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 #define IPOIB_QPN(ha) (be32_to_cpup((__be32 *) ha) & 0xffffff)
54 MODULE_AUTHOR("Roland Dreier");
55 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
56 MODULE_LICENSE("Dual BSD/GPL");
58 int ipoib_sendq_size __read_mostly
= IPOIB_TX_RING_SIZE
;
59 int ipoib_recvq_size __read_mostly
= IPOIB_RX_RING_SIZE
;
61 module_param_named(send_queue_size
, ipoib_sendq_size
, int, 0444);
62 MODULE_PARM_DESC(send_queue_size
, "Number of descriptors in send queue");
63 module_param_named(recv_queue_size
, ipoib_recvq_size
, int, 0444);
64 MODULE_PARM_DESC(recv_queue_size
, "Number of descriptors in receive queue");
66 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
67 int ipoib_debug_level
;
69 module_param_named(debug_level
, ipoib_debug_level
, int, 0644);
70 MODULE_PARM_DESC(debug_level
, "Enable debug tracing if > 0");
73 struct ipoib_path_iter
{
74 struct net_device
*dev
;
75 struct ipoib_path path
;
78 static const u8 ipv4_bcast_addr
[] = {
79 0x00, 0xff, 0xff, 0xff,
80 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
81 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
84 struct workqueue_struct
*ipoib_workqueue
;
86 struct ib_sa_client ipoib_sa_client
;
88 static void ipoib_add_one(struct ib_device
*device
);
89 static void ipoib_remove_one(struct ib_device
*device
);
91 static struct ib_client ipoib_client
= {
94 .remove
= ipoib_remove_one
97 int ipoib_open(struct net_device
*dev
)
99 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
101 ipoib_dbg(priv
, "bringing up interface\n");
103 set_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
105 if (ipoib_pkey_dev_delay_open(dev
))
108 if (ipoib_ib_dev_open(dev
))
111 if (ipoib_ib_dev_up(dev
)) {
112 ipoib_ib_dev_stop(dev
);
116 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
117 struct ipoib_dev_priv
*cpriv
;
119 /* Bring up any child interfaces too */
120 mutex_lock(&priv
->vlan_mutex
);
121 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
124 flags
= cpriv
->dev
->flags
;
128 dev_change_flags(cpriv
->dev
, flags
| IFF_UP
);
130 mutex_unlock(&priv
->vlan_mutex
);
133 netif_start_queue(dev
);
138 static int ipoib_stop(struct net_device
*dev
)
140 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
142 ipoib_dbg(priv
, "stopping interface\n");
144 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
146 netif_stop_queue(dev
);
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
);
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 if (new_mtu
> IPOIB_PACKET_SIZE
- IPOIB_ENCAP_LEN
)
184 priv
->admin_mtu
= new_mtu
;
186 dev
->mtu
= min(priv
->mcast_mtu
, priv
->admin_mtu
);
191 static struct ipoib_path
*__path_find(struct net_device
*dev
, void *gid
)
193 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
194 struct rb_node
*n
= priv
->path_tree
.rb_node
;
195 struct ipoib_path
*path
;
199 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
201 ret
= memcmp(gid
, path
->pathrec
.dgid
.raw
,
202 sizeof (union ib_gid
));
215 static int __path_add(struct net_device
*dev
, struct ipoib_path
*path
)
217 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
218 struct rb_node
**n
= &priv
->path_tree
.rb_node
;
219 struct rb_node
*pn
= NULL
;
220 struct ipoib_path
*tpath
;
225 tpath
= rb_entry(pn
, struct ipoib_path
, rb_node
);
227 ret
= memcmp(path
->pathrec
.dgid
.raw
, tpath
->pathrec
.dgid
.raw
,
228 sizeof (union ib_gid
));
237 rb_link_node(&path
->rb_node
, pn
, n
);
238 rb_insert_color(&path
->rb_node
, &priv
->path_tree
);
240 list_add_tail(&path
->list
, &priv
->path_list
);
245 static void path_free(struct net_device
*dev
, struct ipoib_path
*path
)
247 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
248 struct ipoib_neigh
*neigh
, *tn
;
252 while ((skb
= __skb_dequeue(&path
->queue
)))
253 dev_kfree_skb_irq(skb
);
255 spin_lock_irqsave(&priv
->lock
, flags
);
257 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
259 * It's safe to call ipoib_put_ah() inside priv->lock
260 * here, because we know that path->ah will always
261 * hold one more reference, so ipoib_put_ah() will
262 * never do more than decrement the ref count.
