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>
52 #define DRV_VERSION "1.0.0"
54 const char ipoib_driver_version[] = DRV_VERSION;
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
58 MODULE_LICENSE("Dual BSD/GPL");
59 MODULE_VERSION(DRV_VERSION);
61 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
62 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
64 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
65 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
66 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
67 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
69 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
70 int ipoib_debug_level;
72 module_param_named(debug_level, ipoib_debug_level, int, 0644);
73 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
76 struct ipoib_path_iter {
77 struct net_device *dev;
78 struct ipoib_path path;
81 static const u8 ipv4_bcast_addr[] = {
82 0x00, 0xff, 0xff, 0xff,
83 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
84 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
87 struct workqueue_struct *ipoib_workqueue;
89 struct ib_sa_client ipoib_sa_client;
91 static void ipoib_add_one(struct ib_device *device);
92 static void ipoib_remove_one(struct ib_device *device);
93 static void ipoib_neigh_reclaim(struct rcu_head *rp);
95 static struct ib_client ipoib_client = {
98 .remove = ipoib_remove_one
101 int ipoib_open(struct net_device *dev)
103 struct ipoib_dev_priv *priv = netdev_priv(dev);
105 ipoib_dbg(priv, "bringing up interface\n");
107 netif_carrier_off(dev);
109 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
111 if (ipoib_ib_dev_open(dev)) {
112 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
117 if (ipoib_ib_dev_up(dev))
120 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
121 struct ipoib_dev_priv *cpriv;
123 /* Bring up any child interfaces too */
124 down_read(&priv->vlan_rwsem);
125 list_for_each_entry(cpriv, &priv->child_intfs, list) {
128 flags = cpriv->dev->flags;
132 dev_change_flags(cpriv->dev, flags | IFF_UP);
134 up_read(&priv->vlan_rwsem);
137 netif_start_queue(dev);
142 ipoib_ib_dev_stop(dev);
145 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
150 static int ipoib_stop(struct net_device *dev)
152 struct ipoib_dev_priv *priv = netdev_priv(dev);
154 ipoib_dbg(priv, "stopping interface\n");
156 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
158 netif_stop_queue(dev);
160 ipoib_ib_dev_down(dev);
161 ipoib_ib_dev_stop(dev);
163 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
164 struct ipoib_dev_priv *cpriv;
166 /* Bring down any child interfaces too */
167 down_read(&priv->vlan_rwsem);
168 list_for_each_entry(cpriv, &priv->child_intfs, list) {
171 flags = cpriv->dev->flags;
172 if (!(flags & IFF_UP))
175 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
177 up_read(&priv->vlan_rwsem);
183 static void ipoib_uninit(struct net_device *dev)
185 ipoib_dev_cleanup(dev);
188 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
190 struct ipoib_dev_priv *priv = netdev_priv(dev);
192 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
193 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
198 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
200 struct ipoib_dev_priv *priv = netdev_priv(dev);
202 /* dev->mtu > 2K ==> connected mode */
203 if (ipoib_cm_admin_enabled(dev)) {
204 if (new_mtu > ipoib_cm_max_mtu(dev))
207 if (new_mtu > priv->mcast_mtu)
208 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
215 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
218 priv->admin_mtu = new_mtu;
220 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
225 int ipoib_set_mode(struct net_device *dev, const char *buf)
227 struct ipoib_dev_priv *priv = netdev_priv(dev);
229 /* flush paths if we switch modes so that connections are restarted */
230 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
231 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
232 ipoib_warn(priv, "enabling connected mode "
233 "will cause multicast packet drops\n");
234 netdev_update_features(dev);
236 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
238 ipoib_flush_paths(dev);
243 if (!strcmp(buf, "datagram\n")) {
244 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
245 netdev_update_features(dev);
246 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
248 ipoib_flush_paths(dev);
256 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
258 struct ipoib_dev_priv *priv = netdev_priv(dev);
259 struct rb_node *n = priv->path_tree.rb_node;
260 struct ipoib_path *path;
264 path = rb_entry(n, struct ipoib_path, rb_node);
266 ret = memcmp(gid, path->pathrec.dgid.raw,
267 sizeof (union ib_gid));
280 static int __path_add(struct net_device *dev, struct ipoib_path *path)
282 struct ipoib_dev_priv *priv = netdev_priv(dev);
