This source file includes following definitions.
- ipoib_cm_dma_unmap_rx
- ipoib_cm_post_receive_srq
- ipoib_cm_post_receive_nonsrq
- ipoib_cm_alloc_rx_skb
- ipoib_cm_free_rx_ring
- ipoib_cm_start_rx_drain
- ipoib_cm_rx_event_handler
- ipoib_cm_create_rx_qp
- ipoib_cm_modify_rx_qp
- ipoib_cm_init_rx_wr
- ipoib_cm_nonsrq_init_rx
- ipoib_cm_send_rep
- ipoib_cm_req_handler
- ipoib_cm_rx_handler
- skb_put_frags
- ipoib_cm_handle_rx_wc
- post_send
- ipoib_cm_send
- ipoib_cm_handle_tx_wc
- ipoib_cm_dev_open
- ipoib_cm_free_rx_reap_list
- ipoib_cm_dev_stop
- ipoib_cm_rep_handler
- ipoib_cm_create_tx_qp
- ipoib_cm_send_req
- ipoib_cm_modify_tx_init
- ipoib_cm_tx_init
- ipoib_cm_tx_destroy
- ipoib_cm_tx_handler
- ipoib_cm_create_tx
- ipoib_cm_destroy_tx
- ipoib_cm_tx_start
- ipoib_cm_tx_reap
- ipoib_cm_skb_reap
- ipoib_cm_skb_too_long
- ipoib_cm_rx_reap
- ipoib_cm_stale_task
- show_mode
- set_mode
- ipoib_cm_add_mode_attr
- ipoib_cm_create_srq
- ipoib_cm_dev_init
- ipoib_cm_dev_cleanup
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33 #include <rdma/ib_cm.h>
34 #include <net/dst.h>
35 #include <net/icmp.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>
43
44 #include "ipoib.h"
45
46 int ipoib_max_conn_qp = 128;
47
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)");
52
53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
54 static int data_debug_level;
55
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");
59 #endif
60
61 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
62
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)
67
68 #define IPOIB_CM_RX_RESERVE (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
69
70 static struct ib_qp_attr ipoib_cm_err_attr = {
71 .qp_state = IB_QPS_ERR
72 };
73
74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
75
76 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
77 .opcode = IB_WR_SEND,
78 };
79
80 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
81 const struct ib_cm_event *event);
82
83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
84 u64 mapping[IPOIB_CM_RX_SG])
85 {
86 int i;
87
88 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
89
90 for (i = 0; i < frags; ++i)
91 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
92 }
93
94 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
95 {
96 struct ipoib_dev_priv *priv = ipoib_priv(dev);
97 int i, ret;
98
99 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
100
101 for (i = 0; i < priv->cm.num_frags; ++i)
102 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
103
104 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, NULL);
105 if (unlikely(ret)) {
106 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
107 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
108 priv->cm.srq_ring[id].mapping);
109 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
110 priv->cm.srq_ring[id].skb = NULL;
111 }
112
113 return ret;
114 }
115
116 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
117 struct ipoib_cm_rx *rx,
118 struct ib_recv_wr *wr,
119 struct ib_sge *sge, int id)
120 {
121 struct ipoib_dev_priv *priv = ipoib_priv(dev);
122 int i, ret;
123
124 wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
125
126 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
127 sge[i].addr = rx->rx_ring[id].mapping[i];
128
129 ret = ib_post_recv(rx->qp, wr, NULL);
130 if (unlikely(ret)) {
131 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
132 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
133 rx->rx_ring[id].mapping);
134 dev_kfree_skb_any(rx->rx_ring[id].skb);
135 rx->rx_ring[id].skb = NULL;
136 }
137
138 return ret;
139 }
140
141 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
142 struct ipoib_cm_rx_buf *rx_ring,
143 int id, int frags,
144 u64 mapping[IPOIB_CM_RX_SG],
145 gfp_t gfp)
146 {
147 struct ipoib_dev_priv *priv = ipoib_priv(dev);
148 struct sk_buff *skb;
149 int i;
150
151 skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
152 if (unlikely(!skb))
153 return NULL;
154
155
156
157
158
159 skb_reserve(skb, IPOIB_CM_RX_RESERVE);
160
161 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
162 DMA_FROM_DEVICE);
163 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
164 dev_kfree_skb_any(skb);
165 return NULL;
166 }
167
168 for (i = 0; i < frags; i++) {
169 struct page *page = alloc_page(gfp);
170
171 if (!page)
172 goto partial_error;
173 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
174
175 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
176 0, PAGE_SIZE, DMA_FROM_DEVICE);
177 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
178 goto partial_error;
179 }
180
181 rx_ring[id].skb = skb;
182 return skb;
183
184 partial_error:
185
186 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
187
188 for (; i > 0; --i)
189 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
190
191 dev_kfree_skb_any(skb);
192 return NULL;
193 }
194
195 static void ipoib_cm_free_rx_ring(struct net_device *dev,
196 struct ipoib_cm_rx_buf *rx_ring)
197 {
198 struct ipoib_dev_priv *priv = ipoib_priv(dev);
199 int i;
200
201 for (i = 0; i < ipoib_recvq_size; ++i)
202 if (rx_ring[i].skb) {
203 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
204 rx_ring[i].mapping);
205 dev_kfree_skb_any(rx_ring[i].skb);
206 }
207
208 vfree(rx_ring);
209 }
210
211 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
212 {
