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 #include <linux/sched/signal.h>
42 #include <linux/sched/mm.h>
43 
44 #include "ipoib.h"
45 
46 int ipoib_max_conn_qp = 128;
47 
48 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
49 MODULE_PARM_DESC(max_nonsrq_conn_qp,
50 		 "Max number of connected-mode QPs per interface "
51 		 "(applied only if shared receive queue is not available)");
52 
53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
54 static int data_debug_level;
55 
56 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
57 MODULE_PARM_DESC(cm_data_debug_level,
58 		 "Enable data path debug tracing for connected mode if > 0");
59 #endif
60 
61 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
62 
63 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
64 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
65 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
66 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
67 
68 #define IPOIB_CM_RX_RESERVE     (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
69 
70 static struct ib_qp_attr ipoib_cm_err_attr = {
71 	.qp_state = IB_QPS_ERR
72 };
73 
74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
75 
76 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
77 	.opcode = IB_WR_SEND,
78 };
79 
80 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
81 			       const struct ib_cm_event *event);
82 
83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
84 				  u64 mapping[IPOIB_CM_RX_SG])
85 {
86 	int i;
87 
88 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
89 
90 	for (i = 0; i < frags; ++i)
91 		ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
92 }
93 
94 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
95 {
96 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
97 	int i, ret;
98 
99 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
100 
101 	for (i = 0; i < priv->cm.num_frags; ++i)
102 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
103 
104 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, NULL);
105 	if (unlikely(ret)) {
106 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
107 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
108 				      priv->cm.srq_ring[id].mapping);
109 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
110 		priv->cm.srq_ring[id].skb = NULL;
111 	}
112 
113 	return ret;
114 }
115 
116 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
117 					struct ipoib_cm_rx *rx,
118 					struct ib_recv_wr *wr,
119 					struct ib_sge *sge, int id)
120 {
121 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
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, NULL);
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 = ipoib_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 = ipoib_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 ipoib_cm_rx *p;
214 
215 	/* We only reserved 1 extra slot in CQ for drain WRs, so
216 	 * make sure we have at most 1 outstanding WR. */
217 	if (list_empty(&priv->cm.rx_flush_list) ||
218 	    !list_empty(&priv->cm.rx_drain_list))
219 		return;
220 
221 	/*
222 	 * QPs on flush list are error state.  This way, a "flush
223 	 * error" WC will be immediately generated for each WR we post.
224 	 */
225 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
226 	ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
227 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, NULL))
228 		ipoib_warn(priv, "failed to post drain wr\n");
229 
230 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
231 }
232 
233 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
234 {
235 	struct ipoib_cm_rx *p = ctx;
236 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
237 	unsigned long flags;
238 
239 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
240 		return;
241 
242 	spin_lock_irqsave(&priv->lock, flags);
243 	list_move(&p->list, &priv->cm.rx_flush_list);
244 	p->state = IPOIB_CM_RX_FLUSH;
245 	ipoib_cm_start_rx_drain(priv);
246 	spin_unlock_irqrestore(&priv->lock, flags);
247 }
248 
249 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
250 					   struct ipoib_cm_rx *p)
251 {
252 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
253 	struct ib_qp_init_attr attr = {
254 		.event_handler = ipoib_cm_rx_event_handler,
255 		.send_cq = priv->recv_cq, /* For drain WR */
256 		.recv_cq = priv->recv_cq,
257 		.srq = priv->cm.srq,
258 		.cap.max_send_wr = 1, /* For drain WR */
259 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
260 		.sq_sig_type = IB_SIGNAL_ALL_WR,
261 		.qp_type = IB_QPT_RC,
262 		.qp_context = p,
263 	};
264 
265 	if (!ipoib_cm_has_srq(dev)) {
266 		attr.cap.max_recv_wr  = ipoib_recvq_size;
267 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
268 	}
269 
270 	return ib_create_qp(priv->pd, &attr);
271 }
272 
273 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
274 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
275 				 unsigned int psn)
276 {
277 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
278 	struct ib_qp_attr qp_attr;
279 	int qp_attr_mask, ret;
280 
281 	qp_attr.qp_state = IB_QPS_INIT;
282 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
283 	if (ret) {
284 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
285 		return ret;
286 	}
287 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
288 	if (ret) {
289 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
290 		return ret;
291 	}
292 	qp_attr.qp_state = IB_QPS_RTR;
293 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
294 	if (ret) {
295 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
296 		return ret;
297 	}
298 	qp_attr.rq_psn = psn;
299 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
300 	if (ret) {
301 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
302 		return ret;
303 	}
304 
305 	/*
306 	 * Current Mellanox HCA firmware won't generate completions
307 	 * with error for drain WRs unless the QP has been moved to
308 	 * RTS first. This work-around leaves a window where a QP has
309 	 * moved to error asynchronously, but this will eventually get
310 	 * fixed in firmware, so let's not error out if modify QP
311 	 * fails.
