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 <net/dst.h>
35 #include <net/icmp.h>
36 #include <linux/icmpv6.h>
37 #include <linux/delay.h>
38 #include <linux/slab.h>
39 #include <linux/vmalloc.h>
40 #include <linux/moduleparam.h>
41 
42 #include "ipoib.h"
43 
44 int ipoib_max_conn_qp = 128;
45 
46 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
47 MODULE_PARM_DESC(max_nonsrq_conn_qp,
48 		 "Max number of connected-mode QPs per interface "
49 		 "(applied only if shared receive queue is not available)");
50 
51 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
52 static int data_debug_level;
53 
54 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
55 MODULE_PARM_DESC(cm_data_debug_level,
56 		 "Enable data path debug tracing for connected mode if > 0");
57 #endif
58 
59 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
60 
61 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
62 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
63 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
64 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
65 
66 static struct ib_qp_attr ipoib_cm_err_attr = {
67 	.qp_state = IB_QPS_ERR
68 };
69 
70 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
71 
72 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
73 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
74 	.opcode = IB_WR_SEND,
75 };
76 
77 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
78 			       struct ib_cm_event *event);
79 
80 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
81 				  u64 mapping[IPOIB_CM_RX_SG])
82 {
83 	int i;
84 
85 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
86 
87 	for (i = 0; i < frags; ++i)
88 		ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
89 }
90 
91 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
92 {
93 	struct ipoib_dev_priv *priv = netdev_priv(dev);
94 	struct ib_recv_wr *bad_wr;
95 	int i, ret;
96 
97 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
98 
99 	for (i = 0; i < priv->cm.num_frags; ++i)
100 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
101 
102 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
103 	if (unlikely(ret)) {
104 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
105 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
106 				      priv->cm.srq_ring[id].mapping);
107 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
108 		priv->cm.srq_ring[id].skb = NULL;
109 	}
110 
111 	return ret;
112 }
113 
114 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
115 					struct ipoib_cm_rx *rx,
116 					struct ib_recv_wr *wr,
117 					struct ib_sge *sge, int id)
118 {
119 	struct ipoib_dev_priv *priv = netdev_priv(dev);
120 	struct ib_recv_wr *bad_wr;
121 	int i, ret;
122 
123 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
124 
125 	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
126 		sge[i].addr = rx->rx_ring[id].mapping[i];
127 
128 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
129 	if (unlikely(ret)) {
130 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
131 		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
132 				      rx->rx_ring[id].mapping);
133 		dev_kfree_skb_any(rx->rx_ring[id].skb);
134 		rx->rx_ring[id].skb = NULL;
135 	}
136 
137 	return ret;
138 }
139 
140 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
141 					     struct ipoib_cm_rx_buf *rx_ring,
142 					     int id, int frags,
143 					     u64 mapping[IPOIB_CM_RX_SG])
144 {
145 	struct ipoib_dev_priv *priv = netdev_priv(dev);
146 	struct sk_buff *skb;
147 	int i;
148 
149 	skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
150 	if (unlikely(!skb))
151 		return NULL;
152 
153 	/*
154 	 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
155 	 * IP header to a multiple of 16.
156 	 */
157 	skb_reserve(skb, 12);
158 
159 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
160 				       DMA_FROM_DEVICE);
161 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
162 		dev_kfree_skb_any(skb);
163 		return NULL;
164 	}
165 
166 	for (i = 0; i < frags; i++) {
167 		struct page *page = alloc_page(GFP_ATOMIC);
168 
169 		if (!page)
170 			goto partial_error;
171 		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
172 
173 		mapping[i + 1] = ib_dma_map_page(priv->ca, page,
174 						 0, PAGE_SIZE, DMA_FROM_DEVICE);
175 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
176 			goto partial_error;
177 	}
178 
179 	rx_ring[id].skb = skb;
180 	return skb;
181 
182 partial_error:
183 
184 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
185 
186 	for (; i > 0; --i)
187 		ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
188 
189 	dev_kfree_skb_any(skb);
190 	return NULL;
191 }
192 
193 static void ipoib_cm_free_rx_ring(struct net_device *dev,
194 				  struct ipoib_cm_rx_buf *rx_ring)
195 {
196 	struct ipoib_dev_priv *priv = netdev_priv(dev);
197 	int i;
198 
199 	for (i = 0; i < ipoib_recvq_size; ++i)
200 		if (rx_ring[i].skb) {
201 			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
202 					      rx_ring[i].mapping);
203 			dev_kfree_skb_any(rx_ring[i].skb);
204 		}
205 
206 	vfree(rx_ring);
207 }
208 
209 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
210 {
211 	struct ib_send_wr *bad_wr;
212 	struct ipoib_cm_rx *p;
213 
214 	/* We only reserved 1 extra slot in CQ for drain WRs, so
215 	 * make sure we have at most 1 outstanding WR. */
216 	if (list_empty(&priv->cm.rx_flush_list) ||
217 	    !list_empty(&priv->cm.rx_drain_list))
218 		return;
219 
220 	/*
221 	 * QPs on flush list are error state.  This way, a "flush
222 	 * error" WC will be immediately generated for each WR we post.
