1 /*
2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_cm.h>
34 #include <net/dst.h>
35 #include <net/icmp.h>
36 #include <linux/icmpv6.h>
37 #include <linux/delay.h>
38 #include <linux/slab.h>
39 #include <linux/vmalloc.h>
40 #include <linux/moduleparam.h>
41 
42 #include "ipoib.h"
43 
44 int ipoib_max_conn_qp = 128;
45 
46 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
47 MODULE_PARM_DESC(max_nonsrq_conn_qp,
48 		 "Max number of connected-mode QPs per interface "
49 		 "(applied only if shared receive queue is not available)");
50 
51 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
52 static int data_debug_level;
53 
54 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
55 MODULE_PARM_DESC(cm_data_debug_level,
56 		 "Enable data path debug tracing for connected mode if > 0");
57 #endif
58 
59 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
60 
61 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
62 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
63 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
64 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
65 
66 #define IPOIB_CM_RX_RESERVE     (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
67 
68 static struct ib_qp_attr ipoib_cm_err_attr = {
69 	.qp_state = IB_QPS_ERR
70 };
71 
72 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
73 
74 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
75 	.opcode = IB_WR_SEND,
76 };
77 
78 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
79 			       struct ib_cm_event *event);
80 
81 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
82 				  u64 mapping[IPOIB_CM_RX_SG])
83 {
84 	int i;
85 
86 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
87 
88 	for (i = 0; i < frags; ++i)
89 		ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
90 }
91 
92 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
93 {
94 	struct ipoib_dev_priv *priv = netdev_priv(dev);
95 	struct ib_recv_wr *bad_wr;
96 	int i, ret;
97 
98 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
99 
100 	for (i = 0; i < priv->cm.num_frags; ++i)
101 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
102 
103 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
104 	if (unlikely(ret)) {
105 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
106 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
107 				      priv->cm.srq_ring[id].mapping);
108 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
109 		priv->cm.srq_ring[id].skb = NULL;
110 	}
111 
112 	return ret;
113 }
114 
115 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
116 					struct ipoib_cm_rx *rx,
117 					struct ib_recv_wr *wr,
118 					struct ib_sge *sge, int id)
119 {
120 	struct ipoib_dev_priv *priv = netdev_priv(dev);
121 	struct ib_recv_wr *bad_wr;
122 	int i, ret;
123 
124 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
125 
126 	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
127 		sge[i].addr = rx->rx_ring[id].mapping[i];
128 
129 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
130 	if (unlikely(ret)) {
131 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
132 		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
133 				      rx->rx_ring[id].mapping);
134 		dev_kfree_skb_any(rx->rx_ring[id].skb);
135 		rx->rx_ring[id].skb = NULL;
136 	}
137 
138 	return ret;
139 }
140 
141 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
142 					     struct ipoib_cm_rx_buf *rx_ring,
143 					     int id, int frags,
144 					     u64 mapping[IPOIB_CM_RX_SG],
145 					     gfp_t gfp)
146 {
147 	struct ipoib_dev_priv *priv = netdev_priv(dev);
148 	struct sk_buff *skb;
149 	int i;
150 
151 	skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
152 	if (unlikely(!skb))
153 		return NULL;
154 
155 	/*
156 	 * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
157 	 * IP header to a multiple of 16.
158 	 */
159 	skb_reserve(skb, IPOIB_CM_RX_RESERVE);
160 
161 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
162 				       DMA_FROM_DEVICE);
163 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
164 		dev_kfree_skb_any(skb);
165 		return NULL;
166 	}
167 
168 	for (i = 0; i < frags; i++) {
169 		struct page *page = alloc_page(gfp);
170 
171 		if (!page)
172 			goto partial_error;
173 		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
174 
175 		mapping[i + 1] = ib_dma_map_page(priv->ca, page,
176 						 0, PAGE_SIZE, DMA_FROM_DEVICE);
177 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
178 			goto partial_error;
179 	}
180 
181 	rx_ring[id].skb = skb;
182 	return skb;
183 
184 partial_error:
185 
186 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
187 
188 	for (; i > 0; --i)
189 		ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
190 
191 	dev_kfree_skb_any(skb);
192 	return NULL;
193 }
194 
195 static void ipoib_cm_free_rx_ring(struct net_device *dev,
196 				  struct ipoib_cm_rx_buf *rx_ring)
197 {
198 	struct ipoib_dev_priv *priv = netdev_priv(dev);
199 	int i;
200 
201 	for (i = 0; i < ipoib_recvq_size; ++i)
202 		if (rx_ring[i].skb) {
203 			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
204 					      rx_ring[i].mapping);
205 			dev_kfree_skb_any(rx_ring[i].skb);
206 		}
207 
208 	vfree(rx_ring);
209 }
210 
211 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
212 {
213 	struct ib_send_wr *bad_wr;
214 	struct ipoib_cm_rx *p;
215 
216 	/* We only reserved 1 extra slot in CQ for drain WRs, so
217 	 * make sure we have at most 1 outstanding WR. */
218 	if (list_empty(&priv->cm.rx_flush_list) ||
219 	    !list_empty(&priv->cm.rx_drain_list))
220 		return;
221 
222 	/*
223 	 * QPs on flush list are error state.  This way, a "flush
224 	 * error" WC will be immediately generated for each WR we post.