312 	 */
313 	qp_attr.qp_state = IB_QPS_RTS;
314 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
315 	if (ret) {
316 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
317 		return 0;
318 	}
319 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
320 	if (ret) {
321 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
322 		return 0;
323 	}
324 
325 	return 0;
326 }
327 
328 static void ipoib_cm_init_rx_wr(struct net_device *dev,
329 				struct ib_recv_wr *wr,
330 				struct ib_sge *sge)
331 {
332 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
333 	int i;
334 
335 	for (i = 0; i < priv->cm.num_frags; ++i)
336 		sge[i].lkey = priv->pd->local_dma_lkey;
337 
338 	sge[0].length = IPOIB_CM_HEAD_SIZE;
339 	for (i = 1; i < priv->cm.num_frags; ++i)
340 		sge[i].length = PAGE_SIZE;
341 
342 	wr->next    = NULL;
343 	wr->sg_list = sge;
344 	wr->num_sge = priv->cm.num_frags;
345 }
346 
347 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
348 				   struct ipoib_cm_rx *rx)
349 {
350 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
351 	struct {
352 		struct ib_recv_wr wr;
353 		struct ib_sge sge[IPOIB_CM_RX_SG];
354 	} *t;
355 	int ret;
356 	int i;
357 
358 	rx->rx_ring = vzalloc(array_size(ipoib_recvq_size,
359 					 sizeof(*rx->rx_ring)));
360 	if (!rx->rx_ring)
361 		return -ENOMEM;
362 
363 	t = kmalloc(sizeof(*t), GFP_KERNEL);
364 	if (!t) {
365 		ret = -ENOMEM;
366 		goto err_free_1;
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 					   GFP_KERNEL)) {
387 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
388 			ret = -ENOMEM;
389 			goto err_count;
390 		}
391 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
392 		if (ret) {
393 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
394 				   "failed for buf %d\n", i);
395 			ret = -EIO;
396 			goto err_count;
397 		}
398 	}
399 
400 	rx->recv_count = ipoib_recvq_size;
401 
402 	kfree(t);
403 
404 	return 0;
405 
406 err_count:
407 	spin_lock_irq(&priv->lock);
408 	--priv->cm.nonsrq_conn_qp;
409 	spin_unlock_irq(&priv->lock);
410 
411 err_free:
412 	kfree(t);
413 
414 err_free_1:
415 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
416 
417 	return ret;
418 }
419 
420 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
421 			     struct ib_qp *qp,
422 			     const struct ib_cm_req_event_param *req,
423 			     unsigned int psn)
424 {
425 	struct ipoib_dev_priv *priv = ipoib_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,
443 				const struct ib_cm_event *event)
444 {
445 	struct net_device *dev = cm_id->context;
446 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
447 	struct ipoib_cm_rx *p;
448 	unsigned int psn;
449 	int ret;
450 
451 	ipoib_dbg(priv, "REQ arrived\n");
452 	p = kzalloc(sizeof(*p), GFP_KERNEL);
453 	if (!p)
454 		return -ENOMEM;
455 	p->dev = dev;
456 	p->id = cm_id;
457 	cm_id->context = p;
458 	p->state = IPOIB_CM_RX_LIVE;
459 	p->jiffies = jiffies;
460 	INIT_LIST_HEAD(&p->list);
461 
462 	p->qp = ipoib_cm_create_rx_qp(dev, p);
463 	if (IS_ERR(p->qp)) {
464 		ret = PTR_ERR(p->qp);
465 		goto err_qp;
466 	}
467 
468 	psn = get_random_u32() & 0xffffff;
469 	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
470 	if (ret)
471 		goto err_modify;
472 
473 	if (!ipoib_cm_has_srq(dev)) {
474 		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
475 		if (ret)
476 			goto err_modify;
477 	}
478 
479 	spin_lock_irq(&priv->lock);
480 	queue_delayed_work(priv->wq,
481 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
482 	/* Add this entry to passive ids list head, but do not re-add it
483 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
484 	p->jiffies = jiffies;
485 	if (p->state == IPOIB_CM_RX_LIVE)
486 		list_move(&p->list, &priv->cm.passive_ids);
487 	spin_unlock_irq(&priv->lock);
488 
489 	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
490 	if (ret) {
491 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
492 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
493 			ipoib_warn(priv, "unable to move qp to error state\n");
494 	}
495 	return 0;
496 
497 err_modify:
498 	ib_destroy_qp(p->qp);
499 err_qp:
500 	kfree(p);
501 	return ret;
502 }
503 
504 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
505 			       const struct ib_cm_event *event)
506 {
507 	struct ipoib_cm_rx *p;
508 	struct ipoib_dev_priv *priv;
509 
510 	switch (event->event) {
511 	case IB_CM_REQ_RECEIVED:
512 		return ipoib_cm_req_handler(cm_id, event);
513 	case IB_CM_DREQ_RECEIVED:
514 		ib_send_cm_drep(cm_id, NULL, 0);
515 		fallthrough;
516 	case IB_CM_REJ_RECEIVED:
517 		p = cm_id->context;
518 		priv = ipoib_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 		fallthrough;
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_t(unsigned int, length, 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 = ipoib_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,
598 			  "cm recv error (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 -
644 			   min_t(u32, wc->byte_len, IPOIB_CM_HEAD_SIZE)) /
645 		PAGE_SIZE;
646 
647 	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
648 				       mapping, GFP_ATOMIC);
649 	if (unlikely(!newskb)) {
650 		/*
651 		 * If we can't allocate a new RX buffer, dump
652 		 * this packet and reuse the old buffer.
