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 			       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 	struct ib_recv_wr *bad_wr;
98 	int i, ret;
99 
100 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
101 
102 	for (i = 0; i < priv->cm.num_frags; ++i)
103 		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
104 
105 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
106 	if (unlikely(ret)) {
107 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
108 		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
109 				      priv->cm.srq_ring[id].mapping);
110 		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
111 		priv->cm.srq_ring[id].skb = NULL;
112 	}
113 
114 	return ret;
115 }
116 
117 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
118 					struct ipoib_cm_rx *rx,
119 					struct ib_recv_wr *wr,
120 					struct ib_sge *sge, int id)
121 {
122 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
123 	struct ib_recv_wr *bad_wr;
124 	int i, ret;
125 
126 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
127 
128 	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
129 		sge[i].addr = rx->rx_ring[id].mapping[i];
130 
131 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
132 	if (unlikely(ret)) {
133 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
134 		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
135 				      rx->rx_ring[id].mapping);
136 		dev_kfree_skb_any(rx->rx_ring[id].skb);
137 		rx->rx_ring[id].skb = NULL;
138 	}
139 
140 	return ret;
141 }
142 
143 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
144 					     struct ipoib_cm_rx_buf *rx_ring,
145 					     int id, int frags,
146 					     u64 mapping[IPOIB_CM_RX_SG],
147 					     gfp_t gfp)
148 {
149 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
150 	struct sk_buff *skb;
151 	int i;
152 
153 	skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
154 	if (unlikely(!skb))
155 		return NULL;
156 
157 	/*
158 	 * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
159 	 * IP header to a multiple of 16.
160 	 */
161 	skb_reserve(skb, IPOIB_CM_RX_RESERVE);
162 
163 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
164 				       DMA_FROM_DEVICE);
165 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
166 		dev_kfree_skb_any(skb);
167 		return NULL;
168 	}
169 
170 	for (i = 0; i < frags; i++) {
171 		struct page *page = alloc_page(gfp);
172 
173 		if (!page)
174 			goto partial_error;
175 		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
176 
177 		mapping[i + 1] = ib_dma_map_page(priv->ca, page,
178 						 0, PAGE_SIZE, DMA_FROM_DEVICE);
179 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
180 			goto partial_error;
181 	}
182 
183 	rx_ring[id].skb = skb;
184 	return skb;
185 
186 partial_error:
187 
188 	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
189 
190 	for (; i > 0; --i)
191 		ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
192 
193 	dev_kfree_skb_any(skb);
194 	return NULL;
195 }
196 
197 static void ipoib_cm_free_rx_ring(struct net_device *dev,
198 				  struct ipoib_cm_rx_buf *rx_ring)
199 {
200 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
201 	int i;
202 
203 	for (i = 0; i < ipoib_recvq_size; ++i)
204 		if (rx_ring[i].skb) {
205 			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
206 					      rx_ring[i].mapping);
207 			dev_kfree_skb_any(rx_ring[i].skb);
208 		}
209 
210 	vfree(rx_ring);
211 }
212 
213 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
214 {
215 	struct ib_send_wr *bad_wr;
216 	struct ipoib_cm_rx *p;
217 
218 	/* We only reserved 1 extra slot in CQ for drain WRs, so
219 	 * make sure we have at most 1 outstanding WR. */
220 	if (list_empty(&priv->cm.rx_flush_list) ||
221 	    !list_empty(&priv->cm.rx_drain_list))
222 		return;
223 
224 	/*
225 	 * QPs on flush list are error state.  This way, a "flush
226 	 * error" WC will be immediately generated for each WR we post.
