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 */
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
53 MODULE_LICENSE("Dual BSD/GPL");
55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
63 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
64 int ipoib_debug_level;
66 module_param_named(debug_level, ipoib_debug_level, int, 0644);
67 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
70 struct ipoib_path_iter {
71 struct net_device *dev;
72 struct ipoib_path path;
75 static const u8 ipv4_bcast_addr[] = {
76 0x00, 0xff, 0xff, 0xff,
77 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
78 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
81 struct workqueue_struct *ipoib_workqueue;
83 struct ib_sa_client ipoib_sa_client;
85 static void ipoib_add_one(struct ib_device *device);
86 static void ipoib_remove_one(struct ib_device *device);
88 static struct ib_client ipoib_client = {
91 .remove = ipoib_remove_one
94 int ipoib_open(struct net_device *dev)
96 struct ipoib_dev_priv *priv = netdev_priv(dev);
98 ipoib_dbg(priv, "bringing up interface\n");
100 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
102 if (ipoib_pkey_dev_delay_open(dev))
105 if (ipoib_ib_dev_open(dev))
108 if (ipoib_ib_dev_up(dev))
111 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
112 struct ipoib_dev_priv *cpriv;
114 /* Bring up any child interfaces too */
115 mutex_lock(&priv->vlan_mutex);
116 list_for_each_entry(cpriv, &priv->child_intfs, list) {
119 flags = cpriv->dev->flags;
123 dev_change_flags(cpriv->dev, flags | IFF_UP);
125 mutex_unlock(&priv->vlan_mutex);
128 netif_start_queue(dev);
133 ipoib_ib_dev_stop(dev, 1);
136 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
141 static int ipoib_stop(struct net_device *dev)
143 struct ipoib_dev_priv *priv = netdev_priv(dev);
145 ipoib_dbg(priv, "stopping interface\n");
147 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
149 netif_stop_queue(dev);
151 ipoib_ib_dev_down(dev, 1);
152 ipoib_ib_dev_stop(dev, 0);
154 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
155 struct ipoib_dev_priv *cpriv;
157 /* Bring down any child interfaces too */
158 mutex_lock(&priv->vlan_mutex);
159 list_for_each_entry(cpriv, &priv->child_intfs, list) {
162 flags = cpriv->dev->flags;
163 if (!(flags & IFF_UP))
166 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
168 mutex_unlock(&priv->vlan_mutex);
174 static u32 ipoib_fix_features(struct net_device *dev, u32 features)
176 struct ipoib_dev_priv *priv = netdev_priv(dev);
178 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
179 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
184 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
186 struct ipoib_dev_priv *priv = netdev_priv(dev);
188 /* dev->mtu > 2K ==> connected mode */
189 if (ipoib_cm_admin_enabled(dev)) {
190 if (new_mtu > ipoib_cm_max_mtu(dev))
193 if (new_mtu > priv->mcast_mtu)
194 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
201 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
204 priv->admin_mtu = new_mtu;
206 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
211 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
213 struct ipoib_dev_priv *priv = netdev_priv(dev);
214 struct rb_node *n = priv->path_tree.rb_node;
215 struct ipoib_path *path;
219 path = rb_entry(n, struct ipoib_path, rb_node);
221 ret = memcmp(gid, path->pathrec.dgid.raw,
222 sizeof (union ib_gid));
235 static int __path_add(struct net_device *dev, struct ipoib_path *path)
237 struct ipoib_dev_priv *priv = netdev_priv(dev);
238 struct rb_node **n = &priv->path_tree.rb_node;
239 struct rb_node *pn = NULL;
240 struct ipoib_path *tpath;
245 tpath = rb_entry(pn, struct ipoib_path, rb_node);
247 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
248 sizeof (union ib_gid));
257 rb_link_node(&path->rb_node, pn, n);
258 rb_insert_color(&path->rb_node, &priv->path_tree);
260 list_add_tail(&path->list, &priv->path_list);
265 static void path_free(struct net_device *dev, struct ipoib_path *path)
267 struct ipoib_dev_priv *priv = netdev_priv(dev);
268 struct ipoib_neigh *neigh, *tn;
272 while ((skb = __skb_dequeue(&path->queue)))
273 dev_kfree_skb_irq(skb);
275 spin_lock_irqsave(&priv->lock, flags);
277 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
279 * It's safe to call ipoib_put_ah() inside priv->lock
280 * here, because we know that path->ah will always
281 * hold one more reference, so ipoib_put_ah() will
282 * never do more than decrement the ref count.