283 struct rb_node **n = &priv->path_tree.rb_node;
284 struct rb_node *pn = NULL;
285 struct ipoib_path *tpath;
290 tpath = rb_entry(pn, struct ipoib_path, rb_node);
292 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
293 sizeof (union ib_gid));
302 rb_link_node(&path->rb_node, pn, n);
303 rb_insert_color(&path->rb_node, &priv->path_tree);
305 list_add_tail(&path->list, &priv->path_list);
310 static void path_free(struct net_device *dev, struct ipoib_path *path)
314 while ((skb = __skb_dequeue(&path->queue)))
315 dev_kfree_skb_irq(skb);
317 ipoib_dbg(netdev_priv(dev), "path_free\n");
319 /* remove all neigh connected to this path */
320 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
323 ipoib_put_ah(path->ah);
328 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
330 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
332 struct ipoib_path_iter *iter;
334 iter = kmalloc(sizeof *iter, GFP_KERNEL);
339 memset(iter->path.pathrec.dgid.raw, 0, 16);
341 if (ipoib_path_iter_next(iter)) {
349 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
351 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
353 struct ipoib_path *path;
356 spin_lock_irq(&priv->lock);
358 n = rb_first(&priv->path_tree);
361 path = rb_entry(n, struct ipoib_path, rb_node);
363 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
364 sizeof (union ib_gid)) < 0) {
373 spin_unlock_irq(&priv->lock);
378 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
379 struct ipoib_path *path)
384 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
386 void ipoib_mark_paths_invalid(struct net_device *dev)
388 struct ipoib_dev_priv *priv = netdev_priv(dev);
389 struct ipoib_path *path, *tp;
391 spin_lock_irq(&priv->lock);
393 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
394 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
395 be16_to_cpu(path->pathrec.dlid),
396 path->pathrec.dgid.raw);
400 spin_unlock_irq(&priv->lock);
403 void ipoib_flush_paths(struct net_device *dev)
405 struct ipoib_dev_priv *priv = netdev_priv(dev);
406 struct ipoib_path *path, *tp;
407 LIST_HEAD(remove_list);
410 netif_tx_lock_bh(dev);
411 spin_lock_irqsave(&priv->lock, flags);
413 list_splice_init(&priv->path_list, &remove_list);
415 list_for_each_entry(path, &remove_list, list)
416 rb_erase(&path->rb_node, &priv->path_tree);
418 list_for_each_entry_safe(path, tp, &remove_list, list) {
420 ib_sa_cancel_query(path->query_id, path->query);
421 spin_unlock_irqrestore(&priv->lock, flags);
422 netif_tx_unlock_bh(dev);
423 wait_for_completion(&path->done);
424 path_free(dev, path);
425 netif_tx_lock_bh(dev);
426 spin_lock_irqsave(&priv->lock, flags);
429 spin_unlock_irqrestore(&priv->lock, flags);
430 netif_tx_unlock_bh(dev);
433 static void path_rec_completion(int status,
434 struct ib_sa_path_rec *pathrec,
437 struct ipoib_path *path = path_ptr;
438 struct net_device *dev = path->dev;
439 struct ipoib_dev_priv *priv = netdev_priv(dev);
440 struct ipoib_ah *ah = NULL;
441 struct ipoib_ah *old_ah = NULL;
442 struct ipoib_neigh *neigh, *tn;
443 struct sk_buff_head skqueue;
448 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
449 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
451 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
452 status, path->pathrec.dgid.raw);
454 skb_queue_head_init(&skqueue);
457 struct ib_ah_attr av;
459 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
460 ah = ipoib_create_ah(dev, priv->pd, &av);
463 spin_lock_irqsave(&priv->lock, flags);
465 if (!IS_ERR_OR_NULL(ah)) {
466 path->pathrec = *pathrec;
471 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
472 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
474 while ((skb = __skb_dequeue(&path->queue)))
475 __skb_queue_tail(&skqueue, skb);
477 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
479 WARN_ON(neigh->ah != old_ah);
481 * Dropping the ah reference inside
482 * priv->lock is safe here, because we
483 * will hold one more reference from
484 * the original value of path->ah (ie
487 ipoib_put_ah(neigh->ah);
489 kref_get(&path->ah->ref);
490 neigh->ah = path->ah;
492 if (ipoib_cm_enabled(dev, neigh->daddr)) {
493 if (!ipoib_cm_get(neigh))
494 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
497 if (!ipoib_cm_get(neigh)) {
498 ipoib_neigh_free(neigh);
503 while ((skb = __skb_dequeue(&neigh->queue)))
504 __skb_queue_tail(&skqueue, skb);
510 complete(&path->done);
512 spin_unlock_irqrestore(&priv->lock, flags);
514 if (IS_ERR_OR_NULL(ah))
515 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
518 ipoib_put_ah(old_ah);
520 while ((skb = __skb_dequeue(&skqueue))) {
522 if (dev_queue_xmit(skb))
523 ipoib_warn(priv, "dev_queue_xmit failed "
524 "to requeue packet\n");
528 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
530 struct ipoib_dev_priv *priv = netdev_priv(dev);
531 struct ipoib_path *path;
533 if (!priv->broadcast)
