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 */
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
53 MODULE_LICENSE("Dual BSD/GPL");
55 int ipoib_sendq_size __read_mostly
= IPOIB_TX_RING_SIZE
;
56 int ipoib_recvq_size __read_mostly
= IPOIB_RX_RING_SIZE
;
58 module_param_named(send_queue_size
, ipoib_sendq_size
, int, 0444);
59 MODULE_PARM_DESC(send_queue_size
, "Number of descriptors in send queue");
60 module_param_named(recv_queue_size
, ipoib_recvq_size
, int, 0444);
61 MODULE_PARM_DESC(recv_queue_size
, "Number of descriptors in receive queue");
63 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
64 int ipoib_debug_level
;
66 module_param_named(debug_level
, ipoib_debug_level
, int, 0644);
67 MODULE_PARM_DESC(debug_level
, "Enable debug tracing if > 0");
70 struct ipoib_path_iter
{
71 struct net_device
*dev
;
72 struct ipoib_path path
;
75 static const u8 ipv4_bcast_addr
[] = {
76 0x00, 0xff, 0xff, 0xff,
77 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
78 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
81 struct workqueue_struct
*ipoib_workqueue
;
83 struct ib_sa_client ipoib_sa_client
;
85 static void ipoib_add_one(struct ib_device
*device
);
86 static void ipoib_remove_one(struct ib_device
*device
);
88 static struct ib_client ipoib_client
= {
91 .remove
= ipoib_remove_one
94 int ipoib_open(struct net_device
*dev
)
96 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
98 ipoib_dbg(priv
, "bringing up interface\n");
100 napi_enable(&priv
->napi
);
101 set_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
103 if (ipoib_pkey_dev_delay_open(dev
))
106 if (ipoib_ib_dev_open(dev
)) {
107 napi_disable(&priv
->napi
);
111 if (ipoib_ib_dev_up(dev
)) {
112 ipoib_ib_dev_stop(dev
, 1);
113 napi_disable(&priv
->napi
);
117 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
118 struct ipoib_dev_priv
*cpriv
;
120 /* Bring up any child interfaces too */
121 mutex_lock(&priv
->vlan_mutex
);
122 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
125 flags
= cpriv
->dev
->flags
;
129 dev_change_flags(cpriv
->dev
, flags
| IFF_UP
);
131 mutex_unlock(&priv
->vlan_mutex
);
134 netif_start_queue(dev
);
139 static int ipoib_stop(struct net_device
*dev
)
141 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
143 ipoib_dbg(priv
, "stopping interface\n");
145 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
146 napi_disable(&priv
->napi
);
148 netif_stop_queue(dev
);
151 * Now flush workqueue to make sure a scheduled task doesn't
152 * bring our internal state back up.
154 flush_workqueue(ipoib_workqueue
);
156 ipoib_ib_dev_down(dev
, 1);
157 ipoib_ib_dev_stop(dev
, 1);
159 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
160 struct ipoib_dev_priv
*cpriv
;
162 /* Bring down any child interfaces too */
163 mutex_lock(&priv
->vlan_mutex
);
164 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
167 flags
= cpriv
->dev
->flags
;
168 if (!(flags
& IFF_UP
))
171 dev_change_flags(cpriv
->dev
, flags
& ~IFF_UP
);
173 mutex_unlock(&priv
->vlan_mutex
);
179 static int ipoib_change_mtu(struct net_device
*dev
, int new_mtu
)
181 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
183 /* dev->mtu > 2K ==> connected mode */
184 if (ipoib_cm_admin_enabled(dev
)) {
185 if (new_mtu
> ipoib_cm_max_mtu(dev
))
188 if (new_mtu
> priv
->mcast_mtu
)
189 ipoib_warn(priv
, "mtu > %d will cause multicast packet drops.\n",
196 if (new_mtu
> IPOIB_UD_MTU(priv
->max_ib_mtu
))
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_init(&priv
->path_list
, &remove_list
);
362 list_for_each_entry(path
, &remove_list
, list
)
363 rb_erase(&path
->rb_node
, &priv
->path_tree
);
365 list_for_each_entry_safe(path
, tp
, &remove_list
, list
) {
367 ib_sa_cancel_query(path
->query_id
, path
->query
);
368 spin_unlock(&priv
