2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
37 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
44 #include <linux/if_arp.h> /* For ARPHRD_xxx */
49 #include <linux/jhash.h>
51 #include <net/addrconf.h>
52 #include <linux/inetdevice.h>
53 #include <rdma/ib_cache.h>
55 #define DRV_VERSION "1.0.0"
57 const char ipoib_driver_version[] = DRV_VERSION;
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
61 MODULE_LICENSE("Dual BSD/GPL");
63 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
64 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
72 int ipoib_debug_level;
74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
78 struct ipoib_path_iter {
79 struct net_device *dev;
80 struct ipoib_path path;
83 static const u8 ipv4_bcast_addr[] = {
84 0x00, 0xff, 0xff, 0xff,
85 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
86 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
89 struct workqueue_struct *ipoib_workqueue;
91 struct ib_sa_client ipoib_sa_client;
93 static void ipoib_add_one(struct ib_device *device);
94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
95 static void ipoib_neigh_reclaim(struct rcu_head *rp);
96 static struct net_device *ipoib_get_net_dev_by_params(
97 struct ib_device *dev, u8 port, u16 pkey,
98 const union ib_gid *gid, const struct sockaddr *addr,
100 static int ipoib_set_mac(struct net_device *dev, void *addr);
101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
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 netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
220 struct ipoib_dev_priv *priv = ipoib_priv(dev);
222 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
223 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
228 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
230 struct ipoib_dev_priv *priv = ipoib_priv(dev);
233 /* dev->mtu > 2K ==> connected mode */
234 if (ipoib_cm_admin_enabled(dev)) {
235 if (new_mtu > ipoib_cm_max_mtu(dev))
238 if (new_mtu > priv->mcast_mtu)
239 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
246 if (new_mtu < (ETH_MIN_MTU + IPOIB_ENCAP_LEN) ||
247 new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
250 priv->admin_mtu = new_mtu;
252 if (priv->mcast_mtu < priv->admin_mtu)
253 ipoib_dbg(priv, "MTU must be smaller than the underlying "
254 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
256 new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
258 if (priv->rn_ops->ndo_change_mtu) {
259 bool carrier_status = netif_carrier_ok(dev);
261 netif_carrier_off(dev);
263 /* notify lower level on the real mtu */
264 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
267 netif_carrier_on(dev);
275 static void ipoib_get_stats(struct net_device *dev,
276 struct rtnl_link_stats64 *stats)
278 struct ipoib_dev_priv *priv = ipoib_priv(dev);
280 if (priv->rn_ops->ndo_get_stats64)
281 priv->rn_ops->ndo_get_stats64(dev, stats);
283 netdev_stats_to_stats64(stats, &dev->stats);
286 /* Called with an RCU read lock taken */
287 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
288 struct net_device *dev)
290 struct net *net = dev_net(dev);
291 struct in_device *in_dev;
292 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
293 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
296 switch (addr->sa_family) {
298 in_dev = in_dev_get(dev);
302 ret_addr = inet_confirm_addr(net, in_dev, 0,
303 addr_in->sin_addr.s_addr,
311 if (IS_ENABLED(CONFIG_IPV6) &&
312 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
321 * Find the master net_device on top of the given net_device.
322 * @dev: base IPoIB net_device
324 * Returns the master net_device with a reference held, or the same net_device
325 * if no master exists.
327 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
329 struct net_device *master;
332 master = netdev_master_upper_dev_get_rcu(dev);
344 struct ipoib_walk_data {
345 const struct sockaddr *addr;
346 struct net_device *result;
349 static int ipoib_upper_walk(struct net_device *upper, void *_data)
351 struct ipoib_walk_data *data = _data;
354 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
356 data->result = upper;
364 * Find a net_device matching the given address, which is an upper device of
365 * the given net_device.
366 * @addr: IP address to look for.
367 * @dev: base IPoIB net_device
369 * If found, returns the net_device with a reference held. Otherwise return
372 static struct net_device *ipoib_get_net_dev_match_addr(
373 const struct sockaddr *addr, struct net_device *dev)
375 struct ipoib_walk_data data = {
380 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
386 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
392 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
393 * pkey_index and address, if one exists.
395 * @found_net_dev: contains a matching net_device if the return value >= 1,
396 * with a reference held. */
397 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
398 const union ib_gid *gid,
400 const struct sockaddr *addr,
402 struct net_device **found_net_dev)
404 struct ipoib_dev_priv *child_priv;
405 struct net_device *net_dev = NULL;
408 if (priv->pkey_index == pkey_index &&
409 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
411 net_dev = ipoib_get_master_net_dev(priv->dev);
413 /* Verify the net_device matches the IP address, as
414 * IPoIB child devices currently share a GID. */
415 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
419 *found_net_dev = net_dev;
426 /* Check child interfaces */
427 down_read_nested(&priv->vlan_rwsem, nesting);
428 list_for_each_entry(child_priv, &priv->child_intfs, list) {
429 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
436 up_read(&priv->vlan_rwsem);
441 /* Returns the number of matching net_devs found (between 0 and 2). Also
442 * return the matching net_device in the @net_dev parameter, holding a
443 * reference to the net_device, if the number of matches >= 1 */
444 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
446 const union ib_gid *gid,
447 const struct sockaddr *addr,
448 struct net_device **net_dev)
450 struct ipoib_dev_priv *priv;
455 list_for_each_entry(priv, dev_list, list) {
456 if (priv->port != port)
459 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
