2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
39 #include <linux/module.h>
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/kernel.h>
45 #include <linux/if_arp.h> /* For ARPHRD_xxx */
52 MODULE_AUTHOR("Roland Dreier");
53 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
54 MODULE_LICENSE("Dual BSD/GPL");
56 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
57 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
59 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
60 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
61 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
62 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
64 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
65 int ipoib_debug_level;
67 module_param_named(debug_level, ipoib_debug_level, int, 0644);
68 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
71 struct ipoib_path_iter {
72 struct net_device *dev;
73 struct ipoib_path path;
76 static const u8 ipv4_bcast_addr[] = {
77 0x00, 0xff, 0xff, 0xff,
78 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
79 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
82 struct workqueue_struct *ipoib_workqueue;
84 struct ib_sa_client ipoib_sa_client;
86 static void ipoib_add_one(struct ib_device *device);
87 static void ipoib_remove_one(struct ib_device *device);
89 static struct ib_client ipoib_client = {
92 .remove = ipoib_remove_one
95 int ipoib_open(struct net_device *dev)
97 struct ipoib_dev_priv *priv = netdev_priv(dev);
99 ipoib_dbg(priv, "bringing up interface\n");
101 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
103 if (ipoib_pkey_dev_delay_open(dev))
106 if (ipoib_ib_dev_open(dev))
109 if (ipoib_ib_dev_up(dev)) {
110 ipoib_ib_dev_stop(dev, 1);
114 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
115 struct ipoib_dev_priv *cpriv;
117 /* Bring up any child interfaces too */
118 mutex_lock(&priv->vlan_mutex);
119 list_for_each_entry(cpriv, &priv->child_intfs, list) {
122 flags = cpriv->dev->flags;
126 dev_change_flags(cpriv->dev, flags | IFF_UP);
128 mutex_unlock(&priv->vlan_mutex);
131 netif_start_queue(dev);
136 static int ipoib_stop(struct net_device *dev)
138 struct ipoib_dev_priv *priv = netdev_priv(dev);
140 ipoib_dbg(priv, "stopping interface\n");
142 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
144 netif_stop_queue(dev);
146 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
149 * Now flush workqueue to make sure a scheduled task doesn't
150 * bring our internal state back up.
152 flush_workqueue(ipoib_workqueue);
154 ipoib_ib_dev_down(dev, 1);
155 ipoib_ib_dev_stop(dev, 1);
157 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
158 struct ipoib_dev_priv *cpriv;
160 /* Bring down any child interfaces too */
161 mutex_lock(&priv->vlan_mutex);
162 list_for_each_entry(cpriv, &priv->child_intfs, list) {
165 flags = cpriv->dev->flags;
166 if (!(flags & IFF_UP))
169 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
171 mutex_unlock(&priv->vlan_mutex);
177 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
179 struct ipoib_dev_priv *priv = netdev_priv(dev);
181 /* dev->mtu > 2K ==> connected mode */
182 if (ipoib_cm_admin_enabled(dev) && new_mtu <= IPOIB_CM_MTU) {
183 if (new_mtu > priv->mcast_mtu)
184 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
190 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
194 priv->admin_mtu = new_mtu;
196 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
201 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
203 struct ipoib_dev_priv *priv = netdev_priv(dev);
204 struct rb_node *n = priv->path_tree.rb_node;
205 struct ipoib_path *path;
209 path = rb_entry(n, struct ipoib_path, rb_node);
211 ret = memcmp(gid, path->pathrec.dgid.raw,
212 sizeof (union ib_gid));
225 static int __path_add(struct net_device *dev, struct ipoib_path *path)
227 struct ipoib_dev_priv *priv = netdev_priv(dev);
228 struct rb_node **n = &priv->path_tree.rb_node;
229 struct rb_node *pn = NULL;
230 struct ipoib_path *tpath;
235 tpath = rb_entry(pn, struct ipoib_path, rb_node);
237 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
238 sizeof (union ib_gid));
247 rb_link_node(&path->rb_node, pn, n);
248 rb_insert_color(&path->rb_node, &priv->path_tree);
250 list_add_tail(&path->list, &priv->path_list);
255 static void path_free(struct net_device *dev, struct ipoib_path *path)
257 struct ipoib_dev_priv *priv = netdev_priv(dev);
258 struct ipoib_neigh *neigh, *tn;
262 while ((skb = __skb_dequeue(&path->queue)))
263 dev_kfree_skb_irq(skb);
265 spin_lock_irqsave(&priv->lock, flags);
267 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
269 * It's safe to call ipoib_put_ah() inside priv->lock
270 * here, because we know that path->ah will always
271 * hold one more reference, so ipoib_put_ah() will
272 * never do more than decrement the ref count.
