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 if (ipoib_ib_dev_up(dev))
132 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
133 struct ipoib_dev_priv *cpriv;
135 /* Bring up any child interfaces too */
136 down_read(&priv->vlan_rwsem);
137 list_for_each_entry(cpriv, &priv->child_intfs, list) {
140 flags = cpriv->dev->flags;
144 dev_change_flags(cpriv->dev, flags | IFF_UP);
146 up_read(&priv->vlan_rwsem);
149 netif_start_queue(dev);
154 ipoib_ib_dev_stop(dev);
157 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
162 static int ipoib_stop(struct net_device *dev)
164 struct ipoib_dev_priv *priv = netdev_priv(dev);
166 ipoib_dbg(priv, "stopping interface\n");
168 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
170 netif_stop_queue(dev);
172 ipoib_ib_dev_down(dev);
173 ipoib_ib_dev_stop(dev);
175 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
176 struct ipoib_dev_priv *cpriv;
178 /* Bring down any child interfaces too */
179 down_read(&priv->vlan_rwsem);
180 list_for_each_entry(cpriv, &priv->child_intfs, list) {
183 flags = cpriv->dev->flags;
184 if (!(flags & IFF_UP))
187 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
189 up_read(&priv->vlan_rwsem);
195 static void ipoib_uninit(struct net_device *dev)
197 ipoib_dev_cleanup(dev);
200 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
202 struct ipoib_dev_priv *priv = netdev_priv(dev);
204 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
205 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
210 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
212 struct ipoib_dev_priv *priv = netdev_priv(dev);
214 /* dev->mtu > 2K ==> connected mode */
215 if (ipoib_cm_admin_enabled(dev)) {
216 if (new_mtu > ipoib_cm_max_mtu(dev))
219 if (new_mtu > priv->mcast_mtu)
220 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
227 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
230 priv->admin_mtu = new_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);
296 * Find a net_device matching the given address, which is an upper device of
297 * the given net_device.
298 * @addr: IP address to look for.
299 * @dev: base IPoIB net_device
301 * If found, returns the net_device with a reference held. Otherwise return
304 static struct net_device *ipoib_get_net_dev_match_addr(
305 const struct sockaddr *addr, struct net_device *dev)
307 struct net_device *upper,
309 struct list_head *iter;
312 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
318 netdev_for_each_all_upper_dev_rcu(dev, upper, iter) {
319 if (ipoib_is_dev_match_addr_rcu(addr, upper)) {
330 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
331 * pkey_index and address, if one exists.
333 * @found_net_dev: contains a matching net_device if the return value >= 1,
334 * with a reference held. */
335 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
336 const union ib_gid *gid,
338 const struct sockaddr *addr,
340 struct net_device **found_net_dev)
342 struct ipoib_dev_priv *child_priv;
343 struct net_device *net_dev = NULL;
346 if (priv->pkey_index == pkey_index &&
347 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
349 net_dev = ipoib_get_master_net_dev(priv->dev);
351 /* Verify the net_device matches the IP address, as
352 * IPoIB child devices currently share a GID. */
353 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
357 *found_net_dev = net_dev;
364 /* Check child interfaces */
365 down_read_nested(&priv->vlan_rwsem, nesting);
366 list_for_each_entry(child_priv, &priv->child_intfs, list) {
367 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
374 up_read(&priv->vlan_rwsem);
379 /* Returns the number of matching net_devs found (between 0 and 2). Also
380 * return the matching net_device in the @net_dev parameter, holding a
381 * reference to the net_device, if the number of matches >= 1 */
382 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
384 const union ib_gid *gid,
385 const struct sockaddr *addr,
386 struct net_device **net_dev)
388 struct ipoib_dev_priv *priv;
393 list_for_each_entry(priv, dev_list, list) {
394 if (priv->port != port)
397 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
406 static struct net_device *ipoib_get_net_dev_by_params(
407 struct ib_device *dev, u8 port, u16 pkey,
408 const union ib_gid *gid, const struct sockaddr *addr,
411 struct net_device *net_dev;
412 struct list_head *dev_list = client_data;
417 if (!rdma_protocol_ib(dev, port))
420 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
427 /* See if we can find a unique device matching the L2 parameters */
428 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
429 gid, NULL, &net_dev);
440 /* Couldn't find a unique device with L2 parameters only. Use L3
441 * address to uniquely match the net device */
442 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
443 gid, addr, &net_dev);
448 dev_warn_ratelimited(&dev->dev,
449 "duplicate IP address detected\n");
456 int ipoib_set_mode(struct net_device *dev, const char *buf)
458 struct ipoib_dev_priv *priv = netdev_priv(dev);
460 /* flush paths if we switch modes so that connections are restarted */
461 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
462 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
463 ipoib_warn(priv, "enabling connected mode "