468 static struct net_device *ipoib_get_net_dev_by_params(
469 struct ib_device *dev, u8 port, u16 pkey,
470 const union ib_gid *gid, const struct sockaddr *addr,
473 struct net_device *net_dev;
474 struct list_head *dev_list = client_data;
479 if (!rdma_protocol_ib(dev, port))
482 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
489 /* See if we can find a unique device matching the L2 parameters */
490 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
491 gid, NULL, &net_dev);
502 /* Couldn't find a unique device with L2 parameters only. Use L3
503 * address to uniquely match the net device */
504 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
505 gid, addr, &net_dev);
510 dev_warn_ratelimited(&dev->dev,
511 "duplicate IP address detected\n");
518 int ipoib_set_mode(struct net_device *dev, const char *buf)
520 struct ipoib_dev_priv *priv = ipoib_priv(dev);
522 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
523 !strcmp(buf, "connected\n")) ||
524 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
525 !strcmp(buf, "datagram\n"))) {
529 /* flush paths if we switch modes so that connections are restarted */
530 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
531 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
532 ipoib_warn(priv, "enabling connected mode "
533 "will cause multicast packet drops\n");
534 netdev_update_features(dev);
535 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
537 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
539 ipoib_flush_paths(dev);
540 return (!rtnl_trylock()) ? -EBUSY : 0;
543 if (!strcmp(buf, "datagram\n")) {
544 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
545 netdev_update_features(dev);
546 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
548 ipoib_flush_paths(dev);
549 return (!rtnl_trylock()) ? -EBUSY : 0;
555 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
557 struct ipoib_dev_priv *priv = ipoib_priv(dev);
558 struct rb_node *n = priv->path_tree.rb_node;
559 struct ipoib_path *path;
563 path = rb_entry(n, struct ipoib_path, rb_node);
565 ret = memcmp(gid, path->pathrec.dgid.raw,
566 sizeof (union ib_gid));
579 static int __path_add(struct net_device *dev, struct ipoib_path *path)
581 struct ipoib_dev_priv *priv = ipoib_priv(dev);
582 struct rb_node **n = &priv->path_tree.rb_node;
583 struct rb_node *pn = NULL;
584 struct ipoib_path *tpath;
589 tpath = rb_entry(pn, struct ipoib_path, rb_node);
591 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
592 sizeof (union ib_gid));
601 rb_link_node(&path->rb_node, pn, n);
602 rb_insert_color(&path->rb_node, &priv->path_tree);
604 list_add_tail(&path->list, &priv->path_list);
609 static void path_free(struct net_device *dev, struct ipoib_path *path)
613 while ((skb = __skb_dequeue(&path->queue)))
614 dev_kfree_skb_irq(skb);
616 ipoib_dbg(ipoib_priv(dev), "path_free\n");
618 /* remove all neigh connected to this path */
619 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
622 ipoib_put_ah(path->ah);
627 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
629 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
631 struct ipoib_path_iter *iter;
633 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
638 memset(iter->path.pathrec.dgid.raw, 0, 16);
640 if (ipoib_path_iter_next(iter)) {
648 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
650 struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
652 struct ipoib_path *path;
655 spin_lock_irq(&priv->lock);
657 n = rb_first(&priv->path_tree);
660 path = rb_entry(n, struct ipoib_path, rb_node);
662 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
663 sizeof (union ib_gid)) < 0) {
672 spin_unlock_irq(&priv->lock);
677 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
678 struct ipoib_path *path)
683 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
685 void ipoib_mark_paths_invalid(struct net_device *dev)
687 struct ipoib_dev_priv *priv = ipoib_priv(dev);
688 struct ipoib_path *path, *tp;
690 spin_lock_irq(&priv->lock);
692 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
693 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
694 be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
695 path->pathrec.dgid.raw);
700 spin_unlock_irq(&priv->lock);
703 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
705 struct ipoib_pseudo_header *phdr;
707 phdr = skb_push(skb, sizeof(*phdr));
708 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
711 void ipoib_flush_paths(struct net_device *dev)
713 struct ipoib_dev_priv *priv = ipoib_priv(dev);
714 struct ipoib_path *path, *tp;
715 LIST_HEAD(remove_list);
718 netif_tx_lock_bh(dev);
719 spin_lock_irqsave(&priv->lock, flags);
721 list_splice_init(&priv->path_list, &remove_list);
723 list_for_each_entry(path, &remove_list, list)
724 rb_erase(&path->rb_node, &priv->path_tree);
726 list_for_each_entry_safe(path, tp, &remove_list, list) {
728 ib_sa_cancel_query(path->query_id, path->query);
729 spin_unlock_irqrestore(&priv->lock, flags);
730 netif_tx_unlock_bh(dev);
731 wait_for_completion(&path->done);
732 path_free(dev, path);
733 netif_tx_lock_bh(dev);
734 spin_lock_irqsave(&priv->lock, flags);
737 spin_unlock_irqrestore(&priv->lock, flags);
738 netif_tx_unlock_bh(dev);
741 static void path_rec_completion(int status,
742 struct sa_path_rec *pathrec,
745 struct ipoib_path *path = path_ptr;
746 struct net_device *dev = path->dev;
747 struct ipoib_dev_priv *priv = ipoib_priv(dev);
748 struct ipoib_ah *ah = NULL;
749 struct ipoib_ah *old_ah = NULL;
750 struct ipoib_neigh *neigh, *tn;
751 struct sk_buff_head skqueue;
756 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
757 be32_to_cpu(sa_path_get_dlid(pathrec)),
760 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
761 status, path->pathrec.dgid.raw);
763 skb_queue_head_init(&skqueue);
766 struct rdma_ah_attr av;
768 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
769 pathrec, &av, NULL)) {
770 ah = ipoib_create_ah(dev, priv->pd, &av);
771 rdma_destroy_ah_attr(&av);
775 spin_lock_irqsave(&priv->lock, flags);
777 if (!IS_ERR_OR_NULL(ah)) {
779 * pathrec.dgid is used as the database key from the LLADDR,
780 * it must remain unchanged even if the SA returns a different
781 * GID to use in the AH.
