xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision f15cbe6f1a4b4d9df59142fc8e4abb973302cf44)
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 
33 #include <rdma/ib_cm.h>
34 #include <rdma/ib_cache.h>
35 #include <net/dst.h>
36 #include <net/icmp.h>
37 #include <linux/icmpv6.h>
38 #include <linux/delay.h>
39 #include <linux/vmalloc.h>
40 
41 #include "ipoib.h"
42 
43 int ipoib_max_conn_qp = 128;
44 
45 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
46 MODULE_PARM_DESC(max_nonsrq_conn_qp,
47 		 "Max number of connected-mode QPs per interface "
48 		 "(applied only if shared receive queue is not available)");
49 
50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
51 static int data_debug_level;
52 
53 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
54 MODULE_PARM_DESC(cm_data_debug_level,
55 		 "Enable data path debug tracing for connected mode if > 0");
56 #endif
57 
58 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
59 
60 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
61 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
62 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
63 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
64 
65 static struct ib_qp_attr ipoib_cm_err_attr = {
66 	.qp_state = IB_QPS_ERR
67 };
68 
69 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
70 
71 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
72 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
73 	.opcode = IB_WR_SEND,
74 };
75 
76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
77 			       struct ib_cm_event *event);
78 
79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
80 				  u64 mapping[IPOIB_CM_RX_SG])
81 {
82 	int i;
83 
84 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
85 
86 	for (i = 0; i < frags; ++i)
87 		ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
88 }
89 
90 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
91 {
92 	struct ipoib_dev_priv *priv = netdev_priv(dev);
93 	struct ib_recv_wr *bad_wr;
94 	int i, ret;
95 
96 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
97 
98 	for (i = 0; i < priv->cm.num_frags; ++i)
99 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
100 
101 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
102 	if (unlikely(ret)) {
103 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
104 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
105 				      priv->cm.srq_ring[id].mapping);
106 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
107 		priv->cm.srq_ring[id].skb = NULL;
108 	}
109 
110 	return ret;
111 }
112 
113 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
114 					struct ipoib_cm_rx *rx,
115 					struct ib_recv_wr *wr,
116 					struct ib_sge *sge, int id)
117 {
118 	struct ipoib_dev_priv *priv = netdev_priv(dev);
119 	struct ib_recv_wr *bad_wr;
120 	int i, ret;
121 
122 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
123 
124 	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
125 		sge[i].addr = rx->rx_ring[id].mapping[i];
126 
127 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
128 	if (unlikely(ret)) {
129 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
130 		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
131 				      rx->rx_ring[id].mapping);
132 		dev_kfree_skb_any(rx->rx_ring[id].skb);
133 		rx->rx_ring[id].skb = NULL;
134 	}
135 
136 	return ret;
137 }
138 
139 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
140 					     struct ipoib_cm_rx_buf *rx_ring,
141 					     int id, int frags,
142 					     u64 mapping[IPOIB_CM_RX_SG])
143 {
144 	struct ipoib_dev_priv *priv = netdev_priv(dev);
145 	struct sk_buff *skb;
146 	int i;
147 
148 	skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
149 	if (unlikely(!skb))
150 		return NULL;
151 
152 	/*
153 	 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
154 	 * IP header to a multiple of 16.
155 	 */
156 	skb_reserve(skb, 12);
157 
158 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
159 				       DMA_FROM_DEVICE);
160 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
161 		dev_kfree_skb_any(skb);
162 		return NULL;
163 	}
164 
165 	for (i = 0; i < frags; i++) {
166 		struct page *page = alloc_page(GFP_ATOMIC);
167 
168 		if (!page)
169 			goto partial_error;
170 		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
171 
172 		mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
173 						 0, PAGE_SIZE, DMA_FROM_DEVICE);
174 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
175 			goto partial_error;
176 	}
177 
178 	rx_ring[id].skb = skb;
179 	return skb;
180 
181 partial_error:
182 
183 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
184 
185 	for (; i > 0; --i)
186 		ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
187 
188 	dev_kfree_skb_any(skb);
189 	return NULL;
190 }
191 
192 static void ipoib_cm_free_rx_ring(struct net_device *dev,
193 				  struct ipoib_cm_rx_buf *rx_ring)
194 {
195 	struct ipoib_dev_priv *priv = netdev_priv(dev);
196 	int i;
197 
198 	for (i = 0; i < ipoib_recvq_size; ++i)
199 		if (rx_ring[i].skb) {
200 			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
201 					      rx_ring[i].mapping);
202 			dev_kfree_skb_any(rx_ring[i].skb);
203 		}
204 
205 	kfree(rx_ring);
206 }
207 
208 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
209 {
210 	struct ib_send_wr *bad_wr;
211 	struct ipoib_cm_rx *p;
212 
213 	/* We only reserved 1 extra slot in CQ for drain WRs, so
214 	 * make sure we have at most 1 outstanding WR. */
215 	if (list_empty(&priv->cm.rx_flush_list) ||
216 	    !list_empty(&priv->cm.rx_drain_list))
217 		return;
218 
219 	/*
220 	 * QPs on flush list are error state.  This way, a "flush
221 	 * error" WC will be immediately generated for each WR we post.