275 ipoib_put_ah(neigh->ah);
277 ipoib_neigh_free(dev, neigh);
280 spin_unlock_irqrestore(&priv->lock, flags);
283 ipoib_put_ah(path->ah);
288 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
290 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
292 struct ipoib_path_iter *iter;
294 iter = kmalloc(sizeof *iter, GFP_KERNEL);
299 memset(iter->path.pathrec.dgid.raw, 0, 16);
301 if (ipoib_path_iter_next(iter)) {
309 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
311 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
313 struct ipoib_path *path;
316 spin_lock_irq(&priv->lock);
318 n = rb_first(&priv->path_tree);
321 path = rb_entry(n, struct ipoib_path, rb_node);
323 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
324 sizeof (union ib_gid)) < 0) {
333 spin_unlock_irq(&priv->lock);
338 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
339 struct ipoib_path *path)
344 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
346 void ipoib_flush_paths(struct net_device *dev)
348 struct ipoib_dev_priv *priv = netdev_priv(dev);
349 struct ipoib_path *path, *tp;
350 LIST_HEAD(remove_list);
352 spin_lock_irq(&priv->tx_lock);
353 spin_lock(&priv->lock);
355 list_splice(&priv->path_list, &remove_list);
356 INIT_LIST_HEAD(&priv->path_list);
358 list_for_each_entry(path, &remove_list, list)
359 rb_erase(&path->rb_node, &priv->path_tree);
361 list_for_each_entry_safe(path, tp, &remove_list, list) {
363 ib_sa_cancel_query(path->query_id, path->query);
364 spin_unlock(&priv->lock);
365 spin_unlock_irq(&priv->tx_lock);
366 wait_for_completion(&path->done);
367 path_free(dev, path);
368 spin_lock_irq(&priv->tx_lock);
369 spin_lock(&priv->lock);
371 spin_unlock(&priv->lock);
372 spin_unlock_irq(&priv->tx_lock);
375 static void path_rec_completion(int status,
376 struct ib_sa_path_rec *pathrec,
379 struct ipoib_path *path = path_ptr;
380 struct net_device *dev = path->dev;
381 struct ipoib_dev_priv *priv = netdev_priv(dev);
382 struct ipoib_ah *ah = NULL;
383 struct ipoib_neigh *neigh, *tn;
384 struct sk_buff_head skqueue;
389 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
390 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
392 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
393 status, IPOIB_GID_ARG(path->pathrec.dgid));
395 skb_queue_head_init(&skqueue);
398 struct ib_ah_attr av;
400 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
401 ah = ipoib_create_ah(dev, priv->pd, &av);
404 spin_lock_irqsave(&priv->lock, flags);
409 path->pathrec = *pathrec;
411 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
412 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
414 while ((skb = __skb_dequeue(&path->queue)))
415 __skb_queue_tail(&skqueue, skb);
417 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
418 kref_get(&path->ah->ref);
419 neigh->ah = path->ah;
420 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
421 sizeof(union ib_gid));
423 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
424 if (!ipoib_cm_get(neigh))
425 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
428 if (!ipoib_cm_get(neigh)) {
429 list_del(&neigh->list);
431 ipoib_put_ah(neigh->ah);
432 ipoib_neigh_free(dev, neigh);
437 while ((skb = __skb_dequeue(&neigh->queue)))
438 __skb_queue_tail(&skqueue, skb);
443 complete(&path->done);
445 spin_unlock_irqrestore(&priv->lock, flags);
447 while ((skb = __skb_dequeue(&skqueue))) {
449 if (dev_queue_xmit(skb))
450 ipoib_warn(priv, "dev_queue_xmit failed "
451 "to requeue packet\n");
455 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
457 struct ipoib_dev_priv *priv = netdev_priv(dev);
458 struct ipoib_path *path;
460 path = kzalloc(sizeof *path, GFP_ATOMIC);
466 skb_queue_head_init(&path->queue);
468 INIT_LIST_HEAD(&path->neigh_list);
470 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
471 path->pathrec.sgid = priv->local_gid;
472 path->pathrec.pkey = cpu_to_be16(priv->pkey);
473 path->pathrec.numb_path = 1;
474 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
479 static int path_rec_start(struct net_device *dev,
480 struct ipoib_path *path)
482 struct ipoib_dev_priv *priv = netdev_priv(dev);
484 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
485 IPOIB_GID_ARG(path->pathrec.dgid));
487 init_completion(&path->done);
490 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
492 IB_SA_PATH_REC_DGID |
493 IB_SA_PATH_REC_SGID |
494 IB_SA_PATH_REC_NUMB_PATH |
495 IB_SA_PATH_REC_TRAFFIC_CLASS |
500 if (path->query_id < 0) {
501 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
503 return path->query_id;
509 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
511 struct ipoib_dev_priv *priv = netdev_priv(dev);
512 struct ipoib_path *path;
513 struct ipoib_neigh *neigh;
515 neigh = ipoib_neigh_alloc(skb->dst->neighbour);
517 ++priv->stats.tx_dropped;
518 dev_kfree_skb_any(skb);
523 * We can only be called from ipoib_start_xmit, so we're
524 * inside tx_lock -- no need to save/restore flags.
