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.
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:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
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.
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
37 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
44 #include <linux/if_arp.h> /* For ARPHRD_xxx */
49 #include <linux/jhash.h>
51 #include <net/addrconf.h>
52 #include <linux/inetdevice.h>
53 #include <rdma/ib_cache.h>
55 MODULE_AUTHOR("Roland Dreier");
56 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
57 MODULE_LICENSE("Dual BSD/GPL");
59 int ipoib_sendq_size __read_mostly
= IPOIB_TX_RING_SIZE
;
60 int ipoib_recvq_size __read_mostly
= IPOIB_RX_RING_SIZE
;
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");
67 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
68 int ipoib_debug_level
;
70 module_param_named(debug_level
, ipoib_debug_level
, int, 0644);
71 MODULE_PARM_DESC(debug_level
, "Enable debug tracing if > 0");
74 struct ipoib_path_iter
{
75 struct net_device
*dev
;
76 struct ipoib_path path
;
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
85 struct workqueue_struct
*ipoib_workqueue
;
87 struct ib_sa_client ipoib_sa_client
;
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
,
96 static int ipoib_set_mac(struct net_device
*dev
, void *addr
);
97 static int ipoib_ioctl(struct net_device
*dev
, struct ifreq
*ifr
,
100 static struct ib_client ipoib_client
= {
102 .add
= ipoib_add_one
,
103 .remove
= ipoib_remove_one
,
104 .get_net_dev_by_params
= ipoib_get_net_dev_by_params
,
107 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
108 static int ipoib_netdev_event(struct notifier_block
*this,
109 unsigned long event
, void *ptr
)
111 struct netdev_notifier_info
*ni
= ptr
;
112 struct net_device
*dev
= ni
->dev
;
114 if (dev
->netdev_ops
->ndo_open
!= ipoib_open
)
118 case NETDEV_REGISTER
:
119 ipoib_create_debug_files(dev
);
121 case NETDEV_CHANGENAME
:
122 ipoib_delete_debug_files(dev
);
123 ipoib_create_debug_files(dev
);
125 case NETDEV_UNREGISTER
:
126 ipoib_delete_debug_files(dev
);
134 int ipoib_open(struct net_device
*dev
)
136 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
138 ipoib_dbg(priv
, "bringing up interface\n");
140 netif_carrier_off(dev
);
142 set_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
144 if (ipoib_ib_dev_open(dev
)) {
145 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
))
150 ipoib_ib_dev_up(dev
);
152 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
153 struct ipoib_dev_priv
*cpriv
;
155 /* Bring up any child interfaces too */
156 down_read(&priv
->vlan_rwsem
);
157 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
160 flags
= cpriv
->dev
->flags
;
164 dev_change_flags(cpriv
->dev
, flags
| IFF_UP
, NULL
);
166 up_read(&priv
->vlan_rwsem
);
167 } else if (priv
->parent
) {
168 struct ipoib_dev_priv
*ppriv
= ipoib_priv(priv
->parent
);
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",
174 netif_start_queue(dev
);
179 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
184 static int ipoib_stop(struct net_device
*dev
)
186 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
188 ipoib_dbg(priv
, "stopping interface\n");
190 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
192 netif_stop_queue(dev
);
194 ipoib_ib_dev_down(dev
);
195 ipoib_ib_dev_stop(dev
);
197 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
198 struct ipoib_dev_priv
*cpriv
;
200 /* Bring down any child interfaces too */
201 down_read(&priv
->vlan_rwsem
);
202 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
205 flags
= cpriv
->dev
->flags
;
206 if (!(flags
& IFF_UP
))
209 dev_change_flags(cpriv
->dev
, flags
& ~IFF_UP
, NULL
);
211 up_read(&priv
->vlan_rwsem
);
217 static netdev_features_t
ipoib_fix_features(struct net_device
*dev
, netdev_features_t features
)
219 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
221 if (test_bit(IPOIB_FLAG_ADMIN_CM
, &priv
->flags
))
222 features
&= ~(NETIF_F_IP_CSUM
| NETIF_F_TSO
);
227 static int ipoib_change_mtu(struct net_device
*dev
, int new_mtu
)
229 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
232 /* dev->mtu > 2K ==> connected mode */
233 if (ipoib_cm_admin_enabled(dev
)) {
234 if (new_mtu
> ipoib_cm_max_mtu(dev
))
237 if (new_mtu
> priv
->mcast_mtu
)
238 ipoib_warn(priv
, "mtu > %d will cause multicast packet drops.\n",
245 if (new_mtu
< (ETH_MIN_MTU
+ IPOIB_ENCAP_LEN
) ||
246 new_mtu
> IPOIB_UD_MTU(priv
->max_ib_mtu
))
249 priv
->admin_mtu
= new_mtu
;
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
);
255 new_mtu
= min(priv
->mcast_mtu
, priv
->admin_mtu
);
257 if (priv
->rn_ops
->ndo_change_mtu
) {
258 bool carrier_status
= netif_carrier_ok(dev
);
260 netif_carrier_off(dev
);
262 /* notify lower level on the real mtu */
263 ret
= priv
->rn_ops
->ndo_change_mtu(dev
, new_mtu
);
266 netif_carrier_on(dev
);
274 static void ipoib_get_stats(struct net_device
*dev
,
275 struct rtnl_link_stats64
*stats
)
277 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
279 if (priv
->rn_ops
->ndo_get_stats64
)
280 priv
->rn_ops
->ndo_get_stats64(dev
, stats
);
282 netdev_stats_to_stats64(stats
, &dev
->stats
);
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
)
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
;
295 switch (addr
->sa_family
) {
297 in_dev
= in_dev_get(dev
);
301 ret_addr
= inet_confirm_addr(net
, in_dev
, 0,
302 addr_in
->sin_addr
.s_addr
,
310 if (IS_ENABLED(CONFIG_IPV6
) &&
311 ipv6_chk_addr(net
, &addr_in6
->sin6_addr
, dev
, 1))
320 * Find the master net_device on top of the given net_device.
321 * @dev: base IPoIB net_device
323 * Returns the master net_device with a reference held, or the same net_device
324 * if no master exists.
326 static struct net_device
*ipoib_get_master_net_dev(struct net_device
*dev
)
328 struct net_device
*master
;
331 master
= netdev_master_upper_dev_get_rcu(dev
);
343 struct ipoib_walk_data
{
344 const struct sockaddr
*addr
;
345 struct net_device
*result
;
348 static int ipoib_upper_walk(struct net_device
*upper
,
349 struct netdev_nested_priv
*priv
)
351 struct ipoib_walk_data
*data
= (struct ipoib_walk_data
*)priv
->data
;
354 if (ipoib_is_dev_match_addr_rcu(data
->addr
, upper
)) {
356 data
->result
= upper
;
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.
367 * @addr: IP address to look for.
368 * @dev: base IPoIB net_device
370 * If found, returns the net_device with a reference held. Otherwise return
373 static struct net_device
*ipoib_get_net_dev_match_addr(
374 const struct sockaddr
*addr
, struct net_device
*dev
)
376 struct netdev_nested_priv priv
;
377 struct ipoib_walk_data data
= {
381 priv
.data
= (void *)&data
;
383 if (ipoib_is_dev_match_addr_rcu(addr
, dev
)) {
389 netdev_walk_all_upper_dev_rcu(dev
, ipoib_upper_walk
, &priv
);
395 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
396 * pkey_index and address, if one exists.
