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 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
112 static int ipoib_netdev_event(struct notifier_block *this,
113 unsigned long event, void *ptr)
115 struct netdev_notifier_info *ni = ptr;
116 struct net_device *dev = ni->dev;
118 if (dev->netdev_ops->ndo_open != ipoib_open)
122 case NETDEV_REGISTER:
123 ipoib_create_debug_files(dev);
125 case NETDEV_CHANGENAME:
126 ipoib_delete_debug_files(dev);
127 ipoib_create_debug_files(dev);
129 case NETDEV_UNREGISTER:
130 ipoib_delete_debug_files(dev);
138 int ipoib_open(struct net_device *dev)
140 struct ipoib_dev_priv *priv = ipoib_priv(dev);
142 ipoib_dbg(priv, "bringing up interface\n");
144 netif_carrier_off(dev);
146 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
148 priv->sm_fullmember_sendonly_support = false;
150 if (ipoib_ib_dev_open(dev)) {
151 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
156 ipoib_ib_dev_up(dev);
158 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
159 struct ipoib_dev_priv *cpriv;
161 /* Bring up any child interfaces too */
162 down_read(&priv->vlan_rwsem);
163 list_for_each_entry(cpriv, &priv->child_intfs, list) {
166 flags = cpriv->dev->flags;
170 dev_change_flags(cpriv->dev, flags | IFF_UP);
172 up_read(&priv->vlan_rwsem);
175 netif_start_queue(dev);
180 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
185 static int ipoib_stop(struct net_device *dev)
187 struct ipoib_dev_priv *priv = ipoib_priv(dev);
189 ipoib_dbg(priv, "stopping interface\n");
191 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
193 netif_stop_queue(dev);
195 ipoib_ib_dev_down(dev);
196 ipoib_ib_dev_stop(dev);
198 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
199 struct ipoib_dev_priv *cpriv;
201 /* Bring down any child interfaces too */
202 down_read(&priv->vlan_rwsem);
203 list_for_each_entry(cpriv, &priv->child_intfs, list) {
206 flags = cpriv->dev->flags;
207 if (!(flags & IFF_UP))
210 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
212 up_read(&priv->vlan_rwsem);
218 static void ipoib_uninit(struct net_device *dev)
220 ipoib_dev_cleanup(dev);
223 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
225 struct ipoib_dev_priv *priv = ipoib_priv(dev);
227 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
228 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
233 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
235 struct ipoib_dev_priv *priv = ipoib_priv(dev);
237 /* dev->mtu > 2K ==> connected mode */
238 if (ipoib_cm_admin_enabled(dev)) {
239 if (new_mtu > ipoib_cm_max_mtu(dev))
242 if (new_mtu > priv->mcast_mtu)
243 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
250 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
253 priv->admin_mtu = new_mtu;
255 if (priv->mcast_mtu < priv->admin_mtu)
256 ipoib_dbg(priv, "MTU must be smaller than the underlying "
257 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
259 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
264 /* Called with an RCU read lock taken */
265 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
266 struct net_device *dev)
268 struct net *net = dev_net(dev);
269 struct in_device *in_dev;
270 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
271 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
274 switch (addr->sa_family) {
276 in_dev = in_dev_get(dev);
280 ret_addr = inet_confirm_addr(net, in_dev, 0,
281 addr_in->sin_addr.s_addr,
289 if (IS_ENABLED(CONFIG_IPV6) &&
290 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
299 * Find the master net_device on top of the given net_device.
300 * @dev: base IPoIB net_device
302 * Returns the master net_device with a reference held, or the same net_device
303 * if no master exists.
305 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
307 struct net_device *master;
310 master = netdev_master_upper_dev_get_rcu(dev);
322 struct ipoib_walk_data {
323 const struct sockaddr *addr;
324 struct net_device *result;
327 static int ipoib_upper_walk(struct net_device *upper, void *_data)
329 struct ipoib_walk_data *data = _data;
332 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
334 data->result = upper;
342 * Find a net_device matching the given address, which is an upper device of
343 * the given net_device.
344 * @addr: IP address to look for.
345 * @dev: base IPoIB net_device
347 * If found, returns the net_device with a reference held. Otherwise return
350 static struct net_device *ipoib_get_net_dev_match_addr(
351 const struct sockaddr *addr, struct net_device *dev)
353 struct ipoib_walk_data data = {
358 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
364 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
370 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
371 * pkey_index and address, if one exists.
