2 * Copyright (c) 2006 Mellanox Technologies. All rights reserved
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <rdma/ib_cm.h>
36 #include <linux/icmpv6.h>
37 #include <linux/delay.h>
38 #include <linux/slab.h>
39 #include <linux/vmalloc.h>
40 #include <linux/moduleparam.h>
41 #include <linux/sched/signal.h>
42 #include <linux/sched/mm.h>
46 int ipoib_max_conn_qp
= 128;
48 module_param_named(max_nonsrq_conn_qp
, ipoib_max_conn_qp
, int, 0444);
49 MODULE_PARM_DESC(max_nonsrq_conn_qp
,
50 "Max number of connected-mode QPs per interface "
51 "(applied only if shared receive queue is not available)");
53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
54 static int data_debug_level
;
56 module_param_named(cm_data_debug_level
, data_debug_level
, int, 0644);
57 MODULE_PARM_DESC(cm_data_debug_level
,
58 "Enable data path debug tracing for connected mode if > 0");
61 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
63 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
64 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ)
65 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ)
66 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
68 #define IPOIB_CM_RX_RESERVE (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
70 static struct ib_qp_attr ipoib_cm_err_attr
= {
71 .qp_state
= IB_QPS_ERR
74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
76 static struct ib_send_wr ipoib_cm_rx_drain_wr
= {
80 static int ipoib_cm_tx_handler(struct ib_cm_id
*cm_id
,
81 struct ib_cm_event
*event
);
83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv
*priv
, int frags
,
84 u64 mapping
[IPOIB_CM_RX_SG
])
88 ib_dma_unmap_single(priv
->ca
, mapping
[0], IPOIB_CM_HEAD_SIZE
, DMA_FROM_DEVICE
);
90 for (i
= 0; i
< frags
; ++i
)
91 ib_dma_unmap_page(priv
->ca
, mapping
[i
+ 1], PAGE_SIZE
, DMA_FROM_DEVICE
);
94 static int ipoib_cm_post_receive_srq(struct net_device
*dev
, int id
)
96 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
97 struct ib_recv_wr
*bad_wr
;
100 priv
->cm
.rx_wr
.wr_id
= id
| IPOIB_OP_CM
| IPOIB_OP_RECV
;
102 for (i
= 0; i
< priv
->cm
.num_frags
; ++i
)
103 priv
->cm
.rx_sge
[i
].addr
= priv
->cm
.srq_ring
[id
].mapping
[i
];
105 ret
= ib_post_srq_recv(priv
->cm
.srq
, &priv
->cm
.rx_wr
, &bad_wr
);
107 ipoib_warn(priv
, "post srq failed for buf %d (%d)\n", id
, ret
);
108 ipoib_cm_dma_unmap_rx(priv
, priv
->cm
.num_frags
- 1,
109 priv
->cm
.srq_ring
[id
].mapping
);
110 dev_kfree_skb_any(priv
->cm
.srq_ring
[id
].skb
);
111 priv
->cm
.srq_ring
[id
].skb
= NULL
;
117 static int ipoib_cm_post_receive_nonsrq(struct net_device
*dev
,
118 struct ipoib_cm_rx
*rx
,
119 struct ib_recv_wr
*wr
,
120 struct ib_sge
*sge
, int id
)
122 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
123 struct ib_recv_wr
*bad_wr
;
126 wr
->wr_id
= id
| IPOIB_OP_CM
| IPOIB_OP_RECV
;
128 for (i
= 0; i
< IPOIB_CM_RX_SG
; ++i
)
129 sge
[i
].addr
= rx
->rx_ring
[id
].mapping
[i
];
131 ret
= ib_post_recv(rx
->qp
, wr
, &bad_wr
);
133 ipoib_warn(priv
, "post recv failed for buf %d (%d)\n", id
, ret
);
134 ipoib_cm_dma_unmap_rx(priv
, IPOIB_CM_RX_SG
- 1,
135 rx
->rx_ring
[id
].mapping
);
136 dev_kfree_skb_any(rx
->rx_ring
[id
].skb
);
137 rx
->rx_ring
[id
].skb
= NULL
;
143 static struct sk_buff
*ipoib_cm_alloc_rx_skb(struct net_device
*dev
,
144 struct ipoib_cm_rx_buf
*rx_ring
,
146 u64 mapping
[IPOIB_CM_RX_SG
],
149 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
153 skb
= dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE
+ IPOIB_PSEUDO_LEN
, 16));
158 * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
159 * IP header to a multiple of 16.
161 skb_reserve(skb
, IPOIB_CM_RX_RESERVE
);
163 mapping
[0] = ib_dma_map_single(priv
->ca
, skb
->data
, IPOIB_CM_HEAD_SIZE
,
165 if (unlikely(ib_dma_mapping_error(priv
->ca
, mapping
[0]))) {
166 dev_kfree_skb_any(skb
);
170 for (i
= 0; i
< frags
; i
++) {
171 struct page
*page
= alloc_page(gfp
);
175 skb_fill_page_desc(skb
, i
, page
, 0, PAGE_SIZE
);
177 mapping
[i
+ 1] = ib_dma_map_page(priv
->ca
, page
,
178 0, PAGE_SIZE
, DMA_FROM_DEVICE
);
179 if (unlikely(ib_dma_mapping_error(priv
->ca
, mapping
[i
+ 1])))
183 rx_ring
[id
].skb
= skb
;
188 ib_dma_unmap_single(priv
->ca
, mapping
[0], IPOIB_CM_HEAD_SIZE
, DMA_FROM_DEVICE
);
191 ib_dma_unmap_page(priv
->ca
, mapping
[i
], PAGE_SIZE
, DMA_FROM_DEVICE
);
193 dev_kfree_skb_any(skb
);
197 static void ipoib_cm_free_rx_ring(struct net_device
*dev
,
198 struct ipoib_cm_rx_buf
*rx_ring
)
200 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
203 for (i
= 0; i
< ipoib_recvq_size
; ++i
)
204 if (rx_ring
[i
].skb
) {
205 ipoib_cm_dma_unmap_rx(priv
, IPOIB_CM_RX_SG
- 1,
207 dev_kfree_skb_any(rx_ring
[i
].skb
);
213 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv
*priv
)
215 struct ib_send_wr
*bad_wr
;
216 struct ipoib_cm_rx
*p
;
218 /* We only reserved 1 extra slot in CQ for drain WRs, so
219 * make sure we have at most 1 outstanding WR. */
220 if (list_empty(&priv
->cm
.rx_flush_list
) ||
221 !list_empty(&priv
->cm
.rx_drain_list
))
225 * QPs on flush list are error state. This way, a "flush
226 * error" WC will be immediately generated for each WR we post.
