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>
44 <<<<<<< HEAD
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
46 #include <linux/vmalloc.h>
47 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
49 #include <linux/if_arp.h> /* For ARPHRD_xxx */
56 MODULE_AUTHOR("Roland Dreier");
57 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
58 MODULE_LICENSE("Dual BSD/GPL");
60 int ipoib_sendq_size __read_mostly
= IPOIB_TX_RING_SIZE
;
61 int ipoib_recvq_size __read_mostly
= IPOIB_RX_RING_SIZE
;
63 module_param_named(send_queue_size
, ipoib_sendq_size
, int, 0444);
64 MODULE_PARM_DESC(send_queue_size
, "Number of descriptors in send queue");
65 module_param_named(recv_queue_size
, ipoib_recvq_size
, int, 0444);
66 MODULE_PARM_DESC(recv_queue_size
, "Number of descriptors in receive queue");
68 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
69 int ipoib_debug_level
;
71 module_param_named(debug_level
, ipoib_debug_level
, int, 0644);
72 MODULE_PARM_DESC(debug_level
, "Enable debug tracing if > 0");
75 struct ipoib_path_iter
{
76 struct net_device
*dev
;
77 struct ipoib_path path
;
80 static const u8 ipv4_bcast_addr
[] = {
81 0x00, 0xff, 0xff, 0xff,
82 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
83 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
86 struct workqueue_struct
*ipoib_workqueue
;
88 struct ib_sa_client ipoib_sa_client
;
90 static void ipoib_add_one(struct ib_device
*device
);
91 static void ipoib_remove_one(struct ib_device
*device
);
93 static struct ib_client ipoib_client
= {
96 .remove
= ipoib_remove_one
99 int ipoib_open(struct net_device
*dev
)
101 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
103 ipoib_dbg(priv
, "bringing up interface\n");
105 napi_enable(&priv
->napi
);
106 set_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
108 if (ipoib_pkey_dev_delay_open(dev
))
111 if (ipoib_ib_dev_open(dev
)) {
112 napi_disable(&priv
->napi
);
116 if (ipoib_ib_dev_up(dev
)) {
117 ipoib_ib_dev_stop(dev
, 1);
118 napi_disable(&priv
->napi
);
122 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
123 struct ipoib_dev_priv
*cpriv
;
125 /* Bring up any child interfaces too */
126 mutex_lock(&priv
->vlan_mutex
);
127 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
130 flags
= cpriv
->dev
->flags
;
134 dev_change_flags(cpriv
->dev
, flags
| IFF_UP
);
136 mutex_unlock(&priv
->vlan_mutex
);
139 netif_start_queue(dev
);
144 static int ipoib_stop(struct net_device
*dev
)
146 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
148 ipoib_dbg(priv
, "stopping interface\n");
150 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
151 napi_disable(&priv
->napi
);
153 netif_stop_queue(dev
);
156 * Now flush workqueue to make sure a scheduled task doesn't
157 * bring our internal state back up.
159 flush_workqueue(ipoib_workqueue
);
161 ipoib_ib_dev_down(dev
, 1);
162 ipoib_ib_dev_stop(dev
, 1);
164 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
165 struct ipoib_dev_priv
*cpriv
;
167 /* Bring down any child interfaces too */
168 mutex_lock(&priv
->vlan_mutex
);
169 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
172 flags
= cpriv
->dev
->flags
;
173 if (!(flags
& IFF_UP
))
176 dev_change_flags(cpriv
->dev
, flags
& ~IFF_UP
);
178 mutex_unlock(&priv
->vlan_mutex
);
184 static int ipoib_change_mtu(struct net_device
*dev
, int new_mtu
)
186 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
188 /* dev->mtu > 2K ==> connected mode */
189 if (ipoib_cm_admin_enabled(dev
)) {
190 if (new_mtu
> ipoib_cm_max_mtu(dev
))
193 if (new_mtu
> priv
->mcast_mtu
)
194 ipoib_warn(priv
, "mtu > %d will cause multicast packet drops.\n",
201 if (new_mtu
> IPOIB_PACKET_SIZE
- IPOIB_ENCAP_LEN
)
204 priv
->admin_mtu
= new_mtu
;
206 dev
->mtu
= min(priv
->mcast_mtu
, priv
->admin_mtu
);
211 static struct ipoib_path
*__path_find(struct net_device
*dev
, void *gid
)
213 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
214 struct rb_node
*n
= priv
->path_tree
.rb_node
;
215 struct ipoib_path
*path
;
219 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
221 ret
= memcmp(gid
, path
->pathrec
.dgid
.raw
,
222 sizeof (union ib_gid
));
235 static int __path_add(struct net_device
*dev
, struct ipoib_path
*path
)
237 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
238 struct rb_node
**n
= &priv
->path_tree
.rb_node
;
239 struct rb_node
*pn
= NULL
;
240 struct ipoib_path
*tpath
;
245 tpath
= rb_entry(pn
, struct ipoib_path
, rb_node
);
247 ret
= memcmp(path
->pathrec
.dgid
.raw
, tpath
->pathrec
.dgid
.raw
,
248 sizeof (union ib_gid
));
257 rb_link_node(&path
->rb_node
, pn
, n
);
258 rb_insert_color(&path
->rb_node
, &priv
->path_tree
);
260 list_add_tail(&path
->list
, &priv
->path_list
);
265 static void path_free(struct net_device
*dev
, struct ipoib_path
*path
)
267 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
268 struct ipoib_neigh
*neigh
, *tn
;
272 while ((skb
= __skb_dequeue(&path
->queue
)))
273 dev_kfree_skb_irq(skb
);
275 spin_lock_irqsave(&priv
->lock
, flags
);
277 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
279 * It's safe to call ipoib_put_ah() inside priv->lock
280 * here, because we know that path->ah will always
281 * hold one more reference, so ipoib_put_ah() will
282 * never do more than decrement the ref count.
