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
37 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
44 #include <linux/if_arp.h> /* For ARPHRD_xxx */
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
53 MODULE_LICENSE("Dual BSD/GPL");
55 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
56 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
58 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
59 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
60 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
61 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
64 module_param(lro, bool, 0444);
65 MODULE_PARM_DESC(lro, "Enable LRO (Large Receive Offload)");
67 static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
68 module_param(lro_max_aggr, int, 0644);
69 MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
72 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
73 int ipoib_debug_level;
75 module_param_named(debug_level, ipoib_debug_level, int, 0644);
76 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
79 struct ipoib_path_iter {
80 struct net_device *dev;
81 struct ipoib_path path;
84 static const u8 ipv4_bcast_addr[] = {
85 0x00, 0xff, 0xff, 0xff,
86 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
87 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
90 struct workqueue_struct *ipoib_workqueue;
92 struct ib_sa_client ipoib_sa_client;
94 static void ipoib_add_one(struct ib_device *device);
95 static void ipoib_remove_one(struct ib_device *device);
97 static struct ib_client ipoib_client = {
100 .remove = ipoib_remove_one
103 int ipoib_open(struct net_device *dev)
105 struct ipoib_dev_priv *priv = netdev_priv(dev);
107 ipoib_dbg(priv, "bringing up interface\n");
109 napi_enable(&priv->napi);
110 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
112 if (ipoib_pkey_dev_delay_open(dev))
115 if (ipoib_ib_dev_open(dev)) {
116 napi_disable(&priv->napi);
120 if (ipoib_ib_dev_up(dev)) {
121 ipoib_ib_dev_stop(dev, 1);
122 napi_disable(&priv->napi);
126 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
127 struct ipoib_dev_priv *cpriv;
129 /* Bring up any child interfaces too */
130 mutex_lock(&priv->vlan_mutex);
131 list_for_each_entry(cpriv, &priv->child_intfs, list) {
134 flags = cpriv->dev->flags;
138 dev_change_flags(cpriv->dev, flags | IFF_UP);
140 mutex_unlock(&priv->vlan_mutex);
143 netif_start_queue(dev);
148 static int ipoib_stop(struct net_device *dev)
150 struct ipoib_dev_priv *priv = netdev_priv(dev);
152 ipoib_dbg(priv, "stopping interface\n");
154 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
155 napi_disable(&priv->napi);
157 netif_stop_queue(dev);
159 ipoib_ib_dev_down(dev, 0);
160 ipoib_ib_dev_stop(dev, 0);
162 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
163 struct ipoib_dev_priv *cpriv;
165 /* Bring down any child interfaces too */
166 mutex_lock(&priv->vlan_mutex);
167 list_for_each_entry(cpriv, &priv->child_intfs, list) {
170 flags = cpriv->dev->flags;
171 if (!(flags & IFF_UP))
174 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
176 mutex_unlock(&priv->vlan_mutex);
182 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
184 struct ipoib_dev_priv *priv = netdev_priv(dev);
186 /* dev->mtu > 2K ==> connected mode */
187 if (ipoib_cm_admin_enabled(dev)) {
188 if (new_mtu > ipoib_cm_max_mtu(dev))
191 if (new_mtu > priv->mcast_mtu)
192 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
199 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
202 priv->admin_mtu = new_mtu;
204 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
209 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
211 struct ipoib_dev_priv *priv = netdev_priv(dev);
212 struct rb_node *n = priv->path_tree.rb_node;
213 struct ipoib_path *path;
217 path = rb_entry(n, struct ipoib_path, rb_node);
219 ret = memcmp(gid, path->pathrec.dgid.raw,
220 sizeof (union ib_gid));
233 static int __path_add(struct net_device *dev, struct ipoib_path *path)
235 struct ipoib_dev_priv *priv = netdev_priv(dev);
236 struct rb_node **n = &priv->path_tree.rb_node;
237 struct rb_node *pn = NULL;
238 struct ipoib_path *tpath;
243 tpath = rb_entry(pn, struct ipoib_path, rb_node);
245 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
246 sizeof (union ib_gid));
255 rb_link_node(&path->rb_node, pn, n);
256 rb_insert_color(&path->rb_node, &priv->path_tree);
258 list_add_tail(&path->list, &priv->path_list);
263 static void path_free(struct net_device *dev, struct ipoib_path *path)
265 struct ipoib_dev_priv *priv = netdev_priv(dev);
266 struct ipoib_neigh *neigh, *tn;
270 while ((skb = __skb_dequeue(&path->queue)))
271 dev_kfree_skb_irq(skb);
273 spin_lock_irqsave(&priv->lock, flags);
275 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
277 * It's safe to call ipoib_put_ah() inside priv->lock
278 * here, because we know that path->ah will always
279 * hold one more reference, so ipoib_put_ah() will
280 * never do more than decrement the ref count.
