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>
44 #include <linux/kernel.h>
46 #include <linux/if_arp.h> /* For ARPHRD_xxx */
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
55 MODULE_LICENSE("Dual BSD/GPL");
57 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
58 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
60 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
61 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
62 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
63 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
65 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
66 int ipoib_debug_level;
68 module_param_named(debug_level, ipoib_debug_level, int, 0644);
69 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
72 struct ipoib_path_iter {
73 struct net_device *dev;
74 struct ipoib_path path;
77 static const u8 ipv4_bcast_addr[] = {
78 0x00, 0xff, 0xff, 0xff,
79 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
80 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
83 struct workqueue_struct *ipoib_workqueue;
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)) {
109 ipoib_ib_dev_stop(dev);
113 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
114 struct ipoib_dev_priv *cpriv;
116 /* Bring up any child interfaces too */
117 mutex_lock(&priv->vlan_mutex);
118 list_for_each_entry(cpriv, &priv->child_intfs, list) {
121 flags = cpriv->dev->flags;
125 dev_change_flags(cpriv->dev, flags | IFF_UP);
127 mutex_unlock(&priv->vlan_mutex);
130 netif_start_queue(dev);
135 static int ipoib_stop(struct net_device *dev)
137 struct ipoib_dev_priv *priv = netdev_priv(dev);
139 ipoib_dbg(priv, "stopping interface\n");
141 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
143 netif_stop_queue(dev);
146 * Now flush workqueue to make sure a scheduled task doesn't
147 * bring our internal state back up.
149 flush_workqueue(ipoib_workqueue);
151 ipoib_ib_dev_down(dev, 1);
152 ipoib_ib_dev_stop(dev);
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 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
181 priv->admin_mtu = new_mtu;
183 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
188 static struct ipoib_path *__path_find(struct net_device *dev,
191 struct ipoib_dev_priv *priv = netdev_priv(dev);
192 struct rb_node *n = priv->path_tree.rb_node;
193 struct ipoib_path *path;
197 path = rb_entry(n, struct ipoib_path, rb_node);
199 ret = memcmp(gid->raw, path->pathrec.dgid.raw,
200 sizeof (union ib_gid));
213 static int __path_add(struct net_device *dev, struct ipoib_path *path)
215 struct ipoib_dev_priv *priv = netdev_priv(dev);
216 struct rb_node **n = &priv->path_tree.rb_node;
217 struct rb_node *pn = NULL;
218 struct ipoib_path *tpath;
223 tpath = rb_entry(pn, struct ipoib_path, rb_node);
225 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
226 sizeof (union ib_gid));
235 rb_link_node(&path->rb_node, pn, n);
236 rb_insert_color(&path->rb_node, &priv->path_tree);
238 list_add_tail(&path->list, &priv->path_list);
243 static void path_free(struct net_device *dev, struct ipoib_path *path)
245 struct ipoib_dev_priv *priv = netdev_priv(dev);
246 struct ipoib_neigh *neigh, *tn;
250 while ((skb = __skb_dequeue(&path->queue)))
251 dev_kfree_skb_irq(skb);
253 spin_lock_irqsave(&priv->lock, flags);
255 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
257 * It's safe to call ipoib_put_ah() inside priv->lock
258 * here, because we know that path->ah will always
259 * hold one more reference, so ipoib_put_ah() will
260 * never do more than decrement the ref count.