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
,
403 const struct sockaddr
*addr
,
405 struct net_device
**found_net_dev
)
407 struct ipoib_dev_priv
*child_priv
;
408 struct net_device
*net_dev
= NULL
;
411 if (priv
->pkey_index
== pkey_index
&&
412 (!gid
|| !memcmp(gid
, &priv
->local_gid
, sizeof(*gid
)))) {
414 net_dev
= ipoib_get_master_net_dev(priv
->dev
);
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
);
422 *found_net_dev
= net_dev
;
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
,
439 up_read(&priv
->vlan_rwsem
);
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
,
449 const union ib_gid
*gid
,
450 const struct sockaddr
*addr
,
451 struct net_device
**net_dev
)
453 struct ipoib_dev_priv
*priv
;
458 list_for_each_entry(priv
, dev_list
, list
) {
459 if (priv
->port
!= port
)
462 matches
+= ipoib_match_gid_pkey_addr(priv
, gid
, pkey_index
,
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
,
476 struct net_device
*net_dev
;
477 struct list_head
*dev_list
= client_data
;
482 if (!rdma_protocol_ib(dev
, port
))
485 ret
= ib_find_cached_pkey(dev
, port
, pkey
, &pkey_index
);
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
);
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
);
510 dev_warn_ratelimited(&dev
->dev
,
511 "duplicate IP address detected\n");
518 int ipoib_set_mode(struct net_device
*dev
, const char *buf
)
520 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
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"))) {
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);
538 priv
->tx_wr
.wr
.send_flags
&= ~IB_SEND_IP_CSUM
;
540 ipoib_flush_paths(dev
);
541 return (!rtnl_trylock()) ? -EBUSY
: 0;
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
);
550 ipoib_flush_paths(dev
);
551 return (!rtnl_trylock()) ? -EBUSY
: 0;
557 struct ipoib_path
*__path_find(struct net_device
*dev
, void *gid
)
559 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
560 struct rb_node
*n
= priv
->path_tree
.rb_node
;
561 struct ipoib_path
*path
;
565 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
567 ret
= memcmp(gid
, path
->pathrec
.dgid
.raw
,
568 sizeof (union ib_gid
));
581 static int __path_add(struct net_device
*dev
, struct ipoib_path
*path
)
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
;
591 tpath
= rb_entry(pn
, struct ipoib_path
, rb_node
);
593 ret
= memcmp(path
->pathrec
.dgid
.raw
, tpath
->pathrec
.dgid
.raw
,
594 sizeof (union ib_gid
));
603 rb_link_node(&path
->rb_node
, pn
, n
);
604 rb_insert_color(&path
->rb_node
, &priv
->path_tree
);
606 list_add_tail(&path
->list
, &priv
->path_list
);
611 static void path_free(struct net_device
*dev
, struct ipoib_path
*path
)
615 while ((skb
= __skb_dequeue(&path
->queue
)))
616 dev_kfree_skb_irq(skb
);
618 ipoib_dbg(ipoib_priv(dev
), "%s\n", __func__
);
620 /* remove all neigh connected to this path */
621 ipoib_del_neighs_by_gid(dev
, path
->pathrec
.dgid
.raw
);
624 ipoib_put_ah(path
->ah
);
629 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
631 struct ipoib_path_iter
*ipoib_path_iter_init(struct net_device
*dev
)
633 struct ipoib_path_iter
*iter
;
635 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
640 memset(iter
->path
.pathrec
.dgid
.raw
, 0, 16);
642 if (ipoib_path_iter_next(iter
)) {
650 int ipoib_path_iter_next(struct ipoib_path_iter
*iter
)
652 struct ipoib_dev_priv
*priv
= ipoib_priv(iter
->dev
);
654 struct ipoib_path
*path
;
657 spin_lock_irq(&priv
->lock
);
659 n
= rb_first(&priv
->path_tree
);
662 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
664 if (memcmp(iter
->path
.pathrec
.dgid
.raw
, path
->pathrec
.dgid
.raw
,
665 sizeof (union ib_gid
)) < 0) {
674 spin_unlock_irq(&priv
->lock
);
679 void ipoib_path_iter_read(struct ipoib_path_iter
*iter
,
680 struct ipoib_path
*path
)
685 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
687 void ipoib_mark_paths_invalid(struct net_device
*dev
)
689 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
690 struct ipoib_path
*path
, *tp
;
692 spin_lock_irq(&priv
->lock
);
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
);
702 spin_unlock_irq(&priv
->lock
);
705 static void push_pseudo_header(struct sk_buff
*skb
, const char *daddr
)
707 struct ipoib_pseudo_header
*phdr
;
709 phdr
= skb_push(skb
, sizeof(*phdr
));
710 memcpy(phdr
->hwaddr
, daddr
, INFINIBAND_ALEN
);
713 void ipoib_flush_paths(struct net_device
*dev
)
715 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
716 struct ipoib_path
*path
, *tp
;
717 LIST_HEAD(remove_list
);
720 netif_tx_lock_bh(dev
);
721 spin_lock_irqsave(&priv
->lock
, flags
);
723 list_splice_init(&priv
->path_list
, &remove_list
);
725 list_for_each_entry(path
, &remove_list
, list
)
726 rb_erase(&path
->rb_node
, &priv
->path_tree
);
728 list_for_each_entry_safe(path
, tp
, &remove_list
, list
) {
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
);
739 spin_unlock_irqrestore(&priv
->lock
, flags
);
740 netif_tx_unlock_bh(dev
);
743 static void path_rec_completion(int status
,
744 struct sa_path_rec
*pathrec
,
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
;
758 ipoib_dbg(priv
, "PathRec LID 0x%04x for GID %pI6\n",
759 be32_to_cpu(sa_path_get_dlid(pathrec
)),
762 ipoib_dbg(priv
, "PathRec status %d for GID %pI6\n",
763 status
, path
->pathrec
.dgid
.raw
);
765 skb_queue_head_init(&skqueue
);
768 struct rdma_ah_attr av
;
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
);
777 spin_lock_irqsave(&priv
->lock
, flags
);
779 if (!IS_ERR_OR_NULL(ah
)) {
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.
785 if (memcmp(pathrec
->dgid
.raw
, path
->pathrec
.dgid
.raw
,
786 sizeof(union ib_gid
))) {
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
));
796 path
->pathrec
= *pathrec
;
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
)),
805 while ((skb
= __skb_dequeue(&path
->queue
)))
806 __skb_queue_tail(&skqueue
, skb
);
808 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
810 WARN_ON(neigh
->ah
!= old_ah
);
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
818 ipoib_put_ah(neigh
->ah
);
820 kref_get(&path
->ah
->ref
);
821 neigh
->ah
= path
->ah
;
823 if (ipoib_cm_enabled(dev
, neigh
->daddr
)) {
824 if (!ipoib_cm_get(neigh
))
825 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
,
828 if (!ipoib_cm_get(neigh
)) {
829 ipoib_neigh_free(neigh
);
834 while ((skb
= __skb_dequeue(&neigh
->queue
)))
835 __skb_queue_tail(&skqueue
, skb
);
841 complete(&path
->done
);
843 spin_unlock_irqrestore(&priv
->lock
, flags
);
845 if (IS_ERR_OR_NULL(ah
))
846 ipoib_del_neighs_by_gid(dev
, path
->pathrec
.dgid
.raw
);
849 ipoib_put_ah(old_ah
);
851 while ((skb
= __skb_dequeue(&skqueue
))) {
854 ret
= dev_queue_xmit(skb
);
856 ipoib_warn(priv
, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
861 static void init_path_rec(struct ipoib_dev_priv
*priv
, struct ipoib_path
*path
,
864 path
->dev
= priv
->dev
;
866 if (rdma_cap_opa_ah(priv
->ca
, priv
->port
))
867 path
->pathrec
.rec_type
= SA_PATH_REC_TYPE_OPA
;
869 path
->pathrec
.rec_type
= SA_PATH_REC_TYPE_IB
;
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
;
878 static struct ipoib_path
*path_rec_create(struct net_device
*dev
, void *gid
)
880 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
881 struct ipoib_path
*path
;
883 if (!priv
->broadcast
)
886 path
= kzalloc(sizeof(*path
), GFP_ATOMIC
);
890 skb_queue_head_init(&path
->queue
);
892 INIT_LIST_HEAD(&path
->neigh_list
);
894 init_path_rec(priv
, path
, gid
);
899 static int path_rec_start(struct net_device
*dev
,
900 struct ipoib_path
*path
)
902 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
904 ipoib_dbg(priv
, "Start path record lookup for %pI6\n",
905 path
->pathrec
.dgid
.raw
);
907 init_completion(&path
->done
);
910 ib_sa_path_rec_get(&ipoib_sa_client
, priv
->ca
, priv
->port
,
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
|
920 if (path
->query_id
< 0) {
921 ipoib_warn(priv
, "ib_sa_path_rec_get failed: %d\n", path
->query_id
);
923 complete(&path
->done
);
924 return path
->query_id
;
930 static void neigh_refresh_path(struct ipoib_neigh
*neigh
, u8
*daddr
,
931 struct net_device
*dev
)
933 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
934 struct ipoib_path
*path
;
937 spin_lock_irqsave(&priv
->lock
, flags
);
939 path
= __path_find(dev
, daddr
+ 4);
943 path_rec_start(dev
, path
);
945 spin_unlock_irqrestore(&priv
->lock
, flags
);
948 static struct ipoib_neigh
*neigh_add_path(struct sk_buff
*skb
, u8
*daddr
,
949 struct net_device
*dev
)
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
;
957 spin_lock_irqsave(&priv
->lock
, flags
);
958 neigh
= ipoib_neigh_alloc(daddr
, dev
);
960 spin_unlock_irqrestore(&priv
->lock
, flags
);
961 ++dev
->stats
.tx_dropped
;
962 dev_kfree_skb_any(skb
);
966 /* To avoid race condition, make sure that the
967 * neigh will be added only once.
