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>
54 #include <linux/pci.h>
56 #define DRV_VERSION "1.0.0"
58 const char ipoib_driver_version[] = DRV_VERSION;
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
62 MODULE_LICENSE("Dual BSD/GPL");
63 MODULE_VERSION(DRV_VERSION);
65 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
66 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
68 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
69 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
70 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
71 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
73 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
74 int ipoib_debug_level;
76 module_param_named(debug_level, ipoib_debug_level, int, 0644);
77 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
80 struct ipoib_path_iter {
81 struct net_device *dev;
82 struct ipoib_path path;
85 static const u8 ipv4_bcast_addr[] = {
86 0x00, 0xff, 0xff, 0xff,
87 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
88 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
91 struct workqueue_struct *ipoib_workqueue;
93 struct ib_sa_client ipoib_sa_client;
95 static void ipoib_add_one(struct ib_device *device);
96 static void ipoib_remove_one(struct ib_device *device, void *client_data);
97 static void ipoib_neigh_reclaim(struct rcu_head *rp);
98 static struct net_device *ipoib_get_net_dev_by_params(
99 struct ib_device *dev, u8 port, u16 pkey,
100 const union ib_gid *gid, const struct sockaddr *addr,
102 static int ipoib_set_mac(struct net_device *dev, void *addr);
104 static struct ib_client ipoib_client = {
106 .add = ipoib_add_one,
107 .remove = ipoib_remove_one,
108 .get_net_dev_by_params = ipoib_get_net_dev_by_params,
111 int ipoib_open(struct net_device *dev)
113 struct ipoib_dev_priv *priv = netdev_priv(dev);
115 ipoib_dbg(priv, "bringing up interface\n");
117 netif_carrier_off(dev);
119 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
121 priv->sm_fullmember_sendonly_support = false;
123 if (ipoib_ib_dev_open(dev)) {
124 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
129 ipoib_ib_dev_up(dev);
131 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
132 struct ipoib_dev_priv *cpriv;
134 /* Bring up any child interfaces too */
135 down_read(&priv->vlan_rwsem);
136 list_for_each_entry(cpriv, &priv->child_intfs, list) {
139 flags = cpriv->dev->flags;
143 dev_change_flags(cpriv->dev, flags | IFF_UP);
145 up_read(&priv->vlan_rwsem);
148 netif_start_queue(dev);
153 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
158 static int ipoib_stop(struct net_device *dev)
160 struct ipoib_dev_priv *priv = netdev_priv(dev);
162 ipoib_dbg(priv, "stopping interface\n");
164 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
166 netif_stop_queue(dev);
168 ipoib_ib_dev_down(dev);
169 ipoib_ib_dev_stop(dev);
171 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
172 struct ipoib_dev_priv *cpriv;
174 /* Bring down any child interfaces too */
175 down_read(&priv->vlan_rwsem);
176 list_for_each_entry(cpriv, &priv->child_intfs, list) {
179 flags = cpriv->dev->flags;
180 if (!(flags & IFF_UP))
183 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
185 up_read(&priv->vlan_rwsem);
191 static void ipoib_uninit(struct net_device *dev)
193 ipoib_dev_cleanup(dev);
196 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
198 struct ipoib_dev_priv *priv = netdev_priv(dev);
200 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
201 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
206 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
208 struct ipoib_dev_priv *priv = netdev_priv(dev);
210 /* dev->mtu > 2K ==> connected mode */
211 if (ipoib_cm_admin_enabled(dev)) {
212 if (new_mtu > ipoib_cm_max_mtu(dev))
215 if (new_mtu > priv->mcast_mtu)
216 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
223 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
226 priv->admin_mtu = new_mtu;
228 if (priv->mcast_mtu < priv->admin_mtu)
229 ipoib_dbg(priv, "MTU must be smaller than the underlying "
230 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
232 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
237 /* Called with an RCU read lock taken */
238 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
239 struct net_device *dev)
241 struct net *net = dev_net(dev);
242 struct in_device *in_dev;
243 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
244 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
247 switch (addr->sa_family) {
249 in_dev = in_dev_get(dev);
253 ret_addr = inet_confirm_addr(net, in_dev, 0,
254 addr_in->sin_addr.s_addr,
262 if (IS_ENABLED(CONFIG_IPV6) &&
263 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
272 * Find the master net_device on top of the given net_device.
273 * @dev: base IPoIB net_device
275 * Returns the master net_device with a reference held, or the same net_device
276 * if no master exists.
278 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
280 struct net_device *master;
283 master = netdev_master_upper_dev_get_rcu(dev);
295 struct ipoib_walk_data {
296 const struct sockaddr *addr;
297 struct net_device *result;
300 static int ipoib_upper_walk(struct net_device *upper, void *_data)
302 struct ipoib_walk_data *data = _data;
305 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
307 data->result = upper;
315 * Find a net_device matching the given address, which is an upper device of
316 * the given net_device.
317 * @addr: IP address to look for.
318 * @dev: base IPoIB net_device
320 * If found, returns the net_device with a reference held. Otherwise return
323 static struct net_device *ipoib_get_net_dev_match_addr(
324 const struct sockaddr *addr, struct net_device *dev)
326 struct ipoib_walk_data data = {
331 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
337 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
343 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
344 * pkey_index and address, if one exists.