263 ipoib_put_ah(neigh->ah);
265 ipoib_neigh_free(neigh);
268 spin_unlock_irqrestore(&priv->lock, flags);
271 ipoib_put_ah(path->ah);
276 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
278 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
280 struct ipoib_path_iter *iter;
282 iter = kmalloc(sizeof *iter, GFP_KERNEL);
287 memset(iter->path.pathrec.dgid.raw, 0, 16);
289 if (ipoib_path_iter_next(iter)) {
297 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
299 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
301 struct ipoib_path *path;
304 spin_lock_irq(&priv->lock);
306 n = rb_first(&priv->path_tree);
309 path = rb_entry(n, struct ipoib_path, rb_node);
311 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
312 sizeof (union ib_gid)) < 0) {
321 spin_unlock_irq(&priv->lock);
326 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
327 struct ipoib_path *path)
332 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
334 void ipoib_flush_paths(struct net_device *dev)
336 struct ipoib_dev_priv *priv = netdev_priv(dev);
337 struct ipoib_path *path, *tp;
338 LIST_HEAD(remove_list);
341 spin_lock_irqsave(&priv->lock, flags);
343 list_splice(&priv->path_list, &remove_list);
344 INIT_LIST_HEAD(&priv->path_list);
346 list_for_each_entry(path, &remove_list, list)
347 rb_erase(&path->rb_node, &priv->path_tree);
349 list_for_each_entry_safe(path, tp, &remove_list, list) {
351 ib_sa_cancel_query(path->query_id, path->query);
352 spin_unlock_irqrestore(&priv->lock, flags);
353 wait_for_completion(&path->done);
354 path_free(dev, path);
355 spin_lock_irqsave(&priv->lock, flags);
357 spin_unlock_irqrestore(&priv->lock, flags);
360 static void path_rec_completion(int status,
361 struct ib_sa_path_rec *pathrec,
364 struct ipoib_path *path = path_ptr;
365 struct net_device *dev = path->dev;
366 struct ipoib_dev_priv *priv = netdev_priv(dev);
367 struct ipoib_ah *ah = NULL;
368 struct ipoib_neigh *neigh;
369 struct sk_buff_head skqueue;
374 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
375 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
377 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
378 status, IPOIB_GID_ARG(path->pathrec.dgid));
380 skb_queue_head_init(&skqueue);
383 struct ib_ah_attr av = {
384 .dlid = be16_to_cpu(pathrec->dlid),
386 .port_num = priv->port,
387 .static_rate = pathrec->rate
390 ah = ipoib_create_ah(dev, priv->pd, &av);
393 spin_lock_irqsave(&priv->lock, flags);
398 path->pathrec = *pathrec;
400 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
401 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
403 while ((skb = __skb_dequeue(&path->queue)))
404 __skb_queue_tail(&skqueue, skb);
406 list_for_each_entry(neigh, &path->neigh_list, list) {
407 kref_get(&path->ah->ref);
408 neigh->ah = path->ah;
410 while ((skb = __skb_dequeue(&neigh->queue)))
411 __skb_queue_tail(&skqueue, skb);
416 complete(&path->done);
418 spin_unlock_irqrestore(&priv->lock, flags);
420 while ((skb = __skb_dequeue(&skqueue))) {
422 if (dev_queue_xmit(skb))
423 ipoib_warn(priv, "dev_queue_xmit failed "
424 "to requeue packet\n");
428 static struct ipoib_path *path_rec_create(struct net_device *dev,
431 struct ipoib_dev_priv *priv = netdev_priv(dev);
432 struct ipoib_path *path;
434 path = kzalloc(sizeof *path, GFP_ATOMIC);
440 skb_queue_head_init(&path->queue);
442 INIT_LIST_HEAD(&path->neigh_list);
444 memcpy(path->pathrec.dgid.raw, gid->raw, sizeof (union ib_gid));
