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 napi_enable(&priv->napi);
102 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
104 if (ipoib_pkey_dev_delay_open(dev))
107 if (ipoib_ib_dev_open(dev)) {
108 napi_disable(&priv->napi);
112 if (ipoib_ib_dev_up(dev)) {
113 ipoib_ib_dev_stop(dev, 1);
114 napi_disable(&priv->napi);
118 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
119 struct ipoib_dev_priv *cpriv;
121 /* Bring up any child interfaces too */
122 mutex_lock(&priv->vlan_mutex);
123 list_for_each_entry(cpriv, &priv->child_intfs, list) {
126 flags = cpriv->dev->flags;
130 dev_change_flags(cpriv->dev, flags | IFF_UP);
132 mutex_unlock(&priv->vlan_mutex);
135 netif_start_queue(dev);
140 static int ipoib_stop(struct net_device *dev)
142 struct ipoib_dev_priv *priv = netdev_priv(dev);
144 ipoib_dbg(priv, "stopping interface\n");
146 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
147 napi_disable(&priv->napi);
149 netif_stop_queue(dev);
152 * Now flush workqueue to make sure a scheduled task doesn't
153 * bring our internal state back up.
155 flush_workqueue(ipoib_workqueue);
157 ipoib_ib_dev_down(dev, 1);
158 ipoib_ib_dev_stop(dev, 1);
160 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
161 struct ipoib_dev_priv *cpriv;
163 /* Bring down any child interfaces too */
164 mutex_lock(&priv->vlan_mutex);
165 list_for_each_entry(cpriv, &priv->child_intfs, list) {
168 flags = cpriv->dev->flags;
169 if (!(flags & IFF_UP))
172 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
174 mutex_unlock(&priv->vlan_mutex);
180 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
182 struct ipoib_dev_priv *priv = netdev_priv(dev);
184 /* dev->mtu > 2K ==> connected mode */
185 if (ipoib_cm_admin_enabled(dev)) {
186 if (new_mtu > ipoib_cm_max_mtu(dev))
189 if (new_mtu > priv->mcast_mtu)
190 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
197 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
200 priv->admin_mtu = new_mtu;
202 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
207 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
209 struct ipoib_dev_priv *priv = netdev_priv(dev);
210 struct rb_node *n = priv->path_tree.rb_node;
211 struct ipoib_path *path;
215 path = rb_entry(n, struct ipoib_path, rb_node);
217 ret = memcmp(gid, path->pathrec.dgid.raw,
218 sizeof (union ib_gid));
231 static int __path_add(struct net_device *dev, struct ipoib_path *path)
233 struct ipoib_dev_priv *priv = netdev_priv(dev);
234 struct rb_node **n = &priv->path_tree.rb_node;
235 struct rb_node *pn = NULL;
236 struct ipoib_path *tpath;
241 tpath = rb_entry(pn, struct ipoib_path, rb_node);
243 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
244 sizeof (union ib_gid));
253 rb_link_node(&path->rb_node, pn, n);
254 rb_insert_color(&path->rb_node, &priv->path_tree);
256 list_add_tail(&path->list, &priv->path_list);
261 static void path_free(struct net_device *dev, struct ipoib_path *path)
263 struct ipoib_dev_priv *priv = netdev_priv(dev);
264 struct ipoib_neigh *neigh, *tn;
268 while ((skb = __skb_dequeue(&path->queue)))
269 dev_kfree_skb_irq(skb);
271 spin_lock_irqsave(&priv->lock, flags);
273 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
275 * It's safe to call ipoib_put_ah() inside priv->lock
276 * here, because we know that path->ah will always
277 * hold one more reference, so ipoib_put_ah() will
278 * never do more than decrement the ref count.