653 		 */
654 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
655 		++dev->stats.rx_dropped;
656 		goto repost;
657 	}
658 
659 	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
660 	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof(*mapping));
661 
662 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
663 		       wc->byte_len, wc->slid);
664 
665 	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
666 
667 copied:
668 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
669 	skb_add_pseudo_hdr(skb);
670 
671 	++dev->stats.rx_packets;
672 	dev->stats.rx_bytes += skb->len;
673 
674 	skb->dev = dev;
675 	/* XXX get correct PACKET_ type here */
676 	skb->pkt_type = PACKET_HOST;
677 	netif_receive_skb(skb);
678 
679 repost:
680 	if (has_srq) {
681 		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
682 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
683 				   "for buf %d\n", wr_id);
684 	} else {
685 		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
686 							  &priv->cm.rx_wr,
687 							  priv->cm.rx_sge,
688 							  wr_id))) {
689 			--p->recv_count;
690 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
691 				   "for buf %d\n", wr_id);
692 		}
693 	}
694 }
695 
696 static inline int post_send(struct ipoib_dev_priv *priv,
697 			    struct ipoib_cm_tx *tx,
698 			    unsigned int wr_id,
699 			    struct ipoib_tx_buf *tx_req)
700 {
701 	ipoib_build_sge(priv, tx_req);
702 
703 	priv->tx_wr.wr.wr_id	= wr_id | IPOIB_OP_CM;
704 
705 	return ib_post_send(tx->qp, &priv->tx_wr.wr, NULL);
706 }
707 
708 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
709 {
710 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
711 	struct ipoib_tx_buf *tx_req;
712 	int rc;
713 	unsigned int usable_sge = tx->max_send_sge - !!skb_headlen(skb);
714 
715 	if (unlikely(skb->len > tx->mtu)) {
716 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
717 			   skb->len, tx->mtu);
718 		++dev->stats.tx_dropped;
719 		++dev->stats.tx_errors;
720 		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
721 		return;
722 	}
723 	if (skb_shinfo(skb)->nr_frags > usable_sge) {
724 		if (skb_linearize(skb) < 0) {
725 			ipoib_warn(priv, "skb could not be linearized\n");
726 			++dev->stats.tx_dropped;
727 			++dev->stats.tx_errors;
728 			dev_kfree_skb_any(skb);
729 			return;
730 		}
731 		/* Does skb_linearize return ok without reducing nr_frags? */
732 		if (skb_shinfo(skb)->nr_frags > usable_sge) {
733 			ipoib_warn(priv, "too many frags after skb linearize\n");
734 			++dev->stats.tx_dropped;
735 			++dev->stats.tx_errors;
736 			dev_kfree_skb_any(skb);
737 			return;
738 		}
739 	}
740 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
741 		       tx->tx_head, skb->len, tx->qp->qp_num);
742 
743 	/*
744 	 * We put the skb into the tx_ring _before_ we call post_send()
745 	 * because it's entirely possible that the completion handler will
746 	 * run before we execute anything after the post_send().  That
747 	 * means we have to make sure everything is properly recorded and
748 	 * our state is consistent before we call post_send().
749 	 */
750 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
751 	tx_req->skb = skb;
752 
753 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
754 		++dev->stats.tx_errors;
755 		dev_kfree_skb_any(skb);
756 		return;
757 	}
758 
759 	if ((priv->global_tx_head - priv->global_tx_tail) ==
760 	    ipoib_sendq_size - 1) {
761 		ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
762 			  tx->qp->qp_num);
763 		netif_stop_queue(dev);
764 	}
765 
766 	skb_orphan(skb);
767 	skb_dst_drop(skb);
768 
769 	if (netif_queue_stopped(dev)) {
770 		rc = ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
771 				      IB_CQ_REPORT_MISSED_EVENTS);
772 		if (unlikely(rc < 0))
773 			ipoib_warn(priv, "IPoIB/CM:request notify on send CQ failed\n");
774 		else if (rc)
775 			napi_schedule(&priv->send_napi);
776 	}
777 
778 	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
779 	if (unlikely(rc)) {
780 		ipoib_warn(priv, "IPoIB/CM:post_send failed, error %d\n", rc);
781 		++dev->stats.tx_errors;
782 		ipoib_dma_unmap_tx(priv, tx_req);
783 		dev_kfree_skb_any(skb);
784 
785 		if (netif_queue_stopped(dev))
786 			netif_wake_queue(dev);
787 	} else {
788 		netif_trans_update(dev);
789 		++tx->tx_head;
790 		++priv->global_tx_head;
791 	}
792 }
793 
794 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
795 {
796 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
797 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
798 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
799 	struct ipoib_tx_buf *tx_req;
800 	unsigned long flags;
801 
802 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
803 		       wr_id, wc->status);
804 
805 	if (unlikely(wr_id >= ipoib_sendq_size)) {
806 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
807 			   wr_id, ipoib_sendq_size);
808 		return;
809 	}
810 
811 	tx_req = &tx->tx_ring[wr_id];
812 
813 	ipoib_dma_unmap_tx(priv, tx_req);
814 
815 	/* FIXME: is this right? Shouldn't we only increment on success? */
816 	++dev->stats.tx_packets;
817 	dev->stats.tx_bytes += tx_req->skb->len;
818 
819 	dev_kfree_skb_any(tx_req->skb);
820 
821 	netif_tx_lock(dev);
822 
823 	++tx->tx_tail;
824 	++priv->global_tx_tail;
825 
826 	if (unlikely(netif_queue_stopped(dev) &&
827 		     ((priv->global_tx_head - priv->global_tx_tail) <=
828 		      ipoib_sendq_size >> 1) &&
829 		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
830 		netif_wake_queue(dev);
831 
832 	if (wc->status != IB_WC_SUCCESS &&
833 	    wc->status != IB_WC_WR_FLUSH_ERR) {
834 		struct ipoib_neigh *neigh;
835 
836 		/* IB_WC[_RNR]_RETRY_EXC_ERR error is part of the life cycle,
837 		 * so don't make waves.