526 spin_lock(&priv->lock);
528 path = __path_find(dev, skb->dst->neighbour->ha + 4);
530 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
534 __path_add(dev, path);
537 list_add_tail(&neigh->list, &path->neigh_list);
540 kref_get(&path->ah->ref);
541 neigh->ah = path->ah;
542 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
543 sizeof(union ib_gid));
545 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
546 if (!ipoib_cm_get(neigh))
547 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
548 if (!ipoib_cm_get(neigh)) {
549 list_del(&neigh->list);
551 ipoib_put_ah(neigh->ah);
552 ipoib_neigh_free(dev, neigh);
555 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
556 __skb_queue_tail(&neigh->queue, skb);
558 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
559 skb_queue_len(&neigh->queue));
563 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
567 if (!path->query && path_rec_start(dev, path))
570 __skb_queue_tail(&neigh->queue, skb);
573 spin_unlock(&priv->lock);
577 list_del(&neigh->list);
580 ipoib_neigh_free(dev, neigh);
582 ++priv->stats.tx_dropped;
583 dev_kfree_skb_any(skb);
585 spin_unlock(&priv->lock);
588 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
590 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
592 /* Look up path record for unicasts */
593 if (skb->dst->neighbour->ha[4] != 0xff) {
594 neigh_add_path(skb, dev);
598 /* Add in the P_Key for multicasts */
599 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
600 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
601 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
604 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
605 struct ipoib_pseudoheader *phdr)
607 struct ipoib_dev_priv *priv = netdev_priv(dev);
608 struct ipoib_path *path;
611 * We can only be called from ipoib_start_xmit, so we're
612 * inside tx_lock -- no need to save/restore flags.
614 spin_lock(&priv->lock);
616 path = __path_find(dev, phdr->hwaddr + 4);
618 path = path_rec_create(dev, phdr->hwaddr + 4);
620 /* put pseudoheader back on for next time */
621 skb_push(skb, sizeof *phdr);
622 __skb_queue_tail(&path->queue, skb);
624 if (path_rec_start(dev, path)) {
625 spin_unlock(&priv->lock);
626 path_free(dev, path);
629 __path_add(dev, path);
631 ++priv->stats.tx_dropped;
632 dev_kfree_skb_any(skb);
635 spin_unlock(&priv->lock);
640 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
641 be16_to_cpu(path->pathrec.dlid));
643 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
644 } else if ((path->query || !path_rec_start(dev, path)) &&
645 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
646 /* put pseudoheader back on for next time */
647 skb_push(skb, sizeof *phdr);
648 __skb_queue_tail(&path->queue, skb);
650 ++priv->stats.tx_dropped;
651 dev_kfree_skb_any(skb);
654 spin_unlock(&priv->lock);
657 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
659 struct ipoib_dev_priv *priv = netdev_priv(dev);
660 struct ipoib_neigh *neigh;
663 if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
664 return NETDEV_TX_LOCKED;
667 * Check if our queue is stopped. Since we have the LLTX bit
668 * set, we can't rely on netif_stop_queue() preventing our
669 * xmit function from being called with a full queue.