283 ipoib_put_ah(neigh->ah);
285 ipoib_neigh_free(dev, neigh);
288 spin_unlock_irqrestore(&priv->lock, flags);
291 ipoib_put_ah(path->ah);
296 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
298 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
300 struct ipoib_path_iter *iter;
302 iter = kmalloc(sizeof *iter, GFP_KERNEL);
307 memset(iter->path.pathrec.dgid.raw, 0, 16);
309 if (ipoib_path_iter_next(iter)) {
317 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
319 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
321 struct ipoib_path *path;
324 spin_lock_irq(&priv->lock);
326 n = rb_first(&priv->path_tree);
329 path = rb_entry(n, struct ipoib_path, rb_node);
331 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
332 sizeof (union ib_gid)) < 0) {
341 spin_unlock_irq(&priv->lock);
346 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
347 struct ipoib_path *path)
352 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
354 void ipoib_mark_paths_invalid(struct net_device *dev)
356 struct ipoib_dev_priv *priv = netdev_priv(dev);
357 struct ipoib_path *path, *tp;
359 spin_lock_irq(&priv->lock);
361 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
362 ipoib_dbg(priv, "mark path LID 0x%04x GID %p6 invalid\n",
363 be16_to_cpu(path->pathrec.dlid),
364 path->pathrec.dgid.raw);
368 spin_unlock_irq(&priv->lock);
371 void ipoib_flush_paths(struct net_device *dev)
373 struct ipoib_dev_priv *priv = netdev_priv(dev);
374 struct ipoib_path *path, *tp;
375 LIST_HEAD(remove_list);
378 netif_tx_lock_bh(dev);
379 spin_lock_irqsave(&priv->lock, flags);
381 list_splice_init(&priv->path_list, &remove_list);
383 list_for_each_entry(path, &remove_list, list)
384 rb_erase(&path->rb_node, &priv->path_tree);
386 list_for_each_entry_safe(path, tp, &remove_list, list) {
388 ib_sa_cancel_query(path->query_id, path->query);
389 spin_unlock_irqrestore(&priv->lock, flags);
390 netif_tx_unlock_bh(dev);
391 wait_for_completion(&path->done);
392 path_free(dev, path);
393 netif_tx_lock_bh(dev);
394 spin_lock_irqsave(&priv->lock, flags);
397 spin_unlock_irqrestore(&priv->lock, flags);
398 netif_tx_unlock_bh(dev);
401 static void path_rec_completion(int status,
402 struct ib_sa_path_rec *pathrec,
405 struct ipoib_path *path = path_ptr;
406 struct net_device *dev = path->dev;
407 struct ipoib_dev_priv *priv = netdev_priv(dev);
408 struct ipoib_ah *ah = NULL;
409 struct ipoib_ah *old_ah = NULL;
410 struct ipoib_neigh *neigh, *tn;
411 struct sk_buff_head skqueue;
416 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %p6\n",
417 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
419 ipoib_dbg(priv, "PathRec status %d for GID %p6\n",
420 status, path->pathrec.dgid.raw);
422 skb_queue_head_init(&skqueue);
425 struct ib_ah_attr av;
427 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
428 ah = ipoib_create_ah(dev, priv->pd, &av);
431 spin_lock_irqsave(&priv->lock, flags);
434 path->pathrec = *pathrec;
439 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
440 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
442 while ((skb = __skb_dequeue(&path->queue)))
443 __skb_queue_tail(&skqueue, skb);
445 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
447 WARN_ON(neigh->ah != old_ah);
449 * Dropping the ah reference inside
450 * priv->lock is safe here, because we
451 * will hold one more reference from
452 * the original value of path->ah (ie
455 ipoib_put_ah(neigh->ah);
457 kref_get(&path->ah->ref);
458 neigh->ah = path->ah;
459 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
460 sizeof(union ib_gid));
462 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
463 if (!ipoib_cm_get(neigh))
464 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
467 if (!ipoib_cm_get(neigh)) {
468 list_del(&neigh->list);