285 ipoib_put_ah(neigh->ah);
287 ipoib_neigh_free(dev, neigh);
290 spin_unlock_irqrestore(&priv->lock, flags);
293 ipoib_put_ah(path->ah);
298 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
300 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
302 struct ipoib_path_iter *iter;
304 iter = kmalloc(sizeof *iter, GFP_KERNEL);
309 memset(iter->path.pathrec.dgid.raw, 0, 16);
311 if (ipoib_path_iter_next(iter)) {
319 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
321 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
323 struct ipoib_path *path;
326 spin_lock_irq(&priv->lock);
328 n = rb_first(&priv->path_tree);
331 path = rb_entry(n, struct ipoib_path, rb_node);
333 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
334 sizeof (union ib_gid)) < 0) {
343 spin_unlock_irq(&priv->lock);
348 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
349 struct ipoib_path *path)
354 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
356 void ipoib_mark_paths_invalid(struct net_device *dev)
358 struct ipoib_dev_priv *priv = netdev_priv(dev);
359 struct ipoib_path *path, *tp;
361 spin_lock_irq(&priv->lock);
363 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
364 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
365 be16_to_cpu(path->pathrec.dlid),
366 path->pathrec.dgid.raw);
370 spin_unlock_irq(&priv->lock);
373 void ipoib_flush_paths(struct net_device *dev)
375 struct ipoib_dev_priv *priv = netdev_priv(dev);
376 struct ipoib_path *path, *tp;
377 LIST_HEAD(remove_list);
380 netif_tx_lock_bh(dev);
381 spin_lock_irqsave(&priv->lock, flags);
383 list_splice_init(&priv->path_list, &remove_list);
385 list_for_each_entry(path, &remove_list, list)
386 rb_erase(&path->rb_node, &priv->path_tree);
388 list_for_each_entry_safe(path, tp, &remove_list, list) {
390 ib_sa_cancel_query(path->query_id, path->query);
391 spin_unlock_irqrestore(&priv->lock, flags);
392 netif_tx_unlock_bh(dev);
393 wait_for_completion(&path->done);
394 path_free(dev, path);
395 netif_tx_lock_bh(dev);
396 spin_lock_irqsave(&priv->lock, flags);
399 spin_unlock_irqrestore(&priv->lock, flags);
400 netif_tx_unlock_bh(dev);
403 static void path_rec_completion(int status,
404 struct ib_sa_path_rec *pathrec,
407 struct ipoib_path *path = path_ptr;
408 struct net_device *dev = path->dev;
409 struct ipoib_dev_priv *priv = netdev_priv(dev);
410 struct ipoib_ah *ah = NULL;
411 struct ipoib_ah *old_ah = NULL;
412 struct ipoib_neigh *neigh, *tn;
413 struct sk_buff_head skqueue;
418 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
419 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
421 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
422 status, path->pathrec.dgid.raw);
424 skb_queue_head_init(&skqueue);
427 struct ib_ah_attr av;
429 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
430 ah = ipoib_create_ah(dev, priv->pd, &av);
433 spin_lock_irqsave(&priv->lock, flags);
435 if (!IS_ERR_OR_NULL(ah)) {
436 path->pathrec = *pathrec;
441 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
442 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
444 while ((skb = __skb_dequeue(&path->queue)))
445 __skb_queue_tail(&skqueue, skb);
447 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
449 WARN_ON(neigh->ah != old_ah);
451 * Dropping the ah reference inside
452 * priv->lock is safe here, because we
453 * will hold one more reference from
454 * the original value of path->ah (ie
457 ipoib_put_ah(neigh->ah);
459 kref_get(&path->ah->ref);
460 neigh->ah = path->ah;
461 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
462 sizeof(union ib_gid));
464 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
465 if (!ipoib_cm_get(neigh))
466 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
469 if (!ipoib_cm_get(neigh)) {
