?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[wrt350n-kernel.git] / drivers / infiniband / ulp / ipoib / ipoib_cm.c
blob2948e33039b83b8ad6808b31065f3b13c23a6058
1 /*
2 * Copyright (c) 2006 Mellanox Technologies. All rights reserved
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 * $Id$
33 */
34
35 #include <rdma/ib_cm.h>
36 #include <rdma/ib_cache.h>
37 #include <net/dst.h>
38 #include <net/icmp.h>
39 #include <linux/icmpv6.h>
40 #include <linux/delay.h>
41 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
42 =======
43 #include <linux/vmalloc.h>
44 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
45
46 #include "ipoib.h"
47
48 int ipoib_max_conn_qp = 128;
49
50 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
51 MODULE_PARM_DESC(max_nonsrq_conn_qp,
52 "Max number of connected-mode QPs per interface "
53 "(applied only if shared receive queue is not available)");
54
55 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
56 static int data_debug_level;
57
58 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
59 MODULE_PARM_DESC(cm_data_debug_level,
60 "Enable data path debug tracing for connected mode if > 0");
61 #endif
62
63 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
64
65 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
66 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ)
67 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ)
68 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
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 .wr_id = IPOIB_CM_RX_DRAIN_WRID,
78 .opcode = IB_WR_SEND,
79 };
80
81 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
82 struct ib_cm_event *event);
83
84 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
85 u64 mapping[IPOIB_CM_RX_SG])
86 {
87 int i;
88
89 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
90
91 for (i = 0; i < frags; ++i)
92 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
93 }
94
95 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
96 {
97 struct ipoib_dev_priv *priv = netdev_priv(dev);
98 struct ib_recv_wr *bad_wr;
99 int i, ret;
100
101 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
102
103 for (i = 0; i < priv->cm.num_frags; ++i)
104 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
105
106 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
107 if (unlikely(ret)) {
108 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
109 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
110 priv->cm.srq_ring[id].mapping);
111 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
112 priv->cm.srq_ring[id].skb = NULL;
113 }
114
115 return ret;
116 }
117
118 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
119 struct ipoib_cm_rx *rx, int id)
120 {
121 struct ipoib_dev_priv *priv = netdev_priv(dev);
122 struct ib_recv_wr *bad_wr;
123 int i, ret;
124
125 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
126
127 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
128 priv->cm.rx_sge[i].addr = rx->rx_ring[id].mapping[i];
129
130 ret = ib_post_recv(rx->qp, &priv->cm.rx_wr, &bad_wr);
131 if (unlikely(ret)) {
132 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
133 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
134 rx->rx_ring[id].mapping);
135 dev_kfree_skb_any(rx->rx_ring[id].skb);
136 rx->rx_ring[id].skb = NULL;
137 }
138
139 return ret;
140 }
141
142 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
143 struct ipoib_cm_rx_buf *rx_ring,
144 int id, int frags,
145 u64 mapping[IPOIB_CM_RX_SG])
146 {
147 struct ipoib_dev_priv *priv = netdev_priv(dev);
148 struct sk_buff *skb;
149 int i;
150
151 skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
152 if (unlikely(!skb))
153 return NULL;
154
155 /*
156 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
157 * IP header to a multiple of 16.
158 */
159 skb_reserve(skb, 12);
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_ATOMIC);
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, skb_shinfo(skb)->frags[i].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_single(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 = netdev_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 kfree(rx_ring);
209 }
210
211 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
212 {
213 struct ib_send_wr *bad_wr;
214 struct ipoib_cm_rx *p;
215
216 /* We only reserved 1 extra slot in CQ for drain WRs, so
217 * make sure we have at most 1 outstanding WR. */
218 if (list_empty(&priv->cm.rx_flush_list) ||
219 !list_empty(&priv->cm.rx_drain_list))
220 return;
221
222 /*
223 * QPs on flush list are error state. This way, a "flush
224 * error" WC will be immediately generated for each WR we post.
225 */
226 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
227 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
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 = netdev_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 = netdev_priv(dev);
253 struct ib_qp_init_attr attr = {
254 .event_handler = ipoib_cm_rx_event_handler,
255 .send_cq = priv->cq, /* For drain WR */
256 .recv_cq = priv->cq,
257 .srq = priv->cm.srq,
258 .cap.max_send_wr = 1, /* For drain WR */
259 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
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 psn)
276 {
277 struct ipoib_dev_priv *priv = netdev_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 * Current Mellanox HCA firmware won't generate completions
307 * with error for drain WRs unless the QP has been moved to
308 * RTS first. This work-around leaves a window where a QP has
309 * moved to error asynchronously, but this will eventually get
310 * fixed in firmware, so let's not error out if modify QP
311 * fails.