225 	 */
226 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
227 	ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
228 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
229 		ipoib_warn(priv, "failed to post drain wr\n");
230 
231 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
232 }
233 
234 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
235 {
236 	struct ipoib_cm_rx *p = ctx;
237 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
238 	unsigned long flags;
239 
240 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
241 		return;
242 
243 	spin_lock_irqsave(&priv->lock, flags);
244 	list_move(&p->list, &priv->cm.rx_flush_list);
245 	p->state = IPOIB_CM_RX_FLUSH;
246 	ipoib_cm_start_rx_drain(priv);
247 	spin_unlock_irqrestore(&priv->lock, flags);
248 }
249 
250 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
251 					   struct ipoib_cm_rx *p)
252 {
253 	struct ipoib_dev_priv *priv = netdev_priv(dev);
254 	struct ib_qp_init_attr attr = {
255 		.event_handler = ipoib_cm_rx_event_handler,
256 		.send_cq = priv->recv_cq, /* For drain WR */
257 		.recv_cq = priv->recv_cq,
258 		.srq = priv->cm.srq,
259 		.cap.max_send_wr = 1, /* For drain WR */
260 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
261 		.sq_sig_type = IB_SIGNAL_ALL_WR,
262 		.qp_type = IB_QPT_RC,
263 		.qp_context = p,
264 	};
265 
266 	if (!ipoib_cm_has_srq(dev)) {
267 		attr.cap.max_recv_wr  = ipoib_recvq_size;
268 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
269 	}
270 
271 	return ib_create_qp(priv->pd, &attr);
272 }
273 
274 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
275 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
276 				 unsigned psn)
277 {
278 	struct ipoib_dev_priv *priv = netdev_priv(dev);
279 	struct ib_qp_attr qp_attr;
280 	int qp_attr_mask, ret;
281 
282 	qp_attr.qp_state = IB_QPS_INIT;
283 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
284 	if (ret) {
285 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
286 		return ret;
287 	}
288 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
289 	if (ret) {
290 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
291 		return ret;
292 	}
293 	qp_attr.qp_state = IB_QPS_RTR;
294 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
295 	if (ret) {
296 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
297 		return ret;
298 	}
299 	qp_attr.rq_psn = psn;
300 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
301 	if (ret) {
302 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
303 		return ret;
304 	}
305 
306 	/*
307 	 * Current Mellanox HCA firmware won't generate completions
308 	 * with error for drain WRs unless the QP has been moved to
309 	 * RTS first. This work-around leaves a window where a QP has
310 	 * moved to error asynchronously, but this will eventually get
311 	 * fixed in firmware, so let's not error out if modify QP
312 	 * fails.
313 	 */
314 	qp_attr.qp_state = IB_QPS_RTS;
315 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
316 	if (ret) {
317 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
318 		return 0;
319 	}
320 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
321 	if (ret) {
322 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
323 		return 0;
324 	}
325 
326 	return 0;
327 }
328 
329 static void ipoib_cm_init_rx_wr(struct net_device *dev,
330 				struct ib_recv_wr *wr,
331 				struct ib_sge *sge)
332 {
333 	struct ipoib_dev_priv *priv = netdev_priv(dev);
334 	int i;
335 
336 	for (i = 0; i < priv->cm.num_frags; ++i)
337 		sge[i].lkey = priv->pd->local_dma_lkey;
338 
339 	sge[0].length = IPOIB_CM_HEAD_SIZE;
340 	for (i = 1; i < priv->cm.num_frags; ++i)
341 		sge[i].length = PAGE_SIZE;
342 
343 	wr->next    = NULL;
344 	wr->sg_list = sge;
345 	wr->num_sge = priv->cm.num_frags;
346 }
347 
348 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
349 				   struct ipoib_cm_rx *rx)
350 {
351 	struct ipoib_dev_priv *priv = netdev_priv(dev);
352 	struct {
353 		struct ib_recv_wr wr;
354 		struct ib_sge sge[IPOIB_CM_RX_SG];
355 	} *t;
356 	int ret;
357 	int i;
358 
359 	rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
360 	if (!rx->rx_ring) {
361 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
362 		       priv->ca->name, ipoib_recvq_size);
363 		return -ENOMEM;
364 	}
365 
366 	t = kmalloc(sizeof *t, GFP_KERNEL);
367 	if (!t) {
368 		ret = -ENOMEM;
369 		goto err_free;
370 	}
371 
372 	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
373 
374 	spin_lock_irq(&priv->lock);
375 
376 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
377 		spin_unlock_irq(&priv->lock);
378 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
379 		ret = -EINVAL;
380 		goto err_free;
381 	} else
382 		++priv->cm.nonsrq_conn_qp;
383 
384 	spin_unlock_irq(&priv->lock);
385 
386 	for (i = 0; i < ipoib_recvq_size; ++i) {
387 		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
388 					   rx->rx_ring[i].mapping,
389 					   GFP_KERNEL)) {
390 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
391 			ret = -ENOMEM;
392 			goto err_count;
393 		}
394 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
395 		if (ret) {
396 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
397 				   "failed for buf %d\n", i);
398 			ret = -EIO;
399 			goto err_count;
400 		}
401 	}
402 
403 	rx->recv_count = ipoib_recvq_size;
404 
405 	kfree(t);
406 
407 	return 0;
408 
409 err_count:
410 	spin_lock_irq(&priv->lock);
411 	--priv->cm.nonsrq_conn_qp;
412 	spin_unlock_irq(&priv->lock);
413 
414 err_free:
415 	kfree(t);
416 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
417 
418 	return ret;
419 }
420 
421 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
422 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
423 			     unsigned psn)
424 {
425 	struct ipoib_dev_priv *priv = netdev_priv(dev);
426 	struct ipoib_cm_data data = {};
427 	struct ib_cm_rep_param rep = {};
428 
429 	data.qpn = cpu_to_be32(priv->qp->qp_num);
430 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
431 
432 	rep.private_data = &data;
433 	rep.private_data_len = sizeof data;
434 	rep.flow_control = 0;
435 	rep.rnr_retry_count = req->rnr_retry_count;
436 	rep.srq = ipoib_cm_has_srq(dev);
437 	rep.qp_num = qp->qp_num;
438 	rep.starting_psn = psn;
439 	return ib_send_cm_rep(cm_id, &rep);
440 }
441 
442 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
443 {