285 ipoib_put_ah(neigh
->ah
);
287 ipoib_neigh_free(dev
, neigh
);
290 spin_unlock_irqrestore(&priv
->lock
, flags
);
293 ipoib_put_ah(path
->ah
);
298 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
300 struct ipoib_path_iter
*ipoib_path_iter_init(struct net_device
*dev
)
302 struct ipoib_path_iter
*iter
;
304 iter
= kmalloc(sizeof *iter
, GFP_KERNEL
);
309 memset(iter
->path
.pathrec
.dgid
.raw
, 0, 16);
311 if (ipoib_path_iter_next(iter
)) {
319 int ipoib_path_iter_next(struct ipoib_path_iter
*iter
)
321 struct ipoib_dev_priv
*priv
= netdev_priv(iter
->dev
);
323 struct ipoib_path
*path
;
326 spin_lock_irq(&priv
->lock
);
328 n
= rb_first(&priv
->path_tree
);
331 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
333 if (memcmp(iter
->path
.pathrec
.dgid
.raw
, path
->pathrec
.dgid
.raw
,
334 sizeof (union ib_gid
)) < 0) {
343 spin_unlock_irq(&priv
->lock
);
348 void ipoib_path_iter_read(struct ipoib_path_iter
*iter
,
349 struct ipoib_path
*path
)
354 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
356 void ipoib_flush_paths(struct net_device
*dev
)
358 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
359 struct ipoib_path
*path
, *tp
;
360 LIST_HEAD(remove_list
);
362 spin_lock_irq(&priv
->tx_lock
);
363 spin_lock(&priv
->lock
);
365 list_splice(&priv
->path_list
, &remove_list
);
366 INIT_LIST_HEAD(&priv
->path_list
);
368 list_for_each_entry(path
, &remove_list
, list
)
369 rb_erase(&path
->rb_node
, &priv
->path_tree
);
371 list_for_each_entry_safe(path
, tp
, &remove_list
, list
) {
373 ib_sa_cancel_query(path
->query_id
, path
->query
);
374 spin_unlock(&priv
->lock
);
375 spin_unlock_irq(&priv
->tx_lock
);
376 wait_for_completion(&path
->done
);
377 path_free(dev
, path
);
378 spin_lock_irq(&priv
->tx_lock
);
379 spin_lock(&priv
->lock
);
381 spin_unlock(&priv
->lock
);
382 spin_unlock_irq(&priv
->tx_lock
);
385 static void path_rec_completion(int status
,
386 struct ib_sa_path_rec
*pathrec
,
389 struct ipoib_path
*path
= path_ptr
;
390 struct net_device
*dev
= path
->dev
;
391 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
392 struct ipoib_ah
*ah
= NULL
;
393 struct ipoib_neigh
*neigh
, *tn
;
394 struct sk_buff_head skqueue
;
399 ipoib_dbg(priv
, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT
"\n",
400 be16_to_cpu(pathrec
->dlid
), IPOIB_GID_ARG(pathrec
->dgid
));
402 ipoib_dbg(priv
, "PathRec status %d for GID " IPOIB_GID_FMT
"\n",
403 status
, IPOIB_GID_ARG(path
->pathrec
.dgid
));
405 skb_queue_head_init(&skqueue
);
408 struct ib_ah_attr av
;
410 if (!ib_init_ah_from_path(priv
->ca
, priv
->port
, pathrec
, &av
))
411 ah
= ipoib_create_ah(dev
, priv
->pd
, &av
);
414 spin_lock_irqsave(&priv
->lock
, flags
);
419 path
->pathrec
= *pathrec
;
421 ipoib_dbg(priv
, "created address handle %p for LID 0x%04x, SL %d\n",
422 ah
, be16_to_cpu(pathrec
->dlid
), pathrec
->sl
);
424 while ((skb
= __skb_dequeue(&path
->queue
)))
425 __skb_queue_tail(&skqueue
, skb
);
427 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
428 kref_get(&path
->ah
->ref
);
429 neigh
->ah
= path
->ah
;
430 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
431 sizeof(union ib_gid
));
433 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
434 if (!ipoib_cm_get(neigh
))
435 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
,
438 if (!ipoib_cm_get(neigh
)) {
439 list_del(&neigh
->list
);
441 ipoib_put_ah(neigh
->ah
);
442 ipoib_neigh_free(dev
, neigh
);
447 while ((skb
= __skb_dequeue(&neigh
->queue
)))
448 __skb_queue_tail(&skqueue
, skb
);
453 complete(&path
->done
);
455 spin_unlock_irqrestore(&priv
->lock
, flags
);
457 while ((skb
= __skb_dequeue(&skqueue
))) {
459 if (dev_queue_xmit(skb
))
460 ipoib_warn(priv
, "dev_queue_xmit failed "
