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, void *gid)
190 struct ipoib_dev_priv *priv = netdev_priv(dev);
191 struct rb_node *n = priv->path_tree.rb_node;
192 struct ipoib_path *path;
196 path = rb_entry(n, struct ipoib_path, rb_node);
198 ret = memcmp(gid, path->pathrec.dgid.raw,
199 sizeof (union ib_gid));
212 static int __path_add(struct net_device *dev, struct ipoib_path *path)
214 struct ipoib_dev_priv *priv = netdev_priv(dev);
215 struct rb_node **n = &priv->path_tree.rb_node;
216 struct rb_node *pn = NULL;
217 struct ipoib_path *tpath;
222 tpath = rb_entry(pn, struct ipoib_path, rb_node);
224 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
225 sizeof (union ib_gid));
234 rb_link_node(&path->rb_node, pn, n);
235 rb_insert_color(&path->rb_node, &priv->path_tree);
237 list_add_tail(&path->list, &priv->path_list);
242 static void path_free(struct net_device *dev, struct ipoib_path *path)
244 struct ipoib_dev_priv *priv = netdev_priv(dev);
245 struct ipoib_neigh *neigh, *tn;
249 while ((skb = __skb_dequeue(&path->queue)))
250 dev_kfree_skb_irq(skb);
252 spin_lock_irqsave(&priv->lock, flags);
254 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
256 * It's safe to call ipoib_put_ah() inside priv->lock
257 * here, because we know that path->ah will always
258 * hold one more reference, so ipoib_put_ah() will
259 * never do more than decrement the ref count.
262 ipoib_put_ah(neigh->ah);
264 ipoib_neigh_free(neigh);
267 spin_unlock_irqrestore(&priv->lock, flags);
270 ipoib_put_ah(path->ah);
275 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
277 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
279 struct ipoib_path_iter *iter;
281 iter = kmalloc(sizeof *iter, GFP_KERNEL);
286 memset(iter->path.pathrec.dgid.raw, 0, 16);
288 if (ipoib_path_iter_next(iter)) {
296 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
298 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
300 struct ipoib_path *path;
303 spin_lock_irq(&priv->lock);
305 n = rb_first(&priv->path_tree);
308 path = rb_entry(n, struct ipoib_path, rb_node);
310 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
311 sizeof (union ib_gid)) < 0) {
320 spin_unlock_irq(&priv->lock);
325 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
326 struct ipoib_path *path)
331 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
333 void ipoib_flush_paths(struct net_device *dev)
335 struct ipoib_dev_priv *priv = netdev_priv(dev);
336 struct ipoib_path *path, *tp;
337 LIST_HEAD(remove_list);
339 spin_lock_irq(&priv->lock);
341 list_splice(&priv->path_list, &remove_list);
342 INIT_LIST_HEAD(&priv->path_list);
344 list_for_each_entry(path, &remove_list, list)
345 rb_erase(&path->rb_node, &priv->path_tree);
347 list_for_each_entry_safe(path, tp, &remove_list, list) {
349 ib_sa_cancel_query(path->query_id, path->query);
350 spin_unlock_irq(&priv->lock);
351 wait_for_completion(&path->done);
352 path_free(dev, path);
353 spin_lock_irq(&priv->lock);
355 spin_unlock_irq(&priv->lock);
358 static void path_rec_completion(int status,
359 struct ib_sa_path_rec *pathrec,
362 struct ipoib_path *path = path_ptr;
363 struct net_device *dev = path->dev;
364 struct ipoib_dev_priv *priv = netdev_priv(dev);
365 struct ipoib_ah *ah = NULL;
366 struct ipoib_neigh *neigh;
367 struct sk_buff_head skqueue;
372 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
373 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
375 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
376 status, IPOIB_GID_ARG(path->pathrec.dgid));
378 skb_queue_head_init(&skqueue);
381 struct ib_ah_attr av = {
382 .dlid = be16_to_cpu(pathrec->dlid),
384 .port_num = priv->port,
385 .static_rate = 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, void *gid)
428 struct ipoib_dev_priv *priv = netdev_priv(dev);
429 struct ipoib_path *path;
431 path = kzalloc(sizeof *path, GFP_ATOMIC);
437 skb_queue_head_init(&path->queue);
439 INIT_LIST_HEAD(&path->neigh_list);
441 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
442 path->pathrec.sgid = priv->local_gid;
443 path->pathrec.pkey = cpu_to_be16(priv->pkey);
444 path->pathrec.numb_path = 1;
449 static int path_rec_start(struct net_device *dev,
450 struct ipoib_path *path)
452 struct ipoib_dev_priv *priv = netdev_priv(dev);
454 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
455 IPOIB_GID_ARG(path->pathrec.dgid));
457 init_completion(&path->done);
460 ib_sa_path_rec_get(priv->ca, priv->port,
462 IB_SA_PATH_REC_DGID |
463 IB_SA_PATH_REC_SGID |
464 IB_SA_PATH_REC_NUMB_PATH |
469 if (path->query_id < 0) {
470 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
472 return path->query_id;
478 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
480 struct ipoib_dev_priv *priv = netdev_priv(dev);
481 struct ipoib_path *path;
482 struct ipoib_neigh *neigh;
484 neigh = ipoib_neigh_alloc(skb->dst->neighbour);
486 ++priv->stats.tx_dropped;
487 dev_kfree_skb_any(skb);
491 skb_queue_head_init(&neigh->queue);
494 * We can only be called from ipoib_start_xmit, so we're
495 * inside tx_lock -- no need to save/restore flags.
