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/vmalloc.h>
39 
40 #include "ipoib.h"
41 
42 int ipoib_max_conn_qp = 128;
43 
44 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
45 MODULE_PARM_DESC(max_nonsrq_conn_qp,
46 		 "Max number of connected-mode QPs per interface "
47 		 "(applied only if shared receive queue is not available)");
48 
49 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
50 static int data_debug_level;
51 
52 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
53 MODULE_PARM_DESC(cm_data_debug_level,
54 		 "Enable data path debug tracing for connected mode if > 0");
55 #endif
56 
57 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
58 
59 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
60 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
61 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
62 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
63 
64 static struct ib_qp_attr ipoib_cm_err_attr = {
65 	.qp_state = IB_QPS_ERR
66 };
67 
68 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
69 
70 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
71 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
72 	.opcode = IB_WR_SEND,
73 };
74 
75 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
76 			       struct ib_cm_event *event);
77 
78 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
79 				  u64 mapping[IPOIB_CM_RX_SG])
80 {
81 	int i;
82 
83 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
84 
85 	for (i = 0; i < frags; ++i)
86 		ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
87 }
88 
89 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
90 {
91 	struct ipoib_dev_priv *priv = netdev_priv(dev);
92 	struct ib_recv_wr *bad_wr;
93 	int i, ret;
94 
95 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
96 
97 	for (i = 0; i < priv->cm.num_frags; ++i)
98 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
99 
100 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
101 	if (unlikely(ret)) {
102 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
103 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
104 				      priv->cm.srq_ring[id].mapping);
105 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
106 		priv->cm.srq_ring[id].skb = NULL;
107 	}
108 
109 	return ret;
110 }
111 
112 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
113 					struct ipoib_cm_rx *rx,
114 					struct ib_recv_wr *wr,
115 					struct ib_sge *sge, int id)
116 {
117 	struct ipoib_dev_priv *priv = netdev_priv(dev);
118 	struct ib_recv_wr *bad_wr;
119 	int i, ret;
120 
121 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
122 
123 	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
124 		sge[i].addr = rx->rx_ring[id].mapping[i];
125 
126 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
127 	if (unlikely(ret)) {
128 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
129 		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
130 				      rx->rx_ring[id].mapping);
131 		dev_kfree_skb_any(rx->rx_ring[id].skb);
132 		rx->rx_ring[id].skb = NULL;
133 	}
134 
135 	return ret;
136 }
137 
138 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
139 					     struct ipoib_cm_rx_buf *rx_ring,
140 					     int id, int frags,
141 					     u64 mapping[IPOIB_CM_RX_SG])
142 {
143 	struct ipoib_dev_priv *priv = netdev_priv(dev);
144 	struct sk_buff *skb;
145 	int i;
146 
147 	skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
148 	if (unlikely(!skb))
149 		return NULL;
150 
151 	/*
152 	 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
153 	 * IP header to a multiple of 16.
154 	 */
155 	skb_reserve(skb, 12);
156 
157 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
158 				       DMA_FROM_DEVICE);
159 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
160 		dev_kfree_skb_any(skb);
161 		return NULL;
162 	}
163 
164 	for (i = 0; i < frags; i++) {
165 		struct page *page = alloc_page(GFP_ATOMIC);
166 
167 		if (!page)
168 			goto partial_error;
169 		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
170 
171 		mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
172 						 0, PAGE_SIZE, DMA_FROM_DEVICE);
173 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
174 			goto partial_error;
175 	}
176 
177 	rx_ring[id].skb = skb;
178 	return skb;
179 
180 partial_error:
181 
182 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
183 
184 	for (; i > 0; --i)
185 		ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
186 
187 	dev_kfree_skb_any(skb);
188 	return NULL;
189 }
190 
191 static void ipoib_cm_free_rx_ring(struct net_device *dev,
192 				  struct ipoib_cm_rx_buf *rx_ring)
193 {
194 	struct ipoib_dev_priv *priv = netdev_priv(dev);
195 	int i;
196 
197 	for (i = 0; i < ipoib_recvq_size; ++i)
198 		if (rx_ring[i].skb) {
199 			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
200 					      rx_ring[i].mapping);
201 			dev_kfree_skb_any(rx_ring[i].skb);
202 		}
203 
204 	vfree(rx_ring);
205 }
206 
207 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
208 {
209 	struct ib_send_wr *bad_wr;
210 	struct ipoib_cm_rx *p;
211 
212 	/* We only reserved 1 extra slot in CQ for drain WRs, so
213 	 * make sure we have at most 1 outstanding WR. */
214 	if (list_empty(&priv->cm.rx_flush_list) ||
215 	    !list_empty(&priv->cm.rx_drain_list))
216 		return;
217 
218 	/*
219 	 * QPs on flush list are error state.  This way, a "flush
220 	 * error" WC will be immediately generated for each WR we post.