838 		 */
839 		if (wc->status == IB_WC_RNR_RETRY_EXC_ERR ||
840 		    wc->status == IB_WC_RETRY_EXC_ERR)
841 			ipoib_dbg(priv,
842 				  "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
843 				   __func__, wc->status, wr_id, wc->vendor_err);
844 		else
845 			ipoib_warn(priv,
846 				    "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
847 				   __func__, wc->status, wr_id, wc->vendor_err);
848 
849 		spin_lock_irqsave(&priv->lock, flags);
850 		neigh = tx->neigh;
851 
852 		if (neigh) {
853 			neigh->cm = NULL;
854 			ipoib_neigh_free(neigh);
855 
856 			tx->neigh = NULL;
857 		}
858 
859 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
860 			list_move(&tx->list, &priv->cm.reap_list);
861 			queue_work(priv->wq, &priv->cm.reap_task);
862 		}
863 
864 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
865 
866 		spin_unlock_irqrestore(&priv->lock, flags);
867 	}
868 
869 	netif_tx_unlock(dev);
870 }
871 
872 int ipoib_cm_dev_open(struct net_device *dev)
873 {
874 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
875 	int ret;
876 
877 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
878 		return 0;
879 
880 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
881 	if (IS_ERR(priv->cm.id)) {
882 		pr_warn("%s: failed to create CM ID\n", priv->ca->name);
883 		ret = PTR_ERR(priv->cm.id);
884 		goto err_cm;
885 	}
886 
887 	ret = ib_cm_listen(priv->cm.id,
888 			   cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num));
889 	if (ret) {
890 		pr_warn("%s: failed to listen on ID 0x%llx\n", priv->ca->name,
891 			IPOIB_CM_IETF_ID | priv->qp->qp_num);
892 		goto err_listen;
893 	}
894 
895 	return 0;
896 
897 err_listen:
898 	ib_destroy_cm_id(priv->cm.id);
899 err_cm:
900 	priv->cm.id = NULL;
901 	return ret;
902 }
903 
904 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
905 {
906 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
907 	struct ipoib_cm_rx *rx, *n;
908 	LIST_HEAD(list);
909 
910 	spin_lock_irq(&priv->lock);
911 	list_splice_init(&priv->cm.rx_reap_list, &list);
912 	spin_unlock_irq(&priv->lock);
913 
914 	list_for_each_entry_safe(rx, n, &list, list) {
915 		ib_destroy_cm_id(rx->id);
916 		ib_destroy_qp(rx->qp);
917 		if (!ipoib_cm_has_srq(dev)) {
918 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
919 			spin_lock_irq(&priv->lock);
920 			--priv->cm.nonsrq_conn_qp;
921 			spin_unlock_irq(&priv->lock);
922 		}
923 		kfree(rx);
924 	}
925 }
926 
927 void ipoib_cm_dev_stop(struct net_device *dev)
928 {
929 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
930 	struct ipoib_cm_rx *p;
931 	unsigned long begin;
932 	int ret;
933 
934 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
935 		return;
936 
937 	ib_destroy_cm_id(priv->cm.id);
938 	priv->cm.id = NULL;
939 
940 	spin_lock_irq(&priv->lock);
941 	while (!list_empty(&priv->cm.passive_ids)) {
942 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
943 		list_move(&p->list, &priv->cm.rx_error_list);
944 		p->state = IPOIB_CM_RX_ERROR;
945 		spin_unlock_irq(&priv->lock);
946 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
947 		if (ret)
948 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
949 		spin_lock_irq(&priv->lock);
950 	}
951 
952 	/* Wait for all RX to be drained */
953 	begin = jiffies;
954 
955 	while (!list_empty(&priv->cm.rx_error_list) ||
956 	       !list_empty(&priv->cm.rx_flush_list) ||
957 	       !list_empty(&priv->cm.rx_drain_list)) {
958 		if (time_after(jiffies, begin + 5 * HZ)) {
959 			ipoib_warn(priv, "RX drain timing out\n");
960 
961 			/*
962 			 * assume the HW is wedged and just free up everything.
