
99
100	static 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
108	static 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
134	int 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(nr: IPOIB_FLAG_ADMIN_UP, addr: &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(sem: &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(dev: cpriv->dev, flags: flags | IFF_UP, NULL);
165	}
166	up_read(sem: &priv->vlan_rwsem);
167	} else if (priv->parent) {
168	struct ipoib_dev_priv *ppriv = ipoib_priv(dev: 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
178	err_disable:
179	clear_bit(nr: IPOIB_FLAG_ADMIN_UP, addr: &priv->flags);
180
181	return -EINVAL;
182	}
183
184	static 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(nr: IPOIB_FLAG_ADMIN_UP, addr: &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(sem: &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(dev: cpriv->dev, flags: flags & ~IFF_UP, NULL);
210	}
211	up_read(sem: &priv->vlan_rwsem);
212	}
213
214	return 0;
215	}
216
217	static 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
227	static 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
274	static 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(stats64: stats, netdev_stats: &dev->stats);
283	}
284
285	/* Called with an RCU read lock taken */
286	static 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, dst: 0,
302	local: addr_in->sin_addr.s_addr,
303	scope: RT_SCOPE_HOST);
304	in_dev_put(idev: 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: &addr_in6->sin6_addr, dev, strict: 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	*/
326	static 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(dev: master);
334	rcu_read_unlock();
335
336	if (master)
337	return master;
338
339	dev_hold(dev);
340	return dev;
341	}
342
343	struct ipoib_walk_data {
344	const struct sockaddr *addr;
345	struct net_device *result;
346	};
347
348	static 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(addr: data->addr, dev: upper)) {
355	dev_hold(dev: 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	*/
373	static 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, fn: ipoib_upper_walk, priv: &priv);
390	out:
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. */
400	static 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(p: gid, q: &priv->local_gid, size: sizeof(*gid)))) {
413	if (!addr) {
414	net_dev = ipoib_get_master_net_dev(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, dev: priv->dev);
419	}
420	if (net_dev) {
421	if (!*found_net_dev)
422	*found_net_dev = net_dev;
423	else
424	dev_put(dev: net_dev);
425	++matches;
426	}
427	}
428
429	/* Check child interfaces */
430	down_read_nested(sem: &priv->vlan_rwsem, subclass: nesting);
431	list_for_each_entry(child_priv, &priv->child_intfs, list) {
432	matches += ipoib_match_gid_pkey_addr(priv: child_priv, gid,
433	pkey_index, addr,
434	nesting: nesting + 1,
435	found_net_dev);
436	if (matches > 1)
437	break;
438	}
439	up_read(sem: &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 */
447	static 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, nesting: 0, found_net_dev: net_dev);
464	if (matches > 1)
465	break;
466	}
467
468	return matches;
469	}
470
471	static 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(device: dev, port_num: port))
483	return NULL;
484
485	ret = ib_find_cached_pkey(device: dev, port_num: port, pkey, index: &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: &net_dev);
492
493	switch (matches) {
494	case 0:
495	return NULL;
496	case 1:
497	return net_dev;
498	}
499
500	dev_put(dev: 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: &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
518	int 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(nr: IPOIB_FLAG_ADMIN_CM, addr: &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, txq: 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(nr: IPOIB_FLAG_ADMIN_CM, addr: &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, txq: 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