->lock
);
369 spin_unlock_irq(&priv
->tx_lock
);
370 wait_for_completion(&path
->done
);
371 path_free(dev
, path
);
372 spin_lock_irq(&priv
->tx_lock
);
373 spin_lock(&priv
->lock
);
375 spin_unlock(&priv
->lock
);
376 spin_unlock_irq(&priv
->tx_lock
);
379 static void path_rec_completion(int status
,
380 struct ib_sa_path_rec
*pathrec
,
383 struct ipoib_path
*path
= path_ptr
;
384 struct net_device
*dev
= path
->dev
;
385 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
386 struct ipoib_ah
*ah
= NULL
;
387 struct ipoib_neigh
*neigh
, *tn
;
388 struct sk_buff_head skqueue
;
393 ipoib_dbg(priv
, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT
"\n",
394 be16_to_cpu(pathrec
->dlid
), IPOIB_GID_ARG(pathrec
->dgid
));
396 ipoib_dbg(priv
, "PathRec status %d for GID " IPOIB_GID_FMT
"\n",
397 status
, IPOIB_GID_ARG(path
->pathrec
.dgid
));
399 skb_queue_head_init(&skqueue
);
402 struct ib_ah_attr av
;
404 if (!ib_init_ah_from_path(priv
->ca
, priv
->port
, pathrec
, &av
))
405 ah
= ipoib_create_ah(dev
, priv
->pd
, &av
);
408 spin_lock_irqsave(&priv
->lock
, flags
);
413 path
->pathrec
= *pathrec
;
415 ipoib_dbg(priv
, "created address handle %p for LID 0x%04x, SL %d\n",
416 ah
, be16_to_cpu(pathrec
->dlid
), pathrec
->sl
);
418 while ((skb
= __skb_dequeue(&path
->queue
)))
419 __skb_queue_tail(&skqueue
, skb
);
421 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
422 kref_get(&path
->ah
->ref
);
423 neigh
->ah
= path
->ah
;
424 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
425 sizeof(union ib_gid
));
427 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
428 if (!ipoib_cm_get(neigh
))
429 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
,
432 if (!ipoib_cm_get(neigh
)) {
433 list_del(&neigh
->list
);
435 ipoib_put_ah(neigh
->ah
);
436 ipoib_neigh_free(dev
, neigh
);
441 while ((skb
= __skb_dequeue(&neigh
->queue
)))
442 __skb_queue_tail(&skqueue
, skb
);
447 complete(&path
->done
);
449 spin_unlock_irqrestore(&priv
->lock
, flags
);
451 while ((skb
= __skb_dequeue(&skqueue
))) {
453 if (dev_queue_xmit(skb
))
454 ipoib_warn(priv
, "dev_queue_xmit failed "
455 "to requeue packet\n");
459 static struct ipoib_path
*path_rec_create(struct net_device
*dev
, void *gid
)
461 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
462 struct ipoib_path
*path
;
464 if (!priv
->broadcast
)
467 path
= kzalloc(sizeof *path
, GFP_ATOMIC
);
473 skb_queue_head_init(&path
->queue
);
475 INIT_LIST_HEAD(&path
->neigh_list
);
477 memcpy(path
->pathrec
.dgid
.raw
, gid
, sizeof (union ib_gid
));
478 path
->pathrec
.sgid
= priv
->local_gid
;
479 path
->pathrec
.pkey
= cpu_to_be16(priv
->pkey
);
480 path
->pathrec
.numb_path
= 1;
481 path
->pathrec
.traffic_class
= priv
->broadcast
->mcmember
.traffic_class
;
486 static int path_rec_start(struct net_device
*dev
,
487 struct ipoib_path
*path
)
489 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
491 ipoib_dbg(priv
, "Start path record lookup for " IPOIB_GID_FMT
"\n",
492 IPOIB_GID_ARG(path
->pathrec
.dgid
));
494 init_completion(&path
->done
);
497 ib_sa_path_rec_get(&ipoib_sa_client
, priv
->ca
, priv
->port
,
499 IB_SA_PATH_REC_DGID
|
500 IB_SA_PATH_REC_SGID
|
501 IB_SA_PATH_REC_NUMB_PATH
|
502 IB_SA_PATH_REC_TRAFFIC_CLASS
|
507 if (path
->query_id
< 0) {
508 ipoib_warn(priv
, "ib_sa_path_rec_get failed\n");
510 return path
->query_id
;
516 static void neigh_add_path(struct sk_buff
*skb
, struct net_device
*dev
)
518 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
519 struct ipoib_path
*path
;
520 struct ipoib_neigh
*neigh
;
522 neigh
= ipoib_neigh_alloc(skb
->dst
->neighbour
, skb
->dev
);
524 ++dev
->stats
.tx_dropped
;
525 dev_kfree_skb_any(skb
);
530 * We can only be called from ipoib_start_xmit, so we're
531 * inside tx_lock -- no need to save/restore flags.