445 path->pathrec.sgid = priv->local_gid;
446 path->pathrec.pkey = cpu_to_be16(priv->pkey);
447 path->pathrec.numb_path = 1;
452 static int path_rec_start(struct net_device *dev,
453 struct ipoib_path *path)
455 struct ipoib_dev_priv *priv = netdev_priv(dev);
457 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
458 IPOIB_GID_ARG(path->pathrec.dgid));
460 init_completion(&path->done);
463 ib_sa_path_rec_get(priv->ca, priv->port,
465 IB_SA_PATH_REC_DGID |
466 IB_SA_PATH_REC_SGID |
467 IB_SA_PATH_REC_NUMB_PATH |
472 if (path->query_id < 0) {
473 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
475 return path->query_id;
481 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
483 struct ipoib_dev_priv *priv = netdev_priv(dev);
484 struct ipoib_path *path;
485 struct ipoib_neigh *neigh;
487 neigh = ipoib_neigh_alloc(skb->dst->neighbour);
489 ++priv->stats.tx_dropped;
490 dev_kfree_skb_any(skb);
494 skb_queue_head_init(&neigh->queue);
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 __skb_queue_tail(&neigh->queue, skb);
524 if (!path->query && path_rec_start(dev, path))
528 spin_unlock(&priv->lock);
532 list_del(&neigh->list);
535 ipoib_neigh_free(neigh);
536 ++priv->stats.tx_dropped;
537 dev_kfree_skb_any(skb);
539 spin_unlock(&priv->lock);
542 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
544 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
546 /* Look up path record for unicasts */
547 if (skb->dst->neighbour->ha[4] != 0xff) {
548 neigh_add_path(skb, dev);
552 /* Add in the P_Key for multicasts */
553 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
554 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
555 ipoib_mcast_send(dev, (union ib_gid *) (skb->dst->neighbour->ha + 4), skb);
558 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
559 struct ipoib_pseudoheader *phdr)
561 struct ipoib_dev_priv *priv = netdev_priv(dev);
562 struct ipoib_path *path;
565 * We can only be called from ipoib_start_xmit, so we're
566 * inside tx_lock -- no need to save/restore flags.
568 spin_lock(&priv->lock);
570 path = __path_find(dev, (union ib_gid *) (phdr->hwaddr + 4));
572 path = path_rec_create(dev,
573 (union ib_gid *) (phdr->hwaddr + 4));
575 /* put pseudoheader back on for next time */
576 skb_push(skb, sizeof *phdr);
577 __skb_queue_tail(&path->queue, skb);
579 if (path_rec_start(dev, path)) {
580 spin_unlock(&priv->lock);
581 path_free(dev, path);
584 __path_add(dev, path);
586 ++priv->stats.tx_dropped;
587 dev_kfree_skb_any(skb);
590 spin_unlock(&priv->lock);
595 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
596 be16_to_cpu(path->pathrec.dlid));
598 ipoib_send(dev, skb, path->ah,
599 be32_to_cpup((__be32 *) phdr->hwaddr));
600 } else if ((path->query || !path_rec_start(dev, path)) &&
601 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
602 /* put pseudoheader back on for next time */
603 skb_push(skb, sizeof *phdr);
604 __skb_queue_tail(&path->queue, skb);
606 ++priv->stats.tx_dropped;
607 dev_kfree_skb_any(skb);
610 spin_unlock(&priv->lock);
613 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
615 struct ipoib_dev_priv *priv = netdev_priv(dev);
616 struct ipoib_neigh *neigh;
619 if (!spin_trylock_irqsave(&priv->tx_lock, flags))
620 return NETDEV_TX_LOCKED;
623 * Check if our queue is stopped. Since we have the LLTX bit
624 * set, we can't rely on netif_stop_queue() preventing our
625 * xmit function from being called with a full queue.
