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>
44 #include <linux/vmalloc.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 struct ib_sa_client ipoib_sa_client;
87 static void ipoib_add_one(struct ib_device *device);
88 static void ipoib_remove_one(struct ib_device *device);
90 static struct ib_client ipoib_client = {
93 .remove = ipoib_remove_one
96 int ipoib_open(struct net_device *dev)
98 struct ipoib_dev_priv *priv = netdev_priv(dev);
100 ipoib_dbg(priv, "bringing up interface\n");
102 napi_enable(&priv->napi);
103 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
105 if (ipoib_pkey_dev_delay_open(dev))
108 if (ipoib_ib_dev_open(dev)) {
109 napi_disable(&priv->napi);
113 if (ipoib_ib_dev_up(dev)) {
114 ipoib_ib_dev_stop(dev, 1);
115 napi_disable(&priv->napi);
119 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
120 struct ipoib_dev_priv *cpriv;
122 /* Bring up any child interfaces too */
123 mutex_lock(&priv->vlan_mutex);
124 list_for_each_entry(cpriv, &priv->child_intfs, list) {
127 flags = cpriv->dev->flags;
131 dev_change_flags(cpriv->dev, flags | IFF_UP);
133 mutex_unlock(&priv->vlan_mutex);
136 netif_start_queue(dev);
141 static int ipoib_stop(struct net_device *dev)
143 struct ipoib_dev_priv *priv = netdev_priv(dev);
145 ipoib_dbg(priv, "stopping interface\n");
147 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
148 napi_disable(&priv->napi);
150 netif_stop_queue(dev);
153 * Now flush workqueue to make sure a scheduled task doesn't
154 * bring our internal state back up.
156 flush_workqueue(ipoib_workqueue);
158 ipoib_ib_dev_down(dev, 1);
159 ipoib_ib_dev_stop(dev, 1);
161 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
162 struct ipoib_dev_priv *cpriv;
164 /* Bring down any child interfaces too */
165 mutex_lock(&priv->vlan_mutex);
166 list_for_each_entry(cpriv, &priv->child_intfs, list) {
169 flags = cpriv->dev->flags;
170 if (!(flags & IFF_UP))
173 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
175 mutex_unlock(&priv->vlan_mutex);
181 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
183 struct ipoib_dev_priv *priv = netdev_priv(dev);
185 /* dev->mtu > 2K ==> connected mode */
186 if (ipoib_cm_admin_enabled(dev)) {
187 if (new_mtu > ipoib_cm_max_mtu(dev))
190 if (new_mtu > priv->mcast_mtu)
191 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
198 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN)
201 priv->admin_mtu = new_mtu;
203 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
208 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
210 struct ipoib_dev_priv *priv = netdev_priv(dev);
211 struct rb_node *n = priv->path_tree.rb_node;
212 struct ipoib_path *path;
216 path = rb_entry(n, struct ipoib_path, rb_node);
218 ret = memcmp(gid, path->pathrec.dgid.raw,
219 sizeof (union ib_gid));
232 static int __path_add(struct net_device *dev, struct ipoib_path *path)
234 struct ipoib_dev_priv *priv = netdev_priv(dev);
235 struct rb_node **n = &priv->path_tree.rb_node;
236 struct rb_node *pn = NULL;
237 struct ipoib_path *tpath;
242 tpath = rb_entry(pn, struct ipoib_path, rb_node);
244 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
245 sizeof (union ib_gid));
254 rb_link_node(&path->rb_node, pn, n);
255 rb_insert_color(&path->rb_node, &priv->path_tree);
257 list_add_tail(&path->list, &priv->path_list);
262 static void path_free(struct net_device *dev, struct ipoib_path *path)
264 struct ipoib_dev_priv *priv = netdev_priv(dev);
265 struct ipoib_neigh *neigh, *tn;
269 while ((skb = __skb_dequeue(&path->queue)))
270 dev_kfree_skb_irq(skb);
272 spin_lock_irqsave(&priv->lock, flags);
274 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
276 * It's safe to call ipoib_put_ah() inside priv->lock
277 * here, because we know that path->ah will always
278 * hold one more reference, so ipoib_put_ah() will
279 * never do more than decrement the ref count.