228 p
= list_entry(priv
->cm
.rx_flush_list
.next
, typeof(*p
), list
);
229 ipoib_cm_rx_drain_wr
.wr_id
= IPOIB_CM_RX_DRAIN_WRID
;
230 if (ib_post_send(p
->qp
, &ipoib_cm_rx_drain_wr
, &bad_wr
))
231 ipoib_warn(priv
, "failed to post drain wr\n");
233 list_splice_init(&priv
->cm
.rx_flush_list
, &priv
->cm
.rx_drain_list
);
236 static void ipoib_cm_rx_event_handler(struct ib_event
*event
, void *ctx
)
238 struct ipoib_cm_rx
*p
= ctx
;
239 struct ipoib_dev_priv
*priv
= ipoib_priv(p
->dev
);
242 if (event
->event
!= IB_EVENT_QP_LAST_WQE_REACHED
)
245 spin_lock_irqsave(&priv
->lock
, flags
);
246 list_move(&p
->list
, &priv
->cm
.rx_flush_list
);
247 p
->state
= IPOIB_CM_RX_FLUSH
;
248 ipoib_cm_start_rx_drain(priv
);
249 spin_unlock_irqrestore(&priv
->lock
, flags
);
252 static struct ib_qp
*ipoib_cm_create_rx_qp(struct net_device
*dev
,
253 struct ipoib_cm_rx
*p
)
255 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
256 struct ib_qp_init_attr attr
= {
257 .event_handler
= ipoib_cm_rx_event_handler
,
258 .send_cq
= priv
->recv_cq
, /* For drain WR */
259 .recv_cq
= priv
->recv_cq
,
261 .cap
.max_send_wr
= 1, /* For drain WR */
262 .cap
.max_send_sge
= 1, /* FIXME: 0 Seems not to work */
263 .sq_sig_type
= IB_SIGNAL_ALL_WR
,
264 .qp_type
= IB_QPT_RC
,
268 if (!ipoib_cm_has_srq(dev
)) {
269 attr
.cap
.max_recv_wr
= ipoib_recvq_size
;
270 attr
.cap
.max_recv_sge
= IPOIB_CM_RX_SG
;
273 return ib_create_qp(priv
->pd
, &attr
);
276 static int ipoib_cm_modify_rx_qp(struct net_device
*dev
,
277 struct ib_cm_id
*cm_id
, struct ib_qp
*qp
,
280 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
281 struct ib_qp_attr qp_attr
;
282 int qp_attr_mask
, ret
;
284 qp_attr
.qp_state
= IB_QPS_INIT
;
285 ret
= ib_cm_init_qp_attr(cm_id
, &qp_attr
, &qp_attr_mask
);
287 ipoib_warn(priv
, "failed to init QP attr for INIT: %d\n", ret
);
290 ret
= ib_modify_qp(qp
, &qp_attr
, qp_attr_mask
);
292 ipoib_warn(priv
, "failed to modify QP to INIT: %d\n", ret
);
295 qp_attr
.qp_state
= IB_QPS_RTR
;
296 ret
= ib_cm_init_qp_attr(cm_id
, &qp_attr
, &qp_attr_mask
);
298 ipoib_warn(priv
, "failed to init QP attr for RTR: %d\n", ret
);
301 qp_attr
.rq_psn
= psn
;
302 ret
= ib_modify_qp(qp
, &qp_attr
, qp_attr_mask
);
304 ipoib_warn(priv
, "failed to modify QP to RTR: %d\n", ret
);
309 * Current Mellanox HCA firmware won't generate completions
310 * with error for drain WRs unless the QP has been moved to
311 * RTS first. This work-around leaves a window where a QP has
312 * moved to error asynchronously, but this will eventually get
313 * fixed in firmware, so let's not error out if modify QP
316 qp_attr
.qp_state
= IB_QPS_RTS
;
317 ret
= ib_cm_init_qp_attr(cm_id
, &qp_attr
, &qp_attr_mask
);
319 ipoib_warn(priv
, "failed to init QP attr for RTS: %d\n", ret
);
322 ret
= ib_modify_qp(qp
, &qp_attr
, qp_attr_mask
);
324 ipoib_warn(priv
, "failed to modify QP to RTS: %d\n", ret
);
331 static void ipoib_cm_init_rx_wr(struct net_device
*dev
,
332 struct ib_recv_wr
*wr
,
335 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
338 for (i
= 0; i
< priv
->cm
.num_frags
; ++i
)
339 sge
[i
].lkey
= priv
->pd
->local_dma_lkey
;
341 sge
[0].length
= IPOIB_CM_HEAD_SIZE
;
342 for (i
= 1; i
< priv
->cm
.num_frags
; ++i
)
343 sge
[i
].length
= PAGE_SIZE
;
347 wr
->num_sge
= priv
->cm
.num_frags
;
350 static int ipoib_cm_nonsrq_init_rx(struct net_device
*dev
, struct ib_cm_id
*cm_id
,
351 struct ipoib_cm_rx
*rx
)
353 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
355 struct ib_recv_wr wr
;
356 struct ib_sge sge
[IPOIB_CM_RX_SG
];
361 rx
->rx_ring
= vzalloc(ipoib_recvq_size
* sizeof *rx
->rx_ring
);
365 t
= kmalloc(sizeof *t
, GFP_KERNEL
);
371 ipoib_cm_init_rx_wr(dev
, &t
->wr
, t
->sge
);
373 spin_lock_irq(&priv
->lock
);
375 if (priv
->cm
.nonsrq_conn_qp
>= ipoib_max_conn_qp
) {
376 spin_unlock_irq(&priv
->lock
);
377 ib_send_cm_rej(cm_id
, IB_CM_REJ_NO_QP
, NULL
, 0, NULL
, 0);
381 ++priv
->cm
.nonsrq_conn_qp
;
383 spin_unlock_irq(&priv
->lock
);
385 for (i
= 0; i
< ipoib_recvq_size
; ++i
) {
386 if (!ipoib_cm_alloc_rx_skb(dev
, rx
->rx_ring
, i
, IPOIB_CM_RX_SG
- 1,
387 rx
->rx_ring
[i
].mapping
,
389 ipoib_warn(priv
, "failed to allocate receive buffer %d\n", i
);
393 ret
= ipoib_cm_post_receive_nonsrq(dev
, rx
, &t
->wr
, t
->sge
, i
);
395 ipoib_warn(priv
, "ipoib_cm_post_receive_nonsrq "
396 "failed for buf %d\n", i
);
402 rx
->recv_count
= ipoib_recvq_size
;
409 spin_lock_irq(&priv
->lock
);
410 --priv
->cm
.nonsrq_conn_qp
;
411 spin_unlock_irq(&priv
->lock
);
417 ipoib_cm_free_rx_ring(dev
, rx
->rx_ring
);
422 static int ipoib_cm_send_rep(struct net_device
*dev
, struct ib_cm_id
*cm_id
,
423 struct ib_qp
*qp
, struct ib_cm_req_event_param
*req
,
426 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
427 struct ipoib_cm_data data
= {};
428 struct ib_cm_rep_param rep
= {};
430 data
.qpn
= cpu_to_be32(priv
->qp
->qp_num
);
431 data
.mtu
= cpu_to_be32(IPOIB_CM_BUF_SIZE
);
433 rep
.private_data
= &data
;
434 rep
.private_data_len
= sizeof data
;
435 rep
.flow_control
= 0;
436 rep
.rnr_retry_count
= req
->rnr_retry_count
;
437 rep
.srq
= ipoib_cm_has_srq(dev
);
438 rep
.qp_num
= qp
->qp_num
;
439 rep
.starting_psn
= psn
;
440 return ib_send_cm_rep(cm_id
, &rep
);
443 static int ipoib_cm_req_handler(struct ib_cm_id
*cm_id