963 			 */
964 			list_splice_init(&priv->cm.rx_flush_list,
965 					 &priv->cm.rx_reap_list);
966 			list_splice_init(&priv->cm.rx_error_list,
967 					 &priv->cm.rx_reap_list);
968 			list_splice_init(&priv->cm.rx_drain_list,
969 					 &priv->cm.rx_reap_list);
970 			break;
971 		}
972 		spin_unlock_irq(&priv->lock);
973 		usleep_range(1000, 2000);
974 		ipoib_drain_cq(dev);
975 		spin_lock_irq(&priv->lock);
976 	}
977 
978 	spin_unlock_irq(&priv->lock);
979 
980 	ipoib_cm_free_rx_reap_list(dev);
981 
982 	cancel_delayed_work(&priv->cm.stale_task);
983 }
984 
985 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id,
986 				const struct ib_cm_event *event)
987 {
988 	struct ipoib_cm_tx *p = cm_id->context;
989 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
990 	struct ipoib_cm_data *data = event->private_data;
991 	struct sk_buff_head skqueue;
992 	struct ib_qp_attr qp_attr;
993 	int qp_attr_mask, ret;
994 	struct sk_buff *skb;
995 
996 	p->mtu = be32_to_cpu(data->mtu);
997 
998 	if (p->mtu <= IPOIB_ENCAP_LEN) {
999 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
1000 			   p->mtu, IPOIB_ENCAP_LEN);
1001 		return -EINVAL;
1002 	}
1003 
1004 	qp_attr.qp_state = IB_QPS_RTR;
1005 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1006 	if (ret) {
1007 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
1008 		return ret;
1009 	}
1010 
1011 	qp_attr.rq_psn = 0 /* FIXME */;
1012 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1013 	if (ret) {
1014 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
1015 		return ret;
1016 	}
1017 
1018 	qp_attr.qp_state = IB_QPS_RTS;
1019 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1020 	if (ret) {
1021 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1022 		return ret;
1023 	}
1024 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1025 	if (ret) {
1026 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1027 		return ret;
1028 	}
1029 
1030 	skb_queue_head_init(&skqueue);
1031 
1032 	netif_tx_lock_bh(p->dev);
1033 	spin_lock_irq(&priv->lock);
1034 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1035 	if (p->neigh)
1036 		while ((skb = __skb_dequeue(&p->neigh->queue)))
1037 			__skb_queue_tail(&skqueue, skb);
1038 	spin_unlock_irq(&priv->lock);
1039 	netif_tx_unlock_bh(p->dev);
1040 
1041 	while ((skb = __skb_dequeue(&skqueue))) {
1042 		skb->dev = p->dev;
1043 		ret = dev_queue_xmit(skb);
1044 		if (ret)
1045 			ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1046 				   __func__, ret);
1047 	}
1048 
1049 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1050 	if (ret) {
1051 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1052 		return ret;
1053 	}
1054 	return 0;
1055 }
1056 
1057 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1058 {
1059 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1060 	struct ib_qp_init_attr attr = {
1061 		.send_cq		= priv->send_cq,
1062 		.recv_cq		= priv->recv_cq,
1063 		.srq			= priv->cm.srq,
1064 		.cap.max_send_wr	= ipoib_sendq_size,
1065 		.cap.max_send_sge	= 1,
1066 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1067 		.qp_type		= IB_QPT_RC,
1068 		.qp_context		= tx,
1069 		.create_flags		= 0
1070 	};
1071 	struct ib_qp *tx_qp;
1072 
1073 	if (dev->features & NETIF_F_SG)
1074 		attr.cap.max_send_sge = min_t(u32, priv->ca->attrs.max_send_sge,
1075 					      MAX_SKB_FRAGS + 1);
1076 
1077 	tx_qp = ib_create_qp(priv->pd, &attr);
1078 	tx->max_send_sge = attr.cap.max_send_sge;
1079 	return tx_qp;
1080 }
1081 
1082 static int ipoib_cm_send_req(struct net_device *dev,
1083 			     struct ib_cm_id *id, struct ib_qp *qp,
1084 			     u32 qpn,
1085 			     struct sa_path_rec *pathrec)
1086 {
1087 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1088 	struct ipoib_cm_data data = {};
1089 	struct ib_cm_req_param req = {};
1090 
1091 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1092 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1093 
1094 	req.primary_path		= pathrec;
1095 	req.alternate_path		= NULL;
1096 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1097 	req.qp_num			= qp->qp_num;
1098 	req.qp_type			= qp->qp_type;
1099 	req.private_data		= &data;
1100 	req.private_data_len		= sizeof(data);
1101 	req.flow_control		= 0;
1102 
1103 	req.starting_psn		= 0; /* FIXME */
1104 
1105 	/*
1106 	 * Pick some arbitrary defaults here; we could make these
1107 	 * module parameters if anyone cared about setting them.