444 	struct net_device *dev = cm_id->context;
445 	struct ipoib_dev_priv *priv = netdev_priv(dev);
446 	struct ipoib_cm_rx *p;
447 	unsigned psn;
448 	int ret;
449 
450 	ipoib_dbg(priv, "REQ arrived\n");
451 	p = kzalloc(sizeof *p, GFP_KERNEL);
452 	if (!p)
453 		return -ENOMEM;
454 	p->dev = dev;
455 	p->id = cm_id;
456 	cm_id->context = p;
457 	p->state = IPOIB_CM_RX_LIVE;
458 	p->jiffies = jiffies;
459 	INIT_LIST_HEAD(&p->list);
460 
461 	p->qp = ipoib_cm_create_rx_qp(dev, p);
462 	if (IS_ERR(p->qp)) {
463 		ret = PTR_ERR(p->qp);
464 		goto err_qp;
465 	}
466 
467 	psn = prandom_u32() & 0xffffff;
468 	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
469 	if (ret)
470 		goto err_modify;
471 
472 	if (!ipoib_cm_has_srq(dev)) {
473 		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
474 		if (ret)
475 			goto err_modify;
476 	}
477 
478 	spin_lock_irq(&priv->lock);
479 	queue_delayed_work(priv->wq,
480 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
481 	/* Add this entry to passive ids list head, but do not re-add it
482 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
483 	p->jiffies = jiffies;
484 	if (p->state == IPOIB_CM_RX_LIVE)
485 		list_move(&p->list, &priv->cm.passive_ids);
486 	spin_unlock_irq(&priv->lock);
487 
488 	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
489 	if (ret) {
490 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
491 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
492 			ipoib_warn(priv, "unable to move qp to error state\n");
493 	}
494 	return 0;
495 
496 err_modify:
497 	ib_destroy_qp(p->qp);
498 err_qp:
499 	kfree(p);
500 	return ret;
501 }
502 
503 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
504 			       struct ib_cm_event *event)
505 {
506 	struct ipoib_cm_rx *p;
507 	struct ipoib_dev_priv *priv;
508 
509 	switch (event->event) {
510 	case IB_CM_REQ_RECEIVED:
511 		return ipoib_cm_req_handler(cm_id, event);
512 	case IB_CM_DREQ_RECEIVED:
513 		p = cm_id->context;
514 		ib_send_cm_drep(cm_id, NULL, 0);
515 		/* Fall through */
516 	case IB_CM_REJ_RECEIVED:
517 		p = cm_id->context;
518 		priv = netdev_priv(p->dev);
519 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
520 			ipoib_warn(priv, "unable to move qp to error state\n");
521 		/* Fall through */
522 	default:
523 		return 0;
524 	}
525 }
526 /* Adjust length of skb with fragments to match received data */
527 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
528 			  unsigned int length, struct sk_buff *toskb)
529 {
530 	int i, num_frags;
531 	unsigned int size;
532 
533 	/* put header into skb */
534 	size = min(length, hdr_space);
535 	skb->tail += size;
536 	skb->len += size;
537 	length -= size;
538 
539 	num_frags = skb_shinfo(skb)->nr_frags;
540 	for (i = 0; i < num_frags; i++) {
541 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
542 
543 		if (length == 0) {
544 			/* don't need this page */
545 			skb_fill_page_desc(toskb, i, skb_frag_page(frag),
546 					   0, PAGE_SIZE);
547 			--skb_shinfo(skb)->nr_frags;
548 		} else {
549 			size = min(length, (unsigned) PAGE_SIZE);
550 
551 			skb_frag_size_set(frag, size);
552 			skb->data_len += size;
553 			skb->truesize += size;
554 			skb->len += size;
555 			length -= size;
556 		}
557 	}
558 }
559 
560 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
561 {
562 	struct ipoib_dev_priv *priv = netdev_priv(dev);
563 	struct ipoib_cm_rx_buf *rx_ring;
564 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
565 	struct sk_buff *skb, *newskb;
566 	struct ipoib_cm_rx *p;
567 	unsigned long flags;
568 	u64 mapping[IPOIB_CM_RX_SG];
569 	int frags;
570 	int has_srq;
571 	struct sk_buff *small_skb;
572 
573 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
574 		       wr_id, wc->status);
575 
576 	if (unlikely(wr_id >= ipoib_recvq_size)) {
577 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
578 			spin_lock_irqsave(&priv->lock, flags);
579 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
580 			ipoib_cm_start_rx_drain(priv);
581 			queue_work(priv->wq, &priv->cm.rx_reap_task);
582 			spin_unlock_irqrestore(&priv->lock, flags);
583 		} else
584 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
585 				   wr_id, ipoib_recvq_size);
586 		return;
587 	}
588 
589 	p = wc->qp->qp_context;
590 
591 	has_srq = ipoib_cm_has_srq(dev);
592 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
593 
594 	skb = rx_ring[wr_id].skb;
595 
596 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
597 		ipoib_dbg(priv, "cm recv error "
598 			   "(status=%d, wrid=%d vend_err %x)\n",
599 			   wc->status, wr_id, wc->vendor_err);
600 		++dev->stats.rx_dropped;
601 		if (has_srq)
602 			goto repost;
603 		else {
604 			if (!--p->recv_count) {
605 				spin_lock_irqsave(&priv->lock, flags);
606 				list_move(&p->list, &priv->cm.rx_reap_list);
607 				spin_unlock_irqrestore(&priv->lock, flags);
608 				queue_work(priv->wq, &priv->cm.rx_reap_task);
609 			}
610 			return;
611 		}
612 	}
613 
614 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
615 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
616 			spin_lock_irqsave(&priv->lock, flags);
617 			p->jiffies = jiffies;
618 			/* Move this entry to list head, but do not re-add it
619 			 * if it has been moved out of list. */
620 			if (p->state == IPOIB_CM_RX_LIVE)
621 				list_move(&p->list, &priv->cm.passive_ids);
622 			spin_unlock_irqrestore(&priv->lock, flags);
623 		}
624 	}
625 
626 	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
627 		int dlen = wc->byte_len;
628 
629 		small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
630 		if (small_skb) {
631 			skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
632 			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
633 						   dlen, DMA_FROM_DEVICE);
634 			skb_copy_from_linear_data(skb, small_skb->data, dlen);
635 			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
636 						      dlen, DMA_FROM_DEVICE);
637 			skb_put(small_skb, dlen);
638 			skb = small_skb;
639 			goto copied;
640 		}
641 	}
642 
643 	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
644 					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
645 
646 	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
647 				       mapping, GFP_ATOMIC);
648 	if (unlikely(!newskb)) {
649 		/*
650 		 * If we can't allocate a new RX buffer, dump
651 		 * this packet and reuse the old buffer.