222 	 */
223 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
224 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
225 		ipoib_warn(priv, "failed to post drain wr\n");
226 
227 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
228 }
229 
230 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
231 {
232 	struct ipoib_cm_rx *p = ctx;
233 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
234 	unsigned long flags;
235 
236 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
237 		return;
238 
239 	spin_lock_irqsave(&priv->lock, flags);
240 	list_move(&p->list, &priv->cm.rx_flush_list);
241 	p->state = IPOIB_CM_RX_FLUSH;
242 	ipoib_cm_start_rx_drain(priv);
243 	spin_unlock_irqrestore(&priv->lock, flags);
244 }
245 
246 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
247 					   struct ipoib_cm_rx *p)
248 {
249 	struct ipoib_dev_priv *priv = netdev_priv(dev);
250 	struct ib_qp_init_attr attr = {
251 		.event_handler = ipoib_cm_rx_event_handler,
252 		.send_cq = priv->recv_cq, /* For drain WR */
253 		.recv_cq = priv->recv_cq,
254 		.srq = priv->cm.srq,
255 		.cap.max_send_wr = 1, /* For drain WR */
256 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
257 		.sq_sig_type = IB_SIGNAL_ALL_WR,
258 		.qp_type = IB_QPT_RC,
259 		.qp_context = p,
260 	};
261 
262 	if (!ipoib_cm_has_srq(dev)) {
263 		attr.cap.max_recv_wr  = ipoib_recvq_size;
264 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
265 	}
266 
267 	return ib_create_qp(priv->pd, &attr);
268 }
269 
270 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
271 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
272 				 unsigned psn)
273 {
274 	struct ipoib_dev_priv *priv = netdev_priv(dev);
275 	struct ib_qp_attr qp_attr;
276 	int qp_attr_mask, ret;
277 
278 	qp_attr.qp_state = IB_QPS_INIT;
279 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
280 	if (ret) {
281 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
282 		return ret;
283 	}
284 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
285 	if (ret) {
286 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
287 		return ret;
288 	}
289 	qp_attr.qp_state = IB_QPS_RTR;
290 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
291 	if (ret) {
292 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
293 		return ret;
294 	}
295 	qp_attr.rq_psn = psn;
296 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
297 	if (ret) {
298 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
299 		return ret;
300 	}
301 
302 	/*
303 	 * Current Mellanox HCA firmware won't generate completions
304 	 * with error for drain WRs unless the QP has been moved to
305 	 * RTS first. This work-around leaves a window where a QP has
306 	 * moved to error asynchronously, but this will eventually get
307 	 * fixed in firmware, so let's not error out if modify QP
308 	 * fails.
309 	 */
310 	qp_attr.qp_state = IB_QPS_RTS;
311 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
312 	if (ret) {
313 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
314 		return 0;
315 	}
316 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
317 	if (ret) {
318 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
319 		return 0;
320 	}
321 
322 	return 0;
323 }
324 
325 static void ipoib_cm_init_rx_wr(struct net_device *dev,
326 				struct ib_recv_wr *wr,
327 				struct ib_sge *sge)
328 {
329 	struct ipoib_dev_priv *priv = netdev_priv(dev);
330 	int i;
331 
332 	for (i = 0; i < priv->cm.num_frags; ++i)
333 		sge[i].lkey = priv->mr->lkey;
334 
335 	sge[0].length = IPOIB_CM_HEAD_SIZE;
336 	for (i = 1; i < priv->cm.num_frags; ++i)
337 		sge[i].length = PAGE_SIZE;
338 
339 	wr->next    = NULL;
340 	wr->sg_list = priv->cm.rx_sge;
341 	wr->num_sge = priv->cm.num_frags;
342 }
343 
344 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
345 				   struct ipoib_cm_rx *rx)
346 {
347 	struct ipoib_dev_priv *priv = netdev_priv(dev);
348 	struct {
349 		struct ib_recv_wr wr;
350 		struct ib_sge sge[IPOIB_CM_RX_SG];
351 	} *t;
352 	int ret;
353 	int i;
354 
355 	rx->rx_ring = kcalloc(ipoib_recvq_size, sizeof *rx->rx_ring, GFP_KERNEL);
356 	if (!rx->rx_ring)
357 		return -ENOMEM;
358 
359 	t = kmalloc(sizeof *t, GFP_KERNEL);
360 	if (!t) {
361 		ret = -ENOMEM;
362 		goto err_free;
363 	}
364 
365 	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
366 
367 	spin_lock_irq(&priv->lock);
368 
369 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
370 		spin_unlock_irq(&priv->lock);
371 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
372 		ret = -EINVAL;
373 		goto err_free;
374 	} else
375 		++priv->cm.nonsrq_conn_qp;
376 
377 	spin_unlock_irq(&priv->lock);
378 
379 	for (i = 0; i < ipoib_recvq_size; ++i) {
380 		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
381 					   rx->rx_ring[i].mapping)) {
382 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
383 				ret = -ENOMEM;
384 				goto err_count;
385 		}
386 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
387 		if (ret) {
388 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
389 				   "failed for buf %d\n", i);
390 			ret = -EIO;
391 			goto err_count;
392 		}
393 	}
394 
395 	rx->recv_count = ipoib_recvq_size;
396 
397 	kfree(t);
398 
399 	return 0;
400 
401 err_count:
402 	spin_lock_irq(&priv->lock);
403 	--priv->cm.nonsrq_conn_qp;
404 	spin_unlock_irq(&priv->lock);
405 
406 err_free:
407 	kfree(t);
408 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
409 
410 	return ret;
411 }
412 
413 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
414 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
415 			     unsigned psn)
416 {
417 	struct ipoib_dev_priv *priv = netdev_priv(dev);
418 	struct ipoib_cm_data data = {};
419 	struct ib_cm_rep_param rep = {};
420 
421 	data.qpn = cpu_to_be32(priv->qp->qp_num);
422 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
423 
424 	rep.private_data = &data;
425 	rep.private_data_len = sizeof data;
426 	rep.flow_control = 0;
427 	rep.rnr_retry_count = req->rnr_retry_count;
428 	rep.srq = ipoib_cm_has_srq(dev);
429 	rep.qp_num = qp->qp_num;
430 	rep.starting_psn = psn;
431 	return ib_send_cm_rep(cm_id, &rep);
432 }
433 
434 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
435 {
436 	struct net_device *dev = cm_id->context;
