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
36 #include <linux/delay.h>
37 #include <linux/moduleparam.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
42 #include <linux/tcp.h>
46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
47 static int data_debug_level
;
49 module_param(data_debug_level
, int, 0644);
50 MODULE_PARM_DESC(data_debug_level
,
51 "Enable data path debug tracing if > 0");
54 struct ipoib_ah
*ipoib_create_ah(struct net_device
*dev
,
55 struct ib_pd
*pd
, struct rdma_ah_attr
*attr
)
60 ah
= kmalloc(sizeof *ah
, GFP_KERNEL
);
62 return ERR_PTR(-ENOMEM
);
68 vah
= rdma_create_ah(pd
, attr
);
71 ah
= (struct ipoib_ah
*)vah
;
74 ipoib_dbg(ipoib_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
= ipoib_priv(ah
->dev
);
87 spin_lock_irqsave(&priv
->lock
, flags
);
88 list_add_tail(&ah
->list
, &priv
->dead_ahs
);
89 spin_unlock_irqrestore(&priv
->lock
, flags
);
92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv
*priv
,
93 u64 mapping
[IPOIB_UD_RX_SG
])
95 ib_dma_unmap_single(priv
->ca
, mapping
[0],
96 IPOIB_UD_BUF_SIZE(priv
->max_ib_mtu
),
100 static int ipoib_ib_post_receive(struct net_device
*dev
, int id
)
102 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
103 struct ib_recv_wr
*bad_wr
;
106 priv
->rx_wr
.wr_id
= id
| IPOIB_OP_RECV
;
107 priv
->rx_sge
[0].addr
= priv
->rx_ring
[id
].mapping
[0];
108 priv
->rx_sge
[1].addr
= priv
->rx_ring
[id
].mapping
[1];
111 ret
= ib_post_recv(priv
->qp
, &priv
->rx_wr
, &bad_wr
);
113 ipoib_warn(priv
, "receive failed for buf %d (%d)\n", id
, ret
);
114 ipoib_ud_dma_unmap_rx(priv
, priv
->rx_ring
[id
].mapping
);
115 dev_kfree_skb_any(priv
->rx_ring
[id
].skb
);
116 priv
->rx_ring
[id
].skb
= NULL
;
122 static struct sk_buff
*ipoib_alloc_rx_skb(struct net_device
*dev
, int id
)
124 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
129 buf_size
= IPOIB_UD_BUF_SIZE(priv
->max_ib_mtu
);
131 skb
= dev_alloc_skb(buf_size
+ IPOIB_HARD_LEN
);
136 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is
139 skb_reserve(skb
, sizeof(struct ipoib_pseudo_header
));
141 mapping
= priv
->rx_ring
[id
].mapping
;
142 mapping
[0] = ib_dma_map_single(priv
->ca
, skb
->data
, buf_size
,
144 if (unlikely(ib_dma_mapping_error(priv
->ca
, mapping
[0])))
147 priv
->rx_ring
[id
].skb
= skb
;
150 dev_kfree_skb_any(skb
);
154 static int ipoib_ib_post_receives(struct net_device
*dev
)
156 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
159 for (i
= 0; i
< ipoib_recvq_size
; ++i
) {
160 if (!ipoib_alloc_rx_skb(dev
, i
)) {
161 ipoib_warn(priv
, "failed to allocate receive buffer %d\n", i
);
164 if (ipoib_ib_post_receive(dev
, i
)) {
165 ipoib_warn(priv
, "ipoib_ib_post_receive failed for buf %d\n", i
);
173 static void ipoib_ib_handle_rx_wc(struct net_device
*dev
, struct ib_wc
*wc
)
175 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
176 unsigned int wr_id
= wc
->wr_id
& ~IPOIB_OP_RECV
;
178 u64 mapping
[IPOIB_UD_RX_SG
];
182 ipoib_dbg_data(priv
, "recv completion: id %d, status: %d\n",
185 if (unlikely(wr_id
>= ipoib_recvq_size
)) {
186 ipoib_warn(priv
, "recv completion event with wrid %d (> %d)\n",
187 wr_id
, ipoib_recvq_size
);
191 skb
= priv
->rx_ring
[wr_id
].skb
;
193 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
194 if (wc
->status
!= IB_WC_WR_FLUSH_ERR
)
195 ipoib_warn(priv
, "failed recv event "
196 "(status=%d, wrid=%d vend_err %x)\n",
197 wc
->status
, wr_id
, wc
->vendor_err
);
198 ipoib_ud_dma_unmap_rx(priv
, priv
->rx_ring
[wr_id
].mapping
);
199 dev_kfree_skb_any(skb
);
200 priv
->rx_ring
[wr_id
].skb
= NULL
;
204 memcpy(mapping
, priv
->rx_ring
[wr_id
].mapping
,
205 IPOIB_UD_RX_SG
* sizeof *mapping
);
208 * If we can't allocate a new RX buffer, dump
209 * this packet and reuse the old buffer.