227 	 */
228 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
229 	ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
230 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
231 		ipoib_warn(priv, "failed to post drain wr\n");
232 
233 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
234 }
235 
236 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
237 {
238 	struct ipoib_cm_rx *p = ctx;
239 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
240 	unsigned long flags;
241 
242 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
243 		return;
244 
245 	spin_lock_irqsave(&priv->lock, flags);
246 	list_move(&p->list, &priv->cm.rx_flush_list);
247 	p->state = IPOIB_CM_RX_FLUSH;
248 	ipoib_cm_start_rx_drain(priv);
249 	spin_unlock_irqrestore(&priv->lock, flags);
250 }
251 
252 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
253 					   struct ipoib_cm_rx *p)
254 {
255 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
256 	struct ib_qp_init_attr attr = {
257 		.event_handler = ipoib_cm_rx_event_handler,
258 		.send_cq = priv->recv_cq, /* For drain WR */
259 		.recv_cq = priv->recv_cq,
260 		.srq = priv->cm.srq,
261 		.cap.max_send_wr = 1, /* For drain WR */
262 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
263 		.sq_sig_type = IB_SIGNAL_ALL_WR,
264 		.qp_type = IB_QPT_RC,
265 		.qp_context = p,
266 	};
267 
268 	if (!ipoib_cm_has_srq(dev)) {
269 		attr.cap.max_recv_wr  = ipoib_recvq_size;
270 		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
271 	}
272 
273 	return ib_create_qp(priv->pd, &attr);
274 }
275 
276 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
277 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
278 				 unsigned psn)
279 {
280 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
281 	struct ib_qp_attr qp_attr;
282 	int qp_attr_mask, ret;
283 
284 	qp_attr.qp_state = IB_QPS_INIT;
285 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
286 	if (ret) {
287 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
288 		return ret;
289 	}
290 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
291 	if (ret) {
292 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
293 		return ret;
294 	}
295 	qp_attr.qp_state = IB_QPS_RTR;
296 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
297 	if (ret) {
298 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
299 		return ret;
300 	}
301 	qp_attr.rq_psn = psn;
302 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
303 	if (ret) {
304 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
305 		return ret;
306 	}
307 
308 	/*
309 	 * Current Mellanox HCA firmware won't generate completions
310 	 * with error for drain WRs unless the QP has been moved to
311 	 * RTS first. This work-around leaves a window where a QP has
312 	 * moved to error asynchronously, but this will eventually get
313 	 * fixed in firmware, so let's not error out if modify QP
314 	 * fails.
315 	 */
316 	qp_attr.qp_state = IB_QPS_RTS;
317 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
318 	if (ret) {
319 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
320 		return 0;
321 	}
322 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
323 	if (ret) {
324 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
325 		return 0;
326 	}
327 
328 	return 0;
329 }
330 
331 static void ipoib_cm_init_rx_wr(struct net_device *dev,
332 				struct ib_recv_wr *wr,
333 				struct ib_sge *sge)
334 {
335 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
336 	int i;
337 
338 	for (i = 0; i < priv->cm.num_frags; ++i)
339 		sge[i].lkey = priv->pd->local_dma_lkey;
340 
341 	sge[0].length = IPOIB_CM_HEAD_SIZE;
342 	for (i = 1; i < priv->cm.num_frags; ++i)
343 		sge[i].length = PAGE_SIZE;
344 
345 	wr->next    = NULL;
346 	wr->sg_list = sge;
347 	wr->num_sge = priv->cm.num_frags;
348 }
349 
350 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
351 				   struct ipoib_cm_rx *rx)
352 {
353 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
354 	struct {
355 		struct ib_recv_wr wr;
356 		struct ib_sge sge[IPOIB_CM_RX_SG];
357 	} *t;
358 	int ret;
359 	int i;
360 
361 	rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
362 	if (!rx->rx_ring)
363 		return -ENOMEM;
364 
365 	t = kmalloc(sizeof *t, GFP_KERNEL);
366 	if (!t) {
367 		ret = -ENOMEM;
368 		goto err_free_1;
369 	}
370 
371 	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
372 
373 	spin_lock_irq(&priv->lock);
374 
375 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
376 		spin_unlock_irq(&priv->lock);
377 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
378 		ret = -EINVAL;
379 		goto err_free;
380 	} else
381 		++priv->cm.nonsrq_conn_qp;
382 
383 	spin_unlock_irq(&priv->lock);
384 
385 	for (i = 0; i < ipoib_recvq_size; ++i) {
386 		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
387 					   rx->rx_ring[i].mapping,
388 					   GFP_KERNEL)) {
389 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
390 			ret = -ENOMEM;
391 			goto err_count;
392 		}
393 		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
394 		if (ret) {
395 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
396 				   "failed for buf %d\n", i);
397 			ret = -EIO;
398 			goto err_count;
399 		}
400 	}
401 
402 	rx->recv_count = ipoib_recvq_size;
403 
404 	kfree(t);
405 
406 	return 0;
407 
408 err_count:
409 	spin_lock_irq(&priv->lock);
410 	--priv->cm.nonsrq_conn_qp;
411 	spin_unlock_irq(&priv->lock);
412 
413 err_free:
414 	kfree(t);
415 
416 err_free_1:
417 	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
418 
419 	return ret;
420 }
421 
422 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
423 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
424 			     unsigned psn)
425 {
426 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
427 	struct ipoib_cm_data data = {};
428 	struct ib_cm_rep_param rep = {};
429 
430 	data.qpn = cpu_to_be32(priv->qp->qp_num);
431 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
432 
433 	rep.private_data = &data;
434 	rep.private_data_len = sizeof data;
435 	rep.flow_control = 0;
436 	rep.rnr_retry_count = req->rnr_retry_count;
437 	rep.srq = ipoib_cm_has_srq(dev);
438 	rep.qp_num = qp->qp_num;
439 	rep.starting_psn = psn;
440 	return ib_send_cm_rep(cm_id, &rep);
441 }
442 
443 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, 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 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 = prandom_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 			       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 		/* Fall through */
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 		/* 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 = 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, "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 = ipoib_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 = ipoib_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 		/* IB_WC[_RNR]_RETRY_EXC_ERR error is part of the life cycle,
827 		 * so don't make waves.