373 * @found_net_dev: contains a matching net_device if the return value >= 1,
374 * with a reference held. */
375 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
376 const union ib_gid *gid,
378 const struct sockaddr *addr,
380 struct net_device **found_net_dev)
382 struct ipoib_dev_priv *child_priv;
383 struct net_device *net_dev = NULL;
386 if (priv->pkey_index == pkey_index &&
387 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
389 net_dev = ipoib_get_master_net_dev(priv->dev);
391 /* Verify the net_device matches the IP address, as
392 * IPoIB child devices currently share a GID. */
393 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
397 *found_net_dev = net_dev;
404 /* Check child interfaces */
405 down_read_nested(&priv->vlan_rwsem, nesting);
406 list_for_each_entry(child_priv, &priv->child_intfs, list) {
407 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
414 up_read(&priv->vlan_rwsem);
419 /* Returns the number of matching net_devs found (between 0 and 2). Also
420 * return the matching net_device in the @net_dev parameter, holding a
421 * reference to the net_device, if the number of matches >= 1 */
422 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
424 const union ib_gid *gid,
425 const struct sockaddr *addr,
426 struct net_device **net_dev)
428 struct ipoib_dev_priv *priv;
433 list_for_each_entry(priv, dev_list, list) {
434 if (priv->port != port)
437 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
446 static struct net_device *ipoib_get_net_dev_by_params(
447 struct ib_device *dev, u8 port, u16 pkey,
448 const union ib_gid *gid, const struct sockaddr *addr,
451 struct net_device *net_dev;
452 struct list_head *dev_list = client_data;
457 if (!rdma_protocol_ib(dev, port))
460 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
467 /* See if we can find a unique device matching the L2 parameters */
468 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
469 gid, NULL, &net_dev);
480 /* Couldn't find a unique device with L2 parameters only. Use L3
481 * address to uniquely match the net device */
482 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
483 gid, addr, &net_dev);
488 dev_warn_ratelimited(&dev->dev,
489 "duplicate IP address detected\n");
496 int ipoib_set_mode(struct net_device *dev, const char *buf)
498 struct ipoib_dev_priv *priv = ipoib_priv(dev);
500 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
501 !strcmp(buf, "connected\n")) ||
502 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
503 !strcmp(buf, "datagram\n"))) {
507 /* flush paths if we switch modes so that connections are restarted */
508 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
509 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
510 ipoib_warn(priv, "enabling connected mode "
511 "will cause multicast packet drops\n");
512 netdev_update_features(dev);
513 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
515 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
517 ipoib_flush_paths(dev);
518 return (!rtnl_trylock()) ? -EBUSY : 0;
521 if (!strcmp(buf, "datagram\n")) {
522 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
523 netdev_update_features(dev);
524 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
526 ipoib_flush_paths(dev);
527 return (!rtnl_trylock()) ? -EBUSY : 0;
533 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
535 struct ipoib_dev_priv *priv = ipoib_priv(dev);
536 struct rb_node *n = priv->path_tree.rb_node;
537 struct ipoib_path *path;
541 path = rb_entry(n, struct ipoib_path, rb_node);
543 ret = memcmp(gid, path->pathrec.dgid.raw,
544 sizeof (union ib_gid));
557 static int __path_add(struct net_device *dev, struct ipoib_path *path)
559 struct ipoib_dev_priv *priv = ipoib_priv(dev);
560 struct rb_node **n = &priv->path_tree.rb_node;
561 struct rb_node *pn = NULL;
562 struct ipoib_path *tpath;
567 tpath = rb_entry(pn, struct ipoib_path, rb_node);
569 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
570 sizeof (union ib_gid));
579 rb_link_node(&path->rb_node, pn, n);
580 rb_insert_color(&path->rb_node, &priv->path_tree);
582 list_add_tail(&path->list, &priv->path_list);
587 static void path_free(struct net_device *dev, struct ipoib_path *path)
591 while ((skb = __skb_dequeue(&path->queue)))
592 dev_kfree_skb_irq(skb);
594 ipoib_dbg(ipoib_priv(dev), "path_free\n");
596 /* remove all neigh connected to this path */
597 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
600 ipoib_put_ah(path->ah);
605 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
607 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
609 struct ipoib_path_iter *iter;
611 iter = kmalloc(sizeof *iter, GFP_KERNEL);
616 memset(iter->path.pathrec.dgid.raw, 0, 16);
618 if (ipoib_path_iter_next(iter)) {
626 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
628 struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
630 struct ipoib_path *path;
633 spin_lock_irq(&priv->lock);
635 n = rb_first(&priv->path_tree);
638 path = rb_entry(n, struct ipoib_path, rb_node);
640 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
641 sizeof (union ib_gid)) < 0) {
650 spin_unlock_irq(&priv->lock);
655 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
656 struct ipoib_path *path)
661 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
663 void ipoib_mark_paths_invalid(struct net_device *dev)
665 struct ipoib_dev_priv *priv = ipoib_priv(dev);
666 struct ipoib_path *path, *tp;
668 spin_lock_irq(&priv->lock);
670 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
671 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
672 be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
673 path->pathrec.dgid.raw);
677 spin_unlock_irq(&priv->lock);
680 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
682 struct ipoib_pseudo_header *phdr;
684 phdr = (struct ipoib_pseudo_header *)skb_push(skb, sizeof(*phdr));
685 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
688 void ipoib_flush_paths(struct net_device *dev)
690 struct ipoib_dev_priv *priv = ipoib_priv(dev);
691 struct ipoib_path *path, *tp;
692 LIST_HEAD(remove_list);
695 netif_tx_lock_bh(dev);
696 spin_lock_irqsave(&priv->lock, flags);
698 list_splice_init(&priv->path_list, &remove_list);
700 list_for_each_entry(path, &remove_list, list)
701 rb_erase(&path->rb_node, &priv->path_tree);
703 list_for_each_entry_safe(path, tp, &remove_list, list) {
705 ib_sa_cancel_query(path->query_id, path->query);
706 spin_unlock_irqrestore(&priv->lock, flags);
707 netif_tx_unlock_bh(dev);
708 wait_for_completion(&path->done);
709 path_free(dev, path);
710 netif_tx_lock_bh(dev);
711 spin_lock_irqsave(&priv->lock, flags);
714 spin_unlock_irqrestore(&priv->lock, flags);
715 netif_tx_unlock_bh(dev);
718 static void path_rec_completion(int status,
719 struct sa_path_rec *pathrec,
722 struct ipoib_path *path = path_ptr;
723 struct net_device *dev = path->dev;
724 struct ipoib_dev_priv *priv = ipoib_priv(dev);