536 path = kzalloc(sizeof *path, GFP_ATOMIC);
542 skb_queue_head_init(&path->queue);
544 INIT_LIST_HEAD(&path->neigh_list);
546 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
547 path->pathrec.sgid = priv->local_gid;
548 path->pathrec.pkey = cpu_to_be16(priv->pkey);
549 path->pathrec.numb_path = 1;
550 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
555 static int path_rec_start(struct net_device *dev,
556 struct ipoib_path *path)
558 struct ipoib_dev_priv *priv = netdev_priv(dev);
560 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
561 path->pathrec.dgid.raw);
563 init_completion(&path->done);
566 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
568 IB_SA_PATH_REC_DGID |
569 IB_SA_PATH_REC_SGID |
570 IB_SA_PATH_REC_NUMB_PATH |
571 IB_SA_PATH_REC_TRAFFIC_CLASS |
576 if (path->query_id < 0) {
577 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
579 complete(&path->done);
580 return path->query_id;
586 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
587 struct net_device *dev)
589 struct ipoib_dev_priv *priv = netdev_priv(dev);
590 struct ipoib_path *path;
591 struct ipoib_neigh *neigh;
594 spin_lock_irqsave(&priv->lock, flags);
595 neigh = ipoib_neigh_alloc(daddr, dev);
597 spin_unlock_irqrestore(&priv->lock, flags);
598 ++dev->stats.tx_dropped;
599 dev_kfree_skb_any(skb);
603 path = __path_find(dev, daddr + 4);
605 path = path_rec_create(dev, daddr + 4);
609 __path_add(dev, path);
612 list_add_tail(&neigh->list, &path->neigh_list);
615 kref_get(&path->ah->ref);
616 neigh->ah = path->ah;
618 if (ipoib_cm_enabled(dev, neigh->daddr)) {
619 if (!ipoib_cm_get(neigh))
620 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
621 if (!ipoib_cm_get(neigh)) {
622 ipoib_neigh_free(neigh);
625 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
626 __skb_queue_tail(&neigh->queue, skb);
628 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
629 skb_queue_len(&neigh->queue));
633 spin_unlock_irqrestore(&priv->lock, flags);
634 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
635 ipoib_neigh_put(neigh);
641 if (!path->query && path_rec_start(dev, path))
644 __skb_queue_tail(&neigh->queue, skb);
647 spin_unlock_irqrestore(&priv->lock, flags);
648 ipoib_neigh_put(neigh);
652 ipoib_neigh_free(neigh);
654 ++dev->stats.tx_dropped;
655 dev_kfree_skb_any(skb);
657 spin_unlock_irqrestore(&priv->lock, flags);
658 ipoib_neigh_put(neigh);
661 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
664 struct ipoib_dev_priv *priv = netdev_priv(dev);
665 struct ipoib_path *path;
668 spin_lock_irqsave(&priv->lock, flags);
670 path = __path_find(dev, cb->hwaddr + 4);
671 if (!path || !path->valid) {
675 path = path_rec_create(dev, cb->hwaddr + 4);
679 __skb_queue_tail(&path->queue, skb);
681 if (!path->query && path_rec_start(dev, path)) {
682 spin_unlock_irqrestore(&priv->lock, flags);
684 path_free(dev, path);
687 __path_add(dev, path);
689 ++dev->stats.tx_dropped;
690 dev_kfree_skb_any(skb);
693 spin_unlock_irqrestore(&priv->lock, flags);
698 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
699 be16_to_cpu(path->pathrec.dlid));
701 spin_unlock_irqrestore(&priv->lock, flags);
702 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
704 } else if ((path->query || !path_rec_start(dev, path)) &&
705 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
706 __skb_queue_tail(&path->queue, skb);
708 ++dev->stats.tx_dropped;
709 dev_kfree_skb_any(skb);
712 spin_unlock_irqrestore(&priv->lock, flags);
715 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
717 struct ipoib_dev_priv *priv = netdev_priv(dev);
718 struct ipoib_neigh *neigh;
719 struct ipoib_cb *cb = ipoib_skb_cb(skb);
720 struct ipoib_header *header;
723 header = (struct ipoib_header *) skb->data;
725 if (unlikely(cb->hwaddr[4] == 0xff)) {
726 /* multicast, arrange "if" according to probability */
727 if ((header->proto != htons(ETH_P_IP)) &&
728 (header->proto != htons(ETH_P_IPV6)) &&
729 (header->proto != htons(ETH_P_ARP)) &&
730 (header->proto != htons(ETH_P_RARP)) &&
731 (header->proto != htons(ETH_P_TIPC))) {
732 /* ethertype not supported by IPoIB */
733 ++dev->stats.tx_dropped;
734 dev_kfree_skb_any(skb);
737 /* Add in the P_Key for multicast*/
738 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
739 cb->hwaddr[9] = priv->pkey & 0xff;
741 neigh = ipoib_neigh_get(dev, cb->hwaddr);
743 goto send_using_neigh;
744 ipoib_mcast_send(dev, cb->hwaddr, skb);
748 /* unicast, arrange "switch" according to probability */
749 switch (header->proto) {
750 case htons(ETH_P_IP):
751 case htons(ETH_P_IPV6):
752 case htons(ETH_P_TIPC):
753 neigh = ipoib_neigh_get(dev, cb->hwaddr);
754 if (unlikely(!neigh)) {
755 neigh_add_path(skb, cb->hwaddr, dev);
759 case htons(ETH_P_ARP):
760 case htons(ETH_P_RARP):
761 /* for unicast ARP and RARP should always perform path find */