282 ipoib_put_ah(neigh->ah);
284 ipoib_neigh_free(dev, neigh);
287 spin_unlock_irqrestore(&priv->lock, flags);
290 ipoib_put_ah(path->ah);
295 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
297 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
299 struct ipoib_path_iter *iter;
301 iter = kmalloc(sizeof *iter, GFP_KERNEL);
306 memset(iter->path.pathrec.dgid.raw, 0, 16);
308 if (ipoib_path_iter_next(iter)) {
316 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
318 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
320 struct ipoib_path *path;
323 spin_lock_irq(&priv->lock);
325 n = rb_first(&priv->path_tree);
328 path = rb_entry(n, struct ipoib_path, rb_node);
330 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
331 sizeof (union ib_gid)) < 0) {
340 spin_unlock_irq(&priv->lock);
345 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
346 struct ipoib_path *path)
351 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
353 void ipoib_flush_paths(struct net_device *dev)
355 struct ipoib_dev_priv *priv = netdev_priv(dev);
356 struct ipoib_path *path, *tp;
357 LIST_HEAD(remove_list);
359 spin_lock_irq(&priv->tx_lock);
360 spin_lock(&priv->lock);
362 list_splice(&priv->path_list, &remove_list);
363 INIT_LIST_HEAD(&priv->path_list);
365 list_for_each_entry(path, &remove_list, list)
366 rb_erase(&path->rb_node, &priv->path_tree);
368 list_for_each_entry_safe(path, tp, &remove_list, list) {
370 ib_sa_cancel_query(path->query_id, path->query);
371 spin_unlock(&priv->lock);
372 spin_unlock_irq(&priv->tx_lock);
373 wait_for_completion(&path->done);
374 path_free(dev, path);
375 spin_lock_irq(&priv->tx_lock);
376 spin_lock(&priv->lock);
378 spin_unlock(&priv->lock);
379 spin_unlock_irq(&priv->tx_lock);
382 static void path_rec_completion(int status,
383 struct ib_sa_path_rec *pathrec,
386 struct ipoib_path *path = path_ptr;
387 struct net_device *dev = path->dev;
388 struct ipoib_dev_priv *priv = netdev_priv(dev);
389 struct ipoib_ah *ah = NULL;
390 struct ipoib_neigh *neigh, *tn;
391 struct sk_buff_head skqueue;
396 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
397 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
399 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
400 status, IPOIB_GID_ARG(path->pathrec.dgid));
402 skb_queue_head_init(&skqueue);
405 struct ib_ah_attr av;
407 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
408 ah = ipoib_create_ah(dev, priv->pd, &av);
411 spin_lock_irqsave(&priv->lock, flags);
416 path->pathrec = *pathrec;
418 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
419 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
421 while ((skb = __skb_dequeue(&path->queue)))
422 __skb_queue_tail(&skqueue, skb);
424 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
425 kref_get(&path->ah->ref);
426 neigh->ah = path->ah;
427 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
428 sizeof(union ib_gid));
430 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
431 if (!ipoib_cm_get(neigh))
432 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
435 if (!ipoib_cm_get(neigh)) {
436 list_del(&neigh->list);
438 ipoib_put_ah(neigh->ah);
439 ipoib_neigh_free(dev, neigh);
444 while ((skb = __skb_dequeue(&neigh->queue)))
445 __skb_queue_tail(&skqueue, skb);
450 complete(&path->done);
452 spin_unlock_irqrestore(&priv->lock, flags);
454 while ((skb = __skb_dequeue(&skqueue))) {
456 if (dev_queue_xmit(skb))
457 ipoib_warn(priv, "dev_queue_xmit failed "
458 "to requeue packet\n");
462 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
464 struct ipoib_dev_priv *priv = netdev_priv(dev);
465 struct ipoib_path *path;
467 if (!priv->broadcast)
470 path = kzalloc(sizeof *path, GFP_ATOMIC);
476 skb_queue_head_init(&path->queue);
478 INIT_LIST_HEAD(&path->neigh_list);
480 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
481 path->pathrec.sgid = priv->local_gid;
482 path->pathrec.pkey = cpu_to_be16(priv->pkey);
483 path->pathrec.numb_path = 1;
484 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
489 static int path_rec_start(struct net_device *dev,
490 struct ipoib_path *path)
492 struct ipoib_dev_priv *priv = netdev_priv(dev);
494 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
495 IPOIB_GID_ARG(path->pathrec.dgid));
497 init_completion(&path->done);
500 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
502 IB_SA_PATH_REC_DGID |
503 IB_SA_PATH_REC_SGID |
504 IB_SA_PATH_REC_NUMB_PATH |
505 IB_SA_PATH_REC_TRAFFIC_CLASS |
510 if (path->query_id < 0) {
511 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
513 return path->query_id;
519 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
521 struct ipoib_dev_priv *priv = netdev_priv(dev);
522 struct ipoib_path *path;
523 struct ipoib_neigh *neigh;
525 neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
527 ++dev->stats.tx_dropped;
528 dev_kfree_skb_any(skb);
533 * We can only be called from ipoib_start_xmit, so we're
534 * inside tx_lock -- no need to save/restore flags.