437 	struct ipoib_dev_priv *priv = netdev_priv(dev);
438 	struct ipoib_cm_rx *p;
439 	unsigned psn;
440 	int ret;
441 
442 	ipoib_dbg(priv, "REQ arrived\n");
443 	p = kzalloc(sizeof *p, GFP_KERNEL);
444 	if (!p)
445 		return -ENOMEM;
446 	p->dev = dev;
447 	p->id = cm_id;
448 	cm_id->context = p;
449 	p->state = IPOIB_CM_RX_LIVE;
450 	p->jiffies = jiffies;
451 	INIT_LIST_HEAD(&p->list);
452 
453 	p->qp = ipoib_cm_create_rx_qp(dev, p);
454 	if (IS_ERR(p->qp)) {
455 		ret = PTR_ERR(p->qp);
456 		goto err_qp;
457 	}
458 
459 	psn = random32() & 0xffffff;
460 	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
461 	if (ret)
462 		goto err_modify;
463 
464 	if (!ipoib_cm_has_srq(dev)) {
465 		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
466 		if (ret)
467 			goto err_modify;
468 	}
469 
470 	spin_lock_irq(&priv->lock);
471 	queue_delayed_work(ipoib_workqueue,
472 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
473 	/* Add this entry to passive ids list head, but do not re-add it
474 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
475 	p->jiffies = jiffies;
476 	if (p->state == IPOIB_CM_RX_LIVE)
477 		list_move(&p->list, &priv->cm.passive_ids);
478 	spin_unlock_irq(&priv->lock);
479 
480 	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
481 	if (ret) {
482 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
483 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
484 			ipoib_warn(priv, "unable to move qp to error state\n");
485 	}
486 	return 0;
487 
488 err_modify:
489 	ib_destroy_qp(p->qp);
490 err_qp:
491 	kfree(p);
492 	return ret;
493 }
494 
495 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
496 			       struct ib_cm_event *event)
497 {
498 	struct ipoib_cm_rx *p;
499 	struct ipoib_dev_priv *priv;
500 
501 	switch (event->event) {
502 	case IB_CM_REQ_RECEIVED:
503 		return ipoib_cm_req_handler(cm_id, event);
504 	case IB_CM_DREQ_RECEIVED:
505 		p = cm_id->context;
506 		ib_send_cm_drep(cm_id, NULL, 0);
507 		/* Fall through */
508 	case IB_CM_REJ_RECEIVED:
509 		p = cm_id->context;
510 		priv = netdev_priv(p->dev);
511 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
512 			ipoib_warn(priv, "unable to move qp to error state\n");
513 		/* Fall through */
514 	default:
515 		return 0;
516 	}
517 }
518 /* Adjust length of skb with fragments to match received data */
519 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
520 			  unsigned int length, struct sk_buff *toskb)
521 {
522 	int i, num_frags;
523 	unsigned int size;
524 
525 	/* put header into skb */
526 	size = min(length, hdr_space);
527 	skb->tail += size;
528 	skb->len += size;
529 	length -= size;
530 
531 	num_frags = skb_shinfo(skb)->nr_frags;
532 	for (i = 0; i < num_frags; i++) {
533 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
534 
535 		if (length == 0) {
536 			/* don't need this page */
537 			skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
538 			--skb_shinfo(skb)->nr_frags;
539 		} else {
540 			size = min(length, (unsigned) PAGE_SIZE);
541 
542 			frag->size = size;
543 			skb->data_len += size;
544 			skb->truesize += size;
545 			skb->len += size;
546 			length -= size;
547 		}
548 	}
549 }
550 
551 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
552 {
553 	struct ipoib_dev_priv *priv = netdev_priv(dev);
554 	struct ipoib_cm_rx_buf *rx_ring;
555 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
556 	struct sk_buff *skb, *newskb;
557 	struct ipoib_cm_rx *p;
558 	unsigned long flags;
559 	u64 mapping[IPOIB_CM_RX_SG];
560 	int frags;
561 	int has_srq;
562 	struct sk_buff *small_skb;
563 
564 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
565 		       wr_id, wc->status);
566 
567 	if (unlikely(wr_id >= ipoib_recvq_size)) {
568 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
569 			spin_lock_irqsave(&priv->lock, flags);
570 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
571 			ipoib_cm_start_rx_drain(priv);
572 			queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
573 			spin_unlock_irqrestore(&priv->lock, flags);
574 		} else
575 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
576 				   wr_id, ipoib_recvq_size);
577 		return;
578 	}
579 
580 	p = wc->qp->qp_context;
581 
582 	has_srq = ipoib_cm_has_srq(dev);
583 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
584 
585 	skb = rx_ring[wr_id].skb;
586 
587 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
588 		ipoib_dbg(priv, "cm recv error "
589 			   "(status=%d, wrid=%d vend_err %x)\n",
590 			   wc->status, wr_id, wc->vendor_err);
591 		++dev->stats.rx_dropped;
592 		if (has_srq)
593 			goto repost;
594 		else {
595 			if (!--p->recv_count) {
596 				spin_lock_irqsave(&priv->lock, flags);
597 				list_move(&p->list, &priv->cm.rx_reap_list);
598 				spin_unlock_irqrestore(&priv->lock, flags);
599 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
600 			}
601 			return;
602 		}
603 	}
604 
605 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
606 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
607 			spin_lock_irqsave(&priv->lock, flags);
608 			p->jiffies = jiffies;
609 			/* Move this entry to list head, but do not re-add it
610 			 * if it has been moved out of list. */
611 			if (p->state == IPOIB_CM_RX_LIVE)
612 				list_move(&p->list, &priv->cm.passive_ids);
613 			spin_unlock_irqrestore(&priv->lock, flags);
614 		}
615 	}
616 
617 	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
618 		int dlen = wc->byte_len;
619 
620 		small_skb = dev_alloc_skb(dlen + 12);
621 		if (small_skb) {
622 			skb_reserve(small_skb, 12);
623 			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
624 						   dlen, DMA_FROM_DEVICE);
625 			skb_copy_from_linear_data(skb, small_skb->data, dlen);
626 			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
627 						      dlen, DMA_FROM_DEVICE);
628 			skb_put(small_skb, dlen);
629 			skb = small_skb;
630 			goto copied;
631 		}
632 	}
633 
634 	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
635 					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
636 
637 	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
638 	if (unlikely(!newskb)) {
639 		/*
640 		 * If we can't allocate a new RX buffer, dump
641 		 * this packet and reuse the old buffer.