470 ipoib_put_ah(neigh->ah);
471 ipoib_neigh_free(dev, neigh);
476 while ((skb = __skb_dequeue(&neigh->queue)))
477 __skb_queue_tail(&skqueue, skb);
483 complete(&path->done);
485 spin_unlock_irqrestore(&priv->lock, flags);
488 ipoib_put_ah(old_ah);
490 while ((skb = __skb_dequeue(&skqueue))) {
492 if (dev_queue_xmit(skb))
493 ipoib_warn(priv, "dev_queue_xmit failed "
494 "to requeue packet\n");
498 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
500 struct ipoib_dev_priv *priv = netdev_priv(dev);
501 struct ipoib_path *path;
503 if (!priv->broadcast)
506 path = kzalloc(sizeof *path, GFP_ATOMIC);
512 skb_queue_head_init(&path->queue);
514 INIT_LIST_HEAD(&path->neigh_list);
516 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
517 path->pathrec.sgid = priv->local_gid;
518 path->pathrec.pkey = cpu_to_be16(priv->pkey);
519 path->pathrec.numb_path = 1;
520 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
525 static int path_rec_start(struct net_device *dev,
526 struct ipoib_path *path)
528 struct ipoib_dev_priv *priv = netdev_priv(dev);
530 ipoib_dbg(priv, "Start path record lookup for %p6\n",
531 path->pathrec.dgid.raw);
533 init_completion(&path->done);
536 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
538 IB_SA_PATH_REC_DGID |
539 IB_SA_PATH_REC_SGID |
540 IB_SA_PATH_REC_NUMB_PATH |
541 IB_SA_PATH_REC_TRAFFIC_CLASS |
546 if (path->query_id < 0) {
547 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
549 return path->query_id;
555 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
557 struct ipoib_dev_priv *priv = netdev_priv(dev);
558 struct ipoib_path *path;
559 struct ipoib_neigh *neigh;
562 neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
564 ++dev->stats.tx_dropped;
565 dev_kfree_skb_any(skb);
569 spin_lock_irqsave(&priv->lock, flags);
571 path = __path_find(dev, skb->dst->neighbour->ha + 4);
573 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
577 __path_add(dev, path);
580 list_add_tail(&neigh->list, &path->neigh_list);
583 kref_get(&path->ah->ref);
584 neigh->ah = path->ah;
585 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
586 sizeof(union ib_gid));
588 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
589 if (!ipoib_cm_get(neigh))
590 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
591 if (!ipoib_cm_get(neigh)) {
592 list_del(&neigh->list);
594 ipoib_put_ah(neigh->ah);
595 ipoib_neigh_free(dev, neigh);
598 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
599 __skb_queue_tail(&neigh->queue, skb);
601 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
602 skb_queue_len(&neigh->queue));
606 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
610 if (!path->query && path_rec_start(dev, path))
613 __skb_queue_tail(&neigh->queue, skb);
616 spin_unlock_irqrestore(&priv->lock, flags);
620 list_del(&neigh->list);
623 ipoib_neigh_free(dev, neigh);
625 ++dev->stats.tx_dropped;
626 dev_kfree_skb_any(skb);
628 spin_unlock_irqrestore(&priv->lock, flags);
631 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
633 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
635 /* Look up path record for unicasts */
636 if (skb->dst->neighbour->ha[4] != 0xff) {
637 neigh_add_path(skb, dev);
641 /* Add in the P_Key for multicasts */
642 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
643 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
644 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
647 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
648 struct ipoib_pseudoheader *phdr)
650 struct ipoib_dev_priv *priv = netdev_priv(dev);
651 struct ipoib_path *path;
654 spin_lock_irqsave(&priv->lock, flags);
656 path = __path_find(dev, phdr->hwaddr + 4);
657 if (!path || !path->valid) {
659 path = path_rec_create(dev, phdr->hwaddr + 4);
661 /* put pseudoheader back on for next time */