557	struct 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(p: gid, q: path->pathrec.dgid.raw,
568	size: 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
581	static 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(p: path->pathrec.dgid.raw, q: tpath->pathrec.dgid.raw,
594	size: 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(node: &path->rb_node, parent: pn, rb_link: n);
604	rb_insert_color(&path->rb_node, &priv->path_tree);
605
606	list_add_tail(new: &path->list, head: &priv->path_list);
607
608	return 0;
609	}
610
611	static void path_free(struct net_device *dev, struct ipoib_path *path)
612	{
613	struct sk_buff *skb;
614
615	while ((skb = __skb_dequeue(list: &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, gid: path->pathrec.dgid.raw);
622
623	if (path->ah)
624	ipoib_put_ah(ah: path->ah);
625
626	kfree(objp: path);
627	}
628
629	#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
630
631	struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
632	{
633	struct ipoib_path_iter *iter;
634
635	iter = kmalloc(size: 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(objp: iter);
644	return NULL;
645	}
646
647	return iter;
648	}
649
650	int ipoib_path_iter_next(struct ipoib_path_iter *iter)
651	{
652	struct ipoib_dev_priv *priv = ipoib_priv(dev: iter->dev);
653	struct rb_node *n;
654	struct ipoib_path *path;
655	int ret = 1;
656
657	spin_lock_irq(lock: &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(p: iter->path.pathrec.dgid.raw, q: path->pathrec.dgid.raw,
665	size: 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(lock: &priv->lock);
675
676	return ret;
677	}
678
679	void 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
687	void 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(lock: &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(lock: &priv->lock);
703	}
704
705	static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
706	{
707	struct ipoib_pseudo_header *phdr;
708
709	phdr = skb_push(skb, len: sizeof(*phdr));
710	memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
711	}
712
713	void 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(list: &priv->path_list, head: &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(id: path->query_id, query: path->query);
731	spin_unlock_irqrestore(lock: &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(lock: &priv->lock, flags);
740	netif_tx_unlock_bh(dev);
741	}
742
743	static 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(list: &skqueue);
766
767	if (!status) {
768	struct rdma_ah_attr av;
769
770	if (!ib_init_ah_attr_from_path(device: priv->ca, port_num: priv->port,
771	rec: pathrec, ah_attr: &av, NULL)) {
772	ah = ipoib_create_ah(dev, pd: priv->pd, attr: &av);
773	rdma_destroy_ah_attr(ah_attr: &av);
774	}
775	}
776
777	spin_lock_irqsave(&priv->lock, flags);
778
779	if (!IS_ERR_OR_NULL(ptr: 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(p: pathrec->dgid.raw, q: path->pathrec.dgid.raw,
786	size: 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(list: &path->queue)))
806	__skb_queue_tail(list: &skqueue, newsk: 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(ah: neigh->ah);
819	}
820	kref_get(kref: &path->ah->ref);
821	neigh->ah = path->ah;
822
823	if (ipoib_cm_enabled(dev, hwaddr: neigh->daddr)) {
824	if (!ipoib_cm_get(neigh))
825	ipoib_cm_set(neigh, tx: 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(list: &neigh->queue)))
835	__skb_queue_tail(list: &skqueue, newsk: skb);
836	}
837	path->ah->valid = 1;
838	}
839
840	path->query = NULL;
841	complete(&path->done);
842
843	spin_unlock_irqrestore(lock: &priv->lock, flags);
844
845	if (IS_ERR_OR_NULL(ptr: ah))
846	ipoib_del_neighs_by_gid(dev, gid: path->pathrec.dgid.raw);
847
848	if (old_ah)
849	ipoib_put_ah(ah: old_ah);
850
851	while ((skb = __skb_dequeue(list: &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