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 int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
329 struct ipoib_cm_rx *rx)
330 {
331 struct ipoib_dev_priv *priv = netdev_priv(dev);
332 int ret;
333 int i;
334
335 rx->rx_ring = kcalloc(ipoib_recvq_size, sizeof *rx->rx_ring, GFP_KERNEL);
336 if (!rx->rx_ring)
337 return -ENOMEM;
338
339 spin_lock_irq(&priv->lock);
340
341 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
342 spin_unlock_irq(&priv->lock);
343 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
344 ret = -EINVAL;
345 goto err_free;
346 } else
347 ++priv->cm.nonsrq_conn_qp;
348
349 spin_unlock_irq(&priv->lock);
350
351 for (i = 0; i < ipoib_recvq_size; ++i) {
352 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
353 rx->rx_ring[i].mapping)) {
354 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
355 ret = -ENOMEM;
356 goto err_count;
357 }
358 ret = ipoib_cm_post_receive_nonsrq(dev, rx, i);
359 if (ret) {
360 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
361 "failed for buf %d\n", i);
362 ret = -EIO;
363 goto err_count;
364 }
365 }
366
367 rx->recv_count = ipoib_recvq_size;
368
369 return 0;
370
371 err_count:
372 spin_lock_irq(&priv->lock);
373 --priv->cm.nonsrq_conn_qp;
374 spin_unlock_irq(&priv->lock);
375
376 err_free:
377 ipoib_cm_free_rx_ring(dev, rx->rx_ring);
378
379 return ret;
380 }
381
382 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
383 struct ib_qp *qp, struct ib_cm_req_event_param *req,
384 unsigned psn)
385 {
386 struct ipoib_dev_priv *priv = netdev_priv(dev);
387 struct ipoib_cm_data data = {};
388 struct ib_cm_rep_param rep = {};
389
390 data.qpn = cpu_to_be32(priv->qp->qp_num);
391 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
392
393 rep.private_data = &data;
394 rep.private_data_len = sizeof data;
395 rep.flow_control = 0;
396 rep.rnr_retry_count = req->rnr_retry_count;
397 rep.srq = ipoib_cm_has_srq(dev);
398 rep.qp_num = qp->qp_num;
399 rep.starting_psn = psn;
400 return ib_send_cm_rep(cm_id, &rep);
401 }
402
403 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
404 {
405 struct net_device *dev = cm_id->context;
406 struct ipoib_dev_priv *priv = netdev_priv(dev);
407 struct ipoib_cm_rx *p;
408 unsigned psn;
409 int ret;
410
411 ipoib_dbg(priv, "REQ arrived\n");
412 p = kzalloc(sizeof *p, GFP_KERNEL);
413 if (!p)
414 return -ENOMEM;
415 p->dev = dev;
416 p->id = cm_id;
417 cm_id->context = p;
418 p->state = IPOIB_CM_RX_LIVE;
419 p->jiffies = jiffies;
420 INIT_LIST_HEAD(&p->list);
421
422 p->qp = ipoib_cm_create_rx_qp(dev, p);
423 if (IS_ERR(p->qp)) {
424 ret = PTR_ERR(p->qp);
425 goto err_qp;
426 }
427
428 psn = random32() & 0xffffff;
429 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
430 if (ret)
431 goto err_modify;
432
433 if (!ipoib_cm_has_srq(dev)) {
434 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
435 if (ret)
436 goto err_modify;
437 }
438
439 spin_lock_irq(&priv->lock);
440 queue_delayed_work(ipoib_workqueue,
441 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
442 /* Add this entry to passive ids list head, but do not re-add it
443 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
444 p->jiffies = jiffies;
445 if (p->state == IPOIB_CM_RX_LIVE)
446 list_move(&p->list, &priv->cm.passive_ids);
447 spin_unlock_irq(&priv->lock);
448
449 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
450 if (ret) {
451 ipoib_warn(priv, "failed to send REP: %d\n", ret);
452 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
453 ipoib_warn(priv, "unable to move qp to error state\n");
454 }
455 return 0;
456
457 err_modify:
458 ib_destroy_qp(p->qp);
459 err_qp:
460 kfree(p);
461 return ret;
462 }
463
464 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
465 struct ib_cm_event *event)
466 {
467 struct ipoib_cm_rx *p;
468 struct ipoib_dev_priv *priv;
469
470 switch (event->event) {
471 case IB_CM_REQ_RECEIVED:
472 return ipoib_cm_req_handler(cm_id, event);
473 case IB_CM_DREQ_RECEIVED:
474 p = cm_id->context;
475 ib_send_cm_drep(cm_id, NULL, 0);
476 /* Fall through */
477 case IB_CM_REJ_RECEIVED:
478 p = cm_id->context;
479 priv = netdev_priv(p->dev);
480 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
481 ipoib_warn(priv, "unable to move qp to error state\n");
482 /* Fall through */
483 default:
484 return 0;
485 }
486 }
487 /* Adjust length of skb with fragments to match received data */
488 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
489 unsigned int length, struct sk_buff *toskb)
490 {
491 int i, num_frags;
492 unsigned int size;
493
494 /* put header into skb */
495 size = min(length, hdr_space);
496 skb->tail += size;
497 skb->len += size;
498 length -= size;
499
500 num_frags = skb_shinfo(skb)->nr_frags;
501 for (i = 0; i < num_frags; i++) {
502 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
503
504 if (length == 0) {
505 /* don't need this page */
506 skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
507 --skb_shinfo(skb)->nr_frags;
508 } else {
509 size = min(length, (unsigned) PAGE_SIZE);
