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);
151 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
154 * Now flush workqueue to make sure a scheduled task doesn't
155 * bring our internal state back up.
157 flush_workqueue(ipoib_workqueue);
159 ipoib_ib_dev_down(dev, 1);
160 ipoib_ib_dev_stop(dev, 1);
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) && new_mtu <= IPOIB_CM_MTU) {
188 if (new_mtu > priv->mcast_mtu)
189 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
195 if (new_mtu > IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN) {
199 priv->admin_mtu = new_mtu;
201 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
206 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
208 struct ipoib_dev_priv *priv = netdev_priv(dev);
209 struct rb_node *n = priv->path_tree.rb_node;
210 struct ipoib_path *path;
214 path = rb_entry(n, struct ipoib_path, rb_node);
216 ret = memcmp(gid, path->pathrec.dgid.raw,
217 sizeof (union ib_gid));
230 static int __path_add(struct net_device *dev, struct ipoib_path *path)
232 struct ipoib_dev_priv *priv = netdev_priv(dev);
233 struct rb_node **n = &priv->path_tree.rb_node;
234 struct rb_node *pn = NULL;
235 struct ipoib_path *tpath;
240 tpath = rb_entry(pn, struct ipoib_path, rb_node);
242 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
243 sizeof (union ib_gid));
252 rb_link_node(&path->rb_node, pn, n);
253 rb_insert_color(&path->rb_node, &priv->path_tree);
255 list_add_tail(&path->list, &priv->path_list);
260 static void path_free(struct net_device *dev, struct ipoib_path *path)
262 struct ipoib_dev_priv *priv = netdev_priv(dev);
263 struct ipoib_neigh *neigh, *tn;
267 while ((skb = __skb_dequeue(&path->queue)))
268 dev_kfree_skb_irq(skb);
270 spin_lock_irqsave(&priv->lock, flags);
272 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
274 * It's safe to call ipoib_put_ah() inside priv->lock
275 * here, because we know that path->ah will always
276 * hold one more reference, so ipoib_put_ah() will
277 * never do more than decrement the ref count.
280 ipoib_put_ah(neigh->ah);
282 ipoib_neigh_free(dev, neigh);
285 spin_unlock_irqrestore(&priv->lock, flags);
288 ipoib_put_ah(path->ah);
293 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
295 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
297 struct ipoib_path_iter *iter;
299 iter = kmalloc(sizeof *iter, GFP_KERNEL);
304 memset(iter->path.pathrec.dgid.raw, 0, 16);
306 if (ipoib_path_iter_next(iter)) {
314 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
316 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
318 struct ipoib_path *path;
321 spin_lock_irq(&priv->lock);
323 n = rb_first(&priv->path_tree);
326 path = rb_entry(n, struct ipoib_path, rb_node);
328 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
329 sizeof (union ib_gid)) < 0) {
338 spin_unlock_irq(&priv->lock);
343 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
344 struct ipoib_path *path)
349 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
351 void ipoib_flush_paths(struct net_device *dev)
353 struct ipoib_dev_priv *priv = netdev_priv(dev);
354 struct ipoib_path *path, *tp;
355 LIST_HEAD(remove_list);
357 spin_lock_irq(&priv->tx_lock);
358 spin_lock(&priv->lock);
360 list_splice(&priv->path_list, &remove_list);
361 INIT_LIST_HEAD(&priv->path_list);
363 list_for_each_entry(path, &remove_list, list)
364 rb_erase(&path->rb_node, &priv->path_tree);
366 list_for_each_entry_safe(path, tp, &remove_list, list) {
368 ib_sa_cancel_query(path->query_id, path->query);
