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 	vfree(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 = 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 = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
356 	if (!rx->rx_ring) {
357 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
358 		       priv->ca->name, ipoib_recvq_size);
359 		return -ENOMEM;
360 	}
361 
362 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
363 
364 	t = kmalloc(sizeof *t, GFP_KERNEL);
365 	if (!t) {
366 		ret = -ENOMEM;
367 		goto err_free;
368 	}
369 
370 	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
371 
372 	spin_lock_irq(&priv->lock);
373 
374 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
375 		spin_unlock_irq(&priv->lock);
376 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
377 		ret = -EINVAL;
378 		goto err_free;
379 	} else
380 		++priv->cm.nonsrq_conn_qp;
381 
382 	spin_unlock_irq(&priv->lock);
383 
384 	for (i = 0; i < ipoib_recvq_size; ++i) {
385 		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
386 					   rx->rx_ring[i].mapping)) {
387 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
388 				ret = -ENOMEM;
389 				goto err_count;
390 		}
391 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
392 		if (ret) {
393 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
394 				   "failed for buf %d\n", i);
395 			ret = -EIO;
396 			goto err_count;
397 		}
398 	}
399 
400 	rx->recv_count = ipoib_recvq_size;
401 
402 	kfree(t);
403 
404 	return 0;
405 
406 err_count:
407 	spin_lock_irq(&priv->lock);
408 	--priv->cm.nonsrq_conn_qp;
409 	spin_unlock_irq(&priv->lock);
410 
411 err_free:
412 	kfree(t);
413 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
414 
415 	return ret;
416 }
417 
418 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
419 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
420 			     unsigned psn)
421 {
422 	struct ipoib_dev_priv *priv = netdev_priv(dev);
423 	struct ipoib_cm_data data = {};
424 	struct ib_cm_rep_param rep = {};
425 
426 	data.qpn = cpu_to_be32(priv->qp->qp_num);
427 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
428 
429 	rep.private_data = &data;
430 	rep.private_data_len = sizeof data;
431 	rep.flow_control = 0;
432 	rep.rnr_retry_count = req->rnr_retry_count;
433 	rep.srq = ipoib_cm_has_srq(dev);
434 	rep.qp_num = qp->qp_num;
435 	rep.starting_psn = psn;
436 	return ib_send_cm_rep(cm_id, &rep);
437 }
438 
439 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
440 {
441 	struct net_device *dev = cm_id->context;
442 	struct ipoib_dev_priv *priv = netdev_priv(dev);
443 	struct ipoib_cm_rx *p;
444 	unsigned psn;
445 	int ret;
446 
447 	ipoib_dbg(priv, "REQ arrived\n");
448 	p = kzalloc(sizeof *p, GFP_KERNEL);
449 	if (!p)
450 		return -ENOMEM;
451 	p->dev = dev;
452 	p->id = cm_id;
453 	cm_id->context = p;
454 	p->state = IPOIB_CM_RX_LIVE;
455 	p->jiffies = jiffies;
456 	INIT_LIST_HEAD(&p->list);
457 
458 	p->qp = ipoib_cm_create_rx_qp(dev, p);
459 	if (IS_ERR(p->qp)) {
460 		ret = PTR_ERR(p->qp);
461 		goto err_qp;
462 	}
463 
464 	psn = random32() & 0xffffff;
465 	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
466 	if (ret)
467 		goto err_modify;
468 
469 	if (!ipoib_cm_has_srq(dev)) {
470 		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
471 		if (ret)
472 			goto err_modify;
473 	}
474 
475 	spin_lock_irq(&priv->lock);
476 	queue_delayed_work(ipoib_workqueue,
477 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
478 	/* Add this entry to passive ids list head, but do not re-add it
479 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
480 	p->jiffies = jiffies;
481 	if (p->state == IPOIB_CM_RX_LIVE)
482 		list_move(&p->list, &priv->cm.passive_ids);
483 	spin_unlock_irq(&priv->lock);
484 
485 	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
486 	if (ret) {
487 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
488 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
489 			ipoib_warn(priv, "unable to move qp to error state\n");
490 	}
491 	return 0;
492 