642 		 */
643 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
644 		++dev->stats.rx_dropped;
645 		goto repost;
646 	}
647 
648 	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
649 	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
650 
651 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
652 		       wc->byte_len, wc->slid);
653 
654 	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
655 
656 copied:
657 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
658 	skb_reset_mac_header(skb);
659 	skb_pull(skb, IPOIB_ENCAP_LEN);
660 
661 	dev->last_rx = jiffies;
662 	++dev->stats.rx_packets;
663 	dev->stats.rx_bytes += skb->len;
664 
665 	skb->dev = dev;
666 	/* XXX get correct PACKET_ type here */
667 	skb->pkt_type = PACKET_HOST;
668 	netif_receive_skb(skb);
669 
670 repost:
671 	if (has_srq) {
672 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
673 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
674 				   "for buf %d\n", wr_id);
675 	} else {
676 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
677 							  &priv->cm.rx_wr,
678 							  priv->cm.rx_sge,
679 							  wr_id))) {
680 			--p->recv_count;
681 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
682 				   "for buf %d\n", wr_id);
683 		}
684 	}
685 }
686 
687 static inline int post_send(struct ipoib_dev_priv *priv,
688 			    struct ipoib_cm_tx *tx,
689 			    unsigned int wr_id,
690 			    u64 addr, int len)
691 {
692 	struct ib_send_wr *bad_wr;
693 
694 	priv->tx_sge[0].addr          = addr;
695 	priv->tx_sge[0].length        = len;
696 
697 	priv->tx_wr.num_sge	= 1;
698 	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
699 
700 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
701 }
702 
703 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
704 {
705 	struct ipoib_dev_priv *priv = netdev_priv(dev);
706 	struct ipoib_cm_tx_buf *tx_req;
707 	u64 addr;
708 
709 	if (unlikely(skb->len > tx->mtu)) {
710 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
711 			   skb->len, tx->mtu);
712 		++dev->stats.tx_dropped;
713 		++dev->stats.tx_errors;
714 		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
715 		return;
716 	}
717 
718 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
719 		       tx->tx_head, skb->len, tx->qp->qp_num);
720 
721 	/*
722 	 * We put the skb into the tx_ring _before_ we call post_send()
723 	 * because it's entirely possible that the completion handler will
724 	 * run before we execute anything after the post_send().  That
725 	 * means we have to make sure everything is properly recorded and
726 	 * our state is consistent before we call post_send().
727 	 */
728 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
729 	tx_req->skb = skb;
730 	addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
731 	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
732 		++dev->stats.tx_errors;
733 		dev_kfree_skb_any(skb);
734 		return;
735 	}
736 
737 	tx_req->mapping = addr;
738 
739 	if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
740 			       addr, skb->len))) {
741 		ipoib_warn(priv, "post_send failed\n");
742 		++dev->stats.tx_errors;
743 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
744 		dev_kfree_skb_any(skb);
745 	} else {
746 		dev->trans_start = jiffies;
747 		++tx->tx_head;
748 
749 		if (++priv->tx_outstanding == ipoib_sendq_size) {
750 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
751 				  tx->qp->qp_num);
752 			netif_stop_queue(dev);
753 		}
754 	}
755 }
756 
757 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
758 {
759 	struct ipoib_dev_priv *priv = netdev_priv(dev);
760 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
761 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
762 	struct ipoib_cm_tx_buf *tx_req;
763 	unsigned long flags;
764 
765 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
766 		       wr_id, wc->status);
767 
768 	if (unlikely(wr_id >= ipoib_sendq_size)) {
769 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
770 			   wr_id, ipoib_sendq_size);
771 		return;
772 	}
773 
774 	tx_req = &tx->tx_ring[wr_id];
775 
776 	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
777 
778 	/* FIXME: is this right? Shouldn't we only increment on success? */
779 	++dev->stats.tx_packets;
780 	dev->stats.tx_bytes += tx_req->skb->len;
781 
782 	dev_kfree_skb_any(tx_req->skb);
783 
784 	spin_lock_irqsave(&priv->tx_lock, flags);
785 	++tx->tx_tail;
786 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
787 	    netif_queue_stopped(dev) &&
788 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
789 		netif_wake_queue(dev);
790 
791 	if (wc->status != IB_WC_SUCCESS &&
792 	    wc->status != IB_WC_WR_FLUSH_ERR) {
793 		struct ipoib_neigh *neigh;
794 
795 		ipoib_dbg(priv, "failed cm send event "
796 			   "(status=%d, wrid=%d vend_err %x)\n",
797 			   wc->status, wr_id, wc->vendor_err);
798 
799 		spin_lock(&priv->lock);
800 		neigh = tx->neigh;
801 
802 		if (neigh) {
803 			neigh->cm = NULL;
804 			list_del(&neigh->list);
805 			if (neigh->ah)
806 				ipoib_put_ah(neigh->ah);
807 			ipoib_neigh_free(dev, neigh);
808 
809 			tx->neigh = NULL;
810 		}
811 
812 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
813 			list_move(&tx->list, &priv->cm.reap_list);
814 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
815 		}
816 
817 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
818 
819 		spin_unlock(&priv->lock);
820 	}
821 
822 	spin_unlock_irqrestore(&priv->tx_lock, flags);
823 }
824 
825 int ipoib_cm_dev_open(struct net_device *dev)
826 {
827 	struct ipoib_dev_priv *priv = netdev_priv(dev);
828 	int ret;
829 
830 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
831 		return 0;
832 
833 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
834 	if (IS_ERR(priv->cm.id)) {
835 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
836 		ret = PTR_ERR(priv->cm.id);
837 		goto err_cm;
838 	}
839 
840 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
841 			   0, NULL);
842 	if (ret) {
843 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