861	static 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(device: priv->ca, port_num: 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
878	static 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(size: sizeof(*path), GFP_ATOMIC);
887	if (!path)
888	return NULL;
889
890	skb_queue_head_init(list: &path->queue);
891
892	INIT_LIST_HEAD(list: &path->neigh_list);
893
894	init_path_rec(priv, path, gid);
895
896	return path;
897	}
898
899	static 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(x: &path->done);
908
909	path->query_id =
910	ib_sa_path_rec_get(client: &ipoib_sa_client, device: priv->ca, port_num: priv->port,
911	rec: &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	timeout_ms: 1000, GFP_ATOMIC,
918	callback: path_rec_completion,
919	context: path, query: &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
930	static 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, gid: daddr + 4);
940	if (!path)
941	goto out;
942	if (!path->query)
943	path_rec_start(dev, path);
944	out:
945	spin_unlock_irqrestore(lock: &priv->lock, flags);
946	}
947
948	static 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(lock: &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(lock: &priv->lock, flags);
971	return neigh;
972	}
973
974	path = __path_find(dev, gid: daddr + 4);
975	if (!path) {
976	path = path_rec_create(dev, gid: daddr + 4);
977	if (!path)
978	goto err_path;
979
980	__path_add(dev, path);
981	}
982
983	list_add_tail(new: &neigh->list, head: &path->neigh_list);
984
985	if (path->ah && path->ah->valid) {
986	kref_get(kref: &path->ah->ref);
987	neigh->ah = path->ah;
988
989	if (ipoib_cm_enabled(dev, hwaddr: neigh->daddr)) {
990	if (!ipoib_cm_get(neigh))
991	ipoib_cm_set(neigh, tx: 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(list_: &neigh->queue) <
997	IPOIB_MAX_PATH_REC_QUEUE) {
998	push_pseudo_header(skb, daddr: neigh->daddr);
999	__skb_queue_tail(list: &neigh->queue, newsk: 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(lock: &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(list_: &neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1018	push_pseudo_header(skb, daddr: neigh->daddr);
1019	__skb_queue_tail(list: &neigh->queue, newsk: skb);
1020	} else {
1021	goto err_drop;
1022	}
1023	}
1024
1025	spin_unlock_irqrestore(lock: &priv->lock, flags);
1026	ipoib_neigh_put(neigh);
1027	return NULL;
1028
1029	err_path:
1030	ipoib_neigh_free(neigh);
1031	err_drop:
1032	++dev->stats.tx_dropped;
1033	dev_kfree_skb_any(skb);
1034
1035	spin_unlock_irqrestore(lock: &priv->lock, flags);
1036	ipoib_neigh_put(neigh);
1037
1038	return NULL;
1039	}
1040
1041	static 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, gid: phdr->hwaddr + 4);
1056	if (!path || !path->ah || !path->ah->valid) {
1057	if (!path) {
1058	path = path_rec_create(dev, gid: 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, gid: 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(list_: &path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1074	push_pseudo_header(skb, daddr: phdr->hwaddr);
1075	__skb_queue_tail(list: &path->queue, newsk: skb);
1076	goto unlock;
1077	} else {
1078	goto drop_and_unlock;
1079	}
1080	}
1081
1082	spin_unlock_irqrestore(lock: &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
1089	drop_and_unlock:
1090	++dev->stats.tx_dropped;
1091	dev_kfree_skb_any(skb);
1092	unlock:
1093	spin_unlock_irqrestore(lock: &priv->lock, flags);
1094	}
1095
1096	static 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, len: 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, daddr: phdr->hwaddr);
1126	if (likely(neigh))
1127	goto send_using_neigh;
1128	ipoib_mcast_send(dev, daddr: 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, daddr: phdr->hwaddr);
1138	if (unlikely(!neigh)) {
1139	neigh = neigh_add_path(skb, daddr: 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