627 if (unlikely(netif_queue_stopped(dev))) {
628 spin_unlock_irqrestore(&priv->tx_lock, flags);
629 return NETDEV_TX_BUSY;
632 if (skb->dst && skb->dst->neighbour) {
633 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
634 ipoib_path_lookup(skb, dev);
638 neigh = *to_ipoib_neigh(skb->dst->neighbour);
640 if (likely(neigh->ah)) {
641 ipoib_send(dev, skb, neigh->ah,
642 be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
646 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
647 spin_lock(&priv->lock);
648 __skb_queue_tail(&neigh->queue, skb);
649 spin_unlock(&priv->lock);
651 ++priv->stats.tx_dropped;
652 dev_kfree_skb_any(skb);
655 struct ipoib_pseudoheader *phdr =
656 (struct ipoib_pseudoheader *) skb->data;
657 skb_pull(skb, sizeof *phdr);
659 if (phdr->hwaddr[4] == 0xff) {
660 /* Add in the P_Key for multicast*/
661 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
662 phdr->hwaddr[9] = priv->pkey & 0xff;
664 ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
666 /* unicast GID -- should be ARP or RARP reply */
668 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
669 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
670 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
672 skb->dst ? "neigh" : "dst",
673 be16_to_cpup((__be16 *) skb->data),
674 be32_to_cpup((__be32 *) phdr->hwaddr),
675 IPOIB_GID_ARG(*(union ib_gid *) (phdr->hwaddr + 4)));
676 dev_kfree_skb_any(skb);
677 ++priv->stats.tx_dropped;
681 unicast_arp_send(skb, dev, phdr);
686 spin_unlock_irqrestore(&priv->tx_lock, flags);
691 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
693 struct ipoib_dev_priv *priv = netdev_priv(dev);
698 static void ipoib_timeout(struct net_device *dev)
700 struct ipoib_dev_priv *priv = netdev_priv(dev);
702 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
703 jiffies_to_msecs(jiffies - dev->trans_start));
704 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
705 netif_queue_stopped(dev),
706 priv->tx_head, priv->tx_tail);
707 /* XXX reset QP, etc. */
710 static int ipoib_hard_header(struct sk_buff *skb,
711 struct net_device *dev,
713 void *daddr, void *saddr, unsigned len)
715 struct ipoib_header *header;
717 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
719 header->proto = htons(type);
720 header->reserved = 0;
723 * If we don't have a neighbour structure, stuff the
724 * destination address onto the front of the skb so we can
725 * figure out where to send the packet later.
727 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
728 struct ipoib_pseudoheader *phdr =
729 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
730 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
736 static void ipoib_set_mcast_list(struct net_device *dev)
738 struct ipoib_dev_priv *priv = netdev_priv(dev);
740 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
741 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
745 queue_work(ipoib_workqueue, &priv->restart_task);
748 static void ipoib_neigh_destructor(struct neighbour *n)
750 struct ipoib_neigh *neigh;
751 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
753 struct ipoib_ah *ah = NULL;
756 "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
757 be32_to_cpup((__be32 *) n->ha),
758 IPOIB_GID_ARG(*((union ib_gid *) (n->ha + 4))));
760 spin_lock_irqsave(&priv->lock, flags);
762 neigh = *to_ipoib_neigh(n);
766 list_del(&neigh->list);
767 ipoib_neigh_free(neigh);
770 spin_unlock_irqrestore(&priv->lock, flags);
776 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
778 struct ipoib_neigh *neigh;
780 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
784 neigh->neighbour = neighbour;
785 *to_ipoib_neigh(neighbour) = neigh;
790 void ipoib_neigh_free(struct ipoib_neigh *neigh)
792 *to_ipoib_neigh(neigh->neighbour) = NULL;
796 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
798 parms->neigh_destructor = ipoib_neigh_destructor;
803 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
805 struct ipoib_dev_priv *priv = netdev_priv(dev);
807 /* Allocate RX/TX "rings" to hold queued skbs */
808 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
810 if (!priv->rx_ring) {
811 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
812 ca->name, ipoib_recvq_size);
816 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
818 if (!priv->tx_ring) {
819 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
820 ca->name, ipoib_sendq_size);