211 if (unlikely(!ipoib_alloc_rx_skb(dev
, wr_id
))) {
212 ++dev
->stats
.rx_dropped
;
216 ipoib_dbg_data(priv
, "received %d bytes, SLID 0x%04x\n",
217 wc
->byte_len
, wc
->slid
);
219 ipoib_ud_dma_unmap_rx(priv
, mapping
);
221 skb_put(skb
, wc
->byte_len
);
223 /* First byte of dgid signals multicast when 0xff */
224 dgid
= &((struct ib_grh
*)skb
->data
)->dgid
;
226 if (!(wc
->wc_flags
& IB_WC_GRH
) || dgid
->raw
[0] != 0xff)
227 skb
->pkt_type
= PACKET_HOST
;
228 else if (memcmp(dgid
, dev
->broadcast
+ 4, sizeof(union ib_gid
)) == 0)
229 skb
->pkt_type
= PACKET_BROADCAST
;
231 skb
->pkt_type
= PACKET_MULTICAST
;
233 sgid
= &((struct ib_grh
*)skb
->data
)->sgid
;
236 * Drop packets that this interface sent, ie multicast packets
237 * that the HCA has replicated.
239 if (wc
->slid
== priv
->local_lid
&& wc
->src_qp
== priv
->qp
->qp_num
) {
242 if ((wc
->wc_flags
& IB_WC_GRH
) &&
243 sgid
->global
.interface_id
!= priv
->local_gid
.global
.interface_id
)
247 dev_kfree_skb_any(skb
);
252 skb_pull(skb
, IB_GRH_BYTES
);
254 skb
->protocol
= ((struct ipoib_header
*) skb
->data
)->proto
;
255 skb_add_pseudo_hdr(skb
);
257 ++dev
->stats
.rx_packets
;
258 dev
->stats
.rx_bytes
+= skb
->len
;
261 if ((dev
->features
& NETIF_F_RXCSUM
) &&
262 likely(wc
->wc_flags
& IB_WC_IP_CSUM_OK
))
263 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
265 napi_gro_receive(&priv
->napi
, skb
);
268 if (unlikely(ipoib_ib_post_receive(dev
, wr_id
)))
269 ipoib_warn(priv
, "ipoib_ib_post_receive failed "
270 "for buf %d\n", wr_id
);
273 int ipoib_dma_map_tx(struct ib_device
*ca
, struct ipoib_tx_buf
*tx_req
)
275 struct sk_buff
*skb
= tx_req
->skb
;
276 u64
*mapping
= tx_req
->mapping
;
280 if (skb_headlen(skb
)) {
281 mapping
[0] = ib_dma_map_single(ca
, skb
->data
, skb_headlen(skb
),
283 if (unlikely(ib_dma_mapping_error(ca
, mapping
[0])))
290 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; ++i
) {
291 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
292 mapping
[i
+ off
] = ib_dma_map_page(ca
,
294 frag
->page_offset
, skb_frag_size(frag
),
296 if (unlikely(ib_dma_mapping_error(ca
, mapping
[i
+ off
])))
303 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
- 1];
305 ib_dma_unmap_page(ca
, mapping
[i
- !off
], skb_frag_size(frag
), DMA_TO_DEVICE
);
309 ib_dma_unmap_single(ca
, mapping
[0], skb_headlen(skb
), DMA_TO_DEVICE
);
314 void ipoib_dma_unmap_tx(struct ipoib_dev_priv
*priv
,
315 struct ipoib_tx_buf
*tx_req
)
317 struct sk_buff
*skb
= tx_req
->skb
;
318 u64
*mapping
= tx_req
->mapping
;
322 if (skb_headlen(skb
)) {
323 ib_dma_unmap_single(priv
->ca
, mapping
[0], skb_headlen(skb
),
329 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; ++i
) {
330 const skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
332 ib_dma_unmap_page(priv
->ca
, mapping
[i
+ off
],
333 skb_frag_size(frag
), DMA_TO_DEVICE
);
338 * As the result of a completion error the QP Can be transferred to SQE states.
339 * The function checks if the (send)QP is in SQE state and
340 * moves it back to RTS state, that in order to have it functional again.
