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;
676 * Check if our queue is stopped. Since we have the LLTX bit
677 * set, we can't rely on netif_stop_queue() preventing our
678 * xmit function from being called with a full queue.
680 if (unlikely(netif_queue_stopped(dev))) {
681 spin_unlock_irqrestore(&priv->tx_lock, flags);
682 return NETDEV_TX_BUSY;
685 if (likely(skb->dst && skb->dst->neighbour)) {
686 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
687 ipoib_path_lookup(skb, dev);
691 neigh = *to_ipoib_neigh(skb->dst->neighbour);
693 if (ipoib_cm_get(neigh)) {
694 if (ipoib_cm_up(neigh)) {
695 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
698 } else if (neigh->ah) {
699 if (unlikely((memcmp(&neigh->dgid.raw,
700 skb->dst->neighbour->ha + 4,
701 sizeof(union ib_gid))) ||
702 (neigh->dev != dev))) {
703 spin_lock(&priv->lock);
705 * It's safe to call ipoib_put_ah() inside
706 * priv->lock here, because we know that
707 * path->ah will always hold one more reference,
708 * so ipoib_put_ah() will never do more than
709 * decrement the ref count.
711 ipoib_put_ah(neigh->ah);
712 list_del(&neigh->list);
713 ipoib_neigh_free(dev, neigh);
714 spin_unlock(&priv->lock);
715 ipoib_path_lookup(skb, dev);
719 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
723 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
724 spin_lock(&priv->lock);
725 __skb_queue_tail(&neigh->queue, skb);
726 spin_unlock(&priv->lock);
728 ++dev->stats.tx_dropped;
729 dev_kfree_skb_any(skb);
732 struct ipoib_pseudoheader *phdr =
733 (struct ipoib_pseudoheader *) skb->data;
734 skb_pull(skb, sizeof *phdr);
736 if (phdr->hwaddr[4] == 0xff) {
737 /* Add in the P_Key for multicast*/
738 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
739 phdr->hwaddr[9] = priv->pkey & 0xff;
741 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
743 /* unicast GID -- should be ARP or RARP reply */
745 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
746 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
747 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
749 skb->dst ? "neigh" : "dst",
750 be16_to_cpup((__be16 *) skb->data),
751 IPOIB_QPN(phdr->hwaddr),
752 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
753 dev_kfree_skb_any(skb);
754 ++dev->stats.tx_dropped;
758 unicast_arp_send(skb, dev, phdr);
763 spin_unlock_irqrestore(&priv->tx_lock, flags);
768 static void ipoib_timeout(struct net_device *dev)
770 struct ipoib_dev_priv *priv = netdev_priv(dev);
772 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
773 jiffies_to_msecs(jiffies - dev->trans_start));
774 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
775 netif_queue_stopped(dev),
776 priv->tx_head, priv->tx_tail);
777 /* XXX reset QP, etc. */
780 static int ipoib_hard_header(struct sk_buff *skb,
781 struct net_device *dev,
783 const void *daddr, const void *saddr, unsigned len)
785 struct ipoib_header *header;
787 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
789 header->proto = htons(type);
790 header->reserved = 0;
793 * If we don't have a neighbour structure, stuff the
794 * destination address onto the front of the skb so we can
795 * figure out where to send the packet later.