1156	send_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, tx: 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, daddr: phdr->hwaddr, dev);
1169	}
1170
1171	if (skb_queue_len(list_: &neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1172	push_pseudo_header(skb, daddr: phdr->hwaddr);
1173	spin_lock_irqsave(&priv->lock, flags);
1174	__skb_queue_tail(list: &neigh->queue, newsk: skb);
1175	spin_unlock_irqrestore(lock: &priv->lock, flags);
1176	} else {
1177	++dev->stats.tx_dropped;
1178	dev_kfree_skb_any(skb);
1179	}
1180
1181	unref:
1182	ipoib_neigh_put(neigh);
1183
1184	return NETDEV_TX_OK;
1185	}
1186
1187	static 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(work: &priv->tx_timeout_work);
1205	}
1206
1207	void 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(dev: priv->dev);
1220	err = ipoib_open(dev: 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(dev: priv->dev);
1228	unlock:
1229	rtnl_unlock();
1230
1231	}
1232
1233	static 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, len: 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
1257	static 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(wq: priv->wq, work: &priv->restart_task);
1267	}
1268
1269	static 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 READ_ONCE(dev->ifindex);
1276
1277	/* child/vlan interface */
1278	return READ_ONCE(priv->parent->ifindex);
1279	}
1280
1281	static 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(a: d32[3], b: d32[4], IPOIB_QPN_MASK & d32[0], initval: 0);
1294	return hv & htbl->mask;
1295	}
1296
1297	struct 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(p: daddr, q: neigh->daddr, INFINIBAND_ALEN) == 0) {
1317	/* found, take one ref on behalf of the caller */
1318	if (!refcount_inc_not_zero(r: &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
1330	out_unlock:
1331	rcu_read_unlock_bh();
1332	return neigh;
1333	}
1334
1335	static 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, mgid: neigh->daddr + 4, remove_list: &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(entry: &neigh->list);
1373	call_rcu(head: &neigh->rcu, func: ipoib_neigh_reclaim);
1374	} else {
1375	np = &neigh->hnext;
1376	}
1377
1378	}
1379	}
1380
1381	out_unlock:
1382	spin_unlock_irqrestore(lock: &priv->lock, flags);
1383	ipoib_mcast_remove_list(remove_list: &remove_list);
1384	}
1385
1386	static 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(wq: priv->wq, dwork: &priv->neigh_reap_task,
1394	delay: arp_tbl.gc_interval);
1395	}
1396
1397
1398	static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1399	struct net_device *dev)
1400	{
1401	struct ipoib_neigh *neigh;
1402
1403	neigh = kzalloc(size: 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(list: &neigh->queue);
1410	INIT_LIST_HEAD(list: &neigh->list);
1411	ipoib_cm_set(neigh, NULL);
1412	/* one ref on behalf of the caller */
1413	refcount_set(r: &neigh->refcnt, n: 1);
1414
1415	return neigh;
1416	}
1417
1418	struct 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(p: daddr, q: neigh->daddr, INFINIBAND_ALEN) == 0) {
1444	/* found, take one ref on behalf of the caller */
1445	if (!refcount_inc_not_zero(r: &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(r: &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(v: &ntbl->entries);
1468
1469	out_unlock:
1470
1471	return neigh;
1472	}
1473
1474	void 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(ah: neigh->ah);
1482	while ((skb = __skb_dequeue(list: &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(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(objp: neigh);
1493	if (atomic_dec_and_test(v: &priv->ntbl.entries)) {
1494	if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1495	complete(&priv->ntbl.flushed);
1496	}
1497	}
1498
1499	static 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
1507	void 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, daddr: 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(entry: &neigh->list);
