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 */
49 #include <linux/jhash.h>
52 #define DRV_VERSION "1.0.0"
54 const char ipoib_driver_version[] = DRV_VERSION;
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
58 MODULE_LICENSE("Dual BSD/GPL");
59 MODULE_VERSION(DRV_VERSION);
61 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
62 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
64 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
65 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
66 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
67 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
69 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
70 int ipoib_debug_level;
72 module_param_named(debug_level, ipoib_debug_level, int, 0644);
73 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
76 struct ipoib_path_iter {
77 struct net_device *dev;
78 struct ipoib_path path;
81 static const u8 ipv4_bcast_addr[] = {
82 0x00, 0xff, 0xff, 0xff,
83 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
84 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
87 struct workqueue_struct *ipoib_workqueue;
89 struct ib_sa_client ipoib_sa_client;
91 static void ipoib_add_one(struct ib_device *device);
92 static void ipoib_remove_one(struct ib_device *device);
93 static void ipoib_neigh_reclaim(struct rcu_head *rp);
95 static struct ib_client ipoib_client = {
98 .remove = ipoib_remove_one
101 int ipoib_open(struct net_device *dev)
103 struct ipoib_dev_priv *priv = netdev_priv(dev);
105 ipoib_dbg(priv, "bringing up interface\n");
107 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
109 if (ipoib_pkey_dev_delay_open(dev))
112 if (ipoib_ib_dev_open(dev))
115 if (ipoib_ib_dev_up(dev))
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 ipoib_ib_dev_stop(dev, 1);
143 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
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);
156 netif_stop_queue(dev);
158 ipoib_ib_dev_down(dev, 1);
159 ipoib_ib_dev_stop(dev, 0);
161 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
162 struct ipoib_dev_priv *cpriv;
164 /* Bring down any child interfaces too */
165 mutex_lock(&priv->vlan_mutex);
166 list_for_each_entry(cpriv, &priv->child_intfs, list) {
169 flags = cpriv->dev->flags;
170 if (!(flags & IFF_UP))
173 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
175 mutex_unlock(&priv->vlan_mutex);
181 static void ipoib_uninit(struct net_device *dev)
183 ipoib_dev_cleanup(dev);
186 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
188 struct ipoib_dev_priv *priv = netdev_priv(dev);
190 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
191 features &= ~(NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO);
196 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
198 struct ipoib_dev_priv *priv = netdev_priv(dev);
200 /* dev->mtu > 2K ==> connected mode */
201 if (ipoib_cm_admin_enabled(dev)) {
202 if (new_mtu > ipoib_cm_max_mtu(dev))
205 if (new_mtu > priv->mcast_mtu)
206 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
213 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
216 priv->admin_mtu = new_mtu;
218 dev->mtu = min(priv->mcast_mtu, priv->admin_mtu);
223 int ipoib_set_mode(struct net_device *dev, const char *buf)
225 struct ipoib_dev_priv *priv = netdev_priv(dev);
227 /* flush paths if we switch modes so that connections are restarted */
228 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
229 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
230 ipoib_warn(priv, "enabling connected mode "
231 "will cause multicast packet drops\n");
232 netdev_update_features(dev);
234 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
236 ipoib_flush_paths(dev);
241 if (!strcmp(buf, "datagram\n")) {
242 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
243 netdev_update_features(dev);
244 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
246 ipoib_flush_paths(dev);
254 static struct ipoib_path *__path_find(struct net_device *dev, void *gid)
256 struct ipoib_dev_priv *priv = netdev_priv(dev);
257 struct rb_node *n = priv->path_tree.rb_node;
258 struct ipoib_path *path;
262 path = rb_entry(n, struct ipoib_path, rb_node);
264 ret = memcmp(gid, path->pathrec.dgid.raw,
265 sizeof (union ib_gid));
278 static int __path_add(struct net_device *dev, struct ipoib_path *path)
280 struct ipoib_dev_priv *priv = netdev_priv(dev);
281 struct rb_node **n = &priv->path_tree.rb_node;
282 struct rb_node *pn = NULL;
283 struct ipoib_path *tpath;
288 tpath = rb_entry(pn, struct ipoib_path, rb_node);
290 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
291 sizeof (union ib_gid));
300 rb_link_node(&path->rb_node, pn, n);
301 rb_insert_color(&path->rb_node, &priv->path_tree);
303 list_add_tail(&path->list, &priv->path_list);
308 static void path_free(struct net_device *dev, struct ipoib_path *path)