762 unicast_arp_send(skb, dev, cb);
765 /* ethertype not supported by IPoIB */
766 ++dev->stats.tx_dropped;
767 dev_kfree_skb_any(skb);
772 /* note we now hold a ref to neigh */
773 if (ipoib_cm_get(neigh)) {
774 if (ipoib_cm_up(neigh)) {
775 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
778 } else if (neigh->ah) {
779 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
783 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
784 spin_lock_irqsave(&priv->lock, flags);
785 __skb_queue_tail(&neigh->queue, skb);
786 spin_unlock_irqrestore(&priv->lock, flags);
788 ++dev->stats.tx_dropped;
789 dev_kfree_skb_any(skb);
793 ipoib_neigh_put(neigh);
798 static void ipoib_timeout(struct net_device *dev)
800 struct ipoib_dev_priv *priv = netdev_priv(dev);
802 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
803 jiffies_to_msecs(jiffies - dev->trans_start));
804 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
805 netif_queue_stopped(dev),
806 priv->tx_head, priv->tx_tail);
807 /* XXX reset QP, etc. */
810 static int ipoib_hard_header(struct sk_buff *skb,
811 struct net_device *dev,
813 const void *daddr, const void *saddr, unsigned len)
815 struct ipoib_header *header;
816 struct ipoib_cb *cb = ipoib_skb_cb(skb);
818 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
820 header->proto = htons(type);
821 header->reserved = 0;
824 * we don't rely on dst_entry structure, always stuff the
825 * destination address into skb->cb so we can figure out where
826 * to send the packet later.
828 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
830 return sizeof *header;
833 static void ipoib_set_mcast_list(struct net_device *dev)
835 struct ipoib_dev_priv *priv = netdev_priv(dev);
837 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
838 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
842 queue_work(priv->wq, &priv->restart_task);
845 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
848 * Use only the address parts that contributes to spreading
849 * The subnet prefix is not used as one can not connect to
850 * same remote port (GUID) using the same remote QPN via two
853 /* qpn octets[1:4) & port GUID octets[12:20) */
854 u32 *d32 = (u32 *) daddr;
857 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
858 return hv & htbl->mask;
861 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
863 struct ipoib_dev_priv *priv = netdev_priv(dev);
864 struct ipoib_neigh_table *ntbl = &priv->ntbl;
865 struct ipoib_neigh_hash *htbl;
866 struct ipoib_neigh *neigh = NULL;
871 htbl = rcu_dereference_bh(ntbl->htbl);
876 hash_val = ipoib_addr_hash(htbl, daddr);
877 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
879 neigh = rcu_dereference_bh(neigh->hnext)) {
880 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
881 /* found, take one ref on behalf of the caller */
882 if (!atomic_inc_not_zero(&neigh->refcnt)) {
887 neigh->alive = jiffies;
893 rcu_read_unlock_bh();
897 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
899 struct ipoib_neigh_table *ntbl = &priv->ntbl;
900 struct ipoib_neigh_hash *htbl;
901 unsigned long neigh_obsolete;
906 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
909 spin_lock_irqsave(&priv->lock, flags);
911 htbl = rcu_dereference_protected(ntbl->htbl,
912 lockdep_is_held(&priv->lock));
917 /* neigh is obsolete if it was idle for two GC periods */
918 dt = 2 * arp_tbl.gc_interval;
919 neigh_obsolete = jiffies - dt;
920 /* handle possible race condition */
921 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
924 for (i = 0; i < htbl->size; i++) {
925 struct ipoib_neigh *neigh;
926 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
928 while ((neigh = rcu_dereference_protected(*np,
929 lockdep_is_held(&priv->lock))) != NULL) {
930 /* was the neigh idle for two GC periods */
931 if (time_after(neigh_obsolete, neigh->alive)) {
932 rcu_assign_pointer(*np,
933 rcu_dereference_protected(neigh->hnext,
934 lockdep_is_held(&priv->lock)));
935 /* remove from path/mc list */
936 list_del(&neigh->list);
937 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
946 spin_unlock_irqrestore(&priv->lock, flags);
949 static void ipoib_reap_neigh(struct work_struct *work)
951 struct ipoib_dev_priv *priv =
952 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
954 __ipoib_reap_neigh(priv);
956 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
957 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
958 arp_tbl.gc_interval);
962 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
963 struct net_device *dev)
965 struct ipoib_neigh *neigh;
967 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
972 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
973 skb_queue_head_init(&neigh->queue);
974 INIT_LIST_HEAD(&neigh->list);
975 ipoib_cm_set(neigh, NULL);
976 /* one ref on behalf of the caller */
977 atomic_set(&neigh->refcnt, 1);
