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, 0);
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 int ipoib_change_mtu(struct net_device *dev, int new_mtu)
176 struct ipoib_dev_priv *priv = netdev_priv(dev);
178 /* dev->mtu > 2K ==> connected mode */
179 if (ipoib_cm_admin_enabled(dev)) {
180 if (new_mtu > ipoib_cm_max_mtu(dev))
183 if (new_mtu > priv->mcast_mtu)
184 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
191 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
194 priv->admin_mtu = new_mtu;
196 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
203 struct ipoib_dev_priv *priv = netdev_priv(dev);
204 struct rb_node *n = priv->path_tree.rb_node;
205 struct ipoib_path *path;
209 path = rb_entry(n, struct ipoib_path, rb_node);
211 ret = memcmp(gid, path->pathrec.dgid.raw,
212 sizeof (union ib_gid));
225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
227 struct ipoib_dev_priv *priv = netdev_priv(dev);
228 struct rb_node **n = &priv->path_tree.rb_node;
229 struct rb_node *pn = NULL;
230 struct ipoib_path *tpath;
235 tpath = rb_entry(pn, struct ipoib_path, rb_node);
237 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
238 sizeof (union ib_gid));
247 rb_link_node(&path->rb_node, pn, n);
248 rb_insert_color(&path->rb_node, &priv->path_tree);
250 list_add_tail(&path->list, &priv->path_list);
255 static void path_free(struct net_device *dev, struct ipoib_path *path)
257 struct ipoib_dev_priv *priv = netdev_priv(dev);
258 struct ipoib_neigh *neigh, *tn;
262 while ((skb = __skb_dequeue(&path->queue)))
263 dev_kfree_skb_irq(skb);
265 spin_lock_irqsave(&priv->lock, flags);
267 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
269 * It's safe to call ipoib_put_ah() inside priv->lock
270 * here, because we know that path->ah will always
271 * hold one more reference, so ipoib_put_ah() will
272 * never do more than decrement the ref count.
275 ipoib_put_ah(neigh->ah);
277 ipoib_neigh_free(dev, neigh);
280 spin_unlock_irqrestore(&priv->lock, flags);
283 ipoib_put_ah(path->ah);
288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
292 struct ipoib_path_iter *iter;
294 iter = kmalloc(sizeof *iter, GFP_KERNEL);
299 memset(iter->path.pathrec.dgid.raw, 0, 16);
301 if (ipoib_path_iter_next(iter)) {
309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
311 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
313 struct ipoib_path *path;
316 spin_lock_irq(&priv->lock);
318 n = rb_first(&priv->path_tree);
321 path = rb_entry(n, struct ipoib_path, rb_node);
323 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
324 sizeof (union ib_gid)) < 0) {
333 spin_unlock_irq(&priv->lock);
338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
339 struct ipoib_path *path)
344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
346 void ipoib_mark_paths_invalid(struct net_device *dev)
348 struct ipoib_dev_priv *priv = netdev_priv(dev);
349 struct ipoib_path *path, *tp;
351 spin_lock_irq(&priv->lock);
353 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
354 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
355 be16_to_cpu(path->pathrec.dlid),
356 path->pathrec.dgid.raw);
360 spin_unlock_irq(&priv->lock);
363 void ipoib_flush_paths(struct net_device *dev)
365 struct ipoib_dev_priv *priv = netdev_priv(dev);
366 struct ipoib_path *path, *tp;
367 LIST_HEAD(remove_list);
370 netif_tx_lock_bh(dev);
371 spin_lock_irqsave(&priv->lock, flags);
373 list_splice_init(&priv->path_list, &remove_list);
375 list_for_each_entry(path, &remove_list, list)
376 rb_erase(&path->rb_node, &priv->path_tree);
378 list_for_each_entry_safe(path, tp, &remove_list, list) {
380 ib_sa_cancel_query(path->query_id, path->query);
381 spin_unlock_irqrestore(&priv->lock, flags);