783 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
784 sizeof(union ib_gid))) {
787 "%s got PathRec for gid %pI6 while asked for %pI6\n",
788 dev->name, pathrec->dgid.raw,
789 path->pathrec.dgid.raw);
790 memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
791 sizeof(union ib_gid));
794 path->pathrec = *pathrec;
799 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
800 ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
803 while ((skb = __skb_dequeue(&path->queue)))
804 __skb_queue_tail(&skqueue, skb);
806 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
808 WARN_ON(neigh->ah != old_ah);
810 * Dropping the ah reference inside
811 * priv->lock is safe here, because we
812 * will hold one more reference from
813 * the original value of path->ah (ie
816 ipoib_put_ah(neigh->ah);
818 kref_get(&path->ah->ref);
819 neigh->ah = path->ah;
821 if (ipoib_cm_enabled(dev, neigh->daddr)) {
822 if (!ipoib_cm_get(neigh))
823 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
826 if (!ipoib_cm_get(neigh)) {
827 ipoib_neigh_free(neigh);
832 while ((skb = __skb_dequeue(&neigh->queue)))
833 __skb_queue_tail(&skqueue, skb);
839 complete(&path->done);
841 spin_unlock_irqrestore(&priv->lock, flags);
843 if (IS_ERR_OR_NULL(ah))
844 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
847 ipoib_put_ah(old_ah);
849 while ((skb = __skb_dequeue(&skqueue))) {
852 ret = dev_queue_xmit(skb);
854 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
859 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
862 path->dev = priv->dev;
864 if (rdma_cap_opa_ah(priv->ca, priv->port))
865 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
867 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
869 memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
870 path->pathrec.sgid = priv->local_gid;
871 path->pathrec.pkey = cpu_to_be16(priv->pkey);
872 path->pathrec.numb_path = 1;
873 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
876 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
878 struct ipoib_dev_priv *priv = ipoib_priv(dev);
879 struct ipoib_path *path;
881 if (!priv->broadcast)
884 path = kzalloc(sizeof(*path), GFP_ATOMIC);
888 skb_queue_head_init(&path->queue);
890 INIT_LIST_HEAD(&path->neigh_list);
892 init_path_rec(priv, path, gid);
897 static int path_rec_start(struct net_device *dev,
898 struct ipoib_path *path)
900 struct ipoib_dev_priv *priv = ipoib_priv(dev);
902 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
903 path->pathrec.dgid.raw);
905 init_completion(&path->done);
908 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
910 IB_SA_PATH_REC_DGID |
911 IB_SA_PATH_REC_SGID |
912 IB_SA_PATH_REC_NUMB_PATH |
913 IB_SA_PATH_REC_TRAFFIC_CLASS |
918 if (path->query_id < 0) {
919 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
921 complete(&path->done);
922 return path->query_id;
928 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
929 struct net_device *dev)
931 struct ipoib_dev_priv *priv = ipoib_priv(dev);
932 struct ipoib_path *path;
935 spin_lock_irqsave(&priv->lock, flags);
937 path = __path_find(dev, daddr + 4);
941 path_rec_start(dev, path);
943 spin_unlock_irqrestore(&priv->lock, flags);
946 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
947 struct net_device *dev)
949 struct ipoib_dev_priv *priv = ipoib_priv(dev);
950 struct rdma_netdev *rn = netdev_priv(dev);
951 struct ipoib_path *path;
952 struct ipoib_neigh *neigh;
955 spin_lock_irqsave(&priv->lock, flags);
956 neigh = ipoib_neigh_alloc(daddr, dev);
958 spin_unlock_irqrestore(&priv->lock, flags);
959 ++dev->stats.tx_dropped;
960 dev_kfree_skb_any(skb);
964 /* To avoid race condition, make sure that the
965 * neigh will be added only once.
967 if (unlikely(!list_empty(&neigh->list))) {
968 spin_unlock_irqrestore(&priv->lock, flags);
972 path = __path_find(dev, daddr + 4);
974 path = path_rec_create(dev, daddr + 4);
978 __path_add(dev, path);
981 list_add_tail(&neigh->list, &path->neigh_list);
983 if (path->ah && path->ah->valid) {
984 kref_get(&path->ah->ref);
985 neigh->ah = path->ah;
987 if (ipoib_cm_enabled(dev, neigh->daddr)) {
988 if (!ipoib_cm_get(neigh))
989 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
990 if (!ipoib_cm_get(neigh)) {
991 ipoib_neigh_free(neigh);
994 if (skb_queue_len(&neigh->queue) <
995 IPOIB_MAX_PATH_REC_QUEUE) {
996 push_pseudo_header(skb, neigh->daddr);
997 __skb_queue_tail(&neigh->queue, skb);
999 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
1000 skb_queue_len(&neigh->queue));
1004 spin_unlock_irqrestore(&priv->lock, flags);
1005 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1007 ipoib_neigh_put(neigh);
1013 if (!path->query && path_rec_start(dev, path))
1015 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1016 push_pseudo_header(skb, neigh->daddr);
1017 __skb_queue_tail(&neigh->queue, skb);
1023 spin_unlock_irqrestore(&priv->lock, flags);
1024 ipoib_neigh_put(neigh);
1028 ipoib_neigh_free(neigh);
1030 ++dev->stats.tx_dropped;
1031 dev_kfree_skb_any(skb);
1033 spin_unlock_irqrestore(&priv->lock, flags);
1034 ipoib_neigh_put(neigh);
1039 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1040 struct ipoib_pseudo_header *phdr)
1042 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1043 struct rdma_netdev *rn = netdev_priv(dev);
1044 struct ipoib_path *path;
1045 unsigned long flags;
1047 spin_lock_irqsave(&priv->lock, flags);
1049 /* no broadcast means that all paths are (going to be) not valid */
1050 if (!priv->broadcast)
1051 goto drop_and_unlock;
1053 path = __path_find(dev, phdr->hwaddr + 4);
1054 if (!path || !path->ah || !path->ah->valid) {
1056 path = path_rec_create(dev, phdr->hwaddr + 4);
1058 goto drop_and_unlock;
1059 __path_add(dev, path);
1062 * make sure there are no changes in the existing
1065 init_path_rec(priv, path, phdr->hwaddr + 4);
1067 if (!path->query && path_rec_start(dev, path)) {
1068 goto drop_and_unlock;
1071 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1072 push_pseudo_header(skb, phdr->hwaddr);
1073 __skb_queue_tail(&path->queue, skb);
1076 goto drop_and_unlock;
1080 spin_unlock_irqrestore(&priv->lock, flags);
1081 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1082 be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1083 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1084 IPOIB_QPN(phdr->hwaddr));
1088 ++dev->stats.tx_dropped;
1089 dev_kfree_skb_any(skb);
1091 spin_unlock_irqrestore(&priv->lock, flags);
1094 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1096 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1097 struct rdma_netdev *rn = netdev_priv(dev);
1098 struct ipoib_neigh *neigh;
1099 struct ipoib_pseudo_header *phdr;
1100 struct ipoib_header *header;
1101 unsigned long flags;
1103 phdr = (struct ipoib_pseudo_header *) skb->data;
1104 skb_pull(skb, sizeof(*phdr));
1105 header = (struct ipoib_header *) skb->data;
1107 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1108 /* multicast, arrange "if" according to probability */
1109 if ((header->proto != htons(ETH_P_IP)) &&