461 "to requeue packet\n");
465 static struct ipoib_path
*path_rec_create(struct net_device
*dev
, void *gid
)
467 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
468 struct ipoib_path
*path
;
470 if (!priv
->broadcast
)
473 path
= kzalloc(sizeof *path
, GFP_ATOMIC
);
479 skb_queue_head_init(&path
->queue
);
481 INIT_LIST_HEAD(&path
->neigh_list
);
483 memcpy(path
->pathrec
.dgid
.raw
, gid
, sizeof (union ib_gid
));
484 path
->pathrec
.sgid
= priv
->local_gid
;
485 path
->pathrec
.pkey
= cpu_to_be16(priv
->pkey
);
486 path
->pathrec
.numb_path
= 1;
487 path
->pathrec
.traffic_class
= priv
->broadcast
->mcmember
.traffic_class
;
492 static int path_rec_start(struct net_device
*dev
,
493 struct ipoib_path
*path
)
495 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
497 ipoib_dbg(priv
, "Start path record lookup for " IPOIB_GID_FMT
"\n",
498 IPOIB_GID_ARG(path
->pathrec
.dgid
));
500 init_completion(&path
->done
);
503 ib_sa_path_rec_get(&ipoib_sa_client
, priv
->ca
, priv
->port
,
505 IB_SA_PATH_REC_DGID
|
506 IB_SA_PATH_REC_SGID
|
507 IB_SA_PATH_REC_NUMB_PATH
|
508 IB_SA_PATH_REC_TRAFFIC_CLASS
|
513 if (path
->query_id
< 0) {
514 ipoib_warn(priv
, "ib_sa_path_rec_get failed\n");
516 return path
->query_id
;
522 static void neigh_add_path(struct sk_buff
*skb
, struct net_device
*dev
)
524 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
525 struct ipoib_path
*path
;
526 struct ipoib_neigh
*neigh
;
528 neigh
= ipoib_neigh_alloc(skb
->dst
->neighbour
, skb
->dev
);
530 ++dev
->stats
.tx_dropped
;
531 dev_kfree_skb_any(skb
);
536 * We can only be called from ipoib_start_xmit, so we're
537 * inside tx_lock -- no need to save/restore flags.
539 spin_lock(&priv
->lock
);
541 path
= __path_find(dev
, skb
->dst
->neighbour
->ha
+ 4);
543 path
= path_rec_create(dev
, skb
->dst
->neighbour
->ha
+ 4);
547 __path_add(dev
, path
);
550 list_add_tail(&neigh
->list
, &path
->neigh_list
);
553 kref_get(&path
->ah
->ref
);
554 neigh
->ah
= path
->ah
;
555 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
556 sizeof(union ib_gid
));
558 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
559 if (!ipoib_cm_get(neigh
))
560 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
, path
, neigh
));
561 if (!ipoib_cm_get(neigh
)) {
562 list_del(&neigh
->list
);
564 ipoib_put_ah(neigh
->ah
);
565 ipoib_neigh_free(dev
, neigh
);
568 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
)
569 __skb_queue_tail(&neigh
->queue
, skb
);
571 ipoib_warn(priv
, "queue length limit %d. Packet drop.\n",
572 skb_queue_len(&neigh
->queue
));
576 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
580 if (!path
->query
&& path_rec_start(dev
, path
))
583 __skb_queue_tail(&neigh
->queue
, skb
);
586 spin_unlock(&priv
->lock
);
590 list_del(&neigh
->list
);
593 ipoib_neigh_free(dev
, neigh
);
595 ++dev
->stats
.tx_dropped
;
596 dev_kfree_skb_any(skb
);
598 spin_unlock(&priv
->lock
);
601 static void ipoib_path_lookup(struct sk_buff
*skb
, struct net_device
*dev
)
603 struct ipoib_dev_priv
*priv
= netdev_priv(skb
->dev
);
605 /* Look up path record for unicasts */
606 if (skb
->dst
->neighbour
->ha
[4] != 0xff) {
607 neigh_add_path(skb
, dev
);
611 /* Add in the P_Key for multicasts */
612 skb
->dst
->neighbour
->ha
[8] = (priv
->pkey
>> 8) & 0xff;
613 skb
->dst
->neighbour
->ha
[9] = priv
->pkey
& 0xff;
614 ipoib_mcast_send(dev
, skb
->dst
->neighbour
->ha
+ 4, skb
);
617 static void unicast_arp_send(struct sk_buff
*skb
, struct net_device
*dev
,
618 struct ipoib_pseudoheader
*phdr
)
620 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
621 struct ipoib_path
*path
;
624 * We can only be called from ipoib_start_xmit, so we're
625 * inside tx_lock -- no need to save/restore flags.
