2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
38 #include <linux/delay.h>
39 #include <linux/dma-mapping.h>
41 #include <rdma/ib_cache.h>
45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
46 static int data_debug_level
;
48 module_param(data_debug_level
, int, 0644);
49 MODULE_PARM_DESC(data_debug_level
,
50 "Enable data path debug tracing if > 0");
53 #define IPOIB_OP_RECV (1ul << 31)
55 static DEFINE_MUTEX(pkey_mutex
);
57 struct ipoib_ah
*ipoib_create_ah(struct net_device
*dev
,
58 struct ib_pd
*pd
, struct ib_ah_attr
*attr
)
62 ah
= kmalloc(sizeof *ah
, GFP_KERNEL
);
70 ah
->ah
= ib_create_ah(pd
, attr
);
75 ipoib_dbg(netdev_priv(dev
), "Created ah %p\n", ah
->ah
);
80 void ipoib_free_ah(struct kref
*kref
)
82 struct ipoib_ah
*ah
= container_of(kref
, struct ipoib_ah
, ref
);
83 struct ipoib_dev_priv
*priv
= netdev_priv(ah
->dev
);
87 if ((int) priv
->tx_tail
- (int) ah
->last_send
>= 0) {
88 ipoib_dbg(priv
, "Freeing ah %p\n", ah
->ah
);
89 ib_destroy_ah(ah
->ah
);
92 spin_lock_irqsave(&priv
->lock
, flags
);
93 list_add_tail(&ah
->list
, &priv
->dead_ahs
);
94 spin_unlock_irqrestore(&priv
->lock
, flags
);
98 static int ipoib_ib_post_receive(struct net_device
*dev
, int id
)
100 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
102 struct ib_recv_wr param
;
103 struct ib_recv_wr
*bad_wr
;
106 list
.addr
= priv
->rx_ring
[id
].mapping
;
107 list
.length
= IPOIB_BUF_SIZE
;
108 list
.lkey
= priv
->mr
->lkey
;
111 param
.wr_id
= id
| IPOIB_OP_RECV
;
112 param
.sg_list
= &list
;
115 ret
= ib_post_recv(priv
->qp
, ¶m
, &bad_wr
);
117 ipoib_warn(priv
, "receive failed for buf %d (%d)\n", id
, ret
);
118 dma_unmap_single(priv
->ca
->dma_device
,
119 priv
->rx_ring
[id
].mapping
,
120 IPOIB_BUF_SIZE
, DMA_FROM_DEVICE
);
121 dev_kfree_skb_any(priv
->rx_ring
[id
].skb
);
122 priv
->rx_ring
[id
].skb
= NULL
;
128 static int ipoib_alloc_rx_skb(struct net_device
*dev
, int id
)
130 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
134 skb
= dev_alloc_skb(IPOIB_BUF_SIZE
+ 4);
139 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
140 * header. So we need 4 more bytes to get to 48 and align the
141 * IP header to a multiple of 16.
145 addr
= dma_map_single(priv
->ca
->dma_device
,
146 skb
->data
, IPOIB_BUF_SIZE
,
148 if (unlikely(dma_mapping_error(addr
))) {
149 dev_kfree_skb_any(skb
);
153 priv
->rx_ring
[id
].skb
= skb
;
154 priv
->rx_ring
[id
].mapping
= addr
;
159 static int ipoib_ib_post_receives(struct net_device
*dev
)
161 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
164 for (i
= 0; i
< ipoib_recvq_size
; ++i
) {
165 if (ipoib_alloc_rx_skb(dev
, i
)) {
166 ipoib_warn(priv
, "failed to allocate receive buffer %d\n", i
);
169 if (ipoib_ib_post_receive(dev
, i
)) {
170 ipoib_warn(priv
, "ipoib_ib_post_receive failed for buf %d\n", i
);
178 static void ipoib_ib_handle_wc(struct net_device
*dev
,
181 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
182 unsigned int wr_id
= wc
->wr_id
;
184 ipoib_dbg_data(priv
, "called: id %d, op %d, status: %d\n",
185 wr_id
, wc
->opcode
, wc
->status
);
187 if (wr_id
& IPOIB_OP_RECV
) {
188 wr_id
&= ~IPOIB_OP_RECV
;
190 if (wr_id
< ipoib_recvq_size
) {
191 struct sk_buff
*skb
= priv
->rx_ring
[wr_id
].skb
;
192 dma_addr_t addr
= priv
->rx_ring
[wr_id
].mapping
;
194 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
195 if (wc
->status
!= IB_WC_WR_FLUSH_ERR
)
196 ipoib_warn(priv
, "failed recv event "
197 "(status=%d, wrid=%d vend_err %x)\n",
198 wc
->status
, wr_id
, wc
->vendor_err
);
199 dma_unmap_single(priv
->ca
->dma_device
, addr
,
200 IPOIB_BUF_SIZE
, DMA_FROM_DEVICE
);
201 dev_kfree_skb_any(skb
);
202 priv
->rx_ring
[wr_id
].skb
= NULL
;
207 * If we can't allocate a new RX buffer, dump
208 * this packet and reuse the old buffer.