213 struct ipoib_cm_rx *p;
214
215
216
217 if (list_empty(&priv->cm.rx_flush_list) ||
218 !list_empty(&priv->cm.rx_drain_list))
219 return;
220
221
222
223
224
225 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
226 ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
227 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, NULL))
228 ipoib_warn(priv, "failed to post drain wr\n");
229
230 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
231 }
232
233 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
234 {
235 struct ipoib_cm_rx *p = ctx;
236 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
237 unsigned long flags;
238
239 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
240 return;
241
242 spin_lock_irqsave(&priv->lock, flags);
243 list_move(&p->list, &priv->cm.rx_flush_list);
244 p->state = IPOIB_CM_RX_FLUSH;
245 ipoib_cm_start_rx_drain(priv);
246 spin_unlock_irqrestore(&priv->lock, flags);
247 }
248
249 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
250 struct ipoib_cm_rx *p)
251 {
252 struct ipoib_dev_priv *priv = ipoib_priv(dev);
253 struct ib_qp_init_attr attr = {
254 .event_handler = ipoib_cm_rx_event_handler,
255 .send_cq = priv->recv_cq,
256 .recv_cq = priv->recv_cq,
257 .srq = priv->cm.srq,
258 .cap.max_send_wr = 1,
259 .cap.max_send_sge = 1,
260 .sq_sig_type = IB_SIGNAL_ALL_WR,
261 .qp_type = IB_QPT_RC,
262 .qp_context = p,
263 };
264
265 if (!ipoib_cm_has_srq(dev)) {
266 attr.cap.max_recv_wr = ipoib_recvq_size;
267 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
268 }
269
270 return ib_create_qp(priv->pd, &attr);
271 }
272
273 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
274 struct ib_cm_id *cm_id, struct ib_qp *qp,
275 unsigned int psn)
276 {
277 struct ipoib_dev_priv *priv = ipoib_priv(dev);
278 struct ib_qp_attr qp_attr;
279 int qp_attr_mask, ret;
280
281 qp_attr.qp_state = IB_QPS_INIT;
282 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
283 if (ret) {
284 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
285 return ret;
286 }
287 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
288 if (ret) {
289 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
290 return ret;
291 }
292 qp_attr.qp_state = IB_QPS_RTR;
293 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
294 if (ret) {
295 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
296 return ret;
297 }
298 qp_attr.rq_psn = psn;
299 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
300 if (ret) {
301 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
302 return ret;
303 }
304
305
306
307
308
309
310
311
312
313 qp_attr.qp_state = IB_QPS_RTS;
314 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
315 if (ret) {
316 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
317 return 0;
318 }
319 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
320 if (ret) {
321 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
322 return 0;
323 }
324
325 return 0;
326 }
327
328 static void ipoib_cm_init_rx_wr(struct net_device *dev,
329 struct ib_recv_wr *wr,
330 struct ib_sge *sge)
331 {
332 struct ipoib_dev_priv *priv = ipoib_priv(dev);
333 int i;
334
335 for (i = 0; i < priv->cm.num_frags; ++i)
336 sge[i].lkey = priv->pd->local_dma_lkey;
337
338 sge[0].length = IPOIB_CM_HEAD_SIZE;
339 for (i = 1; i < priv->cm.num_frags; ++i)
340 sge[i].length = PAGE_SIZE;
341
342 wr->next = NULL;
343 wr->sg_list = sge;
344 wr->num_sge = priv->cm.num_frags;
345 }
346
347 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
348 struct ipoib_cm_rx *rx)
349 {
350 struct ipoib_dev_priv *priv = ipoib_priv(dev);
351 struct {
352 struct ib_recv_wr wr;
353 struct ib_sge sge[IPOIB_CM_RX_SG];
354 } *t;
355 int ret;
356 int i;
357
358 rx->rx_ring = vzalloc(array_size(ipoib_recvq_size,
359 sizeof(*rx->rx_ring)));
360 if (!rx->rx_ring)
361 return -ENOMEM;
362
363 t = kmalloc(sizeof(*t), GFP_KERNEL);
364 if (!t) {
365 ret = -ENOMEM;
366 goto err_free_1;
367 }
368
369 ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
370
371 spin_lock_irq(&priv->lock);
372
373 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
374 spin_unlock_irq(&priv->lock);
375 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
376 ret = -EINVAL;
377 goto err_free;
378 } else
379 ++priv->cm.nonsrq_conn_qp;
380
381 spin_unlock_irq(&priv->lock);
382
383 for (i = 0; i < ipoib_recvq_size; ++i) {
384 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
385 rx->rx_ring[i].mapping,
386 GFP_KERNEL)) {
387 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
388 ret = -ENOMEM;
389 goto err_count;
390 }
391 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
392 if (ret) {
393 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
394 "failed for buf %d\n", i);
395 ret = -EIO;
396 goto err_count;
397 }
398 }
399
400 rx->recv_count = ipoib_recvq_size;
401
402 kfree(t);
403
404 return 0;
405
406 err_count:
407 spin_lock_irq(&priv->lock);
408 --priv->cm.nonsrq_conn_qp;
409 spin_unlock_irq(&priv->lock);
410
411 err_free:
412 kfree(t);
413
414 err_free_1:
415 ipoib_cm_free_rx_ring(dev, rx->rx_ring);
416
417 return ret;
418 }
419
420 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
421 struct ib_qp *qp,
422 const struct ib_cm_req_event_param *req,
423 unsigned int psn)
424 {
425 struct ipoib_dev_priv *priv = ipoib_priv(dev);
426 struct ipoib_cm_data data = {};
427 struct ib_cm_rep_param rep = {};
428