470 list_del(&neigh->list);
472 ipoib_put_ah(neigh->ah);
473 ipoib_neigh_free(dev, neigh);
478 while ((skb = __skb_dequeue(&neigh->queue)))
479 __skb_queue_tail(&skqueue, skb);
485 complete(&path->done);
487 spin_unlock_irqrestore(&priv->lock, flags);
490 ipoib_put_ah(old_ah);
492 while ((skb = __skb_dequeue(&skqueue))) {
494 if (dev_queue_xmit(skb))
495 ipoib_warn(priv, "dev_queue_xmit failed "
496 "to requeue packet\n");
500 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
502 struct ipoib_dev_priv *priv = netdev_priv(dev);
503 struct ipoib_path *path;
505 if (!priv->broadcast)
508 path = kzalloc(sizeof *path, GFP_ATOMIC);
514 skb_queue_head_init(&path->queue);
516 INIT_LIST_HEAD(&path->neigh_list);
518 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
519 path->pathrec.sgid = priv->local_gid;
520 path->pathrec.pkey = cpu_to_be16(priv->pkey);
521 path->pathrec.numb_path = 1;
522 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
527 static int path_rec_start(struct net_device *dev,
528 struct ipoib_path *path)
530 struct ipoib_dev_priv *priv = netdev_priv(dev);
532 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
533 path->pathrec.dgid.raw);
535 init_completion(&path->done);
538 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
540 IB_SA_PATH_REC_DGID |
541 IB_SA_PATH_REC_SGID |
542 IB_SA_PATH_REC_NUMB_PATH |
543 IB_SA_PATH_REC_TRAFFIC_CLASS |
548 if (path->query_id < 0) {
549 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
551 complete(&path->done);
552 return path->query_id;
558 /* called with rcu_read_lock */
559 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
561 struct ipoib_dev_priv *priv = netdev_priv(dev);
562 struct ipoib_path *path;
563 struct ipoib_neigh *neigh;
567 n = dst_get_neighbour(skb_dst(skb));
568 neigh = ipoib_neigh_alloc(n, skb->dev);
570 ++dev->stats.tx_dropped;
571 dev_kfree_skb_any(skb);
575 spin_lock_irqsave(&priv->lock, flags);
577 path = __path_find(dev, n->ha + 4);
579 path = path_rec_create(dev, n->ha + 4);
583 __path_add(dev, path);
586 list_add_tail(&neigh->list, &path->neigh_list);
589 kref_get(&path->ah->ref);
590 neigh->ah = path->ah;
591 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
592 sizeof(union ib_gid));
594 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
595 if (!ipoib_cm_get(neigh))
596 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
597 if (!ipoib_cm_get(neigh)) {
598 list_del(&neigh->list);
600 ipoib_put_ah(neigh->ah);
601 ipoib_neigh_free(dev, neigh);
604 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
605 __skb_queue_tail(&neigh->queue, skb);
607 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
608 skb_queue_len(&neigh->queue));
612 spin_unlock_irqrestore(&priv->lock, flags);
613 ipoib_send(dev, skb, path->ah, IPOIB_QPN(n->ha));
619 if (!path->query && path_rec_start(dev, path))
622 __skb_queue_tail(&neigh->queue, skb);
625 spin_unlock_irqrestore(&priv->lock, flags);
629 list_del(&neigh->list);
632 ipoib_neigh_free(dev, neigh);
634 ++dev->stats.tx_dropped;
635 dev_kfree_skb_any(skb);
637 spin_unlock_irqrestore(&priv->lock, flags);
640 /* called with rcu_read_lock */
641 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
643 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
644 struct dst_entry *dst = skb_dst(skb);
647 /* Look up path record for unicasts */
648 n = dst_get_neighbour(dst);
649 if (n->ha[4] != 0xff) {
650 neigh_add_path(skb, dev);
654 /* Add in the P_Key for multicasts */
655 n->ha[8] = (priv->pkey >> 8) & 0xff;
656 n->ha[9] = priv->pkey & 0xff;
657 ipoib_mcast_send(dev, n->ha + 4, skb);
660 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
663 struct ipoib_dev_priv *priv = netdev_priv(dev);
664 struct ipoib_path *path;
667 spin_lock_irqsave(&priv->lock, flags);