312 while ((skb = __skb_dequeue(&path->queue)))
313 dev_kfree_skb_irq(skb);
315 ipoib_dbg(netdev_priv(dev), "path_free\n");
317 /* remove all neigh connected to this path */
318 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
321 ipoib_put_ah(path->ah);
326 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
328 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
330 struct ipoib_path_iter *iter;
332 iter = kmalloc(sizeof *iter, GFP_KERNEL);
337 memset(iter->path.pathrec.dgid.raw, 0, 16);
339 if (ipoib_path_iter_next(iter)) {
347 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
349 struct ipoib_dev_priv *priv = netdev_priv(iter->dev);
351 struct ipoib_path *path;
354 spin_lock_irq(&priv->lock);
356 n = rb_first(&priv->path_tree);
359 path = rb_entry(n, struct ipoib_path, rb_node);
361 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
362 sizeof (union ib_gid)) < 0) {
371 spin_unlock_irq(&priv->lock);
376 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
377 struct ipoib_path *path)
382 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
384 void ipoib_mark_paths_invalid(struct net_device *dev)
386 struct ipoib_dev_priv *priv = netdev_priv(dev);
387 struct ipoib_path *path, *tp;
389 spin_lock_irq(&priv->lock);
391 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
392 ipoib_dbg(priv, "mark path LID 0x%04x GID %pI6 invalid\n",
393 be16_to_cpu(path->pathrec.dlid),
394 path->pathrec.dgid.raw);
398 spin_unlock_irq(&priv->lock);
401 void ipoib_flush_paths(struct net_device *dev)
403 struct ipoib_dev_priv *priv = netdev_priv(dev);
404 struct ipoib_path *path, *tp;
405 LIST_HEAD(remove_list);
408 netif_tx_lock_bh(dev);
409 spin_lock_irqsave(&priv->lock, flags);
411 list_splice_init(&priv->path_list, &remove_list);
413 list_for_each_entry(path, &remove_list, list)
414 rb_erase(&path->rb_node, &priv->path_tree);
416 list_for_each_entry_safe(path, tp, &remove_list, list) {
418 ib_sa_cancel_query(path->query_id, path->query);
419 spin_unlock_irqrestore(&priv->lock, flags);
420 netif_tx_unlock_bh(dev);
421 wait_for_completion(&path->done);
422 path_free(dev, path);
423 netif_tx_lock_bh(dev);
424 spin_lock_irqsave(&priv->lock, flags);
427 spin_unlock_irqrestore(&priv->lock, flags);
428 netif_tx_unlock_bh(dev);
431 static void path_rec_completion(int status,
432 struct ib_sa_path_rec *pathrec,
435 struct ipoib_path *path = path_ptr;
436 struct net_device *dev = path->dev;
437 struct ipoib_dev_priv *priv = netdev_priv(dev);
438 struct ipoib_ah *ah = NULL;
439 struct ipoib_ah *old_ah = NULL;
440 struct ipoib_neigh *neigh, *tn;
441 struct sk_buff_head skqueue;
446 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
447 be16_to_cpu(pathrec->dlid), pathrec->dgid.raw);
449 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
450 status, path->pathrec.dgid.raw);
452 skb_queue_head_init(&skqueue);
455 struct ib_ah_attr av;
457 if (!ib_init_ah_from_path(priv->ca, priv->port, pathrec, &av))
458 ah = ipoib_create_ah(dev, priv->pd, &av);
461 spin_lock_irqsave(&priv->lock, flags);
463 if (!IS_ERR_OR_NULL(ah)) {
464 path->pathrec = *pathrec;
469 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
470 ah, be16_to_cpu(pathrec->dlid), pathrec->sl);
472 while ((skb = __skb_dequeue(&path->queue)))
473 __skb_queue_tail(&skqueue, skb);
475 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
477 WARN_ON(neigh->ah != old_ah);
479 * Dropping the ah reference inside
480 * priv->lock is safe here, because we
481 * will hold one more reference from
482 * the original value of path->ah (ie
485 ipoib_put_ah(neigh->ah);
487 kref_get(&path->ah->ref);
488 neigh->ah = path->ah;
490 if (ipoib_cm_enabled(dev, neigh->daddr)) {
491 if (!ipoib_cm_get(neigh))
492 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
495 if (!ipoib_cm_get(neigh)) {
496 list_del(&neigh->list);
497 ipoib_neigh_free(neigh);
502 while ((skb = __skb_dequeue(&neigh->queue)))
503 __skb_queue_tail(&skqueue, skb);
509 complete(&path->done);
511 spin_unlock_irqrestore(&priv->lock, flags);
513 if (IS_ERR_OR_NULL(ah))
514 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
517 ipoib_put_ah(old_ah);
519 while ((skb = __skb_dequeue(&skqueue))) {
521 if (dev_queue_xmit(skb))
522 ipoib_warn(priv, "dev_queue_xmit failed "
523 "to requeue packet\n");
527 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
529 struct ipoib_dev_priv *priv = netdev_priv(dev);
530 struct ipoib_path *path;
532 if (!priv->broadcast)
535 path = kzalloc(sizeof *path, GFP_ATOMIC);
541 skb_queue_head_init(&path->queue);
543 INIT_LIST_HEAD(&path->neigh_list);
545 memcpy(path->pathrec.dgid.raw, gid, sizeof (union ib_gid));
546 path->pathrec.sgid = priv->local_gid;
547 path->pathrec.pkey = cpu_to_be16(priv->pkey);
548 path->pathrec.numb_path = 1;
549 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