497 spin_lock(&priv->lock);
499 path = __path_find(dev, skb->dst->neighbour->ha + 4);
501 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
505 __path_add(dev, path);
508 list_add_tail(&neigh->list, &path->neigh_list);
511 kref_get(&path->ah->ref);
512 neigh->ah = path->ah;
514 ipoib_send(dev, skb, path->ah,
515 be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
518 __skb_queue_tail(&neigh->queue, skb);
520 if (!path->query && path_rec_start(dev, path))
524 spin_unlock(&priv->lock);
528 list_del(&neigh->list);
531 ipoib_neigh_free(neigh);
532 ++priv->stats.tx_dropped;
533 dev_kfree_skb_any(skb);
535 spin_unlock(&priv->lock);
538 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
540 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
542 /* Look up path record for unicasts */
543 if (skb->dst->neighbour->ha[4] != 0xff) {
544 neigh_add_path(skb, dev);
548 /* Add in the P_Key for multicasts */
549 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
550 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
551 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
554 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
555 struct ipoib_pseudoheader *phdr)
557 struct ipoib_dev_priv *priv = netdev_priv(dev);
558 struct ipoib_path *path;
561 * We can only be called from ipoib_start_xmit, so we're
562 * inside tx_lock -- no need to save/restore flags.
564 spin_lock(&priv->lock);
566 path = __path_find(dev, phdr->hwaddr + 4);
568 path = path_rec_create(dev, phdr->hwaddr + 4);
570 /* put pseudoheader back on for next time */
571 skb_push(skb, sizeof *phdr);
572 __skb_queue_tail(&path->queue, skb);
574 if (path_rec_start(dev, path)) {
575 spin_unlock(&priv->lock);
576 path_free(dev, path);
579 __path_add(dev, path);
581 ++priv->stats.tx_dropped;
582 dev_kfree_skb_any(skb);
585 spin_unlock(&priv->lock);
590 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
591 be16_to_cpu(path->pathrec.dlid));
593 ipoib_send(dev, skb, path->ah,
594 be32_to_cpup((__be32 *) phdr->hwaddr));
595 } else if ((path->query || !path_rec_start(dev, path)) &&
596 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
597 /* put pseudoheader back on for next time */
598 skb_push(skb, sizeof *phdr);
599 __skb_queue_tail(&path->queue, skb);
601 ++priv->stats.tx_dropped;
602 dev_kfree_skb_any(skb);
605 spin_unlock(&priv->lock);
608 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
610 struct ipoib_dev_priv *priv = netdev_priv(dev);
611 struct ipoib_neigh *neigh;
614 if (!spin_trylock_irqsave(&priv->tx_lock, flags))
615 return NETDEV_TX_LOCKED;
618 * Check if our queue is stopped. Since we have the LLTX bit
619 * set, we can't rely on netif_stop_queue() preventing our
620 * xmit function from being called with a full queue.