969 if (unlikely(!list_empty(&neigh
->list
))) {
970 spin_unlock_irqrestore(&priv
->lock
, flags
);
974 path
= __path_find(dev
, daddr
+ 4);
976 path
= path_rec_create(dev
, daddr
+ 4);
980 __path_add(dev
, path
);
983 list_add_tail(&neigh
->list
, &path
->neigh_list
);
985 if (path
->ah
&& path
->ah
->valid
) {
986 kref_get(&path
->ah
->ref
);
987 neigh
->ah
= path
->ah
;
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
);
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
);
1001 ipoib_warn(priv
, "queue length limit %d. Packet drop.\n",
1002 skb_queue_len(&neigh
->queue
));
1006 spin_unlock_irqrestore(&priv
->lock
, flags
);
1007 path
->ah
->last_send
= rn
->send(dev
, skb
, path
->ah
->ah
,
1009 ipoib_neigh_put(neigh
);
1015 if (!path
->query
&& path_rec_start(dev
, 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
);
1025 spin_unlock_irqrestore(&priv
->lock
, flags
);
1026 ipoib_neigh_put(neigh
);
1030 ipoib_neigh_free(neigh
);
1032 ++dev
->stats
.tx_dropped
;
1033 dev_kfree_skb_any(skb
);
1035 spin_unlock_irqrestore(&priv
->lock
, flags
);
1036 ipoib_neigh_put(neigh
);
1041 static void unicast_arp_send(struct sk_buff
*skb
, struct net_device
*dev
,
1042 struct ipoib_pseudo_header
*phdr
)
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
;
1049 spin_lock_irqsave(&priv
->lock
, flags
);
1051 /* no broadcast means that all paths are (going to be) not valid */
1052 if (!priv
->broadcast
)
1053 goto drop_and_unlock
;
1055 path
= __path_find(dev
, phdr
->hwaddr
+ 4);
1056 if (!path
|| !path
->ah
|| !path
->ah
->valid
) {
1058 path
= path_rec_create(dev
, phdr
->hwaddr
+ 4);
1060 goto drop_and_unlock
;
1061 __path_add(dev
, path
);
1064 * make sure there are no changes in the existing
1067 init_path_rec(priv
, path
, phdr
->hwaddr
+ 4);
1069 if (!path
->query
&& path_rec_start(dev
, path
)) {
1070 goto drop_and_unlock
;
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
);
1078 goto drop_and_unlock
;
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
));
1090 ++dev
->stats
.tx_dropped
;
1091 dev_kfree_skb_any(skb
);
1093 spin_unlock_irqrestore(&priv
->lock
, flags
);
1096 static netdev_tx_t
ipoib_start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
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
;
1105 phdr
= (struct ipoib_pseudo_header
*) skb
->data
;
1106 skb_pull(skb
, sizeof(*phdr
));
1107 header
= (struct ipoib_header
*) skb
->data
;
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
;
1121 /* Add in the P_Key for multicast*/
1122 phdr
->hwaddr
[8] = (priv
->pkey
>> 8) & 0xff;
1123 phdr
->hwaddr
[9] = priv
->pkey
& 0xff;
1125 neigh
= ipoib_neigh_get(dev
, phdr
->hwaddr
);
1127 goto send_using_neigh
;
1128 ipoib_mcast_send(dev
, phdr
->hwaddr
, skb
);
1129 return NETDEV_TX_OK
;
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
);
1141 return NETDEV_TX_OK
;
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
;
1150 /* ethertype not supported by IPoIB */
1151 ++dev
->stats
.tx_dropped
;
1152 dev_kfree_skb_any(skb
);
1153 return NETDEV_TX_OK
;
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
));
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
));
1167 } else if (neigh
->ah
) {
1168 neigh_refresh_path(neigh
, phdr
->hwaddr
, dev
);
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
);
1177 ++dev
->stats
.tx_dropped
;
1178 dev_kfree_skb_any(skb
);
1182 ipoib_neigh_put(neigh
);
1184 return NETDEV_TX_OK
;
1187 static void ipoib_timeout(struct net_device
*dev
, unsigned int txqueue
)
1189 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1190 struct rdma_netdev
*rn
= netdev_priv(dev
);
1192 if (rn
->tx_timeout
) {
1193 rn
->tx_timeout(dev
, txqueue
);
1196 ipoib_warn(priv
, "transmit timeout: latency %d msecs\n",
1197 jiffies_to_msecs(jiffies
- dev_trans_start(dev
)));
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
);
1203 /* XXX reset QP, etc. */
1206 static int ipoib_hard_header(struct sk_buff
*skb
,
1207 struct net_device
*dev
,
1208 unsigned short type
,
1213 struct ipoib_header
*header
;
1215 header
= skb_push(skb
, sizeof(*header
));
1217 header
->proto
= htons(type
);
1218 header
->reserved
= 0;
1221 * we don't rely on dst_entry structure, always stuff the
1222 * destination address into skb hard header so we can figure out where
1223 * to send the packet later.
1225 push_pseudo_header(skb
, daddr
);
1227 return IPOIB_HARD_LEN
;
1230 static void ipoib_set_mcast_list(struct net_device
*dev
)
1232 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1234 if (!test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
)) {
1235 ipoib_dbg(priv
, "IPOIB_FLAG_OPER_UP not set");
1239 queue_work(priv
->wq
, &priv
->restart_task
);
1242 static int ipoib_get_iflink(const struct net_device
*dev
)
1244 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1246 /* parent interface */
1247 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
))
1248 return dev
->ifindex
;
1250 /* child/vlan interface */
1251 return priv
->parent
->ifindex
;
1254 static u32
ipoib_addr_hash(struct ipoib_neigh_hash
*htbl
, u8
*daddr
)
1257 * Use only the address parts that contributes to spreading
1258 * The subnet prefix is not used as one can not connect to
1259 * same remote port (GUID) using the same remote QPN via two
1260 * different subnets.
1262 /* qpn octets[1:4) & port GUID octets[12:20) */
1263 u32
*d32
= (u32
*) daddr
;
1266 hv
= jhash_3words(d32
[3], d32
[4], IPOIB_QPN_MASK
& d32
[0], 0);
1267 return hv
& htbl
->mask
;
1270 struct ipoib_neigh
*ipoib_neigh_get(struct net_device
*dev
, u8
*daddr
)
1272 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1273 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1274 struct ipoib_neigh_hash
*htbl
;
1275 struct ipoib_neigh
*neigh
= NULL
;
1280 htbl
= rcu_dereference_bh(ntbl
->htbl
);
1285 hash_val
= ipoib_addr_hash(htbl
, daddr
);
1286 for (neigh
= rcu_dereference_bh(htbl
->buckets
[hash_val
]);
1288 neigh
= rcu_dereference_bh(neigh
->hnext
)) {
1289 if (memcmp(daddr
, neigh
->daddr
, INFINIBAND_ALEN
) == 0) {
1290 /* found, take one ref on behalf of the caller */
1291 if (!refcount_inc_not_zero(&neigh
->refcnt
)) {
1297 if (likely(skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
))
1298 neigh
->alive
= jiffies
;
1304 rcu_read_unlock_bh();
1308 static void __ipoib_reap_neigh(struct ipoib_dev_priv
*priv
)
1310 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1311 struct ipoib_neigh_hash
*htbl
;
1312 unsigned long neigh_obsolete
;
1314 unsigned long flags
;
1316 LIST_HEAD(remove_list
);
1318 spin_lock_irqsave(&priv
->lock
, flags
);
1320 htbl
= rcu_dereference_protected(ntbl
->htbl
,
1321 lockdep_is_held(&priv
->lock
));
1326 /* neigh is obsolete if it was idle for two GC periods */
1327 dt
= 2 * arp_tbl
.gc_interval
;
1328 neigh_obsolete
= jiffies
- dt
;
1330 for (i
= 0; i
< htbl
->size
; i
++) {
1331 struct ipoib_neigh
*neigh
;
1332 struct ipoib_neigh __rcu
**np
= &htbl
->buckets
[i
];
1334 while ((neigh
= rcu_dereference_protected(*np
,
1335 lockdep_is_held(&priv
->lock
))) != NULL
) {
1336 /* was the neigh idle for two GC periods */
1337 if (time_after(neigh_obsolete
, neigh
->alive
)) {
1339 ipoib_check_and_add_mcast_sendonly(priv
, neigh
->daddr
+ 4, &remove_list
);
1341 rcu_assign_pointer(*np
,
1342 rcu_dereference_protected(neigh
->hnext
,
1343 lockdep_is_held(&priv
->lock
)));
1344 /* remove from path/mc list */
1345 list_del_init(&neigh
->list
);
1346 call_rcu(&neigh
->rcu
, ipoib_neigh_reclaim
);
1355 spin_unlock_irqrestore(&priv
->lock
, flags
);
1356 ipoib_mcast_remove_list(&remove_list
);
1359 static void ipoib_reap_neigh(struct work_struct
*work
)
1361 struct ipoib_dev_priv
*priv
=
1362 container_of(work
, struct ipoib_dev_priv
, neigh_reap_task
.work
);
1364 __ipoib_reap_neigh(priv
);
1366 queue_delayed_work(priv
->wq
, &priv
->neigh_reap_task
,
1367 arp_tbl
.gc_interval
);
1371 static struct ipoib_neigh
*ipoib_neigh_ctor(u8
*daddr
,
1372 struct net_device
*dev
)
1374 struct ipoib_neigh
*neigh
;
1376 neigh
= kzalloc(sizeof(*neigh
), GFP_ATOMIC
);
1381 memcpy(&neigh
->daddr
, daddr
, sizeof(neigh
->daddr
));
1382 skb_queue_head_init(&neigh
->queue
);
1383 INIT_LIST_HEAD(&neigh
->list
);
1384 ipoib_cm_set(neigh
, NULL
);
1385 /* one ref on behalf of the caller */
1386 refcount_set(&neigh
->refcnt
, 1);
1391 struct ipoib_neigh
*ipoib_neigh_alloc(u8
*daddr
,
1392 struct net_device
*dev
)
1394 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1395 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1396 struct ipoib_neigh_hash
*htbl
;
1397 struct ipoib_neigh
*neigh
;
1400 htbl
= rcu_dereference_protected(ntbl
->htbl
,
1401 lockdep_is_held(&priv
->lock
));
1407 /* need to add a new neigh, but maybe some other thread succeeded?
