xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
3  *
4  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "ipoib.h"
39 
40 #ifdef CONFIG_INFINIBAND_IPOIB_CM
41 
42 #include <netinet/ip.h>
43 #include <netinet/ip_icmp.h>
44 #include <netinet/icmp6.h>
45 
46 #include <rdma/ib_cm.h>
47 #include <rdma/ib_cache.h>
48 #include <linux/delay.h>
49 
50 int ipoib_max_conn_qp = 128;
51 
52 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
53 MODULE_PARM_DESC(max_nonsrq_conn_qp,
54 		 "Max number of connected-mode QPs per interface "
55 		 "(applied only if shared receive queue is not available)");
56 
57 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
58 static int data_debug_level;
59 
60 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
61 MODULE_PARM_DESC(cm_data_debug_level,
62 		 "Enable data path debug tracing for connected mode if > 0");
63 #endif
64 
65 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
66 
67 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
68 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
69 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
70 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
71 
72 static struct ib_qp_attr ipoib_cm_err_attr = {
73 	.qp_state = IB_QPS_ERR
74 };
75 
76 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
77 
78 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
79 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
80 	.opcode = IB_WR_SEND,
81 };
82 
83 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
84 			       struct ib_cm_event *event);
85 
86 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
87 {
88 
89 	ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req);
90 
91 }
92 
93 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id)
94 {
95 	const struct ib_recv_wr *bad_wr;
96 	struct ipoib_rx_buf *rx_req;
97 	struct mbuf *m;
98 	int ret;
99 	int i;
100 
101 	rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id];
102 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
103 		priv->cm.rx_sge[i].addr = rx_req->mapping[i];
104 		priv->cm.rx_sge[i].length = m->m_len;
105 	}
106 
107 	priv->cm.rx_wr.num_sge = i;
108 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
109 
110 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
111 	if (unlikely(ret)) {
112 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
113 		ipoib_dma_unmap_rx(priv, rx_req);
114 		m_freem(priv->cm.srq_ring[id].mb);
115 		priv->cm.srq_ring[id].mb = NULL;
116 	}
117 
118 	return ret;
119 }
120 
121 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv,
122 					struct ipoib_cm_rx *rx,
123 					struct ib_recv_wr *wr,
124 					struct ib_sge *sge, int id)
125 {
126 	struct ipoib_rx_buf *rx_req;
127 	const struct ib_recv_wr *bad_wr;
128 	struct mbuf *m;
129 	int ret;
130 	int i;
131 
132 	rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id];
133 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
134 		sge[i].addr = rx_req->mapping[i];
135 		sge[i].length = m->m_len;
136 	}
137 
138 	wr->num_sge = i;
139 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
140 
141 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
142 	if (unlikely(ret)) {
143 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
144 		ipoib_dma_unmap_rx(priv, rx_req);
145 		m_freem(rx->rx_ring[id].mb);
146 		rx->rx_ring[id].mb = NULL;
147 	}
148 
149 	return ret;
150 }
151 
152 static struct mbuf *
153 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
154 {
155 	return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req,
156 	    sizeof(struct ipoib_pseudoheader), priv->cm.max_cm_mtu, IPOIB_CM_RX_SG);
157 }
158 
159 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv,
160 				  struct ipoib_cm_rx_buf *rx_ring)
161 {
162 	int i;
163 
164 	for (i = 0; i < ipoib_recvq_size; ++i)
165 		if (rx_ring[i].mb) {
166 			ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]);
167 			m_freem(rx_ring[i].mb);
168 		}
169 
170 	kfree(rx_ring);
171 }
172 
173 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
174 {
175 	const struct ib_send_wr *bad_wr;
176 	struct ipoib_cm_rx *p;
177 
178 	/* We only reserved 1 extra slot in CQ for drain WRs, so
179 	 * make sure we have at most 1 outstanding WR. */
180 	if (list_empty(&priv->cm.rx_flush_list) ||
181 	    !list_empty(&priv->cm.rx_drain_list))
182 		return;
183 
184 	/*
185 	 * QPs on flush list are error state.  This way, a "flush
186 	 * error" WC will be immediately generated for each WR we post.