536 spin_lock(&priv->lock);
538 path = __path_find(dev, skb->dst->neighbour->ha + 4);
540 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
544 __path_add(dev, path);
547 list_add_tail(&neigh->list, &path->neigh_list);
550 kref_get(&path->ah->ref);
551 neigh->ah = path->ah;
552 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
553 sizeof(union ib_gid));
555 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
556 if (!ipoib_cm_get(neigh))
557 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
558 if (!ipoib_cm_get(neigh)) {
559 list_del(&neigh->list);
561 ipoib_put_ah(neigh->ah);
562 ipoib_neigh_free(dev, neigh);
565 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
566 __skb_queue_tail(&neigh->queue, skb);
568 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
569 skb_queue_len(&neigh->queue));
573 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
577 if (!path->query && path_rec_start(dev, path))
580 __skb_queue_tail(&neigh->queue, skb);
583 spin_unlock(&priv->lock);
587 list_del(&neigh->list);
590 ipoib_neigh_free(dev, neigh);
592 ++dev->stats.tx_dropped;
593 dev_kfree_skb_any(skb);
595 spin_unlock(&priv->lock);
598 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
600 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
602 /* Look up path record for unicasts */
603 if (skb->dst->neighbour->ha[4] != 0xff) {
604 neigh_add_path(skb, dev);
608 /* Add in the P_Key for multicasts */
609 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
610 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
611 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
614 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
615 struct ipoib_pseudoheader *phdr)
617 struct ipoib_dev_priv *priv = netdev_priv(dev);
618 struct ipoib_path *path;
621 * We can only be called from ipoib_start_xmit, so we're
622 * inside tx_lock -- no need to save/restore flags.
624 spin_lock(&priv->lock);
626 path = __path_find(dev, phdr->hwaddr + 4);
628 path = path_rec_create(dev, phdr->hwaddr + 4);
630 /* put pseudoheader back on for next time */
631 skb_push(skb, sizeof *phdr);
632 __skb_queue_tail(&path->queue, skb);
634 if (path_rec_start(dev, path)) {
635 spin_unlock(&priv->lock);
636 path_free(dev, path);
639 __path_add(dev, path);
641 ++dev->stats.tx_dropped;
642 dev_kfree_skb_any(skb);
645 spin_unlock(&priv->lock);
650 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
651 be16_to_cpu(path->pathrec.dlid));
653 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
654 } else if ((path->query || !path_rec_start(dev, path)) &&
655 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
656 /* put pseudoheader back on for next time */
657 skb_push(skb, sizeof *phdr);
658 __skb_queue_tail(&path->queue, skb);
660 ++dev->stats.tx_dropped;
661 dev_kfree_skb_any(skb);
664 spin_unlock(&priv->lock);
667 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
669 struct ipoib_dev_priv *priv = netdev_priv(dev);
670 struct ipoib_neigh *neigh;
673 if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
674 return NETDEV_TX_LOCKED;
676 if (likely(skb->dst && skb->dst->neighbour)) {
677 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
678 ipoib_path_lookup(skb, dev);
682 neigh = *to_ipoib_neigh(skb->dst->neighbour);
685 if (unlikely((memcmp(&neigh->dgid.raw,
686 skb->dst->neighbour->ha + 4,
687 sizeof(union ib_gid))) ||
688 (neigh->dev != dev))) {
689 spin_lock(&priv->lock);
691 * It's safe to call ipoib_put_ah() inside
692 * priv->lock here, because we know that
693 * path->ah will always hold one more reference,
694 * so ipoib_put_ah() will never do more than
695 * decrement the ref count.