1110 (header->proto != htons(ETH_P_IPV6)) &&
1111 (header->proto != htons(ETH_P_ARP)) &&
1112 (header->proto != htons(ETH_P_RARP)) &&
1113 (header->proto != htons(ETH_P_TIPC))) {
1114 /* ethertype not supported by IPoIB */
1115 ++dev->stats.tx_dropped;
1116 dev_kfree_skb_any(skb);
1117 return NETDEV_TX_OK;
1119 /* Add in the P_Key for multicast*/
1120 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1121 phdr->hwaddr[9] = priv->pkey & 0xff;
1123 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1125 goto send_using_neigh;
1126 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1127 return NETDEV_TX_OK;
1130 /* unicast, arrange "switch" according to probability */
1131 switch (header->proto) {
1132 case htons(ETH_P_IP):
1133 case htons(ETH_P_IPV6):
1134 case htons(ETH_P_TIPC):
1135 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1136 if (unlikely(!neigh)) {
1137 neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1139 return NETDEV_TX_OK;
1142 case htons(ETH_P_ARP):
1143 case htons(ETH_P_RARP):
1144 /* for unicast ARP and RARP should always perform path find */
1145 unicast_arp_send(skb, dev, phdr);
1146 return NETDEV_TX_OK;
1148 /* ethertype not supported by IPoIB */
1149 ++dev->stats.tx_dropped;
1150 dev_kfree_skb_any(skb);
1151 return NETDEV_TX_OK;
1155 /* note we now hold a ref to neigh */
1156 if (ipoib_cm_get(neigh)) {
1157 if (ipoib_cm_up(neigh)) {
1158 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1161 } else if (neigh->ah && neigh->ah->valid) {
1162 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1163 IPOIB_QPN(phdr->hwaddr));
1165 } else if (neigh->ah) {
1166 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1169 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1170 push_pseudo_header(skb, phdr->hwaddr);
1171 spin_lock_irqsave(&priv->lock, flags);
1172 __skb_queue_tail(&neigh->queue, skb);
1173 spin_unlock_irqrestore(&priv->lock, flags);
1175 ++dev->stats.tx_dropped;
1176 dev_kfree_skb_any(skb);
1180 ipoib_neigh_put(neigh);
1182 return NETDEV_TX_OK;
1185 static void ipoib_timeout(struct net_device *dev)
1187 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1189 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1190 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1192 "queue stopped %d, tx_head %u, tx_tail %u, global_tx_head %u, global_tx_tail %u\n",
1193 netif_queue_stopped(dev), priv->tx_head, priv->tx_tail,
1194 priv->global_tx_head, priv->global_tx_tail);
1196 /* XXX reset QP, etc. */
1199 static int ipoib_hard_header(struct sk_buff *skb,
1200 struct net_device *dev,
1201 unsigned short type,
1206 struct ipoib_header *header;
1208 header = skb_push(skb, sizeof(*header));
1210 header->proto = htons(type);
1211 header->reserved = 0;
1214 * we don't rely on dst_entry structure, always stuff the
1215 * destination address into skb hard header so we can figure out where
1216 * to send the packet later.
1218 push_pseudo_header(skb, daddr);
1220 return IPOIB_HARD_LEN;
1223 static void ipoib_set_mcast_list(struct net_device *dev)
1225 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1227 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1228 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1232 queue_work(priv->wq, &priv->restart_task);
1235 static int ipoib_get_iflink(const struct net_device *dev)
1237 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1239 /* parent interface */
1240 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1241 return dev->ifindex;
1243 /* child/vlan interface */
1244 return priv->parent->ifindex;
1247 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1250 * Use only the address parts that contributes to spreading
1251 * The subnet prefix is not used as one can not connect to
1252 * same remote port (GUID) using the same remote QPN via two
1253 * different subnets.
1255 /* qpn octets[1:4) & port GUID octets[12:20) */
1256 u32 *d32 = (u32 *) daddr;
1259 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1260 return hv & htbl->mask;
1263 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1265 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1266 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1267 struct ipoib_neigh_hash *htbl;
1268 struct ipoib_neigh *neigh = NULL;
1273 htbl = rcu_dereference_bh(ntbl->htbl);
1278 hash_val = ipoib_addr_hash(htbl, daddr);
1279 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1281 neigh = rcu_dereference_bh(neigh->hnext)) {
1282 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1283 /* found, take one ref on behalf of the caller */
1284 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1290 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1291 neigh->alive = jiffies;
1297 rcu_read_unlock_bh();
1301 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1303 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1304 struct ipoib_neigh_hash *htbl;
1305 unsigned long neigh_obsolete;
1307 unsigned long flags;
1309 LIST_HEAD(remove_list);
1311 spin_lock_irqsave(&priv->lock, flags);
1313 htbl = rcu_dereference_protected(ntbl->htbl,
1314 lockdep_is_held(&priv->lock));
1319 /* neigh is obsolete if it was idle for two GC periods */
1320 dt = 2 * arp_tbl.gc_interval;
1321 neigh_obsolete = jiffies - dt;
1323 for (i = 0; i < htbl->size; i++) {
1324 struct ipoib_neigh *neigh;
1325 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1327 while ((neigh = rcu_dereference_protected(*np,
1328 lockdep_is_held(&priv->lock))) != NULL) {
1329 /* was the neigh idle for two GC periods */
1330 if (time_after(neigh_obsolete, neigh->alive)) {
1332 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1334 rcu_assign_pointer(*np,
1335 rcu_dereference_protected(neigh->hnext,
1336 lockdep_is_held(&priv->lock)));
1337 /* remove from path/mc list */
1338 list_del_init(&neigh->list);
1339 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1348 spin_unlock_irqrestore(&priv->lock, flags);
1349 ipoib_mcast_remove_list(&remove_list);
1352 static void ipoib_reap_neigh(struct work_struct *work)
1354 struct ipoib_dev_priv *priv =
1355 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1357 __ipoib_reap_neigh(priv);
1359 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1360 arp_tbl.gc_interval);
1364 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1365 struct net_device *dev)
1367 struct ipoib_neigh *neigh;
1369 neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1374 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1375 skb_queue_head_init(&neigh->queue);
1376 INIT_LIST_HEAD(&neigh->list);
1377 ipoib_cm_set(neigh, NULL);
1378 /* one ref on behalf of the caller */
1379 atomic_set(&neigh->refcnt, 1);
1384 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1385 struct net_device *dev)
1387 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1388 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1389 struct ipoib_neigh_hash *htbl;
1390 struct ipoib_neigh *neigh;
1393 htbl = rcu_dereference_protected(ntbl->htbl,
1394 lockdep_is_held(&priv->lock));
1400 /* need to add a new neigh, but maybe some other thread succeeded?