493 err_modify:
494 	ib_destroy_qp(p->qp);
495 err_qp:
496 	kfree(p);
497 	return ret;
498 }
499 
500 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
501 			       struct ib_cm_event *event)
502 {
503 	struct ipoib_cm_rx *p;
504 	struct ipoib_dev_priv *priv;
505 
506 	switch (event->event) {
507 	case IB_CM_REQ_RECEIVED:
508 		return ipoib_cm_req_handler(cm_id, event);
509 	case IB_CM_DREQ_RECEIVED:
510 		p = cm_id->context;
511 		ib_send_cm_drep(cm_id, NULL, 0);
512 		/* Fall through */
513 	case IB_CM_REJ_RECEIVED:
514 		p = cm_id->context;
515 		priv = netdev_priv(p->dev);
516 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
517 			ipoib_warn(priv, "unable to move qp to error state\n");
518 		/* Fall through */
519 	default:
520 		return 0;
521 	}
522 }
523 /* Adjust length of skb with fragments to match received data */
524 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
525 			  unsigned int length, struct sk_buff *toskb)
526 {
527 	int i, num_frags;
528 	unsigned int size;
529 
530 	/* put header into skb */
531 	size = min(length, hdr_space);
532 	skb->tail += size;
533 	skb->len += size;
534 	length -= size;
535 
536 	num_frags = skb_shinfo(skb)->nr_frags;
537 	for (i = 0; i < num_frags; i++) {
538 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
539 
540 		if (length == 0) {
541 			/* don't need this page */
542 			skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
543 			--skb_shinfo(skb)->nr_frags;
544 		} else {
545 			size = min(length, (unsigned) PAGE_SIZE);
546 
547 			frag->size = 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->last_rx = jiffies;
667 	++dev->stats.rx_packets;
668 	dev->stats.rx_bytes += skb->len;
669 
670 	skb->dev = dev;
671 	/* XXX get correct PACKET_ type here */
672 	skb->pkt_type = PACKET_HOST;
673 	netif_receive_skb(skb);
674 
675 repost:
676 	if (has_srq) {
677 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
678 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
679 				   "for buf %d\n", wr_id);
680 	} else {
681 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
682 							  &priv->cm.rx_wr,
683 							  priv->cm.rx_sge,
684 							  wr_id))) {
685 			--p->recv_count;
686 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
687 				   "for buf %d\n", wr_id);
688 		}
689 	}
690 }
691 
692 static inline int post_send(struct ipoib_dev_priv *priv,
693 			    struct ipoib_cm_tx *tx,
694 			    unsigned int wr_id,
695 			    u64 addr, int len)
696 {
697 	struct ib_send_wr *bad_wr;
698 
699 	priv->tx_sge[0].addr          = addr;
700 	priv->tx_sge[0].length        = len;
701 
702 	priv->tx_wr.num_sge	= 1;
703 	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
704 
705 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
706 }
707 
708 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
709 {
710 	struct ipoib_dev_priv *priv = netdev_priv(dev);
711 	struct ipoib_cm_tx_buf *tx_req;
712 	u64 addr;
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 	if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
745 			       addr, skb->len))) {
746 		ipoib_warn(priv, "post_send failed\n");
747 		++dev->stats.tx_errors;
748 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
749 		dev_kfree_skb_any(skb);
750 	} else {
751 		dev->trans_start = jiffies;
752 		++tx->tx_head;
753 
754 		if (++priv->tx_outstanding == ipoib_sendq_size) {
755 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
756 				  tx->qp->qp_num);
757 			netif_stop_queue(dev);
758 		}
759 	}
760 }
761 
762 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
763 {
764 	struct ipoib_dev_priv *priv = netdev_priv(dev);
765 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
766 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
767 	struct ipoib_cm_tx_buf *tx_req;
768 	unsigned long flags;
769 
770 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
771 		       wr_id, wc->status);
772 
773 	if (unlikely(wr_id >= ipoib_sendq_size)) {
774 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
775 			   wr_id, ipoib_sendq_size);
776 		return;
777 	}