662 skb_push(skb, sizeof *phdr);
663 __skb_queue_tail(&path->queue, skb);
665 if (path_rec_start(dev, path)) {
666 spin_unlock_irqrestore(&priv->lock, flags);
667 path_free(dev, path);
670 __path_add(dev, path);
672 ++dev->stats.tx_dropped;
673 dev_kfree_skb_any(skb);
676 spin_unlock_irqrestore(&priv->lock, flags);
681 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
682 be16_to_cpu(path->pathrec.dlid));
684 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
685 } else if ((path->query || !path_rec_start(dev, path)) &&
686 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
687 /* put pseudoheader back on for next time */
688 skb_push(skb, sizeof *phdr);
689 __skb_queue_tail(&path->queue, skb);
691 ++dev->stats.tx_dropped;
692 dev_kfree_skb_any(skb);
695 spin_unlock_irqrestore(&priv->lock, flags);
698 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
700 struct ipoib_dev_priv *priv = netdev_priv(dev);
701 struct ipoib_neigh *neigh;
704 if (likely(skb->dst && skb->dst->neighbour)) {
705 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
706 ipoib_path_lookup(skb, dev);
710 neigh = *to_ipoib_neigh(skb->dst->neighbour);
713 if (unlikely((memcmp(&neigh->dgid.raw,
714 skb->dst->neighbour->ha + 4,
715 sizeof(union ib_gid))) ||
716 (neigh->dev != dev))) {
717 spin_lock_irqsave(&priv->lock, flags);
719 * It's safe to call ipoib_put_ah() inside
720 * priv->lock here, because we know that
721 * path->ah will always hold one more reference,
722 * so ipoib_put_ah() will never do more than
723 * decrement the ref count.
725 ipoib_put_ah(neigh->ah);
726 list_del(&neigh->list);
727 ipoib_neigh_free(dev, neigh);
728 spin_unlock_irqrestore(&priv->lock, flags);
729 ipoib_path_lookup(skb, dev);
733 if (ipoib_cm_get(neigh)) {
734 if (ipoib_cm_up(neigh)) {
735 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
738 } else if (neigh->ah) {
739 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
743 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
744 spin_lock_irqsave(&priv->lock, flags);
745 __skb_queue_tail(&neigh->queue, skb);
746 spin_unlock_irqrestore(&priv->lock, flags);
748 ++dev->stats.tx_dropped;
749 dev_kfree_skb_any(skb);
752 struct ipoib_pseudoheader *phdr =
753 (struct ipoib_pseudoheader *) skb->data;
754 skb_pull(skb, sizeof *phdr);
756 if (phdr->hwaddr[4] == 0xff) {
757 /* Add in the P_Key for multicast*/
758 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
759 phdr->hwaddr[9] = priv->pkey & 0xff;
761 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
763 /* unicast GID -- should be ARP or RARP reply */
765 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
766 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
767 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x %p6\n",
768 skb->dst ? "neigh" : "dst",
769 be16_to_cpup((__be16 *) skb->data),
770 IPOIB_QPN(phdr->hwaddr),
772 dev_kfree_skb_any(skb);
773 ++dev->stats.tx_dropped;
777 unicast_arp_send(skb, dev, phdr);
784 static void ipoib_timeout(struct net_device *dev)
786 struct ipoib_dev_priv *priv = netdev_priv(dev);
788 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
789 jiffies_to_msecs(jiffies - dev->trans_start));
790 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
791 netif_queue_stopped(dev),
792 priv->tx_head, priv->tx_tail);
793 /* XXX reset QP, etc. */
796 static int ipoib_hard_header(struct sk_buff *skb,
797 struct net_device *dev,
799 const void *daddr, const void *saddr, unsigned len)
801 struct ipoib_header *header;
803 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
805 header->proto = htons(type);
806 header->reserved = 0;
809 * If we don't have a neighbour structure, stuff the
810 * destination address onto the front of the skb so we can
811 * figure out where to send the packet later.