844 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
845 		goto err_listen;
846 	}
847 
848 	return 0;
849 
850 err_listen:
851 	ib_destroy_cm_id(priv->cm.id);
852 err_cm:
853 	priv->cm.id = NULL;
854 	return ret;
855 }
856 
857 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
858 {
859 	struct ipoib_dev_priv *priv = netdev_priv(dev);
860 	struct ipoib_cm_rx *rx, *n;
861 	LIST_HEAD(list);
862 
863 	spin_lock_irq(&priv->lock);
864 	list_splice_init(&priv->cm.rx_reap_list, &list);
865 	spin_unlock_irq(&priv->lock);
866 
867 	list_for_each_entry_safe(rx, n, &list, list) {
868 		ib_destroy_cm_id(rx->id);
869 		ib_destroy_qp(rx->qp);
870 		if (!ipoib_cm_has_srq(dev)) {
871 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
872 			spin_lock_irq(&priv->lock);
873 			--priv->cm.nonsrq_conn_qp;
874 			spin_unlock_irq(&priv->lock);
875 		}
876 		kfree(rx);
877 	}
878 }
879 
880 void ipoib_cm_dev_stop(struct net_device *dev)
881 {
882 	struct ipoib_dev_priv *priv = netdev_priv(dev);
883 	struct ipoib_cm_rx *p;
884 	unsigned long begin;
885 	int ret;
886 
887 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
888 		return;
889 
890 	ib_destroy_cm_id(priv->cm.id);
891 	priv->cm.id = NULL;
892 
893 	spin_lock_irq(&priv->lock);
894 	while (!list_empty(&priv->cm.passive_ids)) {
895 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
896 		list_move(&p->list, &priv->cm.rx_error_list);
897 		p->state = IPOIB_CM_RX_ERROR;
898 		spin_unlock_irq(&priv->lock);
899 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
900 		if (ret)
901 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
902 		spin_lock_irq(&priv->lock);
903 	}
904 
905 	/* Wait for all RX to be drained */
906 	begin = jiffies;
907 
908 	while (!list_empty(&priv->cm.rx_error_list) ||
909 	       !list_empty(&priv->cm.rx_flush_list) ||
910 	       !list_empty(&priv->cm.rx_drain_list)) {
911 		if (time_after(jiffies, begin + 5 * HZ)) {
912 			ipoib_warn(priv, "RX drain timing out\n");
913 
914 			/*
915 			 * assume the HW is wedged and just free up everything.
916 			 */
917 			list_splice_init(&priv->cm.rx_flush_list,
918 					 &priv->cm.rx_reap_list);
919 			list_splice_init(&priv->cm.rx_error_list,
920 					 &priv->cm.rx_reap_list);
921 			list_splice_init(&priv->cm.rx_drain_list,
922 					 &priv->cm.rx_reap_list);
923 			break;
924 		}
925 		spin_unlock_irq(&priv->lock);
926 		msleep(1);
927 		ipoib_drain_cq(dev);
928 		spin_lock_irq(&priv->lock);
929 	}
930 
931 	spin_unlock_irq(&priv->lock);
932 
933 	ipoib_cm_free_rx_reap_list(dev);
934 
935 	cancel_delayed_work(&priv->cm.stale_task);
936 }
937 
938 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
939 {
940 	struct ipoib_cm_tx *p = cm_id->context;
941 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
942 	struct ipoib_cm_data *data = event->private_data;
943 	struct sk_buff_head skqueue;
944 	struct ib_qp_attr qp_attr;
945 	int qp_attr_mask, ret;
946 	struct sk_buff *skb;
947 
948 	p->mtu = be32_to_cpu(data->mtu);
949 
950 	if (p->mtu <= IPOIB_ENCAP_LEN) {
951 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
952 			   p->mtu, IPOIB_ENCAP_LEN);
953 		return -EINVAL;
954 	}
955 
956 	qp_attr.qp_state = IB_QPS_RTR;
957 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
958 	if (ret) {
959 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
960 		return ret;
961 	}
962 
963 	qp_attr.rq_psn = 0 /* FIXME */;
964 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
965 	if (ret) {
966 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
967 		return ret;
968 	}
969 
970 	qp_attr.qp_state = IB_QPS_RTS;
971 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
972 	if (ret) {
973 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
974 		return ret;
975 	}
976 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
977 	if (ret) {
978 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
979 		return ret;
980 	}
981 
982 	skb_queue_head_init(&skqueue);
983 
984 	spin_lock_irq(&priv->lock);
985 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
986 	if (p->neigh)
987 		while ((skb = __skb_dequeue(&p->neigh->queue)))
988 			__skb_queue_tail(&skqueue, skb);
989 	spin_unlock_irq(&priv->lock);
990 
991 	while ((skb = __skb_dequeue(&skqueue))) {
992 		skb->dev = p->dev;
993 		if (dev_queue_xmit(skb))
994 			ipoib_warn(priv, "dev_queue_xmit failed "
995 				   "to requeue packet\n");
996 	}
997 
998 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
999 	if (ret) {
1000 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1001 		return ret;
1002 	}
1003 	return 0;
1004 }
1005 
1006 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1007 {
1008 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1009 	struct ib_qp_init_attr attr = {
1010 		.send_cq		= priv->recv_cq,
1011 		.recv_cq		= priv->recv_cq,
1012 		.srq			= priv->cm.srq,
1013 		.cap.max_send_wr	= ipoib_sendq_size,
1014 		.cap.max_send_sge	= 1,
1015 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1016 		.qp_type		= IB_QPT_RC,
1017 		.qp_context		= tx
1018 	};
1019 
1020 	return ib_create_qp(priv->pd, &attr);
1021 }
1022 
1023 static int ipoib_cm_send_req(struct net_device *dev,
1024 			     struct ib_cm_id *id, struct ib_qp *qp,
1025 			     u32 qpn,
1026 			     struct ib_sa_path_rec *pathrec)
1027 {
1028 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1029 	struct ipoib_cm_data data = {};
1030 	struct ib_cm_req_param req = {};
1031 
1032 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1033 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1034 
1035 	req.primary_path		= pathrec;
1036 	req.alternate_path		= NULL;
1037 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1038 	req.qp_num			= qp->qp_num;
1039 	req.qp_type			= qp->qp_type;
1040 	req.private_data		= &data;
1041 	req.private_data_len		= sizeof data;
1042 	req.flow_control		= 0;
1043 
1044 	req.starting_psn		= 0; /* FIXME */
1045 
1046 	/*
1047 	 * Pick some arbitrary defaults here; we could make these
1048 	 * module parameters if anyone cared about setting them.