382 netif_tx_unlock_bh(dev);
383 wait_for_completion(&path->done);
384 path_free(dev, path);
385 netif_tx_lock_bh(dev);
386 spin_lock_irqsave(&priv->lock, flags);
389 spin_unlock_irqrestore(&priv->lock, flags);
390 netif_tx_unlock_bh(dev);
393 static void path_rec_completion(int status,
394 struct ib_sa_path_rec *pathrec,
397 struct ipoib_path *path = path_ptr;
398 struct net_device *dev = path->dev;
399 struct ipoib_dev_priv *priv = netdev_priv(dev);
400 struct ipoib_ah *ah = NULL;
401 struct ipoib_ah *old_ah = NULL;
402 struct ipoib_neigh *neigh, *tn;
403 struct sk_buff_head skqueue;
408 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
409 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
411 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
412 status, path->pathrec.dgid.raw);
414 skb_queue_head_init(&skqueue);
417 struct ib_ah_attr av;
419 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
420 ah = ipoib_create_ah(dev, priv->pd, &av);
423 spin_lock_irqsave(&priv->lock, flags);
426 path->pathrec = *pathrec;
431 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
432 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
434 while ((skb = __skb_dequeue(&path->queue)))
435 __skb_queue_tail(&skqueue, skb);
437 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
439 WARN_ON(neigh->ah != old_ah);
441 * Dropping the ah reference inside
442 * priv->lock is safe here, because we
443 * will hold one more reference from
444 * the original value of path->ah (ie
447 ipoib_put_ah(neigh->ah);
449 kref_get(&path->ah->ref);
450 neigh->ah = path->ah;
451 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
452 sizeof(union ib_gid));
454 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
455 if (!ipoib_cm_get(neigh))
456 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
459 if (!ipoib_cm_get(neigh)) {
460 list_del(&neigh->list);
462 ipoib_put_ah(neigh->ah);
463 ipoib_neigh_free(dev, neigh);
468 while ((skb = __skb_dequeue(&neigh->queue)))
469 __skb_queue_tail(&skqueue, skb);
475 complete(&path->done);
477 spin_unlock_irqrestore(&priv->lock, flags);
480 ipoib_put_ah(old_ah);
482 while ((skb = __skb_dequeue(&skqueue))) {
484 if (dev_queue_xmit(skb))
485 ipoib_warn(priv, "dev_queue_xmit failed "
486 "to requeue packet\n");
490 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
492 struct ipoib_dev_priv *priv = netdev_priv(dev);
493 struct ipoib_path *path;
495 if (!priv->broadcast)
498 path = kzalloc(sizeof *path, GFP_ATOMIC);
504 skb_queue_head_init(&path->queue);
506 INIT_LIST_HEAD(&path->neigh_list);
508 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
509 path->pathrec.sgid = priv->local_gid;
510 path->pathrec.pkey = cpu_to_be16(priv->pkey);
511 path->pathrec.numb_path = 1;
512 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
517 static int path_rec_start(struct net_device *dev,
518 struct ipoib_path *path)
520 struct ipoib_dev_priv *priv = netdev_priv(dev);
522 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
523 path->pathrec.dgid.raw);
525 init_completion(&path->done);
528 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
530 IB_SA_PATH_REC_DGID |
531 IB_SA_PATH_REC_SGID |
532 IB_SA_PATH_REC_NUMB_PATH |
533 IB_SA_PATH_REC_TRAFFIC_CLASS |
538 if (path->query_id < 0) {
539 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
541 complete(&path->done);
542 return path->query_id;
548 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
550 struct ipoib_dev_priv *priv = netdev_priv(dev);
551 struct ipoib_path *path;
552 struct ipoib_neigh *neigh;
555 neigh = ipoib_neigh_alloc(skb_dst(skb)->neighbour, skb->dev);
557 ++dev->stats.tx_dropped;
558 dev_kfree_skb_any(skb);
562 spin_lock_irqsave(&priv->lock, flags);