1536	call_rcu(head: &neigh->rcu, func: ipoib_neigh_reclaim);
1537	return;
1538	} else {
1539	np = &n->hnext;
1540	}
1541	}
1542	}
1543
1544	static 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(nr: IPOIB_NEIGH_TBL_FLUSH, addr: &priv->flags);
1552	ntbl->htbl = NULL;
1553	htbl = kzalloc(size: sizeof(*htbl), GFP_KERNEL);
1554	if (!htbl)
1555	return -ENOMEM;
1556	size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1557	buckets = kvcalloc(n: size, size: sizeof(*buckets), GFP_KERNEL);
1558	if (!buckets) {
1559	kfree(objp: 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(v: &ntbl->entries, i: 0);
1568
1569	/* start garbage collection */
1570	queue_delayed_work(wq: priv->wq, dwork: &priv->neigh_reap_task,
1571	delay: arp_tbl.gc_interval);
1572
1573	return 0;
1574	}
1575
1576	static 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(addr: buckets);
1585	kfree(objp: htbl);
1586	complete(&ntbl->deleted);
1587	}
1588
1589	void 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(p: gid, q: neigh->daddr + 4, size: 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(entry: &neigh->list);
1619	call_rcu(head: &neigh->rcu, func: ipoib_neigh_reclaim);
1620	} else {
1621	np = &neigh->hnext;
1622	}
1623
1624	}
1625	}
1626	out_unlock:
1627	spin_unlock_irqrestore(lock: &priv->lock, flags);
1628	}
1629
1630	static 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(x: &priv->ntbl.flushed);
1638	set_bit(nr: IPOIB_NEIGH_TBL_FLUSH, addr: &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(v: &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(entry: &neigh->list);
1662	call_rcu(head: &neigh->rcu, func: ipoib_neigh_reclaim);
1663	}
1664	}
1665
1666	free_htbl:
1667	rcu_assign_pointer(ntbl->htbl, NULL);
1668	call_rcu(head: &htbl->rcu, func: neigh_hash_free_rcu);
1669
1670	out_unlock:
1671	spin_unlock_irqrestore(lock: &priv->lock, flags);
1672	if (wait_flushed)
1673	wait_for_completion(&priv->ntbl.flushed);
1674	}
1675
1676	static 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(x: &priv->ntbl.deleted);
1682
1683	cancel_delayed_work_sync(dwork: &priv->neigh_reap_task);
1684
1685	ipoib_flush_neighs(priv);
1686
1687	wait_for_completion(&priv->ntbl.deleted);
1688	}
1689
1690	static 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, napi: &priv->recv_napi, poll: ipoib_rx_poll,
1695	weight: IPOIB_NUM_WC);
1696	netif_napi_add_weight(dev, napi: &priv->send_napi, poll: ipoib_tx_poll,
1697	weight: MAX_SEND_CQE);
1698	}
1699
1700	static void ipoib_napi_del(struct net_device *dev)
1701	{
1702	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1703
1704	netif_napi_del(napi: &priv->recv_napi);
1705	netif_napi_del(napi: &priv->send_napi);
1706	}
1707
1708	static 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(objp: priv->rx_ring);
1719	vfree(addr: priv->tx_ring);
1720
1721	priv->rx_ring = NULL;
1722	priv->tx_ring = NULL;
1723	}
1724
1725	static 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(n: ipoib_recvq_size,
1734	size: 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, ca: 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(dev: priv->dev, offset: 1, addr: addr_mod, len: sizeof(addr_mod));
1760
1761	return 0;
1762
1763	out_tx_ring_cleanup:
1764	vfree(addr: priv->tx_ring);
1765
1766	out_rx_ring_cleanup:
1767	kfree(objp: priv->rx_ring);
1768
1769	out:
1770	ipoib_napi_del(dev);
1771	return -ENOMEM;
1772	}
1773
1774	static 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
1785	static 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(ptr: 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
1831	out_hash_uninit:
1832	ipoib_neigh_hash_uninit(dev);
1833
1834	out_dev_uninit:
1835	ipoib_ib_dev_cleanup(dev);
1836
1837	out_free_pd:
1838	if (priv->pd) {
1839	ib_dealloc_pd(pd: priv->pd);
1840	priv->pd = NULL;
1841	}
1842
1843	clean_wq:
1844	if (priv->wq) {
1845	destroy_workqueue(wq: priv->wq);
1846	priv->wq = NULL;
1847	}
1848
1849	out:
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	*/
1857	static void ipoib_parent_unregister_pre(struct net_device *ndev)
1858	{
1859	struct ipoib_dev_priv *priv = ipoib_priv(dev: 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(dev: priv->dev, flags: priv->dev->flags & ~IFF_UP, NULL);
1867	rtnl_unlock();
1868
1869	/* ipoib_event() cannot be running once this returns */
1870	ib_unregister_event_handler(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
1879	static 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
1894	static int ipoib_parent_init(struct net_device *ndev)
1895	{
1896	struct ipoib_dev_priv *priv = ipoib_priv(dev: ndev);
1897	struct ib_port_attr attr;
1898	int result;
1899
1900	result = ib_query_port(device: priv->ca, port_num: priv->port, port_attr: &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(device: priv->ca, port: priv->port, attr: &attr);
1907
1908	result = ib_query_pkey(device: priv->ca, port_num: priv->port, index: 0, pkey: &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(device: priv->ca, port_num: priv->port, index: 0, gid: &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(dev: priv->dev, offset: 4, addr: priv->local_gid.raw, len: 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
1931	static void ipoib_child_init(struct net_device *ndev)
1932	{
1933	struct ipoib_dev_priv *priv = ipoib_priv(dev: ndev);
1934	struct ipoib_dev_priv *ppriv = ipoib_priv(dev: priv->parent);
1935
1936	priv->max_ib_mtu = ppriv->max_ib_mtu;
1937	set_bit(nr: IPOIB_FLAG_SUBINTERFACE, addr: &priv->flags);
1938	if (memchr_inv(p: priv->dev->dev_addr, c: 0, INFINIBAND_ALEN))
1939	memcpy(&priv->local_gid, priv->dev->dev_addr + 4,
1940	sizeof(priv->local_gid));
1941	else {
1942	__dev_addr_set(dev: priv->dev, addr: ppriv->dev->dev_addr,
1943	INFINIBAND_ALEN);
1944	memcpy(&priv->local_gid, &ppriv->local_gid,
1945	sizeof(priv->local_gid));
1946	}
1947	}
1948
1949	static int ipoib_ndo_init(struct net_device *ndev)
1950	{
1951	struct ipoib_dev_priv *priv = ipoib_priv(dev: ndev);
1952	int rc;
1953	struct rdma_netdev *rn = netdev_priv(dev: 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(nr: IPOIB_FLAG_DEV_ADDR_SET, addr: &priv->flags);
1980
1981	ipoib_set_dev_features(priv);
1982
1983	rc = ipoib_dev_init(dev: 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(dev: priv->parent);
1992
1993	dev_hold(dev: priv->parent);
1994
1995	down_write(sem: &ppriv->vlan_rwsem);
1996	list_add_tail(new: &priv->list, head: &ppriv->child_intfs);
1997	up_write(sem: &ppriv->vlan_rwsem);
1998	}
1999
2000	return 0;
2001	}
2002
2003	static 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(dev: priv->parent);
2017
2018	down_write(sem: &ppriv->vlan_rwsem);
2019	list_del(entry: &priv->list);
2020	up_write(sem: &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(wq: priv->wq);
2032	priv->wq = NULL;
2033	}
2034
2035	dev_put(dev: priv->parent);
2036	}
2037
2038	static 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(device: priv->ca, vf, port: priv->port, state: link_state);
2043	}
2044
2045	static 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(device: priv->ca, vf, port: priv->port, info: 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
2061	static 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(device: priv->ca, vf, port: priv->port, guid, type);
2069	}
2070
2071	static 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(device: priv->ca, vf, port: priv->port, node_guid, port_guid);
2078	}
2079
2080	static 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(device: priv->ca, vf, port: priv->port, stats: vf_stats);
2086	}
2087
2088	static const struct header_ops ipoib_header_ops = {
2089	.create = ipoib_hard_header,
2090	};
2091
2092	static 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
2113	static 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
2128	static 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