671 if (unlikely(netif_queue_stopped(dev))) {
672 spin_unlock_irqrestore(&priv->tx_lock, flags);
673 return NETDEV_TX_BUSY;
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);
684 if (ipoib_cm_get(neigh)) {
685 if (ipoib_cm_up(neigh)) {
686 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
689 } else if (neigh->ah) {
690 if (unlikely(memcmp(&neigh->dgid.raw,
691 skb->dst->neighbour->ha + 4,
692 sizeof(union ib_gid)))) {
693 spin_lock(&priv->lock);
695 * It's safe to call ipoib_put_ah() inside
696 * priv->lock here, because we know that
697 * path->ah will always hold one more reference,
698 * so ipoib_put_ah() will never do more than
699 * decrement the ref count.
701 ipoib_put_ah(neigh->ah);
702 list_del(&neigh->list);
703 ipoib_neigh_free(dev, neigh);
704 spin_unlock(&priv->lock);
705 ipoib_path_lookup(skb, dev);
709 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
713 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
714 spin_lock(&priv->lock);
715 __skb_queue_tail(&neigh->queue, skb);
716 spin_unlock(&priv->lock);
718 ++priv->stats.tx_dropped;
719 dev_kfree_skb_any(skb);
722 struct ipoib_pseudoheader *phdr =
723 (struct ipoib_pseudoheader *) skb->data;
724 skb_pull(skb, sizeof *phdr);
726 if (phdr->hwaddr[4] == 0xff) {
727 /* Add in the P_Key for multicast*/
728 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
729 phdr->hwaddr[9] = priv->pkey & 0xff;
731 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
733 /* unicast GID -- should be ARP or RARP reply */
735 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
736 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
737 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
739 skb->dst ? "neigh" : "dst",
740 be16_to_cpup((__be16 *) skb->data),
741 IPOIB_QPN(phdr->hwaddr),
742 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
743 dev_kfree_skb_any(skb);
744 ++priv->stats.tx_dropped;
748 unicast_arp_send(skb, dev, phdr);
753 spin_unlock_irqrestore(&priv->tx_lock, flags);
758 static struct net_device_stats *ipoib_get_stats(struct net_device *dev)
760 struct ipoib_dev_priv *priv = netdev_priv(dev);
765 static void ipoib_timeout(struct net_device *dev)
767 struct ipoib_dev_priv *priv = netdev_priv(dev);
769 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
770 jiffies_to_msecs(jiffies - dev->trans_start));
771 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
772 netif_queue_stopped(dev),
773 priv->tx_head, priv->tx_tail);
774 /* XXX reset QP, etc. */
777 static int ipoib_hard_header(struct sk_buff *skb,
778 struct net_device *dev,
780 void *daddr, void *saddr, unsigned len)
782 struct ipoib_header *header;
784 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
786 header->proto = htons(type);
787 header->reserved = 0;
790 * If we don't have a neighbour structure, stuff the
791 * destination address onto the front of the skb so we can
792 * figure out where to send the packet later.
794 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
795 struct ipoib_pseudoheader *phdr =
796 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
797 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
803 static void ipoib_set_mcast_list(struct net_device *dev)
805 struct ipoib_dev_priv *priv = netdev_priv(dev);
807 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
808 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
812 queue_work(ipoib_workqueue, &priv->restart_task);
815 static void ipoib_neigh_cleanup(struct neighbour *n)
817 struct ipoib_neigh *neigh;
818 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
820 struct ipoib_ah *ah = NULL;
823 "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
825 IPOIB_GID_RAW_ARG(n->ha + 4));
827 spin_lock_irqsave(&priv->lock, flags);
829 neigh = *to_ipoib_neigh(n);
833 list_del(&neigh->list);
834 ipoib_neigh_free(n->dev, neigh);
837 spin_unlock_irqrestore(&priv->lock, flags);
843 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour)
845 struct ipoib_neigh *neigh;
847 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
851 neigh->neighbour = neighbour;
852 *to_ipoib_neigh(neighbour) = neigh;
853 skb_queue_head_init(&neigh->queue);
854 ipoib_cm_set(neigh, NULL);
859 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
861 struct ipoib_dev_priv *priv = netdev_priv(dev);
863 *to_ipoib_neigh(neigh->neighbour) = NULL;
864 while ((skb = __skb_dequeue(&neigh->queue))) {
865 ++priv->stats.tx_dropped;
866 dev_kfree_skb_any(skb);
868 if (ipoib_cm_get(neigh))
869 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
873 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