1401 * recalc hash, maybe hash resize took place so we do a search
1403 hash_val = ipoib_addr_hash(htbl, daddr);
1404 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1405 lockdep_is_held(&priv->lock));
1407 neigh = rcu_dereference_protected(neigh->hnext,
1408 lockdep_is_held(&priv->lock))) {
1409 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1410 /* found, take one ref on behalf of the caller */
1411 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1416 neigh->alive = jiffies;
1421 neigh = ipoib_neigh_ctor(daddr, dev);
1425 /* one ref on behalf of the hash table */
1426 atomic_inc(&neigh->refcnt);
1427 neigh->alive = jiffies;
1429 rcu_assign_pointer(neigh->hnext,
1430 rcu_dereference_protected(htbl->buckets[hash_val],
1431 lockdep_is_held(&priv->lock)));
1432 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1433 atomic_inc(&ntbl->entries);
1440 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1442 /* neigh reference count was dropprd to zero */
1443 struct net_device *dev = neigh->dev;
1444 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1445 struct sk_buff *skb;
1447 ipoib_put_ah(neigh->ah);
1448 while ((skb = __skb_dequeue(&neigh->queue))) {
1449 ++dev->stats.tx_dropped;
1450 dev_kfree_skb_any(skb);
1452 if (ipoib_cm_get(neigh))
1453 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1454 ipoib_dbg(ipoib_priv(dev),
1455 "neigh free for %06x %pI6\n",
1456 IPOIB_QPN(neigh->daddr),
1459 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1460 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1461 complete(&priv->ntbl.flushed);
1465 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1467 /* Called as a result of removal from hash table */
1468 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1469 /* note TX context may hold another ref */
1470 ipoib_neigh_put(neigh);
1473 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1475 struct net_device *dev = neigh->dev;
1476 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1477 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1478 struct ipoib_neigh_hash *htbl;
1479 struct ipoib_neigh __rcu **np;
1480 struct ipoib_neigh *n;
1483 htbl = rcu_dereference_protected(ntbl->htbl,
1484 lockdep_is_held(&priv->lock));
1488 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1489 np = &htbl->buckets[hash_val];
1490 for (n = rcu_dereference_protected(*np,
1491 lockdep_is_held(&priv->lock));
1493 n = rcu_dereference_protected(*np,
1494 lockdep_is_held(&priv->lock))) {
1497 rcu_assign_pointer(*np,
1498 rcu_dereference_protected(neigh->hnext,
1499 lockdep_is_held(&priv->lock)));
1500 /* remove from parent list */
1501 list_del_init(&neigh->list);
1502 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1510 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1512 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1513 struct ipoib_neigh_hash *htbl;
1514 struct ipoib_neigh __rcu **buckets;
1517 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1519 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1522 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1523 buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1529 htbl->mask = (size - 1);
1530 htbl->buckets = buckets;
1531 RCU_INIT_POINTER(ntbl->htbl, htbl);
1533 atomic_set(&ntbl->entries, 0);
1535 /* start garbage collection */
1536 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1537 arp_tbl.gc_interval);
1542 static void neigh_hash_free_rcu(struct rcu_head *head)
1544 struct ipoib_neigh_hash *htbl = container_of(head,
1545 struct ipoib_neigh_hash,
1547 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1548 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1552 complete(&ntbl->deleted);
1555 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1557 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1558 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1559 struct ipoib_neigh_hash *htbl;
1560 unsigned long flags;
1563 /* remove all neigh connected to a given path or mcast */
1564 spin_lock_irqsave(&priv->lock, flags);
1566 htbl = rcu_dereference_protected(ntbl->htbl,
1567 lockdep_is_held(&priv->lock));
1572 for (i = 0; i < htbl->size; i++) {
1573 struct ipoib_neigh *neigh;
1574 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1576 while ((neigh = rcu_dereference_protected(*np,
1577 lockdep_is_held(&priv->lock))) != NULL) {
1578 /* delete neighs belong to this parent */
1579 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1580 rcu_assign_pointer(*np,
1581 rcu_dereference_protected(neigh->hnext,
1582 lockdep_is_held(&priv->lock)));
1583 /* remove from parent list */
1584 list_del_init(&neigh->list);
1585 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1593 spin_unlock_irqrestore(&priv->lock, flags);
1596 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1598 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1599 struct ipoib_neigh_hash *htbl;
1600 unsigned long flags;
1601 int i, wait_flushed = 0;
1603 init_completion(&priv->ntbl.flushed);
1604 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1606 spin_lock_irqsave(&priv->lock, flags);
1608 htbl = rcu_dereference_protected(ntbl->htbl,
1609 lockdep_is_held(&priv->lock));
1613 wait_flushed = atomic_read(&priv->ntbl.entries);
1617 for (i = 0; i < htbl->size; i++) {
1618 struct ipoib_neigh *neigh;
1619 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1621 while ((neigh = rcu_dereference_protected(*np,
1622 lockdep_is_held(&priv->lock))) != NULL) {
1623 rcu_assign_pointer(*np,
1624 rcu_dereference_protected(neigh->hnext,
1625 lockdep_is_held(&priv->lock)));
1626 /* remove from path/mc list */
1627 list_del_init(&neigh->list);
1628 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1633 rcu_assign_pointer(ntbl->htbl, NULL);
1634 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1637 spin_unlock_irqrestore(&priv->lock, flags);
1639 wait_for_completion(&priv->ntbl.flushed);
1642 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1644 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1646 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1647 init_completion(&priv->ntbl.deleted);
1649 cancel_delayed_work_sync(&priv->neigh_reap_task);
1651 ipoib_flush_neighs(priv);
1653 wait_for_completion(&priv->ntbl.deleted);
1656 static void ipoib_napi_add(struct net_device *dev)
1658 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1660 netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1661 netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1664 static void ipoib_napi_del(struct net_device *dev)
1666 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1668 netif_napi_del(&priv->recv_napi);
1669 netif_napi_del(&priv->send_napi);
1672 static void ipoib_dev_uninit_default(struct net_device *dev)
1674 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1676 ipoib_transport_dev_cleanup(dev);
1678 ipoib_napi_del(dev);
1680 ipoib_cm_dev_cleanup(dev);
1682 kfree(priv->rx_ring);
1683 vfree(priv->tx_ring);
1685 priv->rx_ring = NULL;
1686 priv->tx_ring = NULL;
1689 static int ipoib_dev_init_default(struct net_device *dev)
1691 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1693 ipoib_napi_add(dev);
1695 /* Allocate RX/TX "rings" to hold queued skbs */
1696 priv->rx_ring = kcalloc(ipoib_recvq_size,
1697 sizeof(*priv->rx_ring),
1702 priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1703 sizeof(*priv->tx_ring)));
1704 if (!priv->tx_ring) {
1705 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1706 priv->ca->name, ipoib_sendq_size);
1707 goto out_rx_ring_cleanup;
1710 /* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
1712 if (ipoib_transport_dev_init(dev, priv->ca)) {
1713 pr_warn("%s: ipoib_transport_dev_init failed\n",
1715 goto out_tx_ring_cleanup;
1718 /* after qp created set dev address */
1719 priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1720 priv->dev->dev_addr[2] = (priv->qp->qp_num >> 8) & 0xff;
1721 priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1725 out_tx_ring_cleanup:
1726 vfree(priv->tx_ring);
1728 out_rx_ring_cleanup:
1729 kfree(priv->rx_ring);
1732 ipoib_napi_del(dev);
1736 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1739 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1741 if (!priv->rn_ops->ndo_do_ioctl)
1744 return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1747 static int ipoib_dev_init(struct net_device *dev)
1749 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1755 * the various IPoIB tasks assume they will never race against
1756 * themselves, so always use a single thread workqueue
1758 priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1760 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1764 /* create pd, which used both for control and datapath*/
1765 priv->pd = ib_alloc_pd(priv->ca, 0);
1766 if (IS_ERR(priv->pd)) {
1767 pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1771 ret = priv->rn_ops->ndo_init(dev);
1773 pr_warn("%s failed to init HW resource\n", dev->name);
1777 ret = ipoib_neigh_hash_init(priv);
1779 pr_warn("%s failed to init neigh hash\n", dev->name);
1780 goto out_dev_uninit;
1783 if (dev->flags & IFF_UP) {
1784 if (ipoib_ib_dev_open(dev)) {
1785 pr_warn("%s failed to open device\n", dev->name);
1787 goto out_hash_uninit;
1794 ipoib_neigh_hash_uninit(dev);
1797 ipoib_ib_dev_cleanup(dev);
1801 ib_dealloc_pd(priv->pd);
1807 destroy_workqueue(priv->wq);
1816 * This must be called before doing an unregister_netdev on a parent device to
1817 * shutdown the IB event handler.