510
511 frag->size = size;
512 skb->data_len += size;
513 skb->truesize += size;
514 skb->len += size;
515 length -= size;
516 }
517 }
518 }
519
520 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
521 {
522 struct ipoib_dev_priv *priv = netdev_priv(dev);
523 struct ipoib_cm_rx_buf *rx_ring;
524 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
525 struct sk_buff *skb, *newskb;
526 struct ipoib_cm_rx *p;
527 unsigned long flags;
528 u64 mapping[IPOIB_CM_RX_SG];
529 int frags;
530 int has_srq;
531
532 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
533 wr_id, wc->status);
534
535 if (unlikely(wr_id >= ipoib_recvq_size)) {
536 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
537 spin_lock_irqsave(&priv->lock, flags);
538 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
539 ipoib_cm_start_rx_drain(priv);
540 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
541 spin_unlock_irqrestore(&priv->lock, flags);
542 } else
543 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
544 wr_id, ipoib_recvq_size);
545 return;
546 }
547
548 p = wc->qp->qp_context;
549
550 has_srq = ipoib_cm_has_srq(dev);
551 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
552
553 skb = rx_ring[wr_id].skb;
554
555 if (unlikely(wc->status != IB_WC_SUCCESS)) {
556 ipoib_dbg(priv, "cm recv error "
557 "(status=%d, wrid=%d vend_err %x)\n",
558 wc->status, wr_id, wc->vendor_err);
559 ++dev->stats.rx_dropped;
560 if (has_srq)
561 goto repost;
562 else {
563 if (!--p->recv_count) {
564 spin_lock_irqsave(&priv->lock, flags);
565 list_move(&p->list, &priv->cm.rx_reap_list);
566 spin_unlock_irqrestore(&priv->lock, flags);
567 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
568 }
569 return;
570 }
571 }
572
573 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
574 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
575 spin_lock_irqsave(&priv->lock, flags);
576 p->jiffies = jiffies;
577 /* Move this entry to list head, but do not re-add it
578 * if it has been moved out of list. */
579 if (p->state == IPOIB_CM_RX_LIVE)
580 list_move(&p->list, &priv->cm.passive_ids);
581 spin_unlock_irqrestore(&priv->lock, flags);
582 }
583 }
584
585 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
586 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
587
588 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
589 if (unlikely(!newskb)) {
590 /*
591 * If we can't allocate a new RX buffer, dump
592 * this packet and reuse the old buffer.
593 */
594 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
595 ++dev->stats.rx_dropped;
596 goto repost;
597 }
598
599 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
600 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
601
602 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
603 wc->byte_len, wc->slid);
604
605 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
606
607 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
608 skb_reset_mac_header(skb);
609 skb_pull(skb, IPOIB_ENCAP_LEN);
610
611 dev->last_rx = jiffies;
612 ++dev->stats.rx_packets;
613 dev->stats.rx_bytes += skb->len;
614
615 skb->dev = dev;
616 /* XXX get correct PACKET_ type here */
617 skb->pkt_type = PACKET_HOST;
618 netif_receive_skb(skb);
619
620 repost:
621 if (has_srq) {
622 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
623 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
624 "for buf %d\n", wr_id);
625 } else {
626 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, wr_id))) {
627 --p->recv_count;
628 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
629 "for buf %d\n", wr_id);
630 }
631 }
632 }
633
634 static inline int post_send(struct ipoib_dev_priv *priv,
635 struct ipoib_cm_tx *tx,
636 unsigned int wr_id,
637 u64 addr, int len)
638 {
639 struct ib_send_wr *bad_wr;
640
641 priv->tx_sge[0].addr = addr;
642 priv->tx_sge[0].length = len;
643
644 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
645 =======
646 priv->tx_wr.num_sge = 1;
647 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
648 priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM;
649
650 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
651 }
652
653 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
654 {
655 struct ipoib_dev_priv *priv = netdev_priv(dev);
656 struct ipoib_tx_buf *tx_req;
657 u64 addr;
658
659 if (unlikely(skb->len > tx->mtu)) {
660 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
661 skb->len, tx->mtu);
662 ++dev->stats.tx_dropped;
663 ++dev->stats.tx_errors;
664 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
665 return;
666 }
667
668 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
669 tx->tx_head, skb->len, tx->qp->qp_num);
670
671 /*
672 * We put the skb into the tx_ring _before_ we call post_send()
673 * because it's entirely possible that the completion handler will
674 * run before we execute anything after the post_send(). That
675 * means we have to make sure everything is properly recorded and
676 * our state is consistent before we call post_send().