265 ipoib_put_ah(neigh
->ah
);
267 ipoib_neigh_free(dev
, neigh
);
270 spin_unlock_irqrestore(&priv
->lock
, flags
);
273 ipoib_put_ah(path
->ah
);
278 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
280 struct ipoib_path_iter
*ipoib_path_iter_init(struct net_device
*dev
)
282 struct ipoib_path_iter
*iter
;
284 iter
= kmalloc(sizeof *iter
, GFP_KERNEL
);
289 memset(iter
->path
.pathrec
.dgid
.raw
, 0, 16);
291 if (ipoib_path_iter_next(iter
)) {
299 int ipoib_path_iter_next(struct ipoib_path_iter
*iter
)
301 struct ipoib_dev_priv
*priv
= netdev_priv(iter
->dev
);
303 struct ipoib_path
*path
;
306 spin_lock_irq(&priv
->lock
);
308 n
= rb_first(&priv
->path_tree
);
311 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
313 if (memcmp(iter
->path
.pathrec
.dgid
.raw
, path
->pathrec
.dgid
.raw
,
314 sizeof (union ib_gid
)) < 0) {
323 spin_unlock_irq(&priv
->lock
);
328 void ipoib_path_iter_read(struct ipoib_path_iter
*iter
,
329 struct ipoib_path
*path
)
334 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
336 void ipoib_flush_paths(struct net_device
*dev
)
338 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
339 struct ipoib_path
*path
, *tp
;
340 LIST_HEAD(remove_list
);
342 spin_lock_irq(&priv
->tx_lock
);
343 spin_lock(&priv
->lock
);
345 list_splice(&priv
->path_list
, &remove_list
);
346 INIT_LIST_HEAD(&priv
->path_list
);
348 list_for_each_entry(path
, &remove_list
, list
)
349 rb_erase(&path
->rb_node
, &priv
->path_tree
);
351 list_for_each_entry_safe(path
, tp
, &remove_list
, list
) {
353 ib_sa_cancel_query(path
->query_id
, path
->query
);
354 spin_unlock(&priv
->lock
);
355 spin_unlock_irq(&priv
->tx_lock
);
356 wait_for_completion(&path
->done
);
357 path_free(dev
, path
);
358 spin_lock_irq(&priv
->tx_lock
);
359 spin_lock(&priv
->lock
);
361 spin_unlock(&priv
->lock
);
362 spin_unlock_irq(&priv
->tx_lock
);
365 static void path_rec_completion(int status
,
366 struct ib_sa_path_rec
*pathrec
,
369 struct ipoib_path
*path
= path_ptr
;
370 struct net_device
*dev
= path
->dev
;
371 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
372 struct ipoib_ah
*ah
= NULL
;
373 struct ipoib_neigh
*neigh
;
374 struct sk_buff_head skqueue
;
379 ipoib_dbg(priv
, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT
"\n",
380 be16_to_cpu(pathrec
->dlid
), IPOIB_GID_ARG(pathrec
->dgid
));
382 ipoib_dbg(priv
, "PathRec status %d for GID " IPOIB_GID_FMT
"\n",
383 status
, IPOIB_GID_ARG(path
->pathrec
.dgid
));
385 skb_queue_head_init(&skqueue
);
388 struct ib_ah_attr av
= {
389 .dlid
= be16_to_cpu(pathrec
->dlid
),
391 .port_num
= priv
->port
,
392 .static_rate
= pathrec
->rate
395 ah
= ipoib_create_ah(dev
, priv
->pd
, &av
);
398 spin_lock_irqsave(&priv
->lock
, flags
);
403 path
->pathrec
= *pathrec
;
405 ipoib_dbg(priv
, "created address handle %p for LID 0x%04x, SL %d\n",
406 ah
, be16_to_cpu(pathrec
->dlid
), pathrec
->sl
);
408 while ((skb
= __skb_dequeue(&path
->queue
)))
409 __skb_queue_tail(&skqueue
, skb
);
411 list_for_each_entry(neigh
, &path
->neigh_list
, list
) {
412 kref_get(&path
->ah
->ref
);
413 neigh
->ah
= path
->ah
;
414 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
415 sizeof(union ib_gid
));
417 while ((skb
= __skb_dequeue(&neigh
->queue
)))
418 __skb_queue_tail(&skqueue
, skb
);