652 		 */
653 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
654 		++dev->stats.rx_dropped;
655 		goto repost;
656 	}
657 
658 	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
659 	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
660 
661 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
662 		       wc->byte_len, wc->slid);
663 
664 	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
665 
666 copied:
667 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
668 	skb_add_pseudo_hdr(skb);
669 
670 	++dev->stats.rx_packets;
671 	dev->stats.rx_bytes += skb->len;
672 
673 	skb->dev = dev;
674 	/* XXX get correct PACKET_ type here */
675 	skb->pkt_type = PACKET_HOST;
676 	netif_receive_skb(skb);
677 
678 repost:
679 	if (has_srq) {
680 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
681 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
682 				   "for buf %d\n", wr_id);
683 	} else {
684 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
685 							  &priv->cm.rx_wr,
686 							  priv->cm.rx_sge,
687 							  wr_id))) {
688 			--p->recv_count;
689 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
690 				   "for buf %d\n", wr_id);
691 		}
692 	}
693 }
694 
695 static inline int post_send(struct ipoib_dev_priv *priv,
696 			    struct ipoib_cm_tx *tx,
697 			    unsigned int wr_id,
698 			    struct ipoib_tx_buf *tx_req)
699 {
700 	struct ib_send_wr *bad_wr;
701 
702 	ipoib_build_sge(priv, tx_req);
703 
704 	priv->tx_wr.wr.wr_id	= wr_id | IPOIB_OP_CM;
705 
706 	return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
707 }
708 
709 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
710 {
711 	struct ipoib_dev_priv *priv = netdev_priv(dev);
712 	struct ipoib_tx_buf *tx_req;
713 	int rc;
714 	unsigned usable_sge = tx->max_send_sge - !!skb_headlen(skb);
715 
716 	if (unlikely(skb->len > tx->mtu)) {
717 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
718 			   skb->len, tx->mtu);
719 		++dev->stats.tx_dropped;
720 		++dev->stats.tx_errors;
721 		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
722 		return;
723 	}
724 	if (skb_shinfo(skb)->nr_frags > usable_sge) {
725 		if (skb_linearize(skb) < 0) {
726 			ipoib_warn(priv, "skb could not be linearized\n");
727 			++dev->stats.tx_dropped;
728 			++dev->stats.tx_errors;
729 			dev_kfree_skb_any(skb);
730 			return;
731 		}
732 		/* Does skb_linearize return ok without reducing nr_frags? */
733 		if (skb_shinfo(skb)->nr_frags > usable_sge) {
734 			ipoib_warn(priv, "too many frags after skb linearize\n");
735 			++dev->stats.tx_dropped;
736 			++dev->stats.tx_errors;
737 			dev_kfree_skb_any(skb);
738 			return;
739 		}
740 	}
741 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
742 		       tx->tx_head, skb->len, tx->qp->qp_num);
743 
744 	/*
745 	 * We put the skb into the tx_ring _before_ we call post_send()
746 	 * because it's entirely possible that the completion handler will
747 	 * run before we execute anything after the post_send().  That
748 	 * means we have to make sure everything is properly recorded and
749 	 * our state is consistent before we call post_send().
750 	 */
751 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
752 	tx_req->skb = skb;
753 
754 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
755 		++dev->stats.tx_errors;
756 		dev_kfree_skb_any(skb);
757 		return;
758 	}
759 
760 	skb_orphan(skb);
761 	skb_dst_drop(skb);
762 
763 	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
764 	if (unlikely(rc)) {
765 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
766 		++dev->stats.tx_errors;
767 		ipoib_dma_unmap_tx(priv, tx_req);
768 		dev_kfree_skb_any(skb);
769 	} else {
770 		netif_trans_update(dev);
771 		++tx->tx_head;
772 
773 		if (++priv->tx_outstanding == ipoib_sendq_size) {
774 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
775 				  tx->qp->qp_num);
776 			netif_stop_queue(dev);
777 			rc = ib_req_notify_cq(priv->send_cq,
778 				IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
779 			if (rc < 0)
780 				ipoib_warn(priv, "request notify on send CQ failed\n");
781 			else if (rc)
782 				ipoib_send_comp_handler(priv->send_cq, dev);
783 		}
784 	}
785 }
786 
787 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
788 {
789 	struct ipoib_dev_priv *priv = netdev_priv(dev);
790 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
791 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
792 	struct ipoib_tx_buf *tx_req;
793 	unsigned long flags;
794 
795 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
796 		       wr_id, wc->status);
797 
798 	if (unlikely(wr_id >= ipoib_sendq_size)) {
799 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
800 			   wr_id, ipoib_sendq_size);
801 		return;
802 	}
803 
804 	tx_req = &tx->tx_ring[wr_id];
805 
806 	ipoib_dma_unmap_tx(priv, tx_req);
807 
808 	/* FIXME: is this right? Shouldn't we only increment on success? */
809 	++dev->stats.tx_packets;
810 	dev->stats.tx_bytes += tx_req->skb->len;
811 
812 	dev_kfree_skb_any(tx_req->skb);
813 
814 	netif_tx_lock(dev);
815 
816 	++tx->tx_tail;
817 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
818 	    netif_queue_stopped(dev) &&
819 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
820 		netif_wake_queue(dev);
821 
822 	if (wc->status != IB_WC_SUCCESS &&
823 	    wc->status != IB_WC_WR_FLUSH_ERR) {
824 		struct ipoib_neigh *neigh;
825 
826 		ipoib_dbg(priv, "failed cm send event "
827 			   "(status=%d, wrid=%d vend_err %x)\n",
828 			   wc->status, wr_id, wc->vendor_err);
829 
830 		spin_lock_irqsave(&priv->lock, flags);
831 		neigh = tx->neigh;
832 
833 		if (neigh) {
834 			neigh->cm = NULL;
835 			ipoib_neigh_free(neigh);
836 
837 			tx->neigh = NULL;
838 		}
839 
840 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
841 			list_move(&tx->list, &priv->cm.reap_list);
842 			queue_work(priv->wq, &priv->cm.reap_task);
843 		}
844 
845 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
846 