464 "will cause multicast packet drops\n");
465 netdev_update_features(dev);
466 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
468 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
470 ipoib_flush_paths(dev);
475 if (!strcmp(buf, "datagram\n")) {
476 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
477 netdev_update_features(dev);
478 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
480 ipoib_flush_paths(dev);
488 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
490 struct ipoib_dev_priv *priv = netdev_priv(dev);
491 struct rb_node *n = priv->path_tree.rb_node;
492 struct ipoib_path *path;
496 path = rb_entry(n, struct ipoib_path, rb_node);
498 ret = memcmp(gid, path->pathrec.dgid.raw,
499 sizeof (union ib_gid));
512 static int __path_add(struct net_device *dev, struct ipoib_path *path)
514 struct ipoib_dev_priv *priv = netdev_priv(dev);
515 struct rb_node **n = &priv->path_tree.rb_node;
516 struct rb_node *pn = NULL;
517 struct ipoib_path *tpath;
522 tpath = rb_entry(pn, struct ipoib_path, rb_node);
524 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
525 sizeof (union ib_gid));
534 rb_link_node(&path->rb_node, pn, n);
535 rb_insert_color(&path->rb_node, &priv->path_tree);
537 list_add_tail(&path->list, &priv->path_list);
542 static void path_free(struct net_device *dev, struct ipoib_path *path)
546 while ((skb = __skb_dequeue(&path->queue)))
547 dev_kfree_skb_irq(skb);
549 ipoib_dbg(netdev_priv(dev), "path_free\n");
551 /* remove all neigh connected to this path */
552 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
555 ipoib_put_ah(path->ah);
560 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
562 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
564 struct ipoib_path_iter *iter;
566 iter = kmalloc(sizeof *iter, GFP_KERNEL);
571 memset(iter->path.pathrec.dgid.raw, 0, 16);
573 if (ipoib_path_iter_next(iter)) {
581 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
583 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
585 struct ipoib_path *path;
588 spin_lock_irq(&priv->lock);
590 n = rb_first(&priv->path_tree);
593 path = rb_entry(n, struct ipoib_path, rb_node);
595 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
596 sizeof (union ib_gid)) < 0) {
605 spin_unlock_irq(&priv->lock);
610 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
611 struct ipoib_path *path)
616 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
618 void ipoib_mark_paths_invalid(struct net_device *dev)
620 struct ipoib_dev_priv *priv = netdev_priv(dev);
621 struct ipoib_path *path, *tp;
623 spin_lock_irq(&priv->lock);
625 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
626 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
627 be16_to_cpu(path->pathrec.dlid),
628 path->pathrec.dgid.raw);
632 spin_unlock_irq(&priv->lock);
635 struct classport_info_context {
636 struct ipoib_dev_priv *priv;
637 struct completion done;
638 struct ib_sa_query *sa_query;
641 static void classport_info_query_cb(int status, struct ib_class_port_info *rec,
644 struct classport_info_context *cb_ctx = context;
645 struct ipoib_dev_priv *priv;
651 if (status || !rec) {
652 pr_debug("device: %s failed query classport_info status: %d\n",
653 priv->dev->name, status);
654 /* keeps the default, will try next mcast_restart */
655 priv->sm_fullmember_sendonly_support = false;
659 if (ib_get_cpi_capmask2(rec) &
660 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT) {
661 pr_debug("device: %s enabled fullmember-sendonly for sendonly MCG\n",
663 priv->sm_fullmember_sendonly_support = true;
665 pr_debug("device: %s disabled fullmember-sendonly for sendonly MCG\n",
667 priv->sm_fullmember_sendonly_support = false;
671 complete(&cb_ctx->done);
674 int ipoib_check_sm_sendonly_fullmember_support(struct ipoib_dev_priv *priv)
676 struct classport_info_context *callback_context;
679 callback_context = kmalloc(sizeof(*callback_context), GFP_KERNEL);
680 if (!callback_context)
683 callback_context->priv = priv;
684 init_completion(&callback_context->done);
686 ret = ib_sa_classport_info_rec_query(&ipoib_sa_client,
687 priv->ca, priv->port, 3000,
689 classport_info_query_cb,
691 &callback_context->sa_query);
693 pr_info("%s failed to send ib_sa_classport_info query, ret: %d\n",
694 priv->dev->name, ret);
695 kfree(callback_context);
699 /* waiting for the callback to finish before returnning */
700 wait_for_completion(&callback_context->done);
701 kfree(callback_context);
706 void ipoib_flush_paths(struct net_device *dev)
708 struct ipoib_dev_priv *priv = netdev_priv(dev);
709 struct ipoib_path *path, *tp;
710 LIST_HEAD(remove_list);
713 netif_tx_lock_bh(dev);
714 spin_lock_irqsave(&priv->lock, flags);
716 list_splice_init(&priv->path_list, &remove_list);
718 list_for_each_entry(path, &remove_list, list)
719 rb_erase(&path->rb_node, &priv->path_tree);
721 list_for_each_entry_safe(path, tp, &remove_list, list) {
723 ib_sa_cancel_query(path->query_id, path->query);
724 spin_unlock_irqrestore(&priv->lock, flags);
725 netif_tx_unlock_bh(dev);
726 wait_for_completion(&path->done);
727 path_free(dev, path);
728 netif_tx_lock_bh(dev);
729 spin_lock_irqsave(&priv->lock, flags);
732 spin_unlock_irqrestore(&priv->lock, flags);
733 netif_tx_unlock_bh(dev);