677 */
678 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
679 tx_req->skb = skb;
680 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
681 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
682 ++dev->stats.tx_errors;
683 dev_kfree_skb_any(skb);
684 return;
685 }
686
687 tx_req->mapping[0] = addr;
688
689 if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
690 addr, skb->len))) {
691 ipoib_warn(priv, "post_send failed\n");
692 ++dev->stats.tx_errors;
693 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
694 dev_kfree_skb_any(skb);
695 } else {
696 dev->trans_start = jiffies;
697 ++tx->tx_head;
698
699 if (++priv->tx_outstanding == ipoib_sendq_size) {
700 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
701 tx->qp->qp_num);
702 netif_stop_queue(dev);
703 }
704 }
705 }
706
707 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
708 {
709 struct ipoib_dev_priv *priv = netdev_priv(dev);
710 struct ipoib_cm_tx *tx = wc->qp->qp_context;
711 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
712 struct ipoib_tx_buf *tx_req;
713 unsigned long flags;
714
715 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
716 wr_id, wc->status);
717
718 if (unlikely(wr_id >= ipoib_sendq_size)) {
719 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
720 wr_id, ipoib_sendq_size);
721 return;
722 }
723
724 tx_req = &tx->tx_ring[wr_id];
725
726 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len, DMA_TO_DEVICE);
727
728 /* FIXME: is this right? Shouldn't we only increment on success? */
729 ++dev->stats.tx_packets;
730 dev->stats.tx_bytes += tx_req->skb->len;
731
732 dev_kfree_skb_any(tx_req->skb);
733
734 spin_lock_irqsave(&priv->tx_lock, flags);
735 ++tx->tx_tail;
736 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
737 netif_queue_stopped(dev) &&
738 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
739 netif_wake_queue(dev);
740
741 if (wc->status != IB_WC_SUCCESS &&
742 wc->status != IB_WC_WR_FLUSH_ERR) {
743 struct ipoib_neigh *neigh;
744
745 ipoib_dbg(priv, "failed cm send event "
746 "(status=%d, wrid=%d vend_err %x)\n",
747 wc->status, wr_id, wc->vendor_err);
748
749 spin_lock(&priv->lock);
750 neigh = tx->neigh;
751
752 if (neigh) {
753 neigh->cm = NULL;
754 list_del(&neigh->list);
755 if (neigh->ah)
756 ipoib_put_ah(neigh->ah);
757 ipoib_neigh_free(dev, neigh);
758
759 tx->neigh = NULL;
760 }
761
762 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
763 list_move(&tx->list, &priv->cm.reap_list);
764 queue_work(ipoib_workqueue, &priv->cm.reap_task);
765 }
766
767 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
768
769 spin_unlock(&priv->lock);
770 }
771
772 spin_unlock_irqrestore(&priv->tx_lock, flags);
773 }
774
775 int ipoib_cm_dev_open(struct net_device *dev)
776 {
777 struct ipoib_dev_priv *priv = netdev_priv(dev);
778 int ret;
779
780 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
781 return 0;
782
783 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
784 if (IS_ERR(priv->cm.id)) {
785 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
786 ret = PTR_ERR(priv->cm.id);
787 goto err_cm;
788 }
789
790 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
791 0, NULL);
792 if (ret) {
793 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
794 IPOIB_CM_IETF_ID | priv->qp->qp_num);
795 goto err_listen;
796 }
797
798 return 0;
799
800 err_listen:
801 ib_destroy_cm_id(priv->cm.id);
802 err_cm:
803 priv->cm.id = NULL;
804 return ret;
805 }
806
807 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
808 {
809 struct ipoib_dev_priv *priv = netdev_priv(dev);
810 struct ipoib_cm_rx *rx, *n;
811 LIST_HEAD(list);
812
813 spin_lock_irq(&priv->lock);
814 list_splice_init(&priv->cm.rx_reap_list, &list);
815 spin_unlock_irq(&priv->lock);
816
817 list_for_each_entry_safe(rx, n, &list, list) {
818 ib_destroy_cm_id(rx->id);
819 ib_destroy_qp(rx->qp);
820 if (!ipoib_cm_has_srq(dev)) {
821 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
822 spin_lock_irq(&priv->lock);
823 --priv->cm.nonsrq_conn_qp;
824 spin_unlock_irq(&priv->lock);
825 }
826 kfree(rx);
827 }
828 }
829
830 void ipoib_cm_dev_stop(struct net_device *dev)
831 {
832 struct ipoib_dev_priv *priv = netdev_priv(dev);
833 struct ipoib_cm_rx *p;
834 unsigned long begin;
835 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
836 LIST_HEAD(list);
837 =======
838 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
839 int ret;
840
841 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
842 return;
843
844 ib_destroy_cm_id(priv->cm.id);
845 priv->cm.id = NULL;
846
847 spin_lock_irq(&priv->lock);
848 while (!list_empty(&priv->cm.passive_ids)) {
849 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
850 list_move(&p->list, &priv->cm.rx_error_list);
851 p->state = IPOIB_CM_RX_ERROR;
852 spin_unlock_irq(&priv->lock);
853 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
854 if (ret)
855 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
856 spin_lock_irq(&priv->lock);
857 }
858
859 /* Wait for all RX to be drained */
860 begin = jiffies;
861
862 while (!list_empty(&priv->cm.rx_error_list) ||
863 !list_empty(&priv->cm.rx_flush_list) ||
864 !list_empty(&priv->cm.rx_drain_list)) {
865 if (time_after(jiffies, begin + 5 * HZ)) {
866 ipoib_warn(priv, "RX drain timing out\n");
867
868 /*
869 * assume the HW is wedged and just free up everything.