2135	void 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
2164	static 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(list: &priv->path_list);
2174	INIT_LIST_HEAD(list: &priv->child_intfs);
2175	INIT_LIST_HEAD(list: &priv->dead_ahs);
2176	INIT_LIST_HEAD(list: &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
2190	static 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(device: hca, port_num: port, type: RDMA_NETDEV_IPOIB, name,
2196	NET_NAME_UNKNOWN, setup: ipoib_setup_common);
2197	if (!IS_ERR(ptr: dev) || PTR_ERR(ptr: 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(error: -ENOMEM);
2204	return dev;
2205	}
2206
2207	int 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(size: 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(device: hca, port_num: port, type: RDMA_NETDEV_IPOIB, name,
2222	NET_NAME_UNKNOWN, setup: ipoib_setup_common, netdev: 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, txq: 1);
2233	if (rc)
2234	goto out;
2235
2236	rc = netif_set_real_num_rx_queues(dev, rxq: 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
2261	out:
2262	kfree(objp: priv);
2263	return rc;
2264	}
2265
2266	struct 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(ptr: 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(error: 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
2290	void 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(objp: priv);
2309	}
2310
2311	static 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(dev: ndev);
2316
2317	return sysfs_emit(buf, fmt: "0x%04x\n", priv->pkey);
2318	}
2319	static DEVICE_ATTR_RO(pkey);
2320
2321	static 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(dev: ndev);
2326
2327	return sysfs_emit(buf, fmt: "%d\n",
2328	test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2329	}
2330
2331	void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2332	{
2333	struct ipoib_dev_priv *priv = ipoib_priv(dev: ndev);
2334
2335	if (umcast_val > 0) {
2336	set_bit(nr: IPOIB_FLAG_UMCAST, addr: &priv->flags);
2337	ipoib_warn(priv, "ignoring multicast groups joined directly "
2338	"by userspace\n");
2339	} else
2340	clear_bit(nr: IPOIB_FLAG_UMCAST, addr: &priv->flags);
2341	}
2342
2343	static 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	}
2352	static DEVICE_ATTR_RW(umcast);
2353
2354	int ipoib_add_umcast_attr(struct net_device *dev)
2355	{
2356	return device_create_file(device: &dev->dev, entry: &dev_attr_umcast);
2357	}
2358
2359	static 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(dev: 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(dev: netdev, offset: 4, addr: (u8 *)&priv->local_gid, len: sizeof(priv->local_gid));
2370	clear_bit(nr: IPOIB_FLAG_DEV_ADDR_SET, addr: &priv->flags);
2371
2372	netif_addr_unlock_bh(dev: netdev);
2373
2374	if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2375	down_read(sem: &priv->vlan_rwsem);
2376	list_for_each_entry(child_priv, &priv->child_intfs, list)
2377	set_base_guid(priv: child_priv, gid);
2378	up_read(sem: &priv->vlan_rwsem);
2379	}
2380	}
2381
2382	static 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(p: dev->dev_addr, q: ss->__data,
2394	size: 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
2403	static 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, gid: (union ib_gid *)(ss->__data + 4));
2417
2418	queue_work(wq: ipoib_workqueue, work: &priv->flush_light);
2419
2420	return 0;
2421	}
2422
2423	static 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	}
2440	static DEVICE_ATTR_WO(create_child);
2441
2442	static 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	}
2460	static DEVICE_ATTR_WO(delete_child);
2461
2462	int ipoib_add_pkey_attr(struct net_device *dev)
2463	{
2464	return device_create_file(device: &dev->dev, entry: &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	*/
2476	static 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, fmt: "%#x\n", ndev->dev_id);
2498	}
2499	static DEVICE_ATTR_RO(dev_id);