221 	 */
222 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
223 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
224 		ipoib_warn(priv, "failed to post drain wr\n");
225 
226 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
227 }
228 
229 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
230 {
231 	struct ipoib_cm_rx *p = ctx;
232 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
233 	unsigned long flags;
234 
235 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
236 		return;
237 
238 	spin_lock_irqsave(&priv->lock, flags);
239 	list_move(&p->list, &priv->cm.rx_flush_list);
240 	p->state = IPOIB_CM_RX_FLUSH;
241 	ipoib_cm_start_rx_drain(priv);
242 	spin_unlock_irqrestore(&priv->lock, flags);
243 }
244 
245 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
246 					   struct ipoib_cm_rx *p)
247 {
248 	struct ipoib_dev_priv *priv = netdev_priv(dev);
249 	struct ib_qp_init_attr attr = {
250 		.event_handler = ipoib_cm_rx_event_handler,
251 		.send_cq = priv->recv_cq, /* For drain WR */
252 		.recv_cq = priv->recv_cq,
253 		.srq = priv->cm.srq,
254 		.cap.max_send_wr = 1, /* For drain WR */
255 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
256 		.sq_sig_type = IB_SIGNAL_ALL_WR,
257 		.qp_type = IB_QPT_RC,
258 		.qp_context = p,
259 	};
260 
261 	if (!ipoib_cm_has_srq(dev)) {
262 		attr.cap.max_recv_wr  = ipoib_recvq_size;
263 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
264 	}
265 
266 	return ib_create_qp(priv->pd, &attr);
267 }
268 
269 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
270 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
271 				 unsigned psn)
272 {
273 	struct ipoib_dev_priv *priv = netdev_priv(dev);
274 	struct ib_qp_attr qp_attr;
275 	int qp_attr_mask, ret;
276 
277 	qp_attr.qp_state = IB_QPS_INIT;
278 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
279 	if (ret) {
280 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
281 		return ret;
282 	}
283 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
284 	if (ret) {
285 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
286 		return ret;
287 	}
288 	qp_attr.qp_state = IB_QPS_RTR;
289 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
290 	if (ret) {
291 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
292 		return ret;
293 	}
294 	qp_attr.rq_psn = psn;
295 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
296 	if (ret) {
297 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
298 		return ret;
299 	}
300 
301 	/*
302 	 * Current Mellanox HCA firmware won't generate completions
303 	 * with error for drain WRs unless the QP has been moved to
304 	 * RTS first. This work-around leaves a window where a QP has
305 	 * moved to error asynchronously, but this will eventually get
306 	 * fixed in firmware, so let's not error out if modify QP
307 	 * fails.
308 	 */
309 	qp_attr.qp_state = IB_QPS_RTS;
310 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
311 	if (ret) {
312 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
313 		return 0;
314 	}
315 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
316 	if (ret) {
317 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
318 		return 0;
319 	}
320 
321 	return 0;
322 }
323 
324 static void ipoib_cm_init_rx_wr(struct net_device *dev,
325 				struct ib_recv_wr *wr,
326 				struct ib_sge *sge)
327 {
328 	struct ipoib_dev_priv *priv = netdev_priv(dev);
329 	int i;
330 
331 	for (i = 0; i < priv->cm.num_frags; ++i)
332 		sge[i].lkey = priv->mr->lkey;
333 
334 	sge[0].length = IPOIB_CM_HEAD_SIZE;
335 	for (i = 1; i < priv->cm.num_frags; ++i)
336 		sge[i].length = PAGE_SIZE;
337 
338 	wr->next    = NULL;
339 	wr->sg_list = sge;
340 	wr->num_sge = priv->cm.num_frags;
341 }
342 
343 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
344 				   struct ipoib_cm_rx *rx)
345 {
346 	struct ipoib_dev_priv *priv = netdev_priv(dev);
347 	struct {
348 		struct ib_recv_wr wr;
349 		struct ib_sge sge[IPOIB_CM_RX_SG];
350 	} *t;
351 	int ret;
352 	int i;
353 
354 	rx->rx_ring = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
355 	if (!rx->rx_ring) {
356 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
357 		       priv->ca->name, ipoib_recvq_size);
358 		return -ENOMEM;
359 	}
360 
361 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
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, frag->page, 0, PAGE_SIZE);
542 			--skb_shinfo(skb)->nr_frags;
543 		} else {
544 			size = min(length, (unsigned) PAGE_SIZE);
545 
546 			frag->size = size;
547 			skb->data_len += size;
548 			skb->truesize += size;
549 			skb->len += size;
550 			length -= size;
551 		}
552 	}
553 }
554 
555 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
556 {
557 	struct ipoib_dev_priv *priv = netdev_priv(dev);
558 	struct ipoib_cm_rx_buf *rx_ring;
559 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
560 	struct sk_buff *skb, *newskb;
561 	struct ipoib_cm_rx *p;
562 	unsigned long flags;
563 	u64 mapping[IPOIB_CM_RX_SG];
564 	int frags;
565 	int has_srq;
566 	struct sk_buff *small_skb;
567 
568 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
569 		       wr_id, wc->status);
570 