346 * @found_net_dev: contains a matching net_device if the return value >= 1,
347 * with a reference held. */
348 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
349 const union ib_gid *gid,
351 const struct sockaddr *addr,
353 struct net_device **found_net_dev)
355 struct ipoib_dev_priv *child_priv;
356 struct net_device *net_dev = NULL;
359 if (priv->pkey_index == pkey_index &&
360 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
362 net_dev = ipoib_get_master_net_dev(priv->dev);
364 /* Verify the net_device matches the IP address, as
365 * IPoIB child devices currently share a GID. */
366 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
370 *found_net_dev = net_dev;
377 /* Check child interfaces */
378 down_read_nested(&priv->vlan_rwsem, nesting);
379 list_for_each_entry(child_priv, &priv->child_intfs, list) {
380 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
387 up_read(&priv->vlan_rwsem);
392 /* Returns the number of matching net_devs found (between 0 and 2). Also
393 * return the matching net_device in the @net_dev parameter, holding a
394 * reference to the net_device, if the number of matches >= 1 */
395 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
397 const union ib_gid *gid,
398 const struct sockaddr *addr,
399 struct net_device **net_dev)
401 struct ipoib_dev_priv *priv;
406 list_for_each_entry(priv, dev_list, list) {
407 if (priv->port != port)
410 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
419 static struct net_device *ipoib_get_net_dev_by_params(
420 struct ib_device *dev, u8 port, u16 pkey,
421 const union ib_gid *gid, const struct sockaddr *addr,
424 struct net_device *net_dev;
425 struct list_head *dev_list = client_data;
430 if (!rdma_protocol_ib(dev, port))
433 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
440 /* See if we can find a unique device matching the L2 parameters */
441 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
442 gid, NULL, &net_dev);
453 /* Couldn't find a unique device with L2 parameters only. Use L3
454 * address to uniquely match the net device */
455 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
456 gid, addr, &net_dev);
461 dev_warn_ratelimited(&dev->dev,
462 "duplicate IP address detected\n");
469 int ipoib_set_mode(struct net_device *dev, const char *buf)
471 struct ipoib_dev_priv *priv = netdev_priv(dev);
473 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
474 !strcmp(buf, "connected\n")) ||
475 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
476 !strcmp(buf, "datagram\n"))) {
480 /* flush paths if we switch modes so that connections are restarted */
481 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
482 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
483 ipoib_warn(priv, "enabling connected mode "
484 "will cause multicast packet drops\n");
485 netdev_update_features(dev);
486 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
488 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
490 ipoib_flush_paths(dev);
491 return (!rtnl_trylock()) ? -EBUSY : 0;
494 if (!strcmp(buf, "datagram\n")) {
495 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
496 netdev_update_features(dev);
497 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
499 ipoib_flush_paths(dev);
500 return (!rtnl_trylock()) ? -EBUSY : 0;
506 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
508 struct ipoib_dev_priv *priv = netdev_priv(dev);
509 struct rb_node *n = priv->path_tree.rb_node;
510 struct ipoib_path *path;
514 path = rb_entry(n, struct ipoib_path, rb_node);
516 ret = memcmp(gid, path->pathrec.dgid.raw,
517 sizeof (union ib_gid));
530 static int __path_add(struct net_device *dev, struct ipoib_path *path)
532 struct ipoib_dev_priv *priv = netdev_priv(dev);
533 struct rb_node **n = &priv->path_tree.rb_node;
534 struct rb_node *pn = NULL;
535 struct ipoib_path *tpath;
540 tpath = rb_entry(pn, struct ipoib_path, rb_node);
542 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
543 sizeof (union ib_gid));
552 rb_link_node(&path->rb_node, pn, n);
553 rb_insert_color(&path->rb_node, &priv->path_tree);
555 list_add_tail(&path->list, &priv->path_list);
560 static void path_free(struct net_device *dev, struct ipoib_path *path)
564 while ((skb = __skb_dequeue(&path->queue)))
565 dev_kfree_skb_irq(skb);
567 ipoib_dbg(netdev_priv(dev), "path_free\n");
569 /* remove all neigh connected to this path */
570 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
573 ipoib_put_ah(path->ah);
578 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
580 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
582 struct ipoib_path_iter *iter;
584 iter = kmalloc(sizeof *iter, GFP_KERNEL);
589 memset(iter->path.pathrec.dgid.raw, 0, 16);
591 if (ipoib_path_iter_next(iter)) {
599 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
601 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
603 struct ipoib_path *path;
606 spin_lock_irq(&priv->lock);
608 n = rb_first(&priv->path_tree);
611 path = rb_entry(n, struct ipoib_path, rb_node);
613 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
614 sizeof (union ib_gid)) < 0) {
623 spin_unlock_irq(&priv->lock);
628 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
629 struct ipoib_path *path)
634 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
636 void ipoib_mark_paths_invalid(struct net_device *dev)
638 struct ipoib_dev_priv *priv = netdev_priv(dev);
639 struct ipoib_path *path, *tp;
641 spin_lock_irq(&priv->lock);
643 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
644 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
645 be16_to_cpu(path->pathrec.dlid),
646 path->pathrec.dgid.raw);
650 spin_unlock_irq(&priv->lock);
653 struct classport_info_context {
654 struct ipoib_dev_priv *priv;
655 struct completion done;
656 struct ib_sa_query *sa_query;
659 static void classport_info_query_cb(int status, struct ib_class_port_info *rec,