875 parms->neigh_cleanup = ipoib_neigh_cleanup;
880 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
882 struct ipoib_dev_priv *priv = netdev_priv(dev);
884 /* Allocate RX/TX "rings" to hold queued skbs */
885 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
887 if (!priv->rx_ring) {
888 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
889 ca->name, ipoib_recvq_size);
893 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
895 if (!priv->tx_ring) {
896 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
897 ca->name, ipoib_sendq_size);
898 goto out_rx_ring_cleanup;
901 /* priv->tx_head & tx_tail are already 0 */
903 if (ipoib_ib_dev_init(dev, ca, port))
904 goto out_tx_ring_cleanup;
909 kfree(priv->tx_ring);
912 kfree(priv->rx_ring);
918 void ipoib_dev_cleanup(struct net_device *dev)
920 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
922 ipoib_delete_debug_files(dev);
924 /* Delete any child interfaces first */
925 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
926 unregister_netdev(cpriv->dev);
927 ipoib_dev_cleanup(cpriv->dev);
928 free_netdev(cpriv->dev);
931 ipoib_ib_dev_cleanup(dev);
933 kfree(priv->rx_ring);
934 kfree(priv->tx_ring);
936 priv->rx_ring = NULL;
937 priv->tx_ring = NULL;
940 static void ipoib_setup(struct net_device *dev)
942 struct ipoib_dev_priv *priv = netdev_priv(dev);
944 dev->open = ipoib_open;
945 dev->stop = ipoib_stop;
946 dev->change_mtu = ipoib_change_mtu;
947 dev->hard_start_xmit = ipoib_start_xmit;
948 dev->get_stats = ipoib_get_stats;
949 dev->tx_timeout = ipoib_timeout;
950 dev->hard_header = ipoib_hard_header;
951 dev->set_multicast_list = ipoib_set_mcast_list;
952 dev->neigh_setup = ipoib_neigh_setup_dev;
953 dev->poll = ipoib_poll;
956 dev->watchdog_timeo = HZ;
958 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
961 * We add in INFINIBAND_ALEN to allow for the destination
962 * address "pseudoheader" for skbs without neighbour struct.
964 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
965 dev->addr_len = INFINIBAND_ALEN;
966 dev->type = ARPHRD_INFINIBAND;
967 dev->tx_queue_len = ipoib_sendq_size * 2;
968 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
970 /* MTU will be reset when mcast join happens */
971 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
972 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
974 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
976 netif_carrier_off(dev);
978 SET_MODULE_OWNER(dev);
982 spin_lock_init(&priv->lock);
983 spin_lock_init(&priv->tx_lock);
985 mutex_init(&priv->mcast_mutex);
986 mutex_init(&priv->vlan_mutex);
988 INIT_LIST_HEAD(&priv->path_list);
989 INIT_LIST_HEAD(&priv->child_intfs);
990 INIT_LIST_HEAD(&priv->dead_ahs);
991 INIT_LIST_HEAD(&priv->multicast_list);
993 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
994 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
995 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
996 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
997 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
998 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1001 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1003 struct net_device *dev;
1005 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1010 return netdev_priv(dev);
1013 static ssize_t show_pkey(struct device *dev,
1014 struct device_attribute *attr, char *buf)
1016 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1018 return sprintf(buf, "0x%04x\n", priv->pkey);
1020 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1022 static ssize_t show_umcast(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, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1030 static ssize_t set_umcast(struct device *dev,
1031 struct device_attribute *attr,
1032 const char *buf, size_t count)
1034 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1035 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1037 if (umcast_val > 0) {
1038 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1039 ipoib_warn(priv, "ignoring multicast groups joined directly "
1042 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1046 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1048 int ipoib_add_umcast_attr(struct net_device *dev)
1050 return device_create_file(&dev->dev, &dev_attr_umcast);
1053 static ssize_t create_child(struct device *dev,
1054 struct device_attribute *attr,
1055 const char *buf, size_t count)
1060 if (sscanf(buf, "%i", &pkey) != 1)
1063 if (pkey < 0 || pkey > 0xffff)
1067 * Set the full membership bit, so that we join the right
1068 * broadcast group, etc.