736 static void path_rec_completion(int status,
737 struct ib_sa_path_rec *pathrec,
740 struct ipoib_path *path = path_ptr;
741 struct net_device *dev = path->dev;
742 struct ipoib_dev_priv *priv = netdev_priv(dev);
743 struct ipoib_ah *ah = NULL;
744 struct ipoib_ah *old_ah = NULL;
745 struct ipoib_neigh *neigh, *tn;
746 struct sk_buff_head skqueue;
751 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
752 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
754 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
755 status, path->pathrec.dgid.raw);
757 skb_queue_head_init(&skqueue);
760 struct ib_ah_attr av;
762 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
763 ah = ipoib_create_ah(dev, priv->pd, &av);
766 spin_lock_irqsave(&priv->lock, flags);
768 if (!IS_ERR_OR_NULL(ah)) {
769 path->pathrec = *pathrec;
774 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
775 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
777 while ((skb = __skb_dequeue(&path->queue)))
778 __skb_queue_tail(&skqueue, skb);
780 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
782 WARN_ON(neigh->ah != old_ah);
784 * Dropping the ah reference inside
785 * priv->lock is safe here, because we
786 * will hold one more reference from
787 * the original value of path->ah (ie
790 ipoib_put_ah(neigh->ah);
792 kref_get(&path->ah->ref);
793 neigh->ah = path->ah;
795 if (ipoib_cm_enabled(dev, neigh->daddr)) {
796 if (!ipoib_cm_get(neigh))
797 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
800 if (!ipoib_cm_get(neigh)) {
801 ipoib_neigh_free(neigh);
806 while ((skb = __skb_dequeue(&neigh->queue)))
807 __skb_queue_tail(&skqueue, skb);
813 complete(&path->done);
815 spin_unlock_irqrestore(&priv->lock, flags);
817 if (IS_ERR_OR_NULL(ah))
818 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
821 ipoib_put_ah(old_ah);
823 while ((skb = __skb_dequeue(&skqueue))) {
825 if (dev_queue_xmit(skb))
826 ipoib_warn(priv, "dev_queue_xmit failed "
827 "to requeue packet\n");
831 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
833 struct ipoib_dev_priv *priv = netdev_priv(dev);
834 struct ipoib_path *path;
836 if (!priv->broadcast)
839 path = kzalloc(sizeof *path, GFP_ATOMIC);
845 skb_queue_head_init(&path->queue);
847 INIT_LIST_HEAD(&path->neigh_list);
849 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
850 path->pathrec.sgid = priv->local_gid;
851 path->pathrec.pkey = cpu_to_be16(priv->pkey);
852 path->pathrec.numb_path = 1;
853 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
858 static int path_rec_start(struct net_device *dev,
859 struct ipoib_path *path)
861 struct ipoib_dev_priv *priv = netdev_priv(dev);
863 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
864 path->pathrec.dgid.raw);
866 init_completion(&path->done);
869 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
871 IB_SA_PATH_REC_DGID |
872 IB_SA_PATH_REC_SGID |
873 IB_SA_PATH_REC_NUMB_PATH |
874 IB_SA_PATH_REC_TRAFFIC_CLASS |
879 if (path->query_id < 0) {
880 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
882 complete(&path->done);
883 return path->query_id;
889 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
890 struct net_device *dev)
892 struct ipoib_dev_priv *priv = netdev_priv(dev);
893 struct ipoib_path *path;
894 struct ipoib_neigh *neigh;
897 spin_lock_irqsave(&priv->lock, flags);
898 neigh = ipoib_neigh_alloc(daddr, dev);
900 spin_unlock_irqrestore(&priv->lock, flags);
901 ++dev->stats.tx_dropped;
902 dev_kfree_skb_any(skb);
906 path = __path_find(dev, daddr + 4);
908 path = path_rec_create(dev, daddr + 4);
912 __path_add(dev, path);
915 list_add_tail(&neigh->list, &path->neigh_list);
918 kref_get(&path->ah->ref);
919 neigh->ah = path->ah;
921 if (ipoib_cm_enabled(dev, neigh->daddr)) {
922 if (!ipoib_cm_get(neigh))
923 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
924 if (!ipoib_cm_get(neigh)) {
925 ipoib_neigh_free(neigh);
928 if (skb_queue_len(&neigh->queue) <
929 IPOIB_MAX_PATH_REC_QUEUE) {
930 /* put pseudoheader back on for next time */
931 skb_push(skb, IPOIB_PSEUDO_LEN);
932 __skb_queue_tail(&neigh->queue, skb);
934 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
935 skb_queue_len(&neigh->queue));
939 spin_unlock_irqrestore(&priv->lock, flags);
940 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
941 ipoib_neigh_put(neigh);
947 if (!path->query && path_rec_start(dev, path))
949 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
950 __skb_queue_tail(&neigh->queue, skb);
955 spin_unlock_irqrestore(&priv->lock, flags);
956 ipoib_neigh_put(neigh);
960 ipoib_neigh_free(neigh);
962 ++dev->stats.tx_dropped;
963 dev_kfree_skb_any(skb);
965 spin_unlock_irqrestore(&priv->lock, flags);
966 ipoib_neigh_put(neigh);
969 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
970 struct ipoib_pseudo_header *phdr)
972 struct ipoib_dev_priv *priv = netdev_priv(dev);
973 struct ipoib_path *path;
976 spin_lock_irqsave(&priv->lock, flags);
978 path = __path_find(dev, phdr->hwaddr + 4);
979 if (!path || !path->valid) {
983 path = path_rec_create(dev, phdr->hwaddr + 4);
987 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
988 /* put pseudoheader back on for next time */
989 skb_push(skb, IPOIB_PSEUDO_LEN);
990 __skb_queue_tail(&path->queue, skb);
992 ++dev->stats.tx_dropped;