281 ipoib_put_ah(neigh->ah);
283 ipoib_neigh_free(dev, neigh);
286 spin_unlock_irqrestore(&priv->lock, flags);
289 ipoib_put_ah(path->ah);
294 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
296 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
298 struct ipoib_path_iter *iter;
300 iter = kmalloc(sizeof *iter, GFP_KERNEL);
305 memset(iter->path.pathrec.dgid.raw, 0, 16);
307 if (ipoib_path_iter_next(iter)) {
315 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
317 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
319 struct ipoib_path *path;
322 spin_lock_irq(&priv->lock);
324 n = rb_first(&priv->path_tree);
327 path = rb_entry(n, struct ipoib_path, rb_node);
329 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
330 sizeof (union ib_gid)) < 0) {
339 spin_unlock_irq(&priv->lock);
344 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
345 struct ipoib_path *path)
350 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
352 void ipoib_flush_paths(struct net_device *dev)
354 struct ipoib_dev_priv *priv = netdev_priv(dev);
355 struct ipoib_path *path, *tp;
356 LIST_HEAD(remove_list);
358 spin_lock_irq(&priv->tx_lock);
359 spin_lock(&priv->lock);
361 list_splice(&priv->path_list, &remove_list);
362 INIT_LIST_HEAD(&priv->path_list);
364 list_for_each_entry(path, &remove_list, list)
365 rb_erase(&path->rb_node, &priv->path_tree);
367 list_for_each_entry_safe(path, tp, &remove_list, list) {
369 ib_sa_cancel_query(path->query_id, path->query);
370 spin_unlock(&priv->lock);
371 spin_unlock_irq(&priv->tx_lock);
372 wait_for_completion(&path->done);
373 path_free(dev, path);
374 spin_lock_irq(&priv->tx_lock);
375 spin_lock(&priv->lock);
377 spin_unlock(&priv->lock);
378 spin_unlock_irq(&priv->tx_lock);
381 static void path_rec_completion(int status,
382 struct ib_sa_path_rec *pathrec,
385 struct ipoib_path *path = path_ptr;
386 struct net_device *dev = path->dev;
387 struct ipoib_dev_priv *priv = netdev_priv(dev);
388 struct ipoib_ah *ah = NULL;
389 struct ipoib_neigh *neigh, *tn;
390 struct sk_buff_head skqueue;
395 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
396 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
398 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
399 status, IPOIB_GID_ARG(path->pathrec.dgid));
401 skb_queue_head_init(&skqueue);
404 struct ib_ah_attr av;
406 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
407 ah = ipoib_create_ah(dev, priv->pd, &av);
410 spin_lock_irqsave(&priv->lock, flags);
415 path->pathrec = *pathrec;
417 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
418 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
420 while ((skb = __skb_dequeue(&path->queue)))
421 __skb_queue_tail(&skqueue, skb);
423 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
424 kref_get(&path->ah->ref);
425 neigh->ah = path->ah;
426 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
427 sizeof(union ib_gid));
429 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
430 if (!ipoib_cm_get(neigh))
431 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
434 if (!ipoib_cm_get(neigh)) {
435 list_del(&neigh->list);
437 ipoib_put_ah(neigh->ah);
438 ipoib_neigh_free(dev, neigh);
443 while ((skb = __skb_dequeue(&neigh->queue)))
444 __skb_queue_tail(&skqueue, skb);
449 complete(&path->done);
451 spin_unlock_irqrestore(&priv->lock, flags);
453 while ((skb = __skb_dequeue(&skqueue))) {
455 if (dev_queue_xmit(skb))
456 ipoib_warn(priv, "dev_queue_xmit failed "
457 "to requeue packet\n");
461 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
463 struct ipoib_dev_priv *priv = netdev_priv(dev);
464 struct ipoib_path *path;
466 if (!priv->broadcast)
469 path = kzalloc(sizeof *path, GFP_ATOMIC);
475 skb_queue_head_init(&path->queue);
477 INIT_LIST_HEAD(&path->neigh_list);
479 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
480 path->pathrec.sgid = priv->local_gid;
481 path->pathrec.pkey = cpu_to_be16(priv->pkey);
482 path->pathrec.numb_path = 1;
483 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
488 static int path_rec_start(struct net_device *dev,
489 struct ipoib_path *path)
491 struct ipoib_dev_priv *priv = netdev_priv(dev);
493 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
494 IPOIB_GID_ARG(path->pathrec.dgid));
496 init_completion(&path->done);
499 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
501 IB_SA_PATH_REC_DGID |
502 IB_SA_PATH_REC_SGID |
503 IB_SA_PATH_REC_NUMB_PATH |
504 IB_SA_PATH_REC_TRAFFIC_CLASS |
509 if (path->query_id < 0) {
510 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
512 return path->query_id;
518 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
520 struct ipoib_dev_priv *priv = netdev_priv(dev);
521 struct ipoib_path *path;
522 struct ipoib_neigh *neigh;
524 neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
526 ++dev->stats.tx_dropped;
527 dev_kfree_skb_any(skb);
532 * We can only be called from ipoib_start_xmit, so we're
533 * inside tx_lock -- no need to save/restore flags.