554 static int path_rec_start(struct net_device *dev,
555 struct ipoib_path *path)
557 struct ipoib_dev_priv *priv = netdev_priv(dev);
559 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
560 path->pathrec.dgid.raw);
562 init_completion(&path->done);
565 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
567 IB_SA_PATH_REC_DGID |
568 IB_SA_PATH_REC_SGID |
569 IB_SA_PATH_REC_NUMB_PATH |
570 IB_SA_PATH_REC_TRAFFIC_CLASS |
575 if (path->query_id < 0) {
576 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
578 complete(&path->done);
579 return path->query_id;
585 static void neigh_add_path(struct sk_buff *skb, u8 *daddr,
586 struct net_device *dev)
588 struct ipoib_dev_priv *priv = netdev_priv(dev);
589 struct ipoib_path *path;
590 struct ipoib_neigh *neigh;
593 spin_lock_irqsave(&priv->lock, flags);
594 neigh = ipoib_neigh_alloc(daddr, dev);
596 spin_unlock_irqrestore(&priv->lock, flags);
597 ++dev->stats.tx_dropped;
598 dev_kfree_skb_any(skb);
602 path = __path_find(dev, daddr + 4);
604 path = path_rec_create(dev, daddr + 4);
608 __path_add(dev, path);
611 list_add_tail(&neigh->list, &path->neigh_list);
614 kref_get(&path->ah->ref);
615 neigh->ah = path->ah;
617 if (ipoib_cm_enabled(dev, neigh->daddr)) {
618 if (!ipoib_cm_get(neigh))
619 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
620 if (!ipoib_cm_get(neigh)) {
621 list_del(&neigh->list);
622 ipoib_neigh_free(neigh);
625 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE)
626 __skb_queue_tail(&neigh->queue, skb);
628 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
629 skb_queue_len(&neigh->queue));
633 spin_unlock_irqrestore(&priv->lock, flags);
634 ipoib_send(dev, skb, path->ah, IPOIB_QPN(daddr));
635 ipoib_neigh_put(neigh);
641 if (!path->query && path_rec_start(dev, path))
644 __skb_queue_tail(&neigh->queue, skb);
647 spin_unlock_irqrestore(&priv->lock, flags);
648 ipoib_neigh_put(neigh);
652 list_del(&neigh->list);
655 ipoib_neigh_free(neigh);
657 ++dev->stats.tx_dropped;
658 dev_kfree_skb_any(skb);
660 spin_unlock_irqrestore(&priv->lock, flags);
661 ipoib_neigh_put(neigh);
664 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
667 struct ipoib_dev_priv *priv = netdev_priv(dev);
668 struct ipoib_path *path;
671 spin_lock_irqsave(&priv->lock, flags);
673 path = __path_find(dev, cb->hwaddr + 4);
674 if (!path || !path->valid) {
678 path = path_rec_create(dev, cb->hwaddr + 4);
682 __skb_queue_tail(&path->queue, skb);
684 if (!path->query && path_rec_start(dev, path)) {
685 spin_unlock_irqrestore(&priv->lock, flags);
687 path_free(dev, path);
690 __path_add(dev, path);
692 ++dev->stats.tx_dropped;
693 dev_kfree_skb_any(skb);
696 spin_unlock_irqrestore(&priv->lock, flags);
701 ipoib_dbg(priv, "Send unicast ARP to %04x\n",
702 be16_to_cpu(path->pathrec.dlid));
704 spin_unlock_irqrestore(&priv->lock, flags);
705 ipoib_send(dev, skb, path->ah, IPOIB_QPN(cb->hwaddr));
707 } else if ((path->query || !path_rec_start(dev, path)) &&
708 skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
709 __skb_queue_tail(&path->queue, skb);
711 ++dev->stats.tx_dropped;
712 dev_kfree_skb_any(skb);
715 spin_unlock_irqrestore(&priv->lock, flags);
718 static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
720 struct ipoib_dev_priv *priv = netdev_priv(dev);
721 struct ipoib_neigh *neigh;
722 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
723 struct ipoib_header *header;
726 header = (struct ipoib_header *) skb->data;
728 if (unlikely(cb->hwaddr[4] == 0xff)) {
729 /* multicast, arrange "if" according to probability */
730 if ((header->proto != htons(ETH_P_IP)) &&
731 (header->proto != htons(ETH_P_IPV6)) &&
732 (header->proto != htons(ETH_P_ARP)) &&
733 (header->proto != htons(ETH_P_RARP))) {
734 /* ethertype not supported by IPoIB */
735 ++dev->stats.tx_dropped;
736 dev_kfree_skb_any(skb);
739 /* Add in the P_Key for multicast*/
740 cb->hwaddr[8] = (priv->pkey >> 8) & 0xff;
741 cb->hwaddr[9] = priv->pkey & 0xff;
743 neigh = ipoib_neigh_get(dev, cb->hwaddr);
745 goto send_using_neigh;
746 ipoib_mcast_send(dev, cb->hwaddr, skb);
750 /* unicast, arrange "switch" according to probability */
751 switch (header->proto) {
752 case htons(ETH_P_IP):
753 case htons(ETH_P_IPV6):
754 neigh = ipoib_neigh_get(dev, cb->hwaddr);
755 if (unlikely(!neigh)) {
756 neigh_add_path(skb, cb->hwaddr, dev);
760 case htons(ETH_P_ARP):
761 case htons(ETH_P_RARP):
762 /* for unicast ARP and RARP should always perform path find */
763 unicast_arp_send(skb, dev, cb);
766 /* ethertype not supported by IPoIB */
767 ++dev->stats.tx_dropped;
768 dev_kfree_skb_any(skb);
773 /* note we now hold a ref to neigh */
774 if (ipoib_cm_get(neigh)) {
775 if (ipoib_cm_up(neigh)) {
776 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
779 } else if (neigh->ah) {
780 ipoib_send(dev, skb, neigh->ah, IPOIB_QPN(cb->hwaddr));
784 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
