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
34 * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
39 #include <linux/module.h>
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/vmalloc.h>
45 #include <linux/if_arp.h> /* For ARPHRD_xxx */
52 MODULE_AUTHOR("Roland Dreier");
53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
54 MODULE_LICENSE("Dual BSD/GPL");
56 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
57 int ipoib_debug_level;
59 module_param_named(debug_level, ipoib_debug_level, int, 0644);
60 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
63 struct ipoib_path_iter {
64 struct net_device *dev;
65 struct ipoib_path path;
68 static const u8 ipv4_bcast_addr[] = {
69 0x00, 0xff, 0xff, 0xff,
70 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
71 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
74 struct workqueue_struct *ipoib_workqueue;
76 static void ipoib_add_one(struct ib_device *device);
77 static void ipoib_remove_one(struct ib_device *device);
79 static struct ib_client ipoib_client = {
82 .remove = ipoib_remove_one
85 int ipoib_open(struct net_device *dev)
87 struct ipoib_dev_priv *priv = netdev_priv(dev);
89 ipoib_dbg(priv, "bringing up interface\n");
91 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
93 if (ipoib_pkey_dev_delay_open(dev))
96 if (ipoib_ib_dev_open(dev))
99 if (ipoib_ib_dev_up(dev)) {
100 ipoib_ib_dev_stop(dev);
104 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
105 struct ipoib_dev_priv *cpriv;
107 /* Bring up any child interfaces too */
108 mutex_lock(&priv->vlan_mutex);
109 list_for_each_entry(cpriv, &priv->child_intfs, list) {
112 flags = cpriv->dev->flags;
116 dev_change_flags(cpriv->dev, flags | IFF_UP);
118 mutex_unlock(&priv->vlan_mutex);
121 netif_start_queue(dev);
126 static int ipoib_stop(struct net_device *dev)
128 struct ipoib_dev_priv *priv = netdev_priv(dev);
130 ipoib_dbg(priv, "stopping interface\n");
132 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
134 netif_stop_queue(dev);
137 * Now flush workqueue to make sure a scheduled task doesn't
138 * bring our internal state back up.
140 flush_workqueue(ipoib_workqueue);
142 ipoib_ib_dev_down(dev, 1);
143 ipoib_ib_dev_stop(dev);
145 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
146 struct ipoib_dev_priv *cpriv;
148 /* Bring down any child interfaces too */
149 mutex_lock(&priv->vlan_mutex);
150 list_for_each_entry(cpriv, &priv->child_intfs, list) {
153 flags = cpriv->dev->flags;
154 if (!(flags & IFF_UP))
157 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
159 mutex_unlock(&priv->vlan_mutex);
165 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
167 struct ipoib_dev_priv *priv = netdev_priv(dev);
169 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
172 priv->admin_mtu = new_mtu;
174 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
179 static struct ipoib_path *__path_find(struct net_device *dev,
182 struct ipoib_dev_priv *priv = netdev_priv(dev);
183 struct rb_node *n = priv->path_tree.rb_node;
184 struct ipoib_path *path;
188 path = rb_entry(n, struct ipoib_path, rb_node);
190 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
191 sizeof (union ib_gid));
204 static int __path_add(struct net_device *dev, struct ipoib_path *path)
206 struct ipoib_dev_priv *priv = netdev_priv(dev);
207 struct rb_node **n = &priv->path_tree.rb_node;
208 struct rb_node *pn = NULL;
209 struct ipoib_path *tpath;
214 tpath = rb_entry(pn, struct ipoib_path, rb_node);
216 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
217 sizeof (union ib_gid));
226 rb_link_node(&path->rb_node, pn, n);
227 rb_insert_color(&path->rb_node, &priv->path_tree);
229 list_add_tail(&path->list, &priv->path_list);
234 static void path_free(struct net_device *dev, struct ipoib_path *path)
236 struct ipoib_dev_priv *priv = netdev_priv(dev);
237 struct ipoib_neigh *neigh, *tn;
241 while ((skb = __skb_dequeue(&path->queue)))
242 dev_kfree_skb_irq(skb);
244 spin_lock_irqsave(&priv->lock, flags);
246 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
248 * It's safe to call ipoib_put_ah() inside priv->lock
249 * here, because we know that path->ah will always
250 * hold one more reference, so ipoib_put_ah() will
251 * never do more than decrement the ref count.