725 struct ipoib_ah *ah = NULL;
726 struct ipoib_ah *old_ah = NULL;
727 struct ipoib_neigh *neigh, *tn;
728 struct sk_buff_head skqueue;
733 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
734 be32_to_cpu(sa_path_get_dlid(pathrec)),
737 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
738 status, path->pathrec.dgid.raw);
740 skb_queue_head_init(&skqueue);
743 struct rdma_ah_attr av;
745 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
746 ah = ipoib_create_ah(dev, priv->pd, &av);
749 spin_lock_irqsave(&priv->lock, flags);
751 if (!IS_ERR_OR_NULL(ah)) {
752 path->pathrec = *pathrec;
757 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
758 ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
761 while ((skb = __skb_dequeue(&path->queue)))
762 __skb_queue_tail(&skqueue, skb);
764 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
766 WARN_ON(neigh->ah != old_ah);
768 * Dropping the ah reference inside
769 * priv->lock is safe here, because we
770 * will hold one more reference from
771 * the original value of path->ah (ie
774 ipoib_put_ah(neigh->ah);
776 kref_get(&path->ah->ref);
777 neigh->ah = path->ah;
779 if (ipoib_cm_enabled(dev, neigh->daddr)) {
780 if (!ipoib_cm_get(neigh))
781 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
784 if (!ipoib_cm_get(neigh)) {
785 ipoib_neigh_free(neigh);
790 while ((skb = __skb_dequeue(&neigh->queue)))
791 __skb_queue_tail(&skqueue, skb);
797 complete(&path->done);
799 spin_unlock_irqrestore(&priv->lock, flags);
801 if (IS_ERR_OR_NULL(ah))
802 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
805 ipoib_put_ah(old_ah);
807 while ((skb = __skb_dequeue(&skqueue))) {
810 ret = dev_queue_xmit(skb);
812 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
817 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
819 struct ipoib_dev_priv *priv = ipoib_priv(dev);
820 struct ipoib_path *path;
822 if (!priv->broadcast)
825 path = kzalloc(sizeof *path, GFP_ATOMIC);
831 skb_queue_head_init(&path->queue);
833 INIT_LIST_HEAD(&path->neigh_list);
835 if (rdma_cap_opa_ah(priv->ca, priv->port))
836 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
838 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
839 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
840 path->pathrec.sgid = priv->local_gid;
841 path->pathrec.pkey = cpu_to_be16(priv->pkey);
842 path->pathrec.numb_path = 1;
843 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
848 static int path_rec_start(struct net_device *dev,
849 struct ipoib_path *path)
851 struct ipoib_dev_priv *priv = ipoib_priv(dev);
853 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
854 path->pathrec.dgid.raw);
856 init_completion(&path->done);
859 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
861 IB_SA_PATH_REC_DGID |
862 IB_SA_PATH_REC_SGID |
863 IB_SA_PATH_REC_NUMB_PATH |
864 IB_SA_PATH_REC_TRAFFIC_CLASS |
869 if (path->query_id < 0) {
870 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
872 complete(&path->done);
873 return path->query_id;
879 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
880 struct net_device *dev)
882 struct ipoib_dev_priv *priv = ipoib_priv(dev);
883 struct rdma_netdev *rn = netdev_priv(dev);
884 struct ipoib_path *path;
885 struct ipoib_neigh *neigh;
888 spin_lock_irqsave(&priv->lock, flags);
889 neigh = ipoib_neigh_alloc(daddr, dev);
891 spin_unlock_irqrestore(&priv->lock, flags);
892 ++dev->stats.tx_dropped;
893 dev_kfree_skb_any(skb);
897 path = __path_find(dev, daddr + 4);
899 path = path_rec_create(dev, daddr + 4);
903 __path_add(dev, path);
906 list_add_tail(&neigh->list, &path->neigh_list);
909 kref_get(&path->ah->ref);
910 neigh->ah = path->ah;
912 if (ipoib_cm_enabled(dev, neigh->daddr)) {
913 if (!ipoib_cm_get(neigh))
914 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
915 if (!ipoib_cm_get(neigh)) {
916 ipoib_neigh_free(neigh);
919 if (skb_queue_len(&neigh->queue) <
920 IPOIB_MAX_PATH_REC_QUEUE) {
921 push_pseudo_header(skb, neigh->daddr);
922 __skb_queue_tail(&neigh->queue, skb);
924 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
925 skb_queue_len(&neigh->queue));
929 spin_unlock_irqrestore(&priv->lock, flags);
930 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
932 ipoib_neigh_put(neigh);
938 if (!path->query && path_rec_start(dev, path))
940 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
941 push_pseudo_header(skb, neigh->daddr);
942 __skb_queue_tail(&neigh->queue, skb);
948 spin_unlock_irqrestore(&priv->lock, flags);
949 ipoib_neigh_put(neigh);
953 ipoib_neigh_free(neigh);
955 ++dev->stats.tx_dropped;
956 dev_kfree_skb_any(skb);
958 spin_unlock_irqrestore(&priv->lock, flags);
959 ipoib_neigh_put(neigh);
962 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
963 struct ipoib_pseudo_header *phdr)
965 struct ipoib_dev_priv *priv = ipoib_priv(dev);
966 struct rdma_netdev *rn = netdev_priv(dev);
967 struct ipoib_path *path;
970 spin_lock_irqsave(&priv->lock, flags);
972 path = __path_find(dev, phdr->hwaddr + 4);
973 if (!path || !path->valid) {
977 path = path_rec_create(dev, phdr->hwaddr + 4);
981 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
982 push_pseudo_header(skb, phdr->hwaddr);
983 __skb_queue_tail(&path->queue, skb);
985 ++dev->stats.tx_dropped;
986 dev_kfree_skb_any(skb);
989 if (!path->query && path_rec_start(dev, path)) {
990 spin_unlock_irqrestore(&priv->lock, flags);
992 path_free(dev, path);
995 __path_add(dev, path);
997 ++dev->stats.tx_dropped;
998 dev_kfree_skb_any(skb);
1001 spin_unlock_irqrestore(&priv->lock, flags);
1006 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1007 be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1009 spin_unlock_irqrestore(&priv->lock, flags);
1010 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1011 IPOIB_QPN(phdr->hwaddr));
1013 } else if ((path->query || !path_rec_start(dev, path)) &&
1014 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1015 push_pseudo_header(skb, phdr->hwaddr);
1016 __skb_queue_tail(&path->queue, skb);
1018 ++dev->stats.tx_dropped;
1019 dev_kfree_skb_any(skb);
1022 spin_unlock_irqrestore(&priv->lock, flags);
1025 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1027 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1028 struct rdma_netdev *rn = netdev_priv(dev);
1029 struct ipoib_neigh *neigh;
1030 struct ipoib_pseudo_header *phdr;
1031 struct ipoib_header *header;
1032 unsigned long flags;
1034 phdr = (struct ipoib_pseudo_header *) skb->data;
1035 skb_pull(skb, sizeof(*phdr));
1036 header = (struct ipoib_header *) skb->data;
1038 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1039 /* multicast, arrange "if" according to probability */
1040 if ((header->proto != htons(ETH_P_IP)) &&
1041 (header->proto != htons(ETH_P_IPV6)) &&
1042 (header->proto != htons(ETH_P_ARP)) &&
1043 (header->proto != htons(ETH_P_RARP)) &&
1044 (header->proto != htons(ETH_P_TIPC))) {
1045 /* ethertype not supported by IPoIB */