564 path = __path_find(dev, skb_dst(skb)->neighbour->ha + 4);
566 path = path_rec_create(dev, skb_dst(skb)->neighbour->ha + 4);
570 __path_add(dev, path);
573 list_add_tail(&neigh->list, &path->neigh_list);
576 kref_get(&path->ah->ref);
577 neigh->ah = path->ah;
578 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
579 sizeof(union ib_gid));
581 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
582 if (!ipoib_cm_get(neigh))
583 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
584 if (!ipoib_cm_get(neigh)) {
585 list_del(&neigh->list);
587 ipoib_put_ah(neigh->ah);
588 ipoib_neigh_free(dev, neigh);
591 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
592 __skb_queue_tail(&neigh->queue, skb);
594 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
595 skb_queue_len(&neigh->queue));
599 spin_unlock_irqrestore(&priv->lock, flags);
600 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
606 if (!path->query && path_rec_start(dev, path))
609 __skb_queue_tail(&neigh->queue, skb);
612 spin_unlock_irqrestore(&priv->lock, flags);
616 list_del(&neigh->list);
619 ipoib_neigh_free(dev, neigh);
621 ++dev->stats.tx_dropped;
622 dev_kfree_skb_any(skb);
624 spin_unlock_irqrestore(&priv->lock, flags);
627 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
629 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
631 /* Look up path record for unicasts */
632 if (skb_dst(skb)->neighbour->ha[4] != 0xff) {
633 neigh_add_path(skb, dev);
637 /* Add in the P_Key for multicasts */
638 skb_dst(skb)->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
639 skb_dst(skb)->neighbour->ha[9] = priv->pkey & 0xff;
640 ipoib_mcast_send(dev, skb_dst(skb)->neighbour->ha + 4, skb);
643 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
644 struct ipoib_pseudoheader *phdr)
646 struct ipoib_dev_priv *priv = netdev_priv(dev);
647 struct ipoib_path *path;
650 spin_lock_irqsave(&priv->lock, flags);
652 path = __path_find(dev, phdr->hwaddr + 4);
653 if (!path || !path->valid) {
657 path = path_rec_create(dev, phdr->hwaddr + 4);
661 /* put pseudoheader back on for next time */
662 skb_push(skb, sizeof *phdr);
663 __skb_queue_tail(&path->queue, skb);
665 if (!path->query && path_rec_start(dev, path)) {
666 spin_unlock_irqrestore(&priv->lock, flags);
668 path_free(dev, path);
671 __path_add(dev, path);
673 ++dev->stats.tx_dropped;
674 dev_kfree_skb_any(skb);
677 spin_unlock_irqrestore(&priv->lock, flags);
682 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
683 be16_to_cpu(path->pathrec.dlid));
685 spin_unlock_irqrestore(&priv->lock, flags);
686 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
688 } else if ((path->query || !path_rec_start(dev, path)) &&
689 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
690 /* put pseudoheader back on for next time */
691 skb_push(skb, sizeof *phdr);
692 __skb_queue_tail(&path->queue, skb);
694 ++dev->stats.tx_dropped;
695 dev_kfree_skb_any(skb);
698 spin_unlock_irqrestore(&priv->lock, flags);
701 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
703 struct ipoib_dev_priv *priv = netdev_priv(dev);
704 struct ipoib_neigh *neigh;
707 if (likely(skb_dst(skb) && skb_dst(skb)->neighbour)) {
708 if (unlikely(!*to_ipoib_neigh(skb_dst(skb)->neighbour))) {
709 ipoib_path_lookup(skb, dev);
713 neigh = *to_ipoib_neigh(skb_dst(skb)->neighbour);
715 if (unlikely((memcmp(&neigh->dgid.raw,
716 skb_dst(skb)->neighbour->ha + 4,
717 sizeof(union ib_gid))) ||
718 (neigh->dev != dev))) {
719 spin_lock_irqsave(&priv->lock, flags);
721 * It's safe to call ipoib_put_ah() inside
722 * priv->lock here, because we know that
723 * path->ah will always hold one more reference,
724 * so ipoib_put_ah() will never do more than
725 * decrement the ref count.