785 spin_lock_irqsave(&priv->lock, flags);
786 __skb_queue_tail(&neigh->queue, skb);
787 spin_unlock_irqrestore(&priv->lock, flags);
789 ++dev->stats.tx_dropped;
790 dev_kfree_skb_any(skb);
794 ipoib_neigh_put(neigh);
799 static void ipoib_timeout(struct net_device *dev)
801 struct ipoib_dev_priv *priv = netdev_priv(dev);
803 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
804 jiffies_to_msecs(jiffies - dev->trans_start));
805 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
806 netif_queue_stopped(dev),
807 priv->tx_head, priv->tx_tail);
808 /* XXX reset QP, etc. */
811 static int ipoib_hard_header(struct sk_buff *skb,
812 struct net_device *dev,
814 const void *daddr, const void *saddr, unsigned len)
816 struct ipoib_header *header;
817 struct ipoib_cb *cb = (struct ipoib_cb *) skb->cb;
819 header = (struct ipoib_header *) skb_push(skb, sizeof *header);
821 header->proto = htons(type);
822 header->reserved = 0;
825 * we don't rely on dst_entry structure, always stuff the
826 * destination address into skb->cb so we can figure out where
827 * to send the packet later.
829 memcpy(cb->hwaddr, daddr, INFINIBAND_ALEN);
834 static void ipoib_set_mcast_list(struct net_device *dev)
836 struct ipoib_dev_priv *priv = netdev_priv(dev);
838 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
839 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
843 queue_work(ipoib_workqueue, &priv->restart_task);
846 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
849 * Use only the address parts that contributes to spreading
850 * The subnet prefix is not used as one can not connect to
851 * same remote port (GUID) using the same remote QPN via two
854 /* qpn octets[1:4) & port GUID octets[12:20) */
855 u32 *d32 = (u32 *) daddr;
858 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
859 return hv & htbl->mask;
862 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
864 struct ipoib_dev_priv *priv = netdev_priv(dev);
865 struct ipoib_neigh_table *ntbl = &priv->ntbl;
866 struct ipoib_neigh_hash *htbl;
867 struct ipoib_neigh *neigh = NULL;
872 htbl = rcu_dereference_bh(ntbl->htbl);
877 hash_val = ipoib_addr_hash(htbl, daddr);
878 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
880 neigh = rcu_dereference_bh(neigh->hnext)) {
881 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
882 /* found, take one ref on behalf of the caller */
883 if (!atomic_inc_not_zero(&neigh->refcnt)) {
888 neigh->alive = jiffies;
894 rcu_read_unlock_bh();
898 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
900 struct ipoib_neigh_table *ntbl = &priv->ntbl;
901 struct ipoib_neigh_hash *htbl;
902 unsigned long neigh_obsolete;
907 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
910 spin_lock_irqsave(&priv->lock, flags);
912 htbl = rcu_dereference_protected(ntbl->htbl,
913 lockdep_is_held(&priv->lock));
918 /* neigh is obsolete if it was idle for two GC periods */
919 dt = 2 * arp_tbl.gc_interval;
920 neigh_obsolete = jiffies - dt;
921 /* handle possible race condition */
922 if (test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
925 for (i = 0; i < htbl->size; i++) {
926 struct ipoib_neigh *neigh;
927 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
929 while ((neigh = rcu_dereference_protected(*np,
930 lockdep_is_held(&priv->lock))) != NULL) {
931 /* was the neigh idle for two GC periods */
932 if (time_after(neigh_obsolete, neigh->alive)) {
933 rcu_assign_pointer(*np,
934 rcu_dereference_protected(neigh->hnext,
935 lockdep_is_held(&priv->lock)));
936 /* remove from path/mc list */
937 list_del(&neigh->list);
938 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
947 spin_unlock_irqrestore(&priv->lock, flags);
950 static void ipoib_reap_neigh(struct work_struct *work)
952 struct ipoib_dev_priv *priv =
953 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
955 __ipoib_reap_neigh(priv);
957 if (!test_bit(IPOIB_STOP_NEIGH_GC, &priv->flags))
958 queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
959 arp_tbl.gc_interval);
963 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
964 struct net_device *dev)
966 struct ipoib_neigh *neigh;
968 neigh = kzalloc(sizeof *neigh, GFP_ATOMIC);
973 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
974 skb_queue_head_init(&neigh->queue);
975 INIT_LIST_HEAD(&neigh->list);
976 ipoib_cm_set(neigh, NULL);
977 /* one ref on behalf of the caller */
978 atomic_set(&neigh->refcnt, 1);
983 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
984 struct net_device *dev)
986 struct ipoib_dev_priv *priv = netdev_priv(dev);
987 struct ipoib_neigh_table *ntbl = &priv->ntbl;
988 struct ipoib_neigh_hash *htbl;
989 struct ipoib_neigh *neigh;
992 htbl = rcu_dereference_protected(ntbl->htbl,
993 lockdep_is_held(&priv->lock));
999 /* need to add a new neigh, but maybe some other thread succeeded?