993 dev_kfree_skb_any(skb);
996 if (!path->query && path_rec_start(dev, path)) {
997 spin_unlock_irqrestore(&priv->lock, flags);
999 path_free(dev, path);
1002 __path_add(dev, path);
1004 ++dev->stats.tx_dropped;
1005 dev_kfree_skb_any(skb);
1008 spin_unlock_irqrestore(&priv->lock, flags);
1013 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
1014 be16_to_cpu(path->pathrec.dlid));
1016 spin_unlock_irqrestore(&priv->lock, flags);
1017 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
1019 } else if ((path->query || !path_rec_start(dev, path)) &&
1020 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1021 /* put pseudoheader back on for next time */
1022 skb_push(skb, IPOIB_PSEUDO_LEN);
1023 __skb_queue_tail(&path->queue, skb);
1025 ++dev->stats.tx_dropped;
1026 dev_kfree_skb_any(skb);
1029 spin_unlock_irqrestore(&priv->lock, flags);
1032 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1034 struct ipoib_dev_priv *priv = netdev_priv(dev);
1035 struct ipoib_neigh *neigh;
1036 struct ipoib_pseudo_header *phdr;
1037 struct ipoib_header *header;
1038 unsigned long flags;
1040 phdr = (struct ipoib_pseudo_header *) skb->data;
1041 skb_pull(skb, sizeof(*phdr));
1042 header = (struct ipoib_header *) skb->data;
1044 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1045 /* multicast, arrange "if" according to probability */
1046 if ((header->proto != htons(ETH_P_IP)) &&
1047 (header->proto != htons(ETH_P_IPV6)) &&
1048 (header->proto != htons(ETH_P_ARP)) &&
1049 (header->proto != htons(ETH_P_RARP)) &&
1050 (header->proto != htons(ETH_P_TIPC))) {
1051 /* ethertype not supported by IPoIB */
1052 ++dev->stats.tx_dropped;
1053 dev_kfree_skb_any(skb);
1054 return NETDEV_TX_OK;
1056 /* Add in the P_Key for multicast*/
1057 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1058 phdr->hwaddr[9] = priv->pkey & 0xff;
1060 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1062 goto send_using_neigh;
1063 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1064 return NETDEV_TX_OK;
1067 /* unicast, arrange "switch" according to probability */
1068 switch (header->proto) {
1069 case htons(ETH_P_IP):
1070 case htons(ETH_P_IPV6):
1071 case htons(ETH_P_TIPC):
1072 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1073 if (unlikely(!neigh)) {
1074 neigh_add_path(skb, phdr->hwaddr, dev);
1075 return NETDEV_TX_OK;
1078 case htons(ETH_P_ARP):
1079 case htons(ETH_P_RARP):
1080 /* for unicast ARP and RARP should always perform path find */
1081 unicast_arp_send(skb, dev, phdr);
1082 return NETDEV_TX_OK;
1084 /* ethertype not supported by IPoIB */
1085 ++dev->stats.tx_dropped;
1086 dev_kfree_skb_any(skb);
1087 return NETDEV_TX_OK;
1091 /* note we now hold a ref to neigh */
1092 if (ipoib_cm_get(neigh)) {
1093 if (ipoib_cm_up(neigh)) {
1094 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1097 } else if (neigh->ah) {
1098 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(phdr->hwaddr));
1102 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1103 /* put pseudoheader back on for next time */
1104 skb_push(skb, sizeof(*phdr));
1105 spin_lock_irqsave(&priv->lock, flags);
1106 __skb_queue_tail(&neigh->queue, skb);
1107 spin_unlock_irqrestore(&priv->lock, flags);
1109 ++dev->stats.tx_dropped;
1110 dev_kfree_skb_any(skb);
1114 ipoib_neigh_put(neigh);
1116 return NETDEV_TX_OK;
1119 static void ipoib_timeout(struct net_device *dev)
1121 struct ipoib_dev_priv *priv = netdev_priv(dev);
1123 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1124 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1125 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1126 netif_queue_stopped(dev),
1127 priv->tx_head, priv->tx_tail);
1128 /* XXX reset QP, etc. */
1131 static int ipoib_hard_header(struct sk_buff *skb,
1132 struct net_device *dev,
1133 unsigned short type,
1134 const void *daddr, const void *saddr, unsigned len)
1136 struct ipoib_pseudo_header *phdr;
1137 struct ipoib_header *header;
1139 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1141 header->proto = htons(type);
1142 header->reserved = 0;
1145 * we don't rely on dst_entry structure, always stuff the
1146 * destination address into skb hard header so we can figure out where
1147 * to send the packet later.
1149 phdr = (struct ipoib_pseudo_header *) skb_push(skb, sizeof(*phdr));
1150 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
1152 return IPOIB_HARD_LEN;
1155 static void ipoib_set_mcast_list(struct net_device *dev)
1157 struct ipoib_dev_priv *priv = netdev_priv(dev);
1159 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1160 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1164 queue_work(priv->wq, &priv->restart_task);
1167 static int ipoib_get_iflink(const struct net_device *dev)
1169 struct ipoib_dev_priv *priv = netdev_priv(dev);
1171 /* parent interface */
1172 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1173 return dev->ifindex;
1175 /* child/vlan interface */
1176 return priv->parent->ifindex;
1179 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1182 * Use only the address parts that contributes to spreading
1183 * The subnet prefix is not used as one can not connect to
1184 * same remote port (GUID) using the same remote QPN via two
1185 * different subnets.