533 spin_lock(&priv
->lock
);
535 path
= __path_find(dev
, skb
->dst
->neighbour
->ha
+ 4);
537 path
= path_rec_create(dev
, skb
->dst
->neighbour
->ha
+ 4);
541 __path_add(dev
, path
);
544 list_add_tail(&neigh
->list
, &path
->neigh_list
);
547 kref_get(&path
->ah
->ref
);
548 neigh
->ah
= path
->ah
;
549 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
550 sizeof(union ib_gid
));
552 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
553 if (!ipoib_cm_get(neigh
))
554 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
, path
, neigh
));
555 if (!ipoib_cm_get(neigh
)) {
556 list_del(&neigh
->list
);
558 ipoib_put_ah(neigh
->ah
);
559 ipoib_neigh_free(dev
, neigh
);
562 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
)
563 __skb_queue_tail(&neigh
->queue
, skb
);
565 ipoib_warn(priv
, "queue length limit %d. Packet drop.\n",
566 skb_queue_len(&neigh
->queue
));
570 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
574 if (!path
->query
&& path_rec_start(dev
, path
))
577 __skb_queue_tail(&neigh
->queue
, skb
);
580 spin_unlock(&priv
->lock
);
584 list_del(&neigh
->list
);
587 ipoib_neigh_free(dev
, neigh
);
589 ++dev
->stats
.tx_dropped
;
590 dev_kfree_skb_any(skb
);
592 spin_unlock(&priv
->lock
);
595 static void ipoib_path_lookup(struct sk_buff
*skb
, struct net_device
*dev
)
597 struct ipoib_dev_priv
*priv
= netdev_priv(skb
->dev
);
599 /* Look up path record for unicasts */
600 if (skb
->dst
->neighbour
->ha
[4] != 0xff) {
601 neigh_add_path(skb
, dev
);
605 /* Add in the P_Key for multicasts */
606 skb
->dst
->neighbour
->ha
[8] = (priv
->pkey
>> 8) & 0xff;
607 skb
->dst
->neighbour
->ha
[9] = priv
->pkey
& 0xff;
608 ipoib_mcast_send(dev
, skb
->dst
->neighbour
->ha
+ 4, skb
);
611 static void unicast_arp_send(struct sk_buff
*skb
, struct net_device
*dev
,
612 struct ipoib_pseudoheader
*phdr
)
614 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
615 struct ipoib_path
*path
;
618 * We can only be called from ipoib_start_xmit, so we're
619 * inside tx_lock -- no need to save/restore flags.
621 spin_lock(&priv
->lock
);
623 path
= __path_find(dev
, phdr
->hwaddr
+ 4);
625 path
= path_rec_create(dev
, phdr
->hwaddr
+ 4);
627 /* put pseudoheader back on for next time */
628 skb_push(skb
, sizeof *phdr
);
629 __skb_queue_tail(&path
->queue
, skb
);
631 if (path_rec_start(dev
, path
)) {
632 spin_unlock(&priv
->lock
);
633 path_free(dev
, path
);
636 __path_add(dev
, path
);
638 ++dev
->stats
.tx_dropped
;
639 dev_kfree_skb_any(skb
);
642 spin_unlock(&priv
->lock
);
647 ipoib_dbg(priv
, "Send unicast ARP to %04x\n",
648 be16_to_cpu(path
->pathrec
.dlid
));
650 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(phdr
->hwaddr
));
651 } else if ((path
->query
|| !path_rec_start(dev
, path
)) &&
652 skb_queue_len(&path
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
653 /* put pseudoheader back on for next time */
654 skb_push(skb
, sizeof *phdr
);
655 __skb_queue_tail(&path
->queue
, skb
);
657 ++dev
->stats
.tx_dropped
;
658 dev_kfree_skb_any(skb
);
661 spin_unlock(&priv
->lock
);
664 static int ipoib_start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
666 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
667 struct ipoib_neigh
*neigh
;
670 if (unlikely(!spin_trylock_irqsave(&priv
->tx_lock
, flags
)))
671 return NETDEV_TX_LOCKED
;
673 if (likely(skb
->dst
&& skb
->dst
->neighbour
)) {
674 if (unlikely(!*to_ipoib_neigh(skb
->dst
->neighbour
))) {
675 ipoib_path_lookup(skb
, dev
);
679 neigh
= *to_ipoib_neigh(skb
->dst
->neighbour
);
682 if (unlikely((memcmp(&neigh
->dgid
.raw
,
683 skb
->dst
->neighbour
->ha
+ 4,
684 sizeof(union ib_gid
))) ||
685 (neigh
->dev
!= dev
))) {
686 spin_lock(&priv
->lock
);
688 * It's safe to call ipoib_put_ah() inside
689 * priv->lock here, because we know that
690 * path->ah will always hold one more reference,
691 * so ipoib_put_ah() will never do more than
692 * decrement the ref count.