1046 ++dev->stats.tx_dropped;
1047 dev_kfree_skb_any(skb);
1048 return NETDEV_TX_OK;
1050 /* Add in the P_Key for multicast*/
1051 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1052 phdr->hwaddr[9] = priv->pkey & 0xff;
1054 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1056 goto send_using_neigh;
1057 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1058 return NETDEV_TX_OK;
1061 /* unicast, arrange "switch" according to probability */
1062 switch (header->proto) {
1063 case htons(ETH_P_IP):
1064 case htons(ETH_P_IPV6):
1065 case htons(ETH_P_TIPC):
1066 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1067 if (unlikely(!neigh)) {
1068 neigh_add_path(skb, phdr->hwaddr, dev);
1069 return NETDEV_TX_OK;
1072 case htons(ETH_P_ARP):
1073 case htons(ETH_P_RARP):
1074 /* for unicast ARP and RARP should always perform path find */
1075 unicast_arp_send(skb, dev, phdr);
1076 return NETDEV_TX_OK;
1078 /* ethertype not supported by IPoIB */
1079 ++dev->stats.tx_dropped;
1080 dev_kfree_skb_any(skb);
1081 return NETDEV_TX_OK;
1085 /* note we now hold a ref to neigh */
1086 if (ipoib_cm_get(neigh)) {
1087 if (ipoib_cm_up(neigh)) {
1088 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1091 } else if (neigh->ah) {
1092 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1093 IPOIB_QPN(phdr->hwaddr));
1097 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1098 push_pseudo_header(skb, phdr->hwaddr);
1099 spin_lock_irqsave(&priv->lock, flags);
1100 __skb_queue_tail(&neigh->queue, skb);
1101 spin_unlock_irqrestore(&priv->lock, flags);
1103 ++dev->stats.tx_dropped;
1104 dev_kfree_skb_any(skb);
1108 ipoib_neigh_put(neigh);
1110 return NETDEV_TX_OK;
1113 static void ipoib_timeout(struct net_device *dev)
1115 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1117 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1118 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1119 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1120 netif_queue_stopped(dev),
1121 priv->tx_head, priv->tx_tail);
1122 /* XXX reset QP, etc. */
1125 static int ipoib_hard_header(struct sk_buff *skb,
1126 struct net_device *dev,
1127 unsigned short type,
1128 const void *daddr, const void *saddr, unsigned len)
1130 struct ipoib_header *header;
1132 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
1134 header->proto = htons(type);
1135 header->reserved = 0;
1138 * we don't rely on dst_entry structure, always stuff the
1139 * destination address into skb hard header so we can figure out where
1140 * to send the packet later.
1142 push_pseudo_header(skb, daddr);
1144 return IPOIB_HARD_LEN;
1147 static void ipoib_set_mcast_list(struct net_device *dev)
1149 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1151 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1152 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1156 queue_work(priv->wq, &priv->restart_task);
1159 static int ipoib_get_iflink(const struct net_device *dev)
1161 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1163 /* parent interface */
1164 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1165 return dev->ifindex;
1167 /* child/vlan interface */
1168 return priv->parent->ifindex;
1171 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1174 * Use only the address parts that contributes to spreading
1175 * The subnet prefix is not used as one can not connect to
1176 * same remote port (GUID) using the same remote QPN via two
1177 * different subnets.
1179 /* qpn octets[1:4) & port GUID octets[12:20) */
1180 u32 *d32 = (u32 *) daddr;
1183 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1184 return hv & htbl->mask;
1187 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1189 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1190 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1191 struct ipoib_neigh_hash *htbl;
1192 struct ipoib_neigh *neigh = NULL;
1197 htbl = rcu_dereference_bh(ntbl->htbl);
1202 hash_val = ipoib_addr_hash(htbl, daddr);
1203 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1205 neigh = rcu_dereference_bh(neigh->hnext)) {
1206 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1207 /* found, take one ref on behalf of the caller */
1208 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1214 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1215 neigh->alive = jiffies;
1221 rcu_read_unlock_bh();
1225 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1227 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1228 struct ipoib_neigh_hash *htbl;
1229 unsigned long neigh_obsolete;
1231 unsigned long flags;
1233 LIST_HEAD(remove_list);
1235 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1238 spin_lock_irqsave(&priv->lock, flags);
1240 htbl = rcu_dereference_protected(ntbl->htbl,
1241 lockdep_is_held(&priv->lock));
1246 /* neigh is obsolete if it was idle for two GC periods */
1247 dt = 2 * arp_tbl.gc_interval;
1248 neigh_obsolete = jiffies - dt;
1249 /* handle possible race condition */
1250 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1253 for (i = 0; i < htbl->size; i++) {
1254 struct ipoib_neigh *neigh;
1255 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1257 while ((neigh = rcu_dereference_protected(*np,
1258 lockdep_is_held(&priv->lock))) != NULL) {
1259 /* was the neigh idle for two GC periods */
1260 if (time_after(neigh_obsolete, neigh->alive)) {
1262 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1264 rcu_assign_pointer(*np,
1265 rcu_dereference_protected(neigh->hnext,
1266 lockdep_is_held(&priv->lock)));
1267 /* remove from path/mc list */
1268 list_del_init(&neigh->list);
1269 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1278 spin_unlock_irqrestore(&priv->lock, flags);
1279 ipoib_mcast_remove_list(&remove_list);
1282 static void ipoib_reap_neigh(struct work_struct *work)
1284 struct ipoib_dev_priv *priv =
1285 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1287 __ipoib_reap_neigh(priv);
1289 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
1290 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1291 arp_tbl.gc_interval);
1295 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1296 struct net_device *dev)
1298 struct ipoib_neigh *neigh;
1300 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
1305 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1306 skb_queue_head_init(&neigh->queue);
1307 INIT_LIST_HEAD(&neigh->list);
1308 ipoib_cm_set(neigh, NULL);
1309 /* one ref on behalf of the caller */
1310 atomic_set(&neigh->refcnt, 1);
1315 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1316 struct net_device *dev)
1318 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1319 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1320 struct ipoib_neigh_hash *htbl;
1321 struct ipoib_neigh *neigh;
1324 htbl = rcu_dereference_protected(ntbl->htbl,
1325 lockdep_is_held(&priv->lock));
1331 /* need to add a new neigh, but maybe some other thread succeeded?