982 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
983 struct net_device *dev)
985 struct ipoib_dev_priv *priv = netdev_priv(dev);
986 struct ipoib_neigh_table *ntbl = &priv->ntbl;
987 struct ipoib_neigh_hash *htbl;
988 struct ipoib_neigh *neigh;
991 htbl = rcu_dereference_protected(ntbl->htbl,
992 lockdep_is_held(&priv->lock));
998 /* need to add a new neigh, but maybe some other thread succeeded?
999 * recalc hash, maybe hash resize took place so we do a search
1001 hash_val = ipoib_addr_hash(htbl, daddr);
1002 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1003 lockdep_is_held(&priv->lock));
1005 neigh = rcu_dereference_protected(neigh->hnext,
1006 lockdep_is_held(&priv->lock))) {
1007 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1008 /* found, take one ref on behalf of the caller */
1009 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1014 neigh->alive = jiffies;
1019 neigh = ipoib_neigh_ctor(daddr, dev);
1023 /* one ref on behalf of the hash table */
1024 atomic_inc(&neigh->refcnt);
1025 neigh->alive = jiffies;
1027 rcu_assign_pointer(neigh->hnext,
1028 rcu_dereference_protected(htbl->buckets[hash_val],
1029 lockdep_is_held(&priv->lock)));
1030 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1031 atomic_inc(&ntbl->entries);
1038 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1040 /* neigh reference count was dropprd to zero */
1041 struct net_device *dev = neigh->dev;
1042 struct ipoib_dev_priv *priv = netdev_priv(dev);
1043 struct sk_buff *skb;
1045 ipoib_put_ah(neigh->ah);
1046 while ((skb = __skb_dequeue(&neigh->queue))) {
1047 ++dev->stats.tx_dropped;
1048 dev_kfree_skb_any(skb);
1050 if (ipoib_cm_get(neigh))
1051 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1052 ipoib_dbg(netdev_priv(dev),
1053 "neigh free for %06x %pI6\n",
1054 IPOIB_QPN(neigh->daddr),
1057 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1058 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1059 complete(&priv->ntbl.flushed);
1063 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1065 /* Called as a result of removal from hash table */
1066 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1067 /* note TX context may hold another ref */
1068 ipoib_neigh_put(neigh);
1071 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1073 struct net_device *dev = neigh->dev;
1074 struct ipoib_dev_priv *priv = netdev_priv(dev);
1075 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1076 struct ipoib_neigh_hash *htbl;
1077 struct ipoib_neigh __rcu **np;
1078 struct ipoib_neigh *n;
1081 htbl = rcu_dereference_protected(ntbl->htbl,
1082 lockdep_is_held(&priv->lock));
1086 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1087 np = &htbl->buckets[hash_val];
1088 for (n = rcu_dereference_protected(*np,
1089 lockdep_is_held(&priv->lock));
1091 n = rcu_dereference_protected(*np,
1092 lockdep_is_held(&priv->lock))) {
1095 rcu_assign_pointer(*np,
1096 rcu_dereference_protected(neigh->hnext,
1097 lockdep_is_held(&priv->lock)));
1098 /* remove from parent list */
1099 list_del(&neigh->list);
1100 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1108 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1110 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1111 struct ipoib_neigh_hash *htbl;
1112 struct ipoib_neigh **buckets;
1115 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1117 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1120 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1121 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1122 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1128 htbl->mask = (size - 1);
1129 htbl->buckets = buckets;
1132 atomic_set(&ntbl->entries, 0);
1134 /* start garbage collection */
1135 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1136 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1137 arp_tbl.gc_interval);
1142 static void neigh_hash_free_rcu(struct rcu_head *head)
1144 struct ipoib_neigh_hash *htbl = container_of(head,
1145 struct ipoib_neigh_hash,
1147 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1148 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1152 complete(&ntbl->deleted);
1155 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1157 struct ipoib_dev_priv *priv = netdev_priv(dev);
1158 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1159 struct ipoib_neigh_hash *htbl;
1160 unsigned long flags;
1163 /* remove all neigh connected to a given path or mcast */
1164 spin_lock_irqsave(&priv->lock, flags);
1166 htbl = rcu_dereference_protected(ntbl->htbl,
1167 lockdep_is_held(&priv->lock));
1172 for (i = 0; i < htbl->size; i++) {
1173 struct ipoib_neigh *neigh;
1174 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1176 while ((neigh = rcu_dereference_protected(*np,
1177 lockdep_is_held(&priv->lock))) != NULL) {
1178 /* delete neighs belong to this parent */