187 	 */
188 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
189 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
190 		ipoib_warn(priv, "failed to post drain wr\n");
191 
192 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
193 }
194 
195 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
196 {
197 	struct ipoib_cm_rx *p = ctx;
198 	struct ipoib_dev_priv *priv = p->priv;
199 	unsigned long flags;
200 
201 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
202 		return;
203 
204 	spin_lock_irqsave(&priv->lock, flags);
205 	list_move(&p->list, &priv->cm.rx_flush_list);
206 	p->state = IPOIB_CM_RX_FLUSH;
207 	ipoib_cm_start_rx_drain(priv);
208 	spin_unlock_irqrestore(&priv->lock, flags);
209 }
210 
211 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv,
212 					   struct ipoib_cm_rx *p)
213 {
214 	struct ib_qp_init_attr attr = {
215 		.event_handler = ipoib_cm_rx_event_handler,
216 		.send_cq = priv->recv_cq, /* For drain WR */
217 		.recv_cq = priv->recv_cq,
218 		.srq = priv->cm.srq,
219 		.cap.max_send_wr = 1, /* For drain WR */
220 		.cap.max_send_sge = 1,
221 		.sq_sig_type = IB_SIGNAL_ALL_WR,
222 		.qp_type = IB_QPT_RC,
223 		.qp_context = p,
224 	};
225 
226 	if (!ipoib_cm_has_srq(priv)) {
227 		attr.cap.max_recv_wr  = ipoib_recvq_size;
228 		attr.cap.max_recv_sge = priv->cm.num_frags;
229 	}
230 
231 	return ib_create_qp(priv->pd, &attr);
232 }
233 
234 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv,
235 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
236 				 unsigned psn)
237 {
238 	struct ib_qp_attr qp_attr;
239 	int qp_attr_mask, ret;
240 
241 	qp_attr.qp_state = IB_QPS_INIT;
242 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
243 	if (ret) {
244 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
245 		return ret;
246 	}
247 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
248 	if (ret) {
249 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
250 		return ret;
251 	}
252 	qp_attr.qp_state = IB_QPS_RTR;
253 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
254 	if (ret) {
255 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
256 		return ret;
257 	}
258 	qp_attr.rq_psn = psn;
259 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
260 	if (ret) {
261 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
262 		return ret;
263 	}
264 
265 	/*
266 	 * Current Mellanox HCA firmware won't generate completions
267 	 * with error for drain WRs unless the QP has been moved to
268 	 * RTS first. This work-around leaves a window where a QP has
269 	 * moved to error asynchronously, but this will eventually get
270 	 * fixed in firmware, so let's not error out if modify QP
271 	 * fails.
272 	 */
273 	qp_attr.qp_state = IB_QPS_RTS;
274 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
275 	if (ret) {
276 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
277 		return 0;
278 	}
279 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
280 	if (ret) {
281 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
282 		return 0;
283 	}
284 
285 	return 0;
286 }
287 
288 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv,
289 				struct ib_recv_wr *wr,
290 				struct ib_sge *sge)
291 {
292 	int i;
293 
294 	for (i = 0; i < IPOIB_CM_RX_SG; i++)
295 		sge[i].lkey = priv->pd->local_dma_lkey;
296 
297 	wr->next    = NULL;
298 	wr->sg_list = sge;
299 	wr->num_sge = 1;
300 }
301 
302 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv,
303     struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx)
304 {
305 	struct {
306 		struct ib_recv_wr wr;
307 		struct ib_sge sge[IPOIB_CM_RX_SG];
308 	} *t;
309 	int ret;
310 	int i;
311 
312 	rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL);
313 	if (!rx->rx_ring) {
314 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
315 		       priv->ca->name, ipoib_recvq_size);
316 		return -ENOMEM;
317 	}
318 
319 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
320 
321 	t = kmalloc(sizeof *t, GFP_KERNEL);
322 	if (!t) {
323 		ret = -ENOMEM;
324 		goto err_free;
325 	}
326 
327 	ipoib_cm_init_rx_wr(priv, &t->wr, t->sge);
328 
329 	spin_lock_irq(&priv->lock);
330 
331 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
332 		spin_unlock_irq(&priv->lock);
333 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
334 		ret = -EINVAL;
335 		goto err_free;
336 	} else
337 		++priv->cm.nonsrq_conn_qp;
338 
339 	spin_unlock_irq(&priv->lock);
340 
341 	for (i = 0; i < ipoib_recvq_size; ++i) {
342 		if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) {
343 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