254 ipoib_put_ah(neigh->ah);
255 *to_ipoib_neigh(neigh->neighbour) = NULL;
256 neigh->neighbour->ops->destructor = NULL;
260 spin_unlock_irqrestore(&priv->lock, flags);
263 ipoib_put_ah(path->ah);
268 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
270 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
272 struct ipoib_path_iter *iter;
274 iter = kmalloc(sizeof *iter, GFP_KERNEL);
279 memset(iter->path.pathrec.dgid.raw, 0, 16);
281 if (ipoib_path_iter_next(iter)) {
289 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
291 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
293 struct ipoib_path *path;
296 spin_lock_irq(&priv->lock);
298 n = rb_first(&priv->path_tree);
301 path = rb_entry(n, struct ipoib_path, rb_node);
303 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
304 sizeof (union ib_gid)) < 0) {
313 spin_unlock_irq(&priv->lock);
318 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
319 struct ipoib_path *path)
324 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
326 void ipoib_flush_paths(struct net_device *dev)
328 struct ipoib_dev_priv *priv = netdev_priv(dev);
329 struct ipoib_path *path, *tp;
330 LIST_HEAD(remove_list);
333 spin_lock_irqsave(&priv->lock, flags);
335 list_splice(&priv->path_list, &remove_list);
336 INIT_LIST_HEAD(&priv->path_list);
338 list_for_each_entry(path, &remove_list, list)
339 rb_erase(&path->rb_node, &priv->path_tree);
341 spin_unlock_irqrestore(&priv->lock, flags);
343 list_for_each_entry_safe(path, tp, &remove_list, list) {
345 ib_sa_cancel_query(path->query_id, path->query);
346 wait_for_completion(&path->done);
347 path_free(dev, path);
351 static void path_rec_completion(int status,
352 struct ib_sa_path_rec *pathrec,
355 struct ipoib_path *path = path_ptr;
356 struct net_device *dev = path->dev;
357 struct ipoib_dev_priv *priv = netdev_priv(dev);
358 struct ipoib_ah *ah = NULL;
359 struct ipoib_neigh *neigh;
360 struct sk_buff_head skqueue;
365 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
366 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
368 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
369 status, IPOIB_GID_ARG(path->pathrec.dgid));
371 skb_queue_head_init(&skqueue);
374 struct ib_ah_attr av = {
375 .dlid = be16_to_cpu(pathrec->dlid),
377 .port_num = priv->port
379 int path_rate = ib_sa_rate_enum_to_int(pathrec->rate);
381 if (path_rate > 0 && priv->local_rate > path_rate)
382 av.static_rate = (priv->local_rate - 1) / path_rate;
384 ipoib_dbg(priv, "static_rate %d for local port %dX, path %dX\n",
385 av.static_rate, priv->local_rate,
386 ib_sa_rate_enum_to_int(pathrec->rate));
388 ah = ipoib_create_ah(dev, priv->pd, &av);
391 spin_lock_irqsave(&priv->lock, flags);
396 path->pathrec = *pathrec;
398 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
399 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
401 while ((skb = __skb_dequeue(&path->queue)))
402 __skb_queue_tail(&skqueue, skb);
404 list_for_each_entry(neigh, &path->neigh_list, list) {
405 kref_get(&path->ah->ref);
406 neigh->ah = path->ah;
408 while ((skb = __skb_dequeue(&neigh->queue)))
409 __skb_queue_tail(&skqueue, skb);
414 complete(&path->done);
416 spin_unlock_irqrestore(&priv->lock, flags);
418 while ((skb = __skb_dequeue(&skqueue))) {
420 if (dev_queue_xmit(skb))
421 ipoib_warn(priv, "dev_queue_xmit failed "
422 "to requeue packet\n");
426 static struct ipoib_path *path_rec_create(struct net_device *dev,
429 struct ipoib_dev_priv *priv = netdev_priv(dev);
430 struct ipoib_path *path;
432 path = kzalloc(sizeof *path, GFP_ATOMIC);
438 skb_queue_head_init(&path->queue);
440 INIT_LIST_HEAD(&path->neigh_list);
442 memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
443 path->pathrec.sgid = priv->local_gid;
444 path->pathrec.pkey = cpu_to_be16(priv->pkey);
445 path->pathrec.numb_path = 1;
450 static int path_rec_start(struct net_device *dev,
451 struct ipoib_path *path)
453 struct ipoib_dev_priv *priv = netdev_priv(dev);