1000 * recalc hash, maybe hash resize took place so we do a search
1002 hash_val = ipoib_addr_hash(htbl, daddr);
1003 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1004 lockdep_is_held(&priv->lock));
1006 neigh = rcu_dereference_protected(neigh->hnext,
1007 lockdep_is_held(&priv->lock))) {
1008 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1009 /* found, take one ref on behalf of the caller */
1010 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1015 neigh->alive = jiffies;
1020 neigh = ipoib_neigh_ctor(daddr, dev);
1024 /* one ref on behalf of the hash table */
1025 atomic_inc(&neigh->refcnt);
1026 neigh->alive = jiffies;
1028 rcu_assign_pointer(neigh->hnext,
1029 rcu_dereference_protected(htbl->buckets[hash_val],
1030 lockdep_is_held(&priv->lock)));
1031 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1032 atomic_inc(&ntbl->entries);
1039 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1041 /* neigh reference count was dropprd to zero */
1042 struct net_device *dev = neigh->dev;
1043 struct ipoib_dev_priv *priv = netdev_priv(dev);
1044 struct sk_buff *skb;
1046 ipoib_put_ah(neigh->ah);
1047 while ((skb = __skb_dequeue(&neigh->queue))) {
1048 ++dev->stats.tx_dropped;
1049 dev_kfree_skb_any(skb);
1051 if (ipoib_cm_get(neigh))
1052 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1053 ipoib_dbg(netdev_priv(dev),
1054 "neigh free for %06x %pI6\n",
1055 IPOIB_QPN(neigh->daddr),
1058 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1059 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1060 complete(&priv->ntbl.flushed);
1064 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1066 /* Called as a result of removal from hash table */
1067 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1068 /* note TX context may hold another ref */
1069 ipoib_neigh_put(neigh);
1072 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1074 struct net_device *dev = neigh->dev;
1075 struct ipoib_dev_priv *priv = netdev_priv(dev);
1076 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1077 struct ipoib_neigh_hash *htbl;
1078 struct ipoib_neigh __rcu **np;
1079 struct ipoib_neigh *n;
1082 htbl = rcu_dereference_protected(ntbl->htbl,
1083 lockdep_is_held(&priv->lock));
1087 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1088 np = &htbl->buckets[hash_val];
1089 for (n = rcu_dereference_protected(*np,
1090 lockdep_is_held(&priv->lock));
1092 n = rcu_dereference_protected(*np,
1093 lockdep_is_held(&priv->lock))) {
1096 rcu_assign_pointer(*np,
1097 rcu_dereference_protected(neigh->hnext,
1098 lockdep_is_held(&priv->lock)));
1099 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1107 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1109 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1110 struct ipoib_neigh_hash *htbl;
1111 struct ipoib_neigh **buckets;
1114 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1116 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1119 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1120 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1121 buckets = kzalloc(size * sizeof(*buckets), GFP_KERNEL);
1127 htbl->mask = (size - 1);
1128 htbl->buckets = buckets;
1131 atomic_set(&ntbl->entries, 0);
1133 /* start garbage collection */
1134 clear_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1135 queue_delayed_work(ipoib_workqueue, &priv->neigh_reap_task,
1136 arp_tbl.gc_interval);
1141 static void neigh_hash_free_rcu(struct rcu_head *head)
1143 struct ipoib_neigh_hash *htbl = container_of(head,
1144 struct ipoib_neigh_hash,
1146 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1147 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1151 complete(&ntbl->deleted);
1154 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1156 struct ipoib_dev_priv *priv = netdev_priv(dev);
1157 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1158 struct ipoib_neigh_hash *htbl;
1159 unsigned long flags;
1162 /* remove all neigh connected to a given path or mcast */
1163 spin_lock_irqsave(&priv->lock, flags);
1165 htbl = rcu_dereference_protected(ntbl->htbl,
1166 lockdep_is_held(&priv->lock));
1171 for (i = 0; i < htbl->size; i++) {
1172 struct ipoib_neigh *neigh;
1173 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1175 while ((neigh = rcu_dereference_protected(*np,
1176 lockdep_is_held(&priv->lock))) != NULL) {
1177 /* delete neighs belong to this parent */
1178 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1179 rcu_assign_pointer(*np,
1180 rcu_dereference_protected(neigh->hnext,
1181 lockdep_is_held(&priv->lock)));
1182 /* remove from parent list */
1183 list_del(&neigh->list);
1184 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1192 spin_unlock_irqrestore(&priv->lock, flags);