223 	 */
224 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
225 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
226 		ipoib_warn(priv, "failed to post drain wr\n");
227 
228 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
229 }
230 
231 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
232 {
233 	struct ipoib_cm_rx *p = ctx;
234 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
235 	unsigned long flags;
236 
237 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
238 		return;
239 
240 	spin_lock_irqsave(&priv->lock, flags);
241 	list_move(&p->list, &priv->cm.rx_flush_list);
242 	p->state = IPOIB_CM_RX_FLUSH;
243 	ipoib_cm_start_rx_drain(priv);
244 	spin_unlock_irqrestore(&priv->lock, flags);
245 }
246 
247 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
248 					   struct ipoib_cm_rx *p)
249 {
250 	struct ipoib_dev_priv *priv = netdev_priv(dev);
251 	struct ib_qp_init_attr attr = {
252 		.event_handler = ipoib_cm_rx_event_handler,
253 		.send_cq = priv->recv_cq, /* For drain WR */
254 		.recv_cq = priv->recv_cq,
255 		.srq = priv->cm.srq,
256 		.cap.max_send_wr = 1, /* For drain WR */
257 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
258 		.sq_sig_type = IB_SIGNAL_ALL_WR,
259 		.qp_type = IB_QPT_RC,
260 		.qp_context = p,
261 	};
262 
263 	if (!ipoib_cm_has_srq(dev)) {
264 		attr.cap.max_recv_wr  = ipoib_recvq_size;
265 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
266 	}
267 
268 	return ib_create_qp(priv->pd, &attr);
269 }
270 
271 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
272 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
273 				 unsigned psn)
274 {
275 	struct ipoib_dev_priv *priv = netdev_priv(dev);
276 	struct ib_qp_attr qp_attr;
277 	int qp_attr_mask, ret;
278 
279 	qp_attr.qp_state = IB_QPS_INIT;
280 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
281 	if (ret) {
282 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
283 		return ret;
284 	}
285 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
286 	if (ret) {
287 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
288 		return ret;
289 	}
290 	qp_attr.qp_state = IB_QPS_RTR;
291 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
292 	if (ret) {
293 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
294 		return ret;
295 	}
296 	qp_attr.rq_psn = psn;
297 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
298 	if (ret) {
299 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
300 		return ret;
301 	}
302 
303 	/*
304 	 * Current Mellanox HCA firmware won't generate completions
305 	 * with error for drain WRs unless the QP has been moved to
306 	 * RTS first. This work-around leaves a window where a QP has
307 	 * moved to error asynchronously, but this will eventually get
308 	 * fixed in firmware, so let's not error out if modify QP
309 	 * fails.
310 	 */
311 	qp_attr.qp_state = IB_QPS_RTS;
312 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
313 	if (ret) {
314 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
315 		return 0;
316 	}
317 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
318 	if (ret) {
319 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
320 		return 0;
321 	}
322 
323 	return 0;
324 }
325 
326 static void ipoib_cm_init_rx_wr(struct net_device *dev,
327 				struct ib_recv_wr *wr,
328 				struct ib_sge *sge)
329 {
330 	struct ipoib_dev_priv *priv = netdev_priv(dev);
331 	int i;
332 
333 	for (i = 0; i < priv->cm.num_frags; ++i)
334 		sge[i].lkey = priv->mr->lkey;
335 
336 	sge[0].length = IPOIB_CM_HEAD_SIZE;
337 	for (i = 1; i < priv->cm.num_frags; ++i)
338 		sge[i].length = PAGE_SIZE;
339 
340 	wr->next    = NULL;
341 	wr->sg_list = sge;
342 	wr->num_sge = priv->cm.num_frags;
343 }
344 
345 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
346 				   struct ipoib_cm_rx *rx)
347 {
348 	struct ipoib_dev_priv *priv = netdev_priv(dev);
349 	struct {
350 		struct ib_recv_wr wr;
351 		struct ib_sge sge[IPOIB_CM_RX_SG];
352 	} *t;
353 	int ret;
354 	int i;
355 
356 	rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
357 	if (!rx->rx_ring) {
358 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
359 		       priv->ca->name, ipoib_recvq_size);
360 		return -ENOMEM;
361 	}
362 
363 	t = kmalloc(sizeof *t, GFP_KERNEL);
364 	if (!t) {
365 		ret = -ENOMEM;
366 		goto err_free;
367 	}
368 
369 	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
370 
371 	spin_lock_irq(&priv->lock);
372 
373 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
374 		spin_unlock_irq(&priv->lock);
375 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
376 		ret = -EINVAL;
377 		goto err_free;
378 	} else
379 		++priv->cm.nonsrq_conn_qp;
380 
381 	spin_unlock_irq(&priv->lock);
382 
383 	for (i = 0; i < ipoib_recvq_size; ++i) {
384 		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
385 					   rx->rx_ring[i].mapping)) {
386 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
387 				ret = -ENOMEM;
388 				goto err_count;
389 		}
390 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
391 		if (ret) {
392 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
393 				   "failed for buf %d\n", i);
394 			ret = -EIO;
395 			goto err_count;
396 		}
397 	}
398 
399 	rx->recv_count = ipoib_recvq_size;
400 
401 	kfree(t);
402 
403 	return 0;
404 
405 err_count:
406 	spin_lock_irq(&priv->lock);
407 	--priv->cm.nonsrq_conn_qp;
408 	spin_unlock_irq(&priv->lock);
409 
410 err_free:
411 	kfree(t);
412 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
413 
414 	return ret;
415 }
416 
417 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
418 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
419 			     unsigned psn)
420 {
421 	struct ipoib_dev_priv *priv = netdev_priv(dev);
422 	struct ipoib_cm_data data = {};
423 	struct ib_cm_rep_param rep = {};
424 
425 	data.qpn = cpu_to_be32(priv->qp->qp_num);
426 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
427 
428 	rep.private_data = &data;
429 	rep.private_data_len = sizeof data;
430 	rep.flow_control = 0;
431 	rep.rnr_retry_count = req->rnr_retry_count;
432 	rep.srq = ipoib_cm_has_srq(dev);
433 	rep.qp_num = qp->qp_num;
434 	rep.starting_psn = psn;
435 	return ib_send_cm_rep(cm_id, &rep);
436 }
437 
438 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
439 {