344 				ret = -ENOMEM;
345 				goto err_count;
346 		}
347 		ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i);
348 		if (ret) {
349 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
350 				   "failed for buf %d\n", i);
351 			ret = -EIO;
352 			goto err_count;
353 		}
354 	}
355 
356 	rx->recv_count = ipoib_recvq_size;
357 
358 	kfree(t);
359 
360 	return 0;
361 
362 err_count:
363 	spin_lock_irq(&priv->lock);
364 	--priv->cm.nonsrq_conn_qp;
365 	spin_unlock_irq(&priv->lock);
366 
367 err_free:
368 	kfree(t);
369 	ipoib_cm_free_rx_ring(priv, rx->rx_ring);
370 
371 	return ret;
372 }
373 
374 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id,
375 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
376 			     unsigned psn)
377 {
378 	struct ipoib_cm_data data = {};
379 	struct ib_cm_rep_param rep = {};
380 
381 	data.qpn = cpu_to_be32(priv->qp->qp_num);
382 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
383 
384 	rep.private_data = &data;
385 	rep.private_data_len = sizeof data;
386 	rep.flow_control = 0;
387 	rep.rnr_retry_count = req->rnr_retry_count;
388 	rep.srq = ipoib_cm_has_srq(priv);
389 	rep.qp_num = qp->qp_num;
390 	rep.starting_psn = psn;
391 	return ib_send_cm_rep(cm_id, &rep);
392 }
393 
394 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
395 {
396 	struct ipoib_dev_priv *priv = cm_id->context;
397 	struct ipoib_cm_rx *p;
398 	unsigned psn;
399 	int ret;
400 
401 	ipoib_dbg(priv, "REQ arrived\n");
402 	p = kzalloc(sizeof *p, GFP_KERNEL);
403 	if (!p)
404 		return -ENOMEM;
405 	p->priv = priv;
406 	p->id = cm_id;
407 	cm_id->context = p;
408 	p->state = IPOIB_CM_RX_LIVE;
409 	p->jiffies = jiffies;
410 	INIT_LIST_HEAD(&p->list);
411 
412 	p->qp = ipoib_cm_create_rx_qp(priv, p);
413 	if (IS_ERR(p->qp)) {
414 		ret = PTR_ERR(p->qp);
415 		goto err_qp;
416 	}
417 
418 	psn = random() & 0xffffff;
419 	ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn);
420 	if (ret)
421 		goto err_modify;
422 
423 	if (!ipoib_cm_has_srq(priv)) {
424 		ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p);
425 		if (ret)
426 			goto err_modify;
427 	}
428 
429 	spin_lock_irq(&priv->lock);
430 	queue_delayed_work(ipoib_workqueue,
431 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
432 	/* Add this entry to passive ids list head, but do not re-add it
433 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
434 	p->jiffies = jiffies;
435 	if (p->state == IPOIB_CM_RX_LIVE)
436 		list_move(&p->list, &priv->cm.passive_ids);
437 	spin_unlock_irq(&priv->lock);
438 
439 	ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn);
440 	if (ret) {
441 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
442 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
443 			ipoib_warn(priv, "unable to move qp to error state\n");
444 	}
445 	return 0;
446 
447 err_modify:
448 	ib_destroy_qp(p->qp);
449 err_qp:
450 	kfree(p);
451 	return ret;
452 }
453 
454 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
455 			       struct ib_cm_event *event)
456 {
457 	struct ipoib_cm_rx *p;
458 	struct ipoib_dev_priv *priv;
459 
460 	switch (event->event) {
461 	case IB_CM_REQ_RECEIVED:
462 		return ipoib_cm_req_handler(cm_id, event);
463 	case IB_CM_DREQ_RECEIVED:
464 		p = cm_id->context;
465 		ib_send_cm_drep(cm_id, NULL, 0);
466 		/* Fall through */
467 	case IB_CM_REJ_RECEIVED:
468 		p = cm_id->context;
469 		priv = p->priv;
470 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
471 			ipoib_warn(priv, "unable to move qp to error state\n");
472 		/* Fall through */
473 	default:
474 		return 0;
475 	}
476 }
477 
478 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
479 {
480 	struct ipoib_cm_rx_buf saverx;
481 	struct ipoib_cm_rx_buf *rx_ring;
482 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
483 	if_t dev = priv->dev;
484 	struct mbuf *mb, *newmb;
485 	struct ipoib_cm_rx *p;
486 	int has_srq;
487 
488 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
489 		       wr_id, wc->status);
490 
491 	if (unlikely(wr_id >= ipoib_recvq_size)) {
492 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
493 			spin_lock(&priv->lock);
494 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
495 			ipoib_cm_start_rx_drain(priv);
496 			if (priv->cm.id != NULL)
497 				queue_work(ipoib_workqueue,
498 				    &priv->cm.rx_reap_task);
499 			spin_unlock(&priv->lock);
500 		} else
501 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
502 				   wr_id, ipoib_recvq_size);
503 		goto done;
504 	}
505 
506 	p = wc->qp->qp_context;
507 
508 	has_srq = ipoib_cm_has_srq(priv);