870 */
871 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
872 list_splice_init(&priv->cm.rx_flush_list, &list);
873 list_splice_init(&priv->cm.rx_error_list, &list);
874 list_splice_init(&priv->cm.rx_drain_list, &list);
875 =======
876 list_splice_init(&priv->cm.rx_flush_list,
877 &priv->cm.rx_reap_list);
878 list_splice_init(&priv->cm.rx_error_list,
879 &priv->cm.rx_reap_list);
880 list_splice_init(&priv->cm.rx_drain_list,
881 &priv->cm.rx_reap_list);
882 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
883 break;
884 }
885 spin_unlock_irq(&priv->lock);
886 msleep(1);
887 ipoib_drain_cq(dev);
888 spin_lock_irq(&priv->lock);
889 }
890
891 spin_unlock_irq(&priv->lock);
892
893 ipoib_cm_free_rx_reap_list(dev);
894
895 cancel_delayed_work(&priv->cm.stale_task);
896 }
897
898 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
899 {
900 struct ipoib_cm_tx *p = cm_id->context;
901 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
902 struct ipoib_cm_data *data = event->private_data;
903 struct sk_buff_head skqueue;
904 struct ib_qp_attr qp_attr;
905 int qp_attr_mask, ret;
906 struct sk_buff *skb;
907
908 p->mtu = be32_to_cpu(data->mtu);
909
910 if (p->mtu <= IPOIB_ENCAP_LEN) {
911 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
912 p->mtu, IPOIB_ENCAP_LEN);
913 return -EINVAL;
914 }
915
916 qp_attr.qp_state = IB_QPS_RTR;
917 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
918 if (ret) {
919 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
920 return ret;
921 }
922
923 qp_attr.rq_psn = 0 /* FIXME */;
924 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
925 if (ret) {
926 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
927 return ret;
928 }
929
930 qp_attr.qp_state = IB_QPS_RTS;
931 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
932 if (ret) {
933 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
934 return ret;
935 }
936 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
937 if (ret) {
938 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
939 return ret;
940 }
941
942 skb_queue_head_init(&skqueue);
943
944 spin_lock_irq(&priv->lock);
945 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
946 if (p->neigh)
947 while ((skb = __skb_dequeue(&p->neigh->queue)))
948 __skb_queue_tail(&skqueue, skb);
949 spin_unlock_irq(&priv->lock);
950
951 while ((skb = __skb_dequeue(&skqueue))) {
952 skb->dev = p->dev;
953 if (dev_queue_xmit(skb))
954 ipoib_warn(priv, "dev_queue_xmit failed "
955 "to requeue packet\n");
956 }
957
958 ret = ib_send_cm_rtu(cm_id, NULL, 0);
959 if (ret) {
960 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
961 return ret;
962 }
963 return 0;
964 }
965
966 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
967 {
968 struct ipoib_dev_priv *priv = netdev_priv(dev);
969 struct ib_qp_init_attr attr = {
970 .send_cq = priv->cq,
971 .recv_cq = priv->cq,
972 .srq = priv->cm.srq,
973 .cap.max_send_wr = ipoib_sendq_size,
974 .cap.max_send_sge = 1,
975 .sq_sig_type = IB_SIGNAL_ALL_WR,
976 .qp_type = IB_QPT_RC,
977 .qp_context = tx
978 };
979
980 return ib_create_qp(priv->pd, &attr);
981 }
982
983 static int ipoib_cm_send_req(struct net_device *dev,
984 struct ib_cm_id *id, struct ib_qp *qp,
985 u32 qpn,
986 struct ib_sa_path_rec *pathrec)
987 {
988 struct ipoib_dev_priv *priv = netdev_priv(dev);
989 struct ipoib_cm_data data = {};
990 struct ib_cm_req_param req = {};
991
992 data.qpn = cpu_to_be32(priv->qp->qp_num);
993 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
994
995 req.primary_path = pathrec;
996 req.alternate_path = NULL;
997 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
998 req.qp_num = qp->qp_num;
999 req.qp_type = qp->qp_type;
1000 req.private_data = &data;
1001 req.private_data_len = sizeof data;
1002 req.flow_control = 0;
1003
1004 req.starting_psn = 0; /* FIXME */
1005
1006 /*
1007 * Pick some arbitrary defaults here; we could make these
1008 * module parameters if anyone cared about setting them.