697 ipoib_put_ah(neigh->ah);
698 list_del(&neigh->list);
699 ipoib_neigh_free(dev, neigh);
700 spin_unlock(&priv->lock);
701 ipoib_path_lookup(skb, dev);
705 if (ipoib_cm_get(neigh)) {
706 if (ipoib_cm_up(neigh)) {
707 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
710 } else if (neigh->ah) {
711 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
715 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
716 spin_lock(&priv->lock);
717 __skb_queue_tail(&neigh->queue, skb);
718 spin_unlock(&priv->lock);
720 ++dev->stats.tx_dropped;
721 dev_kfree_skb_any(skb);
724 struct ipoib_pseudoheader *phdr =
725 (struct ipoib_pseudoheader *) skb->data;
726 skb_pull(skb, sizeof *phdr);
728 if (phdr->hwaddr[4] == 0xff) {
729 /* Add in the P_Key for multicast*/
730 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
731 phdr->hwaddr[9] = priv->pkey & 0xff;
733 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
735 /* unicast GID -- should be ARP or RARP reply */
737 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
738 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
739 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
741 skb->dst ? "neigh" : "dst",
742 be16_to_cpup((__be16 *) skb->data),
743 IPOIB_QPN(phdr->hwaddr),
744 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
745 dev_kfree_skb_any(skb);
746 ++dev->stats.tx_dropped;
750 unicast_arp_send(skb, dev, phdr);
755 spin_unlock_irqrestore(&priv->tx_lock, flags);
760 static void ipoib_timeout(struct net_device *dev)
762 struct ipoib_dev_priv *priv = netdev_priv(dev);
764 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
765 jiffies_to_msecs(jiffies - dev->trans_start));
766 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
767 netif_queue_stopped(dev),
768 priv->tx_head, priv->tx_tail);
769 /* XXX reset QP, etc. */
772 static int ipoib_hard_header(struct sk_buff *skb,
773 struct net_device *dev,
775 const void *daddr, const void *saddr, unsigned len)
777 struct ipoib_header *header;
779 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
781 header->proto = htons(type);
782 header->reserved = 0;
785 * If we don't have a neighbour structure, stuff the
786 * destination address onto the front of the skb so we can
787 * figure out where to send the packet later.
789 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
790 struct ipoib_pseudoheader *phdr =
791 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
792 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
798 static void ipoib_set_mcast_list(struct net_device *dev)
800 struct ipoib_dev_priv *priv = netdev_priv(dev);
802 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
803 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
807 queue_work(ipoib_workqueue, &priv->restart_task);
810 static void ipoib_neigh_cleanup(struct neighbour *n)
812 struct ipoib_neigh *neigh;
813 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
815 struct ipoib_ah *ah = NULL;
817 neigh = *to_ipoib_neigh(n);
819 priv = netdev_priv(neigh->dev);
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 = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
892 if (!priv->tx_ring) {
893 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
894 ca->name, ipoib_sendq_size);
895 goto out_rx_ring_cleanup;
897 memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
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 vfree(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 vfree(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 |
973 /* MTU will be reset when mcast join happens */
974 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
975 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
977 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
979 netif_carrier_off(dev);
983 spin_lock_init(&priv->lock);
984 spin_lock_init(&priv->tx_lock);
986 mutex_init(&priv->mcast_mutex);
987 mutex_init(&priv->vlan_mutex);
989 INIT_LIST_HEAD(&priv->path_list);
990 INIT_LIST_HEAD(&priv->child_intfs);
991 INIT_LIST_HEAD(&priv->dead_ahs);
992 INIT_LIST_HEAD(&priv->multicast_list);
994 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
995 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
996 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
997 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
998 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
999 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1002 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1004 struct net_device *dev;
1006 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1011 return netdev_priv(dev);
1014 static ssize_t show_pkey(struct device *dev,
1015 struct device_attribute *attr, char *buf)
1017 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1019 return sprintf(buf, "0x%04x\n", priv->pkey);
1021 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1023 static ssize_t show_umcast(struct device *dev,
1024 struct device_attribute *attr, char *buf)
1026 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1028 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1031 static ssize_t set_umcast(struct device *dev,
1032 struct device_attribute *attr,
1033 const char *buf, size_t count)
1035 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1036 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1038 if (umcast_val > 0) {
1039 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1040 ipoib_warn(priv, "ignoring multicast groups joined directly "
1043 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1047 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1049 int ipoib_add_umcast_attr(struct net_device *dev)
1051 return device_create_file(&dev->dev, &dev_attr_umcast);
1054 static ssize_t create_child(struct device *dev,
1055 struct device_attribute *attr,
1056 const char *buf, size_t count)
1061 if (sscanf(buf, "%i", &pkey) != 1)
1064 if (pkey < 0 || pkey > 0xffff)
1068 * Set the full membership bit, so that we join the right
1069 * broadcast group, etc.