1408 * recalc hash, maybe hash resize took place so we do a search
1410 hash_val
= ipoib_addr_hash(htbl
, daddr
);
1411 for (neigh
= rcu_dereference_protected(htbl
->buckets
[hash_val
],
1412 lockdep_is_held(&priv
->lock
));
1414 neigh
= rcu_dereference_protected(neigh
->hnext
,
1415 lockdep_is_held(&priv
->lock
))) {
1416 if (memcmp(daddr
, neigh
->daddr
, INFINIBAND_ALEN
) == 0) {
1417 /* found, take one ref on behalf of the caller */
1418 if (!refcount_inc_not_zero(&neigh
->refcnt
)) {
1423 neigh
->alive
= jiffies
;
1428 neigh
= ipoib_neigh_ctor(daddr
, dev
);
1432 /* one ref on behalf of the hash table */
1433 refcount_inc(&neigh
->refcnt
);
1434 neigh
->alive
= jiffies
;
1436 rcu_assign_pointer(neigh
->hnext
,
1437 rcu_dereference_protected(htbl
->buckets
[hash_val
],
1438 lockdep_is_held(&priv
->lock
)));
1439 rcu_assign_pointer(htbl
->buckets
[hash_val
], neigh
);
1440 atomic_inc(&ntbl
->entries
);
1447 void ipoib_neigh_dtor(struct ipoib_neigh
*neigh
)
1449 /* neigh reference count was dropprd to zero */
1450 struct net_device
*dev
= neigh
->dev
;
1451 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1452 struct sk_buff
*skb
;
1454 ipoib_put_ah(neigh
->ah
);
1455 while ((skb
= __skb_dequeue(&neigh
->queue
))) {
1456 ++dev
->stats
.tx_dropped
;
1457 dev_kfree_skb_any(skb
);
1459 if (ipoib_cm_get(neigh
))
1460 ipoib_cm_destroy_tx(ipoib_cm_get(neigh
));
1461 ipoib_dbg(ipoib_priv(dev
),
1462 "neigh free for %06x %pI6\n",
1463 IPOIB_QPN(neigh
->daddr
),
1466 if (atomic_dec_and_test(&priv
->ntbl
.entries
)) {
1467 if (test_bit(IPOIB_NEIGH_TBL_FLUSH
, &priv
->flags
))
1468 complete(&priv
->ntbl
.flushed
);
1472 static void ipoib_neigh_reclaim(struct rcu_head
*rp
)
1474 /* Called as a result of removal from hash table */
1475 struct ipoib_neigh
*neigh
= container_of(rp
, struct ipoib_neigh
, rcu
);
1476 /* note TX context may hold another ref */
1477 ipoib_neigh_put(neigh
);
1480 void ipoib_neigh_free(struct ipoib_neigh
*neigh
)
1482 struct net_device
*dev
= neigh
->dev
;
1483 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1484 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1485 struct ipoib_neigh_hash
*htbl
;
1486 struct ipoib_neigh __rcu
**np
;
1487 struct ipoib_neigh
*n
;
1490 htbl
= rcu_dereference_protected(ntbl
->htbl
,
1491 lockdep_is_held(&priv
->lock
));
1495 hash_val
= ipoib_addr_hash(htbl
, neigh
->daddr
);
1496 np
= &htbl
->buckets
[hash_val
];
1497 for (n
= rcu_dereference_protected(*np
,
1498 lockdep_is_held(&priv
->lock
));
1500 n
= rcu_dereference_protected(*np
,
1501 lockdep_is_held(&priv
->lock
))) {
1504 rcu_assign_pointer(*np
,
1505 rcu_dereference_protected(neigh
->hnext
,
1506 lockdep_is_held(&priv
->lock
)));
1507 /* remove from parent list */
1508 list_del_init(&neigh
->list
);
1509 call_rcu(&neigh
->rcu
, ipoib_neigh_reclaim
);
1517 static int ipoib_neigh_hash_init(struct ipoib_dev_priv
*priv
)
1519 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1520 struct ipoib_neigh_hash
*htbl
;
1521 struct ipoib_neigh __rcu
**buckets
;
1524 clear_bit(IPOIB_NEIGH_TBL_FLUSH
, &priv
->flags
);
1526 htbl
= kzalloc(sizeof(*htbl
), GFP_KERNEL
);
1529 size
= roundup_pow_of_two(arp_tbl
.gc_thresh3
);
1530 buckets
= kvcalloc(size
, sizeof(*buckets
), GFP_KERNEL
);
1536 htbl
->mask
= (size
- 1);
1537 htbl
->buckets
= buckets
;
1538 RCU_INIT_POINTER(ntbl
->htbl
, htbl
);
1540 atomic_set(&ntbl
->entries
, 0);
1542 /* start garbage collection */
1543 queue_delayed_work(priv
->wq
, &priv
->neigh_reap_task
,
1544 arp_tbl
.gc_interval
);
1549 static void neigh_hash_free_rcu(struct rcu_head
*head
)
1551 struct ipoib_neigh_hash
*htbl
= container_of(head
,
1552 struct ipoib_neigh_hash
,
1554 struct ipoib_neigh __rcu
**buckets
= htbl
->buckets
;
1555 struct ipoib_neigh_table
*ntbl
= htbl
->ntbl
;
1559 complete(&ntbl
->deleted
);
1562 void ipoib_del_neighs_by_gid(struct net_device
*dev
, u8
*gid
)
1564 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1565 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1566 struct ipoib_neigh_hash
*htbl
;
1567 unsigned long flags
;
1570 /* remove all neigh connected to a given path or mcast */
1571 spin_lock_irqsave(&priv
->lock
, flags
);
1573 htbl
= rcu_dereference_protected(ntbl
->htbl
,
1574 lockdep_is_held(&priv
->lock
));
1579 for (i
= 0; i
< htbl
->size
; i
++) {
1580 struct ipoib_neigh
*neigh
;
1581 struct ipoib_neigh __rcu
**np
= &htbl
->buckets
[i
];
1583 while ((neigh
= rcu_dereference_protected(*np
,
1584 lockdep_is_held(&priv
->lock
))) != NULL
) {
1585 /* delete neighs belong to this parent */
1586 if (!memcmp(gid
, neigh
->daddr
+ 4, sizeof (union ib_gid
))) {
1587 rcu_assign_pointer(*np
,
1588 rcu_dereference_protected(neigh
->hnext
,
1589 lockdep_is_held(&priv
->lock
)));
1590 /* remove from parent list */
1591 list_del_init(&neigh
->list
);
1592 call_rcu(&neigh
->rcu
, ipoib_neigh_reclaim
);
1600 spin_unlock_irqrestore(&priv
->lock
, flags
);
1603 static void ipoib_flush_neighs(struct ipoib_dev_priv
*priv
)
1605 struct ipoib_neigh_table
*ntbl
= &priv
->ntbl
;
1606 struct ipoib_neigh_hash
*htbl
;
1607 unsigned long flags
;
1608 int i
, wait_flushed
= 0;
1610 init_completion(&priv
->ntbl
.flushed
);
1611 set_bit(IPOIB_NEIGH_TBL_FLUSH
, &priv
->flags
);
1613 spin_lock_irqsave(&priv
->lock
, flags
);
1615 htbl
= rcu_dereference_protected(ntbl
->htbl
,
1616 lockdep_is_held(&priv
->lock
));
1620 wait_flushed
= atomic_read(&priv
->ntbl
.entries
);
1624 for (i
= 0; i
< htbl
->size
; i
++) {
1625 struct ipoib_neigh
*neigh
;
1626 struct ipoib_neigh __rcu
**np
= &htbl
->buckets
[i
];
1628 while ((neigh
= rcu_dereference_protected(*np
,
1629 lockdep_is_held(&priv
->lock
))) != NULL
) {
1630 rcu_assign_pointer(*np
,
1631 rcu_dereference_protected(neigh
->hnext
,
1632 lockdep_is_held(&priv
->lock
)));
1633 /* remove from path/mc list */
1634 list_del_init(&neigh
->list
);
1635 call_rcu(&neigh
->rcu
, ipoib_neigh_reclaim
);
1640 rcu_assign_pointer(ntbl
->htbl
, NULL
);
1641 call_rcu(&htbl
->rcu
, neigh_hash_free_rcu
);
1644 spin_unlock_irqrestore(&priv
->lock
, flags
);
1646 wait_for_completion(&priv
->ntbl
.flushed
);
1649 static void ipoib_neigh_hash_uninit(struct net_device
*dev
)
1651 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1653 ipoib_dbg(priv
, "%s\n", __func__
);
1654 init_completion(&priv
->ntbl
.deleted
);
1656 cancel_delayed_work_sync(&priv
->neigh_reap_task
);
1658 ipoib_flush_neighs(priv
);
1660 wait_for_completion(&priv
->ntbl
.deleted
);
1663 static void ipoib_napi_add(struct net_device
*dev
)