828 		 */
829 		if (wc->status == IB_WC_RNR_RETRY_EXC_ERR ||
830 		    wc->status == IB_WC_RETRY_EXC_ERR)
831 			ipoib_dbg(priv,
832 				  "%s: failed cm send event (status=%d, wrid=%d vend_err 0x%x)\n",
833 				   __func__, wc->status, wr_id, wc->vendor_err);
834 		else
835 			ipoib_warn(priv,
836 				    "%s: failed cm send event (status=%d, wrid=%d vend_err 0x%x)\n",
837 				   __func__, wc->status, wr_id, wc->vendor_err);
838 
839 		spin_lock_irqsave(&priv->lock, flags);
840 		neigh = tx->neigh;
841 
842 		if (neigh) {
843 			neigh->cm = NULL;
844 			ipoib_neigh_free(neigh);
845 
846 			tx->neigh = NULL;
847 		}
848 
849 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
850 			list_move(&tx->list, &priv->cm.reap_list);
851 			queue_work(priv->wq, &priv->cm.reap_task);
852 		}
853 
854 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
855 
856 		spin_unlock_irqrestore(&priv->lock, flags);
857 	}
858 
859 	netif_tx_unlock(dev);
860 }
861 
862 int ipoib_cm_dev_open(struct net_device *dev)
863 {
864 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
865 	int ret;
866 
867 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
868 		return 0;
869 
870 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
871 	if (IS_ERR(priv->cm.id)) {
872 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
873 		ret = PTR_ERR(priv->cm.id);
874 		goto err_cm;
875 	}
876 
877 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
878 			   0);
879 	if (ret) {
880 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
881 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
882 		goto err_listen;
883 	}
884 
885 	return 0;
886 
887 err_listen:
888 	ib_destroy_cm_id(priv->cm.id);
889 err_cm:
890 	priv->cm.id = NULL;
891 	return ret;
892 }
893 
894 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
895 {
896 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
897 	struct ipoib_cm_rx *rx, *n;
898 	LIST_HEAD(list);
899 
900 	spin_lock_irq(&priv->lock);
901 	list_splice_init(&priv->cm.rx_reap_list, &list);
902 	spin_unlock_irq(&priv->lock);
903 
904 	list_for_each_entry_safe(rx, n, &list, list) {
905 		ib_destroy_cm_id(rx->id);
906 		ib_destroy_qp(rx->qp);
907 		if (!ipoib_cm_has_srq(dev)) {
908 			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
909 			spin_lock_irq(&priv->lock);
910 			--priv->cm.nonsrq_conn_qp;
911 			spin_unlock_irq(&priv->lock);
912 		}
913 		kfree(rx);
914 	}
915 }
916 
917 void ipoib_cm_dev_stop(struct net_device *dev)
918 {
919 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
920 	struct ipoib_cm_rx *p;
921 	unsigned long begin;
922 	int ret;
923 
924 	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
925 		return;
926 
927 	ib_destroy_cm_id(priv->cm.id);
928 	priv->cm.id = NULL;
929 
930 	spin_lock_irq(&priv->lock);
931 	while (!list_empty(&priv->cm.passive_ids)) {
932 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
933 		list_move(&p->list, &priv->cm.rx_error_list);
934 		p->state = IPOIB_CM_RX_ERROR;
935 		spin_unlock_irq(&priv->lock);
936 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
937 		if (ret)
938 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
939 		spin_lock_irq(&priv->lock);
940 	}
941 
942 	/* Wait for all RX to be drained */
943 	begin = jiffies;
944 
945 	while (!list_empty(&priv->cm.rx_error_list) ||
946 	       !list_empty(&priv->cm.rx_flush_list) ||
947 	       !list_empty(&priv->cm.rx_drain_list)) {
948 		if (time_after(jiffies, begin + 5 * HZ)) {
949 			ipoib_warn(priv, "RX drain timing out\n");
950 
951 			/*
952 			 * assume the HW is wedged and just free up everything.