509 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
510 
511 	mb = rx_ring[wr_id].mb;
512 
513 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
514 		ipoib_dbg(priv, "cm recv error "
515 			   "(status=%d, wrid=%d vend_err %x)\n",
516 			   wc->status, wr_id, wc->vendor_err);
517 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
518 		if (has_srq)
519 			goto repost;
520 		else {
521 			if (!--p->recv_count) {
522 				spin_lock(&priv->lock);
523 				list_move(&p->list, &priv->cm.rx_reap_list);
524 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
525 				spin_unlock(&priv->lock);
526 			}
527 			goto done;
528 		}
529 	}
530 
531 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
532 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
533 			p->jiffies = jiffies;
534 			/* Move this entry to list head, but do not re-add it
535 			 * if it has been moved out of list. */
536 			if (p->state == IPOIB_CM_RX_LIVE)
537 				list_move(&p->list, &priv->cm.passive_ids);
538 		}
539 	}
540 
541 	memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx));
542 	newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]);
543 	if (unlikely(!newmb)) {
544 		/*
545 		 * If we can't allocate a new RX buffer, dump
546 		 * this packet and reuse the old buffer.
547 		 */
548 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
549 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
550 		memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx));
551 		goto repost;
552 	}
553 
554 	ipoib_cm_dma_unmap_rx(priv, &saverx);
555 
556 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
557 		       wc->byte_len, wc->slid);
558 
559 	ipoib_dma_mb(priv, mb, wc->byte_len);
560 
561 	mb->m_pkthdr.rcvif = dev;
562 
563 	M_PREPEND(mb, sizeof(struct ipoib_pseudoheader), M_NOWAIT);
564 	if (likely(mb != NULL)) {
565 		struct ipoib_header *ibh;
566 
567 		if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
568 		if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
569 
570 		/* fixup destination infiniband address */
571 		ibh = mtod(mb, struct ipoib_header *);
572 		memset(ibh->hwaddr, 0, 4);
573 		memcpy(ibh->hwaddr + 4, priv->local_gid.raw, sizeof(union ib_gid));
574 
575 		if_input(dev, mb);
576 	} else {
577 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
578 	}
579 repost:
580 	if (has_srq) {
581 		if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
582 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
583 				   "for buf %d\n", wr_id);
584 	} else {
585 		if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
586 							  &priv->cm.rx_wr,
587 							  priv->cm.rx_sge,
588 							  wr_id))) {
589 			--p->recv_count;
590 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
591 				   "for buf %d\n", wr_id);
592 		}
593 	}
594 done:
595 	return;
596 }
597 
598 static inline int post_send(struct ipoib_dev_priv *priv,
599 			    struct ipoib_cm_tx *tx,
600 			    struct ipoib_cm_tx_buf *tx_req,
601 			    unsigned int wr_id)
602 {
603 	const struct ib_send_wr *bad_wr;
604 	struct mbuf *mb = tx_req->mb;
605 	u64 *mapping = tx_req->mapping;
606 	struct mbuf *m;
607 	int i;
608 
609 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
610 		priv->tx_sge[i].addr = mapping[i];
611 		priv->tx_sge[i].length = m->m_len;
612 	}
613 	priv->tx_wr.wr.num_sge = i;
614 	priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
615 	priv->tx_wr.wr.opcode = IB_WR_SEND;
616 
617 	return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
618 }
619 
620 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
621 {
622 	struct ipoib_cm_tx_buf *tx_req;
623 	if_t dev = priv->dev;
624 
625 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) {
626 		while (ipoib_poll_tx(priv, false))
627 			;	/* nothing */
628 	}
629 
630 	m_adj(mb, sizeof(struct ipoib_pseudoheader));
631 	if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
632 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
633 			   mb->m_pkthdr.len, tx->mtu);
634 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
635 		ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
636 		return;
637 	}
638 
639 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
640 		       tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
641 
642 
643 	/*
644 	 * We put the mb into the tx_ring _before_ we call post_send()
645 	 * because it's entirely possible that the completion handler will
646 	 * run before we execute anything after the post_send().  That
647 	 * means we have to make sure everything is properly recorded and
648 	 * our state is consistent before we call post_send().