821 goto out_rx_ring_cleanup;
824 /* priv->tx_head & tx_tail are already 0 */
826 if (ipoib_ib_dev_init(dev, ca, port))
827 goto out_tx_ring_cleanup;
832 kfree(priv->tx_ring);
835 kfree(priv->rx_ring);
841 void ipoib_dev_cleanup(struct net_device *dev)
843 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
845 ipoib_delete_debug_files(dev);
847 /* Delete any child interfaces first */
848 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
849 unregister_netdev(cpriv->dev);
850 ipoib_dev_cleanup(cpriv->dev);
851 free_netdev(cpriv->dev);
854 ipoib_ib_dev_cleanup(dev);
856 kfree(priv->rx_ring);
857 kfree(priv->tx_ring);
859 priv->rx_ring = NULL;
860 priv->tx_ring = NULL;
863 static void ipoib_setup(struct net_device *dev)
865 struct ipoib_dev_priv *priv = netdev_priv(dev);
867 dev->open = ipoib_open;
868 dev->stop = ipoib_stop;
869 dev->change_mtu = ipoib_change_mtu;
870 dev->hard_start_xmit = ipoib_start_xmit;
871 dev->get_stats = ipoib_get_stats;
872 dev->tx_timeout = ipoib_timeout;
873 dev->hard_header = ipoib_hard_header;
874 dev->set_multicast_list = ipoib_set_mcast_list;
875 dev->neigh_setup = ipoib_neigh_setup_dev;
877 dev->watchdog_timeo = HZ;
879 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
882 * We add in INFINIBAND_ALEN to allow for the destination
883 * address "pseudoheader" for skbs without neighbour struct.
885 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
886 dev->addr_len = INFINIBAND_ALEN;
887 dev->type = ARPHRD_INFINIBAND;
888 dev->tx_queue_len = ipoib_sendq_size * 2;
889 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
891 /* MTU will be reset when mcast join happens */
892 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
893 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
895 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
897 netif_carrier_off(dev);
899 SET_MODULE_OWNER(dev);
903 spin_lock_init(&priv->lock);
904 spin_lock_init(&priv->tx_lock);
906 mutex_init(&priv->mcast_mutex);
907 mutex_init(&priv->vlan_mutex);
909 INIT_LIST_HEAD(&priv->path_list);
910 INIT_LIST_HEAD(&priv->child_intfs);
911 INIT_LIST_HEAD(&priv->dead_ahs);
912 INIT_LIST_HEAD(&priv->multicast_list);
914 INIT_WORK(&priv->pkey_task, ipoib_pkey_poll, priv->dev);
915 INIT_WORK(&priv->mcast_task, ipoib_mcast_join_task, priv->dev);
916 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush, priv->dev);
917 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task, priv->dev);
918 INIT_WORK(&priv->ah_reap_task, ipoib_reap_ah, priv->dev);
921 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
923 struct net_device *dev;
925 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
930 return netdev_priv(dev);
933 static ssize_t show_pkey(struct class_device *cdev, char *buf)
935 struct ipoib_dev_priv *priv =
936 netdev_priv(container_of(cdev, struct net_device, class_dev));
938 return sprintf(buf, "0x%04x\n", priv->pkey);
940 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
942 static ssize_t create_child(struct class_device *cdev,
943 const char *buf, size_t count)
948 if (sscanf(buf, "%i", &pkey) != 1)
951 if (pkey < 0 || pkey > 0xffff)
955 * Set the full membership bit, so that we join the right
956 * broadcast group, etc.
960 ret = ipoib_vlan_add(container_of(cdev, struct net_device, class_dev),
963 return ret ? ret : count;
965 static CLASS_DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
967 static ssize_t delete_child(struct class_device *cdev,
968 const char *buf, size_t count)
973 if (sscanf(buf, "%i", &pkey) != 1)
976 if (pkey < 0 || pkey > 0xffff)
979 ret = ipoib_vlan_delete(container_of(cdev, struct net_device, class_dev),
982 return ret ? ret : count;
985 static CLASS_DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
987 int ipoib_add_pkey_attr(struct net_device *dev)
989 return class_device_create_file(&dev->class_dev,
990 &class_device_attr_pkey);
993 static struct net_device *ipoib_add_port(const char *format,
994 struct ib_device *hca, u8 port)
996 struct ipoib_dev_priv *priv;
997 int result = -ENOMEM;
999 priv = ipoib_intf_alloc(format);
1001 goto alloc_mem_failed;
1003 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1005 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1007 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1008 hca->name, port, result);
1009 goto alloc_mem_failed;
1013 * Set the full membership bit, so that we join the right
1014 * broadcast group, etc.