423 complete(&path
->done
);
425 spin_unlock_irqrestore(&priv
->lock
, flags
);
427 while ((skb
= __skb_dequeue(&skqueue
))) {
429 if (dev_queue_xmit(skb
))
430 ipoib_warn(priv
, "dev_queue_xmit failed "
431 "to requeue packet\n");
435 static struct ipoib_path
*path_rec_create(struct net_device
*dev
, void *gid
)
437 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
438 struct ipoib_path
*path
;
440 path
= kzalloc(sizeof *path
, GFP_ATOMIC
);
446 skb_queue_head_init(&path
->queue
);
448 INIT_LIST_HEAD(&path
->neigh_list
);
450 memcpy(path
->pathrec
.dgid
.raw
, gid
, sizeof (union ib_gid
));
451 path
->pathrec
.sgid
= priv
->local_gid
;
452 path
->pathrec
.pkey
= cpu_to_be16(priv
->pkey
);
453 path
->pathrec
.numb_path
= 1;
458 static int path_rec_start(struct net_device
*dev
,
459 struct ipoib_path
*path
)
461 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
463 ipoib_dbg(priv
, "Start path record lookup for " IPOIB_GID_FMT
"\n",
464 IPOIB_GID_ARG(path
->pathrec
.dgid
));
466 init_completion(&path
->done
);
469 ib_sa_path_rec_get(&ipoib_sa_client
, priv
->ca
, priv
->port
,
471 IB_SA_PATH_REC_DGID
|
472 IB_SA_PATH_REC_SGID
|
473 IB_SA_PATH_REC_NUMB_PATH
|
478 if (path
->query_id
< 0) {
479 ipoib_warn(priv
, "ib_sa_path_rec_get failed\n");
481 return path
->query_id
;
487 static void neigh_add_path(struct sk_buff
*skb
, struct net_device
*dev
)
489 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
490 struct ipoib_path
*path
;
491 struct ipoib_neigh
*neigh
;
493 neigh
= ipoib_neigh_alloc(skb
->dst
->neighbour
);
495 ++priv
->stats
.tx_dropped
;
496 dev_kfree_skb_any(skb
);
501 * We can only be called from ipoib_start_xmit, so we're
502 * inside tx_lock -- no need to save/restore flags.
504 spin_lock(&priv
->lock
);
506 path
= __path_find(dev
, skb
->dst
->neighbour
->ha
+ 4);
508 path
= path_rec_create(dev
, skb
->dst
->neighbour
->ha
+ 4);
512 __path_add(dev
, path
);
515 list_add_tail(&neigh
->list
, &path
->neigh_list
);
518 kref_get(&path
->ah
->ref
);
519 neigh
->ah
= path
->ah
;
520 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
521 sizeof(union ib_gid
));
523 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
527 if (!path
->query
&& path_rec_start(dev
, path
))
530 __skb_queue_tail(&neigh
->queue
, skb
);
533 spin_unlock(&priv
->lock
);
537 list_del(&neigh
->list
);
540 ipoib_neigh_free(dev
, neigh
);
541 ++priv
->stats
.tx_dropped
;
542 dev_kfree_skb_any(skb
);
544 spin_unlock(&priv
->lock
);
547 static void ipoib_path_lookup(struct sk_buff
*skb
, struct net_device
*dev
)
549 struct ipoib_dev_priv
*priv
= netdev_priv(skb
->dev
);
551 /* Look up path record for unicasts */
552 if (skb
->dst
->neighbour
->ha
[4] != 0xff) {
553 neigh_add_path(skb
, dev
);
557 /* Add in the P_Key for multicasts */
558 skb
->dst
->neighbour
->ha
[8] = (priv
->pkey
>> 8) & 0xff;
559 skb
->dst
->neighbour
->ha
[9] = priv
->pkey
& 0xff;
560 ipoib_mcast_send(dev
, skb
->dst
->neighbour
->ha
+ 4, skb
);
563 static void unicast_arp_send(struct sk_buff
*skb
, struct net_device
*dev
,
564 struct ipoib_pseudoheader
*phdr
)
566 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
567 struct ipoib_path
*path
;
570 * We can only be called from ipoib_start_xmit, so we're
571 * inside tx_lock -- no need to save/restore flags.