1665 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1667 netif_napi_add(dev
, &priv
->recv_napi
, ipoib_rx_poll
, IPOIB_NUM_WC
);
1668 netif_napi_add(dev
, &priv
->send_napi
, ipoib_tx_poll
, MAX_SEND_CQE
);
1671 static void ipoib_napi_del(struct net_device
*dev
)
1673 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1675 netif_napi_del(&priv
->recv_napi
);
1676 netif_napi_del(&priv
->send_napi
);
1679 static void ipoib_dev_uninit_default(struct net_device
*dev
)
1681 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1683 ipoib_transport_dev_cleanup(dev
);
1685 ipoib_napi_del(dev
);
1687 ipoib_cm_dev_cleanup(dev
);
1689 kfree(priv
->rx_ring
);
1690 vfree(priv
->tx_ring
);
1692 priv
->rx_ring
= NULL
;
1693 priv
->tx_ring
= NULL
;
1696 static int ipoib_dev_init_default(struct net_device
*dev
)
1698 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1700 ipoib_napi_add(dev
);
1702 /* Allocate RX/TX "rings" to hold queued skbs */
1703 priv
->rx_ring
= kcalloc(ipoib_recvq_size
,
1704 sizeof(*priv
->rx_ring
),
1709 priv
->tx_ring
= vzalloc(array_size(ipoib_sendq_size
,
1710 sizeof(*priv
->tx_ring
)));
1711 if (!priv
->tx_ring
) {
1712 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1713 priv
->ca
->name
, ipoib_sendq_size
);
1714 goto out_rx_ring_cleanup
;
1717 /* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
1719 if (ipoib_transport_dev_init(dev
, priv
->ca
)) {
1720 pr_warn("%s: ipoib_transport_dev_init failed\n",
1722 goto out_tx_ring_cleanup
;
1725 /* after qp created set dev address */
1726 priv
->dev
->dev_addr
[1] = (priv
->qp
->qp_num
>> 16) & 0xff;
1727 priv
->dev
->dev_addr
[2] = (priv
->qp
->qp_num
>> 8) & 0xff;
1728 priv
->dev
->dev_addr
[3] = (priv
->qp
->qp_num
) & 0xff;
1732 out_tx_ring_cleanup
:
1733 vfree(priv
->tx_ring
);
1735 out_rx_ring_cleanup
:
1736 kfree(priv
->rx_ring
);
1739 ipoib_napi_del(dev
);
1743 static int ipoib_ioctl(struct net_device
*dev
, struct ifreq
*ifr
,
1746 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1748 if (!priv
->rn_ops
->ndo_eth_ioctl
)
1751 return priv
->rn_ops
->ndo_eth_ioctl(dev
, ifr
, cmd
);
1754 static int ipoib_dev_init(struct net_device
*dev
)
1756 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1762 * the various IPoIB tasks assume they will never race against
1763 * themselves, so always use a single thread workqueue
1765 priv
->wq
= alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM
);
1767 pr_warn("%s: failed to allocate device WQ\n", dev
->name
);
1771 /* create pd, which used both for control and datapath*/
1772 priv
->pd
= ib_alloc_pd(priv
->ca
, 0);
1773 if (IS_ERR(priv
->pd
)) {
1774 pr_warn("%s: failed to allocate PD\n", priv
->ca
->name
);
1778 ret
= priv
->rn_ops
->ndo_init(dev
);
1780 pr_warn("%s failed to init HW resource\n", dev
->name
);
1784 ret
= ipoib_neigh_hash_init(priv
);
1786 pr_warn("%s failed to init neigh hash\n", dev
->name
);
1787 goto out_dev_uninit
;
1790 if (dev
->flags
& IFF_UP
) {
1791 if (ipoib_ib_dev_open(dev
)) {
1792 pr_warn("%s failed to open device\n", dev
->name
);
1794 goto out_hash_uninit
;
1801 ipoib_neigh_hash_uninit(dev
);
1804 ipoib_ib_dev_cleanup(dev
);
1808 ib_dealloc_pd(priv
->pd
);
1814 destroy_workqueue(priv
->wq
);
1823 * This must be called before doing an unregister_netdev on a parent device to
1824 * shutdown the IB event handler.
1826 static void ipoib_parent_unregister_pre(struct net_device
*ndev
)
1828 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
1831 * ipoib_set_mac checks netif_running before pushing work, clearing
1832 * running ensures the it will not add more work.
1835 dev_change_flags(priv
->dev
, priv
->dev
->flags
& ~IFF_UP
, NULL
);
1838 /* ipoib_event() cannot be running once this returns */
1839 ib_unregister_event_handler(&priv
->event_handler
);
1842 * Work on the queue grabs the rtnl lock, so this cannot be done while
1845 flush_workqueue(ipoib_workqueue
);
1848 static void ipoib_set_dev_features(struct ipoib_dev_priv
*priv
)
1850 priv
->hca_caps
= priv
->ca
->attrs
.device_cap_flags
;
1852 if (priv
->hca_caps
& IB_DEVICE_UD_IP_CSUM
) {
1853 priv
->dev
->hw_features
|= NETIF_F_IP_CSUM
| NETIF_F_RXCSUM
;
1855 if (priv
->hca_caps
& IB_DEVICE_UD_TSO
)
1856 priv
->dev
->hw_features
|= NETIF_F_TSO
;
1858 priv
->dev
->features
|= priv
->dev
->hw_features
;
1862 static int ipoib_parent_init(struct net_device
*ndev
)
1864 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
1865 struct ib_port_attr attr
;
1868 result
= ib_query_port(priv
->ca
, priv
->port
, &attr
);
1870 pr_warn("%s: ib_query_port %d failed\n", priv
->ca
->name
,
1874 priv
->max_ib_mtu
= rdma_mtu_from_attr(priv
->ca
, priv
->port
, &attr
);
1876 result
= ib_query_pkey(priv
->ca
, priv
->port
, 0, &priv
->pkey
);
1878 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1879 priv
->ca
->name
, priv
->port
, result
);
1883 result
= rdma_query_gid(priv
->ca
, priv
->port
, 0, &priv
->local_gid
);
1885 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1886 priv
->ca
->name
, priv
->port
, result
);
1889 memcpy(priv
->dev
->dev_addr
+ 4, priv
->local_gid
.raw
,
1890 sizeof(union ib_gid
));
1892 SET_NETDEV_DEV(priv
->dev
, priv
->ca
->dev
.parent
);
1893 priv
->dev
->dev_port
= priv
->port
- 1;
1894 /* Let's set this one too for backwards compatibility. */
1895 priv
->dev
->dev_id
= priv
->port
- 1;
1900 static void ipoib_child_init(struct net_device
*ndev
)
1902 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
1903 struct ipoib_dev_priv
*ppriv
= ipoib_priv(priv
->parent
);
1905 priv
->max_ib_mtu
= ppriv
->max_ib_mtu
;
1906 set_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
);
1907 if (memchr_inv(priv
->dev
->dev_addr
, 0, INFINIBAND_ALEN
))
1908 memcpy(&priv
->local_gid
, priv
->dev
->dev_addr
+ 4,
1909 sizeof(priv
->local_gid
));
1911 memcpy(priv
->dev
->dev_addr
, ppriv
->dev
->dev_addr
,
1913 memcpy(&priv
->local_gid
, &ppriv
->local_gid
,
1914 sizeof(priv
->local_gid
));
1918 static int ipoib_ndo_init(struct net_device
*ndev
)
1920 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
1922 struct rdma_netdev
*rn
= netdev_priv(ndev
);
1925 ipoib_child_init(ndev
);
1927 rc
= ipoib_parent_init(ndev
);
1932 /* MTU will be reset when mcast join happens */
1933 ndev
->mtu
= IPOIB_UD_MTU(priv
->max_ib_mtu
);
1934 priv
->mcast_mtu
= priv
->admin_mtu
= ndev
->mtu
;
1935 rn
->mtu
= priv
->mcast_mtu
;
1936 ndev
->max_mtu
= IPOIB_CM_MTU
;
1938 ndev
->neigh_priv_len
= sizeof(struct ipoib_neigh
);
1941 * Set the full membership bit, so that we join the right
1942 * broadcast group, etc.