1195 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1197 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1198 struct ipoib_neigh_hash *htbl;
1199 unsigned long flags;
1200 int i, wait_flushed = 0;
1202 init_completion(&priv->ntbl.flushed);
1204 spin_lock_irqsave(&priv->lock, flags);
1206 htbl = rcu_dereference_protected(ntbl->htbl,
1207 lockdep_is_held(&priv->lock));
1211 wait_flushed = atomic_read(&priv->ntbl.entries);
1215 for (i = 0; i < htbl->size; i++) {
1216 struct ipoib_neigh *neigh;
1217 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1219 while ((neigh = rcu_dereference_protected(*np,
1220 lockdep_is_held(&priv->lock))) != NULL) {
1221 rcu_assign_pointer(*np,
1222 rcu_dereference_protected(neigh->hnext,
1223 lockdep_is_held(&priv->lock)));
1224 /* remove from path/mc list */
1225 list_del(&neigh->list);
1226 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1231 rcu_assign_pointer(ntbl->htbl, NULL);
1232 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1235 spin_unlock_irqrestore(&priv->lock, flags);
1237 wait_for_completion(&priv->ntbl.flushed);
1240 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1242 struct ipoib_dev_priv *priv = netdev_priv(dev);
1245 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1246 init_completion(&priv->ntbl.deleted);
1247 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1249 /* Stop GC if called at init fail need to cancel work */
1250 stopped = test_and_set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1252 cancel_delayed_work(&priv->neigh_reap_task);
1254 ipoib_flush_neighs(priv);
1256 wait_for_completion(&priv->ntbl.deleted);
1260 int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
1262 struct ipoib_dev_priv *priv = netdev_priv(dev);
1264 if (ipoib_neigh_hash_init(priv) < 0)
1266 /* Allocate RX/TX "rings" to hold queued skbs */
1267 priv->rx_ring = kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,
1269 if (!priv->rx_ring) {
1270 printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
1271 ca->name, ipoib_recvq_size);
1272 goto out_neigh_hash_cleanup;
1275 priv->tx_ring = vzalloc(ipoib_sendq_size * sizeof *priv->tx_ring);
1276 if (!priv->tx_ring) {
1277 printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
1278 ca->name, ipoib_sendq_size);
1279 goto out_rx_ring_cleanup;
1282 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1284 if (ipoib_ib_dev_init(dev, ca, port))
1285 goto out_tx_ring_cleanup;
1289 out_tx_ring_cleanup:
1290 vfree(priv->tx_ring);
1292 out_rx_ring_cleanup:
1293 kfree(priv->rx_ring);
1295 out_neigh_hash_cleanup:
1296 ipoib_neigh_hash_uninit(dev);
1301 void ipoib_dev_cleanup(struct net_device *dev)
1303 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv, *tcpriv;
1308 ipoib_delete_debug_files(dev);
1310 /* Delete any child interfaces first */
1311 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs, list) {
1312 /* Stop GC on child */
1313 set_bit(IPOIB_STOP_NEIGH_GC, &cpriv->flags);
1314 cancel_delayed_work(&cpriv->neigh_reap_task);
1315 unregister_netdevice_queue(cpriv->dev, &head);
1317 unregister_netdevice_many(&head);
1319 ipoib_ib_dev_cleanup(dev);
1321 kfree(priv->rx_ring);
1322 vfree(priv->tx_ring);
1324 priv->rx_ring = NULL;
1325 priv->tx_ring = NULL;
1327 ipoib_neigh_hash_uninit(dev);
1330 static const struct header_ops ipoib_header_ops = {
1331 .create = ipoib_hard_header,
1334 static const struct net_device_ops ipoib_netdev_ops = {
1335 .ndo_uninit = ipoib_uninit,
1336 .ndo_open = ipoib_open,
1337 .ndo_stop = ipoib_stop,
1338 .ndo_change_mtu = ipoib_change_mtu,
1339 .ndo_fix_features = ipoib_fix_features,
1340 .ndo_start_xmit = ipoib_start_xmit,
1341 .ndo_tx_timeout = ipoib_timeout,
1342 .ndo_set_rx_mode = ipoib_set_mcast_list,
1345 void ipoib_setup(struct net_device *dev)
1347 struct ipoib_dev_priv *priv = netdev_priv(dev);
1349 dev->netdev_ops = &ipoib_netdev_ops;
1350 dev->header_ops = &ipoib_header_ops;
1352 ipoib_set_ethtool_ops(dev);
1354 netif_napi_add(dev, &priv->napi, ipoib_poll, 100);
1356 dev->watchdog_timeo = HZ;
1358 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
1360 dev->hard_header_len = IPOIB_ENCAP_LEN;
1361 dev->addr_len = INFINIBAND_ALEN;
1362 dev->type = ARPHRD_INFINIBAND;
1363 dev->tx_queue_len = ipoib_sendq_size * 2;
1364 dev->features = (NETIF_F_VLAN_CHALLENGED |
1366 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
1368 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
1370 netif_carrier_off(dev);
1374 spin_lock_init(&priv->lock);
1376 mutex_init(&priv->vlan_mutex);
1378 INIT_LIST_HEAD(&priv->path_list);
1379 INIT_LIST_HEAD(&priv->child_intfs);
1380 INIT_LIST_HEAD(&priv->dead_ahs);
1381 INIT_LIST_HEAD(&priv->multicast_list);
1383 INIT_DELAYED_WORK(&priv->pkey_poll_task, ipoib_pkey_poll);