535 spin_lock(&priv->lock);
537 path = __path_find(dev, skb->dst->neighbour->ha + 4);
539 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
543 __path_add(dev, path);
546 list_add_tail(&neigh->list, &path->neigh_list);
549 kref_get(&path->ah->ref);
550 neigh->ah = path->ah;
551 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
552 sizeof(union ib_gid));
554 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
555 if (!ipoib_cm_get(neigh))
556 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
557 if (!ipoib_cm_get(neigh)) {
558 list_del(&neigh->list);
560 ipoib_put_ah(neigh->ah);
561 ipoib_neigh_free(dev, neigh);
564 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
565 __skb_queue_tail(&neigh->queue, skb);
567 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
568 skb_queue_len(&neigh->queue));
572 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
576 if (!path->query && path_rec_start(dev, path))
579 __skb_queue_tail(&neigh->queue, skb);
582 spin_unlock(&priv->lock);
586 list_del(&neigh->list);
589 ipoib_neigh_free(dev, neigh);
591 ++dev->stats.tx_dropped;
592 dev_kfree_skb_any(skb);
594 spin_unlock(&priv->lock);
597 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
599 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
601 /* Look up path record for unicasts */
602 if (skb->dst->neighbour->ha[4] != 0xff) {
603 neigh_add_path(skb, dev);
607 /* Add in the P_Key for multicasts */
608 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
609 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
610 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
613 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
614 struct ipoib_pseudoheader *phdr)
616 struct ipoib_dev_priv *priv = netdev_priv(dev);
617 struct ipoib_path *path;
620 * We can only be called from ipoib_start_xmit, so we're
621 * inside tx_lock -- no need to save/restore flags.
623 spin_lock(&priv->lock);
625 path = __path_find(dev, phdr->hwaddr + 4);
627 path = path_rec_create(dev, phdr->hwaddr + 4);
629 /* put pseudoheader back on for next time */
630 skb_push(skb, sizeof *phdr);
631 __skb_queue_tail(&path->queue, skb);
633 if (path_rec_start(dev, path)) {
634 spin_unlock(&priv->lock);
635 path_free(dev, path);
638 __path_add(dev, path);
640 ++dev->stats.tx_dropped;
641 dev_kfree_skb_any(skb);
644 spin_unlock(&priv->lock);
649 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
650 be16_to_cpu(path->pathrec.dlid));
652 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
653 } else if ((path->query || !path_rec_start(dev, path)) &&
654 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
655 /* put pseudoheader back on for next time */
656 skb_push(skb, sizeof *phdr);
657 __skb_queue_tail(&path->queue, skb);
659 ++dev->stats.tx_dropped;
660 dev_kfree_skb_any(skb);
663 spin_unlock(&priv->lock);
666 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
668 struct ipoib_dev_priv *priv = netdev_priv(dev);
669 struct ipoib_neigh *neigh;
672 if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
673 return NETDEV_TX_LOCKED;
675 if (likely(skb->dst && skb->dst->neighbour)) {
676 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
677 ipoib_path_lookup(skb, dev);
681 neigh = *to_ipoib_neigh(skb->dst->neighbour);
683 if (ipoib_cm_get(neigh)) {
684 if (ipoib_cm_up(neigh)) {
685 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
688 } else if (neigh->ah) {
689 if (unlikely((memcmp(&neigh->dgid.raw,
690 skb->dst->neighbour->ha + 4,
691 sizeof(union ib_gid))) ||
692 (neigh->dev != dev))) {
693 spin_lock(&priv->lock);
695 * It's safe to call ipoib_put_ah() inside
696 * priv->lock here, because we know that
697 * path->ah will always hold one more reference,
698 * so ipoib_put_ah() will never do more than
699 * decrement the ref count.