573 spin_lock(&priv
->lock
);
575 path
= __path_find(dev
, phdr
->hwaddr
+ 4);
577 path
= path_rec_create(dev
, phdr
->hwaddr
+ 4);
579 /* put pseudoheader back on for next time */
580 skb_push(skb
, sizeof *phdr
);
581 __skb_queue_tail(&path
->queue
, skb
);
583 if (path_rec_start(dev
, path
)) {
584 spin_unlock(&priv
->lock
);
585 path_free(dev
, path
);
588 __path_add(dev
, path
);
590 ++priv
->stats
.tx_dropped
;
591 dev_kfree_skb_any(skb
);
594 spin_unlock(&priv
->lock
);
599 ipoib_dbg(priv
, "Send unicast ARP to %04x\n",
600 be16_to_cpu(path
->pathrec
.dlid
));
602 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(phdr
->hwaddr
));
603 } else if ((path
->query
|| !path_rec_start(dev
, path
)) &&
604 skb_queue_len(&path
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
605 /* put pseudoheader back on for next time */
606 skb_push(skb
, sizeof *phdr
);
607 __skb_queue_tail(&path
->queue
, skb
);
609 ++priv
->stats
.tx_dropped
;
610 dev_kfree_skb_any(skb
);
613 spin_unlock(&priv
->lock
);
616 static int ipoib_start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
618 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
619 struct ipoib_neigh
*neigh
;
622 if (unlikely(!spin_trylock_irqsave(&priv
->tx_lock
, flags
)))
623 return NETDEV_TX_LOCKED
;
626 * Check if our queue is stopped. Since we have the LLTX bit
627 * set, we can't rely on netif_stop_queue() preventing our
628 * xmit function from being called with a full queue.
630 if (unlikely(netif_queue_stopped(dev
))) {
631 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
632 return NETDEV_TX_BUSY
;
635 if (likely(skb
->dst
&& skb
->dst
->neighbour
)) {
636 if (unlikely(!*to_ipoib_neigh(skb
->dst
->neighbour
))) {
637 ipoib_path_lookup(skb
, dev
);
641 neigh
= *to_ipoib_neigh(skb
->dst
->neighbour
);
643 if (likely(neigh
->ah
)) {
644 if (unlikely(memcmp(&neigh
->dgid
.raw
,
645 skb
->dst
->neighbour
->ha
+ 4,
646 sizeof(union ib_gid
)))) {
647 spin_lock(&priv
->lock
);
649 * It's safe to call ipoib_put_ah() inside
650 * priv->lock here, because we know that
651 * path->ah will always hold one more reference,
652 * so ipoib_put_ah() will never do more than
653 * decrement the ref count.
655 ipoib_put_ah(neigh
->ah
);
656 list_del(&neigh
->list
);
657 ipoib_neigh_free(dev
, neigh
);
658 spin_unlock(&priv
->lock
);
659 ipoib_path_lookup(skb
, dev
);
663 ipoib_send(dev
, skb
, neigh
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
667 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
668 spin_lock(&priv
->lock
);
669 __skb_queue_tail(&neigh
->queue
, skb
);
670 spin_unlock(&priv
->lock
);
672 ++priv
->stats
.tx_dropped
;
673 dev_kfree_skb_any(skb
);
676 struct ipoib_pseudoheader
*phdr
=
677 (struct ipoib_pseudoheader
*) skb
->data
;
678 skb_pull(skb
, sizeof *phdr
);
680 if (phdr
->hwaddr
[4] == 0xff) {
681 /* Add in the P_Key for multicast*/
682 phdr
->hwaddr
[8] = (priv
->pkey
>> 8) & 0xff;
683 phdr
->hwaddr
[9] = priv
->pkey
& 0xff;
685 ipoib_mcast_send(dev
, phdr
->hwaddr
+ 4, skb
);
687 /* unicast GID -- should be ARP or RARP reply */
689 if ((be16_to_cpup((__be16
*) skb
->data