, struct ib_cm_event
*event
)
445 struct net_device
*dev
= cm_id
->context
;
446 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
447 struct ipoib_cm_rx
*p
;
451 ipoib_dbg(priv
, "REQ arrived\n");
452 p
= kzalloc(sizeof *p
, GFP_KERNEL
);
458 p
->state
= IPOIB_CM_RX_LIVE
;
459 p
->jiffies
= jiffies
;
460 INIT_LIST_HEAD(&p
->list
);
462 p
->qp
= ipoib_cm_create_rx_qp(dev
, p
);
464 ret
= PTR_ERR(p
->qp
);
468 psn
= prandom_u32() & 0xffffff;
469 ret
= ipoib_cm_modify_rx_qp(dev
, cm_id
, p
->qp
, psn
);
473 if (!ipoib_cm_has_srq(dev
)) {
474 ret
= ipoib_cm_nonsrq_init_rx(dev
, cm_id
, p
);
479 spin_lock_irq(&priv
->lock
);
480 queue_delayed_work(priv
->wq
,
481 &priv
->cm
.stale_task
, IPOIB_CM_RX_DELAY
);
482 /* Add this entry to passive ids list head, but do not re-add it
483 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
484 p
->jiffies
= jiffies
;
485 if (p
->state
== IPOIB_CM_RX_LIVE
)
486 list_move(&p
->list
, &priv
->cm
.passive_ids
);
487 spin_unlock_irq(&priv
->lock
);
489 ret
= ipoib_cm_send_rep(dev
, cm_id
, p
->qp
, &event
->param
.req_rcvd
, psn
);
491 ipoib_warn(priv
, "failed to send REP: %d\n", ret
);
492 if (ib_modify_qp(p
->qp
, &ipoib_cm_err_attr
, IB_QP_STATE
))
493 ipoib_warn(priv
, "unable to move qp to error state\n");
498 ib_destroy_qp(p
->qp
);
504 static int ipoib_cm_rx_handler(struct ib_cm_id
*cm_id
,
505 struct ib_cm_event
*event
)
507 struct ipoib_cm_rx
*p
;
508 struct ipoib_dev_priv
*priv
;
510 switch (event
->event
) {
511 case IB_CM_REQ_RECEIVED
:
512 return ipoib_cm_req_handler(cm_id
, event
);
513 case IB_CM_DREQ_RECEIVED
:
514 ib_send_cm_drep(cm_id
, NULL
, 0);
516 case IB_CM_REJ_RECEIVED
:
518 priv
= ipoib_priv(p
->dev
);
519 if (ib_modify_qp(p
->qp
, &ipoib_cm_err_attr
, IB_QP_STATE
))
520 ipoib_warn(priv
, "unable to move qp to error state\n");
526 /* Adjust length of skb with fragments to match received data */
527 static void skb_put_frags(struct sk_buff
*skb
, unsigned int hdr_space
,
528 unsigned int length
, struct sk_buff
*toskb
)
533 /* put header into skb */
534 size
= min(length
, hdr_space
);
539 num_frags
= skb_shinfo(skb
)->nr_frags
;
540 for (i
= 0; i
< num_frags
; i
++) {
541 skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
544 /* don't need this page */
545 skb_fill_page_desc(toskb
, i
, skb_frag_page(frag
),
547 --skb_shinfo(skb
)->nr_frags
;
549 size
= min(length
, (unsigned) PAGE_SIZE
);
551 skb_frag_size_set(frag
, size
);
552 skb
->data_len
+= size
;
553 skb
->truesize
+= size
;
560 void ipoib_cm_handle_rx_wc(struct net_device
*dev
, struct ib_wc
*wc
)
562 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
563 struct ipoib_cm_rx_buf
*rx_ring
;
564 unsigned int wr_id
= wc
->wr_id
& ~(IPOIB_OP_CM
| IPOIB_OP_RECV
);
565 struct sk_buff
*skb
, *newskb
;
566 struct ipoib_cm_rx
*p
;
568 u64 mapping
[IPOIB_CM_RX_SG
];
571 struct sk_buff
*small_skb
;
573 ipoib_dbg_data(priv
, "cm recv completion: id %d, status: %d\n",
576 if (unlikely(wr_id
>= ipoib_recvq_size
)) {
577 if (wr_id
== (IPOIB_CM_RX_DRAIN_WRID
& ~(IPOIB_OP_CM
| IPOIB_OP_RECV
))) {
578 spin_lock_irqsave(&priv
->lock
, flags
);
579 list_splice_init(&priv
->cm
.rx_drain_list
, &priv
->cm
.rx_reap_list
);
580 ipoib_cm_start_rx_drain(priv
);
581 queue_work(priv
->wq
, &priv
->cm
.rx_reap_task
);
582 spin_unlock_irqrestore(&priv
->lock
, flags
);
584 ipoib_warn(priv
, "cm recv completion event with wrid %d (> %d)\n",
585 wr_id
, ipoib_recvq_size
);
589 p
= wc
->qp
->qp_context
;
591 has_srq
= ipoib_cm_has_srq(dev
);
592 rx_ring
= has_srq
? priv
->cm
.srq_ring
: p
->rx_ring
;
594 skb
= rx_ring
[wr_id
].skb
;
596 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
598 "cm recv error (status=%d, wrid=%d vend_err %#x)\n",
599 wc
->status
, wr_id
, wc
->vendor_err
);
600 ++dev
->stats
.rx_dropped
;
604 if (!--p
->recv_count
) {
605 spin_lock_irqsave(&priv
->lock
, flags
);
606 list_move(&p
->list
, &priv
->cm
.rx_reap_list
);
607 spin_unlock_irqrestore(&priv
->lock
, flags
);
608 queue_work(priv
->wq
, &priv
->cm
.rx_reap_task
);
614 if (unlikely(!(wr_id
& IPOIB_CM_RX_UPDATE_MASK
))) {
615 if (p
&& time_after_eq(jiffies
, p
->jiffies
+ IPOIB_CM_RX_UPDATE_TIME
)) {
616 spin_lock_irqsave(&priv
->lock
, flags
);
617 p
->jiffies
= jiffies
;
618 /* Move this entry to list head, but do not re-add it
619 * if it has been moved out of list. */
620 if (p
->state
== IPOIB_CM_RX_LIVE
)
621 list_move(&p
->list
, &priv
->cm
.passive_ids
);
622 spin_unlock_irqrestore(&priv
->lock
, flags
);
626 if (wc
->byte_len
< IPOIB_CM_COPYBREAK
) {
627 int dlen
= wc
->byte_len
;
629 small_skb
= dev_alloc_skb(dlen
+ IPOIB_CM_RX_RESERVE
);
631 skb_reserve(small_skb
, IPOIB_CM_RX_RESERVE
);
632 ib_dma_sync_single_for_cpu(priv
->ca
, rx_ring
[wr_id
].mapping
[0],
633 dlen
, DMA_FROM_DEVICE
);
634 skb_copy_from_linear_data(skb
, small_skb
->data
, dlen
);
635 ib_dma_sync_single_for_device(priv
->ca
, rx_ring
[wr_id
].mapping
[0],
636 dlen
, DMA_FROM_DEVICE
);
637 skb_put(small_skb
, dlen
);
643 frags
= PAGE_ALIGN(wc
->byte_len
- min(wc
->byte_len
,
644 (unsigned)IPOIB_CM_HEAD_SIZE
)) / PAGE_SIZE
;
646 newskb
= ipoib_cm_alloc_rx_skb(dev
, rx_ring
, wr_id
, frags
,
647 mapping
, GFP_ATOMIC
);
648 if (unlikely(!newskb
)) {
650 * If we can't allocate a new RX buffer, dump
651 * this packet and reuse the old buffer.