429 data.qpn = cpu_to_be32(priv->qp->qp_num);
430 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
431
432 rep.private_data = &data;
433 rep.private_data_len = sizeof(data);
434 rep.flow_control = 0;
435 rep.rnr_retry_count = req->rnr_retry_count;
436 rep.srq = ipoib_cm_has_srq(dev);
437 rep.qp_num = qp->qp_num;
438 rep.starting_psn = psn;
439 return ib_send_cm_rep(cm_id, &rep);
440 }
441
442 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id,
443 const struct ib_cm_event *event)
444 {
445 struct net_device *dev = cm_id->context;
446 struct ipoib_dev_priv *priv = ipoib_priv(dev);
447 struct ipoib_cm_rx *p;
448 unsigned int psn;
449 int ret;
450
451 ipoib_dbg(priv, "REQ arrived\n");
452 p = kzalloc(sizeof(*p), GFP_KERNEL);
453 if (!p)
454 return -ENOMEM;
455 p->dev = dev;
456 p->id = cm_id;
457 cm_id->context = p;
458 p->state = IPOIB_CM_RX_LIVE;
459 p->jiffies = jiffies;
460 INIT_LIST_HEAD(&p->list);
461
462 p->qp = ipoib_cm_create_rx_qp(dev, p);
463 if (IS_ERR(p->qp)) {
464 ret = PTR_ERR(p->qp);
465 goto err_qp;
466 }
467
468 psn = prandom_u32() & 0xffffff;
469 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
470 if (ret)
471 goto err_modify;
472
473 if (!ipoib_cm_has_srq(dev)) {
474 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
475 if (ret)
476 goto err_modify;
477 }
478
479 spin_lock_irq(&priv->lock);
480 queue_delayed_work(priv->wq,
481 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
482
483
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);
488
489 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
490 if (ret) {
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");
494 }
495 return 0;
496
497 err_modify:
498 ib_destroy_qp(p->qp);
499 err_qp:
500 kfree(p);
501 return ret;
502 }
503
504 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
505 const struct ib_cm_event *event)
506 {
507 struct ipoib_cm_rx *p;
508 struct ipoib_dev_priv *priv;
509
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);
515
516 case IB_CM_REJ_RECEIVED:
517 p = cm_id->context;
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");
521
522 default:
523 return 0;
524 }
525 }
526
527 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
528 unsigned int length, struct sk_buff *toskb)
529 {
530 int i, num_frags;
531 unsigned int size;
532
533
534 size = min(length, hdr_space);
535 skb->tail += size;
536 skb->len += size;
537 length -= size;
538
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];
542
543 if (length == 0) {
544
545 skb_fill_page_desc(toskb, i, skb_frag_page(frag),
546 0, PAGE_SIZE);
547 --skb_shinfo(skb)->nr_frags;
548 } else {
549 size = min_t(unsigned int, length, PAGE_SIZE);
550
551 skb_frag_size_set(frag, size);
552 skb->data_len += size;
553 skb->truesize += size;
554 skb->len += size;
555 length -= size;
556 }
557 }
558 }
559
560 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
561 {
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;
567 unsigned long flags;
568 u64 mapping[IPOIB_CM_RX_SG];
569 int frags;
570 int has_srq;
571 struct sk_buff *small_skb;
572
573 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
574 wr_id, wc->status);
575
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);
583 } else
584 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
585 wr_id, ipoib_recvq_size);
586 return;
587 }
588
589 p = wc->qp->qp_context;
590
591 has_srq = ipoib_cm_has_srq(dev);
592 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
593
594 skb = rx_ring[wr_id].skb;
595
596 if (unlikely(wc->status != IB_WC_SUCCESS)) {
597 ipoib_dbg(priv,
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;
601 if (has_srq)
602 goto repost;
603 else {
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);
609 }
610 return;
611 }
612 }
613
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
619
620 if (p->state == IPOIB_CM_RX_LIVE)
621 list_move(&p->list, &priv->cm.passive_ids);
622 spin_unlock_irqrestore(&priv->lock, flags);
623 }
624 }
625
626 if (wc->byte_len < IPOIB_CM_COPYBREAK) {
627 int dlen = wc->byte_len;
628
629 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
630 if (small_skb) {
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);
638 skb = small_skb;
639 goto copied;
640 }
641 }
642
643 frags = PAGE_ALIGN(wc->byte_len -
644 min_t(u32, wc->byte_len, IPOIB_CM_HEAD_SIZE)) /
645 PAGE_SIZE;
646
647 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
648 mapping, GFP_ATOMIC);
649 if (unlikely(!newskb)) {
650
651
652
653
654 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
655 ++dev->stats.rx_dropped;
656 goto repost;
657 }
658
659 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
660 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof(*mapping));
661
662 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
663 wc->byte_len, wc->slid);
664
665 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
666
667 copied:
668 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
669 skb_add_pseudo_hdr(skb);
670
671 ++dev->stats.rx_packets;
672 dev->stats.rx_bytes += skb->len;
673
674 skb->dev = dev;
675
676 skb->pkt_type = PACKET_HOST;
677 netif_receive_skb(skb);
678
679 repost:
680 if (has_srq) {
681 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
682 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
683 "for buf %d\n", wr_id);
684 } else {
685 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