342 static void ipoib_qp_state_validate_work(struct work_struct
*work
)
344 struct ipoib_qp_state_validate
*qp_work
=
345 container_of(work
, struct ipoib_qp_state_validate
, work
);
347 struct ipoib_dev_priv
*priv
= qp_work
->priv
;
348 struct ib_qp_attr qp_attr
;
349 struct ib_qp_init_attr query_init_attr
;
352 ret
= ib_query_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
, &query_init_attr
);
354 ipoib_warn(priv
, "%s: Failed to query QP ret: %d\n",
358 pr_info("%s: QP: 0x%x is in state: %d\n",
359 __func__
, priv
->qp
->qp_num
, qp_attr
.qp_state
);
361 /* currently support only in SQE->RTS transition*/
362 if (qp_attr
.qp_state
== IB_QPS_SQE
) {
363 qp_attr
.qp_state
= IB_QPS_RTS
;
365 ret
= ib_modify_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
);
367 pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
368 ret
, priv
->qp
->qp_num
);
371 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
372 __func__
, priv
->qp
->qp_num
);
374 pr_warn("QP (%d) will stay in state: %d\n",
375 priv
->qp
->qp_num
, qp_attr
.qp_state
);
382 static void ipoib_ib_handle_tx_wc(struct net_device
*dev
, struct ib_wc
*wc
)
384 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
385 unsigned int wr_id
= wc
->wr_id
;
386 struct ipoib_tx_buf
*tx_req
;
388 ipoib_dbg_data(priv
, "send completion: id %d, status: %d\n",
391 if (unlikely(wr_id
>= ipoib_sendq_size
)) {
392 ipoib_warn(priv
, "send completion event with wrid %d (> %d)\n",
393 wr_id
, ipoib_sendq_size
);
397 tx_req
= &priv
->tx_ring
[wr_id
];
399 ipoib_dma_unmap_tx(priv
, tx_req
);
401 ++dev
->stats
.tx_packets
;
402 dev
->stats
.tx_bytes
+= tx_req
->skb
->len
;
404 dev_kfree_skb_any(tx_req
->skb
);
407 if (unlikely(--priv
->tx_outstanding
== ipoib_sendq_size
>> 1) &&
408 netif_queue_stopped(dev
) &&
409 test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
))
410 netif_wake_queue(dev
);
412 if (wc
->status
!= IB_WC_SUCCESS
&&
413 wc
->status
!= IB_WC_WR_FLUSH_ERR
) {
414 struct ipoib_qp_state_validate
*qp_work
;
415 ipoib_warn(priv
, "failed send event "
416 "(status=%d, wrid=%d vend_err %x)\n",
417 wc
->status
, wr_id
, wc
->vendor_err
);
418 qp_work
= kzalloc(sizeof(*qp_work
), GFP_ATOMIC
);
422 INIT_WORK(&qp_work
->work
, ipoib_qp_state_validate_work
);
423 qp_work
->priv
= priv
;
424 queue_work(priv
->wq
, &qp_work
->work
);
428 static int poll_tx(struct ipoib_dev_priv
*priv
)
432 n
= ib_poll_cq(priv
->send_cq
, MAX_SEND_CQE
, priv
->send_wc
);
433 for (i
= 0; i
< n
; ++i
)
434 ipoib_ib_handle_tx_wc(priv
->dev
, priv
->send_wc
+ i
);
436 return n
== MAX_SEND_CQE
;
439 int ipoib_poll(struct napi_struct
*napi
, int budget
)
441 struct ipoib_dev_priv
*priv
= container_of(napi
, struct ipoib_dev_priv
, napi
);
442 struct net_device
*dev
= priv
->dev
;
450 while (done
< budget
) {
451 int max
= (budget
- done
);
453 t
= min(IPOIB_NUM_WC
, max
);
454 n
= ib_poll_cq(priv
->recv_cq
, t
, priv
->ibwc
);
456 for (i
= 0; i
< n
; i
++) {
457 struct ib_wc
*wc
= priv
->ibwc
+ i
;
459 if (wc
->wr_id
& IPOIB_OP_RECV
) {
461 if (wc
->wr_id
& IPOIB_OP_CM
)
462 ipoib_cm_handle_rx_wc(dev
, wc
);
464 ipoib_ib_handle_rx_wc(dev
, wc
);
466 ipoib_cm_handle_tx_wc(priv
->dev