1187 /* qpn octets[1:4) & port GUID octets[12:20) */
1188 u32 *d32 = (u32 *) daddr;
1191 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1192 return hv & htbl->mask;
1195 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1197 struct ipoib_dev_priv *priv = netdev_priv(dev);
1198 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1199 struct ipoib_neigh_hash *htbl;
1200 struct ipoib_neigh *neigh = NULL;
1205 htbl = rcu_dereference_bh(ntbl->htbl);
1210 hash_val = ipoib_addr_hash(htbl, daddr);
1211 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1213 neigh = rcu_dereference_bh(neigh->hnext)) {
1214 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1215 /* found, take one ref on behalf of the caller */
1216 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1222 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1223 neigh->alive = jiffies;
1229 rcu_read_unlock_bh();
1233 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1235 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1236 struct ipoib_neigh_hash *htbl;
1237 unsigned long neigh_obsolete;
1239 unsigned long flags;
1241 LIST_HEAD(remove_list);
1243 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1246 spin_lock_irqsave(&priv->lock, flags);
1248 htbl = rcu_dereference_protected(ntbl->htbl,
1249 lockdep_is_held(&priv->lock));
1254 /* neigh is obsolete if it was idle for two GC periods */
1255 dt = 2 * arp_tbl.gc_interval;
1256 neigh_obsolete = jiffies - dt;
1257 /* handle possible race condition */
1258 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1261 for (i = 0; i < htbl->size; i++) {
1262 struct ipoib_neigh *neigh;
1263 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1265 while ((neigh = rcu_dereference_protected(*np,
1266 lockdep_is_held(&priv->lock))) != NULL) {
1267 /* was the neigh idle for two GC periods */
1268 if (time_after(neigh_obsolete, neigh->alive)) {
1270 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1272 rcu_assign_pointer(*np,
1273 rcu_dereference_protected(neigh->hnext,
1274 lockdep_is_held(&priv->lock)));
1275 /* remove from path/mc list */
1276 list_del(&neigh->list);
1277 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1286 spin_unlock_irqrestore(&priv->lock, flags);
1287 ipoib_mcast_remove_list(&remove_list);
1290 static void ipoib_reap_neigh(struct work_struct *work)
1292 struct ipoib_dev_priv *priv =
1293 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1295 __ipoib_reap_neigh(priv);
1297 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1298 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1299 arp_tbl.gc_interval);
1303 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1304 struct net_device *dev)
1306 struct ipoib_neigh *neigh;
1308 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1313 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1314 skb_queue_head_init(&neigh->queue);
1315 INIT_LIST_HEAD(&neigh->list);
1316 ipoib_cm_set(neigh, NULL);
1317 /* one ref on behalf of the caller */
1318 atomic_set(&neigh->refcnt, 1);
1323 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1324 struct net_device *dev)
1326 struct ipoib_dev_priv *priv = netdev_priv(dev);
1327 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1328 struct ipoib_neigh_hash *htbl;
1329 struct ipoib_neigh *neigh;
1332 htbl = rcu_dereference_protected(ntbl->htbl,
1333 lockdep_is_held(&priv->lock));
1339 /* need to add a new neigh, but maybe some other thread succeeded?
1340 * recalc hash, maybe hash resize took place so we do a search
1342 hash_val = ipoib_addr_hash(htbl, daddr);
1343 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1344 lockdep_is_held(&priv->lock));
1346 neigh = rcu_dereference_protected(neigh->hnext,
1347 lockdep_is_held(&priv->lock))) {
1348 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1349 /* found, take one ref on behalf of the caller */
1350 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1355 neigh->alive = jiffies;
1360 neigh = ipoib_neigh_ctor(daddr, dev);
1364 /* one ref on behalf of the hash table */
1365 atomic_inc(&neigh->refcnt);
1366 neigh->alive = jiffies;
1368 rcu_assign_pointer(neigh->hnext,
1369 rcu_dereference_protected(htbl->buckets[hash_val],
1370 lockdep_is_held(&priv->lock)));
1371 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1372 atomic_inc(&ntbl->entries);
1379 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1381 /* neigh reference count was dropprd to zero */
1382 struct net_device *dev = neigh->dev;
1383 struct ipoib_dev_priv *priv = netdev_priv(dev);
1384 struct sk_buff *skb;
1386 ipoib_put_ah(neigh->ah);
1387 while ((skb = __skb_dequeue(&neigh->queue))) {
1388 ++dev->stats.tx_dropped;
1389 dev_kfree_skb_any(skb);
1391 if (ipoib_cm_get(neigh))
1392 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1393 ipoib_dbg(netdev_priv(dev),
1394 "neigh free for %06x %pI6\n",
1395 IPOIB_QPN(neigh->daddr),
1398 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1399 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1400 complete(&priv->ntbl.flushed);
1404 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1406 /* Called as a result of removal from hash table */
1407 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1408 /* note TX context may hold another ref */
1409 ipoib_neigh_put(neigh);
1412 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1414 struct net_device *dev = neigh->dev;
1415 struct ipoib_dev_priv *priv = netdev_priv(dev);
1416 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1417 struct ipoib_neigh_hash *htbl;
1418 struct ipoib_neigh __rcu **np;
1419 struct ipoib_neigh *n;
1422 htbl = rcu_dereference_protected(ntbl->htbl,
1423 lockdep_is_held(&priv->lock));
1427 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1428 np = &htbl->buckets[hash_val];
1429 for (n = rcu_dereference_protected(*np,
1430 lockdep_is_held(&priv->lock));
1432 n = rcu_dereference_protected(*np,
1433 lockdep_is_held(&priv->lock))) {
1436 rcu_assign_pointer(*np,
1437 rcu_dereference_protected(neigh->hnext,
1438 lockdep_is_held(&priv->lock)));
1439 /* remove from parent list */
1440 list_del(&neigh->list);