649 	 */
650 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
651 	tx_req->mb = mb;
652 	if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
653 	    priv->cm.num_frags))) {
654 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
655 		if (tx_req->mb)
656 			m_freem(tx_req->mb);
657 		return;
658 	}
659 
660 	if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
661 		ipoib_warn(priv, "post_send failed\n");
662 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
663 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
664 		m_freem(mb);
665 	} else {
666 		++tx->tx_head;
667 
668 		if (++priv->tx_outstanding == ipoib_sendq_size) {
669 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
670 				  tx->qp->qp_num);
671 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
672 				ipoib_warn(priv, "request notify on send CQ failed\n");
673 			if_setdrvflagbits(dev, IFF_DRV_OACTIVE, 0);
674 		}
675 	}
676 
677 }
678 
679 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
680 {
681 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
682 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
683 	if_t dev = priv->dev;
684 	struct ipoib_cm_tx_buf *tx_req;
685 
686 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
687 		       wr_id, wc->status);
688 
689 	if (unlikely(wr_id >= ipoib_sendq_size)) {
690 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
691 			   wr_id, ipoib_sendq_size);
692 		return;
693 	}
694 
695 	tx_req = &tx->tx_ring[wr_id];
696 
697 	ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
698 
699 	/* FIXME: is this right? Shouldn't we only increment on success? */
700 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
701 
702 	m_freem(tx_req->mb);
703 
704 	++tx->tx_tail;
705 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
706 	    (if_getdrvflags(dev) & IFF_DRV_OACTIVE) != 0 &&
707 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
708 		if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
709 
710 	if (wc->status != IB_WC_SUCCESS &&
711 	    wc->status != IB_WC_WR_FLUSH_ERR) {
712 		struct ipoib_path *path;
713 
714 		ipoib_dbg(priv, "failed cm send event "
715 			   "(status=%d, wrid=%d vend_err %x)\n",
716 			   wc->status, wr_id, wc->vendor_err);
717 
718 		path = tx->path;
719 
720 		if (path) {
721 			path->cm = NULL;
722 			rb_erase(&path->rb_node, &priv->path_tree);
723 			list_del(&path->list);
724 		}
725 
726 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
727 			list_move(&tx->list, &priv->cm.reap_list);
728 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
729 		}
730 
731 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
732 	}
733 
734 }
735 
736 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
737 {
738 	int ret;
739 
740 	if (!IPOIB_CM_SUPPORTED(if_getlladdr(priv->dev)))
741 		return 0;
742 
743 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
744 	if (IS_ERR(priv->cm.id)) {
745 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
746 		ret = PTR_ERR(priv->cm.id);
747 		goto err_cm;
748 	}
749 
750 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 0);
751 	if (ret) {
752 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
753 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
754 		goto err_listen;
755 	}
756 
757 	return 0;
758 
759 err_listen:
760 	ib_destroy_cm_id(priv->cm.id);
761 err_cm:
762 	priv->cm.id = NULL;
763 	return ret;
764 }
765 
766 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
767 {
768 	struct ipoib_cm_rx *rx, *n;
769 	LIST_HEAD(list);
770 
771 	spin_lock_irq(&priv->lock);
772 	list_splice_init(&priv->cm.rx_reap_list, &list);
773 	spin_unlock_irq(&priv->lock);
774 
775 	list_for_each_entry_safe(rx, n, &list, list) {
776 		ib_destroy_cm_id(rx->id);
777 		ib_destroy_qp(rx->qp);
778 		if (!ipoib_cm_has_srq(priv)) {