1384 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
1385 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
1386 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
1387 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
1388 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
1389 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
1390 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
1391 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
1394 struct ipoib_dev_priv *ipoib_intf_alloc(const char *name)
1396 struct net_device *dev;
1398 dev = alloc_netdev((int) sizeof (struct ipoib_dev_priv), name,
1403 return netdev_priv(dev);
1406 static ssize_t show_pkey(struct device *dev,
1407 struct device_attribute *attr, char *buf)
1409 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1411 return sprintf(buf, "0x%04x\n", priv->pkey);
1413 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1415 static ssize_t show_umcast(struct device *dev,
1416 struct device_attribute *attr, char *buf)
1418 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(dev));
1420 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
1423 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
1425 struct ipoib_dev_priv *priv = netdev_priv(ndev);
1427 if (umcast_val > 0) {
1428 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1429 ipoib_warn(priv, "ignoring multicast groups joined directly "
1432 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
1435 static ssize_t set_umcast(struct device *dev,
1436 struct device_attribute *attr,
1437 const char *buf, size_t count)
1439 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
1441 ipoib_set_umcast(to_net_dev(dev), umcast_val);
1445 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
1447 int ipoib_add_umcast_attr(struct net_device *dev)
1449 return device_create_file(&dev->dev, &dev_attr_umcast);
1452 static ssize_t create_child(struct device *dev,
1453 struct device_attribute *attr,
1454 const char *buf, size_t count)
1459 if (sscanf(buf, "%i", &pkey) != 1)
1462 if (pkey < 0 || pkey > 0xffff)
1466 * Set the full membership bit, so that we join the right
1467 * broadcast group, etc.
1471 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
1473 return ret ? ret : count;
1475 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
1477 static ssize_t delete_child(struct device *dev,
1478 struct device_attribute *attr,
1479 const char *buf, size_t count)
1484 if (sscanf(buf, "%i", &pkey) != 1)
1487 if (pkey < 0 || pkey > 0xffff)
1490 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
1492 return ret ? ret : count;
1495 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
1497 int ipoib_add_pkey_attr(struct net_device *dev)
1499 return device_create_file(&dev->dev, &dev_attr_pkey);
1502 int ipoib_set_dev_features(struct ipoib_dev_priv *priv, struct ib_device *hca)
1504 struct ib_device_attr *device_attr;
1505 int result = -ENOMEM;
1507 device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
1509 printk(KERN_WARNING "%s: allocation of %zu bytes failed\n",
1510 hca->name, sizeof *device_attr);
1514 result = ib_query_device(hca, device_attr);
1516 printk(KERN_WARNING "%s: ib_query_device failed (ret = %d)\n",
1521 priv->hca_caps = device_attr->device_cap_flags;
1525 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1526 priv->dev->hw_features = NETIF_F_SG |
1527 NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1529 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1530 priv->dev->hw_features |= NETIF_F_TSO;
1532 priv->dev->features |= priv->dev->hw_features;
1538 static struct net_device *ipoib_add_port(const char *format,
1539 struct ib_device *hca, u8 port)
1541 struct ipoib_dev_priv *priv;
1542 struct ib_port_attr attr;
1543 int result = -ENOMEM;
1545 priv = ipoib_intf_alloc(format);
1547 goto alloc_mem_failed;
1549 SET_NETDEV_DEV(priv->dev, hca->dma_device);
1550 priv->dev->dev_id = port - 1;
1552 if (!ib_query_port(hca, port, &attr))
1553 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1555 printk(KERN_WARNING "%s: ib_query_port %d failed\n",
1557 goto device_init_failed;
1560 /* MTU will be reset when mcast join happens */
1561 priv->dev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1562 priv->mcast_mtu = priv->admin_mtu = priv->dev->mtu;
1564 priv->dev->neigh_priv_len = sizeof(struct ipoib_neigh);
1566 result = ib_query_pkey(hca, port, 0, &priv->pkey);
1568 printk(KERN_WARNING "%s: ib_query_pkey port %d failed (ret = %d)\n",
1569 hca->name, port, result);
1570 goto device_init_failed;
1573 if (ipoib_set_dev_features(priv, hca))
1574 goto device_init_failed;
1577 * Set the full membership bit, so that we join the right
1578 * broadcast group, etc.