571 	if (unlikely(wr_id >= ipoib_recvq_size)) {
572 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
573 			spin_lock_irqsave(&priv->lock, flags);
574 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
575 			ipoib_cm_start_rx_drain(priv);
576 			queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
577 			spin_unlock_irqrestore(&priv->lock, flags);
578 		} else
579 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
580 				   wr_id, ipoib_recvq_size);
581 		return;
582 	}
583 
584 	p = wc->qp->qp_context;
585 
586 	has_srq = ipoib_cm_has_srq(dev);
587 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
588 
589 	skb = rx_ring[wr_id].skb;
590 
591 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
592 		ipoib_dbg(priv, "cm recv error "
593 			   "(status=%d, wrid=%d vend_err %x)\n",
594 			   wc->status, wr_id, wc->vendor_err);
595 		++dev->stats.rx_dropped;
596 		if (has_srq)
597 			goto repost;
598 		else {
599 			if (!--p->recv_count) {
600 				spin_lock_irqsave(&priv->lock, flags);
601 				list_move(&p->list, &priv->cm.rx_reap_list);
602 				spin_unlock_irqrestore(&priv->lock, flags);
603 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
604 			}
605 			return;
606 		}
607 	}
608 
609 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
610 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
611 			spin_lock_irqsave(&priv->lock, flags);
612 			p->jiffies = jiffies;
613 			/* Move this entry to list head, but do not re-add it
614 			 * if it has been moved out of list. */
615 			if (p->state == IPOIB_CM_RX_LIVE)
616 				list_move(&p->list, &priv->cm.passive_ids);
617 			spin_unlock_irqrestore(&priv->lock, flags);
618 		}
619 	}
620 
621 	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
622 		int dlen = wc->byte_len;
623 
624 		small_skb = dev_alloc_skb(dlen + 12);
625 		if (small_skb) {
626 			skb_reserve(small_skb, 12);
627 			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
628 						   dlen, DMA_FROM_DEVICE);
629 			skb_copy_from_linear_data(skb, small_skb->data, dlen);
630 			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
631 						      dlen, DMA_FROM_DEVICE);
632 			skb_put(small_skb, dlen);
633 			skb = small_skb;
634 			goto copied;
635 		}
636 	}
637 
638 	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
639 					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
640 
641 	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
642 	if (unlikely(!newskb)) {
643 		/*
644 		 * If we can't allocate a new RX buffer, dump
645 		 * this packet and reuse the old buffer.
646 		 */
647 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
648 		++dev->stats.rx_dropped;
649 		goto repost;
650 	}
651 
652 	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
653 	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
654 
655 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
656 		       wc->byte_len, wc->slid);
657 
658 	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
659 
660 copied:
661 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
662 	skb_reset_mac_header(skb);
663 	skb_pull(skb, IPOIB_ENCAP_LEN);
664 
665 	++dev->stats.rx_packets;
666 	dev->stats.rx_bytes += skb->len;
667 
668 	skb->dev = dev;
669 	/* XXX get correct PACKET_ type here */
670 	skb->pkt_type = PACKET_HOST;
671 	netif_receive_skb(skb);
672 
673 repost:
674 	if (has_srq) {
675 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
676 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
677 				   "for buf %d\n", wr_id);
678 	} else {
679 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
680 							  &priv->cm.rx_wr,
681 							  priv->cm.rx_sge,
682 							  wr_id))) {
683 			--p->recv_count;
684 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
685 				   "for buf %d\n", wr_id);
686 		}
687 	}
688 }
689 
690 static inline int post_send(struct ipoib_dev_priv *priv,
691 			    struct ipoib_cm_tx *tx,
692 			    unsigned int wr_id,
693 			    u64 addr, int len)
694 {
695 	struct ib_send_wr *bad_wr;
696 
697 	priv->tx_sge[0].addr          = addr;
698 	priv->tx_sge[0].length        = len;
699 
700 	priv->tx_wr.num_sge	= 1;
701 	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
702 
703 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
704 }
705 
706 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
707 {
708 	struct ipoib_dev_priv *priv = netdev_priv(dev);
709 	struct ipoib_cm_tx_buf *tx_req;
710 	u64 addr;
711 
712 	if (unlikely(skb->len > tx->mtu)) {
713 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
714 			   skb->len, tx->mtu);
715 		++dev->stats.tx_dropped;
716 		++dev->stats.tx_errors;
717 		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
718 		return;
719 	}
720 
721 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
722 		       tx->tx_head, skb->len, tx->qp->qp_num);
723 
724 	/*
725 	 * We put the skb into the tx_ring _before_ we call post_send()
726 	 * because it's entirely possible that the completion handler will
727 	 * run before we execute anything after the post_send().  That
728 	 * means we have to make sure everything is properly recorded and
729 	 * our state is consistent before we call post_send().