653 ipoib_dbg(priv
, "failed to allocate receive buffer %d\n", wr_id
);
654 ++dev
->stats
.rx_dropped
;
658 ipoib_cm_dma_unmap_rx(priv
, frags
, rx_ring
[wr_id
].mapping
);
659 memcpy(rx_ring
[wr_id
].mapping
, mapping
, (frags
+ 1) * sizeof *mapping
);
661 ipoib_dbg_data(priv
, "received %d bytes, SLID 0x%04x\n",
662 wc
->byte_len
, wc
->slid
);
664 skb_put_frags(skb
, IPOIB_CM_HEAD_SIZE
, wc
->byte_len
, newskb
);
667 skb
->protocol
= ((struct ipoib_header
*) skb
->data
)->proto
;
668 skb_add_pseudo_hdr(skb
);
670 ++dev
->stats
.rx_packets
;
671 dev
->stats
.rx_bytes
+= skb
->len
;
674 /* XXX get correct PACKET_ type here */
675 skb
->pkt_type
= PACKET_HOST
;
676 netif_receive_skb(skb
);
680 if (unlikely(ipoib_cm_post_receive_srq(dev
, wr_id
)))
681 ipoib_warn(priv
, "ipoib_cm_post_receive_srq failed "
682 "for buf %d\n", wr_id
);
684 if (unlikely(ipoib_cm_post_receive_nonsrq(dev
, p
,
689 ipoib_warn(priv
, "ipoib_cm_post_receive_nonsrq failed "
690 "for buf %d\n", wr_id
);
695 static inline int post_send(struct ipoib_dev_priv
*priv
,
696 struct ipoib_cm_tx
*tx
,
698 struct ipoib_tx_buf
*tx_req
)
700 struct ib_send_wr
*bad_wr
;
702 ipoib_build_sge(priv
, tx_req
);
704 priv
->tx_wr
.wr
.wr_id
= wr_id
| IPOIB_OP_CM
;
706 return ib_post_send(tx
->qp
, &priv
->tx_wr
.wr
, &bad_wr
);
709 void ipoib_cm_send(struct net_device
*dev
, struct sk_buff
*skb
, struct ipoib_cm_tx
*tx
)
711 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
712 struct ipoib_tx_buf
*tx_req
;
714 unsigned usable_sge
= tx
->max_send_sge
- !!skb_headlen(skb
);
716 if (unlikely(skb
->len
> tx
->mtu
)) {
717 ipoib_warn(priv
, "packet len %d (> %d) too long to send, dropping\n",
719 ++dev
->stats
.tx_dropped
;
720 ++dev
->stats
.tx_errors
;
721 ipoib_cm_skb_too_long(dev
, skb
, tx
->mtu
- IPOIB_ENCAP_LEN
);
724 if (skb_shinfo(skb
)->nr_frags
> usable_sge
) {
725 if (skb_linearize(skb
) < 0) {
726 ipoib_warn(priv
, "skb could not be linearized\n");
727 ++dev
->stats
.tx_dropped
;
728 ++dev
->stats
.tx_errors
;
729 dev_kfree_skb_any(skb
);
732 /* Does skb_linearize return ok without reducing nr_frags? */
733 if (skb_shinfo(skb
)->nr_frags
> usable_sge
) {
734 ipoib_warn(priv
, "too many frags after skb linearize\n");
735 ++dev
->stats
.tx_dropped
;
736 ++dev
->stats
.tx_errors
;
737 dev_kfree_skb_any(skb
);
741 ipoib_dbg_data(priv
, "sending packet: head 0x%x length %d connection 0x%x\n",
742 tx
->tx_head
, skb
->len
, tx
->qp
->qp_num
);
745 * We put the skb into the tx_ring _before_ we call post_send()
746 * because it's entirely possible that the completion handler will
747 * run before we execute anything after the post_send(). That
748 * means we have to make sure everything is properly recorded and
749 * our state is consistent before we call post_send().
751 tx_req
= &tx
->tx_ring
[tx
->tx_head
& (ipoib_sendq_size
- 1)];
754 if (unlikely(ipoib_dma_map_tx(priv
->ca
, tx_req
))) {
755 ++dev
->stats
.tx_errors
;
756 dev_kfree_skb_any(skb
);
760 if ((priv
->tx_head
- priv
->tx_tail
) == ipoib_sendq_size
- 1) {
761 ipoib_dbg(priv
, "TX ring 0x%x full, stopping kernel net queue\n",
763 netif_stop_queue(dev
);
769 if (netif_queue_stopped(dev
))
770 if (ib_req_notify_cq(priv
->send_cq
, IB_CQ_NEXT_COMP
|
771 IB_CQ_REPORT_MISSED_EVENTS
)) {
772 ipoib_warn(priv
, "IPoIB/CM:request notify on send CQ failed\n");
773 napi_schedule(&priv
->send_napi
);
776 rc
= post_send(priv
, tx
, tx
->tx_head
& (ipoib_sendq_size
- 1), tx_req
);
778 ipoib_warn(priv
, "IPoIB/CM:post_send failed, error %d\n", rc
);
779 ++dev
->stats
.tx_errors
;
780 ipoib_dma_unmap_tx(priv
, tx_req
);
781 dev_kfree_skb_any(skb
);
783 if (netif_queue_stopped(dev
))
784 netif_wake_queue(dev
);
786 netif_trans_update(dev
);
792 void ipoib_cm_handle_tx_wc(struct net_device
*dev
, struct ib_wc
*wc
)
794 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
795 struct ipoib_cm_tx
*tx
= wc
->qp
->qp_context
;
796 unsigned int wr_id
= wc
->wr_id
& ~IPOIB_OP_CM
;
797 struct ipoib_tx_buf
*tx_req
;
800 ipoib_dbg_data(priv
, "cm send completion: id %d, status: %d\n",
803 if (unlikely(wr_id
>= ipoib_sendq_size
)) {
804 ipoib_warn(priv
, "cm send completion event with wrid %d (> %d)\n",
805 wr_id
, ipoib_sendq_size
);
809 tx_req
= &tx
->tx_ring
[wr_id
];
811 ipoib_dma_unmap_tx(priv
, tx_req
);
813 /* FIXME: is this right? Shouldn't we only increment on success? */
814 ++dev
->stats
.tx_packets
;
815 dev
->stats
.tx_bytes
+= tx_req
->skb
->len
;
817 dev_kfree_skb_any(tx_req
->skb
);
824 if (unlikely(netif_queue_stopped(dev
) &&
825 (priv
->tx_head
- priv
->tx_tail
) <= ipoib_sendq_size
>> 1 &&
826 test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
)))
827 netif_wake_queue(dev
);
829 if (wc
->status
!= IB_WC_SUCCESS
&&
830 wc
->status
!= IB_WC_WR_FLUSH_ERR
) {
831 struct ipoib_neigh
*neigh
;
833 /* IB_WC[_RNR]_RETRY_EXC_ERR error is part of the life cycle,
834 * so don't make waves.