1049 	 */
1050 	req.responder_resources		= 4;
1051 	req.remote_cm_response_timeout	= 20;
1052 	req.local_cm_response_timeout	= 20;
1053 	req.retry_count			= 0; /* RFC draft warns against retries */
1054 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1055 	req.max_cm_retries		= 15;
1056 	req.srq				= ipoib_cm_has_srq(dev);
1057 	return ib_send_cm_req(id, &req);
1058 }
1059 
1060 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1061 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1062 {
1063 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1064 	struct ib_qp_attr qp_attr;
1065 	int qp_attr_mask, ret;
1066 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1067 	if (ret) {
1068 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1069 		return ret;
1070 	}
1071 
1072 	qp_attr.qp_state = IB_QPS_INIT;
1073 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1074 	qp_attr.port_num = priv->port;
1075 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1076 
1077 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1078 	if (ret) {
1079 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1080 		return ret;
1081 	}
1082 	return 0;
1083 }
1084 
1085 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1086 			    struct ib_sa_path_rec *pathrec)
1087 {
1088 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1089 	int ret;
1090 
1091 	p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1092 	if (!p->tx_ring) {
1093 		ipoib_warn(priv, "failed to allocate tx ring\n");
1094 		ret = -ENOMEM;
1095 		goto err_tx;
1096 	}
1097 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1098 
1099 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1100 	if (IS_ERR(p->qp)) {
1101 		ret = PTR_ERR(p->qp);
1102 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1103 		goto err_qp;
1104 	}
1105 
1106 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1107 	if (IS_ERR(p->id)) {
1108 		ret = PTR_ERR(p->id);
1109 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1110 		goto err_id;
1111 	}
1112 
1113 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1114 	if (ret) {
1115 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1116 		goto err_modify;
1117 	}
1118 
1119 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1120 	if (ret) {
1121 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1122 		goto err_send_cm;
1123 	}
1124 
1125 	ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
1126 		  p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
1127 
1128 	return 0;
1129 
1130 err_send_cm:
1131 err_modify:
1132 	ib_destroy_cm_id(p->id);
1133 err_id:
1134 	p->id = NULL;
1135 	ib_destroy_qp(p->qp);
1136 err_qp:
1137 	p->qp = NULL;
1138 	vfree(p->tx_ring);
1139 err_tx:
1140 	return ret;
1141 }
1142 
1143 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1144 {
1145 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1146 	struct ipoib_cm_tx_buf *tx_req;
1147 	unsigned long flags;
1148 	unsigned long begin;
1149 
1150 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1151 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1152 
1153 	if (p->id)
1154 		ib_destroy_cm_id(p->id);
1155 
1156 	if (p->tx_ring) {
1157 		/* Wait for all sends to complete */
1158 		begin = jiffies;
1159 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1160 			if (time_after(jiffies, begin + 5 * HZ)) {
1161 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1162 					   p->tx_head - p->tx_tail);
1163 				goto timeout;
1164 			}
1165 
1166 			msleep(1);
1167 		}
1168 	}
1169 
1170 timeout:
1171 
1172 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1173 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1174 		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1175 				    DMA_TO_DEVICE);
1176 		dev_kfree_skb_any(tx_req->skb);
1177 		++p->tx_tail;
1178 		spin_lock_irqsave(&priv->tx_lock, flags);
1179 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1180 		    netif_queue_stopped(p->dev) &&
1181 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1182 			netif_wake_queue(p->dev);
1183 		spin_unlock_irqrestore(&priv->tx_lock, flags);
1184 	}
1185 
1186 	if (p->qp)
1187 		ib_destroy_qp(p->qp);
1188 
1189 	vfree(p->tx_ring);
1190 	kfree(p);
1191 }
1192 
1193 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1194 			       struct ib_cm_event *event)
1195 {
1196 	struct ipoib_cm_tx *tx = cm_id->context;
1197 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1198 	struct net_device *dev = priv->dev;
1199 	struct ipoib_neigh *neigh;
1200 	int ret;
1201 
1202 	switch (event->event) {
1203 	case IB_CM_DREQ_RECEIVED:
1204 		ipoib_dbg(priv, "DREQ received.\n");
1205 		ib_send_cm_drep(cm_id, NULL, 0);
1206 		break;
1207 	case IB_CM_REP_RECEIVED:
1208 		ipoib_dbg(priv, "REP received.\n");
1209 		ret = ipoib_cm_rep_handler(cm_id, event);
1210 		if (ret)
1211 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1212 				       NULL, 0, NULL, 0);
1213 		break;
1214 	case IB_CM_REQ_ERROR:
1215 	case IB_CM_REJ_RECEIVED:
1216 	case IB_CM_TIMEWAIT_EXIT:
1217 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1218 		spin_lock_irq(&priv->tx_lock);
1219 		spin_lock(&priv->lock);
1220 		neigh = tx->neigh;
1221 
1222 		if (neigh) {
1223 			neigh->cm = NULL;
1224 			list_del(&neigh->list);
1225 			if (neigh->ah)
1226 				ipoib_put_ah(neigh->ah);
1227 			ipoib_neigh_free(dev, neigh);
1228 
1229 			tx->neigh = NULL;
1230 		}
1231 
1232 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1233 			list_move(&tx->list, &priv->cm.reap_list);
1234 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1235 		}
1236 
1237 		spin_unlock(&priv->lock);
1238 		spin_unlock_irq(&priv->tx_lock);