813 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
814 struct ipoib_pseudoheader *phdr =
815 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
816 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
822 static void ipoib_set_mcast_list(struct net_device *dev)
824 struct ipoib_dev_priv *priv = netdev_priv(dev);
826 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
827 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
831 queue_work(ipoib_workqueue, &priv->restart_task);
834 static void ipoib_neigh_cleanup(struct neighbour *n)
836 struct ipoib_neigh *neigh;
837 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
839 struct ipoib_ah *ah = NULL;
841 neigh = *to_ipoib_neigh(n);
843 priv = netdev_priv(neigh->dev);
847 "neigh_cleanup for %06x %p6\n",
851 spin_lock_irqsave(&priv->lock, flags);
855 list_del(&neigh->list);
856 ipoib_neigh_free(n->dev, neigh);
858 spin_unlock_irqrestore(&priv->lock, flags);
864 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
865 struct net_device *dev)
867 struct ipoib_neigh *neigh;
869 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
873 neigh->neighbour = neighbour;
875 *to_ipoib_neigh(neighbour) = neigh;
876 skb_queue_head_init(&neigh->queue);
877 ipoib_cm_set(neigh, NULL);
882 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
885 *to_ipoib_neigh(neigh->neighbour) = NULL;
886 while ((skb = __skb_dequeue(&neigh->queue))) {
887 ++dev->stats.tx_dropped;
888 dev_kfree_skb_any(skb);
890 if (ipoib_cm_get(neigh))
891 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
895 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
897 parms->neigh_cleanup = ipoib_neigh_cleanup;
902 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
904 struct ipoib_dev_priv *priv = netdev_priv(dev);
906 /* Allocate RX/TX "rings" to hold queued skbs */
907 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
909 if (!priv->rx_ring) {
910 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
911 ca->name, ipoib_recvq_size);
915 priv->tx_ring = vmalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
916 if (!priv->tx_ring) {
917 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
918 ca->name, ipoib_sendq_size);
919 goto out_rx_ring_cleanup;
921 memset(priv->tx_ring, 0, ipoib_sendq_size * sizeof *priv->tx_ring);
923 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
925 if (ipoib_ib_dev_init(dev, ca, port))
926 goto out_tx_ring_cleanup;
931 vfree(priv->tx_ring);
934 kfree(priv->rx_ring);
940 void ipoib_dev_cleanup(struct net_device *dev)
942 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
944 ipoib_delete_debug_files(dev);
946 /* Delete any child interfaces first */
947 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
948 unregister_netdev(cpriv->dev);
949 ipoib_dev_cleanup(cpriv->dev);
950 free_netdev(cpriv->dev);
953 ipoib_ib_dev_cleanup(dev);
955 kfree(priv->rx_ring);
956 vfree(priv->tx_ring);
958 priv->rx_ring = NULL;
959 priv->tx_ring = NULL;
962 static const struct header_ops ipoib_header_ops = {
963 .create = ipoib_hard_header,
966 static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
967 void **tcph, u64 *hdr_flags, void *priv)
972 if (unlikely(skb->protocol != htons(ETH_P_IP)))
976 * In the future we may add an else clause that verifies the
977 * checksum and allows devices which do not calculate checksum
980 if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
983 /* Check for non-TCP packet */
984 skb_reset_network_header(skb);
986 if (iph->protocol != IPPROTO_TCP)
989 ip_len = ip_hdrlen(skb);
990 skb_set_transport_header(skb, ip_len);
991 *tcph = tcp_hdr(skb);
993 /* check if IP header and TCP header are complete */
994 if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
997 *hdr_flags = LRO_IPV4 | LRO_TCP;
1003 static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
1005 priv->lro.lro_mgr.max_aggr = lro_max_aggr;
1006 priv->lro.lro_mgr.max_desc = IPOIB_MAX_LRO_DESCRIPTORS;
1007 priv->lro.lro_mgr.lro_arr = priv->lro.lro_desc;
1008 priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
1009 priv->lro.lro_mgr.features = LRO_F_NAPI;
1010 priv->lro.lro_mgr.dev = priv->dev;
1011 priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
1014 static void ipoib_setup(struct net_device *dev)
1016 struct ipoib_dev_priv *priv = netdev_priv(dev);
1018 dev->open = ipoib_open;
1019 dev->stop = ipoib_stop;
1020 dev->change_mtu = ipoib_change_mtu;
1021 dev->hard_start_xmit = ipoib_start_xmit;
1022 dev->tx_timeout = ipoib_timeout;
1023 dev->header_ops = &ipoib_header_ops;
1024 dev->set_multicast_list = ipoib_set_mcast_list;
1025 dev->neigh_setup = ipoib_neigh_setup_dev;
1027 ipoib_set_ethtool_ops(dev);
1029 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1031 dev->watchdog_timeo = HZ;
1033 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1036 * We add in INFINIBAND_ALEN to allow for the destination
1037 * address "pseudoheader" for skbs without neighbour struct.