627 spin_lock(&priv
->lock
);
629 path
= __path_find(dev
, phdr
->hwaddr
+ 4);
631 path
= path_rec_create(dev
, phdr
->hwaddr
+ 4);
633 /* put pseudoheader back on for next time */
634 skb_push(skb
, sizeof *phdr
);
635 __skb_queue_tail(&path
->queue
, skb
);
637 if (path_rec_start(dev
, path
)) {
638 spin_unlock(&priv
->lock
);
639 path_free(dev
, path
);
642 __path_add(dev
, path
);
644 ++dev
->stats
.tx_dropped
;
645 dev_kfree_skb_any(skb
);
648 spin_unlock(&priv
->lock
);
653 ipoib_dbg(priv
, "Send unicast ARP to %04x\n",
654 be16_to_cpu(path
->pathrec
.dlid
));
656 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(phdr
->hwaddr
));
657 } else if ((path
->query
|| !path_rec_start(dev
, path
)) &&
658 skb_queue_len(&path
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
659 /* put pseudoheader back on for next time */
660 skb_push(skb
, sizeof *phdr
);
661 __skb_queue_tail(&path
->queue
, skb
);
663 ++dev
->stats
.tx_dropped
;
664 dev_kfree_skb_any(skb
);
667 spin_unlock(&priv
->lock
);
670 static int ipoib_start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
672 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
673 struct ipoib_neigh
*neigh
;
676 if (unlikely(!spin_trylock_irqsave(&priv
->tx_lock
, flags
)))
677 return NETDEV_TX_LOCKED
;
679 if (likely(skb
->dst
&& skb
->dst
->neighbour
)) {
680 if (unlikely(!*to_ipoib_neigh(skb
->dst
->neighbour
))) {
681 ipoib_path_lookup(skb
, dev
);
685 neigh
= *to_ipoib_neigh(skb
->dst
->neighbour
);
688 if (unlikely((memcmp(&neigh
->dgid
.raw
,
689 skb
->dst
->neighbour
->ha
+ 4,
690 sizeof(union ib_gid
))) ||
691 (neigh
->dev
!= dev
))) {
692 spin_lock(&priv
->lock
);
694 * It's safe to call ipoib_put_ah() inside
695 * priv->lock here, because we know that
696 * path->ah will always hold one more reference,
697 * so ipoib_put_ah() will never do more than
698 * decrement the ref count.
700 ipoib_put_ah(neigh
->ah
);
701 list_del(&neigh
->list
);
702 ipoib_neigh_free(dev
, neigh
);
703 spin_unlock(&priv
->lock
);
704 ipoib_path_lookup(skb
, dev
);
708 if (ipoib_cm_get(neigh
)) {
709 if (ipoib_cm_up(neigh
)) {
710 ipoib_cm_send(dev
, skb
, ipoib_cm_get(neigh
));
713 } else if (neigh
->ah
) {
714 ipoib_send(dev
, skb
, neigh
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
718 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
719 spin_lock(&priv
->lock
);
720 __skb_queue_tail(&neigh
->queue
, skb
);
721 spin_unlock(&priv
->lock
);
723 ++dev
->stats
.tx_dropped
;
724 dev_kfree_skb_any(skb
);
727 struct ipoib_pseudoheader
*phdr
=
728 (struct ipoib_pseudoheader
*) skb
->data
;
729 skb_pull(skb
, sizeof *phdr
);
731 if (phdr
->hwaddr
[4] == 0xff) {
732 /* Add in the P_Key for multicast*/
733 phdr
->hwaddr
[8] = (priv
->pkey
>> 8) & 0xff;
734 phdr
->hwaddr
[9] = priv
->pkey
& 0xff;
736 ipoib_mcast_send(dev
, phdr
->hwaddr
+ 4, skb
);
738 /* unicast GID -- should be ARP or RARP reply */
740 if ((be16_to_cpup((__be16
*) skb
->data
) != ETH_P_ARP
) &&
741 (be16_to_cpup((__be16
*) skb
->data
) != ETH_P_RARP
)) {
742 ipoib_warn(priv
, "Unicast, no %s: type %04x, QPN %06x "
744 skb
->dst
? "neigh" : "dst",
745 be16_to_cpup((__be16
*) skb
->data
),
746 IPOIB_QPN(phdr
->hwaddr
),
747 IPOIB_GID_RAW_ARG(phdr
->hwaddr
+ 4));
748 dev_kfree_skb_any(skb
);
749 ++dev
->stats