) != ETH_P_ARP
) &&
690 (be16_to_cpup((__be16
*) skb
->data
) != ETH_P_RARP
)) {
691 ipoib_warn(priv
, "Unicast, no %s: type %04x, QPN %06x "
693 skb
->dst
? "neigh" : "dst",
694 be16_to_cpup((__be16
*) skb
->data
),
695 IPOIB_QPN(phdr
->hwaddr
),
696 IPOIB_GID_RAW_ARG(phdr
->hwaddr
+ 4));
697 dev_kfree_skb_any(skb
);
698 ++priv
->stats
.tx_dropped
;
702 unicast_arp_send(skb
, dev
, phdr
);
707 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
712 static struct net_device_stats
*ipoib_get_stats(struct net_device
*dev
)
714 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
719 static void ipoib_timeout(struct net_device
*dev
)
721 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
723 ipoib_warn(priv
, "transmit timeout: latency %d msecs\n",
724 jiffies_to_msecs(jiffies
- dev
->trans_start
));
725 ipoib_warn(priv
, "queue stopped %d, tx_head %u, tx_tail %u\n",
726 netif_queue_stopped(dev
),
727 priv
->tx_head
, priv
->tx_tail
);
728 /* XXX reset QP, etc. */
731 static int ipoib_hard_header(struct sk_buff
*skb
,
732 struct net_device
*dev
,
734 void *daddr
, void *saddr
, unsigned len
)
736 struct ipoib_header
*header
;
738 header
= (struct ipoib_header
*) skb_push(skb
, sizeof *header
);
740 header
->proto
= htons(type
);
741 header
->reserved
= 0;
744 * If we don't have a neighbour structure, stuff the
745 * destination address onto the front of the skb so we can
746 * figure out where to send the packet later.
748 if ((!skb
->dst
|| !skb
->dst
->neighbour
) && daddr
) {
749 struct ipoib_pseudoheader
*phdr
=
750 (struct ipoib_pseudoheader
*) skb_push(skb
, sizeof *phdr
);
751 memcpy(phdr
->hwaddr
, daddr
, INFINIBAND_ALEN
);
757 static void ipoib_set_mcast_list(struct net_device
*dev
)
759 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
761 if (!test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
)) {
762 ipoib_dbg(priv
, "IPOIB_FLAG_OPER_UP not set");
766 queue_work(ipoib_workqueue
, &priv
->restart_task
);
769 static void ipoib_neigh_destructor(struct neighbour
*n
)
771 struct ipoib_neigh
*neigh
;
772 struct ipoib_dev_priv
*priv
= netdev_priv(n
->dev
);
774 struct ipoib_ah
*ah
= NULL
;
777 "neigh_destructor for %06x " IPOIB_GID_FMT
"\n",
779 IPOIB_GID_RAW_ARG(n
->ha
+ 4));
781 spin_lock_irqsave(&priv
->lock
, flags
);
783 neigh
= *to_ipoib_neigh(n
);
787 list_del(&neigh
->list
);
788 ipoib_neigh_free(n
->dev
, neigh
);
791 spin_unlock_irqrestore(&priv
->lock
, flags
);
797 struct ipoib_neigh
*ipoib_neigh_alloc(struct neighbour
*neighbour
)
799 struct ipoib_neigh
*neigh
;
801 neigh
= kmalloc(sizeof *neigh
, GFP_ATOMIC
);
805 neigh
->neighbour
= neighbour
;
806 *to_ipoib_neigh(neighbour
) = neigh
;
807 skb_queue_head_init(&neigh
->queue
);
812 void ipoib_neigh_free(struct net_device
*dev
, struct ipoib_neigh
*neigh
)
814 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
816 *to_ipoib_neigh(neigh
->neighbour
) = NULL
;
817 while ((skb
= __skb_dequeue(&neigh
->queue
))) {
818 ++priv
->stats
.tx_dropped
;
819 dev_kfree_skb_any(skb
);
824 static int ipoib_neigh_setup_dev(struct net_device
*dev
, struct neigh_parms
*parms
)
826 parms