1039 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
1040 dev->addr_len = INFINIBAND_ALEN;
1041 dev->type = ARPHRD_INFINIBAND;
1042 dev->tx_queue_len = ipoib_sendq_size * 2;
1043 dev->features = (NETIF_F_VLAN_CHALLENGED |
1046 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1048 netif_carrier_off(dev);
1052 ipoib_lro_setup(priv);
1054 spin_lock_init(&priv->lock);
1056 mutex_init(&priv->vlan_mutex);
1058 INIT_LIST_HEAD(&priv->path_list);
1059 INIT_LIST_HEAD(&priv->child_intfs);
1060 INIT_LIST_HEAD(&priv->dead_ahs);
1061 INIT_LIST_HEAD(&priv->multicast_list);
1063 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1064 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1065 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1066 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1067 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1068 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1069 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1070 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1073 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1075 struct net_device *dev;
1077 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1082 return netdev_priv(dev);
1085 static ssize_t show_pkey(struct device *dev,
1086 struct device_attribute *attr, char *buf)
1088 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1090 return sprintf(buf, "0x%04x\n", priv->pkey);
1092 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1094 static ssize_t show_umcast(struct device *dev,
1095 struct device_attribute *attr, char *buf)
1097 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1099 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1102 static ssize_t set_umcast(struct device *dev,
1103 struct device_attribute *attr,
1104 const char *buf, size_t count)
1106 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1107 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1109 if (umcast_val > 0) {
1110 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1111 ipoib_warn(priv, "ignoring multicast groups joined directly "
1114 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1118 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1120 int ipoib_add_umcast_attr(struct net_device *dev)
1122 return device_create_file(&dev->dev, &dev_attr_umcast);
1125 static ssize_t create_child(struct device *dev,
1126 struct device_attribute *attr,
1127 const char *buf, size_t count)
1132 if (sscanf(buf, "%i", &pkey) != 1)
1135 if (pkey < 0 || pkey > 0xffff)
1139 * Set the full membership bit, so that we join the right
1140 * broadcast group, etc.
1144 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1146 return ret ? ret : count;
1148 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1150 static ssize_t delete_child(struct device *dev,
1151 struct device_attribute *attr,
1152 const char *buf, size_t count)
1157 if (sscanf(buf, "%i", &pkey) != 1)
1160 if (pkey < 0 || pkey > 0xffff)
1163 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1165 return ret ? ret : count;
1168 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1170 int ipoib_add_pkey_attr(struct net_device *dev)
1172 return device_create_file(&dev->dev, &dev_attr_pkey);
1175 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1177 struct ib_device_attr *device_attr;
1178 int result = -ENOMEM;
1180 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1182 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1183 hca->name, sizeof *device_attr);
1187 result = ib_query_device(hca, device_attr);
1189 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1194 priv->hca_caps = device_attr->device_cap_flags;
1198 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1199 set_bit(IPOIB_FLAG_CSUM, &priv->flags);
1200 priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
1204 priv->dev->features |= NETIF_F_LRO;
1206 if (priv->dev->features & NETIF_F_SG && priv->hca_caps & IB_DEVICE_UD_TSO)
1207 priv->dev->features |= NETIF_F_TSO;
1213 static struct net_device *ipoib_add_port(const char *format,
1214 struct ib_device *hca, u8 port)
1216 struct ipoib_dev_priv *priv;
1217 struct ib_port_attr attr;
1218 int result = -ENOMEM;
1220 priv = ipoib_intf_alloc(format);
1222 goto alloc_mem_failed;
1224 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1226 if (!ib_query_port(hca, port, &attr))
1227 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1229 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1231 goto device_init_failed;
1234 /* MTU will be reset when mcast join happens */
1235 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1236 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1238 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1240 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1241 hca->name, port, result);
1242 goto device_init_failed;
1245 if (ipoib_set_dev_features(priv, hca))
1246 goto device_init_failed;
1249 * Set the full membership bit, so that we join the right
1250 * broadcast group, etc.
