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/kernel.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 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
65 int ipoib_debug_level;
67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
71 struct ipoib_path_iter {
72 struct net_device *dev;
73 struct ipoib_path path;
76 static const u8 ipv4_bcast_addr[] = {
77 0x00, 0xff, 0xff, 0xff,
78 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
79 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
82 struct workqueue_struct *ipoib_workqueue;
84 struct ib_sa_client ipoib_sa_client;
86 static void ipoib_add_one(struct ib_device *device);
87 static void ipoib_remove_one(struct ib_device *device);
89 static struct ib_client ipoib_client = {
92 .remove = ipoib_remove_one
95 int ipoib_open(struct net_device *dev)
97 struct ipoib_dev_priv *priv = netdev_priv(dev);
99 ipoib_dbg(priv, "bringing up interface\n");
101 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
103 if (ipoib_pkey_dev_delay_open(dev))
106 if (ipoib_ib_dev_open(dev))
109 if (ipoib_ib_dev_up(dev)) {
110 ipoib_ib_dev_stop(dev, 1);
114 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
115 struct ipoib_dev_priv *cpriv;
117 /* Bring up any child interfaces too */
118 mutex_lock(&priv->vlan_mutex);
119 list_for_each_entry(cpriv, &priv->child_intfs, list) {
122 flags = cpriv->dev->flags;
126 dev_change_flags(cpriv->dev, flags | IFF_UP);
128 mutex_unlock(&priv->vlan_mutex);
131 netif_start_queue(dev);
136 static int ipoib_stop(struct net_device *dev)
138 struct ipoib_dev_priv *priv = netdev_priv(dev);
140 ipoib_dbg(priv, "stopping interface\n");
142 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
144 netif_stop_queue(dev);
146 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
149 * Now flush workqueue to make sure a scheduled task doesn't
150 * bring our internal state back up.
152 flush_workqueue(ipoib_workqueue);
154 ipoib_ib_dev_down(dev, 1);
155 ipoib_ib_dev_stop(dev, 1);
157 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
158 struct ipoib_dev_priv *cpriv;
160 /* Bring down any child interfaces too */
161 mutex_lock(&priv->vlan_mutex);
162 list_for_each_entry(cpriv, &priv->child_intfs, list) {
165 flags = cpriv->dev->flags;
166 if (!(flags & IFF_UP))
169 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
171 mutex_unlock(&priv->vlan_mutex);
177 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
179 struct ipoib_dev_priv *priv = netdev_priv(dev);
181 /* dev->mtu > 2K ==> connected mode */
182 if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
183 if (new_mtu > priv->mcast_mtu)
184 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
190 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
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_flush_paths(struct net_device *dev)
348 struct ipoib_dev_priv *priv = netdev_priv(dev);
349 struct ipoib_path *path, *tp;
350 LIST_HEAD(remove_list);
352 spin_lock_irq(&priv->tx_lock);
353 spin_lock(&priv->lock);
355 list_splice(&priv->path_list, &remove_list);
356 INIT_LIST_HEAD(&priv->path_list);
358 list_for_each_entry(path, &remove_list, list)
359 rb_erase(&path->rb_node, &priv->path_tree);
361 list_for_each_entry_safe(path, tp, &remove_list, list) {
363 ib_sa_cancel_query(path->query_id, path->query);
364 spin_unlock(&priv->lock);
365 spin_unlock_irq(&priv->tx_lock);
366 wait_for_completion(&path->done);
367 path_free(dev, path);
368 spin_lock_irq(&priv->tx_lock);
369 spin_lock(&priv->lock);
371 spin_unlock(&priv->lock);
372 spin_unlock_irq(&priv->tx_lock);
375 static void path_rec_completion(int status,
376 struct ib_sa_path_rec *pathrec,
379 struct ipoib_path *path = path_ptr;
380 struct net_device *dev = path->dev;
381 struct ipoib_dev_priv *priv = netdev_priv(dev);
382 struct ipoib_ah *ah = NULL;
383 struct ipoib_neigh *neigh, *tn;
384 struct sk_buff_head skqueue;
389 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
390 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
392 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
393 status, IPOIB_GID_ARG(path->pathrec.dgid));
395 skb_queue_head_init(&skqueue);
398 struct ib_ah_attr av;
400 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
401 ah = ipoib_create_ah(dev, priv->pd, &av);