797 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
798 struct ipoib_pseudoheader *phdr =
799 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
800 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
806 static void ipoib_set_mcast_list(struct net_device *dev)
808 struct ipoib_dev_priv *priv = netdev_priv(dev);
810 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
811 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
815 queue_work(ipoib_workqueue, &priv->restart_task);
818 static void ipoib_neigh_cleanup(struct neighbour *n)
820 struct ipoib_neigh *neigh;
821 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
823 struct ipoib_ah *ah = NULL;
825 neigh = *to_ipoib_neigh(n);
827 priv = netdev_priv(neigh->dev);
828 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
833 "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
835 IPOIB_GID_RAW_ARG(n->ha + 4));
837 spin_lock_irqsave(&priv->lock, flags);
841 list_del(&neigh->list);
842 ipoib_neigh_free(n->dev, neigh);
844 spin_unlock_irqrestore(&priv->lock, flags);
850 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
851 struct net_device *dev)
853 struct ipoib_neigh *neigh;
855 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
859 neigh->neighbour = neighbour;
861 *to_ipoib_neigh(neighbour) = neigh;
862 skb_queue_head_init(&neigh->queue);
863 ipoib_cm_set(neigh, NULL);
868 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
871 *to_ipoib_neigh(neigh->neighbour) = NULL;
872 while ((skb = __skb_dequeue(&neigh->queue))) {
873 ++dev->stats.tx_dropped;
874 dev_kfree_skb_any(skb);
876 if (ipoib_cm_get(neigh))
877 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
881 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
883 parms->neigh_cleanup = ipoib_neigh_cleanup;
888 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
890 struct ipoib_dev_priv *priv = netdev_priv(dev);
892 /* Allocate RX/TX "rings" to hold queued skbs */
893 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
895 if (!priv->rx_ring) {
896 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
897 ca->name, ipoib_recvq_size);
901 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
903 if (!priv->tx_ring) {
904 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
905 ca->name, ipoib_sendq_size);
906 goto out_rx_ring_cleanup;
909 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
911 if (ipoib_ib_dev_init(dev, ca, port))
912 goto out_tx_ring_cleanup;
917 kfree(priv->tx_ring);
920 kfree(priv->rx_ring);
926 void ipoib_dev_cleanup(struct net_device *dev)
928 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
930 ipoib_delete_debug_files(dev);
932 /* Delete any child interfaces first */
933 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
934 unregister_netdev(cpriv->dev);
935 ipoib_dev_cleanup(cpriv->dev);
936 free_netdev(cpriv->dev);
939 ipoib_ib_dev_cleanup(dev);
941 kfree(priv->rx_ring);
942 kfree(priv->tx_ring);
944 priv->rx_ring = NULL;
945 priv->tx_ring = NULL;
948 static const struct header_ops ipoib_header_ops = {
949 .create = ipoib_hard_header,
952 static void ipoib_setup(struct net_device *dev)
954 struct ipoib_dev_priv *priv = netdev_priv(dev);
956 dev->open = ipoib_open;
957 dev->stop = ipoib_stop;
958 dev->change_mtu = ipoib_change_mtu;
959 dev->hard_start_xmit = ipoib_start_xmit;
960 dev->tx_timeout = ipoib_timeout;
961 dev->header_ops = &ipoib_header_ops;
962 dev->set_multicast_list = ipoib_set_mcast_list;
963 dev->neigh_setup = ipoib_neigh_setup_dev;
965 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
967 dev->watchdog_timeo = HZ;
969 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
972 * We add in INFINIBAND_ALEN to allow for the destination
973 * address "pseudoheader" for skbs without neighbour struct.
975 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
976 dev->addr_len = INFINIBAND_ALEN;
977 dev->type = ARPHRD_INFINIBAND;
978 dev->tx_queue_len = ipoib_sendq_size * 2;
979 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
981 /* MTU will be reset when mcast join happens */
982 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
983 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
985 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
987 netif_carrier_off(dev);
991 spin_lock_init(&priv->lock);
992 spin_lock_init(&priv->tx_lock);
994 mutex_init(&priv->mcast_mutex);
995 mutex_init(&priv->vlan_mutex);
997 INIT_LIST_HEAD(&priv->path_list);
998 INIT_LIST_HEAD(&priv->child_intfs);
999 INIT_LIST_HEAD(&priv->dead_ahs);
1000 INIT_LIST_HEAD(&priv->multicast_list);
1002 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1003 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
1004 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1005 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
1006 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1007 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1010 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1012 struct net_device *dev;
1014 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1019 return netdev_priv(dev);
1022 static ssize_t show_pkey(struct device *dev,
1023 struct device_attribute *attr, char *buf)
1025 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1027 return sprintf(buf, "0x%04x\n", priv->pkey);
1029 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1031 static ssize_t show_umcast(struct device *dev,
1032 struct device_attribute *attr, char *buf)
1034 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1036 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1039 static ssize_t set_umcast(struct device *dev,
1040 struct device_attribute *attr,
1041 const char *buf, size_t count)
1043 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1044 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1046 if (umcast_val > 0) {
1047 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1048 ipoib_warn(priv, "ignoring multicast groups joined directly "
1051 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1055 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1057 int ipoib_add_umcast_attr(struct net_device *dev)
1059 return device_create_file(&dev->dev, &dev_attr_umcast);
1062 static ssize_t create_child(struct device *dev,
1063 struct device_attribute *attr,
1064 const char *buf, size_t count)
1069 if (sscanf(buf, "%i", &pkey) != 1)
1072 if (pkey < 0 || pkey > 0xffff)
1076 * Set the full membership bit, so that we join the right
1077 * broadcast group, etc.