847 		spin_unlock_irqrestore(&priv->lock, flags);
848 	}
849 
850 	netif_tx_unlock(dev);
851 }
852 
853 int ipoib_cm_dev_open(struct net_device *dev)
854 {
855 	struct ipoib_dev_priv *priv = netdev_priv(dev);
856 	int ret;
857 
858 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
859 		return 0;
860 
861 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
862 	if (IS_ERR(priv->cm.id)) {
863 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
864 		ret = PTR_ERR(priv->cm.id);
865 		goto err_cm;
866 	}
867 
868 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
869 			   0);
870 	if (ret) {
871 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
872 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
873 		goto err_listen;
874 	}
875 
876 	return 0;
877 
878 err_listen:
879 	ib_destroy_cm_id(priv->cm.id);
880 err_cm:
881 	priv->cm.id = NULL;
882 	return ret;
883 }
884 
885 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
886 {
887 	struct ipoib_dev_priv *priv = netdev_priv(dev);
888 	struct ipoib_cm_rx *rx, *n;
889 	LIST_HEAD(list);
890 
891 	spin_lock_irq(&priv->lock);
892 	list_splice_init(&priv->cm.rx_reap_list, &list);
893 	spin_unlock_irq(&priv->lock);
894 
895 	list_for_each_entry_safe(rx, n, &list, list) {
896 		ib_destroy_cm_id(rx->id);
897 		ib_destroy_qp(rx->qp);
898 		if (!ipoib_cm_has_srq(dev)) {
899 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
900 			spin_lock_irq(&priv->lock);
901 			--priv->cm.nonsrq_conn_qp;
902 			spin_unlock_irq(&priv->lock);
903 		}
904 		kfree(rx);
905 	}
906 }
907 
908 void ipoib_cm_dev_stop(struct net_device *dev)
909 {
910 	struct ipoib_dev_priv *priv = netdev_priv(dev);
911 	struct ipoib_cm_rx *p;
912 	unsigned long begin;
913 	int ret;
914 
915 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
916 		return;
917 
918 	ib_destroy_cm_id(priv->cm.id);
919 	priv->cm.id = NULL;
920 
921 	spin_lock_irq(&priv->lock);
922 	while (!list_empty(&priv->cm.passive_ids)) {
923 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
924 		list_move(&p->list, &priv->cm.rx_error_list);
925 		p->state = IPOIB_CM_RX_ERROR;
926 		spin_unlock_irq(&priv->lock);
927 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
928 		if (ret)
929 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
930 		spin_lock_irq(&priv->lock);
931 	}
932 
933 	/* Wait for all RX to be drained */
934 	begin = jiffies;
935 
936 	while (!list_empty(&priv->cm.rx_error_list) ||
937 	       !list_empty(&priv->cm.rx_flush_list) ||
938 	       !list_empty(&priv->cm.rx_drain_list)) {
939 		if (time_after(jiffies, begin + 5 * HZ)) {
940 			ipoib_warn(priv, "RX drain timing out\n");
941 
942 			/*
943 			 * assume the HW is wedged and just free up everything.
944 			 */
945 			list_splice_init(&priv->cm.rx_flush_list,
946 					 &priv->cm.rx_reap_list);
947 			list_splice_init(&priv->cm.rx_error_list,
948 					 &priv->cm.rx_reap_list);
949 			list_splice_init(&priv->cm.rx_drain_list,
950 					 &priv->cm.rx_reap_list);
951 			break;
952 		}
953 		spin_unlock_irq(&priv->lock);
954 		msleep(1);
955 		ipoib_drain_cq(dev);
956 		spin_lock_irq(&priv->lock);
957 	}
958 
959 	spin_unlock_irq(&priv->lock);
960 
961 	ipoib_cm_free_rx_reap_list(dev);
962 
963 	cancel_delayed_work(&priv->cm.stale_task);
964 }
965 
966 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
967 {
968 	struct ipoib_cm_tx *p = cm_id->context;
969 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
970 	struct ipoib_cm_data *data = event->private_data;
971 	struct sk_buff_head skqueue;
972 	struct ib_qp_attr qp_attr;
973 	int qp_attr_mask, ret;
974 	struct sk_buff *skb;
975 
976 	p->mtu = be32_to_cpu(data->mtu);
977 
978 	if (p->mtu <= IPOIB_ENCAP_LEN) {
979 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
980 			   p->mtu, IPOIB_ENCAP_LEN);
981 		return -EINVAL;
982 	}
983 
984 	qp_attr.qp_state = IB_QPS_RTR;
985 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
986 	if (ret) {
987 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
988 		return ret;
989 	}
990 
991 	qp_attr.rq_psn = 0 /* FIXME */;
992 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
993 	if (ret) {
994 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
995 		return ret;
996 	}
997 
998 	qp_attr.qp_state = IB_QPS_RTS;
999 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1000 	if (ret) {
1001 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1002 		return ret;
1003 	}
1004 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1005 	if (ret) {
1006 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1007 		return ret;
1008 	}
1009 
1010 	skb_queue_head_init(&skqueue);
1011 
1012 	spin_lock_irq(&priv->lock);
1013 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1014 	if (p->neigh)
1015 		while ((skb = __skb_dequeue(&p->neigh->queue)))
1016 			__skb_queue_tail(&skqueue, skb);
1017 	spin_unlock_irq(&priv->lock);
1018 
1019 	while ((skb = __skb_dequeue(&skqueue))) {
1020 		skb->dev = p->dev;
1021 		if (dev_queue_xmit(skb))
1022 			ipoib_warn(priv, "dev_queue_xmit failed "
1023 				   "to requeue packet\n");
1024 	}
1025 
1026 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1027 	if (ret) {
1028 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1029 		return ret;
1030 	}
1031 	return 0;
1032 }
1033 
1034 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1035 {
1036 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1037 	struct ib_qp_init_attr attr = {
1038 		.send_cq		= priv->recv_cq,
1039 		.recv_cq		= priv->recv_cq,
1040 		.srq			= priv->cm.srq,
1041 		.cap.max_send_wr	= ipoib_sendq_size,
1042 		.cap.max_send_sge	= 1,
1043 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1044 		.qp_type		= IB_QPT_RC,
1045 		.qp_context		= tx,
1046 		.create_flags		= IB_QP_CREATE_USE_GFP_NOIO
1047 	};
1048 
1049 	struct ib_qp *tx_qp;
1050 
1051 	if (dev->features & NETIF_F_SG)
1052 		attr.cap.max_send_sge =
1053 			min_t(u32, priv->ca->attrs.max_sge, MAX_SKB_FRAGS + 1);
1054 
1055 	tx_qp = ib_create_qp(priv->pd, &attr);