, wc
);
475 if (unlikely(ib_req_notify_cq(priv
->recv_cq
,
477 IB_CQ_REPORT_MISSED_EVENTS
)) &&
478 napi_reschedule(napi
))
485 void ipoib_ib_completion(struct ib_cq
*cq
, void *dev_ptr
)
487 struct net_device
*dev
= dev_ptr
;
488 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
490 napi_schedule(&priv
->napi
);
493 static void drain_tx_cq(struct net_device
*dev
)
495 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
498 while (poll_tx(priv
))
501 if (netif_queue_stopped(dev
))
502 mod_timer(&priv
->poll_timer
, jiffies
+ 1);
504 netif_tx_unlock(dev
);
507 void ipoib_send_comp_handler(struct ib_cq
*cq
, void *dev_ptr
)
509 struct ipoib_dev_priv
*priv
= ipoib_priv(dev_ptr
);
511 mod_timer(&priv
->poll_timer
, jiffies
);
514 static inline int post_send(struct ipoib_dev_priv
*priv
,
516 struct ib_ah
*address
, u32 dqpn
,
517 struct ipoib_tx_buf
*tx_req
,
518 void *head
, int hlen
)
520 struct ib_send_wr
*bad_wr
;
521 struct sk_buff
*skb
= tx_req
->skb
;
523 ipoib_build_sge(priv
, tx_req
);
525 priv
->tx_wr
.wr
.wr_id
= wr_id
;
526 priv
->tx_wr
.remote_qpn
= dqpn
;
527 priv
->tx_wr
.ah
= address
;
530 priv
->tx_wr
.mss
= skb_shinfo(skb
)->gso_size
;
531 priv
->tx_wr
.header
= head
;
532 priv
->tx_wr
.hlen
= hlen
;
533 priv
->tx_wr
.wr
.opcode
= IB_WR_LSO
;
535 priv
->tx_wr
.wr
.opcode
= IB_WR_SEND
;
537 return ib_post_send(priv
->qp
, &priv
->tx_wr
.wr
, &bad_wr
);
540 int ipoib_send(struct net_device
*dev
, struct sk_buff
*skb
,
541 struct ib_ah
*address
, u32 dqpn
)
543 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
544 struct ipoib_tx_buf
*tx_req
;
547 unsigned usable_sge
= priv
->max_send_sge
- !!skb_headlen(skb
);
549 if (skb_is_gso(skb
)) {
550 hlen
= skb_transport_offset(skb
) + tcp_hdrlen(skb
);
552 if (unlikely(!skb_pull(skb
, hlen
))) {
553 ipoib_warn(priv
, "linear data too small\n");
554 ++dev
->stats
.tx_dropped
;
555 ++dev
->stats
.tx_errors
;
556 dev_kfree_skb_any(skb
);
560 if (unlikely(skb
->len
> priv
->mcast_mtu
+ IPOIB_ENCAP_LEN
)) {
561 ipoib_warn(priv
, "packet len %d (> %d) too long to send, dropping\n",
562 skb
->len
, priv
->mcast_mtu
+ IPOIB_ENCAP_LEN
);
563 ++dev
->stats
.tx_dropped
;
564 ++dev
->stats
.tx_errors
;
565 ipoib_cm_skb_too_long(dev
, skb
, priv
->mcast_mtu
);
571 if (skb_shinfo(skb
)->nr_frags
> usable_sge
) {
572 if (skb_linearize(skb
) < 0) {
573 ipoib_warn(priv
, "skb could not be linearized\n");
574 ++dev
->stats
.tx_dropped
;
575 ++dev
->stats
.tx_errors
;
576 dev_kfree_skb_any(skb
);
579 /* Does skb_linearize return ok without reducing nr_frags? */
580 if (skb_shinfo(skb
)->nr_frags
> usable_sge
) {
581 ipoib_warn(priv
, "too many frags after skb linearize\n");
582 ++dev
->stats
.tx_dropped
;
583 ++dev
->stats
.tx_errors
;
584 dev_kfree_skb_any(skb
);
590 "sending packet, length=%d address=%p dqpn=0x%06x\n",
591 skb
->len
, address
, dqpn
);
594 * We put the skb into the tx_ring _before_ we call post_send()
595 * because it's entirely possible that the completion handler will
596 * run before we execute anything after the post_send(). That
597 * means we have to make sure everything is properly recorded and
598 * our state is consistent before we call post_send().