669 path = __path_find(dev, cb->hwaddr + 4);
670 if (!path || !path->valid) {
674 path = path_rec_create(dev, cb->hwaddr + 4);
678 __skb_queue_tail(&path->queue, skb);
680 if (!path->query && path_rec_start(dev, path)) {
681 spin_unlock_irqrestore(&priv->lock, flags);
683 path_free(dev, path);
686 __path_add(dev, path);
688 ++dev->stats.tx_dropped;
689 dev_kfree_skb_any(skb);
692 spin_unlock_irqrestore(&priv->lock, flags);
697 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
698 be16_to_cpu(path->pathrec.dlid));
700 spin_unlock_irqrestore(&priv->lock, flags);
701 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
703 } else if ((path->query || !path_rec_start(dev, path)) &&
704 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
705 __skb_queue_tail(&path->queue, skb);
707 ++dev->stats.tx_dropped;
708 dev_kfree_skb_any(skb);
711 spin_unlock_irqrestore(&priv->lock, flags);
714 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
716 struct ipoib_dev_priv *priv = netdev_priv(dev);
717 struct ipoib_neigh *neigh;
718 struct neighbour *n = NULL;
722 if (likely(skb_dst(skb)))
723 n = dst_get_neighbour(skb_dst(skb));
726 if (unlikely(!*to_ipoib_neigh(n))) {
727 ipoib_path_lookup(skb, dev);
731 neigh = *to_ipoib_neigh(n);
733 if (unlikely((memcmp(&neigh->dgid.raw,
735 sizeof(union ib_gid))) ||
736 (neigh->dev != dev))) {
737 spin_lock_irqsave(&priv->lock, flags);
739 * It's safe to call ipoib_put_ah() inside
740 * priv->lock here, because we know that
741 * path->ah will always hold one more reference,
742 * so ipoib_put_ah() will never do more than
743 * decrement the ref count.
746 ipoib_put_ah(neigh->ah);
747 list_del(&neigh->list);
748 ipoib_neigh_free(dev, neigh);
749 spin_unlock_irqrestore(&priv->lock, flags);
750 ipoib_path_lookup(skb, dev);
754 if (ipoib_cm_get(neigh)) {
755 if (ipoib_cm_up(neigh)) {
756 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
759 } else if (neigh->ah) {
760 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(n->ha));
764 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
765 spin_lock_irqsave(&priv->lock, flags);
766 __skb_queue_tail(&neigh->queue, skb);
767 spin_unlock_irqrestore(&priv->lock, flags);
769 ++dev->stats.tx_dropped;
770 dev_kfree_skb_any(skb);
773 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
775 if (cb->hwaddr[4] == 0xff) {
776 /* Add in the P_Key for multicast*/
777 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
778 cb->hwaddr[9] = priv->pkey & 0xff;
780 ipoib_mcast_send(dev, cb->hwaddr + 4, skb);
782 /* unicast GID -- should be ARP or RARP reply */
784 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
785 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
786 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
787 skb_dst(skb) ? "neigh" : "dst",
788 be16_to_cpup((__be16 *) skb->data),
789 IPOIB_QPN(cb->hwaddr),
791 dev_kfree_skb_any(skb);
792 ++dev->stats.tx_dropped;
796 unicast_arp_send(skb, dev, cb);
804 static void ipoib_timeout(struct net_device *dev)
806 struct ipoib_dev_priv *priv = netdev_priv(dev);
808 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
809 jiffies_to_msecs(jiffies - dev->trans_start));
810 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
811 netif_queue_stopped(dev),
812 priv->tx_head, priv->tx_tail);
813 /* XXX reset QP, etc. */
816 static int ipoib_hard_header(struct sk_buff *skb,
817 struct net_device *dev,
819 const void *daddr, const void *saddr, unsigned len)
821 struct ipoib_header *header;
823 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
825 header->proto = htons(type);
826 header->reserved = 0;
829 * If we don't have a dst_entry structure, stuff the
830 * destination address into skb->cb so we can figure out where
831 * to send the packet later.