1944 priv
->pkey
|= 0x8000;
1946 ndev
->broadcast
[8] = priv
->pkey
>> 8;
1947 ndev
->broadcast
[9] = priv
->pkey
& 0xff;
1948 set_bit(IPOIB_FLAG_DEV_ADDR_SET
, &priv
->flags
);
1950 ipoib_set_dev_features(priv
);
1952 rc
= ipoib_dev_init(ndev
);
1954 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1955 priv
->ca
->name
, priv
->dev
->name
, priv
->port
, rc
);
1960 struct ipoib_dev_priv
*ppriv
= ipoib_priv(priv
->parent
);
1962 dev_hold(priv
->parent
);
1964 down_write(&ppriv
->vlan_rwsem
);
1965 list_add_tail(&priv
->list
, &ppriv
->child_intfs
);
1966 up_write(&ppriv
->vlan_rwsem
);
1972 static void ipoib_ndo_uninit(struct net_device
*dev
)
1974 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1979 * ipoib_remove_one guarantees the children are removed before the
1980 * parent, and that is the only place where a parent can be removed.
1982 WARN_ON(!list_empty(&priv
->child_intfs
));
1985 struct ipoib_dev_priv
*ppriv
= ipoib_priv(priv
->parent
);
1987 down_write(&ppriv
->vlan_rwsem
);
1988 list_del(&priv
->list
);
1989 up_write(&ppriv
->vlan_rwsem
);
1992 ipoib_neigh_hash_uninit(dev
);
1994 ipoib_ib_dev_cleanup(dev
);
1996 /* no more works over the priv->wq */
1998 /* See ipoib_mcast_carrier_on_task() */
1999 WARN_ON(test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
));
2000 flush_workqueue(priv
->wq
);
2001 destroy_workqueue(priv
->wq
);
2006 dev_put(priv
->parent
);
2009 static int ipoib_set_vf_link_state(struct net_device
*dev
, int vf
, int link_state
)
2011 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2013 return ib_set_vf_link_state(priv
->ca
, vf
, priv
->port
, link_state
);
2016 static int ipoib_get_vf_config(struct net_device
*dev
, int vf
,
2017 struct ifla_vf_info
*ivf
)
2019 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2022 err
= ib_get_vf_config(priv
->ca
, vf
, priv
->port
, ivf
);
2027 memcpy(ivf
->mac
, dev
->dev_addr
, dev
->addr_len
);
2032 static int ipoib_set_vf_guid(struct net_device
*dev
, int vf
, u64 guid
, int type
)
2034 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2036 if (type
!= IFLA_VF_IB_NODE_GUID
&& type
!= IFLA_VF_IB_PORT_GUID
)
2039 return ib_set_vf_guid(priv
->ca
, vf
, priv
->port
, guid
, type
);
2042 static int ipoib_get_vf_guid(struct net_device
*dev
, int vf
,
2043 struct ifla_vf_guid
*node_guid
,
2044 struct ifla_vf_guid
*port_guid
)
2046 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2048 return ib_get_vf_guid(priv
->ca
, vf
, priv
->port
, node_guid
, port_guid
);
2051 static int ipoib_get_vf_stats(struct net_device
*dev
, int vf
,
2052 struct ifla_vf_stats
*vf_stats
)
2054 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2056 return ib_get_vf_stats(priv
->ca
, vf
, priv
->port
, vf_stats
);
2059 static const struct header_ops ipoib_header_ops
= {
2060 .create
= ipoib_hard_header
,
2063 static const struct net_device_ops ipoib_netdev_ops_pf
= {
2064 .ndo_init
= ipoib_ndo_init
,
2065 .ndo_uninit
= ipoib_ndo_uninit
,
2066 .ndo_open
= ipoib_open
,
2067 .ndo_stop
= ipoib_stop
,
2068 .ndo_change_mtu
= ipoib_change_mtu
,
2069 .ndo_fix_features
= ipoib_fix_features
,
2070 .ndo_start_xmit
= ipoib_start_xmit
,
2071 .ndo_tx_timeout
= ipoib_timeout
,
2072 .ndo_set_rx_mode
= ipoib_set_mcast_list
,
2073 .ndo_get_iflink
= ipoib_get_iflink
,
2074 .ndo_set_vf_link_state
= ipoib_set_vf_link_state
,
2075 .ndo_get_vf_config
= ipoib_get_vf_config
,
2076 .ndo_get_vf_stats
= ipoib_get_vf_stats
,
2077 .ndo_get_vf_guid
= ipoib_get_vf_guid
,
2078 .ndo_set_vf_guid
= ipoib_set_vf_guid
,
2079 .ndo_set_mac_address
= ipoib_set_mac
,
2080 .ndo_get_stats64
= ipoib_get_stats
,
2081 .ndo_eth_ioctl
= ipoib_ioctl
,
2084 static const struct net_device_ops ipoib_netdev_ops_vf
= {
2085 .ndo_init
= ipoib_ndo_init
,
2086 .ndo_uninit
= ipoib_ndo_uninit
,
2087 .ndo_open
= ipoib_open
,
2088 .ndo_stop
= ipoib_stop
,
2089 .ndo_change_mtu
= ipoib_change_mtu
,
2090 .ndo_fix_features
= ipoib_fix_features
,
2091 .ndo_start_xmit
= ipoib_start_xmit
,
2092 .ndo_tx_timeout
= ipoib_timeout
,
2093 .ndo_set_rx_mode
= ipoib_set_mcast_list
,
2094 .ndo_get_iflink
= ipoib_get_iflink
,
2095 .ndo_get_stats64
= ipoib_get_stats
,
2096 .ndo_eth_ioctl
= ipoib_ioctl
,
2099 static const struct net_device_ops ipoib_netdev_default_pf
= {
2100 .ndo_init
= ipoib_dev_init_default
,
2101 .ndo_uninit
= ipoib_dev_uninit_default
,
2102 .ndo_open
= ipoib_ib_dev_open_default
,
2103 .ndo_stop
= ipoib_ib_dev_stop_default
,
2106 void ipoib_setup_common(struct net_device
*dev
)
2108 dev
->header_ops
= &ipoib_header_ops
;
2109 dev
->netdev_ops
= &ipoib_netdev_default_pf
;
2111 ipoib_set_ethtool_ops(dev
);
2113 dev
->watchdog_timeo
= HZ
;
2115 dev
->flags
|= IFF_BROADCAST
| IFF_MULTICAST
;
2117 dev
->hard_header_len
= IPOIB_HARD_LEN
;
2118 dev
->addr_len
= INFINIBAND_ALEN
;
2119 dev
->type
= ARPHRD_INFINIBAND
;
2120 dev
->tx_queue_len
= ipoib_sendq_size
* 2;
2121 dev
->features
= (NETIF_F_VLAN_CHALLENGED
|
2123 netif_keep_dst(dev
);
2125 memcpy(dev
->broadcast
, ipv4_bcast_addr
, INFINIBAND_ALEN
);
2128 * unregister_netdev always frees the netdev, we use this mode
2129 * consistently to unify all the various unregister paths, including
2130 * those connected to rtnl_link_ops which require it.