369 spin_unlock(&priv->lock);
370 spin_unlock_irq(&priv->tx_lock);
371 wait_for_completion(&path->done);
372 path_free(dev, path);
373 spin_lock_irq(&priv->tx_lock);
374 spin_lock(&priv->lock);
376 spin_unlock(&priv->lock);
377 spin_unlock_irq(&priv->tx_lock);
380 static void path_rec_completion(int status,
381 struct ib_sa_path_rec *pathrec,
384 struct ipoib_path *path = path_ptr;
385 struct net_device *dev = path->dev;
386 struct ipoib_dev_priv *priv = netdev_priv(dev);
387 struct ipoib_ah *ah = NULL;
388 struct ipoib_neigh *neigh, *tn;
389 struct sk_buff_head skqueue;
394 ipoib_dbg(priv, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT "\n",
395 be16_to_cpu(pathrec->dlid), IPOIB_GID_ARG(pathrec->dgid));
397 ipoib_dbg(priv, "PathRec status %d for GID " IPOIB_GID_FMT "\n",
398 status, IPOIB_GID_ARG(path->pathrec.dgid));
400 skb_queue_head_init(&skqueue);
403 struct ib_ah_attr av;
405 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
406 ah = ipoib_create_ah(dev, priv->pd, &av);
409 spin_lock_irqsave(&priv->lock, flags);
414 path->pathrec = *pathrec;
416 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
417 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
419 while ((skb = __skb_dequeue(&path->queue)))
420 __skb_queue_tail(&skqueue, skb);
422 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
423 kref_get(&path->ah->ref);
424 neigh->ah = path->ah;
425 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
426 sizeof(union ib_gid));
428 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
429 if (!ipoib_cm_get(neigh))
430 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
433 if (!ipoib_cm_get(neigh)) {
434 list_del(&neigh->list);
436 ipoib_put_ah(neigh->ah);
437 ipoib_neigh_free(dev, neigh);
442 while ((skb = __skb_dequeue(&neigh->queue)))
443 __skb_queue_tail(&skqueue, skb);
448 complete(&path->done);
450 spin_unlock_irqrestore(&priv->lock, flags);
452 while ((skb = __skb_dequeue(&skqueue))) {
454 if (dev_queue_xmit(skb))
455 ipoib_warn(priv, "dev_queue_xmit failed "
456 "to requeue packet\n");
460 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
462 struct ipoib_dev_priv *priv = netdev_priv(dev);
463 struct ipoib_path *path;
465 path = kzalloc(sizeof *path, GFP_ATOMIC);
471 skb_queue_head_init(&path->queue);
473 INIT_LIST_HEAD(&path->neigh_list);
475 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
476 path->pathrec.sgid = priv->local_gid;
477 path->pathrec.pkey = cpu_to_be16(priv->pkey);
478 path->pathrec.numb_path = 1;
479 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
484 static int path_rec_start(struct net_device *dev,
485 struct ipoib_path *path)
487 struct ipoib_dev_priv *priv = netdev_priv(dev);
489 ipoib_dbg(priv, "Start path record lookup for " IPOIB_GID_FMT "\n",
490 IPOIB_GID_ARG(path->pathrec.dgid));
492 init_completion(&path->done);
495 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
497 IB_SA_PATH_REC_DGID |
498 IB_SA_PATH_REC_SGID |
499 IB_SA_PATH_REC_NUMB_PATH |
500 IB_SA_PATH_REC_TRAFFIC_CLASS |
505 if (path->query_id < 0) {
506 ipoib_warn(priv, "ib_sa_path_rec_get failed\n");
508 return path->query_id;
514 static void neigh_add_path(struct sk_buff *skb, struct net_device *dev)
516 struct ipoib_dev_priv *priv = netdev_priv(dev);
517 struct ipoib_path *path;
518 struct ipoib_neigh *neigh;
520 neigh = ipoib_neigh_alloc(skb->dst->neighbour, skb->dev);
522 ++dev->stats.tx_dropped;
523 dev_kfree_skb_any(skb);
528 * We can only be called from ipoib_start_xmit, so we're
529 * inside tx_lock -- no need to save/restore flags.