730 	 */
731 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
732 	tx_req->skb = skb;
733 	addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
734 	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
735 		++dev->stats.tx_errors;
736 		dev_kfree_skb_any(skb);
737 		return;
738 	}
739 
740 	tx_req->mapping = addr;
741 
742 	if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
743 			       addr, skb->len))) {
744 		ipoib_warn(priv, "post_send failed\n");
745 		++dev->stats.tx_errors;
746 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
747 		dev_kfree_skb_any(skb);
748 	} else {
749 		dev->trans_start = jiffies;
750 		++tx->tx_head;
751 
752 		if (++priv->tx_outstanding == ipoib_sendq_size) {
753 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
754 				  tx->qp->qp_num);
755 			netif_stop_queue(dev);
756 		}
757 	}
758 }
759 
760 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
761 {
762 	struct ipoib_dev_priv *priv = netdev_priv(dev);
763 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
764 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
765 	struct ipoib_cm_tx_buf *tx_req;
766 	unsigned long flags;
767 
768 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
769 		       wr_id, wc->status);
770 
771 	if (unlikely(wr_id >= ipoib_sendq_size)) {
772 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
773 			   wr_id, ipoib_sendq_size);
774 		return;
775 	}
776 
777 	tx_req = &tx->tx_ring[wr_id];
778 
779 	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
780 
781 	/* FIXME: is this right? Shouldn't we only increment on success? */
782 	++dev->stats.tx_packets;
783 	dev->stats.tx_bytes += tx_req->skb->len;
784 
785 	dev_kfree_skb_any(tx_req->skb);
786 
787 	netif_tx_lock(dev);
788 
789 	++tx->tx_tail;
790 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
791 	    netif_queue_stopped(dev) &&
792 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
793 		netif_wake_queue(dev);
794 
795 	if (wc->status != IB_WC_SUCCESS &&
796 	    wc->status != IB_WC_WR_FLUSH_ERR) {
797 		struct ipoib_neigh *neigh;
798 
799 		ipoib_dbg(priv, "failed cm send event "
800 			   "(status=%d, wrid=%d vend_err %x)\n",
801 			   wc->status, wr_id, wc->vendor_err);
802 
803 		spin_lock_irqsave(&priv->lock, flags);
804 		neigh = tx->neigh;
805 
806 		if (neigh) {
807 			neigh->cm = NULL;
808 			list_del(&neigh->list);
809 			if (neigh->ah)
810 				ipoib_put_ah(neigh->ah);
811 			ipoib_neigh_free(dev, neigh);
812 
813 			tx->neigh = NULL;
814 		}
815 
816 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
817 			list_move(&tx->list, &priv->cm.reap_list);
818 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
819 		}
820 
821 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
822 
823 		spin_unlock_irqrestore(&priv->lock, flags);
824 	}
825 
826 	netif_tx_unlock(dev);
827 }
828 
829 int ipoib_cm_dev_open(struct net_device *dev)
830 {
831 	struct ipoib_dev_priv *priv = netdev_priv(dev);
832 	int ret;
833 
834 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
835 		return 0;
836 
837 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
838 	if (IS_ERR(priv->cm.id)) {
839 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
840 		ret = PTR_ERR(priv->cm.id);
841 		goto err_cm;
842 	}
843 
844 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
845 			   0, NULL);
846 	if (ret) {
847 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
848 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
849 		goto err_listen;
850 	}
851 
852 	return 0;
853 
854 err_listen:
855 	ib_destroy_cm_id(priv->cm.id);
856 err_cm:
857 	priv->cm.id = NULL;
858 	return ret;
859 }
860 
861 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
862 {
863 	struct ipoib_dev_priv *priv = netdev_priv(dev);
864 	struct ipoib_cm_rx *rx, *n;
865 	LIST_HEAD(list);
866 
867 	spin_lock_irq(&priv->lock);
868 	list_splice_init(&priv->cm.rx_reap_list, &list);
869 	spin_unlock_irq(&priv->lock);
870 
871 	list_for_each_entry_safe(rx, n, &list, list) {
872 		ib_destroy_cm_id(rx->id);
873 		ib_destroy_qp(rx->qp);
874 		if (!ipoib_cm_has_srq(dev)) {
875 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
876 			spin_lock_irq(&priv->lock);
877 			--priv->cm.nonsrq_conn_qp;
878 			spin_unlock_irq(&priv->lock);
879 		}
880 		kfree(rx);
881 	}
882 }
883 
884 void ipoib_cm_dev_stop(struct net_device *dev)
885 {
886 	struct ipoib_dev_priv *priv = netdev_priv(dev);
887 	struct ipoib_cm_rx *p;
888 	unsigned long begin;