600 tx_req
= &priv
->tx_ring
[priv
->tx_head
& (ipoib_sendq_size
- 1)];
602 if (unlikely(ipoib_dma_map_tx(priv
->ca
, tx_req
))) {
603 ++dev
->stats
.tx_errors
;
604 dev_kfree_skb_any(skb
);
608 if (skb
->ip_summed
== CHECKSUM_PARTIAL
)
609 priv
->tx_wr
.wr
.send_flags
|= IB_SEND_IP_CSUM
;
611 priv
->tx_wr
.wr
.send_flags
&= ~IB_SEND_IP_CSUM
;
613 if (++priv
->tx_outstanding
== ipoib_sendq_size
) {
614 ipoib_dbg(priv
, "TX ring full, stopping kernel net queue\n");
615 if (ib_req_notify_cq(priv
->send_cq
, IB_CQ_NEXT_COMP
))
616 ipoib_warn(priv
, "request notify on send CQ failed\n");
617 netif_stop_queue(dev
);
623 rc
= post_send(priv
, priv
->tx_head
& (ipoib_sendq_size
- 1),
624 address
, dqpn
, tx_req
, phead
, hlen
);
626 ipoib_warn(priv
, "post_send failed, error %d\n", rc
);
627 ++dev
->stats
.tx_errors
;
628 --priv
->tx_outstanding
;
629 ipoib_dma_unmap_tx(priv
, tx_req
);
630 dev_kfree_skb_any(skb
);
631 if (netif_queue_stopped(dev
))
632 netif_wake_queue(dev
);
635 netif_trans_update(dev
);
641 if (unlikely(priv
->tx_outstanding
> MAX_SEND_CQE
))
642 while (poll_tx(priv
))
648 static void __ipoib_reap_ah(struct net_device
*dev
)
650 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
651 struct ipoib_ah
*ah
, *tah
;
652 LIST_HEAD(remove_list
);
655 netif_tx_lock_bh(dev
);
656 spin_lock_irqsave(&priv
->lock
, flags
);
658 list_for_each_entry_safe(ah
, tah
, &priv
->dead_ahs
, list
)
659 if ((int) priv
->tx_tail
- (int) ah
->last_send
>= 0) {
661 rdma_destroy_ah(ah
->ah
);
665 spin_unlock_irqrestore(&priv
->lock
, flags
);
666 netif_tx_unlock_bh(dev
);
669 void ipoib_reap_ah(struct work_struct
*work
)
671 struct ipoib_dev_priv
*priv
=
672 container_of(work
, struct ipoib_dev_priv
, ah_reap_task
.work
);
673 struct net_device
*dev
= priv
->dev
;
675 __ipoib_reap_ah(dev
);
677 if (!test_bit(IPOIB_STOP_REAPER
, &priv
->flags
))
678 queue_delayed_work(priv
->wq
, &priv
->ah_reap_task
,
679 round_jiffies_relative(HZ
));
682 static void ipoib_flush_ah(struct net_device
*dev
)
684 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
686 cancel_delayed_work(&priv
->ah_reap_task
);
687 flush_workqueue(priv
->wq
);
688 ipoib_reap_ah(&priv
->ah_reap_task
.work
);
691 static void ipoib_stop_ah(struct net_device
*dev
)
693 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
695 set_bit(IPOIB_STOP_REAPER
, &priv
->flags
);
699 static int recvs_pending(struct net_device
*dev
)
701 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
705 for (i
= 0; i
< ipoib_recvq_size
; ++i
)
706 if (priv
->rx_ring
[i
].skb
)
712 int ipoib_ib_dev_stop_default(struct net_device
*dev
)
714 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
715 struct ib_qp_attr qp_attr
;
717 struct ipoib_tx_buf
*tx_req
;
720 if (test_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
))
721 napi_disable(&priv
->napi
);
723 ipoib_cm_dev_stop(dev
);
726 * Move our QP to the error state and then reinitialize in
727 * when all work requests have completed or have been flushed.
729 qp_attr
.qp_state
= IB_QPS_ERR
;
730 if (ib_modify_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
))
731 ipoib_warn(priv
, "Failed to modify QP to ERROR state\n");
733 /* Wait for all sends and receives to complete */
736 while (priv
->tx_head
!= priv
->tx_tail
|| recvs_pending(dev
)) {
737 if (time_after(jiffies
, begin
+ 5 * HZ
)) {
739 "timing out; %d sends %d receives not completed\n",
740 priv
->tx_head
- priv
->tx_tail
,
744 * assume the HW is wedged and just free up
745 * all our pending work requests.