1056 	if (PTR_ERR(tx_qp) == -EINVAL) {
1057 		ipoib_warn(priv, "can't use GFP_NOIO for QPs on device %s, using GFP_KERNEL\n",
1058 			   priv->ca->name);
1059 		attr.create_flags &= ~IB_QP_CREATE_USE_GFP_NOIO;
1060 		tx_qp = ib_create_qp(priv->pd, &attr);
1061 	}
1062 	tx->max_send_sge = attr.cap.max_send_sge;
1063 	return tx_qp;
1064 }
1065 
1066 static int ipoib_cm_send_req(struct net_device *dev,
1067 			     struct ib_cm_id *id, struct ib_qp *qp,
1068 			     u32 qpn,
1069 			     struct ib_sa_path_rec *pathrec)
1070 {
1071 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1072 	struct ipoib_cm_data data = {};
1073 	struct ib_cm_req_param req = {};
1074 
1075 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1076 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1077 
1078 	req.primary_path		= pathrec;
1079 	req.alternate_path		= NULL;
1080 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1081 	req.qp_num			= qp->qp_num;
1082 	req.qp_type			= qp->qp_type;
1083 	req.private_data		= &data;
1084 	req.private_data_len		= sizeof data;
1085 	req.flow_control		= 0;
1086 
1087 	req.starting_psn		= 0; /* FIXME */
1088 
1089 	/*
1090 	 * Pick some arbitrary defaults here; we could make these
1091 	 * module parameters if anyone cared about setting them.
1092 	 */
1093 	req.responder_resources		= 4;
1094 	req.remote_cm_response_timeout	= 20;
1095 	req.local_cm_response_timeout	= 20;
1096 	req.retry_count			= 0; /* RFC draft warns against retries */
1097 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1098 	req.max_cm_retries		= 15;
1099 	req.srq				= ipoib_cm_has_srq(dev);
1100 	return ib_send_cm_req(id, &req);
1101 }
1102 
1103 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1104 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1105 {
1106 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1107 	struct ib_qp_attr qp_attr;
1108 	int qp_attr_mask, ret;
1109 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1110 	if (ret) {
1111 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1112 		return ret;
1113 	}
1114 
1115 	qp_attr.qp_state = IB_QPS_INIT;
1116 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1117 	qp_attr.port_num = priv->port;
1118 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1119 
1120 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1121 	if (ret) {
1122 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1123 		return ret;
1124 	}
1125 	return 0;
1126 }
1127 
1128 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1129 			    struct ib_sa_path_rec *pathrec)
1130 {
1131 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1132 	int ret;
1133 
1134 	p->tx_ring = __vmalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1135 			       GFP_NOIO, PAGE_KERNEL);
1136 	if (!p->tx_ring) {
1137 		ipoib_warn(priv, "failed to allocate tx ring\n");
1138 		ret = -ENOMEM;
1139 		goto err_tx;
1140 	}
1141 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1142 
1143 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1144 	if (IS_ERR(p->qp)) {
1145 		ret = PTR_ERR(p->qp);
1146 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1147 		goto err_qp;
1148 	}
1149 
1150 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1151 	if (IS_ERR(p->id)) {
1152 		ret = PTR_ERR(p->id);
1153 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1154 		goto err_id;
1155 	}
1156 
1157 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1158 	if (ret) {
1159 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1160 		goto err_modify;
1161 	}
1162 
1163 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1164 	if (ret) {
1165 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1166 		goto err_send_cm;
1167 	}
1168 
1169 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1170 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1171 
1172 	return 0;
1173 
1174 err_send_cm:
1175 err_modify:
1176 	ib_destroy_cm_id(p->id);
1177 err_id:
1178 	p->id = NULL;
1179 	ib_destroy_qp(p->qp);
1180 err_qp:
1181 	p->qp = NULL;
1182 	vfree(p->tx_ring);
1183 err_tx:
1184 	return ret;
1185 }
1186 
1187 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1188 {
1189 	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1190 	struct ipoib_tx_buf *tx_req;
1191 	unsigned long begin;
1192 
1193 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1194 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1195 
1196 	if (p->id)
1197 		ib_destroy_cm_id(p->id);
1198 
1199 	if (p->tx_ring) {
1200 		/* Wait for all sends to complete */
1201 		begin = jiffies;
1202 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1203 			if (time_after(jiffies, begin + 5 * HZ)) {
1204 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1205 					   p->tx_head - p->tx_tail);
1206 				goto timeout;
1207 			}
1208 
1209 			msleep(1);
1210 		}
1211 	}
1212 
1213 timeout:
1214 
1215 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1216 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1217 		ipoib_dma_unmap_tx(priv, tx_req);
1218 		dev_kfree_skb_any(tx_req->skb);
1219 		++p->tx_tail;
1220 		netif_tx_lock_bh(p->dev);
1221 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1222 		    netif_queue_stopped(p->dev) &&
1223 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1224 			netif_wake_queue(p->dev);
1225 		netif_tx_unlock_bh(p->dev);
1226 	}
1227 
1228 	if (p->qp)
1229 		ib_destroy_qp(p->qp);
1230 
1231 	vfree(p->tx_ring);
1232 	kfree(p);
1233 }
1234 
1235 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1236 			       struct ib_cm_event *event)
1237 {
1238 	struct ipoib_cm_tx *tx = cm_id->context;
1239 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1240 	struct net_device *dev = priv->dev;
1241 	struct ipoib_neigh *neigh;
1242 	unsigned long flags;
1243 	int ret;
1244 
1245 	switch (event->event) {
1246 	case IB_CM_DREQ_RECEIVED:
1247 		ipoib_dbg(priv, "DREQ received.\n");
1248 		ib_send_cm_drep(cm_id, NULL, 0);
1249 		break;
1250 	case IB_CM_REP_RECEIVED:
1251 		ipoib_dbg(priv, "REP received.\n");
1252 		ret = ipoib_cm_rep_handler(cm_id, event);
1253 		if (ret)
1254 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1255 				       NULL, 0, NULL, 0);
1256 		break;
1257 	case IB_CM_REQ_ERROR:
1258 	case IB_CM_REJ_RECEIVED:
1259 	case IB_CM_TIMEWAIT_EXIT:
1260 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1261 		netif_tx_lock_bh(dev);
1262 		spin_lock_irqsave(&priv->lock, flags);
1263 		neigh = tx->neigh;
1264 
1265 		if (neigh) {
1266 			neigh->cm = NULL;
1267 			ipoib_neigh_free(neigh);
1268 
1269 			tx->neigh = NULL;
1270 		}
1271 
1272 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1273 			list_move(&tx->list, &priv->cm.reap_list);
1274 			queue_work(priv->wq, &priv->cm.reap_task);
1275 		}
1276 
1277 		spin_unlock_irqrestore(&priv->lock, flags);
1278 		netif_tx_unlock_bh(dev);
1279 		break;
1280 	default:
1281 		break;
1282 	}
1283 
1284 	return 0;
1285 }
1286 
1287 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1288 				       struct ipoib_neigh *neigh)
1289 {
1290 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1291 	struct ipoib_cm_tx *tx;
1292 
1293 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1294 	if (!tx)
1295 		return NULL;
1296 
1297 	neigh->cm = tx;
1298 	tx->neigh = neigh;
1299 	tx->path = path;
1300 	tx->dev = dev;
1301 	list_add(&tx->list, &priv->cm.start_list);
1302 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1303 	queue_work(priv->wq, &priv->cm.start_task);
1304 	return tx;
1305 }
1306 
1307 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1308 {
1309 	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1310 	unsigned long flags;
1311 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1312 		spin_lock_irqsave(&priv->lock, flags);
1313 		list_move(&tx->list, &priv->cm.reap_list);
1314 		queue_work(priv->wq, &priv->cm.reap_task);
1315 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1316 			  tx->neigh->daddr + 4);
1317 		tx->neigh = NULL;
1318 		spin_unlock_irqrestore(&priv->lock, flags);
1319 	}
1320 }
1321 
1322 #define QPN_AND_OPTIONS_OFFSET	4
1323 
1324 static void ipoib_cm_tx_start(struct work_struct *work)
1325 {
1326 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1327 						   cm.start_task);
1328 	struct net_device *dev = priv->dev;
1329 	struct ipoib_neigh *neigh;
1330 	struct ipoib_cm_tx *p;
1331 	unsigned long flags;
1332 	struct ipoib_path *path;
1333 	int ret;
1334 
1335 	struct ib_sa_path_rec pathrec;
1336 	u32 qpn;
1337 
1338 	netif_tx_lock_bh(dev);
1339 	spin_lock_irqsave(&priv->lock, flags);
1340 
1341 	while (!list_empty(&priv->cm.start_list)) {
1342 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1343 		list_del_init(&p->list);
1344 		neigh = p->neigh;
1345 
1346 		qpn = IPOIB_QPN(neigh->daddr);
1347 		/*
1348 		 * As long as the search is with these 2 locks,
1349 		 * path existence indicates its validity.
1350 		 */
1351 		path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1352 		if (!path) {
1353 			pr_info("%s ignore not valid path %pI6\n",
1354 				__func__,
1355 				neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1356 			goto free_neigh;
1357 		}
1358 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1359 
1360 		spin_unlock_irqrestore(&priv->lock, flags);
1361 		netif_tx_unlock_bh(dev);
1362 
1363 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1364 
1365 		netif_tx_lock_bh(dev);
1366 		spin_lock_irqsave(&priv->lock, flags);
1367 
1368 		if (ret) {
1369 free_neigh:
1370 			neigh = p->neigh;
1371 			if (neigh) {
1372 				neigh->cm = NULL;
1373 				ipoib_neigh_free(neigh);
1374 			}
1375 			list_del(&p->list);
1376 			kfree(p);
1377 		}
1378 	}
1379 
1380 	spin_unlock_irqrestore(&priv->lock, flags);
1381 	netif_tx_unlock_bh(dev);
1382 }
1383 
1384 static void ipoib_cm_tx_reap(struct work_struct *work)
1385 {
1386 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1387 						   cm.reap_task);
1388 	struct net_device *dev = priv->dev;
1389 	struct ipoib_cm_tx *p;
1390 	unsigned long flags;
1391 
1392 	netif_tx_lock_bh(dev);
1393 	spin_lock_irqsave(&priv->lock, flags);
1394 
1395 	while (!list_empty(&priv->cm.reap_list)) {
1396 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1397 		list_del(&p->list);
1398 		spin_unlock_irqrestore(&priv->lock, flags);
1399 		netif_tx_unlock_bh(dev);
1400 		ipoib_cm_tx_destroy(p);
1401 		netif_tx_lock_bh(dev);
1402 		spin_lock_irqsave(&priv->lock, flags);
1403 	}
1404 
1405 	spin_unlock_irqrestore(&priv->lock, flags);
1406 	netif_tx_unlock_bh(dev);
1407 }
1408 
1409 static void ipoib_cm_skb_reap(struct work_struct *work)
1410 {
1411 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1412 						   cm.skb_task);
1413 	struct net_device *dev = priv->dev;
1414 	struct sk_buff *skb;
1415 	unsigned long flags;
1416 	unsigned mtu = priv->mcast_mtu;
1417 
1418 	netif_tx_lock_bh(dev);
1419 	spin_lock_irqsave(&priv->lock, flags);
1420 
1421 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1422 		spin_unlock_irqrestore(&priv->lock, flags);
1423 		netif_tx_unlock_bh(dev);
1424 
1425 		if (skb->protocol == htons(ETH_P_IP))
1426 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1427 #if IS_ENABLED(CONFIG_IPV6)
1428 		else if (skb->protocol == htons(ETH_P_IPV6))
1429 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1430 #endif
1431 		dev_kfree_skb_any(skb);
1432 
1433 		netif_tx_lock_bh(dev);
1434 		spin_lock_irqsave(&priv->lock, flags);
1435 	}
1436 
1437 	spin_unlock_irqrestore(&priv->lock, flags);
1438 	netif_tx_unlock_bh(dev);
1439 }
1440 
1441 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1442 			   unsigned int mtu)
1443 {
1444 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1445 	int e = skb_queue_empty(&priv->cm.skb_queue);
1446 
1447 	if (skb_dst(skb))
1448 		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1449 
1450 	skb_queue_tail(&priv->cm.skb_queue, skb);
1451 	if (e)
1452 		queue_work(priv->wq, &priv->cm.skb_task);
1453 }
1454 