1819 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1821 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1824 * ipoib_set_mac checks netif_running before pushing work, clearing
1825 * running ensures the it will not add more work.
1828 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1831 /* ipoib_event() cannot be running once this returns */
1832 ib_unregister_event_handler(&priv->event_handler);
1835 * Work on the queue grabs the rtnl lock, so this cannot be done while
1838 flush_workqueue(ipoib_workqueue);
1841 static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1843 priv->hca_caps = priv->ca->attrs.device_cap_flags;
1845 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1846 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1848 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1849 priv->dev->hw_features |= NETIF_F_TSO;
1851 priv->dev->features |= priv->dev->hw_features;
1855 static int ipoib_parent_init(struct net_device *ndev)
1857 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1858 struct ib_port_attr attr;
1861 result = ib_query_port(priv->ca, priv->port, &attr);
1863 pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1867 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1869 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1871 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1872 priv->ca->name, priv->port, result);
1876 result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1878 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1879 priv->ca->name, priv->port, result);
1882 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
1883 sizeof(union ib_gid));
1885 SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1886 priv->dev->dev_port = priv->port - 1;
1887 /* Let's set this one too for backwards compatibility. */
1888 priv->dev->dev_id = priv->port - 1;
1893 static void ipoib_child_init(struct net_device *ndev)
1895 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1896 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1898 priv->max_ib_mtu = ppriv->max_ib_mtu;
1899 set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1900 memcpy(priv->dev->dev_addr, ppriv->dev->dev_addr, INFINIBAND_ALEN);
1901 memcpy(&priv->local_gid, &ppriv->local_gid, sizeof(priv->local_gid));
1904 static int ipoib_ndo_init(struct net_device *ndev)
1906 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1910 ipoib_child_init(ndev);
1912 rc = ipoib_parent_init(ndev);
1917 /* MTU will be reset when mcast join happens */
1918 ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1919 priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
1920 ndev->max_mtu = IPOIB_CM_MTU;
1922 ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
1925 * Set the full membership bit, so that we join the right
1926 * broadcast group, etc.
1928 priv->pkey |= 0x8000;
1930 ndev->broadcast[8] = priv->pkey >> 8;
1931 ndev->broadcast[9] = priv->pkey & 0xff;
1932 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1934 ipoib_set_dev_features(priv);
1936 rc = ipoib_dev_init(ndev);
1938 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1939 priv->ca->name, priv->dev->name, priv->port, rc);
1944 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1946 dev_hold(priv->parent);
1948 down_write(&ppriv->vlan_rwsem);
1949 list_add_tail(&priv->list, &ppriv->child_intfs);
1950 up_write(&ppriv->vlan_rwsem);
1956 static void ipoib_ndo_uninit(struct net_device *dev)
1958 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1963 * ipoib_remove_one guarantees the children are removed before the
1964 * parent, and that is the only place where a parent can be removed.