1009 */
1010 req.responder_resources = 4;
1011 req.remote_cm_response_timeout = 20;
1012 req.local_cm_response_timeout = 20;
1013 req.retry_count = 0; /* RFC draft warns against retries */
1014 req.rnr_retry_count = 0; /* RFC draft warns against retries */
1015 req.max_cm_retries = 15;
1016 req.srq = ipoib_cm_has_srq(dev);
1017 return ib_send_cm_req(id, &req);
1018 }
1019
1020 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1021 struct ib_cm_id *cm_id, struct ib_qp *qp)
1022 {
1023 struct ipoib_dev_priv *priv = netdev_priv(dev);
1024 struct ib_qp_attr qp_attr;
1025 int qp_attr_mask, ret;
1026 ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1027 if (ret) {
1028 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
1029 return ret;
1030 }
1031
1032 qp_attr.qp_state = IB_QPS_INIT;
1033 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1034 qp_attr.port_num = priv->port;
1035 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1036
1037 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1038 if (ret) {
1039 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1040 return ret;
1041 }
1042 return 0;
1043 }
1044
1045 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1046 struct ib_sa_path_rec *pathrec)
1047 {
1048 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1049 int ret;
1050
1051 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1052 p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1053 GFP_KERNEL);
1054 =======
1055 p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1056 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1057 if (!p->tx_ring) {
1058 ipoib_warn(priv, "failed to allocate tx ring\n");
1059 ret = -ENOMEM;
1060 goto err_tx;
1061 }
1062 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1063 =======
1064 memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1065 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1066
1067 p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1068 if (IS_ERR(p->qp)) {
1069 ret = PTR_ERR(p->qp);
1070 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1071 goto err_qp;
1072 }
1073
1074 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1075 if (IS_ERR(p->id)) {
1076 ret = PTR_ERR(p->id);
1077 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1078 goto err_id;
1079 }
1080
1081 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
1082 if (ret) {
1083 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1084 goto err_modify;
1085 }
1086
1087 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1088 if (ret) {
1089 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1090 goto err_send_cm;
1091 }
1092
1093 ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
1094 p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
1095
1096 return 0;
1097
1098 err_send_cm:
1099 err_modify:
1100 ib_destroy_cm_id(p->id);
1101 err_id:
1102 p->id = NULL;
1103 ib_destroy_qp(p->qp);
1104 err_qp:
1105 p->qp = NULL;
1106 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1107 =======
1108 vfree(p->tx_ring);
1109 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1110 err_tx:
1111 return ret;
1112 }
1113
1114 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1115 {
1116 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1117 struct ipoib_tx_buf *tx_req;
1118 unsigned long flags;
1119 unsigned long begin;
1120
1121 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1122 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1123
1124 if (p->id)
1125 ib_destroy_cm_id(p->id);
1126
1127 if (p->tx_ring) {
1128 /* Wait for all sends to complete */
1129 begin = jiffies;
1130 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1131 if (time_after(jiffies, begin + 5 * HZ)) {
1132 ipoib_warn(priv, "timing out; %d sends not completed\n",
1133 p->tx_head - p->tx_tail);
1134 goto timeout;
1135 }
1136
1137 msleep(1);
1138 }
1139 }
1140
1141 timeout:
1142
1143 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1144 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1145 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
1146 DMA_TO_DEVICE);
1147 dev_kfree_skb_any(tx_req->skb);
1148 ++p->tx_tail;
1149 spin_lock_irqsave(&priv->tx_lock, flags);
1150 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1151 netif_queue_stopped(p->dev) &&
1152 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1153 netif_wake_queue(p->dev);
1154 spin_unlock_irqrestore(&priv->tx_lock, flags);
1155 }
1156
1157 if (p->qp)
1158 ib_destroy_qp(p->qp);
1159
1160 <<<<<<< HEAD:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1161 kfree(p->tx_ring);
1162 =======
1163 vfree(p->tx_ring);
1164 >>>>>>> 264e3e889d86e552b4191d69bb60f4f3b383135a:drivers/infiniband/ulp/ipoib/ipoib_cm.c
1165 kfree(p);
1166 }
1167
1168 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1169 struct ib_cm_event *event)
1170 {
1171 struct ipoib_cm_tx *tx = cm_id->context;
1172 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1173 struct net_device *dev = priv->dev;
1174 struct ipoib_neigh *neigh;
1175 int ret;
1176
1177 switch (event->event) {
1178 case IB_CM_DREQ_RECEIVED:
1179 ipoib_dbg(priv, "DREQ received.\n");
1180 ib_send_cm_drep(cm_id, NULL, 0);
1181 break;
1182 case IB_CM_REP_RECEIVED:
1183 ipoib_dbg(priv, "REP received.\n");
1184 ret = ipoib_cm_rep_handler(cm_id, event);
1185 if (ret)
1186 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1187 NULL, 0, NULL, 0);
1188 break;
1189 case IB_CM_REQ_ERROR:
1190 case IB_CM_REJ_RECEIVED:
1191 case IB_CM_TIMEWAIT_EXIT:
1192 ipoib_dbg(priv, "CM error %d.\n", event->event);
1193 spin_lock_irq(&priv->tx_lock);