2132 dev
->needs_free_netdev
= true;
2135 static void ipoib_build_priv(struct net_device
*dev
)
2137 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2140 spin_lock_init(&priv
->lock
);
2141 init_rwsem(&priv
->vlan_rwsem
);
2142 mutex_init(&priv
->mcast_mutex
);
2144 INIT_LIST_HEAD(&priv
->path_list
);
2145 INIT_LIST_HEAD(&priv
->child_intfs
);
2146 INIT_LIST_HEAD(&priv
->dead_ahs
);
2147 INIT_LIST_HEAD(&priv
->multicast_list
);
2149 INIT_DELAYED_WORK(&priv
->mcast_task
, ipoib_mcast_join_task
);
2150 INIT_WORK(&priv
->carrier_on_task
, ipoib_mcast_carrier_on_task
);
2151 INIT_WORK(&priv
->flush_light
, ipoib_ib_dev_flush_light
);
2152 INIT_WORK(&priv
->flush_normal
, ipoib_ib_dev_flush_normal
);
2153 INIT_WORK(&priv
->flush_heavy
, ipoib_ib_dev_flush_heavy
);
2154 INIT_WORK(&priv
->restart_task
, ipoib_mcast_restart_task
);
2155 INIT_DELAYED_WORK(&priv
->ah_reap_task
, ipoib_reap_ah
);
2156 INIT_DELAYED_WORK(&priv
->neigh_reap_task
, ipoib_reap_neigh
);
2159 static struct net_device
*ipoib_alloc_netdev(struct ib_device
*hca
, u32 port
,
2162 struct net_device
*dev
;
2164 dev
= rdma_alloc_netdev(hca
, port
, RDMA_NETDEV_IPOIB
, name
,
2165 NET_NAME_UNKNOWN
, ipoib_setup_common
);
2166 if (!IS_ERR(dev
) || PTR_ERR(dev
) != -EOPNOTSUPP
)
2169 dev
= alloc_netdev(sizeof(struct rdma_netdev
), name
, NET_NAME_UNKNOWN
,
2170 ipoib_setup_common
);
2172 return ERR_PTR(-ENOMEM
);
2176 int ipoib_intf_init(struct ib_device
*hca
, u32 port
, const char *name
,
2177 struct net_device
*dev
)
2179 struct rdma_netdev
*rn
= netdev_priv(dev
);
2180 struct ipoib_dev_priv
*priv
;
2183 priv
= kzalloc(sizeof(*priv
), GFP_KERNEL
);
2190 rc
= rdma_init_netdev(hca
, port
, RDMA_NETDEV_IPOIB
, name
,
2191 NET_NAME_UNKNOWN
, ipoib_setup_common
, dev
);
2193 if (rc
!= -EOPNOTSUPP
)
2196 rn
->send
= ipoib_send
;
2197 rn
->attach_mcast
= ipoib_mcast_attach
;
2198 rn
->detach_mcast
= ipoib_mcast_detach
;
2202 priv
->rn_ops
= dev
->netdev_ops
;
2204 if (hca
->attrs
.device_cap_flags
& IB_DEVICE_VIRTUAL_FUNCTION
)
2205 dev
->netdev_ops
= &ipoib_netdev_ops_vf
;
2207 dev
->netdev_ops
= &ipoib_netdev_ops_pf
;
2209 rn
->clnt_priv
= priv
;
2211 * Only the child register_netdev flows can handle priv_destructor
2212 * being set, so we force it to NULL here and handle manually until it
2213 * is safe to turn on.
2215 priv
->next_priv_destructor
= dev
->priv_destructor
;
2216 dev
->priv_destructor
= NULL
;
2218 ipoib_build_priv(dev
);
2227 struct net_device
*ipoib_intf_alloc(struct ib_device
*hca
, u32 port
,
2230 struct net_device
*dev
;
2233 dev
= ipoib_alloc_netdev(hca
, port
, name
);
2237 rc
= ipoib_intf_init(hca
, port
, name
, dev
);
2244 * Upon success the caller must ensure ipoib_intf_free is called or
2245 * register_netdevice succeed'd and priv_destructor is set to
2251 void ipoib_intf_free(struct net_device
*dev
)
2253 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2254 struct rdma_netdev
*rn
= netdev_priv(dev
);
2256 dev
->priv_destructor
= priv
->next_priv_destructor
;
2257 if (dev
->priv_destructor
)
2258 dev
->priv_destructor(dev
);
2261 * There are some error flows around register_netdev failing that may
2262 * attempt to call priv_destructor twice, prevent that from happening.
2264 dev
->priv_destructor
= NULL
;
2266 /* unregister/destroy is very complicated. Make bugs more obvious. */
2267 rn
->clnt_priv
= NULL
;
2272 static ssize_t
pkey_show(struct device
*dev
, struct device_attribute
*attr
,
2275 struct net_device
*ndev
= to_net_dev(dev
);
2276 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
2278 return sysfs_emit(buf
, "0x%04x\n", priv
->pkey
);
2280 static DEVICE_ATTR_RO(pkey
);
2282 static ssize_t
umcast_show(struct device
*dev
, struct device_attribute
*attr
,
2285 struct net_device
*ndev
= to_net_dev(dev
);
2286 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
2288 return sysfs_emit(buf
, "%d\n",
2289 test_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
));
2292 void ipoib_set_umcast(struct net_device
*ndev
, int umcast_val
)
2294 struct ipoib_dev_priv
*priv
= ipoib_priv(ndev
);
2296 if (umcast_val
> 0) {
2297 set_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
2298 ipoib_warn(priv
, "ignoring multicast groups joined directly "
2301 clear_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
2304 static ssize_t
umcast_store(struct device
*dev
, struct device_attribute
*attr
,
2305 const char *buf
, size_t count
)
2307 unsigned long umcast_val
= simple_strtoul(buf
, NULL
, 0);
2309 ipoib_set_umcast(to_net_dev(dev
), umcast_val
);
2313 static DEVICE_ATTR_RW(umcast
);
2315 int ipoib_add_umcast_attr(struct net_device
*dev
)
2317 return device_create_file(&dev
->dev
, &dev_attr_umcast
);
2320 static void set_base_guid(struct ipoib_dev_priv
*priv
, union ib_gid
*gid
)
2322 struct ipoib_dev_priv
*child_priv
;
2323 struct net_device
*netdev
= priv
->dev
;
2325 netif_addr_lock_bh(netdev
);
2327 memcpy(&priv
->local_gid
.global
.interface_id
,
2328 &gid
->global
.interface_id
,
2329 sizeof(gid
->global
.interface_id
));
2330 memcpy(netdev
->dev_addr
+ 4, &priv
->local_gid
, sizeof(priv
->local_gid
));
2331 clear_bit(IPOIB_FLAG_DEV_ADDR_SET
, &priv
->flags
);
2333 netif_addr_unlock_bh(netdev
);
2335 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
2336 down_read(&priv
->vlan_rwsem
);
2337 list_for_each_entry(child_priv
, &priv
->child_intfs
, list
)
2338 set_base_guid(child_priv
, gid
);
2339 up_read(&priv
->vlan_rwsem
);
2343 static int ipoib_check_lladdr(struct net_device
*dev
,
2344 struct sockaddr_storage
*ss
)
2346 union ib_gid
*gid
= (union ib_gid
*)(ss
->__data
+ 4);
2349 netif_addr_lock_bh(dev
);
2351 /* Make sure the QPN, reserved and subnet prefix match the current
2352 * lladdr, it also makes sure the lladdr is unicast.
2354 if (memcmp(dev
->dev_addr
, ss
->__data
,
2355 4 + sizeof(gid
->global
.subnet_prefix
)) ||
2356 gid
->global
.interface_id
== 0)
2359 netif_addr_unlock_bh(dev
);
2364 static int ipoib_set_mac(struct net_device
*dev
, void *addr
)
2366 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
2367 struct sockaddr_storage
*ss
= addr
;
2370 if (!(dev
->priv_flags
& IFF_LIVE_ADDR_CHANGE
) && netif_running(dev
))
2373 ret
= ipoib_check_lladdr(dev
, ss
);
2377 set_base_guid(priv
, (union ib_gid
*)(ss
->__data
+ 4));
2379 queue_work(ipoib_workqueue
, &priv
->flush_light
);
2384 static ssize_t
create_child_store(struct device
*dev
,
2385 struct device_attribute
*attr
,
2386 const char *buf
, size_t count
)
2391 if (sscanf(buf
, "%i", &pkey
) != 1)
2394 if (pkey
<= 0 || pkey
> 0xffff || pkey
== 0x8000)
2397 ret
= ipoib_vlan_add(to_net_dev(dev
), pkey
);
2399 return ret
? ret
: count
;
2401 static DEVICE_ATTR_WO(create_child
);
2403 static ssize_t
delete_child_store(struct device
*dev
,
2404 struct device_attribute
*attr
,
2405 const char *buf
, size_t count
)
2410 if (sscanf(buf
, "%i", &pkey
) != 1)
2413 if (pkey
< 0 || pkey
> 0xffff)
2416 ret
= ipoib_vlan_delete(to_net_dev(dev
), pkey
);
2418 return ret
? ret
: count
;
2421 static DEVICE_ATTR_WO(delete_child
);
2423 int ipoib_add_pkey_attr(struct net_device
*dev
)
2425 return device_create_file(&dev
->dev
, &dev_attr_pkey
);
2429 * We erroneously exposed the iface's port number in the dev_id
2430 * sysfs field long after dev_port was introduced for that purpose[1],
2431 * and we need to stop everyone from relying on that.