779 			ipoib_cm_free_rx_ring(priv, rx->rx_ring);
780 			spin_lock_irq(&priv->lock);
781 			--priv->cm.nonsrq_conn_qp;
782 			spin_unlock_irq(&priv->lock);
783 		}
784 		kfree(rx);
785 	}
786 }
787 
788 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
789 {
790 	struct ipoib_cm_rx *p;
791 	unsigned long begin;
792 	int ret;
793 
794 	if (!IPOIB_CM_SUPPORTED(if_getlladdr(priv->dev)) || !priv->cm.id)
795 		return;
796 
797 	ib_destroy_cm_id(priv->cm.id);
798 	priv->cm.id = NULL;
799 
800 	cancel_work_sync(&priv->cm.rx_reap_task);
801 
802 	spin_lock_irq(&priv->lock);
803 	while (!list_empty(&priv->cm.passive_ids)) {
804 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
805 		list_move(&p->list, &priv->cm.rx_error_list);
806 		p->state = IPOIB_CM_RX_ERROR;
807 		spin_unlock_irq(&priv->lock);
808 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
809 		if (ret)
810 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
811 		spin_lock_irq(&priv->lock);
812 	}
813 
814 	/* Wait for all RX to be drained */
815 	begin = jiffies;
816 
817 	while (!list_empty(&priv->cm.rx_error_list) ||
818 	       !list_empty(&priv->cm.rx_flush_list) ||
819 	       !list_empty(&priv->cm.rx_drain_list)) {
820 		if (time_after(jiffies, begin + 5 * HZ)) {
821 			ipoib_warn(priv, "RX drain timing out\n");
822 
823 			/*
824 			 * assume the HW is wedged and just free up everything.
825 			 */
826 			list_splice_init(&priv->cm.rx_flush_list,
827 					 &priv->cm.rx_reap_list);
828 			list_splice_init(&priv->cm.rx_error_list,
829 					 &priv->cm.rx_reap_list);
830 			list_splice_init(&priv->cm.rx_drain_list,
831 					 &priv->cm.rx_reap_list);
832 			break;
833 		}
834 		spin_unlock_irq(&priv->lock);
835 		msleep(1);
836 		ipoib_drain_cq(priv);
837 		spin_lock_irq(&priv->lock);
838 	}
839 
840 	spin_unlock_irq(&priv->lock);
841 
842 	ipoib_cm_free_rx_reap_list(priv);
843 
844 	cancel_delayed_work_sync(&priv->cm.stale_task);
845 }
846 
847 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
848 {
849 	struct ipoib_cm_tx *p = cm_id->context;
850 	struct ipoib_dev_priv *priv = p->priv;
851 	struct ipoib_cm_data *data = event->private_data;
852 	struct epoch_tracker et;
853 	struct ifqueue mbqueue;
854 	struct ib_qp_attr qp_attr;
855 	int qp_attr_mask, ret;
856 	struct mbuf *mb;
857 
858 	ipoib_dbg(priv, "cm rep handler\n");
859 	p->mtu = be32_to_cpu(data->mtu);
860 
861 	if (p->mtu <= IPOIB_ENCAP_LEN) {
862 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
863 			   p->mtu, IPOIB_ENCAP_LEN);
864 		return -EINVAL;
865 	}
866 
867 	qp_attr.qp_state = IB_QPS_RTR;
868 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
869 	if (ret) {
870 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
871 		return ret;
872 	}
873 
874 	qp_attr.rq_psn = 0 /* FIXME */;
875 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
876 	if (ret) {
877 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
878 		return ret;
879 	}
880 
881 	qp_attr.qp_state = IB_QPS_RTS;
882 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
883 	if (ret) {
884 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
885 		return ret;
886 	}
887 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
888 	if (ret) {
889 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
890 		return ret;
891 	}
892 
893 	bzero(&mbqueue, sizeof(mbqueue));
894 
895 	spin_lock_irq(&priv->lock);
896 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
897 	if (p->path)
898 		for (;;) {
899 			_IF_DEQUEUE(&p->path->queue, mb);
900 			if (mb == NULL)
901 				break;
902 			_IF_ENQUEUE(&mbqueue, mb);
903 		}
904 	spin_unlock_irq(&priv->lock);
905 
906 	NET_EPOCH_ENTER(et);
907 	for (;;) {
908 		if_t dev = p->priv->dev;
909 		_IF_DEQUEUE(&mbqueue, mb);
910 		if (mb == NULL)
911 			break;