1455 static void ipoib_cm_rx_reap(struct work_struct *work)
1456 {
1457 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1458 						cm.rx_reap_task)->dev);
1459 }
1460 
1461 static void ipoib_cm_stale_task(struct work_struct *work)
1462 {
1463 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1464 						   cm.stale_task.work);
1465 	struct ipoib_cm_rx *p;
1466 	int ret;
1467 
1468 	spin_lock_irq(&priv->lock);
1469 	while (!list_empty(&priv->cm.passive_ids)) {
1470 		/* List is sorted by LRU, start from tail,
1471 		 * stop when we see a recently used entry */
1472 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1473 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1474 			break;
1475 		list_move(&p->list, &priv->cm.rx_error_list);
1476 		p->state = IPOIB_CM_RX_ERROR;
1477 		spin_unlock_irq(&priv->lock);
1478 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1479 		if (ret)
1480 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1481 		spin_lock_irq(&priv->lock);
1482 	}
1483 
1484 	if (!list_empty(&priv->cm.passive_ids))
1485 		queue_delayed_work(priv->wq,
1486 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1487 	spin_unlock_irq(&priv->lock);
1488 }
1489 
1490 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1491 			 char *buf)
1492 {
1493 	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1494 
1495 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1496 		return sprintf(buf, "connected\n");
1497 	else
1498 		return sprintf(buf, "datagram\n");
1499 }
1500 
1501 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1502 			const char *buf, size_t count)
1503 {
1504 	struct net_device *dev = to_net_dev(d);
1505 	int ret;
1506 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1507 
1508 	if (test_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags))
1509 		return -EPERM;
1510 
1511 	if (!rtnl_trylock())
1512 		return restart_syscall();
1513 
1514 	ret = ipoib_set_mode(dev, buf);
1515 
1516 	rtnl_unlock();
1517 
1518 	if (!ret)
1519 		return count;
1520 
1521 	return ret;
1522 }
1523 
1524 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1525 
1526 int ipoib_cm_add_mode_attr(struct net_device *dev)
1527 {
1528 	return device_create_file(&dev->dev, &dev_attr_mode);
1529 }
1530 
1531 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1532 {
1533 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1534 	struct ib_srq_init_attr srq_init_attr = {
1535 		.srq_type = IB_SRQT_BASIC,
1536 		.attr = {
1537 			.max_wr  = ipoib_recvq_size,
1538 			.max_sge = max_sge
1539 		}
1540 	};
1541 
1542 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1543 	if (IS_ERR(priv->cm.srq)) {
1544 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1545 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1546 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1547 		priv->cm.srq = NULL;
1548 		return;
1549 	}
1550 
1551 	priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1552 	if (!priv->cm.srq_ring) {
1553 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1554 		       priv->ca->name, ipoib_recvq_size);
1555 		ib_destroy_srq(priv->cm.srq);
1556 		priv->cm.srq = NULL;
1557 		return;
1558 	}
1559 
1560 }
1561 
1562 int ipoib_cm_dev_init(struct net_device *dev)
1563 {
1564 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1565 	int max_srq_sge, i;
1566 
1567 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1568 	INIT_LIST_HEAD(&priv->cm.reap_list);
1569 	INIT_LIST_HEAD(&priv->cm.start_list);
1570 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1571 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1572 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1573 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1574 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1575 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1576 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1577 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1578 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1579 
1580 	skb_queue_head_init(&priv->cm.skb_queue);
1581 
1582 	ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1583 
1584 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1585 	ipoib_cm_create_srq(dev, max_srq_sge);
1586 	if (ipoib_cm_has_srq(dev)) {
1587 		priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1588 		priv->cm.num_frags  = max_srq_sge;
1589 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1590 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1591 	} else {
1592 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1593 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1594 	}
1595 
1596 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1597 
1598 	if (ipoib_cm_has_srq(dev)) {
1599 		for (i = 0; i < ipoib_recvq_size; ++i) {
1600 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1601 						   priv->cm.num_frags - 1,
1602 						   priv->cm.srq_ring[i].mapping,
1603 						   GFP_KERNEL)) {
1604 				ipoib_warn(priv, "failed to allocate "
1605 					   "receive buffer %d\n", i);
1606 				ipoib_cm_dev_cleanup(dev);
1607 				return -ENOMEM;
1608 			}
1609 
1610 			if (ipoib_cm_post_receive_srq(dev, i)) {
1611 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1612 					   "failed for buf %d\n", i);
1613 				ipoib_cm_dev_cleanup(dev);
1614 				return -EIO;
1615 			}
1616 		}
1617 	}
1618 
1619 	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1620 	return 0;
1621 }
1622 
1623 void ipoib_cm_dev_cleanup(struct net_device *dev)
1624 {
1625 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1626 	int ret;
1627 
1628 	if (!priv->cm.srq)
1629 		return;
1630 
1631 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1632 
1633 	ret = ib_destroy_srq(priv->cm.srq);
1634 	if (ret)
1635 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1636 
1637 	priv->cm.srq = NULL;
1638 	if (!priv->cm.srq_ring)
1639 		return;
1640 
1641 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1642 	priv->cm.srq_ring = NULL;
1643 }
1644