440 	struct net_device *dev = cm_id->context;
441 	struct ipoib_dev_priv *priv = netdev_priv(dev);
442 	struct ipoib_cm_rx *p;
443 	unsigned psn;
444 	int ret;
445 
446 	ipoib_dbg(priv, "REQ arrived\n");
447 	p = kzalloc(sizeof *p, GFP_KERNEL);
448 	if (!p)
449 		return -ENOMEM;
450 	p->dev = dev;
451 	p->id = cm_id;
452 	cm_id->context = p;
453 	p->state = IPOIB_CM_RX_LIVE;
454 	p->jiffies = jiffies;
455 	INIT_LIST_HEAD(&p->list);
456 
457 	p->qp = ipoib_cm_create_rx_qp(dev, p);
458 	if (IS_ERR(p->qp)) {
459 		ret = PTR_ERR(p->qp);
460 		goto err_qp;
461 	}
462 
463 	psn = random32() & 0xffffff;
464 	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
465 	if (ret)
466 		goto err_modify;
467 
468 	if (!ipoib_cm_has_srq(dev)) {
469 		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
470 		if (ret)
471 			goto err_modify;
472 	}
473 
474 	spin_lock_irq(&priv->lock);
475 	queue_delayed_work(ipoib_workqueue,
476 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
477 	/* Add this entry to passive ids list head, but do not re-add it
478 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
479 	p->jiffies = jiffies;
480 	if (p->state == IPOIB_CM_RX_LIVE)
481 		list_move(&p->list, &priv->cm.passive_ids);
482 	spin_unlock_irq(&priv->lock);
483 
484 	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
485 	if (ret) {
486 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
487 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
488 			ipoib_warn(priv, "unable to move qp to error state\n");
489 	}
490 	return 0;
491 
492 err_modify:
493 	ib_destroy_qp(p->qp);
494 err_qp:
495 	kfree(p);
496 	return ret;
497 }
498 
499 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
500 			       struct ib_cm_event *event)
501 {
502 	struct ipoib_cm_rx *p;
503 	struct ipoib_dev_priv *priv;
504 
505 	switch (event->event) {
506 	case IB_CM_REQ_RECEIVED:
507 		return ipoib_cm_req_handler(cm_id, event);
508 	case IB_CM_DREQ_RECEIVED:
509 		p = cm_id->context;
510 		ib_send_cm_drep(cm_id, NULL, 0);
511 		/* Fall through */
512 	case IB_CM_REJ_RECEIVED:
513 		p = cm_id->context;
514 		priv = netdev_priv(p->dev);
515 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
516 			ipoib_warn(priv, "unable to move qp to error state\n");
517 		/* Fall through */
518 	default:
519 		return 0;
520 	}
521 }
522 /* Adjust length of skb with fragments to match received data */
523 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
524 			  unsigned int length, struct sk_buff *toskb)
525 {
526 	int i, num_frags;
527 	unsigned int size;
528 
529 	/* put header into skb */
530 	size = min(length, hdr_space);
531 	skb->tail += size;
532 	skb->len += size;
533 	length -= size;
534 
535 	num_frags = skb_shinfo(skb)->nr_frags;
536 	for (i = 0; i < num_frags; i++) {
537 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
538 
539 		if (length == 0) {
540 			/* don't need this page */
541 			skb_fill_page_desc(toskb, i, skb_frag_page(frag),
542 					   0, PAGE_SIZE);
543 			--skb_shinfo(skb)->nr_frags;
544 		} else {
545 			size = min(length, (unsigned) PAGE_SIZE);
546 
547 			skb_frag_size_set(frag, size);
548 			skb->data_len += size;
549 			skb->truesize += size;
550 			skb->len += size;
551 			length -= size;
552 		}
553 	}
554 }
555 
556 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
557 {
558 	struct ipoib_dev_priv *priv = netdev_priv(dev);
559 	struct ipoib_cm_rx_buf *rx_ring;
560 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
561 	struct sk_buff *skb, *newskb;
562 	struct ipoib_cm_rx *p;
563 	unsigned long flags;
564 	u64 mapping[IPOIB_CM_RX_SG];
565 	int frags;
566 	int has_srq;
567 	struct sk_buff *small_skb;
568 
569 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
570 		       wr_id, wc->status);
571 
572 	if (unlikely(wr_id >= ipoib_recvq_size)) {
573 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
574 			spin_lock_irqsave(&priv->lock, flags);
575 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
576 			ipoib_cm_start_rx_drain(priv);
577 			queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
578 			spin_unlock_irqrestore(&priv->lock, flags);
579 		} else
580 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
581 				   wr_id, ipoib_recvq_size);
582 		return;
583 	}
584 
585 	p = wc->qp->qp_context;
586 
587 	has_srq = ipoib_cm_has_srq(dev);
588 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
589 
590 	skb = rx_ring[wr_id].skb;
591 
592 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
593 		ipoib_dbg(priv, "cm recv error "
594 			   "(status=%d, wrid=%d vend_err %x)\n",
595 			   wc->status, wr_id, wc->vendor_err);
596 		++dev->stats.rx_dropped;
597 		if (has_srq)
598 			goto repost;
599 		else {
600 			if (!--p->recv_count) {
601 				spin_lock_irqsave(&priv->lock, flags);
602 				list_move(&p->list, &priv->cm.rx_reap_list);
603 				spin_unlock_irqrestore(&priv->lock, flags);
604 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
605 			}
606 			return;
607 		}
608 	}
609 
610 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
611 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
612 			spin_lock_irqsave(&priv->lock, flags);
613 			p->jiffies = jiffies;
614 			/* Move this entry to list head, but do not re-add it
615 			 * if it has been moved out of list. */
616 			if (p->state == IPOIB_CM_RX_LIVE)
617 				list_move(&p->list, &priv->cm.passive_ids);
618 			spin_unlock_irqrestore(&priv->lock, flags);
619 		}
620 	}
621 
622 	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
623 		int dlen = wc->byte_len;
624 
625 		small_skb = dev_alloc_skb(dlen + 12);
626 		if (small_skb) {
627 			skb_reserve(small_skb, 12);
628 			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
629 						   dlen, DMA_FROM_DEVICE);
630 			skb_copy_from_linear_data(skb, small_skb->data, dlen);
631 			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
632 						      dlen, DMA_FROM_DEVICE);
633 			skb_put(small_skb, dlen);
634 			skb = small_skb;
635 			goto copied;
636 		}
637 	}
638 
639 	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
640 					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
641 
642 	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
643 	if (unlikely(!newskb)) {
644 		/*
645 		 * If we can't allocate a new RX buffer, dump
646 		 * this packet and reuse the old buffer.