912 		mb->m_pkthdr.rcvif = dev;
913 		if (if_transmit(dev, mb))
914 			ipoib_warn(priv, "dev_queue_xmit failed "
915 				   "to requeue packet\n");
916 	}
917 	NET_EPOCH_EXIT(et);
918 
919 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
920 	if (ret) {
921 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
922 		return ret;
923 	}
924 	return 0;
925 }
926 
927 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
928     struct ipoib_cm_tx *tx)
929 {
930 	struct ib_qp_init_attr attr = {
931 		.send_cq		= priv->send_cq,
932 		.recv_cq		= priv->recv_cq,
933 		.srq			= priv->cm.srq,
934 		.cap.max_send_wr	= ipoib_sendq_size,
935 		.cap.max_send_sge	= priv->cm.num_frags,
936 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
937 		.qp_type		= IB_QPT_RC,
938 		.qp_context		= tx
939 	};
940 
941 	return ib_create_qp(priv->pd, &attr);
942 }
943 
944 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
945 			     struct ib_cm_id *id, struct ib_qp *qp,
946 			     u32 qpn,
947 			     struct ib_sa_path_rec *pathrec)
948 {
949 	struct ipoib_cm_data data = {};
950 	struct ib_cm_req_param req = {};
951 
952 	ipoib_dbg(priv, "cm send req\n");
953 
954 	data.qpn = cpu_to_be32(priv->qp->qp_num);
955 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
956 
957 	req.primary_path		= pathrec;
958 	req.alternate_path		= NULL;
959 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
960 	req.qp_num			= qp->qp_num;
961 	req.qp_type			= qp->qp_type;
962 	req.private_data		= &data;
963 	req.private_data_len		= sizeof data;
964 	req.flow_control		= 0;
965 
966 	req.starting_psn		= 0; /* FIXME */
967 
968 	/*
969 	 * Pick some arbitrary defaults here; we could make these
970 	 * module parameters if anyone cared about setting them.
971 	 */
972 	req.responder_resources		= 4;
973 	req.remote_cm_response_timeout	= 20;
974 	req.local_cm_response_timeout	= 20;
975 	req.retry_count			= 0; /* RFC draft warns against retries */
976 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
977 	req.max_cm_retries		= 15;
978 	req.srq				= ipoib_cm_has_srq(priv);
979 	return ib_send_cm_req(id, &req);
980 }
981 
982 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
983 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
984 {
985 	struct ib_qp_attr qp_attr;
986 	int qp_attr_mask, ret;
987 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
988 	if (ret) {
989 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
990 		return ret;
991 	}
992 
993 	qp_attr.qp_state = IB_QPS_INIT;
994 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
995 	qp_attr.port_num = priv->port;
996 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
997 
998 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
999 	if (ret) {
1000 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1001 		return ret;
1002 	}
1003 	return 0;
1004 }
1005 
1006 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1007 			    struct ib_sa_path_rec *pathrec)
1008 {
1009 	struct ipoib_dev_priv *priv = p->priv;
1010 	int ret;
1011 
1012 	p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
1013 	if (!p->tx_ring) {
1014 		ipoib_warn(priv, "failed to allocate tx ring\n");
1015 		ret = -ENOMEM;
1016 		goto err_tx;
1017 	}
1018 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1019 
1020 	p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1021 	if (IS_ERR(p->qp)) {
1022 		ret = PTR_ERR(p->qp);
1023 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1024 		goto err_qp;
1025 	}
1026 
1027 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1028 	if (IS_ERR(p->id)) {
1029 		ret = PTR_ERR(p->id);
1030 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1031 		goto err_id;
1032 	}
1033 
1034 	ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1035 	if (ret) {
1036 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1037 		goto err_modify;