455 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
456 IPOIB_GID_ARG(path->pathrec.dgid));
458 init_completion(&path->done);
461 ib_sa_path_rec_get(priv->ca, priv->port,
463 IB_SA_PATH_REC_DGID |
464 IB_SA_PATH_REC_SGID |
465 IB_SA_PATH_REC_NUMB_PATH |
470 if (path->query_id < 0) {
471 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
473 return path->query_id;
479 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
481 struct ipoib_dev_priv *priv = netdev_priv(dev);
482 struct ipoib_path *path;
483 struct ipoib_neigh *neigh;
485 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
487 ++priv->stats.tx_dropped;
488 dev_kfree_skb_any(skb);
492 skb_queue_head_init(&neigh->queue);
493 neigh->neighbour = skb->dst->neighbour;
494 *to_ipoib_neigh(skb->dst->neighbour) = neigh;
497 * We can only be called from ipoib_start_xmit, so we're
498 * inside tx_lock -- no need to save/restore flags.
500 spin_lock(&priv->lock);
502 path = __path_find(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4));
504 path = path_rec_create(dev,
505 (union ib_gid *) (skb->dst->neighbour->ha + 4));
509 __path_add(dev, path);
512 list_add_tail(&neigh->list, &path->neigh_list);
515 kref_get(&path->ah->ref);
516 neigh->ah = path->ah;
518 ipoib_send(dev, skb, path->ah,
519 be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
522 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
523 __skb_queue_tail(&neigh->queue, skb);
525 ++priv->stats.tx_dropped;
526 dev_kfree_skb_any(skb);
529 if (!path->query && path_rec_start(dev, path))
533 spin_unlock(&priv->lock);
537 *to_ipoib_neigh(skb->dst->neighbour) = NULL;
538 list_del(&neigh->list);
539 neigh->neighbour->ops->destructor = NULL;
542 ++priv->stats.tx_dropped;
543 dev_kfree_skb_any(skb);
545 spin_unlock(&priv->lock);
548 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
550 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
552 /* Look up path record for unicasts */
553 if (skb->dst->neighbour->ha[4] != 0xff) {
554 neigh_add_path(skb, dev);
558 /* Add in the P_Key for multicasts */
559 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
560 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
561 ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
564 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
565 struct ipoib_pseudoheader *phdr)
567 struct ipoib_dev_priv *priv = netdev_priv(dev);
568 struct ipoib_path *path;
571 * We can only be called from ipoib_start_xmit, so we're
572 * inside tx_lock -- no need to save/restore flags.
574 spin_lock(&priv->lock);
576 path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
578 path = path_rec_create(dev,
579 (union ib_gid *) (phdr->hwaddr + 4));
581 /* put pseudoheader back on for next time */
582 skb_push(skb, sizeof *phdr);
583 __skb_queue_tail(&path->queue, skb);
585 if (path_rec_start(dev, path)) {
586 spin_unlock(&priv->lock);
587 path_free(dev, path);
590 __path_add(dev, path);
592 ++priv->stats.tx_dropped;
593 dev_kfree_skb_any(skb);
596 spin_unlock(&priv->lock);
601 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
602 be16_to_cpu(path->pathrec.dlid));
604 ipoib_send(dev, skb, path->ah,
605 be32_to_cpup((__be32 *) phdr->hwaddr));
606 } else if ((path->query || !path_rec_start(dev, path)) &&
607 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
608 /* put pseudoheader back on for next time */
609 skb_push(skb, sizeof *phdr);
610 __skb_queue_tail(&path->queue, skb);
612 ++priv->stats.tx_dropped;
613 dev_kfree_skb_any(skb);
616 spin_unlock(&priv->lock);
619 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
621 struct ipoib_dev_priv *priv = netdev_priv(dev);
622 struct ipoib_neigh *neigh;
625 if (!spin_trylock_irqsave(&priv->tx_lock, flags))
626 return NETDEV_TX_LOCKED;
629 * Check if our queue is stopped. Since we have the LLTX bit
630 * set, we can't rely on netif_stop_queue() preventing our
631 * xmit function from being called with a full queue.