647 		 */
648 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
649 		++dev->stats.rx_dropped;
650 		goto repost;
651 	}
652 
653 	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
654 	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
655 
656 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
657 		       wc->byte_len, wc->slid);
658 
659 	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
660 
661 copied:
662 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
663 	skb_reset_mac_header(skb);
664 	skb_pull(skb, IPOIB_ENCAP_LEN);
665 
666 	++dev->stats.rx_packets;
667 	dev->stats.rx_bytes += skb->len;
668 
669 	skb->dev = dev;
670 	/* XXX get correct PACKET_ type here */
671 	skb->pkt_type = PACKET_HOST;
672 	netif_receive_skb(skb);
673 
674 repost:
675 	if (has_srq) {
676 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
677 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
678 				   "for buf %d\n", wr_id);
679 	} else {
680 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
681 							  &priv->cm.rx_wr,
682 							  priv->cm.rx_sge,
683 							  wr_id))) {
684 			--p->recv_count;
685 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
686 				   "for buf %d\n", wr_id);
687 		}
688 	}
689 }
690 
691 static inline int post_send(struct ipoib_dev_priv *priv,
692 			    struct ipoib_cm_tx *tx,
693 			    unsigned int wr_id,
694 			    u64 addr, int len)
695 {
696 	struct ib_send_wr *bad_wr;
697 
698 	priv->tx_sge[0].addr          = addr;
699 	priv->tx_sge[0].length        = len;
700 
701 	priv->tx_wr.num_sge	= 1;
702 	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
703 
704 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
705 }
706 
707 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
708 {
709 	struct ipoib_dev_priv *priv = netdev_priv(dev);
710 	struct ipoib_cm_tx_buf *tx_req;
711 	u64 addr;
712 	int rc;
713 
714 	if (unlikely(skb->len > tx->mtu)) {
715 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
716 			   skb->len, tx->mtu);
717 		++dev->stats.tx_dropped;
718 		++dev->stats.tx_errors;
719 		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
720 		return;
721 	}
722 
723 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
724 		       tx->tx_head, skb->len, tx->qp->qp_num);
725 
726 	/*
727 	 * We put the skb into the tx_ring _before_ we call post_send()
728 	 * because it's entirely possible that the completion handler will
729 	 * run before we execute anything after the post_send().  That
730 	 * means we have to make sure everything is properly recorded and
731 	 * our state is consistent before we call post_send().
732 	 */
733 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
734 	tx_req->skb = skb;
735 	addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
736 	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
737 		++dev->stats.tx_errors;
738 		dev_kfree_skb_any(skb);
739 		return;
740 	}
741 
742 	tx_req->mapping = addr;
743 
744 	skb_orphan(skb);
745 	skb_dst_drop(skb);
746 
747 	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
748 		       addr, skb->len);
749 	if (unlikely(rc)) {
750 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
751 		++dev->stats.tx_errors;
752 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
753 		dev_kfree_skb_any(skb);
754 	} else {
755 		dev->trans_start = jiffies;
756 		++tx->tx_head;
757 
758 		if (++priv->tx_outstanding == ipoib_sendq_size) {
759 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
760 				  tx->qp->qp_num);
761 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
762 				ipoib_warn(priv, "request notify on send CQ failed\n");
763 			netif_stop_queue(dev);
764 		}
765 	}
766 }
767 
768 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
769 {
770 	struct ipoib_dev_priv *priv = netdev_priv(dev);
771 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
772 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
773 	struct ipoib_cm_tx_buf *tx_req;
774 	unsigned long flags;
775 
776 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
777 		       wr_id, wc->status);
778 
779 	if (unlikely(wr_id >= ipoib_sendq_size)) {
780 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
781 			   wr_id, ipoib_sendq_size);
782 		return;
783 	}
784 
785 	tx_req = &tx->tx_ring[wr_id];
786 
787 	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
788 
789 	/* FIXME: is this right? Shouldn't we only increment on success? */
790 	++dev->stats.tx_packets;
791 	dev->stats.tx_bytes += tx_req->skb->len;
792 
793 	dev_kfree_skb_any(tx_req->skb);
794 
795 	netif_tx_lock(dev);
796 
797 	++tx->tx_tail;
798 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
799 	    netif_queue_stopped(dev) &&
800 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
801 		netif_wake_queue(dev);
802 
803 	if (wc->status != IB_WC_SUCCESS &&
804 	    wc->status != IB_WC_WR_FLUSH_ERR) {
805 		struct ipoib_neigh *neigh;
806 
807 		ipoib_dbg(priv, "failed cm send event "
808 			   "(status=%d, wrid=%d vend_err %x)\n",
809 			   wc->status, wr_id, wc->vendor_err);
810 
811 		spin_lock_irqsave(&priv->lock, flags);
812 		neigh = tx->neigh;
813 
814 		if (neigh) {
815 			neigh->cm = NULL;
816 			list_del(&neigh->list);
817 			ipoib_neigh_free(neigh);
818 
819 			tx->neigh = NULL;
820 		}
821 
822 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
823 			list_move(&tx->list, &priv->cm.reap_list);
824 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
825 		}
826 
827 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
828 
829 		spin_unlock_irqrestore(&priv->lock, flags);
830 	}
831 
832 	netif_tx_unlock(dev);
833 }
834 
835 int ipoib_cm_dev_open(struct net_device *dev)
836 {
837 	struct ipoib_dev_priv *priv = netdev_priv(dev);
838 	int ret;
839 
840 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
841 		return 0;
842 
843 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
844 	if (IS_ERR(priv->cm.id)) {
845 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
846 		ret = PTR_ERR(priv->cm.id);
847 		goto err_cm;
848 	}
849 
850 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
851 			   0, NULL);