1239 		break;
1240 	default:
1241 		break;
1242 	}
1243 
1244 	return 0;
1245 }
1246 
1247 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1248 				       struct ipoib_neigh *neigh)
1249 {
1250 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1251 	struct ipoib_cm_tx *tx;
1252 
1253 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1254 	if (!tx)
1255 		return NULL;
1256 
1257 	neigh->cm = tx;
1258 	tx->neigh = neigh;
1259 	tx->path = path;
1260 	tx->dev = dev;
1261 	list_add(&tx->list, &priv->cm.start_list);
1262 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1263 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1264 	return tx;
1265 }
1266 
1267 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1268 {
1269 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1270 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1271 		list_move(&tx->list, &priv->cm.reap_list);
1272 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1273 		ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1274 			  IPOIB_GID_ARG(tx->neigh->dgid));
1275 		tx->neigh = NULL;
1276 	}
1277 }
1278 
1279 static void ipoib_cm_tx_start(struct work_struct *work)
1280 {
1281 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1282 						   cm.start_task);
1283 	struct net_device *dev = priv->dev;
1284 	struct ipoib_neigh *neigh;
1285 	struct ipoib_cm_tx *p;
1286 	unsigned long flags;
1287 	int ret;
1288 
1289 	struct ib_sa_path_rec pathrec;
1290 	u32 qpn;
1291 
1292 	spin_lock_irqsave(&priv->tx_lock, flags);
1293 	spin_lock(&priv->lock);
1294 	while (!list_empty(&priv->cm.start_list)) {
1295 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1296 		list_del_init(&p->list);
1297 		neigh = p->neigh;
1298 		qpn = IPOIB_QPN(neigh->neighbour->ha);
1299 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1300 		spin_unlock(&priv->lock);
1301 		spin_unlock_irqrestore(&priv->tx_lock, flags);
1302 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1303 		spin_lock_irqsave(&priv->tx_lock, flags);
1304 		spin_lock(&priv->lock);
1305 		if (ret) {
1306 			neigh = p->neigh;
1307 			if (neigh) {
1308 				neigh->cm = NULL;
1309 				list_del(&neigh->list);
1310 				if (neigh->ah)
1311 					ipoib_put_ah(neigh->ah);
1312 				ipoib_neigh_free(dev, neigh);
1313 			}
1314 			list_del(&p->list);
1315 			kfree(p);
1316 		}
1317 	}
1318 	spin_unlock(&priv->lock);
1319 	spin_unlock_irqrestore(&priv->tx_lock, flags);
1320 }
1321 
1322 static void ipoib_cm_tx_reap(struct work_struct *work)
1323 {
1324 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1325 						   cm.reap_task);
1326 	struct ipoib_cm_tx *p;
1327 
1328 	spin_lock_irq(&priv->tx_lock);
1329 	spin_lock(&priv->lock);
1330 	while (!list_empty(&priv->cm.reap_list)) {
1331 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1332 		list_del(&p->list);
1333 		spin_unlock(&priv->lock);
1334 		spin_unlock_irq(&priv->tx_lock);
1335 		ipoib_cm_tx_destroy(p);
1336 		spin_lock_irq(&priv->tx_lock);
1337 		spin_lock(&priv->lock);
1338 	}
1339 	spin_unlock(&priv->lock);
1340 	spin_unlock_irq(&priv->tx_lock);
1341 }
1342 
1343 static void ipoib_cm_skb_reap(struct work_struct *work)
1344 {
1345 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1346 						   cm.skb_task);
1347 	struct sk_buff *skb;
1348 
1349 	unsigned mtu = priv->mcast_mtu;
1350 
1351 	spin_lock_irq(&priv->tx_lock);
1352 	spin_lock(&priv->lock);
1353 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1354 		spin_unlock(&priv->lock);
1355 		spin_unlock_irq(&priv->tx_lock);
1356 		if (skb->protocol == htons(ETH_P_IP))
1357 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1358 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1359 		else if (skb->protocol == htons(ETH_P_IPV6))
1360 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1361 #endif
1362 		dev_kfree_skb_any(skb);
1363 		spin_lock_irq(&priv->tx_lock);
1364 		spin_lock(&priv->lock);
1365 	}
1366 	spin_unlock(&priv->lock);
1367 	spin_unlock_irq(&priv->tx_lock);
1368 }
1369 
1370 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1371 			   unsigned int mtu)
1372 {
1373 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1374 	int e = skb_queue_empty(&priv->cm.skb_queue);
1375 
1376 	if (skb->dst)
1377 		skb->dst->ops->update_pmtu(skb->dst, mtu);
1378 
1379 	skb_queue_tail(&priv->cm.skb_queue, skb);
1380 	if (e)
1381 		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1382 }
1383 
1384 static void ipoib_cm_rx_reap(struct work_struct *work)
1385 {
1386 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1387 						cm.rx_reap_task)->dev);
1388 }
1389 
1390 static void ipoib_cm_stale_task(struct work_struct *work)
1391 {
1392 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1393 						   cm.stale_task.work);
1394 	struct ipoib_cm_rx *p;
1395 	int ret;
1396 
1397 	spin_lock_irq(&priv->lock);
1398 	while (!list_empty(&priv->cm.passive_ids)) {
1399 		/* List is sorted by LRU, start from tail,
1400 		 * stop when we see a recently used entry */
1401 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1402 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1403 			break;
1404 		list_move(&p->list, &priv->cm.rx_error_list);
1405 		p->state = IPOIB_CM_RX_ERROR;
1406 		spin_unlock_irq(&priv->lock);
1407 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1408 		if (ret)
1409 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1410 		spin_lock_irq(&priv->lock);
1411 	}
1412 
1413 	if (!list_empty(&priv->cm.passive_ids))
1414 		queue_delayed_work(ipoib_workqueue,
1415 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1416 	spin_unlock_irq(&priv->lock);
1417 }
1418 
1419 