834 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
835 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
841 static void ipoib_set_mcast_list(struct net_device *dev)
843 struct ipoib_dev_priv *priv = netdev_priv(dev);
845 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
846 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
850 queue_work(ipoib_workqueue, &priv->restart_task);
853 static void ipoib_neigh_cleanup(struct neighbour *n)
855 struct ipoib_neigh *neigh;
856 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
858 struct ipoib_ah *ah = NULL;
860 neigh = *to_ipoib_neigh(n);
862 priv = netdev_priv(neigh->dev);
866 "neigh_cleanup for %06x %pI6\n",
870 spin_lock_irqsave(&priv->lock, flags);
874 list_del(&neigh->list);
875 ipoib_neigh_free(n->dev, neigh);
877 spin_unlock_irqrestore(&priv->lock, flags);
883 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
884 struct net_device *dev)
886 struct ipoib_neigh *neigh;
888 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
892 neigh->neighbour = neighbour;
894 memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
895 *to_ipoib_neigh(neighbour) = neigh;
896 skb_queue_head_init(&neigh->queue);
897 ipoib_cm_set(neigh, NULL);
902 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
905 *to_ipoib_neigh(neigh->neighbour) = NULL;
906 while ((skb = __skb_dequeue(&neigh->queue))) {
907 ++dev->stats.tx_dropped;
908 dev_kfree_skb_any(skb);
910 if (ipoib_cm_get(neigh))
911 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
915 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
917 parms->neigh_cleanup = ipoib_neigh_cleanup;
922 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
924 struct ipoib_dev_priv *priv = netdev_priv(dev);
926 /* Allocate RX/TX "rings" to hold queued skbs */
927 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
929 if (!priv->rx_ring) {
930 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
931 ca->name, ipoib_recvq_size);
935 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
936 if (!priv->tx_ring) {
937 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
938 ca->name, ipoib_sendq_size);
939 goto out_rx_ring_cleanup;
942 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
944 if (ipoib_ib_dev_init(dev, ca, port))
945 goto out_tx_ring_cleanup;
950 vfree(priv->tx_ring);
953 kfree(priv->rx_ring);
959 void ipoib_dev_cleanup(struct net_device *dev)
961 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
963 ipoib_delete_debug_files(dev);
965 /* Delete any child interfaces first */
966 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
967 unregister_netdev(cpriv->dev);
968 ipoib_dev_cleanup(cpriv->dev);
969 free_netdev(cpriv->dev);
972 ipoib_ib_dev_cleanup(dev);
974 kfree(priv->rx_ring);
975 vfree(priv->tx_ring);
977 priv->rx_ring = NULL;
978 priv->tx_ring = NULL;
981 static const struct header_ops ipoib_header_ops = {
982 .create = ipoib_hard_header,
985 static const struct net_device_ops ipoib_netdev_ops = {
986 .ndo_open = ipoib_open,
987 .ndo_stop = ipoib_stop,
988 .ndo_change_mtu = ipoib_change_mtu,
989 .ndo_fix_features = ipoib_fix_features,
990 .ndo_start_xmit = ipoib_start_xmit,
991 .ndo_tx_timeout = ipoib_timeout,
992 .ndo_set_rx_mode = ipoib_set_mcast_list,
993 .ndo_neigh_setup = ipoib_neigh_setup_dev,
996 static void ipoib_setup(struct net_device *dev)
998 struct ipoib_dev_priv *priv = netdev_priv(dev);
1000 dev->netdev_ops = &ipoib_netdev_ops;
1001 dev->header_ops = &ipoib_header_ops;
1003 ipoib_set_ethtool_ops(dev);
1005 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1007 dev->watchdog_timeo = HZ;
1009 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1011 dev->hard_header_len = IPOIB_ENCAP_LEN;
1012 dev->addr_len = INFINIBAND_ALEN;
1013 dev->type = ARPHRD_INFINIBAND;
1014 dev->tx_queue_len = ipoib_sendq_size * 2;
1015 dev->features = (NETIF_F_VLAN_CHALLENGED |
1017 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
1019 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1021 netif_carrier_off(dev);
1025 spin_lock_init(&priv->lock);
1027 mutex_init(&priv->vlan_mutex);
1029 INIT_LIST_HEAD(&priv->path_list);
1030 INIT_LIST_HEAD(&priv->child_intfs);
1031 INIT_LIST_HEAD(&priv->dead_ahs);
1032 INIT_LIST_HEAD(&priv->multicast_list);