852 	if (ret) {
853 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
854 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
855 		goto err_listen;
856 	}
857 
858 	return 0;
859 
860 err_listen:
861 	ib_destroy_cm_id(priv->cm.id);
862 err_cm:
863 	priv->cm.id = NULL;
864 	return ret;
865 }
866 
867 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
868 {
869 	struct ipoib_dev_priv *priv = netdev_priv(dev);
870 	struct ipoib_cm_rx *rx, *n;
871 	LIST_HEAD(list);
872 
873 	spin_lock_irq(&priv->lock);
874 	list_splice_init(&priv->cm.rx_reap_list, &list);
875 	spin_unlock_irq(&priv->lock);
876 
877 	list_for_each_entry_safe(rx, n, &list, list) {
878 		ib_destroy_cm_id(rx->id);
879 		ib_destroy_qp(rx->qp);
880 		if (!ipoib_cm_has_srq(dev)) {
881 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
882 			spin_lock_irq(&priv->lock);
883 			--priv->cm.nonsrq_conn_qp;
884 			spin_unlock_irq(&priv->lock);
885 		}
886 		kfree(rx);
887 	}
888 }
889 
890 void ipoib_cm_dev_stop(struct net_device *dev)
891 {
892 	struct ipoib_dev_priv *priv = netdev_priv(dev);
893 	struct ipoib_cm_rx *p;
894 	unsigned long begin;
895 	int ret;
896 
897 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
898 		return;
899 
900 	ib_destroy_cm_id(priv->cm.id);
901 	priv->cm.id = NULL;
902 
903 	spin_lock_irq(&priv->lock);
904 	while (!list_empty(&priv->cm.passive_ids)) {
905 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
906 		list_move(&p->list, &priv->cm.rx_error_list);
907 		p->state = IPOIB_CM_RX_ERROR;
908 		spin_unlock_irq(&priv->lock);
909 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
910 		if (ret)
911 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
912 		spin_lock_irq(&priv->lock);
913 	}
914 
915 	/* Wait for all RX to be drained */
916 	begin = jiffies;
917 
918 	while (!list_empty(&priv->cm.rx_error_list) ||
919 	       !list_empty(&priv->cm.rx_flush_list) ||
920 	       !list_empty(&priv->cm.rx_drain_list)) {
921 		if (time_after(jiffies, begin + 5 * HZ)) {
922 			ipoib_warn(priv, "RX drain timing out\n");
923 
924 			/*
925 			 * assume the HW is wedged and just free up everything.
926 			 */
927 			list_splice_init(&priv->cm.rx_flush_list,
928 					 &priv->cm.rx_reap_list);
929 			list_splice_init(&priv->cm.rx_error_list,
930 					 &priv->cm.rx_reap_list);
931 			list_splice_init(&priv->cm.rx_drain_list,
932 					 &priv->cm.rx_reap_list);
933 			break;
934 		}
935 		spin_unlock_irq(&priv->lock);
936 		msleep(1);
937 		ipoib_drain_cq(dev);
938 		spin_lock_irq(&priv->lock);
939 	}
940 
941 	spin_unlock_irq(&priv->lock);
942 
943 	ipoib_cm_free_rx_reap_list(dev);
944 
945 	cancel_delayed_work(&priv->cm.stale_task);
946 }
947 
948 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
949 {
950 	struct ipoib_cm_tx *p = cm_id->context;
951 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
952 	struct ipoib_cm_data *data = event->private_data;
953 	struct sk_buff_head skqueue;
954 	struct ib_qp_attr qp_attr;
955 	int qp_attr_mask, ret;
956 	struct sk_buff *skb;
957 
958 	p->mtu = be32_to_cpu(data->mtu);
959 
960 	if (p->mtu <= IPOIB_ENCAP_LEN) {
961 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
962 			   p->mtu, IPOIB_ENCAP_LEN);
963 		return -EINVAL;
964 	}
965 
966 	qp_attr.qp_state = IB_QPS_RTR;
967 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
968 	if (ret) {
969 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
970 		return ret;
971 	}
972 
973 	qp_attr.rq_psn = 0 /* FIXME */;
974 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
975 	if (ret) {
976 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
977 		return ret;
978 	}
979 
980 	qp_attr.qp_state = IB_QPS_RTS;
981 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
982 	if (ret) {
983 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
984 		return ret;
985 	}
986 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
987 	if (ret) {
988 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
989 		return ret;
990 	}
991 
992 	skb_queue_head_init(&skqueue);
993 
994 	spin_lock_irq(&priv->lock);
995 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
996 	if (p->neigh)
997 		while ((skb = __skb_dequeue(&p->neigh->queue)))
998 			__skb_queue_tail(&skqueue, skb);
999 	spin_unlock_irq(&priv->lock);
1000 
1001 	while ((skb = __skb_dequeue(&skqueue))) {
1002 		skb->dev = p->dev;
1003 		if (dev_queue_xmit(skb))
1004 			ipoib_warn(priv, "dev_queue_xmit failed "
1005 				   "to requeue packet\n");
1006 	}
1007 
1008 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1009 	if (ret) {
1010 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1011 		return ret;
1012 	}
1013 	return 0;
1014 }
1015 
1016 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1017 {
1018 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1019 	struct ib_qp_init_attr attr = {
1020 		.send_cq		= priv->recv_cq,
1021 		.recv_cq		= priv->recv_cq,
1022 		.srq			= priv->cm.srq,
1023 		.cap.max_send_wr	= ipoib_sendq_size,
1024 		.cap.max_send_sge	= 1,
1025 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1026 		.qp_type		= IB_QPT_RC,
1027 		.qp_context		= tx
1028 	};
1029 
1030 	return ib_create_qp(priv->pd, &attr);
1031 }
1032 
1033 static int ipoib_cm_send_req(struct net_device *dev,
1034 			     struct ib_cm_id *id, struct ib_qp *qp,
1035 			     u32 qpn,
1036 			     struct ib_sa_path_rec *pathrec)
1037 {
1038 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1039 	struct ipoib_cm_data data = {};
1040 	struct ib_cm_req_param req = {};
1041 
1042 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1043 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1044 
1045 	req.primary_path		= pathrec;
1046 	req.alternate_path		= NULL;
1047 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1048 	req.qp_num			= qp->qp_num;
1049 	req.qp_type			= qp->qp_type;
1050 	req.private_data		= &data;
1051 	req.private_data_len		= sizeof data;
1052 	req.flow_control		= 0;
1053 
1054 	req.starting_psn		= 0; /* FIXME */
1055 
1056 	/*
1057 	 * Pick some arbitrary defaults here; we could make these
1058 	 * module parameters if anyone cared about setting them.