1966 WARN_ON(!list_empty(&priv->child_intfs));
1969 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1971 down_write(&ppriv->vlan_rwsem);
1972 list_del(&priv->list);
1973 up_write(&ppriv->vlan_rwsem);
1976 ipoib_neigh_hash_uninit(dev);
1978 ipoib_ib_dev_cleanup(dev);
1980 /* no more works over the priv->wq */
1982 /* See ipoib_mcast_carrier_on_task() */
1983 WARN_ON(test_bit(IPOIB_FLAG_OPER_UP, &priv->flags));
1984 flush_workqueue(priv->wq);
1985 destroy_workqueue(priv->wq);
1990 dev_put(priv->parent);
1993 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1995 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1997 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
2000 static int ipoib_get_vf_config(struct net_device *dev, int vf,
2001 struct ifla_vf_info *ivf)
2003 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2006 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2011 memcpy(ivf->mac, dev->dev_addr, dev->addr_len);
2016 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2018 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2020 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2023 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2026 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2027 struct ifla_vf_stats *vf_stats)
2029 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2031 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2034 static const struct header_ops ipoib_header_ops = {
2035 .create = ipoib_hard_header,
2038 static const struct net_device_ops ipoib_netdev_ops_pf = {
2039 .ndo_init = ipoib_ndo_init,
2040 .ndo_uninit = ipoib_ndo_uninit,
2041 .ndo_open = ipoib_open,
2042 .ndo_stop = ipoib_stop,
2043 .ndo_change_mtu = ipoib_change_mtu,
2044 .ndo_fix_features = ipoib_fix_features,
2045 .ndo_start_xmit = ipoib_start_xmit,
2046 .ndo_tx_timeout = ipoib_timeout,
2047 .ndo_set_rx_mode = ipoib_set_mcast_list,
2048 .ndo_get_iflink = ipoib_get_iflink,
2049 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
2050 .ndo_get_vf_config = ipoib_get_vf_config,
2051 .ndo_get_vf_stats = ipoib_get_vf_stats,
2052 .ndo_set_vf_guid = ipoib_set_vf_guid,
2053 .ndo_set_mac_address = ipoib_set_mac,
2054 .ndo_get_stats64 = ipoib_get_stats,
2055 .ndo_do_ioctl = ipoib_ioctl,
2058 static const struct net_device_ops ipoib_netdev_ops_vf = {
2059 .ndo_init = ipoib_ndo_init,
2060 .ndo_uninit = ipoib_ndo_uninit,
2061 .ndo_open = ipoib_open,
2062 .ndo_stop = ipoib_stop,
2063 .ndo_change_mtu = ipoib_change_mtu,
2064 .ndo_fix_features = ipoib_fix_features,
2065 .ndo_start_xmit = ipoib_start_xmit,
2066 .ndo_tx_timeout = ipoib_timeout,
2067 .ndo_set_rx_mode = ipoib_set_mcast_list,
2068 .ndo_get_iflink = ipoib_get_iflink,
2069 .ndo_get_stats64 = ipoib_get_stats,
2070 .ndo_do_ioctl = ipoib_ioctl,
2073 void ipoib_setup_common(struct net_device *dev)
2075 dev->header_ops = &ipoib_header_ops;
2077 ipoib_set_ethtool_ops(dev);
2079 dev->watchdog_timeo = HZ;
2081 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
2083 dev->hard_header_len = IPOIB_HARD_LEN;
2084 dev->addr_len = INFINIBAND_ALEN;
2085 dev->type = ARPHRD_INFINIBAND;
2086 dev->tx_queue_len = ipoib_sendq_size * 2;
2087 dev->features = (NETIF_F_VLAN_CHALLENGED |
2089 netif_keep_dst(dev);
2091 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2094 * unregister_netdev always frees the netdev, we use this mode
2095 * consistently to unify all the various unregister paths, including
2096 * those connected to rtnl_link_ops which require it.
2098 dev->needs_free_netdev = true;
2101 static void ipoib_build_priv(struct net_device *dev)
2103 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2106 spin_lock_init(&priv->lock);
2107 init_rwsem(&priv->vlan_rwsem);
2108 mutex_init(&priv->mcast_mutex);
2110 INIT_LIST_HEAD(&priv->path_list);
2111 INIT_LIST_HEAD(&priv->child_intfs);
2112 INIT_LIST_HEAD(&priv->dead_ahs);
2113 INIT_LIST_HEAD(&priv->multicast_list);
2115 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
2116 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2117 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
2118 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
2119 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
2120 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2121 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2122 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2125 static const struct net_device_ops ipoib_netdev_default_pf = {
2126 .ndo_init = ipoib_dev_init_default,
2127 .ndo_uninit = ipoib_dev_uninit_default,
2128 .ndo_open = ipoib_ib_dev_open_default,
2129 .ndo_stop = ipoib_ib_dev_stop_default,
2132 static struct net_device
2133 *ipoib_create_netdev_default(struct ib_device *hca,
2135 unsigned char name_assign_type,
2136 void (*setup)(struct net_device *))
2138 struct net_device *dev;
2139 struct rdma_netdev *rn;
2141 dev = alloc_netdev((int)sizeof(struct rdma_netdev),
2143 name_assign_type, setup);
2147 rn = netdev_priv(dev);
2149 rn->send = ipoib_send;
2150 rn->attach_mcast = ipoib_mcast_attach;
2151 rn->detach_mcast = ipoib_mcast_detach;
2153 dev->netdev_ops = &ipoib_netdev_default_pf;
2158 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
2161 struct net_device *dev;
2163 if (hca->alloc_rdma_netdev) {
2164 dev = hca->alloc_rdma_netdev(hca, port,
2165 RDMA_NETDEV_IPOIB, name,
2167 ipoib_setup_common);
2168 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
2172 if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
2173 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
2174 ipoib_setup_common);
2179 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
2182 struct net_device *dev;
2183 struct ipoib_dev_priv *priv;
2184 struct rdma_netdev *rn;
2186 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2193 dev = ipoib_get_netdev(hca, port, name);
2197 priv->rn_ops = dev->netdev_ops;
2199 /* fixme : should be after the query_cap */
2200 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
2201 dev->netdev_ops = &ipoib_netdev_ops_vf;
2203 dev->netdev_ops = &ipoib_netdev_ops_pf;
2205 rn = netdev_priv(dev);
2206 rn->clnt_priv = priv;
2209 * Only the child register_netdev flows can handle priv_destructor
2210 * being set, so we force it to NULL here and handle manually until it
2211 * is safe to turn on.
2213 priv->next_priv_destructor = dev->priv_destructor;
2214 dev->priv_destructor = NULL;
2216 ipoib_build_priv(dev);
2224 void ipoib_intf_free(struct net_device *dev)
2226 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2227 struct rdma_netdev *rn = netdev_priv(dev);
2229 dev->priv_destructor = priv->next_priv_destructor;
2230 if (dev->priv_destructor)
2231 dev->priv_destructor(dev);
2234 * There are some error flows around register_netdev failing that may
2235 * attempt to call priv_destructor twice, prevent that from happening.
2237 dev->priv_destructor = NULL;
2239 /* unregister/destroy is very complicated. Make bugs more obvious. */
2240 rn->clnt_priv = NULL;
2245 static ssize_t show_pkey(struct device *dev,
2246 struct device_attribute *attr, char *buf)
2248 struct net_device *ndev = to_net_dev(dev);
2249 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2251 return sprintf(buf, "0x%04x\n", priv->pkey);
2253 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2255 static ssize_t show_umcast(struct device *dev,
2256 struct device_attribute *attr, char *buf)
2258 struct net_device *ndev = to_net_dev(dev);
2259 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2261 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2264 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2266 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2268 if (umcast_val > 0) {
2269 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2270 ipoib_warn(priv, "ignoring multicast groups joined directly "
2273 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2276 static ssize_t set_umcast(struct device *dev,
2277 struct device_attribute *attr,
2278 const char *buf, size_t count)
2280 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2282 ipoib_set_umcast(to_net_dev(dev), umcast_val);
2286 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2288 int ipoib_add_umcast_attr(struct net_device *dev)
2290 return device_create_file(&dev->dev, &dev_attr_umcast);
2293 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2295 struct ipoib_dev_priv *child_priv;
2296 struct net_device *netdev = priv->dev;
2298 netif_addr_lock_bh(netdev);
2300 memcpy(&priv->local_gid.global.interface_id,
2301 &gid->global.interface_id,
2302 sizeof(gid->global.interface_id));
2303 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2304 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2306 netif_addr_unlock_bh(netdev);
2308 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2309 down_read(&priv->vlan_rwsem);
2310 list_for_each_entry(child_priv, &priv->child_intfs, list)
2311 set_base_guid(child_priv, gid);
2312 up_read(&priv->vlan_rwsem);
2316 static int ipoib_check_lladdr(struct net_device *dev,
2317 struct sockaddr_storage *ss)
2319 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2322 netif_addr_lock_bh(dev);
2324 /* Make sure the QPN, reserved and subnet prefix match the current
2325 * lladdr, it also makes sure the lladdr is unicast.