701 ipoib_put_ah(neigh->ah);
702 list_del(&neigh->list);
703 ipoib_neigh_free(dev, neigh);
704 spin_unlock(&priv->lock);
705 ipoib_path_lookup(skb, dev);
709 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
713 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
714 spin_lock(&priv->lock);
715 __skb_queue_tail(&neigh->queue, skb);
716 spin_unlock(&priv->lock);
718 ++dev->stats.tx_dropped;
719 dev_kfree_skb_any(skb);
722 struct ipoib_pseudoheader *phdr =
723 (struct ipoib_pseudoheader *) skb->data;
724 skb_pull(skb, sizeof *phdr);
726 if (phdr->hwaddr[4] == 0xff) {
727 /* Add in the P_Key for multicast*/
728 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
729 phdr->hwaddr[9] = priv->pkey & 0xff;
731 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
733 /* unicast GID -- should be ARP or RARP reply */
735 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
736 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
737 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
739 skb->dst ? "neigh" : "dst",
740 be16_to_cpup((__be16 *) skb->data),
741 IPOIB_QPN(phdr->hwaddr),
742 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
743 dev_kfree_skb_any(skb);
744 ++dev->stats.tx_dropped;
748 unicast_arp_send(skb, dev, phdr);
753 spin_unlock_irqrestore(&priv->tx_lock, flags);
758 static void ipoib_timeout(struct net_device *dev)
760 struct ipoib_dev_priv *priv = netdev_priv(dev);
762 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
763 jiffies_to_msecs(jiffies - dev->trans_start));
764 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
765 netif_queue_stopped(dev),
766 priv->tx_head, priv->tx_tail);
767 /* XXX reset QP, etc. */
770 static int ipoib_hard_header(struct sk_buff *skb,
771 struct net_device *dev,
773 const void *daddr, const void *saddr, unsigned len)
775 struct ipoib_header *header;
777 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
779 header->proto = htons(type);
780 header->reserved = 0;
783 * If we don't have a neighbour structure, stuff the
784 * destination address onto the front of the skb so we can
785 * figure out where to send the packet later.