953 			 */
954 			list_splice_init(&priv->cm.rx_flush_list,
955 					 &priv->cm.rx_reap_list);
956 			list_splice_init(&priv->cm.rx_error_list,
957 					 &priv->cm.rx_reap_list);
958 			list_splice_init(&priv->cm.rx_drain_list,
959 					 &priv->cm.rx_reap_list);
960 			break;
961 		}
962 		spin_unlock_irq(&priv->lock);
963 		usleep_range(1000, 2000);
964 		ipoib_drain_cq(dev);
965 		spin_lock_irq(&priv->lock);
966 	}
967 
968 	spin_unlock_irq(&priv->lock);
969 
970 	ipoib_cm_free_rx_reap_list(dev);
971 
972 	cancel_delayed_work(&priv->cm.stale_task);
973 }
974 
975 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
976 {
977 	struct ipoib_cm_tx *p = cm_id->context;
978 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
979 	struct ipoib_cm_data *data = event->private_data;
980 	struct sk_buff_head skqueue;
981 	struct ib_qp_attr qp_attr;
982 	int qp_attr_mask, ret;
983 	struct sk_buff *skb;
984 
985 	p->mtu = be32_to_cpu(data->mtu);
986 
987 	if (p->mtu <= IPOIB_ENCAP_LEN) {
988 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
989 			   p->mtu, IPOIB_ENCAP_LEN);
990 		return -EINVAL;
991 	}
992 
993 	qp_attr.qp_state = IB_QPS_RTR;
994 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
995 	if (ret) {
996 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
997 		return ret;
998 	}
999 
1000 	qp_attr.rq_psn = 0 /* FIXME */;
1001 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1002 	if (ret) {
1003 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
1004 		return ret;
1005 	}
1006 
1007 	qp_attr.qp_state = IB_QPS_RTS;
1008 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1009 	if (ret) {
1010 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1011 		return ret;
1012 	}
1013 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1014 	if (ret) {
1015 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1016 		return ret;
1017 	}
1018 
1019 	skb_queue_head_init(&skqueue);
1020 
1021 	spin_lock_irq(&priv->lock);
1022 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1023 	if (p->neigh)
1024 		while ((skb = __skb_dequeue(&p->neigh->queue)))
1025 			__skb_queue_tail(&skqueue, skb);
1026 	spin_unlock_irq(&priv->lock);
1027 
1028 	while ((skb = __skb_dequeue(&skqueue))) {
1029 		skb->dev = p->dev;
1030 		ret = dev_queue_xmit(skb);
1031 		if (ret)
1032 			ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1033 				   __func__, ret);
1034 	}
1035 
1036 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1037 	if (ret) {
1038 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1039 		return ret;
1040 	}
1041 	return 0;
1042 }
1043 
1044 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1045 {
1046 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1047 	struct ib_qp_init_attr attr = {
1048 		.send_cq		= priv->recv_cq,
1049 		.recv_cq		= priv->recv_cq,
1050 		.srq			= priv->cm.srq,
1051 		.cap.max_send_wr	= ipoib_sendq_size,
1052 		.cap.max_send_sge	= 1,
1053 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1054 		.qp_type		= IB_QPT_RC,
1055 		.qp_context		= tx,
1056 		.create_flags		= 0
1057 	};
1058 	struct ib_qp *tx_qp;
1059 
1060 	if (dev->features & NETIF_F_SG)
1061 		attr.cap.max_send_sge =
1062 			min_t(u32, priv->ca->attrs.max_sge, MAX_SKB_FRAGS + 1);
1063 
1064 	tx_qp = ib_create_qp(priv->pd, &attr);
1065 	tx->max_send_sge = attr.cap.max_send_sge;
1066 	return tx_qp;
1067 }
1068 
1069 static int ipoib_cm_send_req(struct net_device *dev,
1070 			     struct ib_cm_id *id, struct ib_qp *qp,
1071 			     u32 qpn,
1072 			     struct sa_path_rec *pathrec)
1073 {
1074 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1075 	struct ipoib_cm_data data = {};
1076 	struct ib_cm_req_param req = {};
1077 
1078 	data.qpn = cpu_to_be32(priv->qp->qp_num);
1079 	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1080 
1081 	req.primary_path		= pathrec;
1082 	req.alternate_path		= NULL;
1083 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1084 	req.qp_num			= qp->qp_num;
1085 	req.qp_type			= qp->qp_type;
1086 	req.private_data		= &data;
1087 	req.private_data_len		= sizeof data;
1088 	req.flow_control		= 0;
1089 
1090 	req.starting_psn		= 0; /* FIXME */
1091 
1092 	/*
1093 	 * Pick some arbitrary defaults here; we could make these
1094 	 * module parameters if anyone cared about setting them.