889 	int ret;
890 
891 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
892 		return;
893 
894 	ib_destroy_cm_id(priv->cm.id);
895 	priv->cm.id = NULL;
896 
897 	spin_lock_irq(&priv->lock);
898 	while (!list_empty(&priv->cm.passive_ids)) {
899 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
900 		list_move(&p->list, &priv->cm.rx_error_list);
901 		p->state = IPOIB_CM_RX_ERROR;
902 		spin_unlock_irq(&priv->lock);
903 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
904 		if (ret)
905 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
906 		spin_lock_irq(&priv->lock);
907 	}
908 
909 	/* Wait for all RX to be drained */
910 	begin = jiffies;
911 
912 	while (!list_empty(&priv->cm.rx_error_list) ||
913 	       !list_empty(&priv->cm.rx_flush_list) ||
914 	       !list_empty(&priv->cm.rx_drain_list)) {
915 		if (time_after(jiffies, begin + 5 * HZ)) {
916 			ipoib_warn(priv, "RX drain timing out\n");
917 
918 			/*
919 			 * assume the HW is wedged and just free up everything.
920 			 */
921 			list_splice_init(&priv->cm.rx_flush_list,
922 					 &priv->cm.rx_reap_list);
923 			list_splice_init(&priv->cm.rx_error_list,
924 					 &priv->cm.rx_reap_list);
925 			list_splice_init(&priv->cm.rx_drain_list,
926 					 &priv->cm.rx_reap_list);
927 			break;
928 		}
929 		spin_unlock_irq(&priv->lock);
930 		msleep(1);
931 		ipoib_drain_cq(dev);
932 		spin_lock_irq(&priv->lock);
933 	}
934 
935 	spin_unlock_irq(&priv->lock);
936 
937 	ipoib_cm_free_rx_reap_list(dev);
938 
939 	cancel_delayed_work(&priv->cm.stale_task);
940 }
941 
942 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
943 {
944 	struct ipoib_cm_tx *p = cm_id->context;
945 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
946 	struct ipoib_cm_data *data = event->private_data;
947 	struct sk_buff_head skqueue;
948 	struct ib_qp_attr qp_attr;
949 	int qp_attr_mask, ret;
950 	struct sk_buff *skb;
951 
952 	p->mtu = be32_to_cpu(data->mtu);
953 
954 	if (p->mtu <= IPOIB_ENCAP_LEN) {
955 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
956 			   p->mtu, IPOIB_ENCAP_LEN);
957 		return -EINVAL;
958 	}
959 
960 	qp_attr.qp_state = IB_QPS_RTR;
961 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
962 	if (ret) {
963 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
964 		return ret;
965 	}
966 
967 	qp_attr.rq_psn = 0 /* FIXME */;
968 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
969 	if (ret) {
970 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
971 		return ret;
972 	}
973 
974 	qp_attr.qp_state = IB_QPS_RTS;
975 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
976 	if (ret) {
977 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
978 		return ret;
979 	}
980 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
981 	if (ret) {
982 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
983 		return ret;
984 	}
985 
986 	skb_queue_head_init(&skqueue);
987 
988 	spin_lock_irq(&priv->lock);
989 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
990 	if (p->neigh)
991 		while ((skb = __skb_dequeue(&p->neigh->queue)))
992 			__skb_queue_tail(&skqueue, skb);
993 	spin_unlock_irq(&priv->lock);
994 
995 	while ((skb = __skb_dequeue(&skqueue))) {
996 		skb->dev = p->dev;
997 		if (dev_queue_xmit(skb))
998 			ipoib_warn(priv, "dev_queue_xmit failed "
999 				   "to requeue packet\n");
1000 	}
1001 
1002 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1003 	if (ret) {
1004 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1005 		return ret;
1006 	}
1007 	return 0;
1008 }
1009 
1010 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1011 {
1012 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1013 	struct ib_qp_init_attr attr = {
1014 		.send_cq		= priv->recv_cq,
1015 		.recv_cq		= priv->recv_cq,
1016 		.srq			= priv->cm.srq,
1017 		.cap.max_send_wr	= ipoib_sendq_size,
1018 		.cap.max_send_sge	= 1,
1019 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1020 		.qp_type		= IB_QPT_RC,
1021 		.qp_context		= tx
1022 	};
1023 
1024 	return ib_create_qp(priv->pd, &attr);
1025 }
1026 
1027 static int ipoib_cm_send_req(struct net_device *dev,
1028 			     struct ib_cm_id *id, struct ib_qp *qp,
1029 			     u32 qpn,
1030 			     struct ib_sa_path_rec *pathrec)
1031 {
1032 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1033 	struct ipoib_cm_data data = {};
1034 	struct ib_cm_req_param req = {};
1035 
1036 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1037 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1038 