787 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
788 struct ipoib_pseudoheader *phdr =
789 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
790 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
796 static void ipoib_set_mcast_list(struct net_device *dev)
798 struct ipoib_dev_priv *priv = netdev_priv(dev);
800 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
801 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
805 queue_work(ipoib_workqueue, &priv->restart_task);
808 static void ipoib_neigh_cleanup(struct neighbour *n)
810 struct ipoib_neigh *neigh;
811 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
813 struct ipoib_ah *ah = NULL;
815 neigh = *to_ipoib_neigh(n);
817 priv = netdev_priv(neigh->dev);
818 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
823 "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
825 IPOIB_GID_RAW_ARG(n->ha + 4));
827 spin_lock_irqsave(&priv->lock, flags);
831 list_del(&neigh->list);
832 ipoib_neigh_free(n->dev, neigh);
834 spin_unlock_irqrestore(&priv->lock, flags);
840 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
841 struct net_device *dev)
843 struct ipoib_neigh *neigh;
845 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
849 neigh->neighbour = neighbour;
851 *to_ipoib_neigh(neighbour) = neigh;
852 skb_queue_head_init(&neigh->queue);
853 ipoib_cm_set(neigh, NULL);
858 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
861 *to_ipoib_neigh(neigh->neighbour) = NULL;
862 while ((skb = __skb_dequeue(&neigh->queue))) {
863 ++dev->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 & tx_outstanding 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 const struct header_ops ipoib_header_ops = {
939 .create = ipoib_hard_header,
942 static void ipoib_setup(struct net_device *dev)
944 struct ipoib_dev_priv *priv = netdev_priv(dev);
946 dev->open = ipoib_open;
947 dev->stop = ipoib_stop;
948 dev->change_mtu = ipoib_change_mtu;
949 dev->hard_start_xmit = ipoib_start_xmit;
950 dev->tx_timeout = ipoib_timeout;
951 dev->header_ops = &ipoib_header_ops;
952 dev->set_multicast_list = ipoib_set_mcast_list;
953 dev->neigh_setup = ipoib_neigh_setup_dev;
955 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
957 dev->watchdog_timeo = HZ;
959 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
962 * We add in INFINIBAND_ALEN to allow for the destination
963 * address "pseudoheader" for skbs without neighbour struct.
965 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
966 dev->addr_len = INFINIBAND_ALEN;
967 dev->type = ARPHRD_INFINIBAND;
968 dev->tx_queue_len = ipoib_sendq_size * 2;
969 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
971 /* MTU will be reset when mcast join happens */
972 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
973 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
975 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
977 netif_carrier_off(dev);
981 spin_lock_init(&priv->lock);
982 spin_lock_init(&priv->tx_lock);
984 mutex_init(&priv->mcast_mutex);
985 mutex_init(&priv->vlan_mutex);
987 INIT_LIST_HEAD(&priv->path_list);
988 INIT_LIST_HEAD(&priv->child_intfs);
989 INIT_LIST_HEAD(&priv->dead_ahs);
990 INIT_LIST_HEAD(&priv->multicast_list);
992 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
993 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
994 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
995 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
996 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
997 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1000 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1002 struct net_device *dev;
1004 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1009 return netdev_priv(dev);
1012 static ssize_t show_pkey(struct device *dev,
1013 struct device_attribute *attr, char *buf)
1015 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1017 return sprintf(buf, "0x%04x\n", priv->pkey);
1019 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1021 static ssize_t show_umcast(struct device *dev,
1022 struct device_attribute *attr, char *buf)
1024 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1026 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1029 static ssize_t set_umcast(struct device *dev,
1030 struct device_attribute *attr,
1031 const char *buf, size_t count)
1033 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1034 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1036 if (umcast_val > 0) {
1037 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1038 ipoib_warn(priv, "ignoring multicast groups joined directly "
1041 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1045 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1047 int ipoib_add_umcast_attr(struct net_device *dev)
1049 return device_create_file(&dev->dev, &dev_attr_umcast);
1052 static ssize_t create_child(struct device *dev,
1053 struct device_attribute *attr,
1054 const char *buf, size_t count)
1059 if (sscanf(buf, "%i", &pkey) != 1)
1062 if (pkey < 0 || pkey > 0xffff)
1066 * Set the full membership bit, so that we join the right
1067 * broadcast group, etc.
1071 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1073 return ret ? ret : count;
1075 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1077 static ssize_t delete_child(struct device *dev,
1078 struct device_attribute *attr,
1079 const char *buf, size_t count)
1084 if (sscanf(buf, "%i", &pkey) != 1)
1087 if (pkey < 0 || pkey > 0xffff)
1090 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1092 return ret ? ret : count;
1095 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1097 int ipoib_add_pkey_attr(struct net_device *dev)
1099 return device_create_file(&dev->dev, &dev_attr_pkey);
1102 static struct net_device *ipoib_add_port(const char *format,
1103 struct ib_device *hca, u8 port)
1105 struct ipoib_dev_priv *priv;
1106 int result = -ENOMEM;
1108 priv = ipoib_intf_alloc(format);
1110 goto alloc_mem_failed;
1112 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1114 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1116 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1117 hca->name, port, result);
1118 goto device_init_failed;
1122 * Set the full membership bit, so that we join the right
1123 * broadcast group, etc.