.tx_dropped
;
753 unicast_arp_send(skb
, dev
, phdr
);
758 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
763 static void ipoib_timeout(struct net_device
*dev
)
765 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
767 ipoib_warn(priv
, "transmit timeout: latency %d msecs\n",
768 jiffies_to_msecs(jiffies
- dev
->trans_start
));
769 ipoib_warn(priv
, "queue stopped %d, tx_head %u, tx_tail %u\n",
770 netif_queue_stopped(dev
),
771 priv
->tx_head
, priv
->tx_tail
);
772 /* XXX reset QP, etc. */
775 static int ipoib_hard_header(struct sk_buff
*skb
,
776 struct net_device
*dev
,
778 const void *daddr
, const void *saddr
, unsigned len
)
780 struct ipoib_header
*header
;
782 header
= (struct ipoib_header
*) skb_push(skb
, sizeof *header
);
784 header
->proto
= htons(type
);
785 header
->reserved
= 0;
788 * If we don't have a neighbour structure, stuff the
789 * destination address onto the front of the skb so we can
790 * figure out where to send the packet later.
792 if ((!skb
->dst
|| !skb
->dst
->neighbour
) && daddr
) {
793 struct ipoib_pseudoheader
*phdr
=
794 (struct ipoib_pseudoheader
*) skb_push(skb
, sizeof *phdr
);
795 memcpy(phdr
->hwaddr
, daddr
, INFINIBAND_ALEN
);
801 static void ipoib_set_mcast_list(struct net_device
*dev
)
803 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
805 if (!test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
)) {
806 ipoib_dbg(priv
, "IPOIB_FLAG_OPER_UP not set");
810 queue_work(ipoib_workqueue
, &priv
->restart_task
);
813 static void ipoib_neigh_cleanup(struct neighbour
*n
)
815 struct ipoib_neigh
*neigh
;
816 struct ipoib_dev_priv
*priv
= netdev_priv(n
->dev
);
818 struct ipoib_ah
*ah
= NULL
;
820 neigh
= *to_ipoib_neigh(n
);
822 priv
= netdev_priv(neigh
->dev
);
826 "neigh_cleanup for %06x " IPOIB_GID_FMT
"\n",
828 IPOIB_GID_RAW_ARG(n
->ha
+ 4));
830 spin_lock_irqsave(&priv
->lock
, flags
);
834 list_del(&neigh
->list
);
835 ipoib_neigh_free(n
->dev
, neigh
);
837 spin_unlock_irqrestore(&priv
->lock
, flags
);
843 struct ipoib_neigh
*ipoib_neigh_alloc(struct neighbour
*neighbour
,
844 struct net_device
*dev
)
846 struct ipoib_neigh
*neigh
;
848 neigh
= kmalloc(sizeof *neigh
, GFP_ATOMIC
);
852 neigh
->neighbour
= neighbour
;
854 *to_ipoib_neigh(neighbour
) = neigh
;
855 skb_queue_head_init(&neigh
->queue
);
856 ipoib_cm_set(neigh
, NULL
);
861 void ipoib_neigh_free(struct net_device
*dev
, struct ipoib_neigh
*neigh
)
864 *to_ipoib_neigh(neigh
->neighbour
) = NULL
;
865 while ((skb
= __skb_dequeue(&neigh
->queue
))) {
866 ++dev
->stats
.tx_dropped
;
867 dev_kfree_skb_any(skb
);
869 if (ipoib_cm_get(neigh
))
870 ipoib_cm_destroy_tx(ipoib_cm_get(neigh
));
874 static int ipoib_neigh_setup_dev(struct net_device
*dev
, struct neigh_parms
*parms
)
876 parms
->neigh_cleanup
= ipoib_neigh_cleanup
;
881 int ipoib_dev_init(struct net_device
*dev
, struct ib_device
*ca
, int port
)
883 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
885 /* Allocate RX/TX "rings" to hold queued skbs */
886 priv
->rx_ring
= kzalloc(ipoib_recvq_size
* sizeof *priv
->rx_ring
,
888 if (!priv
->rx_ring
) {
889 printk(KERN_WARNING
"%s: failed to allocate RX ring (%d entries)\n",
890 ca
->name
, ipoib_recvq_size
);
894 <<<<<<< HEAD
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
895 priv