1034 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1035 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1036 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1037 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1038 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1039 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1040 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1041 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1044 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1046 struct net_device *dev;
1048 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1053 return netdev_priv(dev);
1056 static ssize_t show_pkey(struct device *dev,
1057 struct device_attribute *attr, char *buf)
1059 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1061 return sprintf(buf, "0x%04x\n", priv->pkey);
1063 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1065 static ssize_t show_umcast(struct device *dev,
1066 struct device_attribute *attr, char *buf)
1068 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1070 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1073 static ssize_t set_umcast(struct device *dev,
1074 struct device_attribute *attr,
1075 const char *buf, size_t count)
1077 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1078 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1080 if (umcast_val > 0) {
1081 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1082 ipoib_warn(priv, "ignoring multicast groups joined directly "
1085 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1089 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1091 int ipoib_add_umcast_attr(struct net_device *dev)
1093 return device_create_file(&dev->dev, &dev_attr_umcast);
1096 static ssize_t create_child(struct device *dev,
1097 struct device_attribute *attr,
1098 const char *buf, size_t count)
1103 if (sscanf(buf, "%i", &pkey) != 1)
1106 if (pkey < 0 || pkey > 0xffff)
1110 * Set the full membership bit, so that we join the right
1111 * broadcast group, etc.
1115 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1117 return ret ? ret : count;
1119 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1121 static ssize_t delete_child(struct device *dev,
1122 struct device_attribute *attr,
1123 const char *buf, size_t count)
1128 if (sscanf(buf, "%i", &pkey) != 1)
1131 if (pkey < 0 || pkey > 0xffff)
1134 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1136 return ret ? ret : count;
1139 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1141 int ipoib_add_pkey_attr(struct net_device *dev)
1143 return device_create_file(&dev->dev, &dev_attr_pkey);
1146 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1148 struct ib_device_attr *device_attr;
1149 int result = -ENOMEM;
1151 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1153 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1154 hca->name, sizeof *device_attr);
1158 result = ib_query_device(hca, device_attr);
1160 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1165 priv->hca_caps = device_attr->device_cap_flags;
1169 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1170 priv->dev->hw_features = NETIF_F_SG |
1171 NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1173 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1174 priv->dev->hw_features |= NETIF_F_TSO;
1176 priv->dev->features |= priv->dev->hw_features;
1182 static struct net_device *ipoib_add_port(const char *format,
1183 struct ib_device *hca, u8 port)
1185 struct ipoib_dev_priv *priv;
1186 struct ib_port_attr attr;
1187 int result = -ENOMEM;
1189 priv = ipoib_intf_alloc(format);
1191 goto alloc_mem_failed;
1193 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1194 priv->dev->dev_id = port - 1;
1196 if (!ib_query_port(hca, port, &attr))
1197 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1199 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1201 goto device_init_failed;
1204 /* MTU will be reset when mcast join happens */
1205 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1206 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1208 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1210 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1211 hca->name, port, result);
1212 goto device_init_failed;
1215 if (ipoib_set_dev_features(priv, hca))
1216 goto device_init_failed;
1219 * Set the full membership bit, so that we join the right
1220 * broadcast group, etc.