836 if (wc
->status
== IB_WC_RNR_RETRY_EXC_ERR
||
837 wc
->status
== IB_WC_RETRY_EXC_ERR
)
839 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
840 __func__
, wc
->status
, wr_id
, wc
->vendor_err
);
843 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
844 __func__
, wc
->status
, wr_id
, wc
->vendor_err
);
846 spin_lock_irqsave(&priv
->lock
, flags
);
851 ipoib_neigh_free(neigh
);
856 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED
, &tx
->flags
)) {
857 list_move(&tx
->list
, &priv
->cm
.reap_list
);
858 queue_work(priv
->wq
, &priv
->cm
.reap_task
);
861 clear_bit(IPOIB_FLAG_OPER_UP
, &tx
->flags
);
863 spin_unlock_irqrestore(&priv
->lock
, flags
);
866 netif_tx_unlock(dev
);
869 int ipoib_cm_dev_open(struct net_device
*dev
)
871 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
874 if (!IPOIB_CM_SUPPORTED(dev
->dev_addr
))
877 priv
->cm
.id
= ib_create_cm_id(priv
->ca
, ipoib_cm_rx_handler
, dev
);
878 if (IS_ERR(priv
->cm
.id
)) {
879 printk(KERN_WARNING
"%s: failed to create CM ID\n", priv
->ca
->name
);
880 ret
= PTR_ERR(priv
->cm
.id
);
884 ret
= ib_cm_listen(priv
->cm
.id
, cpu_to_be64(IPOIB_CM_IETF_ID
| priv
->qp
->qp_num
),
887 printk(KERN_WARNING
"%s: failed to listen on ID 0x%llx\n", priv
->ca
->name
,
888 IPOIB_CM_IETF_ID
| priv
->qp
->qp_num
);
895 ib_destroy_cm_id(priv
->cm
.id
);
901 static void ipoib_cm_free_rx_reap_list(struct net_device
*dev
)
903 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
904 struct ipoib_cm_rx
*rx
, *n
;
907 spin_lock_irq(&priv
->lock
);
908 list_splice_init(&priv
->cm
.rx_reap_list
, &list
);
909 spin_unlock_irq(&priv
->lock
);
911 list_for_each_entry_safe(rx
, n
, &list
, list
) {
912 ib_destroy_cm_id(rx
->id
);
913 ib_destroy_qp(rx
->qp
);
914 if (!ipoib_cm_has_srq(dev
)) {
915 ipoib_cm_free_rx_ring(priv
->dev
, rx
->rx_ring
);
916 spin_lock_irq(&priv
->lock
);
917 --priv
->cm
.nonsrq_conn_qp
;
918 spin_unlock_irq(&priv
->lock
);
924 void ipoib_cm_dev_stop(struct net_device
*dev
)
926 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
927 struct ipoib_cm_rx
*p
;
931 if (!IPOIB_CM_SUPPORTED(dev
->dev_addr
) || !priv
->cm
.id
)
934 ib_destroy_cm_id(priv
->cm
.id
);
937 spin_lock_irq(&priv
->lock
);
938 while (!list_empty(&priv
->cm
.passive_ids
)) {
939 p
= list_entry(priv
->cm
.passive_ids
.next
, typeof(*p
), list
);
940 list_move(&p
->list
, &priv
->cm
.rx_error_list
);
941 p
->state
= IPOIB_CM_RX_ERROR
;
942 spin_unlock_irq(&priv
->lock
);
943 ret
= ib_modify_qp(p
->qp
, &ipoib_cm_err_attr
, IB_QP_STATE
);
945 ipoib_warn(priv
, "unable to move qp to error state: %d\n", ret
);
946 spin_lock_irq(&priv
->lock
);
949 /* Wait for all RX to be drained */
952 while (!list_empty(&priv
->cm
.rx_error_list
) ||
953 !list_empty(&priv
->cm
.rx_flush_list
) ||
954 !list_empty(&priv
->cm
.rx_drain_list
)) {
955 if (time_after(jiffies
, begin
+ 5 * HZ
)) {
956 ipoib_warn(priv
, "RX drain timing out\n");
959 * assume the HW is wedged and just free up everything.
961 list_splice_init(&priv
->cm
.rx_flush_list
,
962 &priv
->cm
.rx_reap_list
);
963 list_splice_init(&priv
->cm
.rx_error_list
,
964 &priv
->cm
.rx_reap_list
);
965 list_splice_init(&priv
->cm
.rx_drain_list
,
966 &priv
->cm
.rx_reap_list
);
969 spin_unlock_irq(&priv
->lock
);
970 usleep_range(1000, 2000);
972 spin_lock_irq(&priv
->lock
);
975 spin_unlock_irq(&priv
->lock
);
977 ipoib_cm_free_rx_reap_list(dev
);
979 cancel_delayed_work(&priv
->cm
.stale_task
);
982 static int ipoib_cm_rep_handler(struct ib_cm_id
*cm_id
, struct ib_cm_event
*event
)
984 struct ipoib_cm_tx
*p
= cm_id
->context
;
985 struct ipoib_dev_priv
*priv
= ipoib_priv(p
->dev
);
986 struct ipoib_cm_data
*data
= event
->private_data
;
987 struct sk_buff_head skqueue
;
988 struct ib_qp_attr qp_attr
;
989 int qp_attr_mask
, ret
;
992 p
->mtu
= be32_to_cpu(data
->mtu
);
994 if (p
->mtu
<= IPOIB_ENCAP_LEN
) {
995 ipoib_warn(priv
, "Rejecting connection: mtu %d <= %d\n",
996 p
->mtu
, IPOIB_ENCAP_LEN
);
1000 qp_attr
.qp_state
= IB_QPS_RTR
;
1001 ret
= ib_cm_init_qp_attr(cm_id
, &qp_attr
, &qp_attr_mask
);
1003 ipoib_warn(priv
, "failed to init QP attr for RTR: %d\n", ret
);
1007 qp_attr
.rq_psn
= 0 /* FIXME */;
1008 ret
= ib_modify_qp(p
->qp
, &qp_attr
, qp_attr_mask
);
1010 ipoib_warn(priv
, "failed to modify QP to RTR: %d\n", ret
);
1014 qp_attr
.qp_state
= IB_QPS_RTS
;
1015 ret
= ib_cm_init_qp_attr(cm_id
, &qp_attr
, &qp_attr_mask
);
1017 ipoib_warn(priv
, "failed to init QP attr for RTS: %d\n", ret
);
1020 ret
= ib_modify_qp(p
->qp
, &qp_attr
, qp_attr_mask
);
1022 ipoib_warn(priv
, "failed to modify QP to RTS: %d\n", ret
);
1026 skb_queue_head_init(&skqueue
);
1028 spin_lock_irq(&priv
->lock
);
1029 set_bit(IPOIB_FLAG_OPER_UP
, &p
->flags
);
1031 while ((skb
= __skb_dequeue(&p
->neigh
->queue
)))
1032 __skb_queue_tail(&skqueue
, skb
);
1033 spin_unlock_irq(&priv
->lock
);
1035 while ((skb
= __skb_dequeue(&skqueue
))) {
1037 ret
= dev_queue_xmit(skb
);
1039 ipoib_warn(priv
, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1043 ret
= ib_send_cm_rtu(cm_id
, NULL
, 0);
1045 ipoib_warn(priv
, "failed to send RTU: %d\n", ret
);
1051 static struct ib_qp
*ipoib_cm_create_tx_qp(struct net_device
*dev
, struct ipoib_cm_tx
*tx
)
1053 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1054 struct ib_qp_init_attr attr
= {
1055 .send_cq
= priv
->send_cq
,
1056 .recv_cq
= priv
->recv_cq
,
1057 .srq
= priv
->cm
.srq
,
1058 .cap
.max_send_wr
= ipoib_sendq_size
,
1059 .cap
.max_send_sge
= 1,
1060 .sq_sig_type
= IB_SIGNAL_ALL_WR
,
1061 .qp_type
= IB_QPT_RC
,
1065 struct ib_qp
*tx_qp
;
1067 if (dev
->features
& NETIF_F_SG
)
1068 attr
.cap
.max_send_sge
=
1069 min_t(u32
, priv
->ca
->attrs
.max_sge
, MAX_SKB_FRAGS
+ 1);
1071 tx_qp
= ib_create_qp(priv
->pd
, &attr
);
1072 tx
->max_send_sge
= attr
.cap
.max_send_sge
;
1076 static int ipoib_cm_send_req(struct net_device
*dev
,
1077 struct ib_cm_id
*id
, struct ib_qp
*qp
,
1079 struct sa_path_rec
*pathrec
)
1081 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1082 struct ipoib_cm_data data
= {};
1083 struct ib_cm_req_param req
= {};
1085 data
.qpn
= cpu_to_be32(priv
->qp
->qp_num
);
1086 data
.mtu
= cpu_to_be32(IPOIB_CM_BUF_SIZE
);
1088 req
.primary_path
= pathrec
;
1089 req
.alternate_path
= NULL
;
1090 req
.service_id
= cpu_to_be64(IPOIB_CM_IETF_ID
| qpn
);
1091 req
.qp_num
= qp
->qp_num
;
1092 req
.qp_type
= qp
->qp_type
;
1093 req
.private_data
= &data
;
1094 req
.private_data_len
= sizeof data
;
1095 req
.flow_control
= 0;
1097 req
.starting_psn
= 0; /* FIXME */
1100 * Pick some arbitrary defaults here; we could make these
1101 * module parameters if anyone cared about setting them.