->tx_ring
= kzalloc(ipoib_sendq_size
* sizeof *priv
->tx_ring
,
898 priv
->tx_ring
= vmalloc(ipoib_sendq_size
* sizeof *priv
->tx_ring
);
899 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
900 if (!priv
->tx_ring
) {
901 printk(KERN_WARNING
"%s: failed to allocate TX ring (%d entries)\n",
902 ca
->name
, ipoib_sendq_size
);
903 goto out_rx_ring_cleanup
;
905 <<<<<<< HEAD
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
907 memset(priv
->tx_ring
, 0, ipoib_sendq_size
* sizeof *priv
->tx_ring
);
908 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
910 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
912 if (ipoib_ib_dev_init(dev
, ca
, port
))
913 goto out_tx_ring_cleanup
;
918 <<<<<<< HEAD
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
919 kfree(priv
->tx_ring
);
921 vfree(priv
->tx_ring
);
922 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
925 kfree(priv
->rx_ring
);
931 void ipoib_dev_cleanup(struct net_device
*dev
)
933 struct ipoib_dev_priv
*priv
= netdev_priv(dev
), *cpriv
, *tcpriv
;
935 ipoib_delete_debug_files(dev
);
937 /* Delete any child interfaces first */
938 list_for_each_entry_safe(cpriv
, tcpriv
, &priv
->child_intfs
, list
) {
939 unregister_netdev(cpriv
->dev
);
940 ipoib_dev_cleanup(cpriv
->dev
);
941 free_netdev(cpriv
->dev
);
944 ipoib_ib_dev_cleanup(dev
);
946 kfree(priv
->rx_ring
);
947 <<<<<<< HEAD
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
948 kfree(priv
->tx_ring
);
950 vfree(priv
->tx_ring
);
951 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a
:drivers
/infiniband
/ulp
/ipoib
/ipoib_main
.c
953 priv
->rx_ring
= NULL
;
954 priv
->tx_ring
= NULL
;
957 static const struct header_ops ipoib_header_ops
= {
958 .create
= ipoib_hard_header
,
961 static void ipoib_setup(struct net_device
*dev
)
963 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
965 dev
->open
= ipoib_open
;
966 dev
->stop
= ipoib_stop
;
967 dev
->change_mtu
= ipoib_change_mtu
;
968 dev
->hard_start_xmit
= ipoib_start_xmit
;
969 dev
->tx_timeout
= ipoib_timeout
;
970 dev
->header_ops
= &ipoib_header_ops
;
971 dev
->set_multicast_list
= ipoib_set_mcast_list
;
972 dev
->neigh_setup
= ipoib_neigh_setup_dev
;
974 netif_napi_add(dev
, &priv
->napi
, ipoib_poll
, 100);
976 dev
->watchdog_timeo
= HZ
;
978 dev
->flags
|= IFF_BROADCAST
| IFF_MULTICAST
;
981 * We add in INFINIBAND_ALEN to allow for the destination
982 * address "pseudoheader" for skbs without neighbour struct.
984 dev
->hard_header_len
= IPOIB_ENCAP_LEN
+ INFINIBAND_ALEN
;
985 dev
->addr_len
= INFINIBAND_ALEN
;
986 dev
->type
= ARPHRD_INFINIBAND
;
987 dev
->tx_queue_len
= ipoib_sendq_size
* 2;
988 dev
->features
= (NETIF_F_VLAN_CHALLENGED
|
992 /* MTU will be reset when mcast join happens */
993 dev
->mtu
= IPOIB_PACKET_SIZE
- IPOIB_ENCAP_LEN
;
994 priv
->mcast_mtu
= priv
->admin_mtu
= dev
->mtu
;
996 memcpy(dev
->broadcast
, ipv4_bcast_addr
, INFINIBAND_ALEN
);
998 netif_carrier_off(dev
);
1002 spin_lock_init(&priv
->lock
);
1003 spin_lock_init(&priv
->tx_lock
);
1005 mutex_init(&priv
->mcast_mutex
);
1006 mutex_init(&priv
->vlan_mutex
);
1008 INIT_LIST_HEAD(&priv
->path_list
);
1009 INIT_LIST_HEAD(&priv
->child_intfs
);
1010 INIT_LIST_HEAD(&priv
->dead_ahs
);
1011 INIT_LIST_HEAD(&priv
->multicast_list
);