1441 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1449 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1451 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1452 struct ipoib_neigh_hash *htbl;
1453 struct ipoib_neigh __rcu **buckets;
1456 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1458 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1461 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1462 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1463 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1469 htbl->mask = (size - 1);
1470 htbl->buckets = buckets;
1471 RCU_INIT_POINTER(ntbl->htbl, htbl);
1473 atomic_set(&ntbl->entries, 0);
1475 /* start garbage collection */
1476 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1477 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1478 arp_tbl.gc_interval);
1483 static void neigh_hash_free_rcu(struct rcu_head *head)
1485 struct ipoib_neigh_hash *htbl = container_of(head,
1486 struct ipoib_neigh_hash,
1488 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1489 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1493 complete(&ntbl->deleted);
1496 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1498 struct ipoib_dev_priv *priv = netdev_priv(dev);
1499 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1500 struct ipoib_neigh_hash *htbl;
1501 unsigned long flags;
1504 /* remove all neigh connected to a given path or mcast */
1505 spin_lock_irqsave(&priv->lock, flags);
1507 htbl = rcu_dereference_protected(ntbl->htbl,
1508 lockdep_is_held(&priv->lock));
1513 for (i = 0; i < htbl->size; i++) {
1514 struct ipoib_neigh *neigh;
1515 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1517 while ((neigh = rcu_dereference_protected(*np,
1518 lockdep_is_held(&priv->lock))) != NULL) {
1519 /* delete neighs belong to this parent */
1520 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1521 rcu_assign_pointer(*np,
1522 rcu_dereference_protected(neigh->hnext,
1523 lockdep_is_held(&priv->lock)));
1524 /* remove from parent list */
1525 list_del(&neigh->list);
1526 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1534 spin_unlock_irqrestore(&priv->lock, flags);
1537 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1539 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1540 struct ipoib_neigh_hash *htbl;
1541 unsigned long flags;
1542 int i, wait_flushed = 0;
1544 init_completion(&priv->ntbl.flushed);
1546 spin_lock_irqsave(&priv->lock, flags);
1548 htbl = rcu_dereference_protected(ntbl->htbl,
1549 lockdep_is_held(&priv->lock));
1553 wait_flushed = atomic_read(&priv->ntbl.entries);
1557 for (i = 0; i < htbl->size; i++) {
1558 struct ipoib_neigh *neigh;
1559 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1561 while ((neigh = rcu_dereference_protected(*np,
1562 lockdep_is_held(&priv->lock))) != NULL) {
1563 rcu_assign_pointer(*np,
1564 rcu_dereference_protected(neigh->hnext,
1565 lockdep_is_held(&priv->lock)));
1566 /* remove from path/mc list */
1567 list_del(&neigh->list);
1568 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1573 rcu_assign_pointer(ntbl->htbl, NULL);
1574 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1577 spin_unlock_irqrestore(&priv->lock, flags);
1579 wait_for_completion(&priv->ntbl.flushed);
1582 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1584 struct ipoib_dev_priv *priv = netdev_priv(dev);
1587 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1588 init_completion(&priv->ntbl.deleted);
1589 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1591 /* Stop GC if called at init fail need to cancel work */
1592 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1594 cancel_delayed_work(&priv->neigh_reap_task);
1596 ipoib_flush_neighs(priv);
1598 wait_for_completion(&priv->ntbl.deleted);
1602 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1604 struct ipoib_dev_priv *priv = netdev_priv(dev);
1606 /* Allocate RX/TX "rings" to hold queued skbs */
1607 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1609 if (!priv->rx_ring) {
1610 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1611 ca->name, ipoib_recvq_size);
1615 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1616 if (!priv->tx_ring) {
1617 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1618 ca->name, ipoib_sendq_size);
1619 goto out_rx_ring_cleanup;
1622 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1624 if (ipoib_ib_dev_init(dev, ca, port))
1625 goto out_tx_ring_cleanup;
1628 * Must be after ipoib_ib_dev_init so we can allocate a per
1629 * device wq there and use it here
1631 if (ipoib_neigh_hash_init(priv) < 0)
1632 goto out_dev_uninit;
1637 ipoib_ib_dev_cleanup(dev);
1639 out_tx_ring_cleanup:
1640 vfree(priv->tx_ring);
1642 out_rx_ring_cleanup:
1643 kfree(priv->rx_ring);
1649 void ipoib_dev_cleanup(struct net_device *dev)
1651 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1656 ipoib_delete_debug_files(dev);
1658 /* Delete any child interfaces first */
1659 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1660 /* Stop GC on child */
1661 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1662 cancel_delayed_work(&cpriv->neigh_reap_task);
1663 unregister_netdevice_queue(cpriv->dev, &head);
1665 unregister_netdevice_many(&head);
1668 * Must be before ipoib_ib_dev_cleanup or we delete an in use
1671 ipoib_neigh_hash_uninit(dev);
1673 ipoib_ib_dev_cleanup(dev);
1675 kfree(priv->rx_ring);
1676 vfree(priv->tx_ring);
1678 priv->rx_ring = NULL;
1679 priv->tx_ring = NULL;
1682 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1684 struct ipoib_dev_priv *priv = netdev_priv(dev);
1686 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1689 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1690 struct ifla_vf_info *ivf)
1692 struct ipoib_dev_priv *priv = netdev_priv(dev);
1695 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1704 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1706 struct ipoib_dev_priv *priv = netdev_priv(dev);