2327 if (memcmp(dev->dev_addr, ss->__data,
2328 4 + sizeof(gid->global.subnet_prefix)) ||
2329 gid->global.interface_id == 0)
2332 netif_addr_unlock_bh(dev);
2337 static int ipoib_set_mac(struct net_device *dev, void *addr)
2339 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2340 struct sockaddr_storage *ss = addr;
2343 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2346 ret = ipoib_check_lladdr(dev, ss);
2350 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2352 queue_work(ipoib_workqueue, &priv->flush_light);
2357 static ssize_t create_child(struct device *dev,
2358 struct device_attribute *attr,
2359 const char *buf, size_t count)
2364 if (sscanf(buf, "%i", &pkey) != 1)
2367 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2370 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2372 return ret ? ret : count;
2374 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2376 static ssize_t delete_child(struct device *dev,
2377 struct device_attribute *attr,
2378 const char *buf, size_t count)
2383 if (sscanf(buf, "%i", &pkey) != 1)
2386 if (pkey < 0 || pkey > 0xffff)
2389 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2391 return ret ? ret : count;
2394 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2396 int ipoib_add_pkey_attr(struct net_device *dev)
2398 return device_create_file(&dev->dev, &dev_attr_pkey);
2401 static struct net_device *ipoib_add_port(const char *format,
2402 struct ib_device *hca, u8 port)
2404 struct ipoib_dev_priv *priv;
2405 struct net_device *ndev;
2408 priv = ipoib_intf_alloc(hca, port, format);
2410 pr_warn("%s, %d: ipoib_intf_alloc failed\n", hca->name, port);
2411 return ERR_PTR(-ENOMEM);
2415 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2416 priv->ca, ipoib_event);
2417 ib_register_event_handler(&priv->event_handler);
2419 /* call event handler to ensure pkey in sync */
2420 queue_work(ipoib_workqueue, &priv->flush_heavy);
2422 result = register_netdev(ndev);
2424 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2425 hca->name, port, result);
2427 ipoib_parent_unregister_pre(ndev);
2428 ipoib_intf_free(ndev);
2431 return ERR_PTR(result);
2435 * We cannot set priv_destructor before register_netdev because we
2436 * need priv to be always valid during the error flow to execute
2437 * ipoib_parent_unregister_pre(). Instead handle it manually and only
2438 * enter priv_destructor mode once we are completely registered.
2440 ndev->priv_destructor = ipoib_intf_free;
2442 if (ipoib_cm_add_mode_attr(ndev))
2444 if (ipoib_add_pkey_attr(ndev))
2446 if (ipoib_add_umcast_attr(ndev))
2448 if (device_create_file(&ndev->dev, &dev_attr_create_child))
2450 if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2456 ipoib_parent_unregister_pre(ndev);
2457 unregister_netdev(ndev);
2458 return ERR_PTR(-ENOMEM);
2461 static void ipoib_add_one(struct ib_device *device)
2463 struct list_head *dev_list;
2464 struct net_device *dev;
2465 struct ipoib_dev_priv *priv;
2469 dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2473 INIT_LIST_HEAD(dev_list);
2475 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2476 if (!rdma_protocol_ib(device, p))
2478 dev = ipoib_add_port("ib%d", device, p);
2480 priv = ipoib_priv(dev);
2481 list_add_tail(&priv->list, dev_list);
2491 ib_set_client_data(device, &ipoib_client, dev_list);
2494 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2496 struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2497 struct list_head *dev_list = client_data;
2502 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2504 ipoib_parent_unregister_pre(priv->dev);
2508 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2510 unregister_netdevice_queue(cpriv->dev, &head);
2511 unregister_netdevice_queue(priv->dev, &head);
2512 unregister_netdevice_many(&head);
2520 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2521 static struct notifier_block ipoib_netdev_notifier = {
2522 .notifier_call = ipoib_netdev_event,
2526 static int __init ipoib_init_module(void)
2530 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2531 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2532 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2534 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2535 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2536 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2537 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2538 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2539 ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2543 * When copying small received packets, we only copy from the
2544 * linear data part of the SKB, so we rely on this condition.
2546 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2548 ret = ipoib_register_debugfs();
2553 * We create a global workqueue here that is used for all flush
2554 * operations. However, if you attempt to flush a workqueue
2555 * from a task on that same workqueue, it deadlocks the system.
2556 * We want to be able to flush the tasks associated with a
2557 * specific net device, so we also create a workqueue for each
2558 * netdevice. We queue up the tasks for that device only on
2559 * its private workqueue, and we only queue up flush events
2560 * on our global flush workqueue. This avoids the deadlocks.
2562 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2563 if (!ipoib_workqueue) {
2568 ib_sa_register_client(&ipoib_sa_client);
2570 ret = ib_register_client(&ipoib_client);
2574 ret = ipoib_netlink_init();
2578 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2579 register_netdevice_notifier(&ipoib_netdev_notifier);
2584 ib_unregister_client(&ipoib_client);
2587 ib_sa_unregister_client(&ipoib_sa_client);
2588 destroy_workqueue(ipoib_workqueue);
2591 ipoib_unregister_debugfs();
2596 static void __exit ipoib_cleanup_module(void)
2598 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2599 unregister_netdevice_notifier(&ipoib_netdev_notifier);
2601 ipoib_netlink_fini();
2602 ib_unregister_client(&ipoib_client);
2603 ib_sa_unregister_client(&ipoib_sa_client);
2604 ipoib_unregister_debugfs();
2605 destroy_workqueue(ipoib_workqueue);
2608 module_init(ipoib_init_module);
2609 module_exit(ipoib_cleanup_module);