1125 priv->pkey |= 0x8000;
1127 priv->dev->broadcast[8] = priv->pkey >> 8;
1128 priv->dev->broadcast[9] = priv->pkey & 0xff;
1130 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1132 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1133 hca->name, port, result);
1134 goto device_init_failed;
1136 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1139 result = ipoib_dev_init(priv->dev, hca, port);
1141 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1142 hca->name, port, result);
1143 goto device_init_failed;
1146 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1147 priv->ca, ipoib_event);
1148 result = ib_register_event_handler(&priv->event_handler);
1150 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1151 "port %d (ret = %d)\n",
1152 hca->name, port, result);
1156 result = register_netdev(priv->dev);
1158 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1159 hca->name, port, result);
1160 goto register_failed;
1163 ipoib_create_debug_files(priv->dev);
1165 if (ipoib_cm_add_mode_attr(priv->dev))
1167 if (ipoib_add_pkey_attr(priv->dev))
1169 if (ipoib_add_umcast_attr(priv->dev))
1171 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1173 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1179 ipoib_delete_debug_files(priv->dev);
1180 unregister_netdev(priv->dev);
1183 ib_unregister_event_handler(&priv->event_handler);
1184 flush_scheduled_work();
1187 ipoib_dev_cleanup(priv->dev);
1190 free_netdev(priv->dev);
1193 return ERR_PTR(result);
1196 static void ipoib_add_one(struct ib_device *device)
1198 struct list_head *dev_list;
1199 struct net_device *dev;
1200 struct ipoib_dev_priv *priv;
1203 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1206 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1210 INIT_LIST_HEAD(dev_list);
1212 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1217 e = device->phys_port_cnt;
1220 for (p = s; p <= e; ++p) {
1221 dev = ipoib_add_port("ib%d", device, p);
1223 priv = netdev_priv(dev);
1224 list_add_tail(&priv->list, dev_list);
1228 ib_set_client_data(device, &ipoib_client, dev_list);
1231 static void ipoib_remove_one(struct ib_device *device)
1233 struct ipoib_dev_priv *priv, *tmp;
1234 struct list_head *dev_list;
1236 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1239 dev_list = ib_get_client_data(device, &ipoib_client);
1241 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1242 ib_unregister_event_handler(&priv->event_handler);
1243 flush_scheduled_work();
1245 unregister_netdev(priv->dev);
1246 ipoib_dev_cleanup(priv->dev);
1247 free_netdev(priv->dev);
1253 static int __init ipoib_init_module(void)
1257 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1258 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1259 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1261 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1262 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1263 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1264 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1265 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1268 ret = ipoib_register_debugfs();
1273 * We create our own workqueue mainly because we want to be
1274 * able to flush it when devices are being removed. We can't
1275 * use schedule_work()/flush_scheduled_work() because both
1276 * unregister_netdev() and linkwatch_event take the rtnl lock,
1277 * so flush_scheduled_work() can deadlock during device
1280 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1281 if (!ipoib_workqueue) {
1286 ib_sa_register_client(&ipoib_sa_client);
1288 ret = ib_register_client(&ipoib_client);
1295 ib_sa_unregister_client(&ipoib_sa_client);
1296 destroy_workqueue(ipoib_workqueue);
1299 ipoib_unregister_debugfs();
1304 static void __exit ipoib_cleanup_module(void)
1306 ib_unregister_client(&ipoib_client);
1307 ib_sa_unregister_client(&ipoib_sa_client);
1308 ipoib_unregister_debugfs();
1309 destroy_workqueue(ipoib_workqueue);
1312 module_init(ipoib_init_module);
1313 module_exit(ipoib_cleanup_module);