1039 	req.primary_path		= pathrec;
1040 	req.alternate_path		= NULL;
1041 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1042 	req.qp_num			= qp->qp_num;
1043 	req.qp_type			= qp->qp_type;
1044 	req.private_data		= &data;
1045 	req.private_data_len		= sizeof data;
1046 	req.flow_control		= 0;
1047 
1048 	req.starting_psn		= 0; /* FIXME */
1049 
1050 	/*
1051 	 * Pick some arbitrary defaults here; we could make these
1052 	 * module parameters if anyone cared about setting them.
1053 	 */
1054 	req.responder_resources		= 4;
1055 	req.remote_cm_response_timeout	= 20;
1056 	req.local_cm_response_timeout	= 20;
1057 	req.retry_count			= 0; /* RFC draft warns against retries */
1058 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1059 	req.max_cm_retries		= 15;
1060 	req.srq				= ipoib_cm_has_srq(dev);
1061 	return ib_send_cm_req(id, &req);
1062 }
1063 
1064 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1065 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1066 {
1067 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1068 	struct ib_qp_attr qp_attr;
1069 	int qp_attr_mask, ret;
1070 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1071 	if (ret) {
1072 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1073 		return ret;
1074 	}
1075 
1076 	qp_attr.qp_state = IB_QPS_INIT;
1077 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1078 	qp_attr.port_num = priv->port;
1079 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1080 
1081 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1082 	if (ret) {
1083 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1084 		return ret;
1085 	}
1086 	return 0;
1087 }
1088 
1089 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1090 			    struct ib_sa_path_rec *pathrec)
1091 {
1092 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1093 	int ret;
1094 
1095 	p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1096 	if (!p->tx_ring) {
1097 		ipoib_warn(priv, "failed to allocate tx ring\n");
1098 		ret = -ENOMEM;
1099 		goto err_tx;
1100 	}
1101 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1102 
1103 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1104 	if (IS_ERR(p->qp)) {
1105 		ret = PTR_ERR(p->qp);
1106 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1107 		goto err_qp;
1108 	}
1109 
1110 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1111 	if (IS_ERR(p->id)) {
1112 		ret = PTR_ERR(p->id);
1113 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1114 		goto err_id;
1115 	}
1116 
1117 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1118 	if (ret) {
1119 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1120 		goto err_modify;
1121 	}
1122 
1123 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1124 	if (ret) {
1125 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1126 		goto err_send_cm;
1127 	}
1128 
1129 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1130 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1131 
1132 	return 0;
1133 
1134 err_send_cm:
1135 err_modify:
1136 	ib_destroy_cm_id(p->id);
1137 err_id:
1138 	p->id = NULL;
1139 	ib_destroy_qp(p->qp);
1140 err_qp:
1141 	p->qp = NULL;
1142 	vfree(p->tx_ring);
1143 err_tx:
1144 	return ret;
1145 }
1146 
1147 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1148 {
1149 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1150 	struct ipoib_cm_tx_buf *tx_req;
1151 	unsigned long begin;
1152 
1153 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1154 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1155 
1156 	if (p->id)
1157 		ib_destroy_cm_id(p->id);
1158 
1159 	if (p->tx_ring) {
1160 		/* Wait for all sends to complete */
1161 		begin = jiffies;
1162 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1163 			if (time_after(jiffies, begin + 5 * HZ)) {
1164 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1165 					   p->tx_head - p->tx_tail);
1166 				goto timeout;
1167 			}
1168 
1169 			msleep(1);
1170 		}
1171 	}
1172 
1173 timeout:
1174 
1175 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1176 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1177 		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1178 				    DMA_TO_DEVICE);
1179 		dev_kfree_skb_any(tx_req->skb);
1180 		++p->tx_tail;
1181 		netif_tx_lock_bh(p->dev);
1182 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1183 		    netif_queue_stopped(p->dev) &&
1184 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1185 			netif_wake_queue(p->dev);