778 
779 	tx_req = &tx->tx_ring[wr_id];
780 
781 	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
782 
783 	/* FIXME: is this right? Shouldn't we only increment on success? */
784 	++dev->stats.tx_packets;
785 	dev->stats.tx_bytes += tx_req->skb->len;
786 
787 	dev_kfree_skb_any(tx_req->skb);
788 
789 	netif_tx_lock(dev);
790 
791 	++tx->tx_tail;
792 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
793 	    netif_queue_stopped(dev) &&
794 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
795 		netif_wake_queue(dev);
796 
797 	if (wc->status != IB_WC_SUCCESS &&
798 	    wc->status != IB_WC_WR_FLUSH_ERR) {
799 		struct ipoib_neigh *neigh;
800 
801 		ipoib_dbg(priv, "failed cm send event "
802 			   "(status=%d, wrid=%d vend_err %x)\n",
803 			   wc->status, wr_id, wc->vendor_err);
804 
805 		spin_lock_irqsave(&priv->lock, flags);
806 		neigh = tx->neigh;
807 
808 		if (neigh) {
809 			neigh->cm = NULL;
810 			list_del(&neigh->list);
811 			if (neigh->ah)
812 				ipoib_put_ah(neigh->ah);
813 			ipoib_neigh_free(dev, neigh);
814 
815 			tx->neigh = NULL;
816 		}
817 
818 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
819 			list_move(&tx->list, &priv->cm.reap_list);
820 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
821 		}
822 
823 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
824 
825 		spin_unlock_irqrestore(&priv->lock, flags);
826 	}
827 
828 	netif_tx_unlock(dev);
829 }
830 
831 int ipoib_cm_dev_open(struct net_device *dev)
832 {
833 	struct ipoib_dev_priv *priv = netdev_priv(dev);
834 	int ret;
835 
836 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
837 		return 0;
838 
839 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
840 	if (IS_ERR(priv->cm.id)) {
841 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
842 		ret = PTR_ERR(priv->cm.id);
843 		goto err_cm;
844 	}
845 
846 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
847 			   0, NULL);
848 	if (ret) {
849 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
850 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
851 		goto err_listen;
852 	}
853 
854 	return 0;
855 
856 err_listen:
857 	ib_destroy_cm_id(priv->cm.id);
858 err_cm:
859 	priv->cm.id = NULL;
860 	return ret;
861 }
862 
863 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
864 {
865 	struct ipoib_dev_priv *priv = netdev_priv(dev);
866 	struct ipoib_cm_rx *rx, *n;
867 	LIST_HEAD(list);
868 
869 	spin_lock_irq(&priv->lock);
870 	list_splice_init(&priv->cm.rx_reap_list, &list);
871 	spin_unlock_irq(&priv->lock);
872 
873 	list_for_each_entry_safe(rx, n, &list, list) {
874 		ib_destroy_cm_id(rx->id);
875 		ib_destroy_qp(rx->qp);
876 		if (!ipoib_cm_has_srq(dev)) {
877 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
878 			spin_lock_irq(&priv->lock);
879 			--priv->cm.nonsrq_conn_qp;
880 			spin_unlock_irq(&priv->lock);
881 		}
882 		kfree(rx);
883 	}
884 }
885 
886 void ipoib_cm_dev_stop(struct net_device *dev)
887 {
888 	struct ipoib_dev_priv *priv = netdev_priv(dev);
889 	struct ipoib_cm_rx *p;
890 	unsigned long begin;
891 	int ret;
892 
893 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
894 		return;
895 
896 	ib_destroy_cm_id(priv->cm.id);
897 	priv->cm.id = NULL;
898 
899 	spin_lock_irq(&priv->lock);
900 	while (!list_empty(&priv->cm.passive_ids)) {
901 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
902 		list_move(&p->list, &priv->cm.rx_error_list);
903 		p->state = IPOIB_CM_RX_ERROR;
904 		spin_unlock_irq(&priv->lock);
905 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
906 		if (ret)
907 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
908 		spin_lock_irq(&priv->lock);
909 	}
910 
911 	/* Wait for all RX to be drained */
912 	begin = jiffies;
913 
914 	while (!list_empty(&priv->cm.rx_error_list) ||
915 	       !list_empty(&priv->cm.rx_flush_list) ||
916 	       !list_empty(&priv->cm.rx_drain_list)) {
917 		if (time_after(jiffies, begin + 5 * HZ)) {
918 			ipoib_warn(priv, "RX drain timing out\n");
919 
920 			/*
921 			 * assume the HW is wedged and just free up everything.