1194 spin_lock(&priv->lock);
1195 neigh = tx->neigh;
1196
1197 if (neigh) {
1198 neigh->cm = NULL;
1199 list_del(&neigh->list);
1200 if (neigh->ah)
1201 ipoib_put_ah(neigh->ah);
1202 ipoib_neigh_free(dev, neigh);
1203
1204 tx->neigh = NULL;
1205 }
1206
1207 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1208 list_move(&tx->list, &priv->cm.reap_list);
1209 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1210 }
1211
1212 spin_unlock(&priv->lock);
1213 spin_unlock_irq(&priv->tx_lock);
1214 break;
1215 default:
1216 break;
1217 }
1218
1219 return 0;
1220 }
1221
1222 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1223 struct ipoib_neigh *neigh)
1224 {
1225 struct ipoib_dev_priv *priv = netdev_priv(dev);
1226 struct ipoib_cm_tx *tx;
1227
1228 tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1229 if (!tx)
1230 return NULL;
1231
1232 neigh->cm = tx;
1233 tx->neigh = neigh;
1234 tx->path = path;
1235 tx->dev = dev;
1236 list_add(&tx->list, &priv->cm.start_list);
1237 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1238 queue_work(ipoib_workqueue, &priv->cm.start_task);
1239 return tx;
1240 }
1241
1242 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1243 {
1244 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1245 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1246 list_move(&tx->list, &priv->cm.reap_list);
1247 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1248 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1249 IPOIB_GID_ARG(tx->neigh->dgid));
1250 tx->neigh = NULL;
1251 }
1252 }
1253
1254 static void ipoib_cm_tx_start(struct work_struct *work)
1255 {
1256 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1257 cm.start_task);
1258 struct net_device *dev = priv->dev;
1259 struct ipoib_neigh *neigh;
1260 struct ipoib_cm_tx *p;
1261 unsigned long flags;
1262 int ret;
1263
1264 struct ib_sa_path_rec pathrec;
1265 u32 qpn;
1266
1267 spin_lock_irqsave(&priv->tx_lock, flags);
1268 spin_lock(&priv->lock);
1269 while (!list_empty(&priv->cm.start_list)) {
1270 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1271 list_del_init(&p->list);
1272 neigh = p->neigh;
1273 qpn = IPOIB_QPN(neigh->neighbour->ha);
1274 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1275 spin_unlock(&priv->lock);
1276 spin_unlock_irqrestore(&priv->tx_lock, flags);
1277 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1278 spin_lock_irqsave(&priv->tx_lock, flags);
1279 spin_lock(&priv->lock);
1280 if (ret) {
1281 neigh = p->neigh;
1282 if (neigh) {
1283 neigh->cm = NULL;
1284 list_del(&neigh->list);
1285 if (neigh->ah)
1286 ipoib_put_ah(neigh->ah);
1287 ipoib_neigh_free(dev, neigh);
1288 }
1289 list_del(&p->list);
1290 kfree(p);
1291 }
1292 }
1293 spin_unlock(&priv->lock);
1294 spin_unlock_irqrestore(&priv->tx_lock, flags);
1295 }
1296
1297 static void ipoib_cm_tx_reap(struct work_struct *work)
1298 {
1299 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1300 cm.reap_task);
1301 struct ipoib_cm_tx *p;
1302
1303 spin_lock_irq(&priv->tx_lock);
1304 spin_lock(&priv->lock);
1305 while (!list_empty(&priv->cm.reap_list)) {
1306 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1307 list_del(&p->list);
1308 spin_unlock(&priv->lock);
1309 spin_unlock_irq(&priv->tx_lock);
1310 ipoib_cm_tx_destroy(p);
1311 spin_lock_irq(&priv->tx_lock);
1312 spin_lock(&priv->lock);
1313 }
1314 spin_unlock(&priv->lock);
1315 spin_unlock_irq(&priv->tx_lock);
1316 }
1317
1318 static void ipoib_cm_skb_reap(struct work_struct *work)
1319 {
1320 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1321 cm.skb_task);
1322 struct sk_buff *skb;
1323
1324 unsigned mtu = priv->mcast_mtu;
1325
1326 spin_lock_irq(&priv->tx_lock);
1327 spin_lock(&priv->lock);
1328 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1329 spin_unlock(&priv->lock);
1330 spin_unlock_irq(&priv->tx_lock);
1331 if (skb->protocol == htons(ETH_P_IP))
1332 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1333 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1334 else if (skb->protocol == htons(ETH_P_IPV6))
1335 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1336 #endif
1337 dev_kfree_skb_any(skb);
1338 spin_lock_irq(&priv->tx_lock);
1339 spin_lock(&priv->lock);
1340 }
1341 spin_unlock(&priv->lock);
1342 spin_unlock_irq(&priv->tx_lock);
1343 }
1344
1345 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1346 unsigned int mtu)
1347 {
1348 struct ipoib_dev_priv *priv = netdev_priv(dev);
1349 int e = skb_queue_empty(&priv->cm.skb_queue);
1350
1351 if (skb->dst)
1352 skb->dst->ops->update_pmtu(skb->dst, mtu);
1353
1354 skb_queue_tail(&priv->cm.skb_queue, skb);
1355 if (e)
1356 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1357 }
1358
1359 static void ipoib_cm_rx_reap(struct work_struct *work)
1360 {
1361 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1362 cm.rx_reap_task)->dev);
1363 }
1364
1365 static void ipoib_cm_stale_task(struct work_struct *work)
1366 {
1367 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1368 cm.stale_task.work);
1369 struct ipoib_cm_rx *p;
1370 int ret;
1371
1372 spin_lock_irq(&priv->lock);
1373 while (!list_empty(&priv->cm.passive_ids)) {
1374 /* List is sorted by LRU, start from tail,
1375 * stop when we see a recently used entry */
1376 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1377 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1378 break;
1379 list_move(&p->list, &priv->cm.rx_error_list);
1380 p->state = IPOIB_CM_RX_ERROR;
1381 spin_unlock_irq(&priv->lock);
1382 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1383 if (ret)