1252 priv->pkey |= 0x8000;
1254 priv->dev->broadcast[8] = priv->pkey >> 8;
1255 priv->dev->broadcast[9] = priv->pkey & 0xff;
1257 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1259 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1260 hca->name, port, result);
1261 goto device_init_failed;
1263 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1265 result = ipoib_dev_init(priv->dev, hca, port);
1267 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1268 hca->name, port, result);
1269 goto device_init_failed;
1272 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1273 priv->ca, ipoib_event);
1274 result = ib_register_event_handler(&priv->event_handler);
1276 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1277 "port %d (ret = %d)\n",
1278 hca->name, port, result);
1282 result = register_netdev(priv->dev);
1284 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1285 hca->name, port, result);
1286 goto register_failed;
1289 ipoib_create_debug_files(priv->dev);
1291 if (ipoib_cm_add_mode_attr(priv->dev))
1293 if (ipoib_add_pkey_attr(priv->dev))
1295 if (ipoib_add_umcast_attr(priv->dev))
1297 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1299 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1305 ipoib_delete_debug_files(priv->dev);
1306 unregister_netdev(priv->dev);
1309 ib_unregister_event_handler(&priv->event_handler);
1310 flush_workqueue(ipoib_workqueue);
1313 ipoib_dev_cleanup(priv->dev);
1316 free_netdev(priv->dev);
1319 return ERR_PTR(result);
1322 static void ipoib_add_one(struct ib_device *device)
1324 struct list_head *dev_list;
1325 struct net_device *dev;
1326 struct ipoib_dev_priv *priv;
1329 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1332 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1336 INIT_LIST_HEAD(dev_list);
1338 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1343 e = device->phys_port_cnt;
1346 for (p = s; p <= e; ++p) {
1347 dev = ipoib_add_port("ib%d", device, p);
1349 priv = netdev_priv(dev);
1350 list_add_tail(&priv->list, dev_list);
1354 ib_set_client_data(device, &ipoib_client, dev_list);
1357 static void ipoib_remove_one(struct ib_device *device)
1359 struct ipoib_dev_priv *priv, *tmp;
1360 struct list_head *dev_list;
1362 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1365 dev_list = ib_get_client_data(device, &ipoib_client);
1367 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1368 ib_unregister_event_handler(&priv->event_handler);
1371 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1374 flush_workqueue(ipoib_workqueue);
1376 unregister_netdev(priv->dev);
1377 ipoib_dev_cleanup(priv->dev);
1378 free_netdev(priv->dev);
1384 static int __init ipoib_init_module(void)
1388 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1389 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1390 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1392 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1393 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1394 ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
1395 IPOIB_MIN_QUEUE_SIZE));
1396 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1397 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1401 * When copying small received packets, we only copy from the
1402 * linear data part of the SKB, so we rely on this condition.
1404 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1406 ret = ipoib_register_debugfs();
1411 * We create our own workqueue mainly because we want to be
1412 * able to flush it when devices are being removed. We can't
1413 * use schedule_work()/flush_scheduled_work() because both
1414 * unregister_netdev() and linkwatch_event take the rtnl lock,
1415 * so flush_scheduled_work() can deadlock during device
1418 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1419 if (!ipoib_workqueue) {
1424 ib_sa_register_client(&ipoib_sa_client);
1426 ret = ib_register_client(&ipoib_client);
1433 ib_sa_unregister_client(&ipoib_sa_client);
1434 destroy_workqueue(ipoib_workqueue);
1437 ipoib_unregister_debugfs();
1442 static void __exit ipoib_cleanup_module(void)
1444 ib_unregister_client(&ipoib_client);
1445 ib_sa_unregister_client(&ipoib_sa_client);
1446 ipoib_unregister_debugfs();
1447 destroy_workqueue(ipoib_workqueue);
1450 module_init(ipoib_init_module);
1451 module_exit(ipoib_cleanup_module);