922 			 */
923 			list_splice_init(&priv->cm.rx_flush_list,
924 					 &priv->cm.rx_reap_list);
925 			list_splice_init(&priv->cm.rx_error_list,
926 					 &priv->cm.rx_reap_list);
927 			list_splice_init(&priv->cm.rx_drain_list,
928 					 &priv->cm.rx_reap_list);
929 			break;
930 		}
931 		spin_unlock_irq(&priv->lock);
932 		msleep(1);
933 		ipoib_drain_cq(dev);
934 		spin_lock_irq(&priv->lock);
935 	}
936 
937 	spin_unlock_irq(&priv->lock);
938 
939 	ipoib_cm_free_rx_reap_list(dev);
940 
941 	cancel_delayed_work(&priv->cm.stale_task);
942 }
943 
944 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
945 {
946 	struct ipoib_cm_tx *p = cm_id->context;
947 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
948 	struct ipoib_cm_data *data = event->private_data;
949 	struct sk_buff_head skqueue;
950 	struct ib_qp_attr qp_attr;
951 	int qp_attr_mask, ret;
952 	struct sk_buff *skb;
953 
954 	p->mtu = be32_to_cpu(data->mtu);
955 
956 	if (p->mtu <= IPOIB_ENCAP_LEN) {
957 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
958 			   p->mtu, IPOIB_ENCAP_LEN);
959 		return -EINVAL;
960 	}
961 
962 	qp_attr.qp_state = IB_QPS_RTR;
963 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
964 	if (ret) {
965 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
966 		return ret;
967 	}
968 
969 	qp_attr.rq_psn = 0 /* FIXME */;
970 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
971 	if (ret) {
972 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
973 		return ret;
974 	}
975 
976 	qp_attr.qp_state = IB_QPS_RTS;
977 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
978 	if (ret) {
979 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
980 		return ret;
981 	}
982 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
983 	if (ret) {
984 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
985 		return ret;
986 	}
987 
988 	skb_queue_head_init(&skqueue);
989 
990 	spin_lock_irq(&priv->lock);
991 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
992 	if (p->neigh)
993 		while ((skb = __skb_dequeue(&p->neigh->queue)))
994 			__skb_queue_tail(&skqueue, skb);
995 	spin_unlock_irq(&priv->lock);
996 
997 	while ((skb = __skb_dequeue(&skqueue))) {
998 		skb->dev = p->dev;
999 		if (dev_queue_xmit(skb))
1000 			ipoib_warn(priv, "dev_queue_xmit failed "
1001 				   "to requeue packet\n");
1002 	}
1003 
1004 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1005 	if (ret) {
1006 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1007 		return ret;
1008 	}
1009 	return 0;
1010 }
1011 
1012 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1013 {
1014 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1015 	struct ib_qp_init_attr attr = {
1016 		.send_cq		= priv->recv_cq,
1017 		.recv_cq		= priv->recv_cq,
1018 		.srq			= priv->cm.srq,
1019 		.cap.max_send_wr	= ipoib_sendq_size,
1020 		.cap.max_send_sge	= 1,
1021 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1022 		.qp_type		= IB_QPT_RC,
1023 		.qp_context		= tx
1024 	};
1025 
1026 	return ib_create_qp(priv->pd, &attr);
1027 }
1028 
1029 static int ipoib_cm_send_req(struct net_device *dev,
1030 			     struct ib_cm_id *id, struct ib_qp *qp,
1031 			     u32 qpn,
1032 			     struct ib_sa_path_rec *pathrec)
1033 {
1034 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1035 	struct ipoib_cm_data data = {};
1036 	struct ib_cm_req_param req = {};
1037 
1038 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1039 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1040 
1041 	req.primary_path		= pathrec;
1042 	req.alternate_path		= NULL;
1043 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1044 	req.qp_num			= qp->qp_num;
1045 	req.qp_type			= qp->qp_type;
1046 	req.private_data		= &data;
1047 	req.private_data_len		= sizeof data;
1048 	req.flow_control		= 0;
1049 
1050 	req.starting_psn		= 0; /* FIXME */
1051 
1052 	/*
1053 	 * Pick some arbitrary defaults here; we could make these
1054 	 * module parameters if anyone cared about setting them.