728 ipoib_put_ah(neigh->ah);
729 list_del(&neigh->list);
730 ipoib_neigh_free(dev, neigh);
731 spin_unlock_irqrestore(&priv->lock, flags);
732 ipoib_path_lookup(skb, dev);
736 if (ipoib_cm_get(neigh)) {
737 if (ipoib_cm_up(neigh)) {
738 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
741 } else if (neigh->ah) {
742 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb_dst(skb)->neighbour->ha));
746 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
747 spin_lock_irqsave(&priv->lock, flags);
748 __skb_queue_tail(&neigh->queue, skb);
749 spin_unlock_irqrestore(&priv->lock, flags);
751 ++dev->stats.tx_dropped;
752 dev_kfree_skb_any(skb);
755 struct ipoib_pseudoheader *phdr =
756 (struct ipoib_pseudoheader *) skb->data;
757 skb_pull(skb, sizeof *phdr);
759 if (phdr->hwaddr[4] == 0xff) {
760 /* Add in the P_Key for multicast*/
761 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
762 phdr->hwaddr[9] = priv->pkey & 0xff;
764 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
766 /* unicast GID -- should be ARP or RARP reply */
768 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
769 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
770 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %pI6\n",
771 skb_dst(skb) ? "neigh" : "dst",
772 be16_to_cpup((__be16 *) skb->data),
773 IPOIB_QPN(phdr->hwaddr),
775 dev_kfree_skb_any(skb);
776 ++dev->stats.tx_dropped;
780 unicast_arp_send(skb, dev, phdr);
787 static void ipoib_timeout(struct net_device *dev)
789 struct ipoib_dev_priv *priv = netdev_priv(dev);
791 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
792 jiffies_to_msecs(jiffies - dev->trans_start));
793 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
794 netif_queue_stopped(dev),
795 priv->tx_head, priv->tx_tail);
796 /* XXX reset QP, etc. */
799 static int ipoib_hard_header(struct sk_buff *skb,
800 struct net_device *dev,
802 const void *daddr, const void *saddr, unsigned len)
804 struct ipoib_header *header;
806 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
808 header->proto = htons(type);
809 header->reserved = 0;
812 * If we don't have a neighbour structure, stuff the
813 * destination address onto the front of the skb so we can
814 * figure out where to send the packet later.
816 if ((!skb_dst(skb) || !skb_dst(skb)->neighbour) && daddr) {
817 struct ipoib_pseudoheader *phdr =
818 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
819 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
825 static void ipoib_set_mcast_list(struct net_device *dev)
827 struct ipoib_dev_priv *priv = netdev_priv(dev);
829 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
830 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
834 queue_work(ipoib_workqueue, &priv->restart_task);
837 static void ipoib_neigh_cleanup(struct neighbour *n)
839 struct ipoib_neigh *neigh;
840 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
842 struct ipoib_ah *ah = NULL;
844 neigh = *to_ipoib_neigh(n);
846 priv = netdev_priv(neigh->dev);
850 "neigh_cleanup for %06x %pI6\n",
854 spin_lock_irqsave(&priv->lock, flags);
858 list_del(&neigh->list);
859 ipoib_neigh_free(n->dev, neigh);
861 spin_unlock_irqrestore(&priv->lock, flags);
867 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
868 struct net_device *dev)
870 struct ipoib_neigh *neigh;
872 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
876 neigh->neighbour = neighbour;
878 memset(&neigh->dgid.raw, 0, sizeof (union ib_gid));
879 *to_ipoib_neigh(neighbour) = neigh;
880 skb_queue_head_init(&neigh->queue);
881 ipoib_cm_set(neigh, NULL);
886 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
889 *to_ipoib_neigh(neigh->neighbour) = NULL;
890 while ((skb = __skb_dequeue(&neigh->queue))) {
891 ++dev->stats.tx_dropped;
892 dev_kfree_skb_any(skb);
894 if (ipoib_cm_get(neigh))
895 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
899 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
901 parms->neigh_cleanup = ipoib_neigh_cleanup;
906 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
908 struct ipoib_dev_priv *priv = netdev_priv(dev);
910 /* Allocate RX/TX "rings" to hold queued skbs */
911 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
913 if (!priv->rx_ring) {
914 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
915 ca->name, ipoib_recvq_size);
919 priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
920 if (!priv->tx_ring) {
921 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
922 ca->name, ipoib_sendq_size);
923 goto out_rx_ring_cleanup;
925 memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
927 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
929 if (ipoib_ib_dev_init(dev, ca, port))
930 goto out_tx_ring_cleanup;
935 vfree(priv->tx_ring);
938 kfree(priv->rx_ring);
944 void ipoib_dev_cleanup(struct net_device *dev)
946 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
948 ipoib_delete_debug_files(dev);
950 /* Delete any child interfaces first */
951 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
952 unregister_netdev(cpriv->dev);
953 ipoib_dev_cleanup(cpriv->dev);
954 free_netdev(cpriv->dev);
957 ipoib_ib_dev_cleanup(dev);
959 kfree(priv->rx_ring);
960 vfree(priv->tx_ring);
962 priv->rx_ring = NULL;
963 priv->tx_ring = NULL;
966 static const struct header_ops ipoib_header_ops = {
967 .create = ipoib_hard_header,
970 static const struct net_device_ops ipoib_netdev_ops = {
971 .ndo_open = ipoib_open,
972 .ndo_stop = ipoib_stop,
973 .ndo_change_mtu = ipoib_change_mtu,
974 .ndo_start_xmit = ipoib_start_xmit,
975 .ndo_tx_timeout = ipoib_timeout,
976 .ndo_set_multicast_list = ipoib_set_mcast_list,
977 .ndo_neigh_setup = ipoib_neigh_setup_dev,
980 static void ipoib_setup(struct net_device *dev)
982 struct ipoib_dev_priv *priv = netdev_priv(dev);
984 dev->netdev_ops = &ipoib_netdev_ops;
985 dev->header_ops = &ipoib_header_ops;
987 ipoib_set_ethtool_ops(dev);
989 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
991 dev->watchdog_timeo = HZ;
993 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
996 * We add in INFINIBAND_ALEN to allow for the destination
997 * address "pseudoheader" for skbs without neighbour struct.