404 spin_lock_irqsave(&priv->lock, flags);
409 path->pathrec = *pathrec;
411 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
412 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
414 while ((skb = __skb_dequeue(&path->queue)))
415 __skb_queue_tail(&skqueue, skb);
417 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
418 kref_get(&path->ah->ref);
419 neigh->ah = path->ah;
420 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
421 sizeof(union ib_gid));
423 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
424 if (!ipoib_cm_get(neigh))
425 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
428 if (!ipoib_cm_get(neigh)) {
429 list_del(&neigh->list);
431 ipoib_put_ah(neigh->ah);
432 ipoib_neigh_free(dev, neigh);
437 while ((skb = __skb_dequeue(&neigh->queue)))
438 __skb_queue_tail(&skqueue, skb);
443 complete(&path->done);
445 spin_unlock_irqrestore(&priv->lock, flags);
447 while ((skb = __skb_dequeue(&skqueue))) {
449 if (dev_queue_xmit(skb))
450 ipoib_warn(priv, "dev_queue_xmit failed "
451 "to requeue packet\n");
455 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
457 struct ipoib_dev_priv *priv = netdev_priv(dev);
458 struct ipoib_path *path;
460 path = kzalloc(sizeof *path, GFP_ATOMIC);
466 skb_queue_head_init(&path->queue);
468 INIT_LIST_HEAD(&path->neigh_list);
470 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
471 path->pathrec.sgid = priv->local_gid;
472 path->pathrec.pkey = cpu_to_be16(priv->pkey);
473 path->pathrec.numb_path = 1;
478 static int path_rec_start(struct net_device *dev,
479 struct ipoib_path *path)
481 struct ipoib_dev_priv *priv = netdev_priv(dev);
483 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
484 IPOIB_GID_ARG(path->pathrec.dgid));
486 init_completion(&path->done);
489 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
491 IB_SA_PATH_REC_DGID |
492 IB_SA_PATH_REC_SGID |
493 IB_SA_PATH_REC_NUMB_PATH |
498 if (path->query_id < 0) {
499 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
501 return path->query_id;
507 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
509 struct ipoib_dev_priv *priv = netdev_priv(dev);
510 struct ipoib_path *path;
511 struct ipoib_neigh *neigh;
513 neigh = ipoib_neigh_alloc(skb->dst->neighbour);
515 ++priv->stats.tx_dropped;
516 dev_kfree_skb_any(skb);
521 * We can only be called from ipoib_start_xmit, so we're
522 * inside tx_lock -- no need to save/restore flags.
524 spin_lock(&priv->lock);
526 path = __path_find(dev, skb->dst->neighbour->ha + 4);
528 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
532 __path_add(dev, path);
535 list_add_tail(&neigh->list, &path->neigh_list);
538 kref_get(&path->ah->ref);
539 neigh->ah = path->ah;
540 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
541 sizeof(union ib_gid));
543 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
544 if (!ipoib_cm_get(neigh))
545 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
546 if (!ipoib_cm_get(neigh)) {
547 list_del(&neigh->list);
549 ipoib_put_ah(neigh->ah);
550 ipoib_neigh_free(dev, neigh);
553 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
554 __skb_queue_tail(&neigh->queue, skb);
556 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
557 skb_queue_len(&neigh->queue));
561 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
565 if (!path->query && path_rec_start(dev, path))
568 __skb_queue_tail(&neigh->queue, skb);
571 spin_unlock(&priv->lock);
575 list_del(&neigh->list);
578 ipoib_neigh_free(dev, neigh);
580 ++priv->stats.tx_dropped;
581 dev_kfree_skb_any(skb);
583 spin_unlock(&priv->lock);
586 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
588 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
590 /* Look up path record for unicasts */
591 if (skb->dst->neighbour->ha[4] != 0xff) {
592 neigh_add_path(skb, dev);
596 /* Add in the P_Key for multicasts */
597 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
598 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
599 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
602 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
603 struct ipoib_pseudoheader *phdr)
605 struct ipoib_dev_priv *priv = netdev_priv(dev);
606 struct ipoib_path *path;
609 * We can only be called from ipoib_start_xmit, so we're
610 * inside tx_lock -- no need to save/restore flags.