1103 req
.responder_resources
= 4;
1104 req
.remote_cm_response_timeout
= 20;
1105 req
.local_cm_response_timeout
= 20;
1106 req
.retry_count
= 0; /* RFC draft warns against retries */
1107 req
.rnr_retry_count
= 0; /* RFC draft warns against retries */
1108 req
.max_cm_retries
= 15;
1109 req
.srq
= ipoib_cm_has_srq(dev
);
1110 return ib_send_cm_req(id
, &req
);
1113 static int ipoib_cm_modify_tx_init(struct net_device
*dev
,
1114 struct ib_cm_id
*cm_id
, struct ib_qp
*qp
)
1116 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1117 struct ib_qp_attr qp_attr
;
1118 int qp_attr_mask
, ret
;
1119 ret
= ib_find_pkey(priv
->ca
, priv
->port
, priv
->pkey
, &qp_attr
.pkey_index
);
1121 ipoib_warn(priv
, "pkey 0x%x not found: %d\n", priv
->pkey
, ret
);
1125 qp_attr
.qp_state
= IB_QPS_INIT
;
1126 qp_attr
.qp_access_flags
= IB_ACCESS_LOCAL_WRITE
;
1127 qp_attr
.port_num
= priv
->port
;
1128 qp_attr_mask
= IB_QP_STATE
| IB_QP_ACCESS_FLAGS
| IB_QP_PKEY_INDEX
| IB_QP_PORT
;
1130 ret
= ib_modify_qp(qp
, &qp_attr
, qp_attr_mask
);
1132 ipoib_warn(priv
, "failed to modify tx QP to INIT: %d\n", ret
);
1138 static int ipoib_cm_tx_init(struct ipoib_cm_tx
*p
, u32 qpn
,
1139 struct sa_path_rec
*pathrec
)
1141 struct ipoib_dev_priv
*priv
= ipoib_priv(p
->dev
);
1142 unsigned int noio_flag
;
1145 noio_flag
= memalloc_noio_save();
1146 p
->tx_ring
= vzalloc(ipoib_sendq_size
* sizeof(*p
->tx_ring
));
1148 memalloc_noio_restore(noio_flag
);
1152 memset(p
->tx_ring
, 0, ipoib_sendq_size
* sizeof *p
->tx_ring
);
1154 p
->qp
= ipoib_cm_create_tx_qp(p
->dev
, p
);
1155 memalloc_noio_restore(noio_flag
);
1156 if (IS_ERR(p
->qp
)) {
1157 ret
= PTR_ERR(p
->qp
);
1158 ipoib_warn(priv
, "failed to create tx qp: %d\n", ret
);
1162 p
->id
= ib_create_cm_id(priv
->ca
, ipoib_cm_tx_handler
, p
);
1163 if (IS_ERR(p
->id
)) {
1164 ret
= PTR_ERR(p
->id
);
1165 ipoib_warn(priv
, "failed to create tx cm id: %d\n", ret
);
1169 ret
= ipoib_cm_modify_tx_init(p
->dev
, p
->id
, p
->qp
);
1171 ipoib_warn(priv
, "failed to modify tx qp to rtr: %d\n", ret
);
1172 goto err_modify_send
;
1175 ret
= ipoib_cm_send_req(p
->dev
, p
->id
, p
->qp
, qpn
, pathrec
);
1177 ipoib_warn(priv
, "failed to send cm req: %d\n", ret
);
1178 goto err_modify_send
;
1181 ipoib_dbg(priv
, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1182 p
->qp
->qp_num
, pathrec
->dgid
.raw
, qpn
);
1187 ib_destroy_cm_id(p
->id
);
1190 ib_destroy_qp(p
->qp
);
1198 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx
*p
)
1200 struct ipoib_dev_priv
*priv
= ipoib_priv(p
->dev
);
1201 struct ipoib_tx_buf
*tx_req
;
1202 unsigned long begin
;
1204 ipoib_dbg(priv
, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1205 p
->qp
? p
->qp
->qp_num
: 0, p
->tx_head
, p
->tx_tail
);
1208 ib_destroy_cm_id(p
->id
);
1211 /* Wait for all sends to complete */
1213 while ((int) p
->tx_tail
- (int) p
->tx_head
< 0) {
1214 if (time_after(jiffies
, begin
+ 5 * HZ
)) {
1215 ipoib_warn(priv
, "timing out; %d sends not completed\n",
1216 p
->tx_head
- p
->tx_tail
);
1220 usleep_range(1000, 2000);
1226 while ((int) p
->tx_tail
- (int) p
->tx_head
< 0) {
1227 tx_req
= &p
->tx_ring
[p
->tx_tail
& (ipoib_sendq_size
- 1)];
1228 ipoib_dma_unmap_tx(priv
, tx_req
);
1229 dev_kfree_skb_any(tx_req
->skb
);
1230 netif_tx_lock_bh(p
->dev
);
1233 if (unlikely(priv
->tx_head
- priv
->tx_tail
== ipoib_sendq_size
>> 1) &&
1234 netif_queue_stopped(p
->dev
) &&
1235 test_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
))
1236 netif_wake_queue(p
->dev
);
1237 netif_tx_unlock_bh(p
->dev
);
1241 ib_destroy_qp(p
->qp
);
1247 static int ipoib_cm_tx_handler(struct ib_cm_id
*cm_id
,
1248 struct ib_cm_event
*event
)
1250 struct ipoib_cm_tx
*tx
= cm_id
->context
;
1251 struct ipoib_dev_priv
*priv
= ipoib_priv(tx
->dev
);
1252 struct net_device
*dev
= priv
->dev
;
1253 struct ipoib_neigh
*neigh
;
1254 unsigned long flags
;
1257 switch (event
->event
) {
1258 case IB_CM_DREQ_RECEIVED
:
1259 ipoib_dbg(priv
, "DREQ received.\n");
1260 ib_send_cm_drep(cm_id
, NULL
, 0);
1262 case IB_CM_REP_RECEIVED
:
1263 ipoib_dbg(priv
, "REP received.\n");
1264 ret
= ipoib_cm_rep_handler(cm_id
, event
);
1266 ib_send_cm_rej(cm_id
, IB_CM_REJ_CONSUMER_DEFINED
,
1269 case IB_CM_REQ_ERROR
:
1270 case IB_CM_REJ_RECEIVED
:
1271 case IB_CM_TIMEWAIT_EXIT
:
1272 ipoib_dbg(priv
, "CM error %d.\n", event
->event
);
1273 netif_tx_lock_bh(dev
);
1274 spin_lock_irqsave(&priv
->lock
, flags
);
1279 ipoib_neigh_free(neigh
);
1284 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED
, &tx
->flags
)) {
1285 list_move(&tx
->list
, &priv
->cm
.reap_list
);
1286 queue_work(priv
->wq
, &priv
->cm
.reap_task
);
1289 spin_unlock_irqrestore(&priv
->lock
, flags
);
1290 netif_tx_unlock_bh(dev
);
1299 struct ipoib_cm_tx
*ipoib_cm_create_tx(struct net_device
*dev
, struct ipoib_path
*path
,
1300 struct ipoib_neigh
*neigh
)