686 &priv->cm.rx_wr,
687 priv->cm.rx_sge,
688 wr_id))) {
689 --p->recv_count;
690 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
691 "for buf %d\n", wr_id);
692 }
693 }
694 }
695
696 static inline int post_send(struct ipoib_dev_priv *priv,
697 struct ipoib_cm_tx *tx,
698 unsigned int wr_id,
699 struct ipoib_tx_buf *tx_req)
700 {
701 ipoib_build_sge(priv, tx_req);
702
703 priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
704
705 return ib_post_send(tx->qp, &priv->tx_wr.wr, NULL);
706 }
707
708 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
709 {
710 struct ipoib_dev_priv *priv = ipoib_priv(dev);
711 struct ipoib_tx_buf *tx_req;
712 int rc;
713 unsigned int usable_sge = tx->max_send_sge - !!skb_headlen(skb);
714
715 if (unlikely(skb->len > tx->mtu)) {
716 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
717 skb->len, tx->mtu);
718 ++dev->stats.tx_dropped;
719 ++dev->stats.tx_errors;
720 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
721 return;
722 }
723 if (skb_shinfo(skb)->nr_frags > usable_sge) {
724 if (skb_linearize(skb) < 0) {
725 ipoib_warn(priv, "skb could not be linearized\n");
726 ++dev->stats.tx_dropped;
727 ++dev->stats.tx_errors;
728 dev_kfree_skb_any(skb);
729 return;
730 }
731
732 if (skb_shinfo(skb)->nr_frags > usable_sge) {
733 ipoib_warn(priv, "too many frags after skb linearize\n");
734 ++dev->stats.tx_dropped;
735 ++dev->stats.tx_errors;
736 dev_kfree_skb_any(skb);
737 return;
738 }
739 }
740 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
741 tx->tx_head, skb->len, tx->qp->qp_num);
742
743
744
745
746
747
748
749
750 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
751 tx_req->skb = skb;
752
753 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
754 ++dev->stats.tx_errors;
755 dev_kfree_skb_any(skb);
756 return;
757 }
758
759 if ((priv->global_tx_head - priv->global_tx_tail) ==
760 ipoib_sendq_size - 1) {
761 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
762 tx->qp->qp_num);
763 netif_stop_queue(dev);
764 }
765
766 skb_orphan(skb);
767 skb_dst_drop(skb);
768
769 if (netif_queue_stopped(dev)) {
770 rc = ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
771 IB_CQ_REPORT_MISSED_EVENTS);
772 if (unlikely(rc < 0))
773 ipoib_warn(priv, "IPoIB/CM:request notify on send CQ failed\n");
774 else if (rc)
775 napi_schedule(&priv->send_napi);
776 }
777
778 rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
779 if (unlikely(rc)) {
780 ipoib_warn(priv, "IPoIB/CM:post_send failed, error %d\n", rc);
781 ++dev->stats.tx_errors;
782 ipoib_dma_unmap_tx(priv, tx_req);
783 dev_kfree_skb_any(skb);
784
785 if (netif_queue_stopped(dev))
786 netif_wake_queue(dev);
787 } else {
788 netif_trans_update(dev);
789 ++tx->tx_head;
790 ++priv->global_tx_head;
791 }
792 }
793
794 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
795 {
796 struct ipoib_dev_priv *priv = ipoib_priv(dev);
797 struct ipoib_cm_tx *tx = wc->qp->qp_context;
798 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
799 struct ipoib_tx_buf *tx_req;
800 unsigned long flags;
801
802 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
803 wr_id, wc->status);
804
805 if (unlikely(wr_id >= ipoib_sendq_size)) {
806 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
807 wr_id, ipoib_sendq_size);
808 return;
809 }
810
811 tx_req = &tx->tx_ring[wr_id];
812
813 ipoib_dma_unmap_tx(priv, tx_req);
814
815
816 ++dev->stats.tx_packets;
817 dev->stats.tx_bytes += tx_req->skb->len;
818
819 dev_kfree_skb_any(tx_req->skb);
820
821 netif_tx_lock(dev);
822
823 ++tx->tx_tail;
824 ++priv->global_tx_tail;
825
826 if (unlikely(netif_queue_stopped(dev) &&
827 ((priv->global_tx_head - priv->global_tx_tail) <=
828 ipoib_sendq_size >> 1) &&
829 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
830 netif_wake_queue(dev);
831
832 if (wc->status != IB_WC_SUCCESS &&
833 wc->status != IB_WC_WR_FLUSH_ERR) {
834 struct ipoib_neigh *neigh;
835
836
837
838
839 if (wc->status == IB_WC_RNR_RETRY_EXC_ERR ||
840 wc->status == IB_WC_RETRY_EXC_ERR)
841 ipoib_dbg(priv,
842 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
843 __func__, wc->status, wr_id, wc->vendor_err);
844 else
845 ipoib_warn(priv,
846 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
847 __func__, wc->status, wr_id, wc->vendor_err);
848
849 spin_lock_irqsave(&priv->lock, flags);
850 neigh = tx->neigh;
851
852 if (neigh) {
853 neigh->cm = NULL;
854 ipoib_neigh_free(neigh);
855
856 tx->neigh = NULL;
857 }
858
859 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
860 list_move(&tx->list, &priv->cm.reap_list);
861 queue_work(priv->wq, &priv->cm.reap_task);
862 }
863
864 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
865
866 spin_unlock_irqrestore(&priv->lock, flags);
867 }
868
869 netif_tx_unlock(dev);
870 }
871
872 int ipoib_cm_dev_open(struct net_device *dev)
873 {
874 struct ipoib_dev_priv *priv = ipoib_priv(dev);
875 int ret;
876
877 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
878 return 0;
879
880 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
881 if (IS_ERR(priv->cm.id)) {
882 pr_warn("%s: failed to create CM ID\n", priv->ca->name);
883 ret = PTR_ERR(priv->cm.id);
884 goto err_cm;
885 }
886
887 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
888 0);
889 if (ret) {
890 pr_warn("%s: failed to listen on ID 0x%llx\n", priv->ca->name,
891 IPOIB_CM_IETF_ID | priv->qp->qp_num);