1332 * recalc hash, maybe hash resize took place so we do a search
1334 hash_val = ipoib_addr_hash(htbl, daddr);
1335 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1336 lockdep_is_held(&priv->lock));
1338 neigh = rcu_dereference_protected(neigh->hnext,
1339 lockdep_is_held(&priv->lock))) {
1340 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1341 /* found, take one ref on behalf of the caller */
1342 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1347 neigh->alive = jiffies;
1352 neigh = ipoib_neigh_ctor(daddr, dev);
1356 /* one ref on behalf of the hash table */
1357 atomic_inc(&neigh->refcnt);
1358 neigh->alive = jiffies;
1360 rcu_assign_pointer(neigh->hnext,
1361 rcu_dereference_protected(htbl->buckets[hash_val],
1362 lockdep_is_held(&priv->lock)));
1363 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1364 atomic_inc(&ntbl->entries);
1371 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1373 /* neigh reference count was dropprd to zero */
1374 struct net_device *dev = neigh->dev;
1375 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1376 struct sk_buff *skb;
1378 ipoib_put_ah(neigh->ah);
1379 while ((skb = __skb_dequeue(&neigh->queue))) {
1380 ++dev->stats.tx_dropped;
1381 dev_kfree_skb_any(skb);
1383 if (ipoib_cm_get(neigh))
1384 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1385 ipoib_dbg(ipoib_priv(dev),
1386 "neigh free for %06x %pI6\n",
1387 IPOIB_QPN(neigh->daddr),
1390 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1391 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1392 complete(&priv->ntbl.flushed);
1396 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1398 /* Called as a result of removal from hash table */
1399 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1400 /* note TX context may hold another ref */
1401 ipoib_neigh_put(neigh);
1404 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1406 struct net_device *dev = neigh->dev;
1407 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1408 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1409 struct ipoib_neigh_hash *htbl;
1410 struct ipoib_neigh __rcu **np;
1411 struct ipoib_neigh *n;
1414 htbl = rcu_dereference_protected(ntbl->htbl,
1415 lockdep_is_held(&priv->lock));
1419 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1420 np = &htbl->buckets[hash_val];
1421 for (n = rcu_dereference_protected(*np,
1422 lockdep_is_held(&priv->lock));
1424 n = rcu_dereference_protected(*np,
1425 lockdep_is_held(&priv->lock))) {
1428 rcu_assign_pointer(*np,
1429 rcu_dereference_protected(neigh->hnext,
1430 lockdep_is_held(&priv->lock)));
1431 /* remove from parent list */
1432 list_del_init(&neigh->list);
1433 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1441 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1443 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1444 struct ipoib_neigh_hash *htbl;
1445 struct ipoib_neigh __rcu **buckets;
1448 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1450 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1453 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1454 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1455 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1461 htbl->mask = (size - 1);
1462 htbl->buckets = buckets;
1463 RCU_INIT_POINTER(ntbl->htbl, htbl);
1465 atomic_set(&ntbl->entries, 0);
1467 /* start garbage collection */
1468 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1469 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1470 arp_tbl.gc_interval);
1475 static void neigh_hash_free_rcu(struct rcu_head *head)
1477 struct ipoib_neigh_hash *htbl = container_of(head,
1478 struct ipoib_neigh_hash,
1480 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1481 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1485 complete(&ntbl->deleted);
1488 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1490 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1491 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1492 struct ipoib_neigh_hash *htbl;
1493 unsigned long flags;
1496 /* remove all neigh connected to a given path or mcast */
1497 spin_lock_irqsave(&priv->lock, flags);
1499 htbl = rcu_dereference_protected(ntbl->htbl,
1500 lockdep_is_held(&priv->lock));
1505 for (i = 0; i < htbl->size; i++) {
1506 struct ipoib_neigh *neigh;
1507 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1509 while ((neigh = rcu_dereference_protected(*np,
1510 lockdep_is_held(&priv->lock))) != NULL) {
1511 /* delete neighs belong to this parent */
1512 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1513 rcu_assign_pointer(*np,
1514 rcu_dereference_protected(neigh->hnext,
1515 lockdep_is_held(&priv->lock)));
1516 /* remove from parent list */
1517 list_del_init(&neigh->list);
1518 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1526 spin_unlock_irqrestore(&priv->lock, flags);
1529 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1531 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1532 struct ipoib_neigh_hash *htbl;
1533 unsigned long flags;
1534 int i, wait_flushed = 0;
1536 init_completion(&priv->ntbl.flushed);
1538 spin_lock_irqsave(&priv->lock, flags);
1540 htbl = rcu_dereference_protected(ntbl->htbl,
1541 lockdep_is_held(&priv->lock));
1545 wait_flushed = atomic_read(&priv->ntbl.entries);
1549 for (i = 0; i < htbl->size; i++) {
1550 struct ipoib_neigh *neigh;
1551 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1553 while ((neigh = rcu_dereference_protected(*np,
1554 lockdep_is_held(&priv->lock))) != NULL) {
1555 rcu_assign_pointer(*np,
1556 rcu_dereference_protected(neigh->hnext,
1557 lockdep_is_held(&priv->lock)));
1558 /* remove from path/mc list */
1559 list_del_init(&neigh->list);
1560 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1565 rcu_assign_pointer(ntbl->htbl, NULL);
1566 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1569 spin_unlock_irqrestore(&priv->lock, flags);