694 ipoib_put_ah(neigh
->ah
);
695 list_del(&neigh
->list
);
696 ipoib_neigh_free(dev
, neigh
);
697 spin_unlock(&priv
->lock
);
698 ipoib_path_lookup(skb
, dev
);
702 if (ipoib_cm_get(neigh
)) {
703 if (ipoib_cm_up(neigh
)) {
704 ipoib_cm_send(dev
, skb
, ipoib_cm_get(neigh
));
707 } else if (neigh
->ah
) {
708 ipoib_send(dev
, skb
, neigh
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
712 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
713 spin_lock(&priv
->lock
);
714 __skb_queue_tail(&neigh
->queue
, skb
);
715 spin_unlock(&priv
->lock
);
717 ++dev
->stats
.tx_dropped
;
718 dev_kfree_skb_any(skb
);
721 struct ipoib_pseudoheader
*phdr
=
722 (struct ipoib_pseudoheader
*) skb
->data
;
723 skb_pull(skb
, sizeof *phdr
);
725 if (phdr
->hwaddr
[4] == 0xff) {
726 /* Add in the P_Key for multicast*/
727 phdr
->hwaddr
[8] = (priv
->pkey
>> 8) & 0xff;
728 phdr
->hwaddr
[9] = priv
->pkey
& 0xff;
730 ipoib_mcast_send(dev
, phdr
->hwaddr
+ 4, skb
);
732 /* unicast GID -- should be ARP or RARP reply */
734 if ((be16_to_cpup((__be16
*) skb
->data
) != ETH_P_ARP
) &&
735 (be16_to_cpup((__be16
*) skb
->data
) != ETH_P_RARP
)) {
736 ipoib_warn(priv
, "Unicast, no %s: type %04x, QPN %06x "
738 skb
->dst
? "neigh" : "dst",
739 be16_to_cpup((__be16
*) skb
->data
),
740 IPOIB_QPN(phdr
->hwaddr
),
741 IPOIB_GID_RAW_ARG(phdr
->hwaddr
+ 4));
742 dev_kfree_skb_any(skb
);
743 ++dev
->stats
.tx_dropped
;
747 unicast_arp_send(skb
, dev
, phdr
);
752 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
757 static void ipoib_timeout(struct net_device
*dev
)
759 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
761 ipoib_warn(priv
, "transmit timeout: latency %d msecs\n",
762 jiffies_to_msecs(jiffies
- dev
->trans_start
));
763 ipoib_warn(priv
, "queue stopped %d, tx_head %u, tx_tail %u\n",
764 netif_queue_stopped(dev
),
765 priv
->tx_head
, priv
->tx_tail
);
766 /* XXX reset QP, etc. */
769 static int ipoib_hard_header(struct sk_buff
*skb
,
770 struct net_device
*dev
,
772 const void *daddr
, const void *saddr
, unsigned len
)
774 struct ipoib_header
*header
;
776 header
= (struct ipoib_header
*) skb_push(skb
, sizeof *header
);
778 header
->proto
= htons(type
);
779 header
->reserved
= 0;
782 * If we don't have a neighbour structure, stuff the
783 * destination address onto the front of the skb so we can
784 * figure out where to send the packet later.