892 goto err_listen;
893 }
894
895 return 0;
896
897 err_listen:
898 ib_destroy_cm_id(priv->cm.id);
899 err_cm:
900 priv->cm.id = NULL;
901 return ret;
902 }
903
904 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
905 {
906 struct ipoib_dev_priv *priv = ipoib_priv(dev);
907 struct ipoib_cm_rx *rx, *n;
908 LIST_HEAD(list);
909
910 spin_lock_irq(&priv->lock);
911 list_splice_init(&priv->cm.rx_reap_list, &list);
912 spin_unlock_irq(&priv->lock);
913
914 list_for_each_entry_safe(rx, n, &list, list) {
915 ib_destroy_cm_id(rx->id);
916 ib_destroy_qp(rx->qp);
917 if (!ipoib_cm_has_srq(dev)) {
918 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
919 spin_lock_irq(&priv->lock);
920 --priv->cm.nonsrq_conn_qp;
921 spin_unlock_irq(&priv->lock);
922 }
923 kfree(rx);
924 }
925 }
926
927 void ipoib_cm_dev_stop(struct net_device *dev)
928 {
929 struct ipoib_dev_priv *priv = ipoib_priv(dev);
930 struct ipoib_cm_rx *p;
931 unsigned long begin;
932 int ret;
933
934 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
935 return;
936
937 ib_destroy_cm_id(priv->cm.id);
938 priv->cm.id = NULL;
939
940 spin_lock_irq(&priv->lock);
941 while (!list_empty(&priv->cm.passive_ids)) {
942 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
943 list_move(&p->list, &priv->cm.rx_error_list);
944 p->state = IPOIB_CM_RX_ERROR;
945 spin_unlock_irq(&priv->lock);
946 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
947 if (ret)
948 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
949 spin_lock_irq(&priv->lock);
950 }
951
952
953 begin = jiffies;
954
955 while (!list_empty(&priv->cm.rx_error_list) ||
956 !list_empty(&priv->cm.rx_flush_list) ||
957 !list_empty(&priv->cm.rx_drain_list)) {
958 if (time_after(jiffies, begin + 5 * HZ)) {
959 ipoib_warn(priv, "RX drain timing out\n");
960
961
962
963
964 list_splice_init(&priv->cm.rx_flush_list,
965 &priv->cm.rx_reap_list);
966 list_splice_init(&priv->cm.rx_error_list,
967 &priv->cm.rx_reap_list);
968 list_splice_init(&priv->cm.rx_drain_list,
969 &priv->cm.rx_reap_list);
970 break;
971 }
972 spin_unlock_irq(&priv->lock);
973 usleep_range(1000, 2000);
974 ipoib_drain_cq(dev);
975 spin_lock_irq(&priv->lock);
976 }
977
978 spin_unlock_irq(&priv->lock);
979
980 ipoib_cm_free_rx_reap_list(dev);
981
982 cancel_delayed_work(&priv->cm.stale_task);
983 }
984
985 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id,
986 const struct ib_cm_event *event)
987 {
988 struct ipoib_cm_tx *p = cm_id->context;
989 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
990 struct ipoib_cm_data *data = event->private_data;
991 struct sk_buff_head skqueue;
992 struct ib_qp_attr qp_attr;
993 int qp_attr_mask, ret;
994 struct sk_buff *skb;
995
996 p->mtu = be32_to_cpu(data->mtu);
997
998 if (p->mtu <= IPOIB_ENCAP_LEN) {
999 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
1000 p->mtu, IPOIB_ENCAP_LEN);
1001 return -EINVAL;
1002 }
1003
1004 qp_attr.qp_state = IB_QPS_RTR;
1005 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1006 if (ret) {
1007 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
1008 return ret;
1009 }
1010
1011 qp_attr.rq_psn = 0 ;
1012 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1013 if (ret) {
1014 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
1015 return ret;
1016 }
1017
1018 qp_attr.qp_state = IB_QPS_RTS;
1019 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1020 if (ret) {
1021 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1022 return ret;
1023 }
1024 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1025 if (ret) {
1026 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1027 return ret;
1028 }
1029
1030 skb_queue_head_init(&skqueue);
1031
1032 netif_tx_lock_bh(p->dev);
1033 spin_lock_irq(&priv->lock);
1034 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1035 if (p->neigh)
1036 while ((skb = __skb_dequeue(&p->neigh->queue)))
1037 __skb_queue_tail(&skqueue, skb);
1038 spin_unlock_irq(&priv->lock);
1039 netif_tx_unlock_bh(p->dev);
1040
1041 while ((skb = __skb_dequeue(&skqueue))) {
1042 skb->dev = p->dev;
1043 ret = dev_queue_xmit(skb);
1044 if (ret)
1045 ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1046 __func__, ret);
1047 }
1048
1049 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1050 if (ret) {
1051 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1052 return ret;
1053 }
1054 return 0;
1055 }
1056
1057 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1058 {
1059 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1060 struct ib_qp_init_attr attr = {
1061 .send_cq = priv->send_cq,
1062 .recv_cq = priv->recv_cq,
1063 .srq = priv->cm.srq,
1064 .cap.max_send_wr = ipoib_sendq_size,
1065 .cap.max_send_sge = 1,
1066 .sq_sig_type = IB_SIGNAL_ALL_WR,
1067 .qp_type = IB_QPT_RC,
1068 .qp_context = tx,
1069 .create_flags = 0
1070 };
1071 struct ib_qp *tx_qp;
1072
1073 if (dev->features & NETIF_F_SG)
1074 attr.cap.max_send_sge = min_t(u32, priv->ca->attrs.max_send_sge,
1075 MAX_SKB_FRAGS + 1);
1076
1077 tx_qp = ib_create_qp(priv->pd, &attr);
1078 tx->max_send_sge = attr.cap.max_send_sge;
1079 return tx_qp;
1080 }
1081
1082 static int ipoib_cm_send_req(struct net_device *dev,
1083 struct ib_cm_id *id, struct ib_qp *qp,
1084 u32 qpn,
1085 struct sa_path_rec *pathrec)
1086 {
1087 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1088 struct ipoib_cm_data data = {};
1089 struct ib_cm_req_param req = {};