999 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1000 dev->addr_len = INFINIBAND_ALEN;
1001 dev->type = ARPHRD_INFINIBAND;
1002 dev->tx_queue_len = ipoib_sendq_size * 2;
1003 dev->features = (NETIF_F_VLAN_CHALLENGED |
1005 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
1007 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1009 netif_carrier_off(dev);
1013 spin_lock_init(&priv->lock);
1015 mutex_init(&priv->vlan_mutex);
1017 INIT_LIST_HEAD(&priv->path_list);
1018 INIT_LIST_HEAD(&priv->child_intfs);
1019 INIT_LIST_HEAD(&priv->dead_ahs);
1020 INIT_LIST_HEAD(&priv->multicast_list);
1022 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1023 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1024 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1025 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1026 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1027 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1028 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1029 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1032 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1034 struct net_device *dev;
1036 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1041 return netdev_priv(dev);
1044 static ssize_t show_pkey(struct device *dev,
1045 struct device_attribute *attr, char *buf)
1047 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1049 return sprintf(buf, "0x%04x\n", priv->pkey);
1051 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1053 static ssize_t show_umcast(struct device *dev,
1054 struct device_attribute *attr, char *buf)
1056 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1058 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1061 static ssize_t set_umcast(struct device *dev,
1062 struct device_attribute *attr,
1063 const char *buf, size_t count)
1065 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1066 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1068 if (umcast_val > 0) {
1069 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1070 ipoib_warn(priv, "ignoring multicast groups joined directly "
1073 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1077 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1079 int ipoib_add_umcast_attr(struct net_device *dev)
1081 return device_create_file(&dev->dev, &dev_attr_umcast);
1084 static ssize_t create_child(struct device *dev,
1085 struct device_attribute *attr,
1086 const char *buf, size_t count)
1091 if (sscanf(buf, "%i", &pkey) != 1)
1094 if (pkey < 0 || pkey > 0xffff)
1098 * Set the full membership bit, so that we join the right
1099 * broadcast group, etc.
1103 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1105 return ret ? ret : count;
1107 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1109 static ssize_t delete_child(struct device *dev,
1110 struct device_attribute *attr,
1111 const char *buf, size_t count)
1116 if (sscanf(buf, "%i", &pkey) != 1)
1119 if (pkey < 0 || pkey > 0xffff)
1122 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1124 return ret ? ret : count;
1127 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1129 int ipoib_add_pkey_attr(struct net_device *dev)
1131 return device_create_file(&dev->dev, &dev_attr_pkey);
1134 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1136 struct ib_device_attr *device_attr;
1137 int result = -ENOMEM;
1139 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1141 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1142 hca->name, sizeof *device_attr);
1146 result = ib_query_device(hca, device_attr);
1148 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1153 priv->hca_caps = device_attr->device_cap_flags;
1157 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1158 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1159 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1162 priv->dev->features |= NETIF_F_GRO;
1164 if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1165 priv->dev->features |= NETIF_F_TSO;
1171 static struct net_device *ipoib_add_port(const char *format,
1172 struct ib_device *hca, u8 port)
1174 struct ipoib_dev_priv *priv;
1175 struct ib_port_attr attr;
1176 int result = -ENOMEM;
1178 priv = ipoib_intf_alloc(format);
1180 goto alloc_mem_failed;
1182 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1183 priv->dev->dev_id = port - 1;
1185 if (!ib_query_port(hca, port, &attr))
1186 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1188 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1190 goto device_init_failed;
1193 /* MTU will be reset when mcast join happens */
1194 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1195 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1197 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1199 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1200 hca->name, port, result);
1201 goto device_init_failed;
1204 if (ipoib_set_dev_features(priv, hca))
1205 goto device_init_failed;
1208 * Set the full membership bit, so that we join the right
1209 * broadcast group, etc.