1302 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1303 struct ipoib_cm_tx
*tx
;
1305 tx
= kzalloc(sizeof *tx
, GFP_ATOMIC
);
1313 list_add(&tx
->list
, &priv
->cm
.start_list
);
1314 set_bit(IPOIB_FLAG_INITIALIZED
, &tx
->flags
);
1315 queue_work(priv
->wq
, &priv
->cm
.start_task
);
1319 void ipoib_cm_destroy_tx(struct ipoib_cm_tx
*tx
)
1321 struct ipoib_dev_priv
*priv
= ipoib_priv(tx
->dev
);
1322 unsigned long flags
;
1323 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED
, &tx
->flags
)) {
1324 spin_lock_irqsave(&priv
->lock
, flags
);
1325 list_move(&tx
->list
, &priv
->cm
.reap_list
);
1326 queue_work(priv
->wq
, &priv
->cm
.reap_task
);
1327 ipoib_dbg(priv
, "Reap connection for gid %pI6\n",
1328 tx
->neigh
->daddr
+ 4);
1330 spin_unlock_irqrestore(&priv
->lock
, flags
);
1334 #define QPN_AND_OPTIONS_OFFSET 4
1336 static void ipoib_cm_tx_start(struct work_struct
*work
)
1338 struct ipoib_dev_priv
*priv
= container_of(work
, struct ipoib_dev_priv
,
1340 struct net_device
*dev
= priv
->dev
;
1341 struct ipoib_neigh
*neigh
;
1342 struct ipoib_cm_tx
*p
;
1343 unsigned long flags
;
1344 struct ipoib_path
*path
;
1347 struct sa_path_rec pathrec
;
1350 netif_tx_lock_bh(dev
);
1351 spin_lock_irqsave(&priv
->lock
, flags
);
1353 while (!list_empty(&priv
->cm
.start_list
)) {
1354 p
= list_entry(priv
->cm
.start_list
.next
, typeof(*p
), list
);
1355 list_del_init(&p
->list
);
1358 qpn
= IPOIB_QPN(neigh
->daddr
);
1360 * As long as the search is with these 2 locks,
1361 * path existence indicates its validity.
1363 path
= __path_find(dev
, neigh
->daddr
+ QPN_AND_OPTIONS_OFFSET
);
1365 pr_info("%s ignore not valid path %pI6\n",
1367 neigh
->daddr
+ QPN_AND_OPTIONS_OFFSET
);
1370 memcpy(&pathrec
, &p
->path
->pathrec
, sizeof pathrec
);
1372 spin_unlock_irqrestore(&priv
->lock
, flags
);
1373 netif_tx_unlock_bh(dev
);
1375 ret
= ipoib_cm_tx_init(p
, qpn
, &pathrec
);
1377 netif_tx_lock_bh(dev
);
1378 spin_lock_irqsave(&priv
->lock
, flags
);
1385 ipoib_neigh_free(neigh
);
1392 spin_unlock_irqrestore(&priv
->lock
, flags
);
1393 netif_tx_unlock_bh(dev
);
1396 static void ipoib_cm_tx_reap(struct work_struct
*work
)
1398 struct ipoib_dev_priv
*priv
= container_of(work
, struct ipoib_dev_priv
,
1400 struct net_device
*dev
= priv
->dev
;
1401 struct ipoib_cm_tx
*p
;
1402 unsigned long flags
;
1404 netif_tx_lock_bh(dev
);
1405 spin_lock_irqsave(&priv
->lock
, flags
);
1407 while (!list_empty(&priv
->cm
.reap_list
)) {
1408 p
= list_entry(priv
->cm
.reap_list
.next
, typeof(*p
), list
);
1409 list_del_init(&p
->list
);
1410 spin_unlock_irqrestore(&priv
->lock
, flags
);
1411 netif_tx_unlock_bh(dev
);
1412 ipoib_cm_tx_destroy(p
);
1413 netif_tx_lock_bh(dev
);
1414 spin_lock_irqsave(&priv
->lock
, flags
);
1417 spin_unlock_irqrestore(&priv
->lock
, flags
);
1418 netif_tx_unlock_bh(dev
);
1421 static void ipoib_cm_skb_reap(struct work_struct
*work
)
1423 struct ipoib_dev_priv
*priv
= container_of(work
, struct ipoib_dev_priv
,
1425 struct net_device
*dev
= priv
->dev
;
1426 struct sk_buff
*skb
;
1427 unsigned long flags
;
1428 unsigned mtu
= priv
->mcast_mtu
;
1430 netif_tx_lock_bh(dev
);
1431 spin_lock_irqsave(&priv
->lock
, flags
);
1433 while ((skb
= skb_dequeue(&priv
->cm
.skb_queue
))) {
1434 spin_unlock_irqrestore(&priv
->lock
, flags
);
1435 netif_tx_unlock_bh(dev
);
1437 if (skb
->protocol
== htons(ETH_P_IP
))
1438 icmp_send(skb
, ICMP_DEST_UNREACH
, ICMP_FRAG_NEEDED
, htonl(mtu
));
1439 #if IS_ENABLED(CONFIG_IPV6)
1440 else if (skb
->protocol
== htons(ETH_P_IPV6
))
1441 icmpv6_send(skb
, ICMPV6_PKT_TOOBIG
, 0, mtu
);
1443 dev_kfree_skb_any(skb
);
1445 netif_tx_lock_bh(dev
);
1446 spin_lock_irqsave(&priv
->lock
, flags
);
1449 spin_unlock_irqrestore(&priv
->lock
, flags
);
1450 netif_tx_unlock_bh(dev
);
1453 void ipoib_cm_skb_too_long(struct net_device
*dev
, struct sk_buff
*skb
,
1456 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1457 int e
= skb_queue_empty(&priv
->cm
.skb_queue
);
1460 skb_dst(skb
)->ops
->update_pmtu(skb_dst(skb
), NULL
, skb
, mtu
);
1462 skb_queue_tail(&priv
->cm
.skb_queue
, skb
);
1464 queue_work(priv
->wq
, &priv
->cm
.skb_task
);
1467 static void ipoib_cm_rx_reap(struct work_struct
*work
)
1469 ipoib_cm_free_rx_reap_list(container_of(work
, struct ipoib_dev_priv
,
1470 cm
.rx_reap_task
)->dev
);
1473 static void ipoib_cm_stale_task(struct work_struct
*work
)
1475 struct ipoib_dev_priv
*priv
= container_of(work
, struct ipoib_dev_priv
,
1476 cm
.stale_task
.work
);
1477 struct ipoib_cm_rx
*p
;
1480 spin_lock_irq(&priv
->lock
);
1481 while (!list_empty(&priv
->cm
.passive_ids
)) {
1482 /* List is sorted by LRU, start from tail,
1483 * stop when we see a recently used entry */
1484 p
= list_entry(priv
->cm
.passive_ids
.prev
, typeof(*p
), list
);
1485 if (time_before_eq(jiffies
, p
->jiffies
+ IPOIB_CM_RX_TIMEOUT
))
1487 list_move(&p