747 while ((int)priv
->tx_tail
- (int)priv
->tx_head
< 0) {
748 tx_req
= &priv
->tx_ring
[priv
->tx_tail
&
749 (ipoib_sendq_size
- 1)];
750 ipoib_dma_unmap_tx(priv
, tx_req
);
751 dev_kfree_skb_any(tx_req
->skb
);
753 --priv
->tx_outstanding
;
756 for (i
= 0; i
< ipoib_recvq_size
; ++i
) {
757 struct ipoib_rx_buf
*rx_req
;
759 rx_req
= &priv
->rx_ring
[i
];
762 ipoib_ud_dma_unmap_rx(priv
,
763 priv
->rx_ring
[i
].mapping
);
764 dev_kfree_skb_any(rx_req
->skb
);
776 ipoib_dbg(priv
, "All sends and receives done.\n");
779 del_timer_sync(&priv
->poll_timer
);
780 qp_attr
.qp_state
= IB_QPS_RESET
;
781 if (ib_modify_qp(priv
->qp
, &qp_attr
, IB_QP_STATE
))
782 ipoib_warn(priv
, "Failed to modify QP to RESET state\n");
784 ib_req_notify_cq(priv
->recv_cq
, IB_CQ_NEXT_COMP
);
789 int ipoib_ib_dev_stop(struct net_device
*dev
)
791 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
793 priv
->rn_ops
->ndo_stop(dev
);
795 clear_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
);
801 void ipoib_ib_tx_timer_func(unsigned long ctx
)
803 drain_tx_cq((struct net_device
*)ctx
);
806 int ipoib_ib_dev_open_default(struct net_device
*dev
)
808 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
811 ret
= ipoib_init_qp(dev
);
813 ipoib_warn(priv
, "ipoib_init_qp returned %d\n", ret
);
817 ret
= ipoib_ib_post_receives(dev
);
819 ipoib_warn(priv
, "ipoib_ib_post_receives returned %d\n", ret
);
823 ret
= ipoib_cm_dev_open(dev
);
825 ipoib_warn(priv
, "ipoib_cm_dev_open returned %d\n", ret
);
829 if (!test_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
))
830 napi_enable(&priv
->napi
);
837 int ipoib_ib_dev_open(struct net_device
*dev
)
839 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
841 ipoib_pkey_dev_check_presence(dev
);
843 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
)) {
844 ipoib_warn(priv
, "P_Key 0x%04x is %s\n", priv
->pkey
,
845 (!(priv
->pkey
& 0x7fff) ? "Invalid" : "not found"));
849 clear_bit(IPOIB_STOP_REAPER
, &priv
->flags
);
850 queue_delayed_work(priv
->wq
, &priv
->ah_reap_task
,
851 round_jiffies_relative(HZ
));
853 if (priv
->rn_ops
->ndo_open(dev
)) {
854 pr_warn("%s: Failed to open dev\n", dev
->name
);
858 set_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
);
863 set_bit(IPOIB_STOP_REAPER
, &priv
->flags
);
864 cancel_delayed_work(&priv
->ah_reap_task
);
865 set_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
);
866 ipoib_ib_dev_stop(dev
);
870 void ipoib_pkey_dev_check_presence(struct net_device
*dev
)
872 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
874 if (!(priv
->pkey
& 0x7fff) ||
875 ib_find_pkey(priv
->ca
, priv
->port
, priv
->pkey
,
877 clear_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
);
879 set_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
);
882 void ipoib_ib_dev_up(struct net_device
*dev
)
884 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
886 ipoib_pkey_dev_check_presence(dev
);
888 if (!test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
)) {
889 ipoib_dbg(priv
, "PKEY is not assigned.\n");
893 set_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
895 ipoib_mcast_start_thread(dev
);
898 void ipoib_ib_dev_down(struct net_device
*dev
)
900 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
902 ipoib_dbg(priv
, "downing ib_dev\n");
904 clear_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
905 netif_carrier_off(dev
);
907 ipoib_mcast_stop_thread(dev
);
908 ipoib_mcast_dev_flush(dev
);
910 ipoib_flush_paths(dev
);
913 void ipoib_drain_cq(struct net_device
*dev
)
915 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
919 * We call completion handling routines that expect to be
920 * called from the BH-disabled NAPI poll context, so disable
926 n
= ib_poll_cq(priv
->recv_cq
, IPOIB_NUM_WC
, priv
->ibwc
);
927 for (i
= 0; i
< n
; ++i
) {
929 * Convert any successful completions to flush
930 * errors to avoid passing packets up the
931 * stack after bringing the device down.
933 if (priv
->ibwc
[i
].status
== IB_WC_SUCCESS
)
934 priv
->ibwc
[i
].status
= IB_WC_WR_FLUSH_ERR
;
936 if (priv
->ibwc
[i
].wr_id
& IPOIB_OP_RECV
) {
937 if (priv
->ibwc
[i
].wr_id
& IPOIB_OP_CM
)
938 ipoib_cm_handle_rx_wc(dev
, priv
->ibwc
+ i
);
940 ipoib_ib_handle_rx_wc(dev
, priv
->ibwc
+ i
);
942 ipoib_cm_handle_tx_wc(dev
, priv
->ibwc
+ i
);
944 } while (n
== IPOIB_NUM_WC
);
946 while (poll_tx(priv
))
953 * Takes whatever value which is in pkey index 0 and updates priv->pkey
954 * returns 0 if the pkey value was changed.
956 static inline int update_parent_pkey(struct ipoib_dev_priv
*priv
)
961 prev_pkey
= priv
->pkey
;
962 result
= ib_query_pkey(priv
->ca
, priv
->port
, 0, &priv
->pkey
);
964 ipoib_warn(priv
, "ib_query_pkey port %d failed (ret = %d)\n",
969 priv
->pkey
|= 0x8000;
971 if (prev_pkey
!= priv
->pkey
) {
972 ipoib_dbg(priv
, "pkey changed from 0x%x to 0x%x\n",
973 prev_pkey
, priv
->pkey
);
975 * Update the pkey in the broadcast address, while making sure to set
976 * the full membership bit, so that we join the right broadcast group.