1179 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1180 rcu_assign_pointer(*np,
1181 rcu_dereference_protected(neigh->hnext,
1182 lockdep_is_held(&priv->lock)));
1183 /* remove from parent list */
1184 list_del(&neigh->list);
1185 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1193 spin_unlock_irqrestore(&priv->lock, flags);
1196 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1198 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1199 struct ipoib_neigh_hash *htbl;
1200 unsigned long flags;
1201 int i, wait_flushed = 0;
1203 init_completion(&priv->ntbl.flushed);
1205 spin_lock_irqsave(&priv->lock, flags);
1207 htbl = rcu_dereference_protected(ntbl->htbl,
1208 lockdep_is_held(&priv->lock));
1212 wait_flushed = atomic_read(&priv->ntbl.entries);
1216 for (i = 0; i < htbl->size; i++) {
1217 struct ipoib_neigh *neigh;
1218 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1220 while ((neigh = rcu_dereference_protected(*np,
1221 lockdep_is_held(&priv->lock))) != NULL) {
1222 rcu_assign_pointer(*np,
1223 rcu_dereference_protected(neigh->hnext,
1224 lockdep_is_held(&priv->lock)));
1225 /* remove from path/mc list */
1226 list_del(&neigh->list);
1227 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1232 rcu_assign_pointer(ntbl->htbl, NULL);
1233 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1236 spin_unlock_irqrestore(&priv->lock, flags);
1238 wait_for_completion(&priv->ntbl.flushed);
1241 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1243 struct ipoib_dev_priv *priv = netdev_priv(dev);
1246 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1247 init_completion(&priv->ntbl.deleted);
1248 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1250 /* Stop GC if called at init fail need to cancel work */
1251 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1253 cancel_delayed_work(&priv->neigh_reap_task);
1255 ipoib_flush_neighs(priv);
1257 wait_for_completion(&priv->ntbl.deleted);
1261 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1263 struct ipoib_dev_priv *priv = netdev_priv(dev);
1265 /* Allocate RX/TX "rings" to hold queued skbs */
1266 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1268 if (!priv->rx_ring) {
1269 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1270 ca->name, ipoib_recvq_size);
1274 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1275 if (!priv->tx_ring) {
1276 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1277 ca->name, ipoib_sendq_size);
1278 goto out_rx_ring_cleanup;
1281 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1283 if (ipoib_ib_dev_init(dev, ca, port))
1284 goto out_tx_ring_cleanup;
1287 * Must be after ipoib_ib_dev_init so we can allocate a per
1288 * device wq there and use it here
1290 if (ipoib_neigh_hash_init(priv) < 0)
1291 goto out_dev_uninit;
1296 ipoib_ib_dev_cleanup(dev);
1298 out_tx_ring_cleanup:
1299 vfree(priv->tx_ring);
1301 out_rx_ring_cleanup:
1302 kfree(priv->rx_ring);
1308 void ipoib_dev_cleanup(struct net_device *dev)
1310 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1315 ipoib_delete_debug_files(dev);
1317 /* Delete any child interfaces first */
1318 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1319 /* Stop GC on child */
1320 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1321 cancel_delayed_work(&cpriv->neigh_reap_task);
1322 unregister_netdevice_queue(cpriv->dev, &head);
1324 unregister_netdevice_many(&head);
1327 * Must be before ipoib_ib_dev_cleanup or we delete an in use
1330 ipoib_neigh_hash_uninit(dev);
1332 ipoib_ib_dev_cleanup(dev);
1334 kfree(priv->rx_ring);
1335 vfree(priv->tx_ring);
1337 priv->rx_ring = NULL;
1338 priv->tx_ring = NULL;
1341 static const struct header_ops ipoib_header_ops = {
1342 .create = ipoib_hard_header,
1345 static const struct net_device_ops ipoib_netdev_ops = {
1346 .ndo_uninit = ipoib_uninit,
1347 .ndo_open = ipoib_open,
1348 .ndo_stop = ipoib_stop,
1349 .ndo_change_mtu = ipoib_change_mtu,
1350 .ndo_fix_features = ipoib_fix_features,
1351 .ndo_start_xmit = ipoib_start_xmit,
1352 .ndo_tx_timeout = ipoib_timeout,
1353 .ndo_set_rx_mode = ipoib_set_mcast_list,
1356 void ipoib_setup(struct net_device *dev)
1358 struct ipoib_dev_priv *priv = netdev_priv(dev);
1360 dev->netdev_ops = &ipoib_netdev_ops;
1361 dev->header_ops = &ipoib_header_ops;
1363 ipoib_set_ethtool_ops(dev);
1365 netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
1367 dev->watchdog_timeo = HZ;
1369 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1371 dev->hard_header_len = IPOIB_ENCAP_LEN;
1372 dev->addr_len = INFINIBAND_ALEN;
1373 dev->type = ARPHRD_INFINIBAND;
1374 dev->tx_queue_len = ipoib_sendq_size * 2;
1375 dev->features = (NETIF_F_VLAN_CHALLENGED |
1377 netif_keep_dst(dev);
1379 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1383 spin_lock_init(&priv->lock);
1385 init_rwsem(&priv->vlan_rwsem);