1571 wait_for_completion(&priv->ntbl.flushed);
1574 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1576 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1579 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1580 init_completion(&priv->ntbl.deleted);
1581 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1583 /* Stop GC if called at init fail need to cancel work */
1584 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1586 cancel_delayed_work(&priv->neigh_reap_task);
1588 ipoib_flush_neighs(priv);
1590 wait_for_completion(&priv->ntbl.deleted);
1593 static void ipoib_dev_uninit_default(struct net_device *dev)
1595 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1597 ipoib_transport_dev_cleanup(dev);
1599 ipoib_cm_dev_cleanup(dev);
1601 kfree(priv->rx_ring);
1602 vfree(priv->tx_ring);
1604 priv->rx_ring = NULL;
1605 priv->tx_ring = NULL;
1608 static int ipoib_dev_init_default(struct net_device *dev)
1610 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1612 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1614 /* Allocate RX/TX "rings" to hold queued skbs */
1615 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1620 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1621 if (!priv->tx_ring) {
1622 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1623 priv->ca->name, ipoib_sendq_size);
1624 goto out_rx_ring_cleanup;
1627 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1629 if (ipoib_transport_dev_init(dev, priv->ca)) {
1630 pr_warn("%s: ipoib_transport_dev_init failed\n",
1632 goto out_tx_ring_cleanup;
1635 /* after qp created set dev address */
1636 priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1637 priv->dev->dev_addr[2] = (priv->qp->qp_num >> 8) & 0xff;
1638 priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1640 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
1641 (unsigned long)dev);
1645 out_tx_ring_cleanup:
1646 vfree(priv->tx_ring);
1648 out_rx_ring_cleanup:
1649 kfree(priv->rx_ring);
1655 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1657 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1665 * the various IPoIB tasks assume they will never race against
1666 * themselves, so always use a single thread workqueue
1668 priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1670 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1674 /* create pd, which used both for control and datapath*/
1675 priv->pd = ib_alloc_pd(priv->ca, 0);
1676 if (IS_ERR(priv->pd)) {
1677 pr_warn("%s: failed to allocate PD\n", ca->name);
1681 ret = priv->rn_ops->ndo_init(dev);
1683 pr_warn("%s failed to init HW resource\n", dev->name);
1687 if (ipoib_neigh_hash_init(priv) < 0) {
1688 pr_warn("%s failed to init neigh hash\n", dev->name);
1689 goto out_dev_uninit;
1692 if (dev->flags & IFF_UP) {
1693 if (ipoib_ib_dev_open(dev)) {
1694 pr_warn("%s failed to open device\n", dev->name);
1696 goto out_dev_uninit;
1703 ipoib_ib_dev_cleanup(dev);
1707 ib_dealloc_pd(priv->pd);
1713 destroy_workqueue(priv->wq);
1721 void ipoib_dev_cleanup(struct net_device *dev)
1723 struct ipoib_dev_priv *priv = ipoib_priv(dev), *cpriv, *tcpriv;
1728 /* Delete any child interfaces first */
1729 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1730 /* Stop GC on child */
1731 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1732 cancel_delayed_work(&cpriv->neigh_reap_task);
1733 unregister_netdevice_queue(cpriv->dev, &head);
1735 unregister_netdevice_many(&head);
1737 ipoib_neigh_hash_uninit(dev);
1739 ipoib_ib_dev_cleanup(dev);
1741 /* no more works over the priv->wq */
1743 flush_workqueue(priv->wq);
1744 destroy_workqueue(priv->wq);
1749 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1751 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1753 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1756 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1757 struct ifla_vf_info *ivf)
1759 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1762 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
1771 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
1773 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1775 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
1778 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
1781 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
1782 struct ifla_vf_stats *vf_stats)
1784 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1786 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
1789 static const struct header_ops ipoib_header_ops = {
1790 .create = ipoib_hard_header,
1793 static const struct net_device_ops ipoib_netdev_ops_pf = {
1794 .ndo_uninit = ipoib_uninit,
1795 .ndo_open = ipoib_open,
1796 .ndo_stop = ipoib_stop,
1797 .ndo_change_mtu = ipoib_change_mtu,
1798 .ndo_fix_features = ipoib_fix_features,
1799 .ndo_start_xmit = ipoib_start_xmit,
1800 .ndo_tx_timeout = ipoib_timeout,
1801 .ndo_set_rx_mode = ipoib_set_mcast_list,
1802 .ndo_get_iflink = ipoib_get_iflink,
1803 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
1804 .ndo_get_vf_config = ipoib_get_vf_config,
1805 .ndo_get_vf_stats = ipoib_get_vf_stats,
1806 .ndo_set_vf_guid = ipoib_set_vf_guid,
1807 .ndo_set_mac_address = ipoib_set_mac,
1810 static const struct net_device_ops ipoib_netdev_ops_vf = {
1811 .ndo_uninit = ipoib_uninit,
1812 .ndo_open = ipoib_open,
1813 .ndo_stop = ipoib_stop,
1814 .ndo_change_mtu = ipoib_change_mtu,
1815 .ndo_fix_features = ipoib_fix_features,
1816 .ndo_start_xmit = ipoib_start_xmit,
1817 .ndo_tx_timeout = ipoib_timeout,
1818 .ndo_set_rx_mode = ipoib_set_mcast_list,
1819 .ndo_get_iflink = ipoib_get_iflink,
1822 void ipoib_setup_common(struct net_device *dev)
1824 dev->header_ops = &ipoib_header_ops;
1826 ipoib_set_ethtool_ops(dev);
1828 dev->watchdog_timeo = HZ;
1830 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1832 dev->hard_header_len = IPOIB_HARD_LEN;
1833 dev->addr_len = INFINIBAND_ALEN;