633 if (unlikely(netif_queue_stopped(dev))) {
634 spin_unlock_irqrestore(&priv->tx_lock, flags);
635 return NETDEV_TX_BUSY;
638 if (skb->dst && skb->dst->neighbour) {
639 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
640 ipoib_path_lookup(skb, dev);
644 neigh = *to_ipoib_neigh(skb->dst->neighbour);
646 if (likely(neigh->ah)) {
647 ipoib_send(dev, skb, neigh->ah,
648 be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
652 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
653 spin_lock(&priv->lock);
654 __skb_queue_tail(&neigh->queue, skb);
655 spin_unlock(&priv->lock);
657 ++priv->stats.tx_dropped;
658 dev_kfree_skb_any(skb);
661 struct ipoib_pseudoheader *phdr =
662 (struct ipoib_pseudoheader *) skb->data;
663 skb_pull(skb, sizeof *phdr);
665 if (phdr->hwaddr[4] == 0xff) {
666 /* Add in the P_Key for multicast*/
667 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
668 phdr->hwaddr[9] = priv->pkey & 0xff;
670 ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
672 /* unicast GID -- should be ARP or RARP reply */
674 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
675 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
676 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
678 skb->dst ? "neigh" : "dst",
679 be16_to_cpup((__be16 *) skb->data),
680 be32_to_cpup((__be32 *) phdr->hwaddr),
681 IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
682 dev_kfree_skb_any(skb);
683 ++priv->stats.tx_dropped;
687 unicast_arp_send(skb, dev, phdr);
692 spin_unlock_irqrestore(&priv->tx_lock, flags);
697 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
699 struct ipoib_dev_priv *priv = netdev_priv(dev);
704 static void ipoib_timeout(struct net_device *dev)
706 struct ipoib_dev_priv *priv = netdev_priv(dev);
708 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
709 jiffies_to_msecs(jiffies - dev->trans_start));
710 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
711 netif_queue_stopped(dev),
712 priv->tx_head, priv->tx_tail);
713 /* XXX reset QP, etc. */
716 static int ipoib_hard_header(struct sk_buff *skb,
717 struct net_device *dev,
719 void *daddr, void *saddr, unsigned len)
721 struct ipoib_header *header;
723 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
725 header->proto = htons(type);
726 header->reserved = 0;
729 * If we don't have a neighbour structure, stuff the
730 * destination address onto the front of the skb so we can
731 * figure out where to send the packet later.