622 if (unlikely(netif_queue_stopped(dev))) {
623 spin_unlock_irqrestore(&priv->tx_lock, flags);
624 return NETDEV_TX_BUSY;
627 if (skb->dst && skb->dst->neighbour) {
628 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
629 ipoib_path_lookup(skb, dev);
633 neigh = *to_ipoib_neigh(skb->dst->neighbour);
635 if (likely(neigh->ah)) {
636 ipoib_send(dev, skb, neigh->ah,
637 be32_to_cpup((__be32 *) skb->dst->neighbour->ha));
641 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
642 spin_lock(&priv->lock);
643 __skb_queue_tail(&neigh->queue, skb);
644 spin_unlock(&priv->lock);
646 ++priv->stats.tx_dropped;
647 dev_kfree_skb_any(skb);
650 struct ipoib_pseudoheader *phdr =
651 (struct ipoib_pseudoheader *) skb->data;
652 skb_pull(skb, sizeof *phdr);
654 if (phdr->hwaddr[4] == 0xff) {
655 /* Add in the P_Key for multicast*/
656 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
657 phdr->hwaddr[9] = priv->pkey & 0xff;
659 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
661 /* unicast GID -- should be ARP or RARP reply */
663 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
664 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
665 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
667 skb->dst ? "neigh" : "dst",
668 be16_to_cpup((__be16 *) skb->data),
669 be32_to_cpup((__be32 *) phdr->hwaddr),
670 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
671 dev_kfree_skb_any(skb);
672 ++priv->stats.tx_dropped;
676 unicast_arp_send(skb, dev, phdr);
681 spin_unlock_irqrestore(&priv->tx_lock, flags);
686 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
688 struct ipoib_dev_priv *priv = netdev_priv(dev);
693 static void ipoib_timeout(struct net_device *dev)
695 struct ipoib_dev_priv *priv = netdev_priv(dev);
697 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
698 jiffies_to_msecs(jiffies - dev->trans_start));
699 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
700 netif_queue_stopped(dev),
701 priv->tx_head, priv->tx_tail);
702 /* XXX reset QP, etc. */
705 static int ipoib_hard_header(struct sk_buff *skb,
706 struct net_device *dev,
708 void *daddr, void *saddr, unsigned len)
710 struct ipoib_header *header;
712 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
714 header->proto = htons(type);
715 header->reserved = 0;
718 * If we don't have a neighbour structure, stuff the
719 * destination address onto the front of the skb so we can
720 * figure out where to send the packet later.
722 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
723 struct ipoib_pseudoheader *phdr =
724 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
725 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
731 static void ipoib_set_mcast_list(struct net_device *dev)
733 struct ipoib_dev_priv *priv = netdev_priv(dev);
735 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
736 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
740 queue_work(ipoib_workqueue, &priv->restart_task);
743 static void ipoib_neigh_destructor(struct neighbour *n)
745 struct ipoib_neigh *neigh;
746 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
748 struct ipoib_ah *ah = NULL;
751 "neigh_destructor for %06x " IPOIB_GID_FMT "\n",
752 be32_to_cpup((__be32 *) n->ha),
753 IPOIB_GID_RAW_ARG(n->ha + 4));
755 spin_lock_irqsave(&priv->lock, flags);
757 neigh = *to_ipoib_neigh(n);
761 list_del(&neigh->list);
762 ipoib_neigh_free(neigh);
765 spin_unlock_irqrestore(&priv->lock, flags);
771 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
773 struct ipoib_neigh *neigh;
775 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
779 neigh->neighbour = neighbour;
780 *to_ipoib_neigh(neighbour) = neigh;
785 void ipoib_neigh_free(struct ipoib_neigh *neigh)
787 *to_ipoib_neigh(neigh->neighbour) = NULL;
791 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
793 parms->neigh_destructor = ipoib_neigh_destructor;
798 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
800 struct ipoib_dev_priv *priv = netdev_priv(dev);
802 /* Allocate RX/TX "rings" to hold queued skbs */
803 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
805 if (!priv->rx_ring) {
806 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
807 ca->name, ipoib_recvq_size);
811 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
813 if (!priv->tx_ring) {
814 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
815 ca->name, ipoib_sendq_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_sendq_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 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1136 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1137 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1139 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1140 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1141 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1143 ret = ipoib_register_debugfs();
1148 * We create our own workqueue mainly because we want to be
1149 * able to flush it when devices are being removed. We can't
1150 * use schedule_work()/flush_scheduled_work() because both
1151 * unregister_netdev() and linkwatch_event take the rtnl lock,
1152 * so flush_scheduled_work() can deadlock during device
1155 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1156 if (!ipoib_workqueue) {
1161 ret = ib_register_client(&ipoib_client);
1168 destroy_workqueue(ipoib_workqueue);
1171 ipoib_unregister_debugfs();
1176 static void __exit ipoib_cleanup_module(void)
1178 ib_unregister_client(&ipoib_client);
1179 ipoib_unregister_debugfs();
1180 destroy_workqueue(ipoib_workqueue);
1183 module_init(ipoib_init_module);
1184 module_exit(ipoib_cleanup_module);