662 struct classport_info_context *cb_ctx = context;
663 struct ipoib_dev_priv *priv;
669 if (status || !rec) {
670 pr_debug("device: %s failed query classport_info status: %d\n",
671 priv->dev->name, status);
672 /* keeps the default, will try next mcast_restart */
673 priv->sm_fullmember_sendonly_support = false;
677 if (ib_get_cpi_capmask2(rec) &
678 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) {
679 pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n",
681 priv->sm_fullmember_sendonly_support = true;
683 pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n",
685 priv->sm_fullmember_sendonly_support = false;
689 complete(&cb_ctx->done);
692 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv)
694 struct classport_info_context *callback_context;
697 callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL);
698 if (!callback_context)
701 callback_context->priv = priv;
702 init_completion(&callback_context->done);
704 ret = ib_sa_classport_info_rec_query(&ipoib_sa_client,
705 priv->ca, priv->port, 3000,
707 classport_info_query_cb,
709 &callback_context->sa_query);
711 pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n",
712 priv->dev->name, ret);
713 kfree(callback_context);
717 /* waiting for the callback to finish before returnning */
718 wait_for_completion(&callback_context->done);
719 kfree(callback_context);
724 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
726 struct ipoib_pseudo_header *phdr;
728 phdr = (struct ipoib_pseudo_header *)skb_push(skb, sizeof(*phdr));
729 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
732 void ipoib_flush_paths(struct net_device *dev)
734 struct ipoib_dev_priv *priv = netdev_priv(dev);
735 struct ipoib_path *path, *tp;
736 LIST_HEAD(remove_list);
739 netif_tx_lock_bh(dev);
740 spin_lock_irqsave(&priv->lock, flags);
742 list_splice_init(&priv->path_list, &remove_list);
744 list_for_each_entry(path, &remove_list, list)
745 rb_erase(&path->rb_node, &priv->path_tree);
747 list_for_each_entry_safe(path, tp, &remove_list, list) {
749 ib_sa_cancel_query(path->query_id, path->query);
750 spin_unlock_irqrestore(&priv->lock, flags);
751 netif_tx_unlock_bh(dev);
752 wait_for_completion(&path->done);
753 path_free(dev, path);
754 netif_tx_lock_bh(dev);
755 spin_lock_irqsave(&priv->lock, flags);
758 spin_unlock_irqrestore(&priv->lock, flags);
759 netif_tx_unlock_bh(dev);
762 static void path_rec_completion(int status,
763 struct ib_sa_path_rec *pathrec,
766 struct ipoib_path *path = path_ptr;
767 struct net_device *dev = path->dev;
768 struct ipoib_dev_priv *priv = netdev_priv(dev);
769 struct ipoib_ah *ah = NULL;
770 struct ipoib_ah *old_ah = NULL;
771 struct ipoib_neigh *neigh, *tn;
772 struct sk_buff_head skqueue;
777 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
778 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
780 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
781 status, path->pathrec.dgid.raw);
783 skb_queue_head_init(&skqueue);
786 struct ib_ah_attr av;
788 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
789 ah = ipoib_create_ah(dev, priv->pd, &av);
792 spin_lock_irqsave(&priv->lock, flags);
794 if (!IS_ERR_OR_NULL(ah)) {
795 path->pathrec = *pathrec;
800 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
801 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
803 while ((skb = __skb_dequeue(&path->queue)))
804 __skb_queue_tail(&skqueue, skb);
806 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
808 WARN_ON(neigh->ah != old_ah);
810 * Dropping the ah reference inside
811 * priv->lock is safe here, because we
812 * will hold one more reference from
813 * the original value of path->ah (ie
816 ipoib_put_ah(neigh->ah);
818 kref_get(&path->ah->ref);
819 neigh->ah = path->ah;
821 if (ipoib_cm_enabled(dev, neigh->daddr)) {
822 if (!ipoib_cm_get(neigh))
823 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
826 if (!ipoib_cm_get(neigh)) {
827 ipoib_neigh_free(neigh);
832 while ((skb = __skb_dequeue(&neigh->queue)))
833 __skb_queue_tail(&skqueue, skb);
839 complete(&path->done);
841 spin_unlock_irqrestore(&priv->lock, flags);
843 if (IS_ERR_OR_NULL(ah))
844 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
847 ipoib_put_ah(old_ah);
849 while ((skb = __skb_dequeue(&skqueue))) {
852 ret = dev_queue_xmit(skb);
854 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
859 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
861 struct ipoib_dev_priv *priv = netdev_priv(dev);
862 struct ipoib_path *path;
864 if (!priv->broadcast)
867 path = kzalloc(sizeof *path, GFP_ATOMIC);
873 skb_queue_head_init(&path->queue);
875 INIT_LIST_HEAD(&path->neigh_list);
877 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
878 path->pathrec.sgid = priv->local_gid;
879 path->pathrec.pkey = cpu_to_be16(priv->pkey);
880 path->pathrec.numb_path = 1;
881 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
886 static int path_rec_start(struct net_device *dev,
887 struct ipoib_path *path)
889 struct ipoib_dev_priv *priv = netdev_priv(dev);
891 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
892 path->pathrec.dgid.raw);
894 init_completion(&path->done);
897 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
899 IB_SA_PATH_REC_DGID |
900 IB_SA_PATH_REC_SGID |
901 IB_SA_PATH_REC_NUMB_PATH |
902 IB_SA_PATH_REC_TRAFFIC_CLASS |
907 if (path->query_id < 0) {
908 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
910 complete(&path->done);
911 return path->query_id;
917 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
918 struct net_device *dev)
920 struct ipoib_dev_priv *priv = netdev_priv(dev);
921 struct ipoib_path *path;
922 struct ipoib_neigh *neigh;
925 spin_lock_irqsave(&priv->lock, flags);
926 neigh = ipoib_neigh_alloc(daddr, dev);
928 spin_unlock_irqrestore(&priv->lock, flags);
929 ++dev->stats.tx_dropped;
930 dev_kfree_skb_any(skb);
934 path = __path_find(dev, daddr + 4);