978 priv
->dev
->broadcast
[8] = priv
->pkey
>> 8;
979 priv
->dev
->broadcast
[9] = priv
->pkey
& 0xff;
982 * Update the broadcast address in the priv->broadcast object,
983 * in case it already exists, otherwise no one will do that.
985 if (priv
->broadcast
) {
986 spin_lock_irq(&priv
->lock
);
987 memcpy(priv
->broadcast
->mcmember
.mgid
.raw
,
988 priv
->dev
->broadcast
+ 4,
989 sizeof(union ib_gid
));
990 spin_unlock_irq(&priv
->lock
);
999 * returns 0 if pkey value was found in a different slot.
1001 static inline int update_child_pkey(struct ipoib_dev_priv
*priv
)
1003 u16 old_index
= priv
->pkey_index
;
1005 priv
->pkey_index
= 0;
1006 ipoib_pkey_dev_check_presence(priv
->dev
);
1008 if (test_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
) &&
1009 (old_index
== priv
->pkey_index
))
1015 * returns true if the device address of the ipoib interface has changed and the
1016 * new address is a valid one (i.e in the gid table), return false otherwise.
1018 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv
*priv
)
1020 union ib_gid search_gid
;
1022 union ib_gid
*netdev_gid
;
1028 netdev_gid
= (union ib_gid
*)(priv
->dev
->dev_addr
+ 4);
1029 if (ib_query_gid(priv
->ca
, priv
->port
, 0, &gid0
, NULL
))
1032 netif_addr_lock_bh(priv
->dev
);
1034 /* The subnet prefix may have changed, update it now so we won't have
1037 priv
->local_gid
.global
.subnet_prefix
= gid0
.global
.subnet_prefix
;
1038 netdev_gid
->global
.subnet_prefix
= gid0
.global
.subnet_prefix
;
1039 search_gid
.global
.subnet_prefix
= gid0
.global
.subnet_prefix
;
1041 search_gid
.global
.interface_id
= priv
->local_gid
.global
.interface_id
;
1043 netif_addr_unlock_bh(priv
->dev
);
1045 err
= ib_find_gid(priv
->ca
, &search_gid
, IB_GID_TYPE_IB
,
1046 priv
->dev
, &port
, &index
);
1048 netif_addr_lock_bh(priv
->dev
);
1050 if (search_gid
.global
.interface_id
!=
1051 priv
->local_gid
.global
.interface_id
)
1052 /* There was a change while we were looking up the gid, bail
1053 * here and let the next work sort this out
1057 /* The next section of code needs some background:
1058 * Per IB spec the port GUID can't change if the HCA is powered on.
1059 * port GUID is the basis for GID at index 0 which is the basis for
1060 * the default device address of a ipoib interface.
1062 * so it seems the flow should be:
1063 * if user_changed_dev_addr && gid in gid tbl
1064 * set bit dev_addr_set
1069 * The issue is that there are devices that don't follow the spec,
1070 * they change the port GUID when the HCA is powered, so in order
1071 * not to break userspace applications, We need to check if the
1072 * user wanted to control the device address and we assume that
1073 * if he sets the device address back to be based on GID index 0,
1074 * he no longer wishs to control it.
1076 * If the user doesn't control the the device address,
1077 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1078 * the port GUID has changed and GID at index 0 has changed
1079 * so we need to change priv->local_gid and priv->dev->dev_addr
1080 * to reflect the new GID.
1082 if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET
, &priv
->flags
)) {
1083 if (!err
&& port
== priv
->port
) {
1084 set_bit(IPOIB_FLAG_DEV_ADDR_SET
, &priv
->flags
);
1086 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL
,
1089 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL
, &priv
->flags
);
1095 if (!err
&& port
== priv
->port
) {
1098 if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL
, &priv
->flags
)) {
1099 memcpy(&priv
->local_gid
, &gid0
,
1100 sizeof(priv
->local_gid
));
1101 memcpy(priv
->dev
->dev_addr
+ 4, &gid0
,
1102 sizeof(priv
->local_gid
));
1109 netif_addr_unlock_bh(priv
->dev
);
1114 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv
*priv
,
1115 enum ipoib_flush_level level
,
1118 struct ipoib_dev_priv
*cpriv
;
1119 struct net_device
*dev
= priv
->dev
;
1122 down_read_nested(&priv
->vlan_rwsem
, nesting
);
1125 * Flush any child interfaces too -- they might be up even if
1126 * the parent is down.