210 if (unlikely(ipoib_alloc_rx_skb(dev
, wr_id
))) {
211 ++priv
->stats
.rx_dropped
;
215 ipoib_dbg_data(priv
, "received %d bytes, SLID 0x%04x\n",
216 wc
->byte_len
, wc
->slid
);
218 dma_unmap_single(priv
->ca
->dma_device
, addr
,
219 IPOIB_BUF_SIZE
, DMA_FROM_DEVICE
);
221 skb_put(skb
, wc
->byte_len
);
222 skb_pull(skb
, IB_GRH_BYTES
);
224 if (wc
->slid
!= priv
->local_lid
||
225 wc
->src_qp
!= priv
->qp
->qp_num
) {
226 skb
->protocol
= ((struct ipoib_header
*) skb
->data
)->proto
;
227 skb
->mac
.raw
= skb
->data
;
228 skb_pull(skb
, IPOIB_ENCAP_LEN
);
230 dev
->last_rx
= jiffies
;
231 ++priv
->stats
.rx_packets
;
232 priv
->stats
.rx_bytes
+= skb
->len
;
235 /* XXX get correct PACKET_ type here */
236 skb
->pkt_type
= PACKET_HOST
;
239 ipoib_dbg_data(priv
, "dropping loopback packet\n");
240 dev_kfree_skb_any(skb
);
244 if (unlikely(ipoib_ib_post_receive(dev
, wr_id
)))
245 ipoib_warn(priv
, "ipoib_ib_post_receive failed "
246 "for buf %d\n", wr_id
);
248 ipoib_warn(priv
, "completion event with wrid %d\n",
252 struct ipoib_tx_buf
*tx_req
;
255 if (wr_id
>= ipoib_sendq_size
) {
256 ipoib_warn(priv
, "completion event with wrid %d (> %d)\n",
257 wr_id
, ipoib_sendq_size
);
261 ipoib_dbg_data(priv
, "send complete, wrid %d\n", wr_id
);
263 tx_req
= &priv
->tx_ring
[wr_id
];
265 dma_unmap_single(priv
->ca
->dma_device
,
266 pci_unmap_addr(tx_req
, mapping
),
270 ++priv
->stats
.tx_packets
;
271 priv
->stats
.tx_bytes
+= tx_req
->skb
->len
;
273 dev_kfree_skb_any(tx_req
->skb
);
275 spin_lock_irqsave(&priv
->tx_lock
, flags
);
277 if (netif_queue_stopped(dev
) &&
278 test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
) &&
279 priv
->tx_head
- priv
->tx_tail
<= ipoib_sendq_size
>> 1)
280 netif_wake_queue(dev
);
281 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
283 if (wc
->status
!= IB_WC_SUCCESS
&&
284 wc
->status
!= IB_WC_WR_FLUSH_ERR
)
285 ipoib_warn(priv
, "failed send event "
286 "(status=%d, wrid=%d vend_err %x)\n",
287 wc
->status
, wr_id
, wc
->vendor_err
);
291 void ipoib_ib_completion(struct ib_cq
*cq
, void *dev_ptr
)
293 struct net_device
*dev
= (struct net_device
*) dev_ptr
;
294 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
297 ib_req_notify_cq(cq
, IB_CQ_NEXT_COMP
);
299 n
= ib_poll_cq(cq
, IPOIB_NUM_WC
, priv
->ibwc
);
300 for (i
= 0; i
< n
; ++i
)
301 ipoib_ib_handle_wc(dev
, priv
->ibwc
+ i
);
302 } while (n
== IPOIB_NUM_WC
);
305 static inline int post_send(struct ipoib_dev_priv
*priv
,
307 struct ib_ah
*address
, u32 qpn
,
308 dma_addr_t addr
, int len
)
310 struct ib_send_wr
*bad_wr
;
312 priv
->tx_sge
.addr
= addr
;
313 priv
->tx_sge
.length
= len
;
315 priv
->tx_wr
.wr_id
= wr_id
;
316 priv
->tx_wr
.wr
.ud
.remote_qpn
= qpn
;
317 priv
->tx_wr
.wr
.ud
.ah
= address
;
319 return ib_post_send(priv