1108 	 */
1109 	req.responder_resources		= 4;
1110 	req.remote_cm_response_timeout	= 20;
1111 	req.local_cm_response_timeout	= 20;
1112 	req.retry_count			= 0; /* RFC draft warns against retries */
1113 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1114 	req.max_cm_retries		= 15;
1115 	req.srq				= ipoib_cm_has_srq(dev);
1116 	return ib_send_cm_req(id, &req);
1117 }
1118 
1119 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1120 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1121 {
1122 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1123 	struct ib_qp_attr qp_attr;
1124 	int qp_attr_mask, ret;
1125 
1126 	qp_attr.pkey_index = priv->pkey_index;
1127 	qp_attr.qp_state = IB_QPS_INIT;
1128 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1129 	qp_attr.port_num = priv->port;
1130 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1131 
1132 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1133 	if (ret) {
1134 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1135 		return ret;
1136 	}
1137 	return 0;
1138 }
1139 
1140 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1141 			    struct sa_path_rec *pathrec)
1142 {
1143 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1144 	unsigned int noio_flag;
1145 	int ret;
1146 
1147 	noio_flag = memalloc_noio_save();
1148 	p->tx_ring = vzalloc(array_size(ipoib_sendq_size, sizeof(*p->tx_ring)));
1149 	if (!p->tx_ring) {
1150 		memalloc_noio_restore(noio_flag);
1151 		ret = -ENOMEM;
1152 		goto err_tx;
1153 	}
1154 
1155 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1156 	memalloc_noio_restore(noio_flag);
1157 	if (IS_ERR(p->qp)) {
1158 		ret = PTR_ERR(p->qp);
1159 		ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1160 		goto err_qp;
1161 	}
1162 
1163 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1164 	if (IS_ERR(p->id)) {
1165 		ret = PTR_ERR(p->id);
1166 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1167 		goto err_id;
1168 	}
1169 
1170 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1171 	if (ret) {
1172 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1173 		goto err_modify_send;
1174 	}
1175 
1176 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1177 	if (ret) {
1178 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1179 		goto err_modify_send;
1180 	}
1181 
1182 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1183 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1184 
1185 	return 0;
1186 
1187 err_modify_send:
1188 	ib_destroy_cm_id(p->id);
1189 err_id:
1190 	p->id = NULL;
1191 	ib_destroy_qp(p->qp);
1192 err_qp:
1193 	p->qp = NULL;
1194 	vfree(p->tx_ring);
1195 err_tx:
1196 	return ret;
1197 }
1198 
1199 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1200 {
1201 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1202 	struct ipoib_tx_buf *tx_req;
1203 	unsigned long begin;
1204 
1205 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1206 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1207 
1208 	if (p->id)
1209 		ib_destroy_cm_id(p->id);
1210 
1211 	if (p->tx_ring) {
1212 		/* Wait for all sends to complete */
1213 		begin = jiffies;
1214 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1215 			if (time_after(jiffies, begin + 5 * HZ)) {
1216 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1217 					   p->tx_head - p->tx_tail);
1218 				goto timeout;
1219 			}
1220 
1221 			usleep_range(1000, 2000);
1222 		}
1223 	}
1224 
1225 timeout:
1226 
1227 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1228 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1229 		ipoib_dma_unmap_tx(priv, tx_req);
1230 		dev_kfree_skb_any(tx_req->skb);
1231 		netif_tx_lock_bh(p->dev);
1232 		++p->tx_tail;
1233 		++priv->global_tx_tail;
1234 		if (unlikely((priv->global_tx_head - priv->global_tx_tail) <=
1235 			     ipoib_sendq_size >> 1) &&
1236 		    netif_queue_stopped(p->dev) &&
1237 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1238 			netif_wake_queue(p->dev);
1239 		netif_tx_unlock_bh(p->dev);
1240 	}
1241 
1242 	if (p->qp)
1243 		ib_destroy_qp(p->qp);
1244 
1245 	vfree(p->tx_ring);
1246 	kfree(p);
1247 }
1248 
1249 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1250 			       const struct ib_cm_event *event)
1251 {
1252 	struct ipoib_cm_tx *tx = cm_id->context;
1253 	struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1254 	struct net_device *dev = priv->dev;
1255 	struct ipoib_neigh *neigh;
1256 	unsigned long flags;
1257 	int ret;
1258 
1259 	switch (event->event) {
1260 	case IB_CM_DREQ_RECEIVED:
1261 		ipoib_dbg(priv, "DREQ received.\n");
1262 		ib_send_cm_drep(cm_id, NULL, 0);
1263 		break;
1264 	case IB_CM_REP_RECEIVED:
1265 		ipoib_dbg(priv, "REP received.\n");
1266 		ret = ipoib_cm_rep_handler(cm_id, event);
1267 		if (ret)
1268 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1269 				       NULL, 0, NULL, 0);
1270 		break;
1271 	case IB_CM_REQ_ERROR:
1272 	case IB_CM_REJ_RECEIVED:
1273 	case IB_CM_TIMEWAIT_EXIT:
1274 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1275 		netif_tx_lock_bh(dev);
1276 		spin_lock_irqsave(&priv->lock, flags);
1277 		neigh = tx->neigh;
1278 
1279 		if (neigh) {
1280 			neigh->cm = NULL;
1281 			ipoib_neigh_free(neigh);
1282 
1283 			tx->neigh = NULL;
1284 		}
1285 
1286 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1287 			list_move(&tx->list, &priv->cm.reap_list);
1288 			queue_work(priv->wq, &priv->cm.reap_task);
1289 		}
1290 
1291 		spin_unlock_irqrestore(&priv->lock, flags);
1292 		netif_tx_unlock_bh(dev);
1293 		break;
1294 	default:
1295 		break;
1296 	}
1297 
1298 	return 0;
1299 }
1300 
1301 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1302 				       struct ipoib_neigh *neigh)
1303 {
1304 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1305 	struct ipoib_cm_tx *tx;
1306 
1307 	tx = kzalloc(sizeof(*tx), GFP_ATOMIC);
1308 	if (!tx)
1309 		return NULL;
1310 
1311 	neigh->cm = tx;
1312 	tx->neigh = neigh;
1313 	tx->dev = dev;
1314 	list_add(&tx->list, &priv->cm.start_list);
1315 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1316 	queue_work(priv->wq, &priv->cm.start_task);
1317 	return tx;
1318 }
1319 
1320 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1321 {
1322 	struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1323 	unsigned long flags;
1324 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1325 		spin_lock_irqsave(&priv->lock, flags);
1326 		list_move(&tx->list, &priv->cm.reap_list);
1327 		queue_work(priv->wq, &priv->cm.reap_task);
1328 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1329 			  tx->neigh->daddr + 4);
1330 		tx->neigh = NULL;
1331 		spin_unlock_irqrestore(&priv->lock, flags);
1332 	}
1333 }
1334 
1335 #define QPN_AND_OPTIONS_OFFSET	4
1336 
1337 static void ipoib_cm_tx_start(struct work_struct *work)
1338 {
1339 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1340 						   cm.start_task);
1341 	struct net_device *dev = priv->dev;
1342 	struct ipoib_neigh *neigh;
1343 	struct ipoib_cm_tx *p;
1344 	unsigned long flags;
1345 	struct ipoib_path *path;
1346 	int ret;
1347 
1348 	struct sa_path_rec pathrec;
1349 	u32 qpn;
1350 
1351 	netif_tx_lock_bh(dev);
1352 	spin_lock_irqsave(&priv->lock, flags);
1353 
1354 	while (!list_empty(&priv->cm.start_list)) {
1355 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1356 		list_del_init(&p->list);
1357 		neigh = p->neigh;
1358 
1359 		qpn = IPOIB_QPN(neigh->daddr);
1360 		/*
1361 		 * As long as the search is with these 2 locks,
1362 		 * path existence indicates its validity.