612 spin_lock(&priv->lock);
614 path = __path_find(dev, phdr->hwaddr + 4);
616 path = path_rec_create(dev, phdr->hwaddr + 4);
618 /* put pseudoheader back on for next time */
619 skb_push(skb, sizeof *phdr);
620 __skb_queue_tail(&path->queue, skb);
622 if (path_rec_start(dev, path)) {
623 spin_unlock(&priv->lock);
624 path_free(dev, path);
627 __path_add(dev, path);
629 ++priv->stats.tx_dropped;
630 dev_kfree_skb_any(skb);
633 spin_unlock(&priv->lock);
638 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
639 be16_to_cpu(path->pathrec.dlid));
641 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
642 } else if ((path->query || !path_rec_start(dev, path)) &&
643 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
644 /* put pseudoheader back on for next time */
645 skb_push(skb, sizeof *phdr);
646 __skb_queue_tail(&path->queue, skb);
648 ++priv->stats.tx_dropped;
649 dev_kfree_skb_any(skb);
652 spin_unlock(&priv->lock);
655 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
657 struct ipoib_dev_priv *priv = netdev_priv(dev);
658 struct ipoib_neigh *neigh;
661 if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
662 return NETDEV_TX_LOCKED;
665 * Check if our queue is stopped. Since we have the LLTX bit
666 * set, we can't rely on netif_stop_queue() preventing our
667 * xmit function from being called with a full queue.
669 if (unlikely(netif_queue_stopped(dev))) {
670 spin_unlock_irqrestore(&priv->tx_lock, flags);
671 return NETDEV_TX_BUSY;
674 if (likely(skb->dst && skb->dst->neighbour)) {
675 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
676 ipoib_path_lookup(skb, dev);
680 neigh = *to_ipoib_neigh(skb->dst->neighbour);
682 if (ipoib_cm_get(neigh)) {
683 if (ipoib_cm_up(neigh)) {
684 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
687 } else if (neigh->ah) {
688 if (unlikely(memcmp(&neigh->dgid.raw,
689 skb->dst->neighbour->ha + 4,
690 sizeof(union ib_gid)))) {
691 spin_lock(&priv->lock);
693 * It's safe to call ipoib_put_ah() inside
694 * priv->lock here, because we know that
695 * path->ah will always hold one more reference,
696 * so ipoib_put_ah() will never do more than
697 * decrement the ref count.
699 ipoib_put_ah(neigh->ah);
700 list_del(&neigh->list);
701 ipoib_neigh_free(dev, neigh);
702 spin_unlock(&priv->lock);
703 ipoib_path_lookup(skb, dev);
707 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
711 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
712 spin_lock(&priv->lock);
713 __skb_queue_tail(&neigh->queue, skb);
714 spin_unlock(&priv->lock);
716 ++priv->stats.tx_dropped;
717 dev_kfree_skb_any(skb);
720 struct ipoib_pseudoheader *phdr =
721 (struct ipoib_pseudoheader *) skb->data;
722 skb_pull(skb, sizeof *phdr);
724 if (phdr->hwaddr[4] == 0xff) {
725 /* Add in the P_Key for multicast*/
726 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
727 phdr->hwaddr[9] = priv->pkey & 0xff;
729 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
731 /* unicast GID -- should be ARP or RARP reply */
733 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
734 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
735 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
737 skb->dst ? "neigh" : "dst",
738 be16_to_cpup((__be16 *) skb->data),
739 IPOIB_QPN(phdr->hwaddr),
740 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
741 dev_kfree_skb_any(skb);
742 ++priv->stats.tx_dropped;
746 unicast_arp_send(skb, dev, phdr);
751 spin_unlock_irqrestore(&priv->tx_lock, flags);
756 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
758 struct ipoib_dev_priv *priv = netdev_priv(dev);
763 static void ipoib_timeout(struct net_device *dev)
765 struct ipoib_dev_priv *priv = netdev_priv(dev);
767 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
768 jiffies_to_msecs(jiffies - dev->trans_start));
769 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
770 netif_queue_stopped(dev),
771 priv->tx_head, priv->tx_tail);
772 /* XXX reset QP, etc. */
775 static int ipoib_hard_header(struct sk_buff *skb,
776 struct net_device *dev,
778 void *daddr, void *saddr, unsigned len)
780 struct ipoib_header *header;
782 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
784 header->proto = htons(type);
785 header->reserved = 0;
788 * If we don't have a neighbour structure, stuff the
789 * destination address onto the front of the skb so we can
790 * figure out where to send the packet later.