936 path = path_rec_create(dev, daddr + 4);
940 __path_add(dev, path);
943 list_add_tail(&neigh->list, &path->neigh_list);
946 kref_get(&path->ah->ref);
947 neigh->ah = path->ah;
949 if (ipoib_cm_enabled(dev, neigh->daddr)) {
950 if (!ipoib_cm_get(neigh))
951 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
952 if (!ipoib_cm_get(neigh)) {
953 ipoib_neigh_free(neigh);
956 if (skb_queue_len(&neigh->queue) <
957 IPOIB_MAX_PATH_REC_QUEUE) {
958 push_pseudo_header(skb, neigh->daddr);
959 __skb_queue_tail(&neigh->queue, skb);
961 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
962 skb_queue_len(&neigh->queue));
966 spin_unlock_irqrestore(&priv->lock, flags);
967 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
968 ipoib_neigh_put(neigh);
974 if (!path->query && path_rec_start(dev, path))
976 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
977 push_pseudo_header(skb, neigh->daddr);
978 __skb_queue_tail(&neigh->queue, skb);
984 spin_unlock_irqrestore(&priv->lock, flags);
985 ipoib_neigh_put(neigh);
989 ipoib_neigh_free(neigh);
991 ++dev->stats.tx_dropped;
992 dev_kfree_skb_any(skb);
994 spin_unlock_irqrestore(&priv->lock, flags);
995 ipoib_neigh_put(neigh);
998 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
999 struct ipoib_pseudo_header *phdr)
1001 struct ipoib_dev_priv *priv = netdev_priv(dev);
1002 struct ipoib_path *path;
1003 unsigned long flags;
1005 spin_lock_irqsave(&priv->lock, flags);
1007 path = __path_find(dev, phdr->hwaddr + 4);
1008 if (!path || !path->valid) {
1012 path = path_rec_create(dev, phdr->hwaddr + 4);
1016 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1017 push_pseudo_header(skb, phdr->hwaddr);
1018 __skb_queue_tail(&path->queue, skb);
1020 ++dev->stats.tx_dropped;
1021 dev_kfree_skb_any(skb);
1024 if (!path->query && path_rec_start(dev, path)) {
1025 spin_unlock_irqrestore(&priv->lock, flags);
1027 path_free(dev, path);
1030 __path_add(dev, path);
1032 ++dev->stats.tx_dropped;
1033 dev_kfree_skb_any(skb);
1036 spin_unlock_irqrestore(&priv->lock, flags);
1041 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
1042 be16_to_cpu(path->pathrec.dlid));
1044 spin_unlock_irqrestore(&priv->lock, flags);
1045 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
1047 } else if ((path->query || !path_rec_start(dev, path)) &&
1048 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1049 push_pseudo_header(skb, phdr->hwaddr);
1050 __skb_queue_tail(&path->queue, skb);
1052 ++dev->stats.tx_dropped;
1053 dev_kfree_skb_any(skb);
1056 spin_unlock_irqrestore(&priv->lock, flags);
1059 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1061 struct ipoib_dev_priv *priv = netdev_priv(dev);
1062 struct ipoib_neigh *neigh;
1063 struct ipoib_pseudo_header *phdr;
1064 struct ipoib_header *header;
1065 unsigned long flags;
1067 phdr = (struct ipoib_pseudo_header *) skb->data;
1068 skb_pull(skb, sizeof(*phdr));
1069 header = (struct ipoib_header *) skb->data;
1071 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1072 /* multicast, arrange "if" according to probability */
1073 if ((header->proto != htons(ETH_P_IP)) &&
1074 (header->proto != htons(ETH_P_IPV6)) &&
1075 (header->proto != htons(ETH_P_ARP)) &&
1076 (header->proto != htons(ETH_P_RARP)) &&
1077 (header->proto != htons(ETH_P_TIPC))) {
1078 /* ethertype not supported by IPoIB */
1079 ++dev->stats.tx_dropped;
1080 dev_kfree_skb_any(skb);
1081 return NETDEV_TX_OK;
1083 /* Add in the P_Key for multicast*/
1084 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1085 phdr->hwaddr[9] = priv->pkey & 0xff;
1087 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1089 goto send_using_neigh;
1090 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1091 return NETDEV_TX_OK;
1094 /* unicast, arrange "switch" according to probability */
1095 switch (header->proto) {
1096 case htons(ETH_P_IP):
1097 case htons(ETH_P_IPV6):
1098 case htons(ETH_P_TIPC):
1099 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1100 if (unlikely(!neigh)) {
1101 neigh_add_path(skb, phdr->hwaddr, dev);
1102 return NETDEV_TX_OK;
1105 case htons(ETH_P_ARP):
1106 case htons(ETH_P_RARP):
1107 /* for unicast ARP and RARP should always perform path find */
1108 unicast_arp_send(skb, dev, phdr);
1109 return NETDEV_TX_OK;
1111 /* ethertype not supported by IPoIB */
1112 ++dev->stats.tx_dropped;
1113 dev_kfree_skb_any(skb);
1114 return NETDEV_TX_OK;
1118 /* note we now hold a ref to neigh */
1119 if (ipoib_cm_get(neigh)) {
1120 if (ipoib_cm_up(neigh)) {
1121 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1124 } else if (neigh->ah) {
1125 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(phdr->hwaddr));
1129 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1130 push_pseudo_header(skb, phdr->hwaddr);
1131 spin_lock_irqsave(&priv->lock, flags);
1132 __skb_queue_tail(&neigh->queue, skb);
1133 spin_unlock_irqrestore(&priv->lock, flags);
1135 ++dev->stats.tx_dropped;
1136 dev_kfree_skb_any(skb);
1140 ipoib_neigh_put(neigh);
1142 return NETDEV_TX_OK;
1145 static void ipoib_timeout(struct net_device *dev)
1147 struct ipoib_dev_priv *priv = netdev_priv(dev);
1149 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1150 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1151 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1152 netif_queue_stopped(dev),
1153 priv->tx_head, priv->tx_tail);
1154 /* XXX reset QP, etc. */
1157 static int ipoib_hard_header(struct sk_buff *skb,
1158 struct net_device *dev,
1159 unsigned short type,
1160 const void *daddr, const void *saddr, unsigned len)
1162 struct ipoib_header *header;
1164 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1166 header->proto = htons(type);
1167 header->reserved = 0;
1170 * we don't rely on dst_entry structure, always stuff the
1171 * destination address into skb hard header so we can figure out where
1172 * to send the packet later.