733 if (!skb->dst || !skb->dst->neighbour) {
734 struct ipoib_pseudoheader *phdr =
735 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
736 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
742 static void ipoib_set_mcast_list(struct net_device *dev)
744 struct ipoib_dev_priv *priv = netdev_priv(dev);
746 queue_work(ipoib_workqueue, &priv->restart_task);
749 static void ipoib_neigh_destructor(struct neighbour *n)
751 struct ipoib_neigh *neigh;
752 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
754 struct ipoib_ah *ah = NULL;
757 "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
758 be32_to_cpup((__be32 *) n->ha),
759 IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
761 spin_lock_irqsave(&priv->lock, flags);
763 neigh = *to_ipoib_neigh(n);
767 list_del(&neigh->list);
768 *to_ipoib_neigh(n) = NULL;
772 spin_unlock_irqrestore(&priv->lock, flags);
778 static int ipoib_neigh_setup(struct neighbour *neigh)
781 * Is this kosher? I can't find anybody in the kernel that
782 * sets neigh->destructor, so we should be able to set it here
785 neigh->ops->destructor = ipoib_neigh_destructor;
790 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
792 parms->neigh_setup = ipoib_neigh_setup;
797 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
799 struct ipoib_dev_priv *priv = netdev_priv(dev);
801 /* Allocate RX/TX "rings" to hold queued skbs */
803 priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
805 if (!priv->rx_ring) {
806 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
807 ca->name, IPOIB_RX_RING_SIZE);
811 priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
813 if (!priv->tx_ring) {
814 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
815 ca->name, IPOIB_TX_RING_SIZE);
816 goto out_rx_ring_cleanup;
819 /* priv->tx_head & tx_tail are already 0 */
821 if (ipoib_ib_dev_init(dev, ca, port))
822 goto out_tx_ring_cleanup;
827 kfree(priv->tx_ring);
830 kfree(priv->rx_ring);
836 void ipoib_dev_cleanup(struct net_device *dev)
838 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
840 ipoib_delete_debug_files(dev);
842 /* Delete any child interfaces first */
843 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
844 unregister_netdev(cpriv->dev);
845 ipoib_dev_cleanup(cpriv->dev);
846 free_netdev(cpriv->dev);
849 ipoib_ib_dev_cleanup(dev);
851 kfree(priv->rx_ring);
852 kfree(priv->tx_ring);
854 priv->rx_ring = NULL;
855 priv->tx_ring = NULL;
858 static void ipoib_setup(struct net_device *dev)
860 struct ipoib_dev_priv *priv = netdev_priv(dev);
862 dev->open = ipoib_open;
863 dev->stop = ipoib_stop;
864 dev->change_mtu = ipoib_change_mtu;
865 dev->hard_start_xmit = ipoib_start_xmit;
866 dev->get_stats = ipoib_get_stats;
867 dev->tx_timeout = ipoib_timeout;
868 dev->hard_header = ipoib_hard_header;
869 dev->set_multicast_list = ipoib_set_mcast_list;
870 dev->neigh_setup = ipoib_neigh_setup_dev;
872 dev->watchdog_timeo = HZ;
874 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
877 * We add in INFINIBAND_ALEN to allow for the destination
878 * address "pseudoheader" for skbs without neighbour struct.
880 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
881 dev->addr_len = INFINIBAND_ALEN;
882 dev->type = ARPHRD_INFINIBAND;
883 dev->tx_queue_len = IPOIB_TX_RING_SIZE * 2;
884 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
886 /* MTU will be reset when mcast join happens */
887 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
888 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
890 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
892 netif_carrier_off(dev);
894 SET_MODULE_OWNER(dev);
898 spin_lock_init(&priv->lock);
899 spin_lock_init(&priv->tx_lock);
901 mutex_init(&priv->mcast_mutex);
902 mutex_init(&priv->vlan_mutex);
904 INIT_LIST_HEAD(&priv->path_list);
905 INIT_LIST_HEAD(&priv->child_intfs);
906 INIT_LIST_HEAD(&priv->dead_ahs);
907 INIT_LIST_HEAD(&priv->multicast_list);
909 INIT_WORK(&priv->pkey_task, ipoib_pkey_poll, priv->dev);
910 INIT_WORK(&priv->mcast_task, ipoib_mcast_join_task, priv->dev);
911 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush, priv->dev);
912 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
913 INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah, priv->dev);
916 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
918 struct net_device *dev;
920 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
925 return netdev_priv(dev);
928 static ssize_t show_pkey(struct class_device *cdev, char *buf)
930 struct ipoib_dev_priv *priv =
931 netdev_priv(container_of(cdev, struct net_device, class_dev));
933 return sprintf(buf, "0x%04x\n", priv->pkey);
935 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
937 static ssize_t create_child(struct class_device *cdev,
938 const char *buf, size_t count)
943 if (sscanf(buf, "%i", &pkey) != 1)
946 if (pkey < 0 || pkey > 0xffff)
950 * Set the full membership bit, so that we join the right
951 * broadcast group, etc.