1013 INIT_DELAYED_WORK(&priv
->pkey_poll_task
, ipoib_pkey_poll
);
1014 INIT_WORK(&priv
->pkey_event_task
, ipoib_pkey_event
);
1015 INIT_DELAYED_WORK(&priv
->mcast_task
, ipoib_mcast_join_task
);
1016 INIT_WORK(&priv
->flush_task
, ipoib_ib_dev_flush
);
1017 INIT_WORK(&priv
->restart_task
, ipoib_mcast_restart_task
);
1018 INIT_DELAYED_WORK(&priv
->ah_reap_task
, ipoib_reap_ah
);
1021 struct ipoib_dev_priv
*ipoib_intf_alloc(const char *name
)
1023 struct net_device
*dev
;
1025 dev
= alloc_netdev((int) sizeof (struct ipoib_dev_priv
), name
,
1030 return netdev_priv(dev
);
1033 static ssize_t
show_pkey(struct device
*dev
,
1034 struct device_attribute
*attr
, char *buf
)
1036 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1038 return sprintf(buf
, "0x%04x\n", priv
->pkey
);
1040 static DEVICE_ATTR(pkey
, S_IRUGO
, show_pkey
, NULL
);
1042 static ssize_t
show_umcast(struct device
*dev
,
1043 struct device_attribute
*attr
, char *buf
)
1045 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1047 return sprintf(buf
, "%d\n", test_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
));
1050 static ssize_t
set_umcast(struct device
*dev
,
1051 struct device_attribute
*attr
,
1052 const char *buf
, size_t count
)
1054 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1055 unsigned long umcast_val
= simple_strtoul(buf
, NULL
, 0);
1057 if (umcast_val
> 0) {
1058 set_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1059 ipoib_warn(priv
, "ignoring multicast groups joined directly "
1062 clear_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1066 static DEVICE_ATTR(umcast
, S_IWUSR
| S_IRUGO
, show_umcast
, set_umcast
);
1068 int ipoib_add_umcast_attr(struct net_device
*dev
)
1070 return device_create_file(&dev
->dev
, &dev_attr_umcast
);
1073 static ssize_t
create_child(struct device
*dev
,
1074 struct device_attribute
*attr
,
1075 const char *buf
, size_t count
)
1080 if (sscanf(buf
, "%i", &pkey
) != 1)
1083 if (pkey
< 0 || pkey
> 0xffff)
1087 * Set the full membership bit, so that we join the right
1088 * broadcast group, etc.
1092 ret
= ipoib_vlan_add(to_net_dev(dev
), pkey
);
1094 return ret
? ret
: count
;
1096 static DEVICE_ATTR(create_child
, S_IWUGO
, NULL
, create_child
);
1098 static ssize_t
delete_child(struct device
*dev
,
1099 struct device_attribute
*attr
,
1100 const char *buf
, size_t count
)
1105 if (sscanf(buf
, "%i", &pkey
) != 1)
1108 if (pkey
< 0 || pkey
> 0xffff)
1111 ret
= ipoib_vlan_delete(to_net_dev(dev
), pkey
);
1113 return ret
? ret
: count
;
1116 static DEVICE_ATTR(delete_child
, S_IWUGO
, NULL
, delete_child
);
1118 int ipoib_add_pkey_attr(struct net_device
*dev
)
1120 return device_create_file(&dev
->dev
, &dev_attr_pkey
);
1123 static struct net_device
*ipoib_add_port(const char *format
,
1124 struct ib_device
*hca
, u8 port
)
1126 struct ipoib_dev_priv
*priv
;
1127 int result
= -ENOMEM
;
1129 priv
= ipoib_intf_alloc(format
);
1131 goto alloc_mem_failed
;
1133 SET_NETDEV_DEV(priv
->dev
, hca
->dma_device
);
1135 result
= ib_query_pkey(hca
, port
, 0, &priv
->pkey
);
1137 printk(KERN_WARNING
"%s: ib_query_pkey port %d failed (ret = %d)\n",
1138 hca
->name
, port
, result
);
1139 goto device_init_failed
;
1143 * Set the full membership bit, so that we join the right
1144 * broadcast group, etc.