1174 push_pseudo_header(skb, daddr);
1176 return IPOIB_HARD_LEN;
1179 static void ipoib_set_mcast_list(struct net_device *dev)
1181 struct ipoib_dev_priv *priv = netdev_priv(dev);
1183 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1184 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1188 queue_work(priv->wq, &priv->restart_task);
1191 static int ipoib_get_iflink(const struct net_device *dev)
1193 struct ipoib_dev_priv *priv = netdev_priv(dev);
1195 /* parent interface */
1196 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1197 return dev->ifindex;
1199 /* child/vlan interface */
1200 return priv->parent->ifindex;
1203 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1206 * Use only the address parts that contributes to spreading
1207 * The subnet prefix is not used as one can not connect to
1208 * same remote port (GUID) using the same remote QPN via two
1209 * different subnets.
1211 /* qpn octets[1:4) & port GUID octets[12:20) */
1212 u32 *d32 = (u32 *) daddr;
1215 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1216 return hv & htbl->mask;
1219 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1221 struct ipoib_dev_priv *priv = netdev_priv(dev);
1222 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1223 struct ipoib_neigh_hash *htbl;
1224 struct ipoib_neigh *neigh = NULL;
1229 htbl = rcu_dereference_bh(ntbl->htbl);
1234 hash_val = ipoib_addr_hash(htbl, daddr);
1235 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1237 neigh = rcu_dereference_bh(neigh->hnext)) {
1238 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1239 /* found, take one ref on behalf of the caller */
1240 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1246 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1247 neigh->alive = jiffies;
1253 rcu_read_unlock_bh();
1257 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1259 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1260 struct ipoib_neigh_hash *htbl;
1261 unsigned long neigh_obsolete;
1263 unsigned long flags;
1265 LIST_HEAD(remove_list);
1267 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1270 spin_lock_irqsave(&priv->lock, flags);
1272 htbl = rcu_dereference_protected(ntbl->htbl,
1273 lockdep_is_held(&priv->lock));
1278 /* neigh is obsolete if it was idle for two GC periods */
1279 dt = 2 * arp_tbl.gc_interval;
1280 neigh_obsolete = jiffies - dt;
1281 /* handle possible race condition */
1282 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1285 for (i = 0; i < htbl->size; i++) {
1286 struct ipoib_neigh *neigh;
1287 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1289 while ((neigh = rcu_dereference_protected(*np,
1290 lockdep_is_held(&priv->lock))) != NULL) {
1291 /* was the neigh idle for two GC periods */
1292 if (time_after(neigh_obsolete, neigh->alive)) {
1294 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1296 rcu_assign_pointer(*np,
1297 rcu_dereference_protected(neigh->hnext,
1298 lockdep_is_held(&priv->lock)));
1299 /* remove from path/mc list */
1300 list_del_init(&neigh->list);
1301 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1310 spin_unlock_irqrestore(&priv->lock, flags);
1311 ipoib_mcast_remove_list(&remove_list);
1314 static void ipoib_reap_neigh(struct work_struct *work)
1316 struct ipoib_dev_priv *priv =
1317 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1319 __ipoib_reap_neigh(priv);
1321 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1322 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1323 arp_tbl.gc_interval);
1327 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1328 struct net_device *dev)
1330 struct ipoib_neigh *neigh;
1332 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1337 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1338 skb_queue_head_init(&neigh->queue);
1339 INIT_LIST_HEAD(&neigh->list);
1340 ipoib_cm_set(neigh, NULL);
1341 /* one ref on behalf of the caller */
1342 atomic_set(&neigh->refcnt, 1);
1347 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1348 struct net_device *dev)
1350 struct ipoib_dev_priv *priv = netdev_priv(dev);
1351 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1352 struct ipoib_neigh_hash *htbl;
1353 struct ipoib_neigh *neigh;
1356 htbl = rcu_dereference_protected(ntbl->htbl,
1357 lockdep_is_held(&priv->lock));
1363 /* need to add a new neigh, but maybe some other thread succeeded?
1364 * recalc hash, maybe hash resize took place so we do a search
1366 hash_val = ipoib_addr_hash(htbl, daddr);
1367 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1368 lockdep_is_held(&priv->lock));
1370 neigh = rcu_dereference_protected(neigh->hnext,
1371 lockdep_is_held(&priv->lock))) {
1372 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1373 /* found, take one ref on behalf of the caller */
1374 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1379 neigh->alive = jiffies;
1384 neigh = ipoib_neigh_ctor(daddr, dev);
1388 /* one ref on behalf of the hash table */
1389 atomic_inc(&neigh->refcnt);
1390 neigh->alive = jiffies;
1392 rcu_assign_pointer(neigh->hnext,
1393 rcu_dereference_protected(htbl->buckets[hash_val],
1394 lockdep_is_held(&priv->lock)));
1395 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1396 atomic_inc(&ntbl->entries);
1403 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1405 /* neigh reference count was dropprd to zero */
1406 struct net_device *dev = neigh->dev;
1407 struct ipoib_dev_priv *priv = netdev_priv(dev);
1408 struct sk_buff *skb;
1410 ipoib_put_ah(neigh->ah);
1411 while ((skb = __skb_dequeue(&neigh->queue))) {
1412 ++dev->stats.tx_dropped;
1413 dev_kfree_skb_any(skb);
1415 if (ipoib_cm_get(neigh))
1416 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1417 ipoib_dbg(netdev_priv(dev),
1418 "neigh free for %06x %pI6\n",
1419 IPOIB_QPN(neigh->daddr),
1422 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1423 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1424 complete(&priv->ntbl.flushed);
1428 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1430 /* Called as a result of removal from hash table */
1431 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1432 /* note TX context may hold another ref */
1433 ipoib_neigh_put(neigh);
1436 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1438 struct net_device *dev = neigh->dev;
1439 struct ipoib_dev_priv *priv = netdev_priv(dev);