2432 * Let's overload the shower routine for the dev_id file here
2433 * to gently bring the issue up.
2435 * [1] https://www.spinics.net/lists/netdev/msg272123.html
2437 static ssize_t
dev_id_show(struct device
*dev
,
2438 struct device_attribute
*attr
, char *buf
)
2440 struct net_device
*ndev
= to_net_dev(dev
);
2443 * ndev->dev_port will be equal to 0 in old kernel prior to commit
2444 * 9b8b2a323008 ("IB/ipoib: Use dev_port to expose network interface
2445 * port numbers") Zero was chosen as special case for user space
2446 * applications to fallback and query dev_id to check if it has
2447 * different value or not.
2449 * Don't print warning in such scenario.
2451 * https://github.com/systemd/systemd/blob/master/src/udev/udev-builtin-net_id.c#L358
2453 if (ndev
->dev_port
&& ndev
->dev_id
== ndev
->dev_port
)
2454 netdev_info_once(ndev
,
2455 "\"%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",
2458 return sysfs_emit(buf
, "%#x\n", ndev
->dev_id
);
2460 static DEVICE_ATTR_RO(dev_id
);
2462 static int ipoib_intercept_dev_id_attr(struct net_device
*dev
)
2464 device_remove_file(&dev
->dev
, &dev_attr_dev_id
);
2465 return device_create_file(&dev
->dev
, &dev_attr_dev_id
);
2468 static struct net_device
*ipoib_add_port(const char *format
,
2469 struct ib_device
*hca
, u32 port
)
2471 struct rtnl_link_ops
*ops
= ipoib_get_link_ops();
2472 struct rdma_netdev_alloc_params params
;
2473 struct ipoib_dev_priv
*priv
;
2474 struct net_device
*ndev
;
2477 ndev
= ipoib_intf_alloc(hca
, port
, format
);
2479 pr_warn("%s, %d: ipoib_intf_alloc failed %ld\n", hca
->name
, port
,
2483 priv
= ipoib_priv(ndev
);
2485 INIT_IB_EVENT_HANDLER(&priv
->event_handler
,
2486 priv
->ca
, ipoib_event
);
2487 ib_register_event_handler(&priv
->event_handler
);
2489 /* call event handler to ensure pkey in sync */
2490 queue_work(ipoib_workqueue
, &priv
->flush_heavy
);
2492 ndev
->rtnl_link_ops
= ipoib_get_link_ops();
2494 result
= register_netdev(ndev
);
2496 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2497 hca
->name
, port
, result
);
2499 ipoib_parent_unregister_pre(ndev
);
2500 ipoib_intf_free(ndev
);
2503 return ERR_PTR(result
);
2506 if (hca
->ops
.rdma_netdev_get_params
) {
2507 int rc
= hca
->ops
.rdma_netdev_get_params(hca
, port
,
2511 if (!rc
&& ops
->priv_size
< params
.sizeof_priv
)
2512 ops
->priv_size
= params
.sizeof_priv
;
2515 * We cannot set priv_destructor before register_netdev because we
2516 * need priv to be always valid during the error flow to execute
2517 * ipoib_parent_unregister_pre(). Instead handle it manually and only
2518 * enter priv_destructor mode once we are completely registered.
2520 ndev
->priv_destructor
= ipoib_intf_free
;
2522 if (ipoib_intercept_dev_id_attr(ndev
))
2524 if (ipoib_cm_add_mode_attr(ndev
))
2526 if (ipoib_add_pkey_attr(ndev
))
2528 if (ipoib_add_umcast_attr(ndev
))
2530 if (device_create_file(&ndev
->dev
, &dev_attr_create_child
))
2532 if (device_create_file(&ndev
->dev
, &dev_attr_delete_child
))
2538 ipoib_parent_unregister_pre(ndev
);
2539 unregister_netdev(ndev
);
2540 return ERR_PTR(-ENOMEM
);
2543 static int ipoib_add_one(struct ib_device
*device
)
2545 struct list_head
*dev_list
;
2546 struct net_device
*dev
;
2547 struct ipoib_dev_priv
*priv
;
2551 dev_list
= kmalloc(sizeof(*dev_list
), GFP_KERNEL
);
2555 INIT_LIST_HEAD(dev_list
);
2557 rdma_for_each_port (device
, p
) {
2558 if (!rdma_protocol_ib(device
, p
))
2560 dev
= ipoib_add_port("ib%d", device
, p
);
2562 priv
= ipoib_priv(dev
);
2563 list_add_tail(&priv
->list
, dev_list
);
2573 ib_set_client_data(device
, &ipoib_client
, dev_list
);
2577 static void ipoib_remove_one(struct ib_device
*device
, void *client_data
)
2579 struct ipoib_dev_priv
*priv
, *tmp
, *cpriv
, *tcpriv
;
2580 struct list_head
*dev_list
= client_data
;
2582 list_for_each_entry_safe(priv
, tmp
, dev_list
, list
) {
2584 ipoib_parent_unregister_pre(priv
->dev
);
2588 list_for_each_entry_safe(cpriv
, tcpriv
, &priv
->child_intfs
,
2590 unregister_netdevice_queue(cpriv
->dev
, &head
);
2591 unregister_netdevice_queue(priv
->dev
, &head
);
2592 unregister_netdevice_many(&head
);
2600 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2601 static struct notifier_block ipoib_netdev_notifier
= {
2602 .notifier_call
= ipoib_netdev_event
,
2606 static int __init
ipoib_init_module(void)
2610 ipoib_recvq_size
= roundup_pow_of_two(ipoib_recvq_size
);
2611 ipoib_recvq_size
= min(ipoib_recvq_size
, IPOIB_MAX_QUEUE_SIZE
);
2612 ipoib_recvq_size
= max(ipoib_recvq_size
, IPOIB_MIN_QUEUE_SIZE
);
2614 ipoib_sendq_size
= roundup_pow_of_two(ipoib_sendq_size
);
2615 ipoib_sendq_size
= min(ipoib_sendq_size
, IPOIB_MAX_QUEUE_SIZE
);
2616 ipoib_sendq_size
= max3(ipoib_sendq_size
, 2 * MAX_SEND_CQE
, IPOIB_MIN_QUEUE_SIZE
);
2617 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2618 ipoib_max_conn_qp
= min(ipoib_max_conn_qp
, IPOIB_CM_MAX_CONN_QP
);
2619 ipoib_max_conn_qp
= max(ipoib_max_conn_qp
, 0);
2623 * When copying small received packets, we only copy from the
2624 * linear data part of the SKB, so we rely on this condition.
2626 BUILD_BUG_ON(IPOIB_CM_COPYBREAK
> IPOIB_CM_HEAD_SIZE
);
2628 ipoib_register_debugfs();
2631 * We create a global workqueue here that is used for all flush
2632 * operations. However, if you attempt to flush a workqueue
2633 * from a task on that same workqueue, it deadlocks the system.
2634 * We want to be able to flush the tasks associated with a
2635 * specific net device, so we also create a workqueue for each
2636 * netdevice. We queue up the tasks for that device only on
2637 * its private workqueue, and we only queue up flush events
2638 * on our global flush workqueue. This avoids the deadlocks.
2640 ipoib_workqueue
= alloc_ordered_workqueue("ipoib_flush", 0);
2641 if (!ipoib_workqueue
) {
2646 ib_sa_register_client(&ipoib_sa_client
);
2648 ret
= ib_register_client(&ipoib_client
);
2652 ret
= ipoib_netlink_init();
2656 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2657 register_netdevice_notifier(&ipoib_netdev_notifier
);
2662 ib_unregister_client(&ipoib_client
);
2665 ib_sa_unregister_client(&ipoib_sa_client
);
2666 destroy_workqueue(ipoib_workqueue
);
2669 ipoib_unregister_debugfs();
2674 static void __exit
ipoib_cleanup_module(void)
2676 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2677 unregister_netdevice_notifier(&ipoib_netdev_notifier
);
2679 ipoib_netlink_fini();
2680 ib_unregister_client(&ipoib_client
);
2681 ib_sa_unregister_client(&ipoib_sa_client
);
2682 ipoib_unregister_debugfs();
2683 destroy_workqueue(ipoib_workqueue
);
2686 module_init(ipoib_init_module
);
2687 module_exit(ipoib_cleanup_module
);