->qp
, &priv
->tx_wr
, &bad_wr
);
322 void ipoib_send(struct net_device
*dev
, struct sk_buff
*skb
,
323 struct ipoib_ah
*address
, u32 qpn
)
325 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
326 struct ipoib_tx_buf
*tx_req
;
329 if (skb
->len
> dev
->mtu
+ INFINIBAND_ALEN
) {
330 ipoib_warn(priv
, "packet len %d (> %d) too long to send, dropping\n",
331 skb
->len
, dev
->mtu
+ INFINIBAND_ALEN
);
332 ++priv
->stats
.tx_dropped
;
333 ++priv
->stats
.tx_errors
;
334 dev_kfree_skb_any(skb
);
338 ipoib_dbg_data(priv
, "sending packet, length=%d address=%p qpn=0x%06x\n",
339 skb
->len
, address
, qpn
);
342 * We put the skb into the tx_ring _before_ we call post_send()
343 * because it's entirely possible that the completion handler will
344 * run before we execute anything after the post_send(). That
345 * means we have to make sure everything is properly recorded and
346 * our state is consistent before we call post_send().
348 tx_req
= &priv
->tx_ring
[priv
->tx_head
& (ipoib_sendq_size
- 1)];
350 addr
= dma_map_single(priv
->ca
->dma_device
, skb
->data
, skb
->len
,
352 pci_unmap_addr_set(tx_req
, mapping
, addr
);
354 if (unlikely(post_send(priv
, priv
->tx_head
& (ipoib_sendq_size
- 1),
355 address
->ah
, qpn
, addr
, skb
->len
))) {
356 ipoib_warn(priv
, "post_send failed\n");
357 ++priv
->stats
.tx_errors
;
358 dma_unmap_single(priv
->ca
->dma_device
, addr
, skb
->len
,
360 dev_kfree_skb_any(skb
);
362 dev
->trans_start
= jiffies
;
364 address
->last_send
= priv
->tx_head
;
367 if (priv
->tx_head
- priv
->tx_tail
== ipoib_sendq_size
) {
368 ipoib_dbg(priv
, "TX ring full, stopping kernel net queue\n");
369 netif_stop_queue(dev
);
374 static void __ipoib_reap_ah(struct net_device
*dev
)
376 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
377 struct ipoib_ah
*ah
, *tah
;
378 LIST_HEAD(remove_list
);
380 spin_lock_irq(&priv
->lock
);
381 list_for_each_entry_safe(ah
, tah
, &priv
->dead_ahs
, list
)
382 if ((int) priv
->tx_tail
- (int) ah
->last_send
>= 0) {
384 list_add_tail(&ah
->list
, &remove_list
);
386 spin_unlock_irq(&priv
->lock
);
388 list_for_each_entry_safe(ah
, tah
, &remove_list
, list
) {
389 ipoib_dbg(priv
, "Reaping ah %p\n", ah
->ah
);
390 ib_destroy_ah(ah
->ah
);
395 void ipoib_reap_ah(void *dev_ptr
)
397 struct net_device
*dev
= dev_ptr
;
398 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
400 __ipoib_reap_ah(dev
);
402 if (!test_bit(IPOIB_STOP_REAPER
, &priv
->flags
))
403 queue_delayed_work(ipoib_workqueue
, &priv
->ah_reap_task
, HZ
);
406 int ipoib_ib_dev_open(struct net_device
*dev
)
408 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
411 ret
= ipoib_init_qp(dev
);
413 ipoib_warn(priv
, "ipoib_init_qp returned %d\n", ret
);
417 ret
= ipoib_ib_post_receives(dev
);