1440 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1441 struct ipoib_neigh_hash *htbl;
1442 struct ipoib_neigh __rcu **np;
1443 struct ipoib_neigh *n;
1446 htbl = rcu_dereference_protected(ntbl->htbl,
1447 lockdep_is_held(&priv->lock));
1451 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1452 np = &htbl->buckets[hash_val];
1453 for (n = rcu_dereference_protected(*np,
1454 lockdep_is_held(&priv->lock));
1456 n = rcu_dereference_protected(*np,
1457 lockdep_is_held(&priv->lock))) {
1460 rcu_assign_pointer(*np,
1461 rcu_dereference_protected(neigh->hnext,
1462 lockdep_is_held(&priv->lock)));
1463 /* remove from parent list */
1464 list_del_init(&neigh->list);
1465 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1473 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1475 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1476 struct ipoib_neigh_hash *htbl;
1477 struct ipoib_neigh __rcu **buckets;
1480 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1482 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1485 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1486 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1487 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1493 htbl->mask = (size - 1);
1494 htbl->buckets = buckets;
1495 RCU_INIT_POINTER(ntbl->htbl, htbl);
1497 atomic_set(&ntbl->entries, 0);
1499 /* start garbage collection */
1500 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1501 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1502 arp_tbl.gc_interval);
1507 static void neigh_hash_free_rcu(struct rcu_head *head)
1509 struct ipoib_neigh_hash *htbl = container_of(head,
1510 struct ipoib_neigh_hash,
1512 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1513 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1517 complete(&ntbl->deleted);
1520 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1522 struct ipoib_dev_priv *priv = netdev_priv(dev);
1523 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1524 struct ipoib_neigh_hash *htbl;
1525 unsigned long flags;
1528 /* remove all neigh connected to a given path or mcast */
1529 spin_lock_irqsave(&priv->lock, flags);
1531 htbl = rcu_dereference_protected(ntbl->htbl,
1532 lockdep_is_held(&priv->lock));
1537 for (i = 0; i < htbl->size; i++) {
1538 struct ipoib_neigh *neigh;
1539 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1541 while ((neigh = rcu_dereference_protected(*np,
1542 lockdep_is_held(&priv->lock))) != NULL) {
1543 /* delete neighs belong to this parent */
1544 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1545 rcu_assign_pointer(*np,
1546 rcu_dereference_protected(neigh->hnext,
1547 lockdep_is_held(&priv->lock)));
1548 /* remove from parent list */
1549 list_del_init(&neigh->list);
1550 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1558 spin_unlock_irqrestore(&priv->lock, flags);
1561 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1563 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1564 struct ipoib_neigh_hash *htbl;
1565 unsigned long flags;
1566 int i, wait_flushed = 0;
1568 init_completion(&priv->ntbl.flushed);
1570 spin_lock_irqsave(&priv->lock, flags);
1572 htbl = rcu_dereference_protected(ntbl->htbl,
1573 lockdep_is_held(&priv->lock));
1577 wait_flushed = atomic_read(&priv->ntbl.entries);
1581 for (i = 0; i < htbl->size; i++) {
1582 struct ipoib_neigh *neigh;
1583 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1585 while ((neigh = rcu_dereference_protected(*np,
1586 lockdep_is_held(&priv->lock))) != NULL) {
1587 rcu_assign_pointer(*np,
1588 rcu_dereference_protected(neigh->hnext,
1589 lockdep_is_held(&priv->lock)));
1590 /* remove from path/mc list */
1591 list_del_init(&neigh->list);
1592 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1597 rcu_assign_pointer(ntbl->htbl, NULL);
1598 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1601 spin_unlock_irqrestore(&priv->lock, flags);
1603 wait_for_completion(&priv->ntbl.flushed);
1606 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1608 struct ipoib_dev_priv *priv = netdev_priv(dev);
1611 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1612 init_completion(&priv->ntbl.deleted);
1613 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1615 /* Stop GC if called at init fail need to cancel work */
1616 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1618 cancel_delayed_work(&priv->neigh_reap_task);
1620 ipoib_flush_neighs(priv);
1622 wait_for_completion(&priv->ntbl.deleted);
1626 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1628 struct ipoib_dev_priv *priv = netdev_priv(dev);
1630 /* Allocate RX/TX "rings" to hold queued skbs */
1631 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1636 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1637 if (!priv->tx_ring) {
1638 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1639 ca->name, ipoib_sendq_size);
1640 goto out_rx_ring_cleanup;
1643 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1645 if (ipoib_ib_dev_init(dev, ca, port))
1646 goto out_tx_ring_cleanup;
1649 * Must be after ipoib_ib_dev_init so we can allocate a per
1650 * device wq there and use it here
1652 if (ipoib_neigh_hash_init(priv) < 0)
1653 goto out_dev_uninit;
1658 ipoib_ib_dev_cleanup(dev);
1660 out_tx_ring_cleanup:
1661 vfree(priv->tx_ring);
1663 out_rx_ring_cleanup:
1664 kfree(priv->rx_ring);
1670 void ipoib_dev_cleanup(struct net_device *dev)
1672 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1677 ipoib_delete_debug_files(dev);
1679 /* Delete any child interfaces first */
1680 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1681 /* Stop GC on child */
1682 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1683 cancel_delayed_work(&cpriv->neigh_reap_task);
1684 unregister_netdevice_queue(cpriv->dev, &head);
1686 unregister_netdevice_many(&head);
1689 * Must be before ipoib_ib_dev_cleanup or we delete an in use
1692 ipoib_neigh_hash_uninit(dev);
1694 ipoib_ib_dev_cleanup(dev);
1696 kfree(priv->rx_ring);
1697 vfree(priv->tx_ring);
1699 priv->rx_ring = NULL;
1700 priv->tx_ring = NULL;
1703 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1705 struct ipoib_dev_priv *priv = netdev_priv(dev);
1707 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1710 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1711 struct ifla_vf_info *ivf)
1713 struct ipoib_dev_priv *priv = netdev_priv(dev);