786 if ((!skb
->dst
|| !skb
->dst
->neighbour
) && daddr
) {
787 struct ipoib_pseudoheader
*phdr
=
788 (struct ipoib_pseudoheader
*) skb_push(skb
, sizeof *phdr
);
789 memcpy(phdr
->hwaddr
, daddr
, INFINIBAND_ALEN
);
795 static void ipoib_set_mcast_list(struct net_device
*dev
)
797 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
799 if (!test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
)) {
800 ipoib_dbg(priv
, "IPOIB_FLAG_OPER_UP not set");
804 queue_work(ipoib_workqueue
, &priv
->restart_task
);
807 static void ipoib_neigh_cleanup(struct neighbour
*n
)
809 struct ipoib_neigh
*neigh
;
810 struct ipoib_dev_priv
*priv
= netdev_priv(n
->dev
);
812 struct ipoib_ah
*ah
= NULL
;
814 neigh
= *to_ipoib_neigh(n
);
816 priv
= netdev_priv(neigh
->dev
);
820 "neigh_cleanup for %06x " IPOIB_GID_FMT
"\n",
822 IPOIB_GID_RAW_ARG(n
->ha
+ 4));
824 spin_lock_irqsave(&priv
->lock
, flags
);
828 list_del(&neigh
->list
);
829 ipoib_neigh_free(n
->dev
, neigh
);
831 spin_unlock_irqrestore(&priv
->lock
, flags
);
837 struct ipoib_neigh
*ipoib_neigh_alloc(struct neighbour
*neighbour
,
838 struct net_device
*dev
)
840 struct ipoib_neigh
*neigh
;
842 neigh
= kmalloc(sizeof *neigh
, GFP_ATOMIC
);
846 neigh
->neighbour
= neighbour
;
848 *to_ipoib_neigh(neighbour
) = neigh
;
849 skb_queue_head_init(&neigh
->queue
);
850 ipoib_cm_set(neigh
, NULL
);
855 void ipoib_neigh_free(struct net_device
*dev
, struct ipoib_neigh
*neigh
)
858 *to_ipoib_neigh(neigh
->neighbour
) = NULL
;
859 while ((skb
= __skb_dequeue(&neigh
->queue
))) {
860 ++dev
->stats
.tx_dropped
;
861 dev_kfree_skb_any(skb
);
863 if (ipoib_cm_get(neigh
))
864 ipoib_cm_destroy_tx(ipoib_cm_get(neigh
));
868 static int ipoib_neigh_setup_dev(struct net_device
*dev
, struct neigh_parms
*parms
)
870 parms
->neigh_cleanup
= ipoib_neigh_cleanup
;
875 int ipoib_dev_init(struct net_device
*dev
, struct ib_device
*ca
, int port
)
877 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
879 /* Allocate RX/TX "rings" to hold queued skbs */
880 priv
->rx_ring
= kzalloc(ipoib_recvq_size
* sizeof *priv
->rx_ring
,
882 if (!priv
->rx_ring
) {
883 printk(KERN_WARNING
"%s: failed to allocate RX ring (%d entries)\n",
884 ca
->name
, ipoib_recvq_size
);
888 priv
->tx_ring
= vmalloc(ipoib_sendq_size
* sizeof *priv
->tx_ring
);
889 if (!priv
->tx_ring
) {
890 printk(KERN_WARNING
"%s: failed to allocate TX ring (%d entries)\n",
891 ca
->name
, ipoib_sendq_size
);
892 goto out_rx_ring_cleanup
;
894 memset(priv
->tx_ring
, 0, ipoib_sendq_size
* sizeof *priv
->tx_ring
);
896 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
898 if (ipoib_ib_dev_init(dev
, ca
, port
))
899 goto out_tx_ring_cleanup
;
904 vfree(priv
->tx_ring
);
907 kfree(priv
->rx_ring
);
913 void ipoib_dev_cleanup(struct net_device
*dev
)
915 struct ipoib_dev_priv
*priv
= netdev_priv(dev
), *cpriv
, *tcpriv
;
917 ipoib_delete_debug_files(dev
);
919 /* Delete any child interfaces first */
920 list_for_each_entry_safe(cpriv
, tcpriv
, &priv
->child_intfs
, list
) {
921 unregister_netdev(cpriv
->dev
);
922 ipoib_dev_cleanup(cpriv
->dev
);
923 free_netdev(cpriv
->dev
);
926 ipoib_ib_dev_cleanup(dev
);
928 kfree(priv
->rx_ring
);
929 vfree(priv
->tx_ring
);
931 priv
->rx_ring
= NULL
;
932 priv
->tx_ring
= NULL
;
935 static const struct header_ops ipoib_header_ops
= {
936 .create
= ipoib_hard_header
,
939 static void ipoib_setup(struct net_device
*dev
)
941 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
943 dev
->open
= ipoib_open
;
944 dev
->stop
= ipoib_stop
;
945 dev
->change_mtu
= ipoib_change_mtu
;
946 dev
->hard_start_xmit
= ipoib_start_xmit
;
947 dev
->tx_timeout
= ipoib_timeout
;
948 dev
->header_ops
= &ipoib_header_ops
;
949 dev
->set_multicast_list
= ipoib_set_mcast_list
;
950 dev
->neigh_setup
= ipoib_neigh_setup_dev
;
952 ipoib_set_ethtool_ops(dev
);
954 netif_napi_add(dev
, &priv
->napi
, ipoib_poll
, 100);
956 dev
->watchdog_timeo
= HZ
;
958 dev
->flags
|= IFF_BROADCAST
| IFF_MULTICAST
;
961 * We add in INFINIBAND_ALEN to allow for the destination
962 * address "pseudoheader" for skbs without neighbour struct.