1387 INIT_LIST_HEAD(&priv->path_list);
1388 INIT_LIST_HEAD(&priv->child_intfs);
1389 INIT_LIST_HEAD(&priv->dead_ahs);
1390 INIT_LIST_HEAD(&priv->multicast_list);
1392 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1393 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1394 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1395 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1396 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1397 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1398 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1399 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1402 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1404 struct net_device *dev;
1406 dev = alloc_netdev((int)sizeof(struct ipoib_dev_priv), name,
1407 NET_NAME_UNKNOWN, ipoib_setup);
1411 return netdev_priv(dev);
1414 static ssize_t show_pkey(struct device *dev,
1415 struct device_attribute *attr, char *buf)
1417 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1419 return sprintf(buf, "0x%04x\n", priv->pkey);
1421 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1423 static ssize_t show_umcast(struct device *dev,
1424 struct device_attribute *attr, char *buf)
1426 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1428 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1431 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1433 struct ipoib_dev_priv *priv = netdev_priv(ndev);
1435 if (umcast_val > 0) {
1436 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1437 ipoib_warn(priv, "ignoring multicast groups joined directly "
1440 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1443 static ssize_t set_umcast(struct device *dev,
1444 struct device_attribute *attr,
1445 const char *buf, size_t count)
1447 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1449 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1453 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1455 int ipoib_add_umcast_attr(struct net_device *dev)
1457 return device_create_file(&dev->dev, &dev_attr_umcast);
1460 static ssize_t create_child(struct device *dev,
1461 struct device_attribute *attr,
1462 const char *buf, size_t count)
1467 if (sscanf(buf, "%i", &pkey) != 1)
1470 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
1474 * Set the full membership bit, so that we join the right
1475 * broadcast group, etc.
1479 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1481 return ret ? ret : count;
1483 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1485 static ssize_t delete_child(struct device *dev,
1486 struct device_attribute *attr,
1487 const char *buf, size_t count)
1492 if (sscanf(buf, "%i", &pkey) != 1)
1495 if (pkey < 0 || pkey > 0xffff)
1498 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1500 return ret ? ret : count;
1503 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1505 int ipoib_add_pkey_attr(struct net_device *dev)
1507 return device_create_file(&dev->dev, &dev_attr_pkey);
1510 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1512 struct ib_device_attr *device_attr;
1513 int result = -ENOMEM;
1515 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1517 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1518 hca->name, sizeof *device_attr);
1522 result = ib_query_device(hca, device_attr);
1524 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1529 priv->hca_caps = device_attr->device_cap_flags;
1533 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1534 priv->dev->hw_features = NETIF_F_SG |
1535 NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1537 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1538 priv->dev->hw_features |= NETIF_F_TSO;
1540 priv->dev->features |= priv->dev->hw_features;
1546 static struct net_device *ipoib_add_port(const char *format,
1547 struct ib_device *hca, u8 port)
1549 struct ipoib_dev_priv *priv;
1550 struct ib_port_attr attr;
1551 int result = -ENOMEM;
1553 priv = ipoib_intf_alloc(format);
1555 goto alloc_mem_failed;
1557 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1558 priv->dev->dev_id = port - 1;
1560 if (!ib_query_port(hca, port, &attr))
1561 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1563 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1565 goto device_init_failed;
1568 /* MTU will be reset when mcast join happens */
1569 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1570 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1572 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1574 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1576 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1577 hca->name, port, result);
1578 goto device_init_failed;
1581 if (ipoib_set_dev_features(priv, hca))
1582 goto device_init_failed;
1585 * Set the full membership bit, so that we join the right
1586 * broadcast group, etc.