531 spin_lock(&priv->lock);
533 path = __path_find(dev, skb->dst->neighbour->ha + 4);
535 path = path_rec_create(dev, skb->dst->neighbour->ha + 4);
539 __path_add(dev, path);
542 list_add_tail(&neigh->list, &path->neigh_list);
545 kref_get(&path->ah->ref);
546 neigh->ah = path->ah;
547 memcpy(&neigh->dgid.raw, &path->pathrec.dgid.raw,
548 sizeof(union ib_gid));
550 if (ipoib_cm_enabled(dev, neigh->neighbour)) {
551 if (!ipoib_cm_get(neigh))
552 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
553 if (!ipoib_cm_get(neigh)) {
554 list_del(&neigh->list);
556 ipoib_put_ah(neigh->ah);
557 ipoib_neigh_free(dev, neigh);
560 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
561 __skb_queue_tail(&neigh->queue, skb);
563 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
564 skb_queue_len(&neigh->queue));
568 ipoib_send(dev, skb, path->ah, IPOIB_QPN(skb->dst->neighbour->ha));
572 if (!path->query && path_rec_start(dev, path))
575 __skb_queue_tail(&neigh->queue, skb);
578 spin_unlock(&priv->lock);
582 list_del(&neigh->list);
585 ipoib_neigh_free(dev, neigh);
587 ++dev->stats.tx_dropped;
588 dev_kfree_skb_any(skb);
590 spin_unlock(&priv->lock);
593 static void ipoib_path_lookup(struct sk_buff *skb, struct net_device *dev)
595 struct ipoib_dev_priv *priv = netdev_priv(skb->dev);
597 /* Look up path record for unicasts */
598 if (skb->dst->neighbour->ha[4] != 0xff) {
599 neigh_add_path(skb, dev);
603 /* Add in the P_Key for multicasts */
604 skb->dst->neighbour->ha[8] = (priv->pkey >> 8) & 0xff;
605 skb->dst->neighbour->ha[9] = priv->pkey & 0xff;
606 ipoib_mcast_send(dev, skb->dst->neighbour->ha + 4, skb);
609 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
610 struct ipoib_pseudoheader *phdr)
612 struct ipoib_dev_priv *priv = netdev_priv(dev);
613 struct ipoib_path *path;
616 * We can only be called from ipoib_start_xmit, so we're
617 * inside tx_lock -- no need to save/restore flags.
619 spin_lock(&priv->lock);
621 path = __path_find(dev, phdr->hwaddr + 4);
623 path = path_rec_create(dev, phdr->hwaddr + 4);
625 /* put pseudoheader back on for next time */
626 skb_push(skb, sizeof *phdr);
627 __skb_queue_tail(&path->queue, skb);
629 if (path_rec_start(dev, path)) {
630 spin_unlock(&priv->lock);
631 path_free(dev, path);
634 __path_add(dev, path);
636 ++dev->stats.tx_dropped;
637 dev_kfree_skb_any(skb);
640 spin_unlock(&priv->lock);
645 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
646 be16_to_cpu(path->pathrec.dlid));
648 ipoib_send(dev, skb, path->ah, IPOIB_QPN(phdr->hwaddr));
649 } else if ((path->query || !path_rec_start(dev, path)) &&
650 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
651 /* put pseudoheader back on for next time */
652 skb_push(skb, sizeof *phdr);
653 __skb_queue_tail(&path->queue, skb);
655 ++dev->stats.tx_dropped;
656 dev_kfree_skb_any(skb);
659 spin_unlock(&priv->lock);
662 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
664 struct ipoib_dev_priv *priv = netdev_priv(dev);
665 struct ipoib_neigh *neigh;
668 if (unlikely(!spin_trylock_irqsave(&priv->tx_lock, flags)))
669 return NETDEV_TX_LOCKED;
672 * Check if our queue is stopped. Since we have the LLTX bit
673 * set, we can't rely on netif_stop_queue() preventing our
674 * xmit function from being called with a full queue.
676 if (unlikely(netif_queue_stopped(dev))) {
677 spin_unlock_irqrestore(&priv->tx_lock, flags);
678 return NETDEV_TX_BUSY;
681 if (likely(skb->dst && skb->dst->neighbour)) {
682 if (unlikely(!*to_ipoib_neigh(skb->dst->neighbour))) {
683 ipoib_path_lookup(skb, dev);
687 neigh = *to_ipoib_neigh(skb->dst->neighbour);
689 if (ipoib_cm_get(neigh)) {
690 if (ipoib_cm_up(neigh)) {
691 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
694 } else if (neigh->ah) {
695 if (unlikely((memcmp(&neigh->dgid.raw,
696 skb->dst->neighbour->ha + 4,
697 sizeof(union ib_gid))) ||
698 (neigh->dev != dev))) {
699 spin_lock(&priv->lock);
701 * It's safe to call ipoib_put_ah() inside
702 * priv->lock here, because we know that
703 * path->ah will always hold one more reference,
704 * so ipoib_put_ah() will never do more than
705 * decrement the ref count.