1420 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1421 			 char *buf)
1422 {
1423 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1424 
1425 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1426 		return sprintf(buf, "connected\n");
1427 	else
1428 		return sprintf(buf, "datagram\n");
1429 }
1430 
1431 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1432 			const char *buf, size_t count)
1433 {
1434 	struct net_device *dev = to_net_dev(d);
1435 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1436 
1437 	/* flush paths if we switch modes so that connections are restarted */
1438 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1439 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1440 		ipoib_warn(priv, "enabling connected mode "
1441 			   "will cause multicast packet drops\n");
1442 
1443 		rtnl_lock();
1444 		dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1445 		rtnl_unlock();
1446 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1447 
1448 		ipoib_flush_paths(dev);
1449 		return count;
1450 	}
1451 
1452 	if (!strcmp(buf, "datagram\n")) {
1453 		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1454 
1455 		rtnl_lock();
1456 		if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1457 			dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1458 			if (priv->hca_caps & IB_DEVICE_UD_TSO)
1459 				dev->features |= NETIF_F_TSO;
1460 		}
1461 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1462 		rtnl_unlock();
1463 		ipoib_flush_paths(dev);
1464 
1465 		return count;
1466 	}
1467 
1468 	return -EINVAL;
1469 }
1470 
1471 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1472 
1473 int ipoib_cm_add_mode_attr(struct net_device *dev)
1474 {
1475 	return device_create_file(&dev->dev, &dev_attr_mode);
1476 }
1477 
1478 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1479 {
1480 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1481 	struct ib_srq_init_attr srq_init_attr = {
1482 		.attr = {
1483 			.max_wr  = ipoib_recvq_size,
1484 			.max_sge = max_sge
1485 		}
1486 	};
1487 
1488 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1489 	if (IS_ERR(priv->cm.srq)) {
1490 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1491 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1492 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1493 		priv->cm.srq = NULL;
1494 		return;
1495 	}
1496 
1497 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1498 				    GFP_KERNEL);
1499 	if (!priv->cm.srq_ring) {
1500 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1501 		       priv->ca->name, ipoib_recvq_size);
1502 		ib_destroy_srq(priv->cm.srq);
1503 		priv->cm.srq = NULL;
1504 	}
1505 }
1506 
1507 int ipoib_cm_dev_init(struct net_device *dev)
1508 {
1509 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1510 	int i, ret;
1511 	struct ib_device_attr attr;
1512 
1513 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1514 	INIT_LIST_HEAD(&priv->cm.reap_list);
1515 	INIT_LIST_HEAD(&priv->cm.start_list);
1516 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1517 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1518 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1519 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1520 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1521 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1522 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1523 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1524 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1525 
1526 	skb_queue_head_init(&priv->cm.skb_queue);
1527 
1528 	ret = ib_query_device(priv->ca, &attr);
1529 	if (ret) {
1530 		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1531 		return ret;
1532 	}
1533 
1534 	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1535 
1536 	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1537 	ipoib_cm_create_srq(dev, attr.max_srq_sge);
1538 	if (ipoib_cm_has_srq(dev)) {
1539 		priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1540 		priv->cm.num_frags  = attr.max_srq_sge;
1541 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1542 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1543 	} else {
1544 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1545 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1546 	}
1547 
1548 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1549 
1550 	if (ipoib_cm_has_srq(dev)) {
1551 		for (i = 0; i < ipoib_recvq_size; ++i) {
1552 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1553 						   priv->cm.num_frags - 1,
1554 						   priv->cm.srq_ring[i].mapping)) {
1555 				ipoib_warn(priv, "failed to allocate "
1556 					   "receive buffer %d\n", i);
1557 				ipoib_cm_dev_cleanup(dev);
1558 				return -ENOMEM;
1559 			}
1560 
1561 			if (ipoib_cm_post_receive_srq(dev, i)) {
1562 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1563 					   "failed for buf %d\n", i);
1564 				ipoib_cm_dev_cleanup(dev);
1565 				return -EIO;
1566 			}
1567 		}
1568 	}
1569 
1570 	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1571 	return 0;
1572 }
1573 
1574 void ipoib_cm_dev_cleanup(struct net_device *dev)
1575 {
1576 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1577 	int ret;
1578 
1579 	if (!priv->cm.srq)
1580 		return;
1581 
1582 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1583 
1584 	ret = ib_destroy_srq(priv->cm.srq);
1585 	if (ret)
1586 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1587 
1588 	priv->cm.srq = NULL;
1589 	if (!priv->cm.srq_ring)
1590 		return;
1591 
1592 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1593 	priv->cm.srq_ring = NULL;
1594 }
1595