1081 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1083 return ret ? ret : count;
1085 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1087 static ssize_t delete_child(struct device *dev,
1088 struct device_attribute *attr,
1089 const char *buf, size_t count)
1094 if (sscanf(buf, "%i", &pkey) != 1)
1097 if (pkey < 0 || pkey > 0xffff)
1100 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1102 return ret ? ret : count;
1105 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1107 int ipoib_add_pkey_attr(struct net_device *dev)
1109 return device_create_file(&dev->dev, &dev_attr_pkey);
1112 static struct net_device *ipoib_add_port(const char *format,
1113 struct ib_device *hca, u8 port)
1115 struct ipoib_dev_priv *priv;
1116 int result = -ENOMEM;
1118 priv = ipoib_intf_alloc(format);
1120 goto alloc_mem_failed;
1122 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1124 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1126 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1127 hca->name, port, result);
1128 goto device_init_failed;
1132 * Set the full membership bit, so that we join the right
1133 * broadcast group, etc.
1135 priv->pkey |= 0x8000;
1137 priv->dev->broadcast[8] = priv->pkey >> 8;
1138 priv->dev->broadcast[9] = priv->pkey & 0xff;
1140 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1142 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1143 hca->name, port, result);
1144 goto device_init_failed;
1146 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1149 result = ipoib_dev_init(priv->dev, hca, port);
1151 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1152 hca->name, port, result);
1153 goto device_init_failed;
1156 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1157 priv->ca, ipoib_event);
1158 result = ib_register_event_handler(&priv->event_handler);
1160 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1161 "port %d (ret = %d)\n",
1162 hca->name, port, result);
1166 result = register_netdev(priv->dev);
1168 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1169 hca->name, port, result);
1170 goto register_failed;
1173 ipoib_create_debug_files(priv->dev);
1175 if (ipoib_cm_add_mode_attr(priv->dev))
1177 if (ipoib_add_pkey_attr(priv->dev))
1179 if (ipoib_add_umcast_attr(priv->dev))
1181 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1183 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1189 ipoib_delete_debug_files(priv->dev);
1190 unregister_netdev(priv->dev);
1193 ib_unregister_event_handler(&priv->event_handler);
1194 flush_scheduled_work();
1197 ipoib_dev_cleanup(priv->dev);
1200 free_netdev(priv->dev);
1203 return ERR_PTR(result);
1206 static void ipoib_add_one(struct ib_device *device)
1208 struct list_head *dev_list;
1209 struct net_device *dev;
1210 struct ipoib_dev_priv *priv;
1213 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1216 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1220 INIT_LIST_HEAD(dev_list);
1222 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1227 e = device->phys_port_cnt;
1230 for (p = s; p <= e; ++p) {
1231 dev = ipoib_add_port("ib%d", device, p);
1233 priv = netdev_priv(dev);
1234 list_add_tail(&priv->list, dev_list);
1238 ib_set_client_data(device, &ipoib_client, dev_list);
1241 static void ipoib_remove_one(struct ib_device *device)
1243 struct ipoib_dev_priv *priv, *tmp;
1244 struct list_head *dev_list;
1246 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1249 dev_list = ib_get_client_data(device, &ipoib_client);
1251 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1252 ib_unregister_event_handler(&priv->event_handler);
1253 flush_scheduled_work();
1255 unregister_netdev(priv->dev);
1256 ipoib_dev_cleanup(priv->dev);
1257 free_netdev(priv->dev);
1263 static int __init ipoib_init_module(void)
1267 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1268 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1269 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1271 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1272 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1273 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1274 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1275 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1278 ret = ipoib_register_debugfs();
1283 * We create our own workqueue mainly because we want to be
1284 * able to flush it when devices are being removed. We can't
1285 * use schedule_work()/flush_scheduled_work() because both
1286 * unregister_netdev() and linkwatch_event take the rtnl lock,
1287 * so flush_scheduled_work() can deadlock during device
1290 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1291 if (!ipoib_workqueue) {
1296 ib_sa_register_client(&ipoib_sa_client);
1298 ret = ib_register_client(&ipoib_client);
1305 ib_sa_unregister_client(&ipoib_sa_client);
1306 destroy_workqueue(ipoib_workqueue);
1309 ipoib_unregister_debugfs();
1314 static void __exit ipoib_cleanup_module(void)
1316 ib_unregister_client(&ipoib_client);
1317 ib_sa_unregister_client(&ipoib_sa_client);
1318 ipoib_unregister_debugfs();
1319 destroy_workqueue(ipoib_workqueue);
1322 module_init(ipoib_init_module);
1323 module_exit(ipoib_cleanup_module);