1090
1091 data.qpn = cpu_to_be32(priv->qp->qp_num);
1092 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1093
1094 req.primary_path = pathrec;
1095 req.alternate_path = NULL;
1096 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1097 req.qp_num = qp->qp_num;
1098 req.qp_type = qp->qp_type;
1099 req.private_data = &data;
1100 req.private_data_len = sizeof(data);
1101 req.flow_control = 0;
1102
1103 req.starting_psn = 0;
1104
1105
1106
1107
1108
1109 req.responder_resources = 4;
1110 req.remote_cm_response_timeout = 20;
1111 req.local_cm_response_timeout = 20;
1112 req.retry_count = 0;
1113 req.rnr_retry_count = 0;
1114 req.max_cm_retries = 15;
1115 req.srq = ipoib_cm_has_srq(dev);
1116 return ib_send_cm_req(id, &req);
1117 }
1118
1119 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1120 struct ib_cm_id *cm_id, struct ib_qp *qp)
1121 {
1122 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1123 struct ib_qp_attr qp_attr;
1124 int qp_attr_mask, ret;
1125 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1126 if (ret) {
1127 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1128 return ret;
1129 }
1130
1131 qp_attr.qp_state = IB_QPS_INIT;
1132 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1133 qp_attr.port_num = priv->port;
1134 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1135
1136 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1137 if (ret) {
1138 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1139 return ret;
1140 }
1141 return 0;
1142 }
1143
1144 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1145 struct sa_path_rec *pathrec)
1146 {
1147 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1148 unsigned int noio_flag;
1149 int ret;
1150
1151 noio_flag = memalloc_noio_save();
1152 p->tx_ring = vzalloc(array_size(ipoib_sendq_size, sizeof(*p->tx_ring)));
1153 if (!p->tx_ring) {
1154 memalloc_noio_restore(noio_flag);
1155 ret = -ENOMEM;
1156 goto err_tx;
1157 }
1158
1159 p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1160 memalloc_noio_restore(noio_flag);
1161 if (IS_ERR(p->qp)) {
1162 ret = PTR_ERR(p->qp);
1163 ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1164 goto err_qp;
1165 }
1166
1167 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1168 if (IS_ERR(p->id)) {
1169 ret = PTR_ERR(p->id);
1170 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1171 goto err_id;
1172 }
1173
1174 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
1175 if (ret) {
1176 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1177 goto err_modify_send;
1178 }
1179
1180 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1181 if (ret) {
1182 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1183 goto err_modify_send;
1184 }
1185
1186 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1187 p->qp->qp_num, pathrec->dgid.raw, qpn);
1188
1189 return 0;
1190
1191 err_modify_send:
1192 ib_destroy_cm_id(p->id);
1193 err_id:
1194 p->id = NULL;
1195 ib_destroy_qp(p->qp);
1196 err_qp:
1197 p->qp = NULL;
1198 vfree(p->tx_ring);
1199 err_tx:
1200 return ret;
1201 }
1202
1203 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1204 {
1205 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1206 struct ipoib_tx_buf *tx_req;
1207 unsigned long begin;
1208
1209 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1210 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1211
1212 if (p->id)
1213 ib_destroy_cm_id(p->id);
1214
1215 if (p->tx_ring) {
1216
1217 begin = jiffies;
1218 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1219 if (time_after(jiffies, begin + 5 * HZ)) {
1220 ipoib_warn(priv, "timing out; %d sends not completed\n",
1221 p->tx_head - p->tx_tail);
1222 goto timeout;
1223 }
1224
1225 usleep_range(1000, 2000);
1226 }
1227 }
1228
1229 timeout:
1230
1231 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1232 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1233 ipoib_dma_unmap_tx(priv, tx_req);
1234 dev_kfree_skb_any(tx_req->skb);
1235 netif_tx_lock_bh(p->dev);
1236 ++p->tx_tail;
1237 ++priv->global_tx_tail;
1238 if (unlikely((priv->global_tx_head - priv->global_tx_tail) <=
1239 ipoib_sendq_size >> 1) &&
1240 netif_queue_stopped(p->dev) &&
1241 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1242 netif_wake_queue(p->dev);
1243 netif_tx_unlock_bh(p->dev);
1244 }
1245
1246 if (p->qp)
1247 ib_destroy_qp(p->qp);
1248
1249 vfree(p->tx_ring);
1250 kfree(p);
1251 }
1252
1253 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1254 const struct ib_cm_event *event)
1255 {
1256 struct ipoib_cm_tx *tx = cm_id->context;
1257 struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1258 struct net_device *dev = priv->dev;
1259 struct ipoib_neigh *neigh;
1260 unsigned long flags;
1261 int ret;
1262
1263 switch (event->event) {
1264 case IB_CM_DREQ_RECEIVED:
1265 ipoib_dbg(priv, "DREQ received.\n");
1266 ib_send_cm_drep(cm_id, NULL, 0);
1267 break;
1268 case IB_CM_REP_RECEIVED:
1269 ipoib_dbg(priv, "REP received.\n");
1270 ret = ipoib_cm_rep_handler(cm_id, event);
1271 if (ret)
1272 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1273 NULL, 0, NULL, 0);
1274 break;
1275 case IB_CM_REQ_ERROR:
1276 case IB_CM_REJ_RECEIVED:
1277 case IB_CM_TIMEWAIT_EXIT:
1278 ipoib_dbg(priv, "CM error %d.\n", event->event);
1279 netif_tx_lock_bh(dev);
1280 spin_lock_irqsave(&priv->lock, flags);
1281 neigh = tx->neigh;
1282