1186 		netif_tx_unlock_bh(p->dev);
1187 	}
1188 
1189 	if (p->qp)
1190 		ib_destroy_qp(p->qp);
1191 
1192 	vfree(p->tx_ring);
1193 	kfree(p);
1194 }
1195 
1196 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1197 			       struct ib_cm_event *event)
1198 {
1199 	struct ipoib_cm_tx *tx = cm_id->context;
1200 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1201 	struct net_device *dev = priv->dev;
1202 	struct ipoib_neigh *neigh;
1203 	unsigned long flags;
1204 	int ret;
1205 
1206 	switch (event->event) {
1207 	case IB_CM_DREQ_RECEIVED:
1208 		ipoib_dbg(priv, "DREQ received.\n");
1209 		ib_send_cm_drep(cm_id, NULL, 0);
1210 		break;
1211 	case IB_CM_REP_RECEIVED:
1212 		ipoib_dbg(priv, "REP received.\n");
1213 		ret = ipoib_cm_rep_handler(cm_id, event);
1214 		if (ret)
1215 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1216 				       NULL, 0, NULL, 0);
1217 		break;
1218 	case IB_CM_REQ_ERROR:
1219 	case IB_CM_REJ_RECEIVED:
1220 	case IB_CM_TIMEWAIT_EXIT:
1221 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1222 		netif_tx_lock_bh(dev);
1223 		spin_lock_irqsave(&priv->lock, flags);
1224 		neigh = tx->neigh;
1225 
1226 		if (neigh) {
1227 			neigh->cm = NULL;
1228 			list_del(&neigh->list);
1229 			if (neigh->ah)
1230 				ipoib_put_ah(neigh->ah);
1231 			ipoib_neigh_free(dev, neigh);
1232 
1233 			tx->neigh = NULL;
1234 		}
1235 
1236 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1237 			list_move(&tx->list, &priv->cm.reap_list);
1238 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1239 		}
1240 
1241 		spin_unlock_irqrestore(&priv->lock, flags);
1242 		netif_tx_unlock_bh(dev);
1243 		break;
1244 	default:
1245 		break;
1246 	}
1247 
1248 	return 0;
1249 }
1250 
1251 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1252 				       struct ipoib_neigh *neigh)
1253 {
1254 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1255 	struct ipoib_cm_tx *tx;
1256 
1257 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1258 	if (!tx)
1259 		return NULL;
1260 
1261 	neigh->cm = tx;
1262 	tx->neigh = neigh;
1263 	tx->path = path;
1264 	tx->dev = dev;
1265 	list_add(&tx->list, &priv->cm.start_list);
1266 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1267 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1268 	return tx;
1269 }
1270 
1271 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1272 {
1273 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1274 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1275 		list_move(&tx->list, &priv->cm.reap_list);
1276 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1277 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1278 			  tx->neigh->dgid.raw);
1279 		tx->neigh = NULL;
1280 	}
1281 }
1282 
1283 static void ipoib_cm_tx_start(struct work_struct *work)
1284 {
1285 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1286 						   cm.start_task);
1287 	struct net_device *dev = priv->dev;
1288 	struct ipoib_neigh *neigh;
1289 	struct ipoib_cm_tx *p;
1290 	unsigned long flags;
1291 	int ret;
1292 
1293 	struct ib_sa_path_rec pathrec;
1294 	u32 qpn;
1295 
1296 	netif_tx_lock_bh(dev);
1297 	spin_lock_irqsave(&priv->lock, flags);
1298 
1299 	while (!list_empty(&priv->cm.start_list)) {
1300 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1301 		list_del_init(&p->list);
1302 		neigh = p->neigh;
1303 		qpn = IPOIB_QPN(neigh->neighbour->ha);
1304 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1305 
1306 		spin_unlock_irqrestore(&priv->lock, flags);
1307 		netif_tx_unlock_bh(dev);
1308 
1309 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1310 
1311 		netif_tx_lock_bh(dev);
1312 		spin_lock_irqsave(&priv->lock, flags);
1313 
1314 		if (ret) {
1315 			neigh = p->neigh;
1316 			if (neigh) {
1317 				neigh->cm = NULL;
1318 				list_del(&neigh->list);
1319 				if (neigh->ah)
1320 					ipoib_put_ah(neigh->ah);
1321 				ipoib_neigh_free(dev, neigh);
1322 			}
1323 			list_del(&p->list);
1324 			kfree(p);
1325 		}
1326 	}
1327 
1328 	spin_unlock_irqrestore(&priv->lock, flags);
1329 	netif_tx_unlock_bh(dev);
1330 }
1331 
1332 static void ipoib_cm_tx_reap(struct work_struct *work)
1333 {
1334 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1335 						   cm.reap_task);
1336 	struct net_device *dev = priv->dev;
1337 	struct ipoib_cm_tx *p;
1338 	unsigned long flags;
1339 
1340 	netif_tx_lock_bh(dev);
1341 	spin_lock_irqsave(&priv->lock, flags);
1342 