964 dev
->hard_header_len
= IPOIB_ENCAP_LEN
+ INFINIBAND_ALEN
;
965 dev
->addr_len
= INFINIBAND_ALEN
;
966 dev
->type
= ARPHRD_INFINIBAND
;
967 dev
->tx_queue_len
= ipoib_sendq_size
* 2;
968 dev
->features
= (NETIF_F_VLAN_CHALLENGED
|
972 memcpy(dev
->broadcast
, ipv4_bcast_addr
, INFINIBAND_ALEN
);
974 netif_carrier_off(dev
);
978 spin_lock_init(&priv
->lock
);
979 spin_lock_init(&priv
->tx_lock
);
981 mutex_init(&priv
->mcast_mutex
);
982 mutex_init(&priv
->vlan_mutex
);
984 INIT_LIST_HEAD(&priv
->path_list
);
985 INIT_LIST_HEAD(&priv
->child_intfs
);
986 INIT_LIST_HEAD(&priv
->dead_ahs
);
987 INIT_LIST_HEAD(&priv
->multicast_list
);
989 INIT_DELAYED_WORK(&priv
->pkey_poll_task
, ipoib_pkey_poll
);
990 INIT_WORK(&priv
->pkey_event_task
, ipoib_pkey_event
);
991 INIT_DELAYED_WORK(&priv
->mcast_task
, ipoib_mcast_join_task
);
992 INIT_WORK(&priv
->flush_task
, ipoib_ib_dev_flush
);
993 INIT_WORK(&priv
->restart_task
, ipoib_mcast_restart_task
);
994 INIT_DELAYED_WORK(&priv
->ah_reap_task
, ipoib_reap_ah
);
997 struct ipoib_dev_priv
*ipoib_intf_alloc(const char *name
)
999 struct net_device
*dev
;
1001 dev
= alloc_netdev((int) sizeof (struct ipoib_dev_priv
), name
,
1006 return netdev_priv(dev
);
1009 static ssize_t
show_pkey(struct device
*dev
,
1010 struct device_attribute
*attr
, char *buf
)
1012 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1014 return sprintf(buf
, "0x%04x\n", priv
->pkey
);
1016 static DEVICE_ATTR(pkey
, S_IRUGO
, show_pkey
, NULL
);
1018 static ssize_t
show_umcast(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
, "%d\n", test_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
));
1026 static ssize_t
set_umcast(struct device
*dev
,
1027 struct device_attribute
*attr
,
1028 const char *buf
, size_t count
)
1030 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1031 unsigned long umcast_val
= simple_strtoul(buf
, NULL
, 0);
1033 if (umcast_val
> 0) {
1034 set_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1035 ipoib_warn(priv
, "ignoring multicast groups joined directly "
1038 clear_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1042 static DEVICE_ATTR(umcast
, S_IWUSR
| S_IRUGO
, show_umcast
, set_umcast
);
1044 int ipoib_add_umcast_attr(struct net_device
*dev
)
1046 return device_create_file(&dev
->dev
, &dev_attr_umcast
);
1049 static ssize_t
create_child(struct device
*dev
,
1050 struct device_attribute
*attr
,
1051 const char *buf
, size_t count
)
1056 if (sscanf(buf
, "%i", &pkey
) != 1)
1059 if (pkey
< 0 || pkey
> 0xffff)
1063 * Set the full membership bit, so that we join the right
1064 * broadcast group, etc.