707 ipoib_put_ah(neigh->ah);
708 list_del(&neigh->list);
709 ipoib_neigh_free(dev, neigh);
710 spin_unlock(&priv->lock);
711 ipoib_path_lookup(skb, dev);
715 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(skb->dst->neighbour->ha));
719 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
720 spin_lock(&priv->lock);
721 __skb_queue_tail(&neigh->queue, skb);
722 spin_unlock(&priv->lock);
724 ++dev->stats.tx_dropped;
725 dev_kfree_skb_any(skb);
728 struct ipoib_pseudoheader *phdr =
729 (struct ipoib_pseudoheader *) skb->data;
730 skb_pull(skb, sizeof *phdr);
732 if (phdr->hwaddr[4] == 0xff) {
733 /* Add in the P_Key for multicast*/
734 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
735 phdr->hwaddr[9] = priv->pkey & 0xff;
737 ipoib_mcast_send(dev, phdr->hwaddr + 4, skb);
739 /* unicast GID -- should be ARP or RARP reply */
741 if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
742 (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
743 ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
745 skb->dst ? "neigh" : "dst",
746 be16_to_cpup((__be16 *) skb->data),
747 IPOIB_QPN(phdr->hwaddr),
748 IPOIB_GID_RAW_ARG(phdr->hwaddr + 4));
749 dev_kfree_skb_any(skb);
750 ++dev->stats.tx_dropped;
754 unicast_arp_send(skb, dev, phdr);
759 spin_unlock_irqrestore(&priv->tx_lock, flags);
764 static void ipoib_timeout(struct net_device *dev)
766 struct ipoib_dev_priv *priv = netdev_priv(dev);
768 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
769 jiffies_to_msecs(jiffies - dev->trans_start));
770 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
771 netif_queue_stopped(dev),
772 priv->tx_head, priv->tx_tail);
773 /* XXX reset QP, etc. */
776 static int ipoib_hard_header(struct sk_buff *skb,
777 struct net_device *dev,
779 const void *daddr, const void *saddr, unsigned len)
781 struct ipoib_header *header;
783 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
785 header->proto = htons(type);
786 header->reserved = 0;
789 * If we don't have a neighbour structure, stuff the
790 * destination address onto the front of the skb so we can
791 * figure out where to send the packet later.
793 if ((!skb->dst || !skb->dst->neighbour) && daddr) {
794 struct ipoib_pseudoheader *phdr =
795 (struct ipoib_pseudoheader *) skb_push(skb, sizeof *phdr);
796 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
802 static void ipoib_set_mcast_list(struct net_device *dev)
804 struct ipoib_dev_priv *priv = netdev_priv(dev);
806 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
807 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
811 queue_work(ipoib_workqueue, &priv->restart_task);
814 static void ipoib_neigh_cleanup(struct neighbour *n)
816 struct ipoib_neigh *neigh;
817 struct ipoib_dev_priv *priv = netdev_priv(n->dev);
819 struct ipoib_ah *ah = NULL;
821 neigh = *to_ipoib_neigh(n);
823 priv = netdev_priv(neigh->dev);
824 ipoib_dbg(priv, "neigh_destructor for bonding device: %s\n",
829 "neigh_cleanup for %06x " IPOIB_GID_FMT "\n",
831 IPOIB_GID_RAW_ARG(n->ha + 4));
833 spin_lock_irqsave(&priv->lock, flags);
837 list_del(&neigh->list);
838 ipoib_neigh_free(n->dev, neigh);
840 spin_unlock_irqrestore(&priv->lock, flags);
846 struct ipoib_neigh *ipoib_neigh_alloc(struct neighbour *neighbour,
847 struct net_device *dev)
849 struct ipoib_neigh *neigh;
851 neigh = kmalloc(sizeof *neigh, GFP_ATOMIC);
855 neigh->neighbour = neighbour;
857 *to_ipoib_neigh(neighbour) = neigh;
858 skb_queue_head_init(&neigh->queue);
859 ipoib_cm_set(neigh, NULL);
864 void ipoib_neigh_free(struct net_device *dev, struct ipoib_neigh *neigh)
867 *to_ipoib_neigh(neigh->neighbour) = NULL;
868 while ((skb = __skb_dequeue(&neigh->queue))) {
869 ++dev->stats.tx_dropped;
870 dev_kfree_skb_any(skb);
872 if (ipoib_cm_get(neigh))
873 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
877 static int ipoib_neigh_setup_dev(struct net_device *dev, struct neigh_parms *parms)
879 parms->neigh_cleanup = ipoib_neigh_cleanup;
884 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
886 struct ipoib_dev_priv *priv = netdev_priv(dev);
888 /* Allocate RX/TX "rings" to hold queued skbs */
889 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
891 if (!priv->rx_ring) {
892 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
893 ca->name, ipoib_recvq_size);