1834 dev->type = ARPHRD_INFINIBAND;
1835 dev->tx_queue_len = ipoib_sendq_size * 2;
1836 dev->features = (NETIF_F_VLAN_CHALLENGED |
1838 netif_keep_dst(dev);
1840 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1843 static void ipoib_build_priv(struct net_device *dev)
1845 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1848 spin_lock_init(&priv->lock);
1849 init_rwsem(&priv->vlan_rwsem);
1851 INIT_LIST_HEAD(&priv->path_list);
1852 INIT_LIST_HEAD(&priv->child_intfs);
1853 INIT_LIST_HEAD(&priv->dead_ahs);
1854 INIT_LIST_HEAD(&priv->multicast_list);
1856 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1857 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1858 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1859 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1860 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1861 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1862 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1863 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1866 static const struct net_device_ops ipoib_netdev_default_pf = {
1867 .ndo_init = ipoib_dev_init_default,
1868 .ndo_uninit = ipoib_dev_uninit_default,
1869 .ndo_open = ipoib_ib_dev_open_default,
1870 .ndo_stop = ipoib_ib_dev_stop_default,
1873 static struct net_device
1874 *ipoib_create_netdev_default(struct ib_device *hca,
1876 unsigned char name_assign_type,
1877 void (*setup)(struct net_device *))
1879 struct net_device *dev;
1880 struct rdma_netdev *rn;
1882 dev = alloc_netdev((int)sizeof(struct rdma_netdev),
1884 name_assign_type, setup);
1888 rn = netdev_priv(dev);
1890 rn->send = ipoib_send;
1891 rn->attach_mcast = ipoib_mcast_attach;
1892 rn->detach_mcast = ipoib_mcast_detach;
1895 dev->netdev_ops = &ipoib_netdev_default_pf;
1900 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
1903 struct net_device *dev;
1905 if (hca->alloc_rdma_netdev) {
1906 dev = hca->alloc_rdma_netdev(hca, port,
1907 RDMA_NETDEV_IPOIB, name,
1909 ipoib_setup_common);
1910 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
1914 if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
1915 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
1916 ipoib_setup_common);
1921 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
1924 struct net_device *dev;
1925 struct ipoib_dev_priv *priv;
1926 struct rdma_netdev *rn;
1928 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
1932 dev = ipoib_get_netdev(hca, port, name);
1936 priv->rn_ops = dev->netdev_ops;
1938 /* fixme : should be after the query_cap */
1939 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
1940 dev->netdev_ops = &ipoib_netdev_ops_vf;
1942 dev->netdev_ops = &ipoib_netdev_ops_pf;
1944 rn = netdev_priv(dev);
1945 rn->clnt_priv = priv;
1946 ipoib_build_priv(dev);
1954 static ssize_t show_pkey(struct device *dev,
1955 struct device_attribute *attr, char *buf)
1957 struct net_device *ndev = to_net_dev(dev);
1958 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1960 return sprintf(buf, "0x%04x\n", priv->pkey);
1962 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1964 static ssize_t show_umcast(struct device *dev,
1965 struct device_attribute *attr, char *buf)
1967 struct net_device *ndev = to_net_dev(dev);
1968 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1970 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1973 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1975 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1977 if (umcast_val > 0) {
1978 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1979 ipoib_warn(priv, "ignoring multicast groups joined directly "
1982 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1985 static ssize_t set_umcast(struct device *dev,
1986 struct device_attribute *attr,
1987 const char *buf, size_t count)
1989 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1991 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1995 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1997 int ipoib_add_umcast_attr(struct net_device *dev)
1999 return device_create_file(&dev->dev, &dev_attr_umcast);
2002 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2004 struct ipoib_dev_priv *child_priv;
2005 struct net_device *netdev = priv->dev;
2007 netif_addr_lock_bh(netdev);
2009 memcpy(&priv->local_gid.global.interface_id,
2010 &gid->global.interface_id,
2011 sizeof(gid->global.interface_id));
2012 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2013 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2015 netif_addr_unlock_bh(netdev);
2017 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2018 down_read(&priv->vlan_rwsem);
2019 list_for_each_entry(child_priv, &priv->child_intfs, list)
2020 set_base_guid(child_priv, gid);
2021 up_read(&priv->vlan_rwsem);
2025 static int ipoib_check_lladdr(struct net_device *dev,
2026 struct sockaddr_storage *ss)
2028 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2031 netif_addr_lock_bh(dev);
2033 /* Make sure the QPN, reserved and subnet prefix match the current
2034 * lladdr, it also makes sure the lladdr is unicast.
2036 if (memcmp(dev->dev_addr, ss->__data,
2037 4 + sizeof(gid->global.subnet_prefix)) ||
2038 gid->global.interface_id == 0)
2041 netif_addr_unlock_bh(dev);
2046 static int ipoib_set_mac(struct net_device *dev, void *addr)
2048 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2049 struct sockaddr_storage *ss = addr;
2052 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2055 ret = ipoib_check_lladdr(dev, ss);
2059 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2061 queue_work(ipoib_workqueue, &priv->flush_light);
2066 static ssize_t create_child(struct device *dev,
2067 struct device_attribute *attr,
2068 const char *buf, size_t count)
2073 if (sscanf(buf, "%i", &pkey) != 1)
2076 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2080 * Set the full membership bit, so that we join the right
2081 * broadcast group, etc.