1072 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1074 return ret ? ret : count;
1076 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1078 static ssize_t delete_child(struct device *dev,
1079 struct device_attribute *attr,
1080 const char *buf, size_t count)
1085 if (sscanf(buf, "%i", &pkey) != 1)
1088 if (pkey < 0 || pkey > 0xffff)
1091 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1093 return ret ? ret : count;
1096 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1098 int ipoib_add_pkey_attr(struct net_device *dev)
1100 return device_create_file(&dev->dev, &dev_attr_pkey);
1103 static struct net_device *ipoib_add_port(const char *format,
1104 struct ib_device *hca, u8 port)
1106 struct ipoib_dev_priv *priv;
1107 int result = -ENOMEM;
1109 priv = ipoib_intf_alloc(format);
1111 goto alloc_mem_failed;
1113 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1115 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1117 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1118 hca->name, port, result);
1119 goto device_init_failed;
1123 * Set the full membership bit, so that we join the right
1124 * broadcast group, etc.
1126 priv->pkey |= 0x8000;
1128 priv->dev->broadcast[8] = priv->pkey >> 8;
1129 priv->dev->broadcast[9] = priv->pkey & 0xff;
1131 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1133 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1134 hca->name, port, result);
1135 goto device_init_failed;
1137 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1140 result = ipoib_dev_init(priv->dev, hca, port);
1142 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1143 hca->name, port, result);
1144 goto device_init_failed;
1147 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1148 priv->ca, ipoib_event);
1149 result = ib_register_event_handler(&priv->event_handler);
1151 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1152 "port %d (ret = %d)\n",
1153 hca->name, port, result);
1157 result = register_netdev(priv->dev);
1159 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1160 hca->name, port, result);
1161 goto register_failed;
1164 ipoib_create_debug_files(priv->dev);
1166 if (ipoib_cm_add_mode_attr(priv->dev))
1168 if (ipoib_add_pkey_attr(priv->dev))
1170 if (ipoib_add_umcast_attr(priv->dev))
1172 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1174 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1180 ipoib_delete_debug_files(priv->dev);
1181 unregister_netdev(priv->dev);
1184 ib_unregister_event_handler(&priv->event_handler);
1185 flush_scheduled_work();
1188 ipoib_dev_cleanup(priv->dev);
1191 free_netdev(priv->dev);
1194 return ERR_PTR(result);
1197 static void ipoib_add_one(struct ib_device *device)
1199 struct list_head *dev_list;
1200 struct net_device *dev;
1201 struct ipoib_dev_priv *priv;
1204 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1207 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1211 INIT_LIST_HEAD(dev_list);
1213 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1218 e = device->phys_port_cnt;
1221 for (p = s; p <= e; ++p) {
1222 dev = ipoib_add_port("ib%d", device, p);
1224 priv = netdev_priv(dev);
1225 list_add_tail(&priv->list, dev_list);
1229 ib_set_client_data(device, &ipoib_client, dev_list);
1232 static void ipoib_remove_one(struct ib_device *device)
1234 struct ipoib_dev_priv *priv, *tmp;
1235 struct list_head *dev_list;
1237 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1240 dev_list = ib_get_client_data(device, &ipoib_client);
1242 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1243 ib_unregister_event_handler(&priv->event_handler);
1244 flush_scheduled_work();
1246 unregister_netdev(priv->dev);
1247 ipoib_dev_cleanup(priv->dev);
1248 free_netdev(priv->dev);
1254 static int __init ipoib_init_module(void)
1258 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1259 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1260 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1262 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1263 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1264 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1266 ret = ipoib_register_debugfs();
1271 * We create our own workqueue mainly because we want to be
1272 * able to flush it when devices are being removed. We can't
1273 * use schedule_work()/flush_scheduled_work() because both
1274 * unregister_netdev() and linkwatch_event take the rtnl lock,
1275 * so flush_scheduled_work() can deadlock during device
1278 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1279 if (!ipoib_workqueue) {
1284 ib_sa_register_client(&ipoib_sa_client);
1286 ret = ib_register_client(&ipoib_client);
1293 ib_sa_unregister_client(&ipoib_sa_client);
1294 destroy_workqueue(ipoib_workqueue);
1297 ipoib_unregister_debugfs();
1302 static void __exit ipoib_cleanup_module(void)
1304 ib_unregister_client(&ipoib_client);
1305 ib_sa_unregister_client(&ipoib_sa_client);
1306 ipoib_unregister_debugfs();
1307 destroy_workqueue(ipoib_workqueue);
1310 module_init(ipoib_init_module);
1311 module_exit(ipoib_cleanup_module);