1068 ret
= ipoib_vlan_add(to_net_dev(dev
), pkey
);
1070 return ret
? ret
: count
;
1072 static DEVICE_ATTR(create_child
, S_IWUGO
, NULL
, create_child
);
1074 static ssize_t
delete_child(struct device
*dev
,
1075 struct device_attribute
*attr
,
1076 const char *buf
, size_t count
)
1081 if (sscanf(buf
, "%i", &pkey
) != 1)
1084 if (pkey
< 0 || pkey
> 0xffff)
1087 ret
= ipoib_vlan_delete(to_net_dev(dev
), pkey
);
1089 return ret
? ret
: count
;
1092 static DEVICE_ATTR(delete_child
, S_IWUGO
, NULL
, delete_child
);
1094 int ipoib_add_pkey_attr(struct net_device
*dev
)
1096 return device_create_file(&dev
->dev
, &dev_attr_pkey
);
1099 static struct net_device
*ipoib_add_port(const char *format
,
1100 struct ib_device
*hca
, u8 port
)
1102 struct ipoib_dev_priv
*priv
;
1103 struct ib_device_attr
*device_attr
;
1104 struct ib_port_attr attr
;
1105 int result
= -ENOMEM
;
1107 priv
= ipoib_intf_alloc(format
);
1109 goto alloc_mem_failed
;
1111 SET_NETDEV_DEV(priv
->dev
, hca
->dma_device
);
1113 if (!ib_query_port(hca
, port
, &attr
))
1114 priv
->max_ib_mtu
= ib_mtu_enum_to_int(attr
.max_mtu
);
1116 printk(KERN_WARNING
"%s: ib_query_port %d failed\n",
1118 goto device_init_failed
;
1121 /* MTU will be reset when mcast join happens */
1122 priv
->dev
->mtu
= IPOIB_UD_MTU(priv
->max_ib_mtu
);
1123 priv
->mcast_mtu
= priv
->admin_mtu
= priv
->dev
->mtu
;
1125 result
= ib_query_pkey(hca
, port
, 0, &priv
->pkey
);
1127 printk(KERN_WARNING
"%s: ib_query_pkey port %d failed (ret = %d)\n",
1128 hca
->name
, port
, result
);
1129 goto device_init_failed
;
1132 device_attr
= kmalloc(sizeof *device_attr
, GFP_KERNEL
);
1134 printk(KERN_WARNING
"%s: allocation of %zu bytes failed\n",
1135 hca
->name
, sizeof *device_attr
);
1136 goto device_init_failed
;
1139 result
= ib_query_device(hca
, device_attr
);
1141 printk(KERN_WARNING
"%s: ib_query_device failed (ret = %d)\n",
1144 goto device_init_failed
;
1146 priv
->hca_caps
= device_attr
->device_cap_flags
;
1150 if (priv
->hca_caps
& IB_DEVICE_UD_IP_CSUM
) {
1151 set_bit(IPOIB_FLAG_CSUM
, &priv
->flags
);
1152 priv
->dev
->features
|= NETIF_F_SG
| NETIF_F_IP_CSUM
;
1156 * Set the full membership bit, so that we join the right
1157 * broadcast group, etc.
1159 priv
->pkey
|= 0x8000;
1161 priv
->dev
->broadcast
[8] = priv
->pkey
>> 8;
1162 priv
->dev
->broadcast
[9] = priv
->pkey
& 0xff;
1164 result
= ib_query_gid(hca
, port
, 0, &priv
->local_gid
);
1166 printk(KERN_WARNING
"%s: ib_query_gid port %d failed (ret = %d)\n",
1167 hca
->name
, port
, result
);
1168 goto device_init_failed
;
1170 memcpy(priv
->dev
->dev_addr
+ 4, priv
->local_gid
.raw
, sizeof (union ib_gid
));
1172 result
= ipoib_dev_init(priv
->dev
, hca
, port
);
1174 printk(KERN_WARNING
"%s: failed to initialize port %d (ret = %d)\n",
1175 hca
->name
, port
, result
);
1176 goto device_init_failed
;
1179 INIT_IB_EVENT_HANDLER(&priv
->event_handler
,
1180 priv
->ca
, ipoib_event
);
1181 result
= ib_register_event_handler(&priv
->event_handler
);
1183 printk(KERN_WARNING
"%s: ib_register_event_handler failed for "
1184 "port %d (ret = %d)\n",
1185 hca
->name
, port
, result
);
1189 if (priv
->dev
->features
& NETIF_F_SG
&& priv
->hca_caps
& IB_DEVICE_UD_TSO
)
1190 priv
->dev
->features
|= NETIF_F_TSO
;
1192 result
= register_netdev(priv
->dev
);
1194 printk(KERN_WARNING
"%s: couldn't register ipoib port %d; error %d\n",