1073 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1075 return ret ? ret : count;
1077 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1079 static ssize_t delete_child(struct device *dev,
1080 struct device_attribute *attr,
1081 const char *buf, size_t count)
1086 if (sscanf(buf, "%i", &pkey) != 1)
1089 if (pkey < 0 || pkey > 0xffff)
1092 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1094 return ret ? ret : count;
1097 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1099 int ipoib_add_pkey_attr(struct net_device *dev)
1101 return device_create_file(&dev->dev, &dev_attr_pkey);
1104 static struct net_device *ipoib_add_port(const char *format,
1105 struct ib_device *hca, u8 port)
1107 struct ipoib_dev_priv *priv;
1108 int result = -ENOMEM;
1110 priv = ipoib_intf_alloc(format);
1112 goto alloc_mem_failed;
1114 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1116 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1118 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1119 hca->name, port, result);
1120 goto device_init_failed;
1124 * Set the full membership bit, so that we join the right
1125 * broadcast group, etc.
1127 priv->pkey |= 0x8000;
1129 priv->dev->broadcast[8] = priv->pkey >> 8;
1130 priv->dev->broadcast[9] = priv->pkey & 0xff;
1132 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1134 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1135 hca->name, port, result);
1136 goto device_init_failed;
1138 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1141 result = ipoib_dev_init(priv->dev, hca, port);
1143 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1144 hca->name, port, result);
1145 goto device_init_failed;
1148 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1149 priv->ca, ipoib_event);
1150 result = ib_register_event_handler(&priv->event_handler);
1152 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1153 "port %d (ret = %d)\n",
1154 hca->name, port, result);
1158 result = register_netdev(priv->dev);
1160 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1161 hca->name, port, result);
1162 goto register_failed;
1165 ipoib_create_debug_files(priv->dev);
1167 if (ipoib_cm_add_mode_attr(priv->dev))
1169 if (ipoib_add_pkey_attr(priv->dev))
1171 if (ipoib_add_umcast_attr(priv->dev))
1173 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1175 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1181 ipoib_delete_debug_files(priv->dev);
1182 unregister_netdev(priv->dev);
1185 ib_unregister_event_handler(&priv->event_handler);
1186 flush_scheduled_work();
1189 ipoib_dev_cleanup(priv->dev);
1192 free_netdev(priv->dev);
1195 return ERR_PTR(result);
1198 static void ipoib_add_one(struct ib_device *device)
1200 struct list_head *dev_list;
1201 struct net_device *dev;
1202 struct ipoib_dev_priv *priv;
1205 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1208 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1212 INIT_LIST_HEAD(dev_list);
1214 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1219 e = device->phys_port_cnt;
1222 for (p = s; p <= e; ++p) {
1223 dev = ipoib_add_port("ib%d", device, p);
1225 priv = netdev_priv(dev);
1226 list_add_tail(&priv->list, dev_list);
1230 ib_set_client_data(device, &ipoib_client, dev_list);
1233 static void ipoib_remove_one(struct ib_device *device)
1235 struct ipoib_dev_priv *priv, *tmp;
1236 struct list_head *dev_list;
1238 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1241 dev_list = ib_get_client_data(device, &ipoib_client);
1243 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1244 ib_unregister_event_handler(&priv->event_handler);
1245 flush_scheduled_work();
1247 unregister_netdev(priv->dev);
1248 ipoib_dev_cleanup(priv->dev);
1249 free_netdev(priv->dev);
1255 static int __init ipoib_init_module(void)
1259 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1260 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1261 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1263 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1264 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1265 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1266 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1267 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1270 ret = ipoib_register_debugfs();
1275 * We create our own workqueue mainly because we want to be
1276 * able to flush it when devices are being removed. We can't
1277 * use schedule_work()/flush_scheduled_work() because both
1278 * unregister_netdev() and linkwatch_event take the rtnl lock,
1279 * so flush_scheduled_work() can deadlock during device
1282 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1283 if (!ipoib_workqueue) {
1288 ib_sa_register_client(&ipoib_sa_client);
1290 ret = ib_register_client(&ipoib_client);
1297 ib_sa_unregister_client(&ipoib_sa_client);
1298 destroy_workqueue(ipoib_workqueue);
1301 ipoib_unregister_debugfs();
1306 static void __exit ipoib_cleanup_module(void)
1308 ib_unregister_client(&ipoib_client);
1309 ib_sa_unregister_client(&ipoib_sa_client);
1310 ipoib_unregister_debugfs();
1311 destroy_workqueue(ipoib_workqueue);
1314 module_init(ipoib_init_module);
1315 module_exit(ipoib_cleanup_module);