1384 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1385 spin_lock_irq(&priv->lock);
1386 }
1387
1388 if (!list_empty(&priv->cm.passive_ids))
1389 queue_delayed_work(ipoib_workqueue,
1390 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1391 spin_unlock_irq(&priv->lock);
1392 }
1393
1394
1395 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1396 char *buf)
1397 {
1398 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1399
1400 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1401 return sprintf(buf, "connected\n");
1402 else
1403 return sprintf(buf, "datagram\n");
1404 }
1405
1406 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1407 const char *buf, size_t count)
1408 {
1409 struct net_device *dev = to_net_dev(d);
1410 struct ipoib_dev_priv *priv = netdev_priv(dev);
1411
1412 /* flush paths if we switch modes so that connections are restarted */
1413 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1414 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1415 ipoib_warn(priv, "enabling connected mode "
1416 "will cause multicast packet drops\n");
1417 ipoib_flush_paths(dev);
1418 return count;
1419 }
1420
1421 if (!strcmp(buf, "datagram\n")) {
1422 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1423 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1424 ipoib_flush_paths(dev);
1425 return count;
1426 }
1427
1428 return -EINVAL;
1429 }
1430
1431 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1432
1433 int ipoib_cm_add_mode_attr(struct net_device *dev)
1434 {
1435 return device_create_file(&dev->dev, &dev_attr_mode);
1436 }
1437
1438 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1439 {
1440 struct ipoib_dev_priv *priv = netdev_priv(dev);
1441 struct ib_srq_init_attr srq_init_attr = {
1442 .attr = {
1443 .max_wr = ipoib_recvq_size,
1444 .max_sge = max_sge
1445 }
1446 };
1447
1448 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1449 if (IS_ERR(priv->cm.srq)) {
1450 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1451 printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1452 priv->ca->name, PTR_ERR(priv->cm.srq));
1453 priv->cm.srq = NULL;
1454 return;
1455 }
1456
1457 priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1458 GFP_KERNEL);
1459 if (!priv->cm.srq_ring) {
1460 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1461 priv->ca->name, ipoib_recvq_size);
1462 ib_destroy_srq(priv->cm.srq);
1463 priv->cm.srq = NULL;
1464 }
1465 }
1466
1467 int ipoib_cm_dev_init(struct net_device *dev)
1468 {
1469 struct ipoib_dev_priv *priv = netdev_priv(dev);
1470 int i, ret;
1471 struct ib_device_attr attr;
1472
1473 INIT_LIST_HEAD(&priv->cm.passive_ids);
1474 INIT_LIST_HEAD(&priv->cm.reap_list);
1475 INIT_LIST_HEAD(&priv->cm.start_list);
1476 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1477 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1478 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1479 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1480 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1481 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1482 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1483 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1484 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1485
1486 skb_queue_head_init(&priv->cm.skb_queue);
1487
1488 ret = ib_query_device(priv->ca, &attr);
1489 if (ret) {
1490 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1491 return ret;
1492 }
1493
1494 ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1495
1496 attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1497 ipoib_cm_create_srq(dev, attr.max_srq_sge);
1498 if (ipoib_cm_has_srq(dev)) {
1499 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1500 priv->cm.num_frags = attr.max_srq_sge;
1501 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1502 priv->cm.max_cm_mtu, priv->cm.num_frags);
1503 } else {
1504 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1505 priv->cm.num_frags = IPOIB_CM_RX_SG;
1506 }
1507
1508 for (i = 0; i < priv->cm.num_frags; ++i)
1509 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1510
1511 priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1512 for (i = 1; i < priv->cm.num_frags; ++i)
1513 priv->cm.rx_sge[i].length = PAGE_SIZE;
1514 priv->cm.rx_wr.next = NULL;
1515 priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1516 priv->cm.rx_wr.num_sge = priv->cm.num_frags;
1517
1518 if (ipoib_cm_has_srq(dev)) {
1519 for (i = 0; i < ipoib_recvq_size; ++i) {
1520 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1521 priv->cm.num_frags - 1,
1522 priv->cm.srq_ring[i].mapping)) {
1523 ipoib_warn(priv, "failed to allocate "
1524 "receive buffer %d\n", i);
1525 ipoib_cm_dev_cleanup(dev);
1526 return -ENOMEM;
1527 }
1528
1529 if (ipoib_cm_post_receive_srq(dev, i)) {
1530 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1531 "failed for buf %d\n", i);
1532 ipoib_cm_dev_cleanup(dev);
1533 return -EIO;
1534 }
1535 }
1536 }
1537
1538 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1539 return 0;
1540 }
1541
1542 void ipoib_cm_dev_cleanup(struct net_device *dev)
1543 {
1544 struct ipoib_dev_priv *priv = netdev_priv(dev);
1545 int ret;
1546
1547 if (!priv->cm.srq)
1548 return;
1549
1550 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1551
1552 ret = ib_destroy_srq(priv->cm.srq);
1553 if (ret)
1554 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1555
1556 priv->cm.srq = NULL;
1557 if (!priv->cm.srq_ring)
1558 return;
1559
1560 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1561 priv->cm.srq_ring = NULL;
1562 }
WRT350N Kernel Patches