2085 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2087 return ret ? ret : count;
2089 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2091 static ssize_t delete_child(struct device *dev,
2092 struct device_attribute *attr,
2093 const char *buf, size_t count)
2098 if (sscanf(buf, "%i", &pkey) != 1)
2101 if (pkey < 0 || pkey > 0xffff)
2104 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2106 return ret ? ret : count;
2109 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2111 int ipoib_add_pkey_attr(struct net_device *dev)
2113 return device_create_file(&dev->dev, &dev_attr_pkey);
2116 void ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
2118 priv->hca_caps = hca->attrs.device_cap_flags;
2120 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
2121 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
2123 if (priv->hca_caps & IB_DEVICE_UD_TSO)
2124 priv->dev->hw_features |= NETIF_F_TSO;
2126 priv->dev->features |= priv->dev->hw_features;
2130 static struct net_device *ipoib_add_port(const char *format,
2131 struct ib_device *hca, u8 port)
2133 struct ipoib_dev_priv *priv;
2134 struct ib_port_attr attr;
2135 int result = -ENOMEM;
2137 priv = ipoib_intf_alloc(hca, port, format);
2139 goto alloc_mem_failed;
2141 SET_NETDEV_DEV(priv->dev, hca->dev.parent);
2142 priv->dev->dev_id = port - 1;
2144 result = ib_query_port(hca, port, &attr);
2146 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
2148 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
2150 goto device_init_failed;
2153 /* MTU will be reset when mcast join happens */
2154 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
2155 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
2156 priv->dev->max_mtu = IPOIB_CM_MTU;
2158 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
2160 result = ib_query_pkey(hca, port, 0, &priv->pkey);
2162 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
2163 hca->name, port, result);
2164 goto device_init_failed;
2167 ipoib_set_dev_features(priv, hca);
2170 * Set the full membership bit, so that we join the right
2171 * broadcast group, etc.
2173 priv->pkey |= 0x8000;
2175 priv->dev->broadcast[8] = priv->pkey >> 8;
2176 priv->dev->broadcast[9] = priv->pkey & 0xff;
2178 result = ib_query_gid(hca, port, 0, &priv->local_gid, NULL);
2180 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
2181 hca->name, port, result);
2182 goto device_init_failed;
2184 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
2185 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2187 result = ipoib_dev_init(priv->dev, hca, port);
2189 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
2190 hca->name, port, result);
2191 goto device_init_failed;
2194 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2195 priv->ca, ipoib_event);
2196 result = ib_register_event_handler(&priv->event_handler);
2198 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
2199 "port %d (ret = %d)\n",
2200 hca->name, port, result);
2204 result = register_netdev(priv->dev);
2206 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
2207 hca->name, port, result);
2208 goto register_failed;
2211 if (ipoib_cm_add_mode_attr(priv->dev))
2213 if (ipoib_add_pkey_attr(priv->dev))
2215 if (ipoib_add_umcast_attr(priv->dev))
2217 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
2219 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
2225 unregister_netdev(priv->dev);
2228 ib_unregister_event_handler(&priv->event_handler);
2229 flush_workqueue(ipoib_workqueue);
2230 /* Stop GC if started before flush */
2231 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2232 cancel_delayed_work(&priv->neigh_reap_task);
2233 flush_workqueue(priv->wq);
2236 ipoib_dev_cleanup(priv->dev);
2239 free_netdev(priv->dev);
2242 return ERR_PTR(result);
2245 static void ipoib_add_one(struct ib_device *device)
2247 struct list_head *dev_list;
2248 struct net_device *dev;
2249 struct ipoib_dev_priv *priv;
2253 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
2257 INIT_LIST_HEAD(dev_list);
2259 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2260 if (!rdma_protocol_ib(device, p))
2262 dev = ipoib_add_port("ib%d", device, p);
2264 priv = ipoib_priv(dev);
2265 list_add_tail(&priv->list, dev_list);
2275 ib_set_client_data(device, &ipoib_client, dev_list);
2278 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2280 struct ipoib_dev_priv *priv, *tmp;
2281 struct list_head *dev_list = client_data;
2286 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2287 ib_unregister_event_handler(&priv->event_handler);
2288 flush_workqueue(ipoib_workqueue);
2290 /* mark interface in the middle of destruction */
2291 set_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags);
2294 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
2298 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
2299 cancel_delayed_work(&priv->neigh_reap_task);
2300 flush_workqueue(priv->wq);
2302 unregister_netdev(priv->dev);
2303 free_netdev(priv->dev);
2310 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2311 static struct notifier_block ipoib_netdev_notifier = {
2312 .notifier_call = ipoib_netdev_event,
2316 static int __init ipoib_init_module(void)
2320 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2321 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2322 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2324 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2325 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2326 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2327 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2328 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2332 * When copying small received packets, we only copy from the
2333 * linear data part of the SKB, so we rely on this condition.
2335 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2337 ret = ipoib_register_debugfs();
2342 * We create a global workqueue here that is used for all flush
2343 * operations. However, if you attempt to flush a workqueue
2344 * from a task on that same workqueue, it deadlocks the system.
2345 * We want to be able to flush the tasks associated with a
2346 * specific net device, so we also create a workqueue for each
2347 * netdevice. We queue up the tasks for that device only on
2348 * its private workqueue, and we only queue up flush events
2349 * on our global flush workqueue. This avoids the deadlocks.
2351 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush",
2353 if (!ipoib_workqueue) {
2358 ib_sa_register_client(&ipoib_sa_client);
2360 ret = ib_register_client(&ipoib_client);
2364 ret = ipoib_netlink_init();
2368 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2369 register_netdevice_notifier(&ipoib_netdev_notifier);
2374 ib_unregister_client(&ipoib_client);
2377 ib_sa_unregister_client(&ipoib_sa_client);
2378 destroy_workqueue(ipoib_workqueue);
2381 ipoib_unregister_debugfs();
2386 static void __exit ipoib_cleanup_module(void)
2388 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2389 unregister_netdevice_notifier(&ipoib_netdev_notifier);
2391 ipoib_netlink_fini();
2392 ib_unregister_client(&ipoib_client);
2393 ib_sa_unregister_client(&ipoib_sa_client);
2394 ipoib_unregister_debugfs();
2395 destroy_workqueue(ipoib_workqueue);
2398 module_init(ipoib_init_module);
2399 module_exit(ipoib_cleanup_module);