1708 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1711 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1714 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1715 struct ifla_vf_stats *vf_stats)
1717 struct ipoib_dev_priv *priv = netdev_priv(dev);
1719 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1722 static const struct header_ops ipoib_header_ops = {
1723 .create = ipoib_hard_header,
1726 static const struct net_device_ops ipoib_netdev_ops_pf = {
1727 .ndo_uninit = ipoib_uninit,
1728 .ndo_open = ipoib_open,
1729 .ndo_stop = ipoib_stop,
1730 .ndo_change_mtu = ipoib_change_mtu,
1731 .ndo_fix_features = ipoib_fix_features,
1732 .ndo_start_xmit = ipoib_start_xmit,
1733 .ndo_tx_timeout = ipoib_timeout,
1734 .ndo_set_rx_mode = ipoib_set_mcast_list,
1735 .ndo_get_iflink = ipoib_get_iflink,
1736 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
1737 .ndo_get_vf_config = ipoib_get_vf_config,
1738 .ndo_get_vf_stats = ipoib_get_vf_stats,
1739 .ndo_set_vf_guid = ipoib_set_vf_guid,
1740 .ndo_set_mac_address = ipoib_set_mac,
1743 static const struct net_device_ops ipoib_netdev_ops_vf = {
1744 .ndo_uninit = ipoib_uninit,
1745 .ndo_open = ipoib_open,
1746 .ndo_stop = ipoib_stop,
1747 .ndo_change_mtu = ipoib_change_mtu,
1748 .ndo_fix_features = ipoib_fix_features,
1749 .ndo_start_xmit = ipoib_start_xmit,
1750 .ndo_tx_timeout = ipoib_timeout,
1751 .ndo_set_rx_mode = ipoib_set_mcast_list,
1752 .ndo_get_iflink = ipoib_get_iflink,
1755 void ipoib_setup(struct net_device *dev)
1757 struct ipoib_dev_priv *priv = netdev_priv(dev);
1759 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1760 dev->netdev_ops = &ipoib_netdev_ops_vf;
1762 dev->netdev_ops = &ipoib_netdev_ops_pf;
1764 dev->header_ops = &ipoib_header_ops;
1766 ipoib_set_ethtool_ops(dev);
1768 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1770 dev->watchdog_timeo = HZ;
1772 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1774 dev->hard_header_len = IPOIB_HARD_LEN;
1775 dev->addr_len = INFINIBAND_ALEN;
1776 dev->type = ARPHRD_INFINIBAND;
1777 dev->tx_queue_len = ipoib_sendq_size * 2;
1778 dev->features = (NETIF_F_VLAN_CHALLENGED |
1780 netif_keep_dst(dev);
1782 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1786 spin_lock_init(&priv->lock);
1788 init_rwsem(&priv->vlan_rwsem);
1790 INIT_LIST_HEAD(&priv->path_list);
1791 INIT_LIST_HEAD(&priv->child_intfs);
1792 INIT_LIST_HEAD(&priv->dead_ahs);
1793 INIT_LIST_HEAD(&priv->multicast_list);
1795 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1796 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1797 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1798 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1799 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1800 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1801 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1802 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1805 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1807 struct net_device *dev;
1809 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1810 NET_NAME_UNKNOWN, ipoib_setup);
1814 return netdev_priv(dev);
1817 static ssize_t show_pkey(struct device *dev,
1818 struct device_attribute *attr, char *buf)
1820 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1822 return sprintf(buf, "0x%04x\n", priv->pkey);
1824 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1826 static ssize_t show_umcast(struct device *dev,
1827 struct device_attribute *attr, char *buf)
1829 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1831 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1834 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1836 struct ipoib_dev_priv *priv = netdev_priv(ndev);
1838 if (umcast_val > 0) {
1839 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1840 ipoib_warn(priv, "ignoring multicast groups joined directly "
1843 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1846 static ssize_t set_umcast(struct device *dev,
1847 struct device_attribute *attr,
1848 const char *buf, size_t count)
1850 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1852 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1856 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1858 int ipoib_add_umcast_attr(struct net_device *dev)
1860 return device_create_file(&dev->dev, &dev_attr_umcast);
1863 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
1865 struct ipoib_dev_priv *child_priv;
1866 struct net_device *netdev = priv->dev;
1868 netif_addr_lock_bh(netdev);
1870 memcpy(&priv->local_gid.global.interface_id,
1871 &gid->global.interface_id,
1872 sizeof(gid->global.interface_id));
1873 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
1874 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1876 netif_addr_unlock_bh(netdev);
1878 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1879 down_read(&priv->vlan_rwsem);
1880 list_for_each_entry(child_priv, &priv->child_intfs, list)
1881 set_base_guid(child_priv, gid);
1882 up_read(&priv->vlan_rwsem);
1886 static int ipoib_check_lladdr(struct net_device *dev,
1887 struct sockaddr_storage *ss)
1889 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
1892 netif_addr_lock_bh(dev);
1894 /* Make sure the QPN, reserved and subnet prefix match the current
1895 * lladdr, it also makes sure the lladdr is unicast.
1897 if (memcmp(dev->dev_addr, ss->__data,
1898 4 + sizeof(gid->global.subnet_prefix)) ||
1899 gid->global.interface_id == 0)
1902 netif_addr_unlock_bh(dev);
1907 static int ipoib_set_mac(struct net_device *dev, void *addr)
1909 struct ipoib_dev_priv *priv = netdev_priv(dev);
1910 struct sockaddr_storage *ss = addr;
1913 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
1916 ret = ipoib_check_lladdr(dev, ss);
1920 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
1922 queue_work(ipoib_workqueue, &priv->flush_light);
1927 static ssize_t create_child(struct device *dev,
1928 struct device_attribute *attr,
1929 const char *buf, size_t count)
1934 if (sscanf(buf, "%i", &pkey) != 1)
1937 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1941 * Set the full membership bit, so that we join the right
1942 * broadcast group, etc.