1363 		 */
1364 		path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1365 		if (!path) {
1366 			pr_info("%s ignore not valid path %pI6\n",
1367 				__func__,
1368 				neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1369 			goto free_neigh;
1370 		}
1371 		memcpy(&pathrec, &path->pathrec, sizeof(pathrec));
1372 
1373 		spin_unlock_irqrestore(&priv->lock, flags);
1374 		netif_tx_unlock_bh(dev);
1375 
1376 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1377 
1378 		netif_tx_lock_bh(dev);
1379 		spin_lock_irqsave(&priv->lock, flags);
1380 
1381 		if (ret) {
1382 free_neigh:
1383 			neigh = p->neigh;
1384 			if (neigh) {
1385 				neigh->cm = NULL;
1386 				ipoib_neigh_free(neigh);
1387 			}
1388 			list_del(&p->list);
1389 			kfree(p);
1390 		}
1391 	}
1392 
1393 	spin_unlock_irqrestore(&priv->lock, flags);
1394 	netif_tx_unlock_bh(dev);
1395 }
1396 
1397 static void ipoib_cm_tx_reap(struct work_struct *work)
1398 {
1399 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1400 						   cm.reap_task);
1401 	struct net_device *dev = priv->dev;
1402 	struct ipoib_cm_tx *p;
1403 	unsigned long flags;
1404 
1405 	netif_tx_lock_bh(dev);
1406 	spin_lock_irqsave(&priv->lock, flags);
1407 
1408 	while (!list_empty(&priv->cm.reap_list)) {
1409 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1410 		list_del_init(&p->list);
1411 		spin_unlock_irqrestore(&priv->lock, flags);
1412 		netif_tx_unlock_bh(dev);
1413 		ipoib_cm_tx_destroy(p);
1414 		netif_tx_lock_bh(dev);
1415 		spin_lock_irqsave(&priv->lock, flags);
1416 	}
1417 
1418 	spin_unlock_irqrestore(&priv->lock, flags);
1419 	netif_tx_unlock_bh(dev);
1420 }
1421 
1422 static void ipoib_cm_skb_reap(struct work_struct *work)
1423 {
1424 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1425 						   cm.skb_task);
1426 	struct net_device *dev = priv->dev;
1427 	struct sk_buff *skb;
1428 	unsigned long flags;
1429 	unsigned int mtu = priv->mcast_mtu;
1430 
1431 	netif_tx_lock_bh(dev);
1432 	spin_lock_irqsave(&priv->lock, flags);
1433 
1434 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1435 		spin_unlock_irqrestore(&priv->lock, flags);
1436 		netif_tx_unlock_bh(dev);
1437 
1438 		if (skb->protocol == htons(ETH_P_IP)) {
1439 			memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1440 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1441 		}
1442 #if IS_ENABLED(CONFIG_IPV6)
1443 		else if (skb->protocol == htons(ETH_P_IPV6)) {
1444 			memset(IP6CB(skb), 0, sizeof(*IP6CB(skb)));
1445 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1446 		}
1447 #endif
1448 		dev_kfree_skb_any(skb);
1449 
1450 		netif_tx_lock_bh(dev);
1451 		spin_lock_irqsave(&priv->lock, flags);
1452 	}
1453 
1454 	spin_unlock_irqrestore(&priv->lock, flags);
1455 	netif_tx_unlock_bh(dev);
1456 }
1457 
1458 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1459 			   unsigned int mtu)
1460 {
1461 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1462 	int e = skb_queue_empty(&priv->cm.skb_queue);
1463 
1464 	skb_dst_update_pmtu(skb, mtu);
1465 
1466 	skb_queue_tail(&priv->cm.skb_queue, skb);
1467 	if (e)
1468 		queue_work(priv->wq, &priv->cm.skb_task);
1469 }
1470 
1471 static void ipoib_cm_rx_reap(struct work_struct *work)
1472 {
1473 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1474 						cm.rx_reap_task)->dev);
1475 }
1476 
1477 static void ipoib_cm_stale_task(struct work_struct *work)
1478 {
1479 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1480 						   cm.stale_task.work);
1481 	struct ipoib_cm_rx *p;
1482 	int ret;
1483 
1484 	spin_lock_irq(&priv->lock);
1485 	while (!list_empty(&priv->cm.passive_ids)) {
1486 		/* List is sorted by LRU, start from tail,
1487 		 * stop when we see a recently used entry */
1488 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1489 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1490 			break;
1491 		list_move(&p->list, &priv->cm.rx_error_list);
1492 		p->state = IPOIB_CM_RX_ERROR;
1493 		spin_unlock_irq(&priv->lock);
1494 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1495 		if (ret)
1496 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1497 		spin_lock_irq(&priv->lock);
1498 	}
1499 
1500 	if (!list_empty(&priv->cm.passive_ids))
1501 		queue_delayed_work(priv->wq,
1502 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1503 	spin_unlock_irq(&priv->lock);
1504 }
1505 
1506 static ssize_t mode_show(struct device *d, struct device_attribute *attr,
1507 			 char *buf)
1508 {
1509 	struct net_device *dev = to_net_dev(d);
1510 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1511 