1195 hca
->name
, port
, result
);
1196 goto register_failed
;
1199 ipoib_create_debug_files(priv
->dev
);
1201 if (ipoib_cm_add_mode_attr(priv
->dev
))
1203 if (ipoib_add_pkey_attr(priv
->dev
))
1205 if (ipoib_add_umcast_attr(priv
->dev
))
1207 if (device_create_file(&priv
->dev
->dev
, &dev_attr_create_child
))
1209 if (device_create_file(&priv
->dev
->dev
, &dev_attr_delete_child
))
1215 ipoib_delete_debug_files(priv
->dev
);
1216 unregister_netdev(priv
->dev
);
1219 ib_unregister_event_handler(&priv
->event_handler
);
1220 flush_scheduled_work();
1223 ipoib_dev_cleanup(priv
->dev
);
1226 free_netdev(priv
->dev
);
1229 return ERR_PTR(result
);
1232 static void ipoib_add_one(struct ib_device
*device
)
1234 struct list_head
*dev_list
;
1235 struct net_device
*dev
;
1236 struct ipoib_dev_priv
*priv
;
1239 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1242 dev_list
= kmalloc(sizeof *dev_list
, GFP_KERNEL
);
1246 INIT_LIST_HEAD(dev_list
);
1248 if (device
->node_type
== RDMA_NODE_IB_SWITCH
) {
1253 e
= device
->phys_port_cnt
;
1256 for (p
= s
; p
<= e
; ++p
) {
1257 dev
= ipoib_add_port("ib%d", device
, p
);
1259 priv
= netdev_priv(dev
);
1260 list_add_tail(&priv
->list
, dev_list
);
1264 ib_set_client_data(device
, &ipoib_client
, dev_list
);
1267 static void ipoib_remove_one(struct ib_device
*device
)
1269 struct ipoib_dev_priv
*priv
, *tmp
;
1270 struct list_head
*dev_list
;
1272 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1275 dev_list
= ib_get_client_data(device
, &ipoib_client
);
1277 list_for_each_entry_safe(priv
, tmp
, dev_list
, list
) {
1278 ib_unregister_event_handler(&priv
->event_handler
);
1279 flush_scheduled_work();
1281 unregister_netdev(priv
->dev
);
1282 ipoib_dev_cleanup(priv
->dev
);
1283 free_netdev(priv
->dev
);
1289 static int __init
ipoib_init_module(void)
1293 ipoib_recvq_size
= roundup_pow_of_two(ipoib_recvq_size
);
1294 ipoib_recvq_size
= min(ipoib_recvq_size
, IPOIB_MAX_QUEUE_SIZE
);
1295 ipoib_recvq_size
= max(ipoib_recvq_size
, IPOIB_MIN_QUEUE_SIZE
);
1297 ipoib_sendq_size
= roundup_pow_of_two(ipoib_sendq_size
);
1298 ipoib_sendq_size
= min(ipoib_sendq_size
, IPOIB_MAX_QUEUE_SIZE
);
1299 ipoib_sendq_size
= max(ipoib_sendq_size
, max(2 * MAX_SEND_CQE
,
1300 IPOIB_MIN_QUEUE_SIZE
));
1301 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1302 ipoib_max_conn_qp
= min(ipoib_max_conn_qp
, IPOIB_CM_MAX_CONN_QP
);
1305 ret
= ipoib_register_debugfs();
1310 * We create our own workqueue mainly because we want to be
1311 * able to flush it when devices are being removed. We can't
1312 * use schedule_work()/flush_scheduled_work() because both
1313 * unregister_netdev() and linkwatch_event take the rtnl lock,
1314 * so flush_scheduled_work() can deadlock during device
1317 ipoib_workqueue
= create_singlethread_workqueue("ipoib");
1318 if (!ipoib_workqueue
) {
1323 ib_sa_register_client(&ipoib_sa_client
);
1325 ret
= ib_register_client(&ipoib_client
);
1332 ib_sa_unregister_client(&ipoib_sa_client
);
1333 destroy_workqueue(ipoib_workqueue
);
1336 ipoib_unregister_debugfs();
1341 static void __exit
ipoib_cleanup_module(void)
1343 ib_unregister_client(&ipoib_client
);
1344 ib_sa_unregister_client(&ipoib_sa_client
);
1345 ipoib_unregister_debugfs();
1346 destroy_workqueue(ipoib_workqueue
);
1349 module_init(ipoib_init_module
);
1350 module_exit(ipoib_cleanup_module
);