792 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
793 struct ipoib_pseudoheader *phdr =
794 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
795 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
801 static void ipoib_set_mcast_list(struct net_device *dev)
803 struct ipoib_dev_priv *priv = netdev_priv(dev);
805 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
806 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
810 queue_work(ipoib_workqueue, &priv->restart_task);
813 static void ipoib_neigh_cleanup(struct neighbour *n)
815 struct ipoib_neigh *neigh;
816 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
818 struct ipoib_ah *ah = NULL;
821 "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
823 IPOIB_GID_RAW_ARG(n->ha + 4));
825 spin_lock_irqsave(&priv->lock, flags);
827 neigh = *to_ipoib_neigh(n);
831 list_del(&neigh->list);
832 ipoib_neigh_free(n->dev, neigh);
835 spin_unlock_irqrestore(&priv->lock, flags);
841 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
843 struct ipoib_neigh *neigh;
845 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
849 neigh->neighbour = neighbour;
850 *to_ipoib_neigh(neighbour) = neigh;
851 skb_queue_head_init(&neigh->queue);
852 ipoib_cm_set(neigh, NULL);
857 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
859 struct ipoib_dev_priv *priv = netdev_priv(dev);
861 *to_ipoib_neigh(neigh->neighbour) = NULL;
862 while ((skb = __skb_dequeue(&neigh->queue))) {
863 ++priv->stats.tx_dropped;
864 dev_kfree_skb_any(skb);
866 if (ipoib_cm_get(neigh))
867 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
871 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
873 parms->neigh_cleanup = ipoib_neigh_cleanup;
878 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
880 struct ipoib_dev_priv *priv = netdev_priv(dev);
882 /* Allocate RX/TX "rings" to hold queued skbs */
883 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
885 if (!priv->rx_ring) {
886 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
887 ca->name, ipoib_recvq_size);
891 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
893 if (!priv->tx_ring) {
894 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
895 ca->name, ipoib_sendq_size);
896 goto out_rx_ring_cleanup;
899 /* priv->tx_head & tx_tail are already 0 */
901 if (ipoib_ib_dev_init(dev, ca, port))
902 goto out_tx_ring_cleanup;
907 kfree(priv->tx_ring);
910 kfree(priv->rx_ring);
916 void ipoib_dev_cleanup(struct net_device *dev)
918 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
920 ipoib_delete_debug_files(dev);
922 /* Delete any child interfaces first */
923 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
924 unregister_netdev(cpriv->dev);
925 ipoib_dev_cleanup(cpriv->dev);
926 free_netdev(cpriv->dev);
929 ipoib_ib_dev_cleanup(dev);
931 kfree(priv->rx_ring);
932 kfree(priv->tx_ring);
934 priv->rx_ring = NULL;
935 priv->tx_ring = NULL;
938 static void ipoib_setup(struct net_device *dev)
940 struct ipoib_dev_priv *priv = netdev_priv(dev);
942 dev->open = ipoib_open;
943 dev->stop = ipoib_stop;
944 dev->change_mtu = ipoib_change_mtu;
945 dev->hard_start_xmit = ipoib_start_xmit;
946 dev->get_stats = ipoib_get_stats;
947 dev->tx_timeout = ipoib_timeout;
948 dev->hard_header = ipoib_hard_header;
949 dev->set_multicast_list = ipoib_set_mcast_list;
950 dev->neigh_setup = ipoib_neigh_setup_dev;
951 dev->poll = ipoib_poll;
954 dev->watchdog_timeo = HZ;
956 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
959 * We add in INFINIBAND_ALEN to allow for the destination
960 * address "pseudoheader" for skbs without neighbour struct.
962 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
963 dev->addr_len = INFINIBAND_ALEN;
964 dev->type = ARPHRD_INFINIBAND;
965 dev->tx_queue_len = ipoib_sendq_size * 2;
966 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
968 /* MTU will be reset when mcast join happens */
969 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
970 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
972 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
974 netif_carrier_off(dev);
976 SET_MODULE_OWNER(dev);
980 spin_lock_init(&priv->lock);
981 spin_lock_init(&priv->tx_lock);
983 mutex_init(&priv->mcast_mutex);
984 mutex_init(&priv->vlan_mutex);
986 INIT_LIST_HEAD(&priv->path_list);
987 INIT_LIST_HEAD(&priv->child_intfs);
988 INIT_LIST_HEAD(&priv->dead_ahs);
989 INIT_LIST_HEAD(&priv->multicast_list);
991 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
992 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
993 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
994 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
995 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
996 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
999 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1001 struct net_device *dev;
1003 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1008 return netdev_priv(dev);
1011 static ssize_t show_pkey(struct device *dev,
1012 struct device_attribute *attr, char *buf)
1014 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1016 return sprintf(buf, "0x%04x\n", priv->pkey);
1018 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1020 static ssize_t create_child(struct device *dev,
1021 struct device_attribute *attr,
1022 const char *buf, size_t count)
1027 if (sscanf(buf, "%i", &pkey) != 1)
1030 if (pkey < 0 || pkey > 0xffff)
1034 * Set the full membership bit, so that we join the right
1035 * broadcast group, etc.