1059 	 */
1060 	req.responder_resources		= 4;
1061 	req.remote_cm_response_timeout	= 20;
1062 	req.local_cm_response_timeout	= 20;
1063 	req.retry_count			= 0; /* RFC draft warns against retries */
1064 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1065 	req.max_cm_retries		= 15;
1066 	req.srq				= ipoib_cm_has_srq(dev);
1067 	return ib_send_cm_req(id, &req);
1068 }
1069 
1070 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1071 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1072 {
1073 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1074 	struct ib_qp_attr qp_attr;
1075 	int qp_attr_mask, ret;
1076 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1077 	if (ret) {
1078 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1079 		return ret;
1080 	}
1081 
1082 	qp_attr.qp_state = IB_QPS_INIT;
1083 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1084 	qp_attr.port_num = priv->port;
1085 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1086 
1087 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1088 	if (ret) {
1089 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1090 		return ret;
1091 	}
1092 	return 0;
1093 }
1094 
1095 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1096 			    struct ib_sa_path_rec *pathrec)
1097 {
1098 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1099 	int ret;
1100 
1101 	p->tx_ring = vzalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1102 	if (!p->tx_ring) {
1103 		ipoib_warn(priv, "failed to allocate tx ring\n");
1104 		ret = -ENOMEM;
1105 		goto err_tx;
1106 	}
1107 
1108 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1109 	if (IS_ERR(p->qp)) {
1110 		ret = PTR_ERR(p->qp);
1111 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1112 		goto err_qp;
1113 	}
1114 
1115 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1116 	if (IS_ERR(p->id)) {
1117 		ret = PTR_ERR(p->id);
1118 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1119 		goto err_id;
1120 	}
1121 
1122 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1123 	if (ret) {
1124 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1125 		goto err_modify;
1126 	}
1127 
1128 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1129 	if (ret) {
1130 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1131 		goto err_send_cm;
1132 	}
1133 
1134 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1135 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1136 
1137 	return 0;
1138 
1139 err_send_cm:
1140 err_modify:
1141 	ib_destroy_cm_id(p->id);
1142 err_id:
1143 	p->id = NULL;
1144 	ib_destroy_qp(p->qp);
1145 err_qp:
1146 	p->qp = NULL;
1147 	vfree(p->tx_ring);
1148 err_tx:
1149 	return ret;
1150 }
1151 
1152 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1153 {
1154 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1155 	struct ipoib_cm_tx_buf *tx_req;
1156 	unsigned long begin;
1157 
1158 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1159 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1160 
1161 	if (p->id)
1162 		ib_destroy_cm_id(p->id);
1163 
1164 	if (p->tx_ring) {
1165 		/* Wait for all sends to complete */
1166 		begin = jiffies;
1167 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1168 			if (time_after(jiffies, begin + 5 * HZ)) {
1169 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1170 					   p->tx_head - p->tx_tail);
1171 				goto timeout;
1172 			}
1173 
1174 			msleep(1);
1175 		}
1176 	}
1177 
1178 timeout:
1179 
1180 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1181 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1182 		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1183 				    DMA_TO_DEVICE);
1184 		dev_kfree_skb_any(tx_req->skb);
1185 		++p->tx_tail;
1186 		netif_tx_lock_bh(p->dev);
1187 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1188 		    netif_queue_stopped(p->dev) &&
1189 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1190 			netif_wake_queue(p->dev);
1191 		netif_tx_unlock_bh(p->dev);
1192 	}
1193 
1194 	if (p->qp)
1195 		ib_destroy_qp(p->qp);
1196 
1197 	vfree(p->tx_ring);
1198 	kfree(p);
1199 }
1200 
1201 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1202 			       struct ib_cm_event *event)
1203 {
1204 	struct ipoib_cm_tx *tx = cm_id->context;
1205 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1206 	struct net_device *dev = priv->dev;
1207 	struct ipoib_neigh *neigh;
1208 	unsigned long flags;
1209 	int ret;
1210 
1211 	switch (event->event) {
1212 	case IB_CM_DREQ_RECEIVED:
1213 		ipoib_dbg(priv, "DREQ received.\n");
1214 		ib_send_cm_drep(cm_id, NULL, 0);
1215 		break;
1216 	case IB_CM_REP_RECEIVED:
1217 		ipoib_dbg(priv, "REP received.\n");
1218 		ret = ipoib_cm_rep_handler(cm_id, event);
1219 		if (ret)
1220 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1221 				       NULL, 0, NULL, 0);
1222 		break;
1223 	case IB_CM_REQ_ERROR:
1224 	case IB_CM_REJ_RECEIVED:
1225 	case IB_CM_TIMEWAIT_EXIT:
1226 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1227 		netif_tx_lock_bh(dev);
1228 		spin_lock_irqsave(&priv->lock, flags);
1229 		neigh = tx->neigh;
1230 
1231 		if (neigh) {
1232 			neigh->cm = NULL;
1233 			list_del(&neigh->list);
1234 			ipoib_neigh_free(neigh);
1235 
1236 			tx->neigh = NULL;
1237 		}
1238 
1239 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1240 			list_move(&tx->list, &priv->cm.reap_list);
1241 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1242 		}