1055 	 */
1056 	req.responder_resources		= 4;
1057 	req.remote_cm_response_timeout	= 20;
1058 	req.local_cm_response_timeout	= 20;
1059 	req.retry_count			= 0; /* RFC draft warns against retries */
1060 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1061 	req.max_cm_retries		= 15;
1062 	req.srq				= ipoib_cm_has_srq(dev);
1063 	return ib_send_cm_req(id, &req);
1064 }
1065 
1066 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1067 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1068 {
1069 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1070 	struct ib_qp_attr qp_attr;
1071 	int qp_attr_mask, ret;
1072 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1073 	if (ret) {
1074 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1075 		return ret;
1076 	}
1077 
1078 	qp_attr.qp_state = IB_QPS_INIT;
1079 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1080 	qp_attr.port_num = priv->port;
1081 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1082 
1083 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1084 	if (ret) {
1085 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1086 		return ret;
1087 	}
1088 	return 0;
1089 }
1090 
1091 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1092 			    struct ib_sa_path_rec *pathrec)
1093 {
1094 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1095 	int ret;
1096 
1097 	p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1098 	if (!p->tx_ring) {
1099 		ipoib_warn(priv, "failed to allocate tx ring\n");
1100 		ret = -ENOMEM;
1101 		goto err_tx;
1102 	}
1103 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1104 
1105 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1106 	if (IS_ERR(p->qp)) {
1107 		ret = PTR_ERR(p->qp);
1108 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1109 		goto err_qp;
1110 	}
1111 
1112 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1113 	if (IS_ERR(p->id)) {
1114 		ret = PTR_ERR(p->id);
1115 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1116 		goto err_id;
1117 	}
1118 
1119 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1120 	if (ret) {
1121 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1122 		goto err_modify;
1123 	}
1124 
1125 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1126 	if (ret) {
1127 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1128 		goto err_send_cm;
1129 	}
1130 
1131 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1132 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1133 
1134 	return 0;
1135 
1136 err_send_cm:
1137 err_modify:
1138 	ib_destroy_cm_id(p->id);
1139 err_id:
1140 	p->id = NULL;
1141 	ib_destroy_qp(p->qp);
1142 err_qp:
1143 	p->qp = NULL;
1144 	vfree(p->tx_ring);
1145 err_tx:
1146 	return ret;
1147 }
1148 
1149 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1150 {
1151 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1152 	struct ipoib_cm_tx_buf *tx_req;
1153 	unsigned long begin;
1154 
1155 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1156 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1157 
1158 	if (p->id)
1159 		ib_destroy_cm_id(p->id);
1160 
1161 	if (p->tx_ring) {
1162 		/* Wait for all sends to complete */
1163 		begin = jiffies;
1164 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1165 			if (time_after(jiffies, begin + 5 * HZ)) {
1166 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1167 					   p->tx_head - p->tx_tail);
1168 				goto timeout;
1169 			}
1170 
1171 			msleep(1);
1172 		}
1173 	}
1174 
1175 timeout:
1176 
1177 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1178 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1179 		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1180 				    DMA_TO_DEVICE);
1181 		dev_kfree_skb_any(tx_req->skb);
1182 		++p->tx_tail;
1183 		netif_tx_lock_bh(p->dev);
1184 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1185 		    netif_queue_stopped(p->dev) &&
1186 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1187 			netif_wake_queue(p->dev);
1188 		netif_tx_unlock_bh(p->dev);
1189 	}
1190 
1191 	if (p->qp)
1192 		ib_destroy_qp(p->qp);
1193 
1194 	vfree(p->tx_ring);
1195 	kfree(p);
1196 }
1197 
1198 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1199 			       struct ib_cm_event *event)
1200 {
1201 	struct ipoib_cm_tx *tx = cm_id->context;