->list
, &priv
->cm
.rx_error_list
);
1488 p
->state
= IPOIB_CM_RX_ERROR
;
1489 spin_unlock_irq(&priv
->lock
);
1490 ret
= ib_modify_qp(p
->qp
, &ipoib_cm_err_attr
, IB_QP_STATE
);
1492 ipoib_warn(priv
, "unable to move qp to error state: %d\n", ret
);
1493 spin_lock_irq(&priv
->lock
);
1496 if (!list_empty(&priv
->cm
.passive_ids
))
1497 queue_delayed_work(priv
->wq
,
1498 &priv
->cm
.stale_task
, IPOIB_CM_RX_DELAY
);
1499 spin_unlock_irq(&priv
->lock
);
1502 static ssize_t
show_mode(struct device
*d
, struct device_attribute
*attr
,
1505 struct net_device
*dev
= to_net_dev(d
);
1506 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1508 if (test_bit(IPOIB_FLAG_ADMIN_CM
, &priv
->flags
))
1509 return sprintf(buf
, "connected\n");
1511 return sprintf(buf
, "datagram\n");
1514 static ssize_t
set_mode(struct device
*d
, struct device_attribute
*attr
,
1515 const char *buf
, size_t count
)
1517 struct net_device
*dev
= to_net_dev(d
);
1519 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1521 if (test_bit(IPOIB_FLAG_GOING_DOWN
, &priv
->flags
))
1524 if (!mutex_trylock(&priv
->sysfs_mutex
))
1525 return restart_syscall();
1527 if (!rtnl_trylock()) {
1528 mutex_unlock(&priv
->sysfs_mutex
);
1529 return restart_syscall();
1532 ret
= ipoib_set_mode(dev
, buf
);
1534 /* The assumption is that the function ipoib_set_mode returned
1535 * with the rtnl held by it, if not the value -EBUSY returned,
1536 * then no need to rtnl_unlock
1540 mutex_unlock(&priv
->sysfs_mutex
);
1542 return (!ret
|| ret
== -EBUSY
) ? count
: ret
;
1545 static DEVICE_ATTR(mode
, S_IWUSR
| S_IRUGO
, show_mode
, set_mode
);
1547 int ipoib_cm_add_mode_attr(struct net_device
*dev
)
1549 return device_create_file(&dev
->dev
, &dev_attr_mode
);
1552 static void ipoib_cm_create_srq(struct net_device
*dev
, int max_sge
)
1554 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1555 struct ib_srq_init_attr srq_init_attr
= {
1556 .srq_type
= IB_SRQT_BASIC
,
1558 .max_wr
= ipoib_recvq_size
,
1563 priv
->cm
.srq
= ib_create_srq(priv
->pd
, &srq_init_attr
);
1564 if (IS_ERR(priv
->cm
.srq
)) {
1565 if (PTR_ERR(priv
->cm
.srq
) != -ENOSYS
)
1566 printk(KERN_WARNING
"%s: failed to allocate SRQ, error %ld\n",
1567 priv
->ca
->name
, PTR_ERR(priv
->cm
.srq
));
1568 priv
->cm
.srq
= NULL
;
1572 priv
->cm
.srq_ring
= vzalloc(ipoib_recvq_size
* sizeof *priv
->cm
.srq_ring
);
1573 if (!priv
->cm
.srq_ring
) {
1574 ib_destroy_srq(priv
->cm
.srq
);
1575 priv
->cm
.srq
= NULL
;
1581 int ipoib_cm_dev_init(struct net_device
*dev
)
1583 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1586 INIT_LIST_HEAD(&priv
->cm
.passive_ids
);
1587 INIT_LIST_HEAD(&priv
->cm
.reap_list
);
1588 INIT_LIST_HEAD(&priv
->cm
.start_list
);
1589 INIT_LIST_HEAD(&priv
->cm
.rx_error_list
);
1590 INIT_LIST_HEAD(&priv
->cm
.rx_flush_list
);
1591 INIT_LIST_HEAD(&priv
->cm
.rx_drain_list
);
1592 INIT_LIST_HEAD(&priv
->cm
.rx_reap_list
);
1593 INIT_WORK(&priv
->cm
.start_task
, ipoib_cm_tx_start
);
1594 INIT_WORK(&priv
->cm
.reap_task
, ipoib_cm_tx_reap
);
1595 INIT_WORK(&priv
->cm
.skb_task
, ipoib_cm_skb_reap
);
1596 INIT_WORK(&priv
->cm
.rx_reap_task
, ipoib_cm_rx_reap
);
1597 INIT_DELAYED_WORK(&priv
->cm
.stale_task
, ipoib_cm_stale_task
);
1599 skb_queue_head_init(&priv
->cm
.skb_queue
);
1601 ipoib_dbg(priv
, "max_srq_sge=%d\n", priv
->ca
->attrs
.max_srq_sge
);
1603 max_srq_sge
= min_t(int, IPOIB_CM_RX_SG
, priv
->ca
->attrs
.max_srq_sge
);
1604 ipoib_cm_create_srq(dev
, max_srq_sge
);
1605 if (ipoib_cm_has_srq(dev
)) {
1606 priv
->cm
.max_cm_mtu
= max_srq_sge
* PAGE_SIZE
- 0x10;
1607 priv
->cm
.num_frags
= max_srq_sge
;
1608 ipoib_dbg(priv
, "max_cm_mtu = 0x%x, num_frags=%d\n",
1609 priv
->cm
.max_cm_mtu
, priv
->cm
.num_frags
);
1611 priv
->cm
.max_cm_mtu
= IPOIB_CM_MTU
;
1612 priv
->cm
.num_frags
= IPOIB_CM_RX_SG
;
1615 ipoib_cm_init_rx_wr(dev
, &priv
->cm
.rx_wr
, priv
->cm
.rx_sge
);
1617 if (ipoib_cm_has_srq(dev
)) {
1618 for (i
= 0; i
< ipoib_recvq_size
; ++i
) {
1619 if (!ipoib_cm_alloc_rx_skb(dev
, priv
->cm
.srq_ring
, i
,
1620 priv
->cm
.num_frags
- 1,
1621 priv
->cm
.srq_ring
[i
].mapping
,
1623 ipoib_warn(priv
, "failed to allocate "
1624 "receive buffer %d\n", i
);
1625 ipoib_cm_dev_cleanup(dev
);
1629 if (ipoib_cm_post_receive_srq(dev
, i
)) {
1630 ipoib_warn(priv
, "ipoib_cm_post_receive_srq "
1631 "failed for buf %d\n", i
);
1632 ipoib_cm_dev_cleanup(dev
);
1638 priv
->dev
->dev_addr
[0] = IPOIB_FLAGS_RC
;
1642 void ipoib_cm_dev_cleanup(struct net_device
*dev
)
1644 struct ipoib_dev_priv
*priv
= ipoib_priv(dev
);
1650 ipoib_dbg(priv
, "Cleanup ipoib connected mode.\n");
1652 ret
= ib_destroy_srq(priv
->cm
.srq
);
1654 ipoib_warn(priv
, "ib_destroy_srq failed: %d\n", ret
);
1656 priv
->cm
.srq
= NULL
;
1657 if (!priv
->cm
.srq_ring
)
1660 ipoib_cm_free_rx_ring(dev
, priv
->cm
.srq_ring
);
1661 priv
->cm
.srq_ring
= NULL
;