1211 priv->pkey |= 0x8000;
1213 priv->dev->broadcast[8] = priv->pkey >> 8;
1214 priv->dev->broadcast[9] = priv->pkey & 0xff;
1216 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1218 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1219 hca->name, port, result);
1220 goto device_init_failed;
1222 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1224 result = ipoib_dev_init(priv->dev, hca, port);
1226 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1227 hca->name, port, result);
1228 goto device_init_failed;
1231 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1232 priv->ca, ipoib_event);
1233 result = ib_register_event_handler(&priv->event_handler);
1235 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1236 "port %d (ret = %d)\n",
1237 hca->name, port, result);
1241 result = register_netdev(priv->dev);
1243 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1244 hca->name, port, result);
1245 goto register_failed;
1248 ipoib_create_debug_files(priv->dev);
1250 if (ipoib_cm_add_mode_attr(priv->dev))
1252 if (ipoib_add_pkey_attr(priv->dev))
1254 if (ipoib_add_umcast_attr(priv->dev))
1256 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1258 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1264 ipoib_delete_debug_files(priv->dev);
1265 unregister_netdev(priv->dev);
1268 ib_unregister_event_handler(&priv->event_handler);
1269 flush_workqueue(ipoib_workqueue);
1272 ipoib_dev_cleanup(priv->dev);
1275 free_netdev(priv->dev);
1278 return ERR_PTR(result);
1281 static void ipoib_add_one(struct ib_device *device)
1283 struct list_head *dev_list;
1284 struct net_device *dev;
1285 struct ipoib_dev_priv *priv;
1288 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1291 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1295 INIT_LIST_HEAD(dev_list);
1297 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1302 e = device->phys_port_cnt;
1305 for (p = s; p <= e; ++p) {
1306 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1308 dev = ipoib_add_port("ib%d", device, p);
1310 priv = netdev_priv(dev);
1311 list_add_tail(&priv->list, dev_list);
1315 ib_set_client_data(device, &ipoib_client, dev_list);
1318 static void ipoib_remove_one(struct ib_device *device)
1320 struct ipoib_dev_priv *priv, *tmp;
1321 struct list_head *dev_list;
1323 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1326 dev_list = ib_get_client_data(device, &ipoib_client);
1328 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1329 ib_unregister_event_handler(&priv->event_handler);
1332 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1335 flush_workqueue(ipoib_workqueue);
1337 unregister_netdev(priv->dev);
1338 ipoib_dev_cleanup(priv->dev);
1339 free_netdev(priv->dev);
1345 static int __init ipoib_init_module(void)
1349 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1350 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1351 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1353 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1354 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1355 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1356 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1357 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1361 * When copying small received packets, we only copy from the
1362 * linear data part of the SKB, so we rely on this condition.
1364 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1366 ret = ipoib_register_debugfs();
1371 * We create our own workqueue mainly because we want to be
1372 * able to flush it when devices are being removed. We can't
1373 * use schedule_work()/flush_scheduled_work() because both
1374 * unregister_netdev() and linkwatch_event take the rtnl lock,
1375 * so flush_scheduled_work() can deadlock during device
1378 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1379 if (!ipoib_workqueue) {
1384 ib_sa_register_client(&ipoib_sa_client);
1386 ret = ib_register_client(&ipoib_client);
1393 ib_sa_unregister_client(&ipoib_sa_client);
1394 destroy_workqueue(ipoib_workqueue);
1397 ipoib_unregister_debugfs();
1402 static void __exit ipoib_cleanup_module(void)
1404 ib_unregister_client(&ipoib_client);
1405 ib_sa_unregister_client(&ipoib_sa_client);
1406 ipoib_unregister_debugfs();
1407 destroy_workqueue(ipoib_workqueue);
1410 module_init(ipoib_init_module);
1411 module_exit(ipoib_cleanup_module);