1716 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1725 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1727 struct ipoib_dev_priv *priv = netdev_priv(dev);
1729 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1732 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1735 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1736 struct ifla_vf_stats *vf_stats)
1738 struct ipoib_dev_priv *priv = netdev_priv(dev);
1740 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1743 static const struct header_ops ipoib_header_ops = {
1744 .create = ipoib_hard_header,
1747 static const struct net_device_ops ipoib_netdev_ops_pf = {
1748 .ndo_uninit = ipoib_uninit,
1749 .ndo_open = ipoib_open,
1750 .ndo_stop = ipoib_stop,
1751 .ndo_change_mtu = ipoib_change_mtu,
1752 .ndo_fix_features = ipoib_fix_features,
1753 .ndo_start_xmit = ipoib_start_xmit,
1754 .ndo_tx_timeout = ipoib_timeout,
1755 .ndo_set_rx_mode = ipoib_set_mcast_list,
1756 .ndo_get_iflink = ipoib_get_iflink,
1757 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
1758 .ndo_get_vf_config = ipoib_get_vf_config,
1759 .ndo_get_vf_stats = ipoib_get_vf_stats,
1760 .ndo_set_vf_guid = ipoib_set_vf_guid,
1761 .ndo_set_mac_address = ipoib_set_mac,
1764 static const struct net_device_ops ipoib_netdev_ops_vf = {
1765 .ndo_uninit = ipoib_uninit,
1766 .ndo_open = ipoib_open,
1767 .ndo_stop = ipoib_stop,
1768 .ndo_change_mtu = ipoib_change_mtu,
1769 .ndo_fix_features = ipoib_fix_features,
1770 .ndo_start_xmit = ipoib_start_xmit,
1771 .ndo_tx_timeout = ipoib_timeout,
1772 .ndo_set_rx_mode = ipoib_set_mcast_list,
1773 .ndo_get_iflink = ipoib_get_iflink,
1776 void ipoib_setup(struct net_device *dev)
1778 struct ipoib_dev_priv *priv = netdev_priv(dev);
1780 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1781 dev->netdev_ops = &ipoib_netdev_ops_vf;
1783 dev->netdev_ops = &ipoib_netdev_ops_pf;
1785 dev->header_ops = &ipoib_header_ops;
1787 ipoib_set_ethtool_ops(dev);
1789 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1791 dev->watchdog_timeo = HZ;
1793 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1795 dev->hard_header_len = IPOIB_HARD_LEN;
1796 dev->addr_len = INFINIBAND_ALEN;
1797 dev->type = ARPHRD_INFINIBAND;
1798 dev->tx_queue_len = ipoib_sendq_size * 2;
1799 dev->features = (NETIF_F_VLAN_CHALLENGED |
1801 netif_keep_dst(dev);
1803 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1807 spin_lock_init(&priv->lock);
1809 init_rwsem(&priv->vlan_rwsem);
1811 INIT_LIST_HEAD(&priv->path_list);
1812 INIT_LIST_HEAD(&priv->child_intfs);
1813 INIT_LIST_HEAD(&priv->dead_ahs);
1814 INIT_LIST_HEAD(&priv->multicast_list);
1816 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1817 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1818 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1819 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1820 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1821 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1822 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1823 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1826 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1828 struct net_device *dev;
1830 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1831 NET_NAME_UNKNOWN, ipoib_setup);
1835 return netdev_priv(dev);
1838 static ssize_t show_pkey(struct device *dev,
1839 struct device_attribute *attr, char *buf)
1841 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1843 return sprintf(buf, "0x%04x\n", priv->pkey);
1845 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1847 static ssize_t show_umcast(struct device *dev,
1848 struct device_attribute *attr, char *buf)
1850 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1852 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1855 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1857 struct ipoib_dev_priv *priv = netdev_priv(ndev);
1859 if (umcast_val > 0) {
1860 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1861 ipoib_warn(priv, "ignoring multicast groups joined directly "
1864 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1867 static ssize_t set_umcast(struct device *dev,
1868 struct device_attribute *attr,
1869 const char *buf, size_t count)
1871 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1873 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1877 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1879 int ipoib_add_umcast_attr(struct net_device *dev)
1881 return device_create_file(&dev->dev, &dev_attr_umcast);
1884 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
1886 struct ipoib_dev_priv *child_priv;
1887 struct net_device *netdev = priv->dev;
1889 netif_addr_lock_bh(netdev);
1891 memcpy(&priv->local_gid.global.interface_id,
1892 &gid->global.interface_id,
1893 sizeof(gid->global.interface_id));
1894 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
1895 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1897 netif_addr_unlock_bh(netdev);
1899 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1900 down_read(&priv->vlan_rwsem);
1901 list_for_each_entry(child_priv, &priv->child_intfs, list)
1902 set_base_guid(child_priv, gid);
1903 up_read(&priv->vlan_rwsem);
1907 static int ipoib_check_lladdr(struct net_device *dev,
1908 struct sockaddr_storage *ss)
1910 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
1913 netif_addr_lock_bh(dev);
1915 /* Make sure the QPN, reserved and subnet prefix match the current
1916 * lladdr, it also makes sure the lladdr is unicast.
1918 if (memcmp(dev->dev_addr, ss->__data,
1919 4 + sizeof(gid->global.subnet_prefix)) ||
1920 gid->global.interface_id == 0)
1923 netif_addr_unlock_bh(dev);
1928 static int ipoib_set_mac(struct net_device *dev, void *addr)
1930 struct ipoib_dev_priv *priv = netdev_priv(dev);
1931 struct sockaddr_storage *ss = addr;
1934 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
1937 ret = ipoib_check_lladdr(dev, ss);
1941 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
1943 queue_work(ipoib_workqueue, &priv->flush_light);
1948 static ssize_t create_child(struct device *dev,
1949 struct device_attribute *attr,
1950 const char *buf, size_t count)
1955 if (sscanf(buf, "%i", &pkey) != 1)
1958 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1962 * Set the full membership bit, so that we join the right
1963 * broadcast group, etc.