1283 if (neigh) {
1284 neigh->cm = NULL;
1285 ipoib_neigh_free(neigh);
1286
1287 tx->neigh = NULL;
1288 }
1289
1290 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1291 list_move(&tx->list, &priv->cm.reap_list);
1292 queue_work(priv->wq, &priv->cm.reap_task);
1293 }
1294
1295 spin_unlock_irqrestore(&priv->lock, flags);
1296 netif_tx_unlock_bh(dev);
1297 break;
1298 default:
1299 break;
1300 }
1301
1302 return 0;
1303 }
1304
1305 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1306 struct ipoib_neigh *neigh)
1307 {
1308 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1309 struct ipoib_cm_tx *tx;
1310
1311 tx = kzalloc(sizeof(*tx), GFP_ATOMIC);
1312 if (!tx)
1313 return NULL;
1314
1315 neigh->cm = tx;
1316 tx->neigh = neigh;
1317 tx->dev = dev;
1318 list_add(&tx->list, &priv->cm.start_list);
1319 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1320 queue_work(priv->wq, &priv->cm.start_task);
1321 return tx;
1322 }
1323
1324 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1325 {
1326 struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1327 unsigned long flags;
1328 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1329 spin_lock_irqsave(&priv->lock, flags);
1330 list_move(&tx->list, &priv->cm.reap_list);
1331 queue_work(priv->wq, &priv->cm.reap_task);
1332 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1333 tx->neigh->daddr + 4);
1334 tx->neigh = NULL;
1335 spin_unlock_irqrestore(&priv->lock, flags);
1336 }
1337 }
1338
1339 #define QPN_AND_OPTIONS_OFFSET 4
1340
1341 static void ipoib_cm_tx_start(struct work_struct *work)
1342 {
1343 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1344 cm.start_task);
1345 struct net_device *dev = priv->dev;
1346 struct ipoib_neigh *neigh;
1347 struct ipoib_cm_tx *p;
1348 unsigned long flags;
1349 struct ipoib_path *path;
1350 int ret;
1351
1352 struct sa_path_rec pathrec;
1353 u32 qpn;
1354
1355 netif_tx_lock_bh(dev);
1356 spin_lock_irqsave(&priv->lock, flags);
1357
1358 while (!list_empty(&priv->cm.start_list)) {
1359 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1360 list_del_init(&p->list);
1361 neigh = p->neigh;
1362
1363 qpn = IPOIB_QPN(neigh->daddr);
1364
1365
1366
1367
1368 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1369 if (!path) {
1370 pr_info("%s ignore not valid path %pI6\n",
1371 __func__,
1372 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1373 goto free_neigh;
1374 }
1375 memcpy(&pathrec, &path->pathrec, sizeof(pathrec));
1376
1377 spin_unlock_irqrestore(&priv->lock, flags);
1378 netif_tx_unlock_bh(dev);
1379
1380 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1381
1382 netif_tx_lock_bh(dev);
1383 spin_lock_irqsave(&priv->lock, flags);
1384
1385 if (ret) {
1386 free_neigh:
1387 neigh = p->neigh;
1388 if (neigh) {
1389 neigh->cm = NULL;
1390 ipoib_neigh_free(neigh);
1391 }
1392 list_del(&p->list);
1393 kfree(p);
1394 }
1395 }
1396
1397 spin_unlock_irqrestore(&priv->lock, flags);
1398 netif_tx_unlock_bh(dev);
1399 }
1400
1401 static void ipoib_cm_tx_reap(struct work_struct *work)
1402 {
1403 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1404 cm.reap_task);
1405 struct net_device *dev = priv->dev;
1406 struct ipoib_cm_tx *p;
1407 unsigned long flags;
1408
1409 netif_tx_lock_bh(dev);
1410 spin_lock_irqsave(&priv->lock, flags);
1411
1412 while (!list_empty(&priv->cm.reap_list)) {
1413 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1414 list_del_init(&p->list);
1415 spin_unlock_irqrestore(&priv->lock, flags);
1416 netif_tx_unlock_bh(dev);
1417 ipoib_cm_tx_destroy(p);
1418 netif_tx_lock_bh(dev);
1419 spin_lock_irqsave(&priv->lock, flags);
1420 }
1421
1422 spin_unlock_irqrestore(&priv->lock, flags);
1423 netif_tx_unlock_bh(dev);
1424 }
1425
1426 static void ipoib_cm_skb_reap(struct work_struct *work)
1427 {
1428 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1429 cm.skb_task);
1430 struct net_device *dev = priv->dev;
1431 struct sk_buff *skb;
1432 unsigned long flags;
1433 unsigned int mtu = priv->mcast_mtu;
1434
1435 netif_tx_lock_bh(dev);
1436 spin_lock_irqsave(&priv->lock, flags);
1437
1438 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1439 spin_unlock_irqrestore(&priv->lock, flags);
1440 netif_tx_unlock_bh(dev);
1441
1442 if (skb->protocol == htons(ETH_P_IP)) {
1443 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1444 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1445 }
1446 #if IS_ENABLED(CONFIG_IPV6)
1447 else if (skb->protocol == htons(ETH_P_IPV6)) {
1448 memset(IP6CB(skb), 0, sizeof(*IP6CB(skb)));
1449 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1450 }
1451 #endif
1452 dev_kfree_skb_any(skb);
1453
1454 netif_tx_lock_bh(dev);
1455 spin_lock_irqsave(&priv->lock, flags);
1456 }
1457
1458 spin_unlock_irqrestore(&priv->lock, flags);
1459 netif_tx_unlock_bh(dev);
1460 }
1461
1462 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1463 unsigned int mtu)
1464 {
1465 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1466 int e = skb_queue_empty(&priv->cm.skb_queue);
1467
1468 skb_dst_update_pmtu(skb, mtu);
1469
1470 skb_queue_tail(&priv->cm.skb_queue, skb);
1471 if (e)
1472 queue_work(priv->wq, &priv->cm.skb_task);
1473 }
1474
1475 static void ipoib_cm_rx_reap(struct work_struct *work)
1476 {
1477 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1478 cm.rx_reap_task)->dev);
1479 }
1480
1481 static void ipoib_cm_stale_task(struct work_struct *work)