1222 priv->pkey |= 0x8000;
1224 priv->dev->broadcast[8] = priv->pkey >> 8;
1225 priv->dev->broadcast[9] = priv->pkey & 0xff;
1227 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1229 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1230 hca->name, port, result);
1231 goto device_init_failed;
1233 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1235 result = ipoib_dev_init(priv->dev, hca, port);
1237 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1238 hca->name, port, result);
1239 goto device_init_failed;
1242 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1243 priv->ca, ipoib_event);
1244 result = ib_register_event_handler(&priv->event_handler);
1246 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1247 "port %d (ret = %d)\n",
1248 hca->name, port, result);
1252 result = register_netdev(priv->dev);
1254 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1255 hca->name, port, result);
1256 goto register_failed;
1259 ipoib_create_debug_files(priv->dev);
1261 if (ipoib_cm_add_mode_attr(priv->dev))
1263 if (ipoib_add_pkey_attr(priv->dev))
1265 if (ipoib_add_umcast_attr(priv->dev))
1267 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1269 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1275 ipoib_delete_debug_files(priv->dev);
1276 unregister_netdev(priv->dev);
1279 ib_unregister_event_handler(&priv->event_handler);
1280 flush_workqueue(ipoib_workqueue);
1283 ipoib_dev_cleanup(priv->dev);
1286 free_netdev(priv->dev);
1289 return ERR_PTR(result);
1292 static void ipoib_add_one(struct ib_device *device)
1294 struct list_head *dev_list;
1295 struct net_device *dev;
1296 struct ipoib_dev_priv *priv;
1299 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1302 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1306 INIT_LIST_HEAD(dev_list);
1308 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1313 e = device->phys_port_cnt;
1316 for (p = s; p <= e; ++p) {
1317 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1319 dev = ipoib_add_port("ib%d", device, p);
1321 priv = netdev_priv(dev);
1322 list_add_tail(&priv->list, dev_list);
1326 ib_set_client_data(device, &ipoib_client, dev_list);
1329 static void ipoib_remove_one(struct ib_device *device)
1331 struct ipoib_dev_priv *priv, *tmp;
1332 struct list_head *dev_list;
1334 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1337 dev_list = ib_get_client_data(device, &ipoib_client);
1339 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1340 ib_unregister_event_handler(&priv->event_handler);
1343 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1346 flush_workqueue(ipoib_workqueue);
1348 unregister_netdev(priv->dev);
1349 ipoib_dev_cleanup(priv->dev);
1350 free_netdev(priv->dev);
1356 static int __init ipoib_init_module(void)
1360 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1361 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1362 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1364 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1365 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1366 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1367 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1368 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1372 * When copying small received packets, we only copy from the
1373 * linear data part of the SKB, so we rely on this condition.
1375 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1377 ret = ipoib_register_debugfs();
1382 * We create our own workqueue mainly because we want to be
1383 * able to flush it when devices are being removed. We can't
1384 * use schedule_work()/flush_scheduled_work() because both
1385 * unregister_netdev() and linkwatch_event take the rtnl lock,
1386 * so flush_scheduled_work() can deadlock during device
1389 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1390 if (!ipoib_workqueue) {
1395 ib_sa_register_client(&ipoib_sa_client);
1397 ret = ib_register_client(&ipoib_client);
1404 ib_sa_unregister_client(&ipoib_sa_client);
1405 destroy_workqueue(ipoib_workqueue);
1408 ipoib_unregister_debugfs();
1413 static void __exit ipoib_cleanup_module(void)
1415 ib_unregister_client(&ipoib_client);
1416 ib_sa_unregister_client(&ipoib_sa_client);
1417 ipoib_unregister_debugfs();
1418 destroy_workqueue(ipoib_workqueue);
1421 module_init(ipoib_init_module);
1422 module_exit(ipoib_cleanup_module);