419 ipoib_warn(priv
, "ipoib_ib_post_receives returned %d\n", ret
);
420 ipoib_ib_dev_stop(dev
);
424 clear_bit(IPOIB_STOP_REAPER
, &priv
->flags
);
425 queue_delayed_work(ipoib_workqueue
, &priv
->ah_reap_task
, HZ
);
427 set_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
);
432 static void ipoib_pkey_dev_check_presence(struct net_device
*dev
)
434 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
437 if (ib_find_cached_pkey(priv
->ca
, priv
->port
, priv
->pkey
, &pkey_index
))
438 clear_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
);
440 set_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
);
443 int ipoib_ib_dev_up(struct net_device
*dev
)
445 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
447 ipoib_pkey_dev_check_presence(dev
);
449 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
)) {
450 ipoib_dbg(priv
, "PKEY is not assigned.\n");
454 set_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
456 return ipoib_mcast_start_thread(dev
);
459 int ipoib_ib_dev_down(struct net_device
*dev
, int flush
)
461 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
463 ipoib_dbg(priv
, "downing ib_dev\n");
465 clear_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
466 netif_carrier_off(dev
);
468 /* Shutdown the P_Key thread if still active */
469 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
)) {
470 mutex_lock(&pkey_mutex
);
471 set_bit(IPOIB_PKEY_STOP
, &priv
->flags
);
472 cancel_delayed_work(&priv
->pkey_task
);
473 mutex_unlock(&pkey_mutex
);
475 flush_workqueue(ipoib_workqueue
);
478 ipoib_mcast_stop_thread(dev
, flush
);
479 ipoib_mcast_dev_flush(dev
);
481 ipoib_flush_paths(dev
);
486 static int recvs_pending(struct net_device
*dev
)
488 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
492 for (i
= 0; i
< ipoib_recvq_size
; ++i
)
493 if (priv
->rx_ring
[i
].skb
)
499 int ipoib_ib_dev_stop(struct net_device
*dev
)
501 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
502 struct ib_qp_attr qp_attr
;
504 struct ipoib_tx_buf
*tx_req
;
507 clear_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
);
510 * Move our QP to the error state and then reinitialize in
511 * when all work requests have completed or have been flushed.
513 qp_attr
.qp_state
= IB_QPS_ERR
;
514 if (ib_modify_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
))
515 ipoib_warn(priv
, "Failed to modify QP to ERROR state\n");
517 /* Wait for all sends and receives to complete */
520 while (priv
->tx_head
!= priv
->tx_tail
|| recvs_pending(dev
)) {
521 if (time_after(jiffies
, begin
+ 5 * HZ
)) {
522 ipoib_warn(priv
, "timing out; %d sends %d receives not completed\n",
523 priv
->tx_head
- priv
->tx_tail
, recvs_pending(dev
));
526 * assume the HW is wedged and just free up
527 * all our pending work requests.