1946 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1948 return ret ? ret : count;
1950 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1952 static ssize_t delete_child(struct device *dev,
1953 struct device_attribute *attr,
1954 const char *buf, size_t count)
1959 if (sscanf(buf, "%i", &pkey) != 1)
1962 if (pkey < 0 || pkey > 0xffff)
1965 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1967 return ret ? ret : count;
1970 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1972 int ipoib_add_pkey_attr(struct net_device *dev)
1974 return device_create_file(&dev->dev, &dev_attr_pkey);
1977 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1979 priv->hca_caps = hca->attrs.device_cap_flags;
1981 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1982 priv->dev->hw_features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1984 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1985 priv->dev->hw_features |= NETIF_F_TSO;
1987 priv->dev->features |= priv->dev->hw_features;
1993 static struct net_device *ipoib_add_port(const char *format,
1994 struct ib_device *hca, u8 port)
1996 struct ipoib_dev_priv *priv;
1997 struct ib_port_attr attr;
1998 int result = -ENOMEM;
2000 priv = ipoib_intf_alloc(format);
2002 goto alloc_mem_failed;
2004 SET_NETDEV_DEV(priv->dev, hca->dma_device);
2005 priv->dev->dev_id = port - 1;
2007 result = ib_query_port(hca, port, &attr);
2009 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2011 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2013 goto device_init_failed;
2016 /* MTU will be reset when mcast join happens */
2017 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
2018 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
2020 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2022 result = ib_query_pkey(hca, port, 0, &priv->pkey);
2024 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2025 hca->name, port, result);
2026 goto device_init_failed;
2029 result = ipoib_set_dev_features(priv, hca);
2031 goto device_init_failed;
2034 * Set the full membership bit, so that we join the right
2035 * broadcast group, etc.
2037 priv->pkey |= 0x8000;
2039 priv->dev->broadcast[8] = priv->pkey >> 8;
2040 priv->dev->broadcast[9] = priv->pkey & 0xff;
2042 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2044 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2045 hca->name, port, result);
2046 goto device_init_failed;
2048 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2049 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2051 result = ipoib_dev_init(priv->dev, hca, port);
2053 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2054 hca->name, port, result);
2055 goto device_init_failed;
2058 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2059 priv->ca, ipoib_event);
2060 result = ib_register_event_handler(&priv->event_handler);
2062 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2063 "port %d (ret = %d)\n",
2064 hca->name, port, result);
2068 result = register_netdev(priv->dev);
2070 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2071 hca->name, port, result);
2072 goto register_failed;
2075 ipoib_create_debug_files(priv->dev);
2077 if (ipoib_cm_add_mode_attr(priv->dev))
2079 if (ipoib_add_pkey_attr(priv->dev))
2081 if (ipoib_add_umcast_attr(priv->dev))
2083 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2085 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2091 ipoib_delete_debug_files(priv->dev);
2092 unregister_netdev(priv->dev);
2095 ib_unregister_event_handler(&priv->event_handler);
2096 flush_workqueue(ipoib_workqueue);
2097 /* Stop GC if started before flush */
2098 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2099 cancel_delayed_work(&priv->neigh_reap_task);
2100 flush_workqueue(priv->wq);
2103 ipoib_dev_cleanup(priv->dev);
2106 free_netdev(priv->dev);
2109 return ERR_PTR(result);
2112 static void ipoib_add_one(struct ib_device *device)
2114 struct list_head *dev_list;
2115 struct net_device *dev;
2116 struct ipoib_dev_priv *priv;
2120 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2124 INIT_LIST_HEAD(dev_list);
2126 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2127 if (!rdma_protocol_ib(device, p))
2129 dev = ipoib_add_port("ib%d", device, p);
2131 priv = netdev_priv(dev);
2132 list_add_tail(&priv->list, dev_list);
2142 ib_set_client_data(device, &ipoib_client, dev_list);
2145 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2147 struct ipoib_dev_priv *priv, *tmp;
2148 struct list_head *dev_list = client_data;
2153 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2154 ib_unregister_event_handler(&priv->event_handler);
2155 flush_workqueue(ipoib_workqueue);
2157 /* mark interface in the middle of destruction */
2158 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2161 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2165 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2166 cancel_delayed_work(&priv->neigh_reap_task);
2167 flush_workqueue(priv->wq);
2169 unregister_netdev(priv->dev);
2170 free_netdev(priv->dev);
2176 static int __init ipoib_init_module(void)
2180 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2181 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2182 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2184 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2185 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2186 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2187 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2188 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2192 * When copying small received packets, we only copy from the
2193 * linear data part of the SKB, so we rely on this condition.
2195 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2197 ret = ipoib_register_debugfs();
2202 * We create a global workqueue here that is used for all flush
2203 * operations. However, if you attempt to flush a workqueue
2204 * from a task on that same workqueue, it deadlocks the system.
2205 * We want to be able to flush the tasks associated with a
2206 * specific net device, so we also create a workqueue for each
2207 * netdevice. We queue up the tasks for that device only on
2208 * its private workqueue, and we only queue up flush events
2209 * on our global flush workqueue. This avoids the deadlocks.
2211 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2213 if (!ipoib_workqueue) {
2218 ib_sa_register_client(&ipoib_sa_client);
2220 ret = ib_register_client(&ipoib_client);
2224 ret = ipoib_netlink_init();
2231 ib_unregister_client(&ipoib_client);
2234 ib_sa_unregister_client(&ipoib_sa_client);
2235 destroy_workqueue(ipoib_workqueue);
2238 ipoib_unregister_debugfs();
2243 static void __exit ipoib_cleanup_module(void)
2245 ipoib_netlink_fini();
2246 ib_unregister_client(&ipoib_client);
2247 ib_sa_unregister_client(&ipoib_sa_client);
2248 ipoib_unregister_debugfs();
2249 destroy_workqueue(ipoib_workqueue);
2252 module_init(ipoib_init_module);
2253 module_exit(ipoib_cleanup_module);