->neigh_destructor
= ipoib_neigh_destructor
;
831 int ipoib_dev_init(struct net_device
*dev
, struct ib_device
*ca
, int port
)
833 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
835 /* Allocate RX/TX "rings" to hold queued skbs */
836 priv
->rx_ring
= kzalloc(ipoib_recvq_size
* sizeof *priv
->rx_ring
,
838 if (!priv
->rx_ring
) {
839 printk(KERN_WARNING
"%s: failed to allocate RX ring (%d entries)\n",
840 ca
->name
, ipoib_recvq_size
);
844 priv
->tx_ring
= kzalloc(ipoib_sendq_size
* sizeof *priv
->tx_ring
,
846 if (!priv
->tx_ring
) {
847 printk(KERN_WARNING
"%s: failed to allocate TX ring (%d entries)\n",
848 ca
->name
, ipoib_sendq_size
);
849 goto out_rx_ring_cleanup
;
852 /* priv->tx_head & tx_tail are already 0 */
854 if (ipoib_ib_dev_init(dev
, ca
, port
))
855 goto out_tx_ring_cleanup
;
860 kfree(priv
->tx_ring
);
863 kfree(priv
->rx_ring
);
869 void ipoib_dev_cleanup(struct net_device
*dev
)
871 struct ipoib_dev_priv
*priv
= netdev_priv(dev
), *cpriv
, *tcpriv
;
873 ipoib_delete_debug_files(dev
);
875 /* Delete any child interfaces first */
876 list_for_each_entry_safe(cpriv
, tcpriv
, &priv
->child_intfs
, list
) {
877 unregister_netdev(cpriv
->dev
);
878 ipoib_dev_cleanup(cpriv
->dev
);
879 free_netdev(cpriv
->dev
);
882 ipoib_ib_dev_cleanup(dev
);
884 kfree(priv
->rx_ring
);
885 kfree(priv
->tx_ring
);
887 priv
->rx_ring
= NULL
;
888 priv
->tx_ring
= NULL
;
891 static void ipoib_setup(struct net_device
*dev
)
893 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
895 dev
->open
= ipoib_open
;
896 dev
->stop
= ipoib_stop
;
897 dev
->change_mtu
= ipoib_change_mtu
;
898 dev
->hard_start_xmit
= ipoib_start_xmit
;
899 dev
->get_stats
= ipoib_get_stats
;
900 dev
->tx_timeout
= ipoib_timeout
;
901 dev
->hard_header
= ipoib_hard_header
;
902 dev
->set_multicast_list
= ipoib_set_mcast_list
;
903 dev
->neigh_setup
= ipoib_neigh_setup_dev
;
905 dev
->watchdog_timeo
= HZ
;
907 dev
->flags
|= IFF_BROADCAST
| IFF_MULTICAST
;
910 * We add in INFINIBAND_ALEN to allow for the destination
911 * address "pseudoheader" for skbs without neighbour struct.
913 dev
->hard_header_len
= IPOIB_ENCAP_LEN
+ INFINIBAND_ALEN
;
914 dev
->addr_len
= INFINIBAND_ALEN
;
915 dev
->type
= ARPHRD_INFINIBAND
;
916 dev
->tx_queue_len
= ipoib_sendq_size
* 2;
917 dev
->features
= NETIF_F_VLAN_CHALLENGED
| NETIF_F_LLTX
;
919 /* MTU will be reset when mcast join happens */
920 dev
->mtu
= IPOIB_PACKET_SIZE
- IPOIB_ENCAP_LEN
;
921 priv
->mcast_mtu
= priv
->admin_mtu
= dev
->mtu
;
923 memcpy(dev
->broadcast
, ipv4_bcast_addr
, INFINIBAND_ALEN
);
925 netif_carrier_off(dev
);
927 SET_MODULE_OWNER(dev
);
931 spin_lock_init(&priv
->lock
);
932 spin_lock_init(&priv
->tx_lock
);
934 mutex_init(&priv
->mcast_mutex
);
935 mutex_init(&priv
->vlan_mutex
);
937 INIT_LIST_HEAD(&priv
->path_list
);
938 INIT_LIST_HEAD(&priv
->child_intfs
);
939 INIT_LIST_HEAD(&priv
->dead_ahs
);
940 INIT_LIST_HEAD(&priv
->multicast_list
);
942 INIT_DELAYED_WORK(&priv
->pkey_task
, ipoib_pkey_poll
);
943 INIT_DELAYED_WORK(&priv
->mcast_task
, ipoib_mcast_join_task
);
944 INIT_WORK(&priv
->flush_task
, ipoib_ib_dev_flush