1095 	 */
1096 	req.responder_resources		= 4;
1097 	req.remote_cm_response_timeout	= 20;
1098 	req.local_cm_response_timeout	= 20;
1099 	req.retry_count			= 0; /* RFC draft warns against retries */
1100 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1101 	req.max_cm_retries		= 15;
1102 	req.srq				= ipoib_cm_has_srq(dev);
1103 	return ib_send_cm_req(id, &req);
1104 }
1105 
1106 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1107 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1108 {
1109 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1110 	struct ib_qp_attr qp_attr;
1111 	int qp_attr_mask, ret;
1112 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1113 	if (ret) {
1114 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1115 		return ret;
1116 	}
1117 
1118 	qp_attr.qp_state = IB_QPS_INIT;
1119 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1120 	qp_attr.port_num = priv->port;
1121 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1122 
1123 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1124 	if (ret) {
1125 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1126 		return ret;
1127 	}
1128 	return 0;
1129 }
1130 
1131 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1132 			    struct sa_path_rec *pathrec)
1133 {
1134 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1135 	unsigned int noio_flag;
1136 	int ret;
1137 
1138 	noio_flag = memalloc_noio_save();
1139 	p->tx_ring = vzalloc(ipoib_sendq_size * sizeof(*p->tx_ring));
1140 	if (!p->tx_ring) {
1141 		ret = -ENOMEM;
1142 		goto err_tx;
1143 	}
1144 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1145 
1146 	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1147 	memalloc_noio_restore(noio_flag);
1148 	if (IS_ERR(p->qp)) {
1149 		ret = PTR_ERR(p->qp);
1150 		ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1151 		goto err_qp;
1152 	}
1153 
1154 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1155 	if (IS_ERR(p->id)) {
1156 		ret = PTR_ERR(p->id);
1157 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1158 		goto err_id;
1159 	}
1160 
1161 	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1162 	if (ret) {
1163 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1164 		goto err_modify_send;
1165 	}
1166 
1167 	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1168 	if (ret) {
1169 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1170 		goto err_modify_send;
1171 	}
1172 
1173 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1174 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1175 
1176 	return 0;
1177 
1178 err_modify_send:
1179 	ib_destroy_cm_id(p->id);
1180 err_id:
1181 	p->id = NULL;
1182 	ib_destroy_qp(p->qp);
1183 err_qp:
1184 	p->qp = NULL;
1185 	vfree(p->tx_ring);
1186 err_tx:
1187 	return ret;
1188 }
1189 
1190 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1191 {
1192 	struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1193 	struct ipoib_tx_buf *tx_req;
1194 	unsigned long begin;
1195 
1196 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1197 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1198 
1199 	if (p->id)
1200 		ib_destroy_cm_id(p->id);
1201 
1202 	if (p->tx_ring) {
1203 		/* Wait for all sends to complete */
1204 		begin = jiffies;
1205 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1206 			if (time_after(jiffies, begin + 5 * HZ)) {
1207 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1208 					   p->tx_head - p->tx_tail);
1209 				goto timeout;
1210 			}
1211 
1212 			usleep_range(1000, 2000);
1213 		}
1214 	}
1215 
1216 timeout:
1217 
1218 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1219 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1220 		ipoib_dma_unmap_tx(priv, tx_req);
1221 		dev_kfree_skb_any(tx_req->skb);
1222 		++p->tx_tail;