1243 
1244 		spin_unlock_irqrestore(&priv->lock, flags);
1245 		netif_tx_unlock_bh(dev);
1246 		break;
1247 	default:
1248 		break;
1249 	}
1250 
1251 	return 0;
1252 }
1253 
1254 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1255 				       struct ipoib_neigh *neigh)
1256 {
1257 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1258 	struct ipoib_cm_tx *tx;
1259 
1260 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1261 	if (!tx)
1262 		return NULL;
1263 
1264 	neigh->cm = tx;
1265 	tx->neigh = neigh;
1266 	tx->path = path;
1267 	tx->dev = dev;
1268 	list_add(&tx->list, &priv->cm.start_list);
1269 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1270 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1271 	return tx;
1272 }
1273 
1274 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1275 {
1276 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1277 	unsigned long flags;
1278 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1279 		spin_lock_irqsave(&priv->lock, flags);
1280 		list_move(&tx->list, &priv->cm.reap_list);
1281 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1282 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1283 			  tx->neigh->daddr + 4);
1284 		tx->neigh = NULL;
1285 		spin_unlock_irqrestore(&priv->lock, flags);
1286 	}
1287 }
1288 
1289 static void ipoib_cm_tx_start(struct work_struct *work)
1290 {
1291 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1292 						   cm.start_task);
1293 	struct net_device *dev = priv->dev;
1294 	struct ipoib_neigh *neigh;
1295 	struct ipoib_cm_tx *p;
1296 	unsigned long flags;
1297 	int ret;
1298 
1299 	struct ib_sa_path_rec pathrec;
1300 	u32 qpn;
1301 
1302 	netif_tx_lock_bh(dev);
1303 	spin_lock_irqsave(&priv->lock, flags);
1304 
1305 	while (!list_empty(&priv->cm.start_list)) {
1306 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1307 		list_del_init(&p->list);
1308 		neigh = p->neigh;
1309 		qpn = IPOIB_QPN(neigh->daddr);
1310 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1311 
1312 		spin_unlock_irqrestore(&priv->lock, flags);
1313 		netif_tx_unlock_bh(dev);
1314 
1315 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1316 
1317 		netif_tx_lock_bh(dev);
1318 		spin_lock_irqsave(&priv->lock, flags);
1319 
1320 		if (ret) {
1321 			neigh = p->neigh;
1322 			if (neigh) {
1323 				neigh->cm = NULL;
1324 				list_del(&neigh->list);
1325 				ipoib_neigh_free(neigh);
1326 			}
1327 			list_del(&p->list);
1328 			kfree(p);
1329 		}
1330 	}
1331 
1332 	spin_unlock_irqrestore(&priv->lock, flags);
1333 	netif_tx_unlock_bh(dev);
1334 }
1335 
1336 static void ipoib_cm_tx_reap(struct work_struct *work)
1337 {
1338 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1339 						   cm.reap_task);
1340 	struct net_device *dev = priv->dev;
1341 	struct ipoib_cm_tx *p;
1342 	unsigned long flags;
1343 
1344 	netif_tx_lock_bh(dev);
1345 	spin_lock_irqsave(&priv->lock, flags);
1346 
1347 	while (!list_empty(&priv->cm.reap_list)) {
1348 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1349 		list_del(&p->list);
1350 		spin_unlock_irqrestore(&priv->lock, flags);
1351 		netif_tx_unlock_bh(dev);
1352 		ipoib_cm_tx_destroy(p);
1353 		netif_tx_lock_bh(dev);
1354 		spin_lock_irqsave(&priv->lock, flags);
1355 	}
1356 
1357 	spin_unlock_irqrestore(&priv->lock, flags);
1358 	netif_tx_unlock_bh(dev);
1359 }
1360 
1361 static void ipoib_cm_skb_reap(struct work_struct *work)
1362 {
1363 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1364 						   cm.skb_task);
1365 	struct net_device *dev = priv->dev;
1366 	struct sk_buff *skb;
1367 	unsigned long flags;
1368 	unsigned mtu = priv->mcast_mtu;
1369 
1370 	netif_tx_lock_bh(dev);
1371 	spin_lock_irqsave(&priv->lock, flags);
1372 
1373 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1374 		spin_unlock_irqrestore(&priv->lock, flags);
1375 		netif_tx_unlock_bh(dev);
1376 
1377 		if (skb->protocol == htons(ETH_P_IP))
1378 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1379 #if IS_ENABLED(CONFIG_IPV6)
1380 		else if (skb->protocol == htons(ETH_P_IPV6))
1381 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1382 #endif
1383 		dev_kfree_skb_any(skb);
1384 
1385 		netif_tx_lock_bh(dev);
1386 		spin_lock_irqsave(&priv->lock, flags);
1387 	}
1388 
1389 	spin_unlock_irqrestore(&priv->lock, flags);
1390 	netif_tx_unlock_bh(dev);
1391 }
1392 
1393 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1394 			   unsigned int mtu)
1395 {
1396 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1397 	int e = skb_queue_empty(&priv->cm.skb_queue);
1398 
1399 	if (skb_dst(skb))
1400 		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1401 
1402 	skb_queue_tail(&priv->cm.skb_queue, skb);
1403 	if (e)
1404 		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1405 }
1406 
1407 static void ipoib_cm_rx_reap(struct work_struct *work)
1408 {
1409 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1410 						cm.rx_reap_task)->dev);
1411 }
1412 
1413 static void ipoib_cm_stale_task(struct work_struct *work)
1414 {
1415 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1416 						   cm.stale_task.work);
1417 	struct ipoib_cm_rx *p;
1418 	int ret;
1419 
1420 	spin_lock_irq(&priv->lock);
1421 	while (!list_empty(&priv->cm.passive_ids)) {
1422 		/* List is sorted by LRU, start from tail,
1423 		 * stop when we see a recently used entry */