529 while ((int) priv
->tx_tail
- (int) priv
->tx_head
< 0) {
530 tx_req
= &priv
->tx_ring
[priv
->tx_tail
&
531 (ipoib_sendq_size
- 1)];
532 dma_unmap_single(priv
->ca
->dma_device
,
533 pci_unmap_addr(tx_req
, mapping
),
536 dev_kfree_skb_any(tx_req
->skb
);
540 for (i
= 0; i
< ipoib_recvq_size
; ++i
)
541 if (priv
->rx_ring
[i
].skb
) {
542 dma_unmap_single(priv
->ca
->dma_device
,
543 pci_unmap_addr(&priv
->rx_ring
[i
],
547 dev_kfree_skb_any(priv
->rx_ring
[i
].skb
);
548 priv
->rx_ring
[i
].skb
= NULL
;
557 ipoib_dbg(priv
, "All sends and receives done.\n");
560 qp_attr
.qp_state
= IB_QPS_RESET
;
561 if (ib_modify_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
))
562 ipoib_warn(priv
, "Failed to modify QP to RESET state\n");
564 /* Wait for all AHs to be reaped */
565 set_bit(IPOIB_STOP_REAPER
, &priv
->flags
);
566 cancel_delayed_work(&priv
->ah_reap_task
);
567 flush_workqueue(ipoib_workqueue
);
571 while (!list_empty(&priv
->dead_ahs
)) {
572 __ipoib_reap_ah(dev
);
574 if (time_after(jiffies
, begin
+ HZ
)) {
575 ipoib_warn(priv
, "timing out; will leak address handles\n");
585 int ipoib_ib_dev_init(struct net_device
*dev
, struct ib_device
*ca
, int port
)
587 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
593 if (ipoib_transport_dev_init(dev
, ca
)) {
594 printk(KERN_WARNING
"%s: ipoib_transport_dev_init failed\n", ca
->name
);
598 if (dev
->flags
& IFF_UP
) {
599 if (ipoib_ib_dev_open(dev
)) {
600 ipoib_transport_dev_cleanup(dev
);
608 void ipoib_ib_dev_flush(void *_dev
)
610 struct net_device
*dev
= (struct net_device
*)_dev
;
611 struct ipoib_dev_priv
*priv
= netdev_priv(dev
), *cpriv
;
613 if (!test_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
) ) {
614 ipoib_dbg(priv
, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
618 if (!test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
)) {
619 ipoib_dbg(priv
, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
623 ipoib_dbg(priv
, "flushing\n");
625 ipoib_ib_dev_down(dev
, 0);
628 * The device could have been brought down between the start and when
629 * we get here, don't bring it back up if it's not configured up
631 if (test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
))
632 ipoib_ib_dev_up(dev
);
634 mutex_lock(&priv
->vlan_mutex
);
636 /* Flush any child interfaces too */
637 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
)
638 ipoib_ib_dev_flush(cpriv
->dev
);
640 mutex_unlock(&priv
->vlan_mutex
);
643 void ipoib_ib_dev_cleanup(struct net_device
*dev
)
645 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
647 ipoib_dbg(priv
, "cleaning up ib_dev\n");
649 ipoib_mcast_stop_thread(dev
, 1);
650 ipoib_mcast_dev_flush(dev
);
652 ipoib_transport_dev_cleanup(dev
);
656 * Delayed P_Key Assigment Interim Support
658 * The following is initial implementation of delayed P_Key assigment
659 * mechanism. It is using the same approach implemented for the multicast
660 * group join. The single goal of this implementation is to quickly address
661 * Bug #2507. This implementation will probably be removed when the P_Key
662 * change async notification is available.
665 void ipoib_pkey_poll(void *dev_ptr
)
667 struct net_device
*dev
= dev_ptr
;
668 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
670 ipoib_pkey_dev_check_presence(dev
);
672 if (test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
))
675 mutex_lock(&pkey_mutex
);
676 if (!test_bit(IPOIB_PKEY_STOP
, &priv
->flags
))
677 queue_delayed_work(ipoib_workqueue
,
680 mutex_unlock(&pkey_mutex
);
684 int ipoib_pkey_dev_delay_open(struct net_device
*dev
)
686 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
688 /* Look for the interface pkey value in the IB Port P_Key table and */
689 /* set the interface pkey assigment flag */
690 ipoib_pkey_dev_check_presence(dev
);
692 /* P_Key value not assigned yet - start polling */
693 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
)) {
694 mutex_lock(&pkey_mutex
);
695 clear_bit(IPOIB_PKEY_STOP
, &priv
->flags
);
696 queue_delayed_work(ipoib_workqueue
,
699 mutex_unlock(&pkey_mutex
);