1223 		netif_tx_lock_bh(p->dev);
1224 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1225 		    netif_queue_stopped(p->dev) &&
1226 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1227 			netif_wake_queue(p->dev);
1228 		netif_tx_unlock_bh(p->dev);
1229 	}
1230 
1231 	if (p->qp)
1232 		ib_destroy_qp(p->qp);
1233 
1234 	vfree(p->tx_ring);
1235 	kfree(p);
1236 }
1237 
1238 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1239 			       struct ib_cm_event *event)
1240 {
1241 	struct ipoib_cm_tx *tx = cm_id->context;
1242 	struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1243 	struct net_device *dev = priv->dev;
1244 	struct ipoib_neigh *neigh;
1245 	unsigned long flags;
1246 	int ret;
1247 
1248 	switch (event->event) {
1249 	case IB_CM_DREQ_RECEIVED:
1250 		ipoib_dbg(priv, "DREQ received.\n");
1251 		ib_send_cm_drep(cm_id, NULL, 0);
1252 		break;
1253 	case IB_CM_REP_RECEIVED:
1254 		ipoib_dbg(priv, "REP received.\n");
1255 		ret = ipoib_cm_rep_handler(cm_id, event);
1256 		if (ret)
1257 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1258 				       NULL, 0, NULL, 0);
1259 		break;
1260 	case IB_CM_REQ_ERROR:
1261 	case IB_CM_REJ_RECEIVED:
1262 	case IB_CM_TIMEWAIT_EXIT:
1263 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1264 		netif_tx_lock_bh(dev);
1265 		spin_lock_irqsave(&priv->lock, flags);
1266 		neigh = tx->neigh;
1267 
1268 		if (neigh) {
1269 			neigh->cm = NULL;
1270 			ipoib_neigh_free(neigh);
1271 
1272 			tx->neigh = NULL;
1273 		}
1274 
1275 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1276 			list_move(&tx->list, &priv->cm.reap_list);
1277 			queue_work(priv->wq, &priv->cm.reap_task);
1278 		}
1279 
1280 		spin_unlock_irqrestore(&priv->lock, flags);
1281 		netif_tx_unlock_bh(dev);
1282 		break;
1283 	default:
1284 		break;
1285 	}
1286 
1287 	return 0;
1288 }
1289 
1290 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1291 				       struct ipoib_neigh *neigh)
1292 {
1293 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1294 	struct ipoib_cm_tx *tx;
1295 
1296 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1297 	if (!tx)
1298 		return NULL;
1299 
1300 	neigh->cm = tx;
1301 	tx->neigh = neigh;
1302 	tx->path = path;
1303 	tx->dev = dev;
1304 	list_add(&tx->list, &priv->cm.start_list);
1305 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1306 	queue_work(priv->wq, &priv->cm.start_task);
1307 	return tx;
1308 }
1309 
1310 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1311 {
1312 	struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1313 	unsigned long flags;
1314 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1315 		spin_lock_irqsave(&priv->lock, flags);
1316 		list_move(&tx->list, &priv->cm.reap_list);
1317 		queue_work(priv->wq, &priv->cm.reap_task);
1318 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1319 			  tx->neigh->daddr + 4);
1320 		tx->neigh = NULL;
1321 		spin_unlock_irqrestore(&priv->lock, flags);
1322 	}
1323 }
1324 
1325 #define QPN_AND_OPTIONS_OFFSET	4
1326 
1327 static void ipoib_cm_tx_start(struct work_struct *work)
1328 {
1329 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1330 						   cm.start_task);
1331 	struct net_device *dev = priv->dev;
1332 	struct ipoib_neigh *neigh;
1333 	struct ipoib_cm_tx *p;
1334 	unsigned long flags;
1335 	struct ipoib_path *path;
1336 	int ret;
1337 
1338 	struct sa_path_rec pathrec;
1339 	u32 qpn;
1340 
1341 	netif_tx_lock_bh(dev);
1342 	spin_lock_irqsave(&priv->lock, flags);
1343 
1344 	while (!list_empty(&priv->cm.start_list)) {
1345 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1346 		list_del_init(&p->list);
1347 		neigh = p->neigh;
1348 
1349 		qpn = IPOIB_QPN(neigh->daddr);
1350 		/*
1351 		 * As long as the search is with these 2 locks,
1352 		 * path existence indicates its validity.