1424 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1425 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1426 			break;
1427 		list_move(&p->list, &priv->cm.rx_error_list);
1428 		p->state = IPOIB_CM_RX_ERROR;
1429 		spin_unlock_irq(&priv->lock);
1430 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1431 		if (ret)
1432 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1433 		spin_lock_irq(&priv->lock);
1434 	}
1435 
1436 	if (!list_empty(&priv->cm.passive_ids))
1437 		queue_delayed_work(ipoib_workqueue,
1438 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1439 	spin_unlock_irq(&priv->lock);
1440 }
1441 
1442 
1443 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1444 			 char *buf)
1445 {
1446 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1447 
1448 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1449 		return sprintf(buf, "connected\n");
1450 	else
1451 		return sprintf(buf, "datagram\n");
1452 }
1453 
1454 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1455 			const char *buf, size_t count)
1456 {
1457 	struct net_device *dev = to_net_dev(d);
1458 	int ret;
1459 
1460 	if (!rtnl_trylock())
1461 		return restart_syscall();
1462 
1463 	ret = ipoib_set_mode(dev, buf);
1464 
1465 	rtnl_unlock();
1466 
1467 	if (!ret)
1468 		return count;
1469 
1470 	return ret;
1471 }
1472 
1473 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1474 
1475 int ipoib_cm_add_mode_attr(struct net_device *dev)
1476 {
1477 	return device_create_file(&dev->dev, &dev_attr_mode);
1478 }
1479 
1480 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1481 {
1482 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1483 	struct ib_srq_init_attr srq_init_attr = {
1484 		.srq_type = IB_SRQT_BASIC,
1485 		.attr = {
1486 			.max_wr  = ipoib_recvq_size,
1487 			.max_sge = max_sge
1488 		}
1489 	};
1490 
1491 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1492 	if (IS_ERR(priv->cm.srq)) {
1493 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1494 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1495 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1496 		priv->cm.srq = NULL;
1497 		return;
1498 	}
1499 
1500 	priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1501 	if (!priv->cm.srq_ring) {
1502 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1503 		       priv->ca->name, ipoib_recvq_size);
1504 		ib_destroy_srq(priv->cm.srq);
1505 		priv->cm.srq = NULL;
1506 		return;
1507 	}
1508 
1509 }
1510 
1511 int ipoib_cm_dev_init(struct net_device *dev)
1512 {
1513 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1514 	int i, ret;
1515 	struct ib_device_attr attr;
1516 
1517 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1518 	INIT_LIST_HEAD(&priv->cm.reap_list);
1519 	INIT_LIST_HEAD(&priv->cm.start_list);
1520 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1521 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1522 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1523 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1524 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1525 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1526 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1527 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1528 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1529 
1530 	skb_queue_head_init(&priv->cm.skb_queue);
1531 
1532 	ret = ib_query_device(priv->ca, &attr);
1533 	if (ret) {
1534 		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1535 		return ret;
1536 	}
1537 
1538 	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1539 
1540 	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1541 	ipoib_cm_create_srq(dev, attr.max_srq_sge);
1542 	if (ipoib_cm_has_srq(dev)) {
1543 		priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1544 		priv->cm.num_frags  = attr.max_srq_sge;
1545 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1546 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1547 	} else {
1548 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1549 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1550 	}
1551 
1552 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1553 
1554 	if (ipoib_cm_has_srq(dev)) {
1555 		for (i = 0; i < ipoib_recvq_size; ++i) {
1556 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1557 						   priv->cm.num_frags - 1,
1558 						   priv->cm.srq_ring[i].mapping)) {
1559 				ipoib_warn(priv, "failed to allocate "
1560 					   "receive buffer %d\n", i);
1561 				ipoib_cm_dev_cleanup(dev);
1562 				return -ENOMEM;
1563 			}
1564 
1565 			if (ipoib_cm_post_receive_srq(dev, i)) {
1566 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1567 					   "failed for buf %d\n", i);
1568 				ipoib_cm_dev_cleanup(dev);
1569 				return -EIO;
1570 			}
1571 		}
1572 	}
1573 
1574 	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1575 	return 0;
1576 }
1577 
1578 void ipoib_cm_dev_cleanup(struct net_device *dev)
1579 {
1580 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1581 	int ret;
1582 
1583 	if (!priv->cm.srq)
1584 		return;
1585 
1586 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1587 
1588 	ret = ib_destroy_srq(priv->cm.srq);
1589 	if (ret)
1590 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1591 
1592 	priv->cm.srq = NULL;
1593 	if (!priv->cm.srq_ring)
1594 		return;
1595 
1596 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1597 	priv->cm.srq_ring = NULL;
1598 }
1599