1202 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1203 	struct net_device *dev = priv->dev;
1204 	struct ipoib_neigh *neigh;
1205 	unsigned long flags;
1206 	int ret;
1207 
1208 	switch (event->event) {
1209 	case IB_CM_DREQ_RECEIVED:
1210 		ipoib_dbg(priv, "DREQ received.\n");
1211 		ib_send_cm_drep(cm_id, NULL, 0);
1212 		break;
1213 	case IB_CM_REP_RECEIVED:
1214 		ipoib_dbg(priv, "REP received.\n");
1215 		ret = ipoib_cm_rep_handler(cm_id, event);
1216 		if (ret)
1217 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1218 				       NULL, 0, NULL, 0);
1219 		break;
1220 	case IB_CM_REQ_ERROR:
1221 	case IB_CM_REJ_RECEIVED:
1222 	case IB_CM_TIMEWAIT_EXIT:
1223 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1224 		netif_tx_lock_bh(dev);
1225 		spin_lock_irqsave(&priv->lock, flags);
1226 		neigh = tx->neigh;
1227 
1228 		if (neigh) {
1229 			neigh->cm = NULL;
1230 			list_del(&neigh->list);
1231 			if (neigh->ah)
1232 				ipoib_put_ah(neigh->ah);
1233 			ipoib_neigh_free(dev, neigh);
1234 
1235 			tx->neigh = NULL;
1236 		}
1237 
1238 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1239 			list_move(&tx->list, &priv->cm.reap_list);
1240 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1241 		}
1242 
1243 		spin_unlock_irqrestore(&priv->lock, flags);
1244 		netif_tx_unlock_bh(dev);
1245 		break;
1246 	default:
1247 		break;
1248 	}
1249 
1250 	return 0;
1251 }
1252 
1253 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1254 				       struct ipoib_neigh *neigh)
1255 {
1256 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1257 	struct ipoib_cm_tx *tx;
1258 
1259 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1260 	if (!tx)
1261 		return NULL;
1262 
1263 	neigh->cm = tx;
1264 	tx->neigh = neigh;
1265 	tx->path = path;
1266 	tx->dev = dev;
1267 	list_add(&tx->list, &priv->cm.start_list);
1268 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1269 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1270 	return tx;
1271 }
1272 
1273 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1274 {
1275 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1276 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1277 		list_move(&tx->list, &priv->cm.reap_list);
1278 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1279 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1280 			  tx->neigh->dgid.raw);
1281 		tx->neigh = NULL;
1282 	}
1283 }
1284 
1285 static void ipoib_cm_tx_start(struct work_struct *work)
1286 {
1287 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1288 						   cm.start_task);
1289 	struct net_device *dev = priv->dev;
1290 	struct ipoib_neigh *neigh;
1291 	struct ipoib_cm_tx *p;
1292 	unsigned long flags;
1293 	int ret;
1294 
1295 	struct ib_sa_path_rec pathrec;
1296 	u32 qpn;
1297 
1298 	netif_tx_lock_bh(dev);
1299 	spin_lock_irqsave(&priv->lock, flags);
1300 
1301 	while (!list_empty(&priv->cm.start_list)) {
1302 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1303 		list_del_init(&p->list);
1304 		neigh = p->neigh;
1305 		qpn = IPOIB_QPN(neigh->neighbour->ha);
1306 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1307 
1308 		spin_unlock_irqrestore(&priv->lock, flags);
1309 		netif_tx_unlock_bh(dev);
1310 
1311 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1312 
1313 		netif_tx_lock_bh(dev);
1314 		spin_lock_irqsave(&priv->lock, flags);
1315 
1316 		if (ret) {
1317 			neigh = p->neigh;
1318 			if (neigh) {
1319 				neigh->cm = NULL;
1320 				list_del(&neigh->list);
1321 				if (neigh->ah)
1322 					ipoib_put_ah(neigh->ah);
1323 				ipoib_neigh_free(dev, neigh);
1324 			}
1325 			list_del(&p->list);
1326 			kfree(p);
1327 		}
1328 	}
1329 
1330 	spin_unlock_irqrestore(&priv->lock, flags);
1331 	netif_tx_unlock_bh(dev);
1332 }
1333 
1334 static void ipoib_cm_tx_reap(struct work_struct *work)
1335 {
1336 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1337 						   cm.reap_task);
1338 	struct net_device *dev = priv->dev;
1339 	struct ipoib_cm_tx *p;
1340 	unsigned long flags;
1341 
1342 	netif_tx_lock_bh(dev);
1343 	spin_lock_irqsave(&priv->lock, flags);
1344 
1345 	while (!list_empty(&priv->cm.reap_list)) {
1346 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1347 		list_del(&p->list);
1348 		spin_unlock_irqrestore(&priv->lock, flags);
1349 		netif_tx_unlock_bh(dev);