);
945 INIT_WORK(&priv
->restart_task
, ipoib_mcast_restart_task
);
946 INIT_DELAYED_WORK(&priv
->ah_reap_task
, ipoib_reap_ah
);
949 struct ipoib_dev_priv
*ipoib_intf_alloc(const char *name
)
951 struct net_device
*dev
;
953 dev
= alloc_netdev((int) sizeof (struct ipoib_dev_priv
), name
,
958 return netdev_priv(dev
);
961 static ssize_t
show_pkey(struct device
*dev
,
962 struct device_attribute
*attr
, char *buf
)
964 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
966 return sprintf(buf
, "0x%04x\n", priv
->pkey
);
968 static DEVICE_ATTR(pkey
, S_IRUGO
, show_pkey
, NULL
);
970 static ssize_t
create_child(struct device
*dev
,
971 struct device_attribute
*attr
,
972 const char *buf
, size_t count
)
977 if (sscanf(buf
, "%i", &pkey
) != 1)
980 if (pkey
< 0 || pkey
> 0xffff)
984 * Set the full membership bit, so that we join the right
985 * broadcast group, etc.
989 ret
= ipoib_vlan_add(to_net_dev(dev
), pkey
);
991 return ret
? ret
: count
;
993 static DEVICE_ATTR(create_child
, S_IWUGO
, NULL
, create_child
);
995 static ssize_t
delete_child(struct device
*dev
,
996 struct device_attribute
*attr
,
997 const char *buf
, size_t count
)
1002 if (sscanf(buf
, "%i", &pkey
) != 1)
1005 if (pkey
< 0 || pkey
> 0xffff)
1008 ret
= ipoib_vlan_delete(to_net_dev(dev
), pkey
);
1010 return ret
? ret
: count
;
1013 static DEVICE_ATTR(delete_child
, S_IWUGO
, NULL
, delete_child
);
1015 int ipoib_add_pkey_attr(struct net_device
*dev
)
1017 return device_create_file(&dev
->dev
, &dev_attr_pkey
);
1020 static struct net_device
*ipoib_add_port(const char *format
,
1021 struct ib_device
*hca
, u8 port
)
1023 struct ipoib_dev_priv
*priv
;
1024 int result
= -ENOMEM
;
1026 priv
= ipoib_intf_alloc(format
);
1028 goto alloc_mem_failed
;
1030 SET_NETDEV_DEV(priv
->dev
, hca
->dma_device
);
1032 result
= ib_query_pkey(hca
, port
, 0, &priv
->pkey
);
1034 printk(KERN_WARNING
"%s: ib_query_pkey port %d failed (ret = %d)\n",
1035 hca
->name
, port
, result
);
1036 goto alloc_mem_failed
;
1040 * Set the full membership bit, so that we join the right
1041 * broadcast group, etc.
1043 priv
->pkey
|= 0x8000;
1045 priv
->dev
->broadcast
[8] = priv
->pkey
>> 8;
1046 priv
->dev
->broadcast
[9] = priv
->pkey
& 0xff;
1048 result
= ib_query_gid(hca
, port
, 0, &priv
->local_gid
);
1050 printk(KERN_WARNING
"%s: ib_query_gid port %d failed (ret = %d)\n",
1051 hca
->name
, port
, result
);
1052 goto alloc_mem_failed
;
1054 memcpy(priv
->dev
->dev_addr
+ 4, priv
->local_gid
.raw
, sizeof (union ib_gid
));
1057 result
= ipoib_dev_init(priv
->dev
, hca
, port
);
1059 printk(KERN_WARNING
"%s: failed to initialize port %d (ret = %d)\n",
1060 hca
->name
, port
, result
);
1061 goto device_init_failed
;
1064 INIT_IB_EVENT_HANDLER(&priv
->event_handler
,
1065 priv
->ca
, ipoib_event
);
1066 result
= ib_register_event_handler(&priv
->event_handler
);
1068 printk(KERN_WARNING
"%s: ib_register_event_handler failed for "
1069 "port %d (ret = %d)\n",
1070 hca
->name
, port
, result
);
1074 result
= register_netdev(priv
->dev
);
1076 printk(KERN_WARNING
"%s: couldn't register ipoib port %d; error %d\n",