897 priv->tx_ring = kzalloc(ipoib_sendq_size * sizeof *priv->tx_ring,
899 if (!priv->tx_ring) {
900 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
901 ca->name, ipoib_sendq_size);
902 goto out_rx_ring_cleanup;
905 /* priv->tx_head & tx_tail are already 0 */
907 if (ipoib_ib_dev_init(dev, ca, port))
908 goto out_tx_ring_cleanup;
913 kfree(priv->tx_ring);
916 kfree(priv->rx_ring);
922 void ipoib_dev_cleanup(struct net_device *dev)
924 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
926 ipoib_delete_debug_files(dev);
928 /* Delete any child interfaces first */
929 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
930 unregister_netdev(cpriv->dev);
931 ipoib_dev_cleanup(cpriv->dev);
932 free_netdev(cpriv->dev);
935 ipoib_ib_dev_cleanup(dev);
937 kfree(priv->rx_ring);
938 kfree(priv->tx_ring);
940 priv->rx_ring = NULL;
941 priv->tx_ring = NULL;
944 static const struct header_ops ipoib_header_ops = {
945 .create = ipoib_hard_header,
948 static void ipoib_setup(struct net_device *dev)
950 struct ipoib_dev_priv *priv = netdev_priv(dev);
952 dev->open = ipoib_open;
953 dev->stop = ipoib_stop;
954 dev->change_mtu = ipoib_change_mtu;
955 dev->hard_start_xmit = ipoib_start_xmit;
956 dev->tx_timeout = ipoib_timeout;
957 dev->header_ops = &ipoib_header_ops;
958 dev->set_multicast_list = ipoib_set_mcast_list;
959 dev->neigh_setup = ipoib_neigh_setup_dev;
961 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
963 dev->watchdog_timeo = HZ;
965 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
968 * We add in INFINIBAND_ALEN to allow for the destination
969 * address "pseudoheader" for skbs without neighbour struct.
971 dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
972 dev->addr_len = INFINIBAND_ALEN;
973 dev->type = ARPHRD_INFINIBAND;
974 dev->tx_queue_len = ipoib_sendq_size * 2;
975 dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
977 /* MTU will be reset when mcast join happens */
978 dev->mtu = IPOIB_PACKET_SIZE - IPOIB_ENCAP_LEN;
979 priv->mcast_mtu = priv->admin_mtu = dev->mtu;
981 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
983 netif_carrier_off(dev);
987 spin_lock_init(&priv->lock);
988 spin_lock_init(&priv->tx_lock);
990 mutex_init(&priv->mcast_mutex);
991 mutex_init(&priv->vlan_mutex);
993 INIT_LIST_HEAD(&priv->path_list);
994 INIT_LIST_HEAD(&priv->child_intfs);
995 INIT_LIST_HEAD(&priv->dead_ahs);
996 INIT_LIST_HEAD(&priv->multicast_list);
998 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
999 INIT_WORK(&priv->pkey_event_task, ipoib_pkey_event);
1000 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1001 INIT_WORK(&priv->flush_task, ipoib_ib_dev_flush);
1002 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1003 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1006 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1008 struct net_device *dev;
1010 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1015 return netdev_priv(dev);
1018 static ssize_t show_pkey(struct device *dev,
1019 struct device_attribute *attr, char *buf)
1021 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1023 return sprintf(buf, "0x%04x\n", priv->pkey);
1025 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1027 static ssize_t show_umcast(struct device *dev,
1028 struct device_attribute *attr, char *buf)
1030 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1032 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1035 static ssize_t set_umcast(struct device *dev,
1036 struct device_attribute *attr,
1037 const char *buf, size_t count)
1039 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1040 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1042 if (umcast_val > 0) {
1043 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1044 ipoib_warn(priv, "ignoring multicast groups joined directly "
1047 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1051 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1053 int ipoib_add_umcast_attr(struct net_device *dev)
1055 return device_create_file(&dev->dev, &dev_attr_umcast);
1058 static ssize_t create_child(struct device *dev,
1059 struct device_attribute *attr,
1060 const char *buf, size_t count)
1065 if (sscanf(buf, "%i", &pkey) != 1)
1068 if (pkey < 0 || pkey > 0xffff)
1072 * Set the full membership bit, so that we join the right
1073 * broadcast group, etc.