1350 		ipoib_cm_tx_destroy(p);
1351 		netif_tx_lock_bh(dev);
1352 		spin_lock_irqsave(&priv->lock, flags);
1353 	}
1354 
1355 	spin_unlock_irqrestore(&priv->lock, flags);
1356 	netif_tx_unlock_bh(dev);
1357 }
1358 
1359 static void ipoib_cm_skb_reap(struct work_struct *work)
1360 {
1361 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1362 						   cm.skb_task);
1363 	struct net_device *dev = priv->dev;
1364 	struct sk_buff *skb;
1365 	unsigned long flags;
1366 	unsigned mtu = priv->mcast_mtu;
1367 
1368 	netif_tx_lock_bh(dev);
1369 	spin_lock_irqsave(&priv->lock, flags);
1370 
1371 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1372 		spin_unlock_irqrestore(&priv->lock, flags);
1373 		netif_tx_unlock_bh(dev);
1374 
1375 		if (skb->protocol == htons(ETH_P_IP))
1376 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1377 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1378 		else if (skb->protocol == htons(ETH_P_IPV6))
1379 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1380 #endif
1381 		dev_kfree_skb_any(skb);
1382 
1383 		netif_tx_lock_bh(dev);
1384 		spin_lock_irqsave(&priv->lock, flags);
1385 	}
1386 
1387 	spin_unlock_irqrestore(&priv->lock, flags);
1388 	netif_tx_unlock_bh(dev);
1389 }
1390 
1391 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1392 			   unsigned int mtu)
1393 {
1394 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1395 	int e = skb_queue_empty(&priv->cm.skb_queue);
1396 
1397 	if (skb->dst)
1398 		skb->dst->ops->update_pmtu(skb->dst, mtu);
1399 
1400 	skb_queue_tail(&priv->cm.skb_queue, skb);
1401 	if (e)
1402 		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1403 }
1404 
1405 static void ipoib_cm_rx_reap(struct work_struct *work)
1406 {
1407 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1408 						cm.rx_reap_task)->dev);
1409 }
1410 
1411 static void ipoib_cm_stale_task(struct work_struct *work)
1412 {
1413 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1414 						   cm.stale_task.work);
1415 	struct ipoib_cm_rx *p;
1416 	int ret;
1417 
1418 	spin_lock_irq(&priv->lock);
1419 	while (!list_empty(&priv->cm.passive_ids)) {
1420 		/* List is sorted by LRU, start from tail,
1421 		 * stop when we see a recently used entry */
1422 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1423 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1424 			break;
1425 		list_move(&p->list, &priv->cm.rx_error_list);
1426 		p->state = IPOIB_CM_RX_ERROR;
1427 		spin_unlock_irq(&priv->lock);
1428 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1429 		if (ret)
1430 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1431 		spin_lock_irq(&priv->lock);
1432 	}
1433 
1434 	if (!list_empty(&priv->cm.passive_ids))
1435 		queue_delayed_work(ipoib_workqueue,
1436 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1437 	spin_unlock_irq(&priv->lock);
1438 }
1439 
1440 
1441 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1442 			 char *buf)
1443 {
1444 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1445 
1446 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1447 		return sprintf(buf, "connected\n");
1448 	else
1449 		return sprintf(buf, "datagram\n");
1450 }
1451 
1452 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1453 			const char *buf, size_t count)
1454 {
1455 	struct net_device *dev = to_net_dev(d);
1456 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1457 
1458 	/* flush paths if we switch modes so that connections are restarted */
1459 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1460 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1461 		ipoib_warn(priv, "enabling connected mode "
1462 			   "will cause multicast packet drops\n");
1463 
1464 		rtnl_lock();
1465 		dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1466 		rtnl_unlock();
1467 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1468 
1469 		ipoib_flush_paths(dev);
1470 		return count;
1471 	}
1472 
1473 	if (!strcmp(buf, "datagram\n")) {
1474 		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1475 
1476 		rtnl_lock();
1477 		if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1478 			dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1479 			if (priv->hca_caps & IB_DEVICE_UD_TSO)
1480 				dev->features |= NETIF_F_TSO;
1481 		}