1146 priv
->pkey
|= 0x8000;
1148 priv
->dev
->broadcast
[8] = priv
->pkey
>> 8;
1149 priv
->dev
->broadcast
[9] = priv
->pkey
& 0xff;
1151 result
= ib_query_gid(hca
, port
, 0, &priv
->local_gid
);
1153 printk(KERN_WARNING
"%s: ib_query_gid port %d failed (ret = %d)\n",
1154 hca
->name
, port
, result
);
1155 goto device_init_failed
;
1157 memcpy(priv
->dev
->dev_addr
+ 4, priv
->local_gid
.raw
, sizeof (union ib_gid
));
1160 result
= ipoib_dev_init(priv
->dev
, hca
, port
);
1162 printk(KERN_WARNING
"%s: failed to initialize port %d (ret = %d)\n",
1163 hca
->name
, port
, result
);
1164 goto device_init_failed
;
1167 INIT_IB_EVENT_HANDLER(&priv
->event_handler
,
1168 priv
->ca
, ipoib_event
);
1169 result
= ib_register_event_handler(&priv
->event_handler
);
1171 printk(KERN_WARNING
"%s: ib_register_event_handler failed for "
1172 "port %d (ret = %d)\n",
1173 hca
->name
, port
, result
);
1177 result
= register_netdev(priv
->dev
);
1179 printk(KERN_WARNING
"%s: couldn't register ipoib port %d; error %d\n",
1180 hca
->name
, port
, result
);
1181 goto register_failed
;
1184 ipoib_create_debug_files(priv
->dev
);
1186 if (ipoib_cm_add_mode_attr(priv
->dev
))
1188 if (ipoib_add_pkey_attr(priv
->dev
))
1190 if (ipoib_add_umcast_attr(priv
->dev
))
1192 if (device_create_file(&priv
->dev
->dev
, &dev_attr_create_child
))
1194 if (device_create_file(&priv
->dev
->dev
, &dev_attr_delete_child
))
1200 ipoib_delete_debug_files(priv
->dev
);
1201 unregister_netdev(priv
->dev
);
1204 ib_unregister_event_handler(&priv
->event_handler
);
1205 flush_scheduled_work();
1208 ipoib_dev_cleanup(priv
->dev
);
1211 free_netdev(priv
->dev
);
1214 return ERR_PTR(result
);
1217 static void ipoib_add_one(struct ib_device
*device
)
1219 struct list_head
*dev_list
;
1220 struct net_device
*dev
;
1221 struct ipoib_dev_priv
*priv
;
1224 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1227 dev_list
= kmalloc(sizeof *dev_list
, GFP_KERNEL
);
1231 INIT_LIST_HEAD(dev_list
);
1233 if (device
->node_type
== RDMA_NODE_IB_SWITCH
) {
1238 e
= device
->phys_port_cnt
;
1241 for (p
= s
; p
<= e
; ++p
) {
1242 dev
= ipoib_add_port("ib%d", device
, p
);
1244 priv
= netdev_priv(dev
);
1245 list_add_tail(&priv
->list
, dev_list
);
1249 ib_set_client_data(device
, &ipoib_client
, dev_list
);
1252 static void ipoib_remove_one(struct ib_device
*device
)
1254 struct ipoib_dev_priv
*priv
, *tmp
;
1255 struct list_head
*dev_list
;
1257 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1260 dev_list
= ib_get_client_data(device
, &ipoib_client
);
1262 list_for_each_entry_safe(priv
, tmp
, dev_list
, list
) {
1263 ib_unregister_event_handler(&priv
->event_handler
);
1264 flush_scheduled_work();
1266 unregister_netdev(priv
->dev
);
1267 ipoib_dev_cleanup(priv
->dev
);
1268 free_netdev(priv
->dev
);
1274 static int __init
ipoib_init_module(void)
1278 ipoib_recvq_size
= roundup_pow_of_two(ipoib_recvq_size
);
1279 ipoib_recvq_size
= min(ipoib_recvq_size
, IPOIB_MAX_QUEUE_SIZE
);
1280 ipoib_recvq_size
= max(ipoib_recvq_size
, IPOIB_MIN_QUEUE_SIZE
);
1282 ipoib_sendq_size
= roundup_pow_of_two(ipoib_sendq_size
);
1283 ipoib_sendq_size
= min(ipoib_sendq_size
, IPOIB_MAX_QUEUE_SIZE
);
1284 ipoib_sendq_size
= max(ipoib_sendq_size
, IPOIB_MIN_QUEUE_SIZE
);
1285 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1286 ipoib_max_conn_qp
= min(ipoib_max_conn_qp
, IPOIB_CM_MAX_CONN_QP
);
1289 ret
= ipoib_register_debugfs();
1294 * We create our own workqueue mainly because we want to be
1295 * able to flush it when devices are being removed. We can't
1296 * use schedule_work()/flush_scheduled_work() because both
1297 * unregister_netdev() and linkwatch_event take the rtnl lock,
1298 * so flush_scheduled_work() can deadlock during device
1301 ipoib_workqueue
= create_singlethread_workqueue("ipoib");
1302 if (!ipoib_workqueue
) {
1307 ib_sa_register_client(&ipoib_sa_client
);
1309 ret
= ib_register_client(&ipoib_client
);
1316 ib_sa_unregister_client(&ipoib_sa_client
);
1317 destroy_workqueue(ipoib_workqueue
);
1320 ipoib_unregister_debugfs();
1325 static void __exit
ipoib_cleanup_module(void)
1327 ib_unregister_client(&ipoib_client
);
1328 ib_sa_unregister_client(&ipoib_sa_client
);
1329 ipoib_unregister_debugfs();
1330 destroy_workqueue(ipoib_workqueue
);
1333 module_init(ipoib_init_module
);
1334 module_exit(ipoib_cleanup_module
);