1016 priv->pkey |= 0x8000;
1018 priv->dev->broadcast[8] = priv->pkey >> 8;
1019 priv->dev->broadcast[9] = priv->pkey & 0xff;
1021 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1023 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1024 hca->name, port, result);
1025 goto alloc_mem_failed;
1027 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1030 result = ipoib_dev_init(priv->dev, hca, port);
1032 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1033 hca->name, port, result);
1034 goto device_init_failed;
1037 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1038 priv->ca, ipoib_event);
1039 result = ib_register_event_handler(&priv->event_handler);
1041 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1042 "port %d (ret = %d)\n",
1043 hca->name, port, result);
1047 result = register_netdev(priv->dev);
1049 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1050 hca->name, port, result);
1051 goto register_failed;
1054 ipoib_create_debug_files(priv->dev);
1056 if (ipoib_add_pkey_attr(priv->dev))
1058 if (class_device_create_file(&priv->dev->class_dev,
1059 &class_device_attr_create_child))
1061 if (class_device_create_file(&priv->dev->class_dev,
1062 &class_device_attr_delete_child))
1068 ipoib_delete_debug_files(priv->dev);
1069 unregister_netdev(priv->dev);
1072 ib_unregister_event_handler(&priv->event_handler);
1073 flush_scheduled_work();
1076 ipoib_dev_cleanup(priv->dev);
1079 free_netdev(priv->dev);
1082 return ERR_PTR(result);
1085 static void ipoib_add_one(struct ib_device *device)
1087 struct list_head *dev_list;
1088 struct net_device *dev;
1089 struct ipoib_dev_priv *priv;
1092 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1096 INIT_LIST_HEAD(dev_list);
1098 if (device->node_type == IB_NODE_SWITCH) {
1103 e = device->phys_port_cnt;
1106 for (p = s; p <= e; ++p) {
1107 dev = ipoib_add_port("ib%d", device, p);
1109 priv = netdev_priv(dev);
1110 list_add_tail(&priv->list, dev_list);
1114 ib_set_client_data(device, &ipoib_client, dev_list);
1117 static void ipoib_remove_one(struct ib_device *device)
1119 struct ipoib_dev_priv *priv, *tmp;
1120 struct list_head *dev_list;
1122 dev_list = ib_get_client_data(device, &ipoib_client);
1124 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1125 ib_unregister_event_handler(&priv->event_handler);
1126 flush_scheduled_work();
1128 unregister_netdev(priv->dev);
1129 ipoib_dev_cleanup(priv->dev);
1130 free_netdev(priv->dev);
1136 static int __init ipoib_init_module(void)
1140 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1141 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1142 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1144 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1145 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1146 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1148 ret = ipoib_register_debugfs();
1153 * We create our own workqueue mainly because we want to be
1154 * able to flush it when devices are being removed. We can't
1155 * use schedule_work()/flush_scheduled_work() because both
1156 * unregister_netdev() and linkwatch_event take the rtnl lock,
1157 * so flush_scheduled_work() can deadlock during device
1160 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1161 if (!ipoib_workqueue) {
1166 ret = ib_register_client(&ipoib_client);
1173 destroy_workqueue(ipoib_workqueue);
1176 ipoib_unregister_debugfs();
1181 static void __exit ipoib_cleanup_module(void)
1183 ib_unregister_client(&ipoib_client);
1184 ipoib_unregister_debugfs();
1185 destroy_workqueue(ipoib_workqueue);
1188 module_init(ipoib_init_module);
1189 module_exit(ipoib_cleanup_module);