1343 	while (!list_empty(&priv->cm.reap_list)) {
1344 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1345 		list_del(&p->list);
1346 		spin_unlock_irqrestore(&priv->lock, flags);
1347 		netif_tx_unlock_bh(dev);
1348 		ipoib_cm_tx_destroy(p);
1349 		netif_tx_lock_bh(dev);
1350 		spin_lock_irqsave(&priv->lock, flags);
1351 	}
1352 
1353 	spin_unlock_irqrestore(&priv->lock, flags);
1354 	netif_tx_unlock_bh(dev);
1355 }
1356 
1357 static void ipoib_cm_skb_reap(struct work_struct *work)
1358 {
1359 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1360 						   cm.skb_task);
1361 	struct net_device *dev = priv->dev;
1362 	struct sk_buff *skb;
1363 	unsigned long flags;
1364 	unsigned mtu = priv->mcast_mtu;
1365 
1366 	netif_tx_lock_bh(dev);
1367 	spin_lock_irqsave(&priv->lock, flags);
1368 
1369 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1370 		spin_unlock_irqrestore(&priv->lock, flags);
1371 		netif_tx_unlock_bh(dev);
1372 
1373 		if (skb->protocol == htons(ETH_P_IP))
1374 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1375 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1376 		else if (skb->protocol == htons(ETH_P_IPV6))
1377 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1378 #endif
1379 		dev_kfree_skb_any(skb);
1380 
1381 		netif_tx_lock_bh(dev);
1382 		spin_lock_irqsave(&priv->lock, flags);
1383 	}
1384 
1385 	spin_unlock_irqrestore(&priv->lock, flags);
1386 	netif_tx_unlock_bh(dev);
1387 }
1388 
1389 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1390 			   unsigned int mtu)
1391 {
1392 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1393 	int e = skb_queue_empty(&priv->cm.skb_queue);
1394 
1395 	if (skb_dst(skb))
1396 		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1397 
1398 	skb_queue_tail(&priv->cm.skb_queue, skb);
1399 	if (e)
1400 		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1401 }
1402 
1403 static void ipoib_cm_rx_reap(struct work_struct *work)
1404 {
1405 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1406 						cm.rx_reap_task)->dev);
1407 }
1408 
1409 static void ipoib_cm_stale_task(struct work_struct *work)
1410 {
1411 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1412 						   cm.stale_task.work);
1413 	struct ipoib_cm_rx *p;
1414 	int ret;
1415 
1416 	spin_lock_irq(&priv->lock);
1417 	while (!list_empty(&priv->cm.passive_ids)) {
1418 		/* List is sorted by LRU, start from tail,
1419 		 * stop when we see a recently used entry */
1420 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1421 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1422 			break;
1423 		list_move(&p->list, &priv->cm.rx_error_list);
1424 		p->state = IPOIB_CM_RX_ERROR;
1425 		spin_unlock_irq(&priv->lock);
1426 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1427 		if (ret)
1428 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1429 		spin_lock_irq(&priv->lock);
1430 	}
1431 
1432 	if (!list_empty(&priv->cm.passive_ids))
1433 		queue_delayed_work(ipoib_workqueue,
1434 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1435 	spin_unlock_irq(&priv->lock);
1436 }
1437 
1438 
1439 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1440 			 char *buf)
1441 {
1442 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1443 
1444 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1445 		return sprintf(buf, "connected\n");
1446 	else
1447 		return sprintf(buf, "datagram\n");
1448 }
1449 
1450 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1451 			const char *buf, size_t count)
1452 {
1453 	struct net_device *dev = to_net_dev(d);
1454 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1455 
1456 	if (!rtnl_trylock())
1457 		return restart_syscall();
1458 
1459 	/* flush paths if we switch modes so that connections are restarted */
1460 	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1461 		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1462 		ipoib_warn(priv, "enabling connected mode "
1463 			   "will cause multicast packet drops\n");
1464 
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 		if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1477 			dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1478 			if (priv->hca_caps & IB_DEVICE_UD_TSO)
1479 				dev->features |= NETIF_F_TSO;
1480 		}
1481 		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1482 		rtnl_unlock();
1483 		ipoib_flush_paths(dev);
1484 
1485 		return count;
1486 	}
1487 	rtnl_unlock();
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