1077 hca
->name
, port
, result
);
1078 goto register_failed
;
1081 ipoib_create_debug_files(priv
->dev
);
1083 if (ipoib_add_pkey_attr(priv
->dev
))
1085 if (device_create_file(&priv
->dev
->dev
, &dev_attr_create_child
))
1087 if (device_create_file(&priv
->dev
->dev
, &dev_attr_delete_child
))
1093 ipoib_delete_debug_files(priv
->dev
);
1094 unregister_netdev(priv
->dev
);
1097 ib_unregister_event_handler(&priv
->event_handler
);
1098 flush_scheduled_work();
1101 ipoib_dev_cleanup(priv
->dev
);
1104 free_netdev(priv
->dev
);
1107 return ERR_PTR(result
);
1110 static void ipoib_add_one(struct ib_device
*device
)
1112 struct list_head
*dev_list
;
1113 struct net_device
*dev
;
1114 struct ipoib_dev_priv
*priv
;
1117 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1120 dev_list
= kmalloc(sizeof *dev_list
, GFP_KERNEL
);
1124 INIT_LIST_HEAD(dev_list
);
1126 if (device
->node_type
== RDMA_NODE_IB_SWITCH
) {
1131 e
= device
->phys_port_cnt
;
1134 for (p
= s
; p
<= e
; ++p
) {
1135 dev
= ipoib_add_port("ib%d", device
, p
);
1137 priv
= netdev_priv(dev
);
1138 list_add_tail(&priv
->list
, dev_list
);
1142 ib_set_client_data(device
, &ipoib_client
, dev_list
);
1145 static void ipoib_remove_one(struct ib_device
*device
)
1147 struct ipoib_dev_priv
*priv
, *tmp
;
1148 struct list_head
*dev_list
;
1150 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1153 dev_list
= ib_get_client_data(device
, &ipoib_client
);
1155 list_for_each_entry_safe(priv
, tmp
, dev_list
, list
) {
1156 ib_unregister_event_handler(&priv
->event_handler
);
1157 flush_scheduled_work();
1159 unregister_netdev(priv
->dev
);
1160 ipoib_dev_cleanup(priv
->dev
);
1161 free_netdev(priv
->dev
);
1167 static int __init
ipoib_init_module(void)
1171 ipoib_recvq_size
= roundup_pow_of_two(ipoib_recvq_size
);
1172 ipoib_recvq_size
= min(ipoib_recvq_size
, IPOIB_MAX_QUEUE_SIZE
);
1173 ipoib_recvq_size
= max(ipoib_recvq_size
, IPOIB_MIN_QUEUE_SIZE
);
1175 ipoib_sendq_size
= roundup_pow_of_two(ipoib_sendq_size
);
1176 ipoib_sendq_size
= min(ipoib_sendq_size
, IPOIB_MAX_QUEUE_SIZE
);
1177 ipoib_sendq_size
= max(ipoib_sendq_size
, IPOIB_MIN_QUEUE_SIZE
);
1179 ret
= ipoib_register_debugfs();
1184 * We create our own workqueue mainly because we want to be
1185 * able to flush it when devices are being removed. We can't
1186 * use schedule_work()/flush_scheduled_work() because both
1187 * unregister_netdev() and linkwatch_event take the rtnl lock,
1188 * so flush_scheduled_work() can deadlock during device
1191 ipoib_workqueue
= create_singlethread_workqueue("ipoib");
1192 if (!ipoib_workqueue
) {
1197 ib_sa_register_client(&ipoib_sa_client
);
1199 ret
= ib_register_client(&ipoib_client
);
1206 ib_sa_unregister_client(&ipoib_sa_client
);
1207 destroy_workqueue(ipoib_workqueue
);
1210 ipoib_unregister_debugfs();
1215 static void __exit
ipoib_cleanup_module(void)
1217 ib_unregister_client(&ipoib_client
);
1218 ib_sa_unregister_client(&ipoib_sa_client
);
1219 ipoib_unregister_debugfs();
1220 destroy_workqueue(ipoib_workqueue
);
1223 module_init(ipoib_init_module
);
1224 module_exit(ipoib_cleanup_module
);