955 ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
958 return ret ? ret : count;
960 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
962 static ssize_t delete_child(struct class_device *cdev,
963 const char *buf, size_t count)
968 if (sscanf(buf, "%i", &pkey) != 1)
971 if (pkey < 0 || pkey > 0xffff)
974 ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
977 return ret ? ret : count;
980 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
982 int ipoib_add_pkey_attr(struct net_device *dev)
984 return class_device_create_file(&dev->class_dev,
985 &class_device_attr_pkey);
988 static struct net_device *ipoib_add_port(const char *format,
989 struct ib_device *hca, u8 port)
991 struct ipoib_dev_priv *priv;
992 int result = -ENOMEM;
994 priv = ipoib_intf_alloc(format);
996 goto alloc_mem_failed;
998 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1000 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1002 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1003 hca->name, port, result);
1004 goto alloc_mem_failed;
1008 * Set the full membership bit, so that we join the right
1009 * broadcast group, etc.
1011 priv->pkey |= 0x8000;
1013 priv->dev->broadcast[8] = priv->pkey >> 8;
1014 priv->dev->broadcast[9] = priv->pkey & 0xff;
1016 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1018 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1019 hca->name, port, result);
1020 goto alloc_mem_failed;
1022 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1025 result = ipoib_dev_init(priv->dev, hca, port);
1027 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1028 hca->name, port, result);
1029 goto device_init_failed;
1032 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1033 priv->ca, ipoib_event);
1034 result = ib_register_event_handler(&priv->event_handler);
1036 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1037 "port %d (ret = %d)\n",
1038 hca->name, port, result);
1042 result = register_netdev(priv->dev);
1044 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1045 hca->name, port, result);
1046 goto register_failed;
1049 ipoib_create_debug_files(priv->dev);
1051 if (ipoib_add_pkey_attr(priv->dev))
1053 if (class_device_create_file(&priv->dev->class_dev,
1054 &class_device_attr_create_child))
1056 if (class_device_create_file(&priv->dev->class_dev,
1057 &class_device_attr_delete_child))
1063 ipoib_delete_debug_files(priv->dev);
1064 unregister_netdev(priv->dev);
1067 ib_unregister_event_handler(&priv->event_handler);
1068 flush_scheduled_work();
1071 ipoib_dev_cleanup(priv->dev);
1074 free_netdev(priv->dev);
1077 return ERR_PTR(result);
1080 static void ipoib_add_one(struct ib_device *device)
1082 struct list_head *dev_list;
1083 struct net_device *dev;
1084 struct ipoib_dev_priv *priv;
1087 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1091 INIT_LIST_HEAD(dev_list);
1093 if (device->node_type == IB_NODE_SWITCH) {
1098 e = device->phys_port_cnt;
1101 for (p = s; p <= e; ++p) {
1102 dev = ipoib_add_port("ib%d", device, p);
1104 priv = netdev_priv(dev);
1105 list_add_tail(&priv->list, dev_list);
1109 ib_set_client_data(device, &ipoib_client, dev_list);
1112 static void ipoib_remove_one(struct ib_device *device)
1114 struct ipoib_dev_priv *priv, *tmp;
1115 struct list_head *dev_list;
1117 dev_list = ib_get_client_data(device, &ipoib_client);
1119 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1120 ib_unregister_event_handler(&priv->event_handler);
1121 flush_scheduled_work();
1123 unregister_netdev(priv->dev);
1124 ipoib_dev_cleanup(priv->dev);
1125 free_netdev(priv->dev);
1131 static int __init ipoib_init_module(void)
1135 ret = ipoib_register_debugfs();
1140 * We create our own workqueue mainly because we want to be
1141 * able to flush it when devices are being removed. We can't
1142 * use schedule_work()/flush_scheduled_work() because both
1143 * unregister_netdev() and linkwatch_event take the rtnl lock,
1144 * so flush_scheduled_work() can deadlock during device
1147 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1148 if (!ipoib_workqueue) {
1153 ret = ib_register_client(&ipoib_client);
1160 destroy_workqueue(ipoib_workqueue);
1163 ipoib_unregister_debugfs();
1168 static void __exit ipoib_cleanup_module(void)
1170 ib_unregister_client(&ipoib_client);
1171 ipoib_unregister_debugfs();
1172 destroy_workqueue(ipoib_workqueue);
1175 module_init(ipoib_init_module);
1176 module_exit(ipoib_cleanup_module);