1077 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1079 return ret ? ret : count;
1081 static DEVICE_ATTR(create_child, S_IWUGO, NULL, create_child);
1083 static ssize_t delete_child(struct device *dev,
1084 struct device_attribute *attr,
1085 const char *buf, size_t count)
1090 if (sscanf(buf, "%i", &pkey) != 1)
1093 if (pkey < 0 || pkey > 0xffff)
1096 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1098 return ret ? ret : count;
1101 static DEVICE_ATTR(delete_child, S_IWUGO, NULL, delete_child);
1103 int ipoib_add_pkey_attr(struct net_device *dev)
1105 return device_create_file(&dev->dev, &dev_attr_pkey);
1108 static struct net_device *ipoib_add_port(const char *format,
1109 struct ib_device *hca, u8 port)
1111 struct ipoib_dev_priv *priv;
1112 int result = -ENOMEM;
1114 priv = ipoib_intf_alloc(format);
1116 goto alloc_mem_failed;
1118 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1120 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1122 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1123 hca->name, port, result);
1124 goto device_init_failed;
1128 * Set the full membership bit, so that we join the right
1129 * broadcast group, etc.
1131 priv->pkey |= 0x8000;
1133 priv->dev->broadcast[8] = priv->pkey >> 8;
1134 priv->dev->broadcast[9] = priv->pkey & 0xff;
1136 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1138 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1139 hca->name, port, result);
1140 goto device_init_failed;
1142 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1145 result = ipoib_dev_init(priv->dev, hca, port);
1147 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1148 hca->name, port, result);
1149 goto device_init_failed;
1152 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1153 priv->ca, ipoib_event);
1154 result = ib_register_event_handler(&priv->event_handler);
1156 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1157 "port %d (ret = %d)\n",
1158 hca->name, port, result);
1162 result = register_netdev(priv->dev);
1164 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1165 hca->name, port, result);
1166 goto register_failed;
1169 ipoib_create_debug_files(priv->dev);
1171 if (ipoib_cm_add_mode_attr(priv->dev))
1173 if (ipoib_add_pkey_attr(priv->dev))
1175 if (ipoib_add_umcast_attr(priv->dev))
1177 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1179 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1185 ipoib_delete_debug_files(priv->dev);
1186 unregister_netdev(priv->dev);
1189 ib_unregister_event_handler(&priv->event_handler);
1190 flush_scheduled_work();
1193 ipoib_dev_cleanup(priv->dev);
1196 free_netdev(priv->dev);
1199 return ERR_PTR(result);
1202 static void ipoib_add_one(struct ib_device *device)
1204 struct list_head *dev_list;
1205 struct net_device *dev;
1206 struct ipoib_dev_priv *priv;
1209 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1212 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1216 INIT_LIST_HEAD(dev_list);
1218 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1223 e = device->phys_port_cnt;
1226 for (p = s; p <= e; ++p) {
1227 dev = ipoib_add_port("ib%d", device, p);
1229 priv = netdev_priv(dev);
1230 list_add_tail(&priv->list, dev_list);
1234 ib_set_client_data(device, &ipoib_client, dev_list);
1237 static void ipoib_remove_one(struct ib_device *device)
1239 struct ipoib_dev_priv *priv, *tmp;
1240 struct list_head *dev_list;
1242 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1245 dev_list = ib_get_client_data(device, &ipoib_client);
1247 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1248 ib_unregister_event_handler(&priv->event_handler);
1249 flush_scheduled_work();
1251 unregister_netdev(priv->dev);
1252 ipoib_dev_cleanup(priv->dev);
1253 free_netdev(priv->dev);
1259 static int __init ipoib_init_module(void)
1263 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1264 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1265 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1267 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1268 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1269 ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
1271 ret = ipoib_register_debugfs();
1276 * We create our own workqueue mainly because we want to be
1277 * able to flush it when devices are being removed. We can't
1278 * use schedule_work()/flush_scheduled_work() because both
1279 * unregister_netdev() and linkwatch_event take the rtnl lock,
1280 * so flush_scheduled_work() can deadlock during device
1283 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1284 if (!ipoib_workqueue) {
1289 ib_sa_register_client(&ipoib_sa_client);
1291 ret = ib_register_client(&ipoib_client);
1298 ib_sa_unregister_client(&ipoib_sa_client);
1299 destroy_workqueue(ipoib_workqueue);
1302 ipoib_unregister_debugfs();
1307 static void __exit ipoib_cleanup_module(void)
1309 ib_unregister_client(&ipoib_client);
1310 ib_sa_unregister_client(&ipoib_sa_client);
1311 ipoib_unregister_debugfs();
1312 destroy_workqueue(ipoib_workqueue);
1315 module_init(ipoib_init_module);
1316 module_exit(ipoib_cleanup_module);