1039 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1041 return ret ? ret : count;
1043 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1045 static ssize_t delete_child(struct device *dev,
1046 struct device_attribute *attr,
1047 const char *buf, size_t count)
1052 if (sscanf(buf, "%i", &pkey) != 1)
1055 if (pkey < 0 || pkey > 0xffff)
1058 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1060 return ret ? ret : count;
1063 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1065 int ipoib_add_pkey_attr(struct net_device *dev)
1067 return device_create_file(&dev->dev, &dev_attr_pkey);
1070 static struct net_device *ipoib_add_port(const char *format,
1071 struct ib_device *hca, u8 port)
1073 struct ipoib_dev_priv *priv;
1074 int result = -ENOMEM;
1076 priv = ipoib_intf_alloc(format);
1078 goto alloc_mem_failed;
1080 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1082 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1084 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1085 hca->name, port, result);
1086 goto alloc_mem_failed;
1090 * Set the full membership bit, so that we join the right
1091 * broadcast group, etc.
1093 priv->pkey |= 0x8000;
1095 priv->dev->broadcast[8] = priv->pkey >> 8;
1096 priv->dev->broadcast[9] = priv->pkey & 0xff;
1098 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1100 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1101 hca->name, port, result);
1102 goto alloc_mem_failed;
1104 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1107 result = ipoib_dev_init(priv->dev, hca, port);
1109 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1110 hca->name, port, result);
1111 goto device_init_failed;
1114 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1115 priv->ca, ipoib_event);
1116 result = ib_register_event_handler(&priv->event_handler);
1118 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1119 "port %d (ret = %d)\n",
1120 hca->name, port, result);
1124 result = register_netdev(priv->dev);
1126 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1127 hca->name, port, result);
1128 goto register_failed;
1131 ipoib_create_debug_files(priv->dev);
1133 if (ipoib_cm_add_mode_attr(priv->dev))
1135 if (ipoib_add_pkey_attr(priv->dev))
1137 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1139 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1145 ipoib_delete_debug_files(priv->dev);
1146 unregister_netdev(priv->dev);
1149 ib_unregister_event_handler(&priv->event_handler);
1150 flush_scheduled_work();
1153 ipoib_dev_cleanup(priv->dev);
1156 free_netdev(priv->dev);
1159 return ERR_PTR(result);
1162 static void ipoib_add_one(struct ib_device *device)
1164 struct list_head *dev_list;
1165 struct net_device *dev;
1166 struct ipoib_dev_priv *priv;
1169 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1172 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1176 INIT_LIST_HEAD(dev_list);
1178 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1183 e = device->phys_port_cnt;
1186 for (p = s; p <= e; ++p) {
1187 dev = ipoib_add_port("ib%d", device, p);
1189 priv = netdev_priv(dev);
1190 list_add_tail(&priv->list, dev_list);
1194 ib_set_client_data(device, &ipoib_client, dev_list);
1197 static void ipoib_remove_one(struct ib_device *device)
1199 struct ipoib_dev_priv *priv, *tmp;
1200 struct list_head *dev_list;
1202 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1205 dev_list = ib_get_client_data(device, &ipoib_client);
1207 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1208 ib_unregister_event_handler(&priv->event_handler);
1209 flush_scheduled_work();
1211 unregister_netdev(priv->dev);
1212 ipoib_dev_cleanup(priv->dev);
1213 free_netdev(priv->dev);
1219 static int __init ipoib_init_module(void)
1223 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1224 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1225 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1227 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1228 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1229 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1231 ret = ipoib_register_debugfs();
1236 * We create our own workqueue mainly because we want to be
1237 * able to flush it when devices are being removed. We can't
1238 * use schedule_work()/flush_scheduled_work() because both
1239 * unregister_netdev() and linkwatch_event take the rtnl lock,
1240 * so flush_scheduled_work() can deadlock during device
1243 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1244 if (!ipoib_workqueue) {
1249 ib_sa_register_client(&ipoib_sa_client);
1251 ret = ib_register_client(&ipoib_client);
1258 ib_sa_unregister_client(&ipoib_sa_client);
1259 destroy_workqueue(ipoib_workqueue);
1262 ipoib_unregister_debugfs();
1267 static void __exit ipoib_cleanup_module(void)
1269 ib_unregister_client(&ipoib_client);
1270 ib_sa_unregister_client(&ipoib_sa_client);
1271 ipoib_unregister_debugfs();
1272 destroy_workqueue(ipoib_workqueue);
1275 module_init(ipoib_init_module);
1276 module_exit(ipoib_cleanup_module);