1353 		 */
1354 		path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1355 		if (!path) {
1356 			pr_info("%s ignore not valid path %pI6\n",
1357 				__func__,
1358 				neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1359 			goto free_neigh;
1360 		}
1361 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1362 
1363 		spin_unlock_irqrestore(&priv->lock, flags);
1364 		netif_tx_unlock_bh(dev);
1365 
1366 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1367 
1368 		netif_tx_lock_bh(dev);
1369 		spin_lock_irqsave(&priv->lock, flags);
1370 
1371 		if (ret) {
1372 free_neigh:
1373 			neigh = p->neigh;
1374 			if (neigh) {
1375 				neigh->cm = NULL;
1376 				ipoib_neigh_free(neigh);
1377 			}
1378 			list_del(&p->list);
1379 			kfree(p);
1380 		}
1381 	}
1382 
1383 	spin_unlock_irqrestore(&priv->lock, flags);
1384 	netif_tx_unlock_bh(dev);
1385 }
1386 
1387 static void ipoib_cm_tx_reap(struct work_struct *work)
1388 {
1389 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1390 						   cm.reap_task);
1391 	struct net_device *dev = priv->dev;
1392 	struct ipoib_cm_tx *p;
1393 	unsigned long flags;
1394 
1395 	netif_tx_lock_bh(dev);
1396 	spin_lock_irqsave(&priv->lock, flags);
1397 
1398 	while (!list_empty(&priv->cm.reap_list)) {
1399 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1400 		list_del_init(&p->list);
1401 		spin_unlock_irqrestore(&priv->lock, flags);
1402 		netif_tx_unlock_bh(dev);
1403 		ipoib_cm_tx_destroy(p);
1404 		netif_tx_lock_bh(dev);
1405 		spin_lock_irqsave(&priv->lock, flags);
1406 	}
1407 
1408 	spin_unlock_irqrestore(&priv->lock, flags);
1409 	netif_tx_unlock_bh(dev);
1410 }
1411 
1412 static void ipoib_cm_skb_reap(struct work_struct *work)
1413 {
1414 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1415 						   cm.skb_task);
1416 	struct net_device *dev = priv->dev;
1417 	struct sk_buff *skb;
1418 	unsigned long flags;
1419 	unsigned mtu = priv->mcast_mtu;
1420 
1421 	netif_tx_lock_bh(dev);
1422 	spin_lock_irqsave(&priv->lock, flags);
1423 
1424 	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1425 		spin_unlock_irqrestore(&priv->lock, flags);
1426 		netif_tx_unlock_bh(dev);
1427 
1428 		if (skb->protocol == htons(ETH_P_IP))
1429 			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1430 #if IS_ENABLED(CONFIG_IPV6)
1431 		else if (skb->protocol == htons(ETH_P_IPV6))
1432 			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1433 #endif
1434 		dev_kfree_skb_any(skb);
1435 
1436 		netif_tx_lock_bh(dev);
1437 		spin_lock_irqsave(&priv->lock, flags);
1438 	}
1439 
1440 	spin_unlock_irqrestore(&priv->lock, flags);
1441 	netif_tx_unlock_bh(dev);
1442 }
1443 
1444 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1445 			   unsigned int mtu)
1446 {
1447 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1448 	int e = skb_queue_empty(&priv->cm.skb_queue);
1449 
1450 	if (skb_dst(skb))
1451 		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1452 
1453 	skb_queue_tail(&priv->cm.skb_queue, skb);
1454 	if (e)
1455 		queue_work(priv->wq, &priv->cm.skb_task);
1456 }
1457 
1458 static void ipoib_cm_rx_reap(struct work_struct *work)
1459 {
1460 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1461 						cm.rx_reap_task)->dev);
1462 }
1463 
1464 static void ipoib_cm_stale_task(struct work_struct *work)
1465 {
1466 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1467 						   cm.stale_task.work);
1468 	struct ipoib_cm_rx *p;
1469 	int ret;
1470 
1471 	spin_lock_irq(&priv->lock);
1472 	while (!list_empty(&priv->cm.passive_ids)) {
1473 		/* List is sorted by LRU, start from tail,
1474 		 * stop when we see a recently used entry */
1475 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1476 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1477 			break;
1478 		list_move(&p->list, &priv->cm.rx_error_list);
1479 		p->state = IPOIB_CM_RX_ERROR;
1480 		spin_unlock_irq(&priv->lock);
1481 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1482 		if (ret)
1483 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1484 		spin_lock_irq(&priv->lock);
1485 	}
1486 
1487 	if (!list_empty(&priv->cm.passive_ids))
1488 		queue_delayed_work(priv->wq,
1489 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1490 	spin_unlock_irq(&priv->lock);
1491 }
1492 
1493 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1494 			 char *buf)
1495 {
1496 	struct net_device *dev = to_net_dev(d);
1497 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1498 
1499 	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1500 		return sprintf(buf, "connected\n");
1501 	else
1502 		return sprintf(buf, "datagram\n");
1503 }
1504 
1505 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1506 			const char *buf, size_t count)