1482 {
1483 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1484 cm.stale_task.work);
1485 struct ipoib_cm_rx *p;
1486 int ret;
1487
1488 spin_lock_irq(&priv->lock);
1489 while (!list_empty(&priv->cm.passive_ids)) {
1490
1491
1492 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1493 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1494 break;
1495 list_move(&p->list, &priv->cm.rx_error_list);
1496 p->state = IPOIB_CM_RX_ERROR;
1497 spin_unlock_irq(&priv->lock);
1498 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1499 if (ret)
1500 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1501 spin_lock_irq(&priv->lock);
1502 }
1503
1504 if (!list_empty(&priv->cm.passive_ids))
1505 queue_delayed_work(priv->wq,
1506 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1507 spin_unlock_irq(&priv->lock);
1508 }
1509
1510 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1511 char *buf)
1512 {
1513 struct net_device *dev = to_net_dev(d);
1514 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1515
1516 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1517 return sprintf(buf, "connected\n");
1518 else
1519 return sprintf(buf, "datagram\n");
1520 }
1521
1522 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1523 const char *buf, size_t count)
1524 {
1525 struct net_device *dev = to_net_dev(d);
1526 int ret;
1527
1528 if (!rtnl_trylock()) {
1529 return restart_syscall();
1530 }
1531
1532 if (dev->reg_state != NETREG_REGISTERED) {
1533 rtnl_unlock();
1534 return -EPERM;
1535 }
1536
1537 ret = ipoib_set_mode(dev, buf);
1538
1539
1540
1541
1542
1543 if (ret != -EBUSY)
1544 rtnl_unlock();
1545
1546 return (!ret || ret == -EBUSY) ? count : ret;
1547 }
1548
1549 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1550
1551 int ipoib_cm_add_mode_attr(struct net_device *dev)
1552 {
1553 return device_create_file(&dev->dev, &dev_attr_mode);
1554 }
1555
1556 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1557 {
1558 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1559 struct ib_srq_init_attr srq_init_attr = {
1560 .srq_type = IB_SRQT_BASIC,
1561 .attr = {
1562 .max_wr = ipoib_recvq_size,
1563 .max_sge = max_sge
1564 }
1565 };
1566
1567 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1568 if (IS_ERR(priv->cm.srq)) {
1569 if (PTR_ERR(priv->cm.srq) != -EOPNOTSUPP)
1570 pr_warn("%s: failed to allocate SRQ, error %ld\n",
1571 priv->ca->name, PTR_ERR(priv->cm.srq));
1572 priv->cm.srq = NULL;
1573 return;
1574 }
1575
1576 priv->cm.srq_ring = vzalloc(array_size(ipoib_recvq_size,
1577 sizeof(*priv->cm.srq_ring)));
1578 if (!priv->cm.srq_ring) {
1579 ib_destroy_srq(priv->cm.srq);
1580 priv->cm.srq = NULL;
1581 return;
1582 }
1583
1584 }
1585
1586 int ipoib_cm_dev_init(struct net_device *dev)
1587 {
1588 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1589 int max_srq_sge, i;
1590
1591 INIT_LIST_HEAD(&priv->cm.passive_ids);
1592 INIT_LIST_HEAD(&priv->cm.reap_list);
1593 INIT_LIST_HEAD(&priv->cm.start_list);
1594 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1595 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1596 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1597 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1598 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1599 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1600 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1601 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1602 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1603
1604 skb_queue_head_init(&priv->cm.skb_queue);
1605
1606 ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1607
1608 max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1609 ipoib_cm_create_srq(dev, max_srq_sge);
1610 if (ipoib_cm_has_srq(dev)) {
1611 priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1612 priv->cm.num_frags = max_srq_sge;
1613 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1614 priv->cm.max_cm_mtu, priv->cm.num_frags);
1615 } else {
1616 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1617 priv->cm.num_frags = IPOIB_CM_RX_SG;
1618 }
1619
1620 ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1621
1622 if (ipoib_cm_has_srq(dev)) {
1623 for (i = 0; i < ipoib_recvq_size; ++i) {
1624 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1625 priv->cm.num_frags - 1,
1626 priv->cm.srq_ring[i].mapping,
1627 GFP_KERNEL)) {
1628 ipoib_warn(priv, "failed to allocate "
1629 "receive buffer %d\n", i);
1630 ipoib_cm_dev_cleanup(dev);
1631 return -ENOMEM;
1632 }
1633
1634 if (ipoib_cm_post_receive_srq(dev, i)) {
1635 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1636 "failed for buf %d\n", i);
1637 ipoib_cm_dev_cleanup(dev);
1638 return -EIO;
1639 }
1640 }
1641 }
1642
1643 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1644 return 0;
1645 }
1646
1647 void ipoib_cm_dev_cleanup(struct net_device *dev)
1648 {
1649 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1650 int ret;
1651
1652 if (!priv->cm.srq)
1653 return;
1654
1655 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1656
1657 ret = ib_destroy_srq(priv->cm.srq);
1658 if (ret)
1659 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1660
1661 priv->cm.srq = NULL;
1662 if (!priv->cm.srq_ring)
1663 return;
1664
1665 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1666 priv->cm.srq_ring = NULL;
1667 }