1588 priv->pkey |= 0x8000;
1590 priv->dev->broadcast[8] = priv->pkey >> 8;
1591 priv->dev->broadcast[9] = priv->pkey & 0xff;
1593 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1595 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1596 hca->name, port, result);
1597 goto device_init_failed;
1599 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1601 result = ipoib_dev_init(priv->dev, hca, port);
1603 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1604 hca->name, port, result);
1605 goto device_init_failed;
1608 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1609 priv->ca, ipoib_event);
1610 result = ib_register_event_handler(&priv->event_handler);
1612 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1613 "port %d (ret = %d)\n",
1614 hca->name, port, result);
1618 result = register_netdev(priv->dev);
1620 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1621 hca->name, port, result);
1622 goto register_failed;
1625 ipoib_create_debug_files(priv->dev);
1627 if (ipoib_cm_add_mode_attr(priv->dev))
1629 if (ipoib_add_pkey_attr(priv->dev))
1631 if (ipoib_add_umcast_attr(priv->dev))
1633 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1635 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1641 ipoib_delete_debug_files(priv->dev);
1642 unregister_netdev(priv->dev);
1645 ib_unregister_event_handler(&priv->event_handler);
1646 /* Stop GC if started before flush */
1647 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1648 cancel_delayed_work(&priv->neigh_reap_task);
1649 flush_workqueue(priv->wq);
1652 ipoib_dev_cleanup(priv->dev);
1655 free_netdev(priv->dev);
1658 return ERR_PTR(result);
1661 static void ipoib_add_one(struct ib_device *device)
1663 struct list_head *dev_list;
1664 struct net_device *dev;
1665 struct ipoib_dev_priv *priv;
1668 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1671 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1675 INIT_LIST_HEAD(dev_list);
1677 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1682 e = device->phys_port_cnt;
1685 for (p = s; p <= e; ++p) {
1686 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1688 dev = ipoib_add_port("ib%d", device, p);
1690 priv = netdev_priv(dev);
1691 list_add_tail(&priv->list, dev_list);
1695 ib_set_client_data(device, &ipoib_client, dev_list);
1698 static void ipoib_remove_one(struct ib_device *device)
1700 struct ipoib_dev_priv *priv, *tmp;
1701 struct list_head *dev_list;
1703 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1706 dev_list = ib_get_client_data(device, &ipoib_client);
1710 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1711 ib_unregister_event_handler(&priv->event_handler);
1714 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1718 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1719 cancel_delayed_work(&priv->neigh_reap_task);
1720 flush_workqueue(priv->wq);
1722 unregister_netdev(priv->dev);
1723 free_netdev(priv->dev);
1729 static int __init ipoib_init_module(void)
1733 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1734 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1735 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1737 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1738 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1739 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1740 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1741 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1745 * When copying small received packets, we only copy from the
1746 * linear data part of the SKB, so we rely on this condition.
1748 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1750 ret = ipoib_register_debugfs();
1755 * We create our own workqueue mainly because we want to be
1756 * able to flush it when devices are being removed. We can't
1757 * use schedule_work()/flush_scheduled_work() because both
1758 * unregister_netdev() and linkwatch_event take the rtnl lock,
1759 * so flush_scheduled_work() can deadlock during device
1762 * In addition, bringing one device up and another down at the
1763 * same time can deadlock a single workqueue, so we have this
1764 * global fallback workqueue, but we also attempt to open a
1765 * per device workqueue each time we bring an interface up
1767 ipoib_workqueue = create_singlethread_workqueue("ipoib_flush");
1768 if (!ipoib_workqueue) {
1773 ib_sa_register_client(&ipoib_sa_client);
1775 ret = ib_register_client(&ipoib_client);
1779 ret = ipoib_netlink_init();
1786 ib_unregister_client(&ipoib_client);
1789 ib_sa_unregister_client(&ipoib_sa_client);
1790 destroy_workqueue(ipoib_workqueue);
1793 ipoib_unregister_debugfs();
1798 static void __exit ipoib_cleanup_module(void)
1800 ipoib_netlink_fini();
1801 ib_unregister_client(&ipoib_client);
1802 ib_sa_unregister_client(&ipoib_sa_client);
1803 ipoib_unregister_debugfs();
1804 destroy_workqueue(ipoib_workqueue);
1807 module_init(ipoib_init_module);
1808 module_exit(ipoib_cleanup_module);