1967 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1969 return ret ? ret : count;
1971 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1973 static ssize_t delete_child(struct device *dev,
1974 struct device_attribute *attr,
1975 const char *buf, size_t count)
1980 if (sscanf(buf, "%i", &pkey) != 1)
1983 if (pkey < 0 || pkey > 0xffff)
1986 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1988 return ret ? ret : count;
1991 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1993 int ipoib_add_pkey_attr(struct net_device *dev)
1995 return device_create_file(&dev->dev, &dev_attr_pkey);
1998 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2000 priv->hca_caps = hca->attrs.device_cap_flags;
2002 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2003 priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2005 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2006 priv->dev->hw_features |= NETIF_F_TSO;
2008 priv->dev->features |= priv->dev->hw_features;
2012 static struct net_device *ipoib_add_port(const char *format,
2013 struct ib_device *hca, u8 port)
2015 struct ipoib_dev_priv *priv;
2016 struct ib_port_attr attr;
2017 int result = -ENOMEM;
2019 priv = ipoib_intf_alloc(format);
2021 goto alloc_mem_failed;
2023 SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2024 priv->dev->dev_id = port - 1;
2026 result = ib_query_port(hca, port, &attr);
2028 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2030 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2032 goto device_init_failed;
2035 /* MTU will be reset when mcast join happens */
2036 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
2037 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
2038 priv->dev->max_mtu = IPOIB_CM_MTU;
2040 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2042 result = ib_query_pkey(hca, port, 0, &priv->pkey);
2044 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2045 hca->name, port, result);
2046 goto device_init_failed;
2049 ipoib_set_dev_features(priv, hca);
2052 * Set the full membership bit, so that we join the right
2053 * broadcast group, etc.
2055 priv->pkey |= 0x8000;
2057 priv->dev->broadcast[8] = priv->pkey >> 8;
2058 priv->dev->broadcast[9] = priv->pkey & 0xff;
2060 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2062 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2063 hca->name, port, result);
2064 goto device_init_failed;
2066 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2067 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2069 result = ipoib_dev_init(priv->dev, hca, port);
2071 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2072 hca->name, port, result);
2073 goto device_init_failed;
2076 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2077 priv->ca, ipoib_event);
2078 result = ib_register_event_handler(&priv->event_handler);
2080 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2081 "port %d (ret = %d)\n",
2082 hca->name, port, result);
2086 result = register_netdev(priv->dev);
2088 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2089 hca->name, port, result);
2090 goto register_failed;
2093 ipoib_create_debug_files(priv->dev);
2095 if (ipoib_cm_add_mode_attr(priv->dev))
2097 if (ipoib_add_pkey_attr(priv->dev))
2099 if (ipoib_add_umcast_attr(priv->dev))
2101 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2103 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2109 ipoib_delete_debug_files(priv->dev);
2110 unregister_netdev(priv->dev);
2113 ib_unregister_event_handler(&priv->event_handler);
2114 flush_workqueue(ipoib_workqueue);
2115 /* Stop GC if started before flush */
2116 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2117 cancel_delayed_work(&priv->neigh_reap_task);
2118 flush_workqueue(priv->wq);
2121 ipoib_dev_cleanup(priv->dev);
2124 free_netdev(priv->dev);
2127 return ERR_PTR(result);
2130 static void ipoib_add_one(struct ib_device *device)
2132 struct list_head *dev_list;
2133 struct net_device *dev;
2134 struct ipoib_dev_priv *priv;
2138 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2142 INIT_LIST_HEAD(dev_list);
2144 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2145 if (!rdma_protocol_ib(device, p))
2147 dev = ipoib_add_port("ib%d", device, p);
2149 priv = netdev_priv(dev);
2150 list_add_tail(&priv->list, dev_list);
2160 ib_set_client_data(device, &ipoib_client, dev_list);
2163 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2165 struct ipoib_dev_priv *priv, *tmp;
2166 struct list_head *dev_list = client_data;
2171 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2172 ib_unregister_event_handler(&priv->event_handler);
2173 flush_workqueue(ipoib_workqueue);
2175 /* mark interface in the middle of destruction */
2176 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2179 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2183 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2184 cancel_delayed_work(&priv->neigh_reap_task);
2185 flush_workqueue(priv->wq);
2187 unregister_netdev(priv->dev);
2188 free_netdev(priv->dev);
2194 static int __init ipoib_init_module(void)
2198 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2199 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2200 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2202 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2203 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2204 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2205 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2206 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2210 * When copying small received packets, we only copy from the
2211 * linear data part of the SKB, so we rely on this condition.
2213 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2215 ret = ipoib_register_debugfs();
2220 * We create a global workqueue here that is used for all flush
2221 * operations. However, if you attempt to flush a workqueue
2222 * from a task on that same workqueue, it deadlocks the system.
2223 * We want to be able to flush the tasks associated with a
2224 * specific net device, so we also create a workqueue for each
2225 * netdevice. We queue up the tasks for that device only on
2226 * its private workqueue, and we only queue up flush events
2227 * on our global flush workqueue. This avoids the deadlocks.
2229 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2231 if (!ipoib_workqueue) {
2236 ib_sa_register_client(&ipoib_sa_client);
2238 ret = ib_register_client(&ipoib_client);
2242 ret = ipoib_netlink_init();
2249 ib_unregister_client(&ipoib_client);
2252 ib_sa_unregister_client(&ipoib_sa_client);
2253 destroy_workqueue(ipoib_workqueue);
2256 ipoib_unregister_debugfs();
2261 static void __exit ipoib_cleanup_module(void)
2263 ipoib_netlink_fini();
2264 ib_unregister_client(&ipoib_client);
2265 ib_sa_unregister_client(&ipoib_sa_client);
2266 ipoib_unregister_debugfs();
2267 destroy_workqueue(ipoib_workqueue);
2270 module_init(ipoib_init_module);
2271 module_exit(ipoib_cleanup_module);