1482 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1483 		rtnl_unlock();
1484 		ipoib_flush_paths(dev);
1485 
1486 		return count;
1487 	}
1488 
1489 	return -EINVAL;
1490 }
1491 
1492 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1493 
1494 int ipoib_cm_add_mode_attr(struct net_device *dev)
1495 {
1496 	return device_create_file(&dev->dev, &dev_attr_mode);
1497 }
1498 
1499 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1500 {
1501 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1502 	struct ib_srq_init_attr srq_init_attr = {
1503 		.attr = {
1504 			.max_wr  = ipoib_recvq_size,
1505 			.max_sge = max_sge
1506 		}
1507 	};
1508 
1509 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1510 	if (IS_ERR(priv->cm.srq)) {
1511 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1512 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1513 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1514 		priv->cm.srq = NULL;
1515 		return;
1516 	}
1517 
1518 	priv->cm.srq_ring = vmalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1519 	if (!priv->cm.srq_ring) {
1520 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1521 		       priv->ca->name, ipoib_recvq_size);
1522 		ib_destroy_srq(priv->cm.srq);
1523 		priv->cm.srq = NULL;
1524 		return;
1525 	}
1526 
1527 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1528 }
1529 
1530 int ipoib_cm_dev_init(struct net_device *dev)
1531 {
1532 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1533 	int i, ret;
1534 	struct ib_device_attr attr;
1535 
1536 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1537 	INIT_LIST_HEAD(&priv->cm.reap_list);
1538 	INIT_LIST_HEAD(&priv->cm.start_list);
1539 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1540 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1541 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1542 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1543 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1544 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1545 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1546 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1547 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1548 
1549 	skb_queue_head_init(&priv->cm.skb_queue);
1550 
1551 	ret = ib_query_device(priv->ca, &attr);
1552 	if (ret) {
1553 		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1554 		return ret;
1555 	}
1556 
1557 	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1558 
1559 	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1560 	ipoib_cm_create_srq(dev, attr.max_srq_sge);
1561 	if (ipoib_cm_has_srq(dev)) {
1562 		priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1563 		priv->cm.num_frags  = attr.max_srq_sge;
1564 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1565 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1566 	} else {
1567 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1568 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1569 	}
1570 
1571 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1572 
1573 	if (ipoib_cm_has_srq(dev)) {
1574 		for (i = 0; i < ipoib_recvq_size; ++i) {
1575 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1576 						   priv->cm.num_frags - 1,
1577 						   priv->cm.srq_ring[i].mapping)) {
1578 				ipoib_warn(priv, "failed to allocate "
1579 					   "receive buffer %d\n", i);
1580 				ipoib_cm_dev_cleanup(dev);
1581 				return -ENOMEM;
1582 			}
1583 
1584 			if (ipoib_cm_post_receive_srq(dev, i)) {
1585 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1586 					   "failed for buf %d\n", i);
1587 				ipoib_cm_dev_cleanup(dev);
1588 				return -EIO;
1589 			}
1590 		}
1591 	}
1592 
1593 	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1594 	return 0;
1595 }
1596 
1597 void ipoib_cm_dev_cleanup(struct net_device *dev)
1598 {
1599 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1600 	int ret;
1601 
1602 	if (!priv->cm.srq)
1603 		return;
1604 
1605 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1606 
1607 	ret = ib_destroy_srq(priv->cm.srq);
1608 	if (ret)
1609 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1610 
1611 	priv->cm.srq = NULL;
1612 	if (!priv->cm.srq_ring)
1613 		return;
1614 
1615 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1616 	priv->cm.srq_ring = NULL;
1617 }
1618