1507 {
1508 	struct net_device *dev = to_net_dev(d);
1509 	int ret;
1510 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1511 
1512 	if (test_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags))
1513 		return -EPERM;
1514 
1515 	if (!mutex_trylock(&priv->sysfs_mutex))
1516 		return restart_syscall();
1517 
1518 	if (!rtnl_trylock()) {
1519 		mutex_unlock(&priv->sysfs_mutex);
1520 		return restart_syscall();
1521 	}
1522 
1523 	ret = ipoib_set_mode(dev, buf);
1524 
1525 	/* The assumption is that the function ipoib_set_mode returned
1526 	 * with the rtnl held by it, if not the value -EBUSY returned,
1527 	 * then no need to rtnl_unlock
1528 	 */
1529 	if (ret != -EBUSY)
1530 		rtnl_unlock();
1531 	mutex_unlock(&priv->sysfs_mutex);
1532 
1533 	return (!ret || ret == -EBUSY) ? count : ret;
1534 }
1535 
1536 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1537 
1538 int ipoib_cm_add_mode_attr(struct net_device *dev)
1539 {
1540 	return device_create_file(&dev->dev, &dev_attr_mode);
1541 }
1542 
1543 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1544 {
1545 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1546 	struct ib_srq_init_attr srq_init_attr = {
1547 		.srq_type = IB_SRQT_BASIC,
1548 		.attr = {
1549 			.max_wr  = ipoib_recvq_size,
1550 			.max_sge = max_sge
1551 		}
1552 	};
1553 
1554 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1555 	if (IS_ERR(priv->cm.srq)) {
1556 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1557 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1558 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1559 		priv->cm.srq = NULL;
1560 		return;
1561 	}
1562 
1563 	priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1564 	if (!priv->cm.srq_ring) {
1565 		ib_destroy_srq(priv->cm.srq);
1566 		priv->cm.srq = NULL;
1567 		return;
1568 	}
1569 
1570 }
1571 
1572 int ipoib_cm_dev_init(struct net_device *dev)
1573 {
1574 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1575 	int max_srq_sge, i;
1576 
1577 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1578 	INIT_LIST_HEAD(&priv->cm.reap_list);
1579 	INIT_LIST_HEAD(&priv->cm.start_list);
1580 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1581 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1582 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1583 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1584 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1585 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1586 	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1587 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1588 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1589 
1590 	skb_queue_head_init(&priv->cm.skb_queue);
1591 
1592 	ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1593 
1594 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1595 	ipoib_cm_create_srq(dev, max_srq_sge);
1596 	if (ipoib_cm_has_srq(dev)) {
1597 		priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1598 		priv->cm.num_frags  = max_srq_sge;
1599 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1600 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1601 	} else {
1602 		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1603 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1604 	}
1605 
1606 	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1607 
1608 	if (ipoib_cm_has_srq(dev)) {
1609 		for (i = 0; i < ipoib_recvq_size; ++i) {
1610 			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1611 						   priv->cm.num_frags - 1,
1612 						   priv->cm.srq_ring[i].mapping,
1613 						   GFP_KERNEL)) {
1614 				ipoib_warn(priv, "failed to allocate "
1615 					   "receive buffer %d\n", i);
1616 				ipoib_cm_dev_cleanup(dev);
1617 				return -ENOMEM;
1618 			}
1619 
1620 			if (ipoib_cm_post_receive_srq(dev, i)) {
1621 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1622 					   "failed for buf %d\n", i);
1623 				ipoib_cm_dev_cleanup(dev);
1624 				return -EIO;
1625 			}
1626 		}
1627 	}
1628 
1629 	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1630 	return 0;
1631 }
1632 
1633 void ipoib_cm_dev_cleanup(struct net_device *dev)
1634 {
1635 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1636 	int ret;
1637 
1638 	if (!priv->cm.srq)
1639 		return;
1640 
1641 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1642 
1643 	ret = ib_destroy_srq(priv->cm.srq);
1644 	if (ret)
1645 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1646 
1647 	priv->cm.srq = NULL;
1648 	if (!priv->cm.srq_ring)
1649 		return;
1650 
1651 	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1652 	priv->cm.srq_ring = NULL;
1653 }
1654