1512 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1513 		return sysfs_emit(buf, "connected\n");
1514 	else
1515 		return sysfs_emit(buf, "datagram\n");
1516 }
1517 
1518 static ssize_t mode_store(struct device *d, struct device_attribute *attr,
1519 			  const char *buf, size_t count)
1520 {
1521 	struct net_device *dev = to_net_dev(d);
1522 	int ret;
1523 
1524 	if (!rtnl_trylock()) {
1525 		return restart_syscall();
1526 	}
1527 
1528 	if (dev->reg_state != NETREG_REGISTERED) {
1529 		rtnl_unlock();
1530 		return -EPERM;
1531 	}
1532 
1533 	ret = ipoib_set_mode(dev, buf);
1534 
1535 	/* The assumption is that the function ipoib_set_mode returned
1536 	 * with the rtnl held by it, if not the value -EBUSY returned,
1537 	 * then no need to rtnl_unlock
1538 	 */
1539 	if (ret != -EBUSY)
1540 		rtnl_unlock();
1541 
1542 	return (!ret || ret == -EBUSY) ? count : ret;
1543 }
1544 
1545 static DEVICE_ATTR_RW(mode);
1546 
1547 int ipoib_cm_add_mode_attr(struct net_device *dev)
1548 {
1549 	return device_create_file(&dev->dev, &dev_attr_mode);
1550 }
1551 
1552 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1553 {
1554 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1555 	struct ib_srq_init_attr srq_init_attr = {
1556 		.srq_type = IB_SRQT_BASIC,
1557 		.attr = {
1558 			.max_wr  = ipoib_recvq_size,
1559 			.max_sge = max_sge
1560 		}
1561 	};
1562 
1563 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1564 	if (IS_ERR(priv->cm.srq)) {
1565 		if (PTR_ERR(priv->cm.srq) != -EOPNOTSUPP)
1566 			pr_warn("%s: failed to allocate SRQ, error %ld\n",
1567 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1568 		priv->cm.srq = NULL;
1569 		return;
1570 	}
1571 
1572 	priv->cm.srq_ring = vzalloc(array_size(ipoib_recvq_size,
1573 					       sizeof(*priv->cm.srq_ring)));
1574 	if (!priv->cm.srq_ring) {
1575 		ib_destroy_srq(priv->cm.srq);
1576 		priv->cm.srq = NULL;
1577 		return;
1578 	}
1579 
1580 }
1581 
1582 int ipoib_cm_dev_init(struct net_device *dev)
1583 {
1584 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1585 	int max_srq_sge, i;
1586 	u8 addr;
1587 
1588 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1589 	INIT_LIST_HEAD(&priv->cm.reap_list);
1590 	INIT_LIST_HEAD(&priv->cm.start_list);
1591 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1592 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1593 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1594 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1595 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1596 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1597 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1598 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1599 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1600 
1601 	skb_queue_head_init(&priv->cm.skb_queue);
1602 
1603 	ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1604 
1605 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1606 	ipoib_cm_create_srq(dev, max_srq_sge);
1607 	if (ipoib_cm_has_srq(dev)) {
1608 		priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1609 		priv->cm.num_frags  = max_srq_sge;
1610 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1611 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1612 	} else {
1613 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1614 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1615 	}
1616 
1617 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1618 
1619 	if (ipoib_cm_has_srq(dev)) {
1620 		for (i = 0; i < ipoib_recvq_size; ++i) {
1621 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1622 						   priv->cm.num_frags - 1,
1623 						   priv->cm.srq_ring[i].mapping,
1624 						   GFP_KERNEL)) {
1625 				ipoib_warn(priv, "failed to allocate "
1626 					   "receive buffer %d\n", i);
1627 				ipoib_cm_dev_cleanup(dev);
1628 				return -ENOMEM;
1629 			}
1630 
1631 			if (ipoib_cm_post_receive_srq(dev, i)) {
1632 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1633 					   "failed for buf %d\n", i);
1634 				ipoib_cm_dev_cleanup(dev);
1635 				return -EIO;
1636 			}
1637 		}
1638 	}
1639 
1640 	addr = IPOIB_FLAGS_RC;
1641 	dev_addr_mod(dev, 0, &addr, 1);
1642 	return 0;
1643 }
1644 
1645 void ipoib_cm_dev_cleanup(struct net_device *dev)
1646 {
1647 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1648 
1649 	if (!priv->cm.srq)
1650 		return;
1651 
1652 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1653 
1654 	ib_destroy_srq(priv->cm.srq);
1655 	priv->cm.srq = NULL;
1656 	if (!priv->cm.srq_ring)
1657 		return;
1658 
1659 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1660 	priv->cm.srq_ring = NULL;
1661 }
1662