1580 priv->pkey |= 0x8000;
1582 priv->dev->broadcast[8] = priv->pkey >> 8;
1583 priv->dev->broadcast[9] = priv->pkey & 0xff;
1585 result = ib_query_gid(hca, port, 0, &priv->local_gid);
1587 printk(KERN_WARNING "%s: ib_query_gid port %d failed (ret = %d)\n",
1588 hca->name, port, result);
1589 goto device_init_failed;
1591 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw, sizeof (union ib_gid));
1593 result = ipoib_dev_init(priv->dev, hca, port);
1595 printk(KERN_WARNING "%s: failed to initialize port %d (ret = %d)\n",
1596 hca->name, port, result);
1597 goto device_init_failed;
1600 INIT_IB_EVENT_HANDLER(&priv->event_handler,
1601 priv->ca, ipoib_event);
1602 result = ib_register_event_handler(&priv->event_handler);
1604 printk(KERN_WARNING "%s: ib_register_event_handler failed for "
1605 "port %d (ret = %d)\n",
1606 hca->name, port, result);
1610 result = register_netdev(priv->dev);
1612 printk(KERN_WARNING "%s: couldn't register ipoib port %d; error %d\n",
1613 hca->name, port, result);
1614 goto register_failed;
1617 ipoib_create_debug_files(priv->dev);
1619 if (ipoib_cm_add_mode_attr(priv->dev))
1621 if (ipoib_add_pkey_attr(priv->dev))
1623 if (ipoib_add_umcast_attr(priv->dev))
1625 if (device_create_file(&priv->dev->dev, &dev_attr_create_child))
1627 if (device_create_file(&priv->dev->dev, &dev_attr_delete_child))
1633 ipoib_delete_debug_files(priv->dev);
1634 unregister_netdev(priv->dev);
1637 ib_unregister_event_handler(&priv->event_handler);
1638 /* Stop GC if started before flush */
1639 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1640 cancel_delayed_work(&priv->neigh_reap_task);
1641 flush_workqueue(ipoib_workqueue);
1644 ipoib_dev_cleanup(priv->dev);
1647 free_netdev(priv->dev);
1650 return ERR_PTR(result);
1653 static void ipoib_add_one(struct ib_device *device)
1655 struct list_head *dev_list;
1656 struct net_device *dev;
1657 struct ipoib_dev_priv *priv;
1660 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1663 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1667 INIT_LIST_HEAD(dev_list);
1669 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1674 e = device->phys_port_cnt;
1677 for (p = s; p <= e; ++p) {
1678 if (rdma_port_get_link_layer(device, p) != IB_LINK_LAYER_INFINIBAND)
1680 dev = ipoib_add_port("ib%d", device, p);
1682 priv = netdev_priv(dev);
1683 list_add_tail(&priv->list, dev_list);
1687 ib_set_client_data(device, &ipoib_client, dev_list);
1690 static void ipoib_remove_one(struct ib_device *device)
1692 struct ipoib_dev_priv *priv, *tmp;
1693 struct list_head *dev_list;
1695 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
1698 dev_list = ib_get_client_data(device, &ipoib_client);
1702 list_for_each_entry_safe(priv, tmp, dev_list, list) {
1703 ib_unregister_event_handler(&priv->event_handler);
1706 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1710 set_bit(IPOIB_STOP_NEIGH_GC, &priv->flags);
1711 cancel_delayed_work(&priv->neigh_reap_task);
1712 flush_workqueue(ipoib_workqueue);
1714 unregister_netdev(priv->dev);
1715 free_netdev(priv->dev);
1721 static int __init ipoib_init_module(void)
1725 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
1726 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
1727 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
1729 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
1730 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
1731 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
1732 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1733 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
1737 * When copying small received packets, we only copy from the
1738 * linear data part of the SKB, so we rely on this condition.
1740 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
1742 ret = ipoib_register_debugfs();
1747 * We create our own workqueue mainly because we want to be
1748 * able to flush it when devices are being removed. We can't
1749 * use schedule_work()/flush_scheduled_work() because both
1750 * unregister_netdev() and linkwatch_event take the rtnl lock,
1751 * so flush_scheduled_work() can deadlock during device
1754 ipoib_workqueue = create_singlethread_workqueue("ipoib");
1755 if (!ipoib_workqueue) {
1760 ib_sa_register_client(&ipoib_sa_client);
1762 ret = ib_register_client(&ipoib_client);
1766 ret = ipoib_netlink_init();
1773 ib_unregister_client(&ipoib_client);
1776 ib_sa_unregister_client(&ipoib_sa_client);
1777 destroy_workqueue(ipoib_workqueue);
1780 ipoib_unregister_debugfs();
1785 static void __exit ipoib_cleanup_module(void)
1787 ipoib_netlink_fini();
1788 ib_unregister_client(&ipoib_client);
1789 ib_sa_unregister_client(&ipoib_sa_client);
1790 ipoib_unregister_debugfs();
1791 destroy_workqueue(ipoib_workqueue);
1794 module_init(ipoib_init_module);
1795 module_exit(ipoib_cleanup_module);