2500
2501	static int ipoib_intercept_dev_id_attr(struct net_device *dev)
2502	{
2503	device_remove_file(dev: &dev->dev, attr: &dev_attr_dev_id);
2504	return device_create_file(device: &dev->dev, entry: &dev_attr_dev_id);
2505	}
2506
2507	static 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, name: format);
2517	if (IS_ERR(ptr: 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(dev: ndev);
2523
2524	INIT_IB_EVENT_HANDLER(&priv->event_handler,
2525	priv->ca, ipoib_event);
2526	ib_register_event_handler(event_handler: &priv->event_handler);
2527
2528	/* call event handler to ensure pkey in sync */
2529	queue_work(wq: ipoib_workqueue, work: &priv->flush_heavy);
2530
2531	ndev->rtnl_link_ops = ipoib_get_link_ops();
2532
2533	result = register_netdev(dev: 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(dev: ndev);
2540	free_netdev(dev: ndev);
2541
2542	return ERR_PTR(error: 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	&params);
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(dev: ndev))
2562	goto sysfs_failed;
2563	if (ipoib_cm_add_mode_attr(dev: ndev))
2564	goto sysfs_failed;
2565	if (ipoib_add_pkey_attr(dev: ndev))
2566	goto sysfs_failed;
2567	if (ipoib_add_umcast_attr(dev: ndev))
2568	goto sysfs_failed;
2569	if (device_create_file(device: &ndev->dev, entry: &dev_attr_create_child))
2570	goto sysfs_failed;
2571	if (device_create_file(device: &ndev->dev, entry: &dev_attr_delete_child))
2572	goto sysfs_failed;
2573
2574	return ndev;
2575
2576	sysfs_failed:
2577	ipoib_parent_unregister_pre(ndev);
2578	unregister_netdev(dev: ndev);
2579	return ERR_PTR(error: -ENOMEM);
2580	}
2581
2582	static 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(size: sizeof(*dev_list), GFP_KERNEL);
2591	if (!dev_list)
2592	return -ENOMEM;
2593
2594	INIT_LIST_HEAD(list: dev_list);
2595
2596	rdma_for_each_port (device, p) {
2597	if (!rdma_protocol_ib(device, port_num: p))
2598	continue;
2599	dev = ipoib_add_port(format: "ib%d", hca: device, port: p);
2600	if (!IS_ERR(ptr: dev)) {
2601	priv = ipoib_priv(dev);
2602	list_add_tail(new: &priv->list, head: dev_list);
2603	count++;
2604	}
2605	}
2606
2607	if (!count) {
2608	kfree(objp: dev_list);
2609	return -EOPNOTSUPP;
2610	}
2611
2612	ib_set_client_data(device, client: &ipoib_client, data: dev_list);
2613	return 0;
2614	}
2615
2616	static 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(ndev: priv->dev);
2624
2625	rtnl_lock();
2626
2627	list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2628	list)
2629	unregister_netdevice_queue(dev: cpriv->dev, head: &head);
2630	unregister_netdevice_queue(dev: priv->dev, head: &head);
2631	unregister_netdevice_many(head: &head);
2632
2633	rtnl_unlock();
2634	}
2635
2636	kfree(objp: dev_list);
2637	}
2638
2639	#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2640	static struct notifier_block ipoib_netdev_notifier = {
2641	.notifier_call = ipoib_netdev_event,
2642	};
2643	#endif
2644
2645	static 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(client: &ipoib_sa_client);
2686
2687	ret = ib_register_client(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(nb: &ipoib_netdev_notifier);
2697	#endif
2698	return 0;
2699
2700	err_client:
2701	ib_unregister_client(client: &ipoib_client);
2702
2703	err_sa:
2704	ib_sa_unregister_client(client: &ipoib_sa_client);
2705	destroy_workqueue(wq: ipoib_workqueue);
2706
2707	err_fs:
2708	ipoib_unregister_debugfs();
2709
2710	return ret;
2711	}
2712
2713	static void __exit ipoib_cleanup_module(void)
2714	{
2715	#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2716	unregister_netdevice_notifier(nb: &ipoib_netdev_notifier);
2717	#endif
2718	ipoib_netlink_fini();
2719	ib_unregister_client(client: &ipoib_client);
2720	ib_sa_unregister_client(client: &ipoib_sa_client);
2721	ipoib_unregister_debugfs();
2722	destroy_workqueue(wq: ipoib_workqueue);
2723	}
2724
2725	module_init(ipoib_init_module);
2726	module_exit(ipoib_cleanup_module);
2727