1128 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
)
1129 __ipoib_ib_dev_flush(cpriv
, level
, nesting
+ 1);
1131 up_read(&priv
->vlan_rwsem
);
1133 if (!test_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
) &&
1134 level
!= IPOIB_FLUSH_HEAVY
) {
1135 /* Make sure the dev_addr is set even if not flushing */
1136 if (level
== IPOIB_FLUSH_LIGHT
)
1137 ipoib_dev_addr_changed_valid(priv
);
1138 ipoib_dbg(priv
, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1142 if (!test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
)) {
1143 /* interface is down. update pkey and leave. */
1144 if (level
== IPOIB_FLUSH_HEAVY
) {
1145 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
))
1146 update_parent_pkey(priv
);
1148 update_child_pkey(priv
);
1149 } else if (level
== IPOIB_FLUSH_LIGHT
)
1150 ipoib_dev_addr_changed_valid(priv
);
1151 ipoib_dbg(priv
, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1155 if (level
== IPOIB_FLUSH_HEAVY
) {
1156 /* child devices chase their origin pkey value, while non-child
1157 * (parent) devices should always takes what present in pkey index 0
1159 if (test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
1160 result
= update_child_pkey(priv
);
1162 /* restart QP only if P_Key index is changed */
1163 ipoib_dbg(priv
, "Not flushing - P_Key index not changed.\n");
1168 result
= update_parent_pkey(priv
);
1169 /* restart QP only if P_Key value changed */
1171 ipoib_dbg(priv
, "Not flushing - P_Key value not changed.\n");
1177 if (level
== IPOIB_FLUSH_LIGHT
) {
1179 ipoib_mark_paths_invalid(dev
);
1180 /* Set IPoIB operation as down to prevent races between:
1181 * the flush flow which leaves MCG and on the fly joins
1182 * which can happen during that time. mcast restart task
1183 * should deal with join requests we missed.
1185 oper_up
= test_and_clear_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
1186 ipoib_mcast_dev_flush(dev
);
1188 set_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
);
1189 ipoib_flush_ah(dev
);
1192 if (level
>= IPOIB_FLUSH_NORMAL
)
1193 ipoib_ib_dev_down(dev
);
1195 if (level
== IPOIB_FLUSH_HEAVY
) {
1196 if (test_bit(IPOIB_FLAG_INITIALIZED
, &priv
->flags
))
1197 ipoib_ib_dev_stop(dev
);
1198 if (ipoib_ib_dev_open(dev
) != 0)
1200 if (netif_queue_stopped(dev
))
1201 netif_start_queue(dev
);
1205 * The device could have been brought down between the start and when
1206 * we get here, don't bring it back up if it's not configured up
1208 if (test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
)) {
1209 if (level
>= IPOIB_FLUSH_NORMAL
)
1210 ipoib_ib_dev_up(dev
);
1211 if (ipoib_dev_addr_changed_valid(priv
))
1212 ipoib_mcast_restart_task(&priv
->restart_task
);
1216 void ipoib_ib_dev_flush_light(struct work_struct
*work
)
1218 struct ipoib_dev_priv
*priv
=
1219 container_of(work
, struct ipoib_dev_priv
, flush_light
);
1221 __ipoib_ib_dev_flush(priv
, IPOIB_FLUSH_LIGHT
, 0);
1224 void ipoib_ib_dev_flush_normal(struct work_struct
*work
)
1226 struct ipoib_dev_priv
*priv
=
1227 container_of(work
, struct ipoib_dev_priv
, flush_normal
);
1229 __ipoib_ib_dev_flush(priv
, IPOIB_FLUSH_NORMAL
, 0);
1232 void ipoib_ib_dev_flush_heavy(struct work_struct
*work
)
1234 struct ipoib_dev_priv
*priv
=
1235 container_of(work
, struct ipoib_dev_priv
, flush_heavy
);
1237 __ipoib_ib_dev_flush(priv
, IPOIB_FLUSH_HEAVY
, 0);
1240 void ipoib_ib_dev_cleanup(struct net_device
*dev
)
1242 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1244 ipoib_dbg(priv
, "cleaning up ib_dev\n");
1246 * We must make sure there are no more (path) completions
1247 * that may wish to touch priv fields that are no longer valid
1249 ipoib_flush_paths(dev
);
1251 ipoib_mcast_stop_thread(dev
);
1252 ipoib_mcast_dev_flush(dev
);
1255 * All of our ah references aren't free until after
1256 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1257 * the neighbor garbage collection is stopped and reaped.
1258 * That should all be done now, so make a final ah flush.
1262 clear_bit(IPOIB_PKEY_ASSIGNED
, &priv
->flags
);
1264 priv
->rn_ops
->ndo_uninit(dev
);
1267 ib_dealloc_pd(priv
->pd
);