1038 	}
1039 
1040 	ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1041 	if (ret) {
1042 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1043 		goto err_send_cm;
1044 	}
1045 
1046 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1047 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1048 
1049 	return 0;
1050 
1051 err_send_cm:
1052 err_modify:
1053 	ib_destroy_cm_id(p->id);
1054 err_id:
1055 	p->id = NULL;
1056 	ib_destroy_qp(p->qp);
1057 err_qp:
1058 	p->qp = NULL;
1059 	kfree(p->tx_ring);
1060 err_tx:
1061 	return ret;
1062 }
1063 
1064 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1065 {
1066 	struct ipoib_dev_priv *priv = p->priv;
1067 	if_t dev = priv->dev;
1068 	struct ipoib_cm_tx_buf *tx_req;
1069 	unsigned long begin;
1070 
1071 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1072 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1073 
1074 	if (p->path)
1075 		ipoib_path_free(priv, p->path);
1076 
1077 	if (p->id)
1078 		ib_destroy_cm_id(p->id);
1079 
1080 	if (p->tx_ring) {
1081 		/* Wait for all sends to complete */
1082 		begin = jiffies;
1083 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1084 			if (time_after(jiffies, begin + 5 * HZ)) {
1085 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1086 					   p->tx_head - p->tx_tail);
1087 				goto timeout;
1088 			}
1089 
1090 			msleep(1);
1091 		}
1092 	}
1093 
1094 timeout:
1095 
1096 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1097 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1098 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1099 		m_freem(tx_req->mb);
1100 		++p->tx_tail;
1101 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1102 		    (if_getdrvflags(dev) & IFF_DRV_OACTIVE) != 0 &&
1103 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1104 			if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
1105 	}
1106 
1107 	if (p->qp)
1108 		ib_destroy_qp(p->qp);
1109 
1110 	kfree(p->tx_ring);
1111 	kfree(p);
1112 }
1113 
1114 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1115 			       struct ib_cm_event *event)
1116 {
1117 	struct ipoib_cm_tx *tx = cm_id->context;
1118 	struct ipoib_dev_priv *priv = tx->priv;
1119 	struct ipoib_path *path;
1120 	unsigned long flags;
1121 	int ret;
1122 
1123 	switch (event->event) {
1124 	case IB_CM_DREQ_RECEIVED:
1125 		ipoib_dbg(priv, "DREQ received.\n");
1126 		ib_send_cm_drep(cm_id, NULL, 0);
1127 		break;
1128 	case IB_CM_REP_RECEIVED:
1129 		ipoib_dbg(priv, "REP received.\n");
1130 		ret = ipoib_cm_rep_handler(cm_id, event);
1131 		if (ret)
1132 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1133 				       NULL, 0, NULL, 0);
1134 		break;
1135 	case IB_CM_REQ_ERROR:
1136 	case IB_CM_REJ_RECEIVED:
1137 	case IB_CM_TIMEWAIT_EXIT:
1138 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1139 		spin_lock_irqsave(&priv->lock, flags);
1140 		path = tx->path;
1141 
1142 		if (path) {
1143 			path->cm = NULL;
1144 			tx->path = NULL;
1145 			rb_erase(&path->rb_node, &priv->path_tree);
1146 			list_del(&path->list);
1147 		}
1148 
1149 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1150 			list_move(&tx->list, &priv->cm.reap_list);
1151 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1152 		}
1153 
1154 		spin_unlock_irqrestore(&priv->lock, flags);
1155 		if (path)
1156 			ipoib_path_free(tx->priv, path);
1157 		break;
1158 	default:
1159 		break;
1160 	}
1161 
1162 	return 0;
1163 }
1164 
1165 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1166     struct ipoib_path *path)
1167 {
1168 	struct ipoib_cm_tx *tx;
1169 
1170 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1171 	if (!tx)
1172 		return NULL;
1173 
1174 	ipoib_dbg(priv, "Creating cm tx\n");
1175 	path->cm = tx;
1176 	tx->path = path;
1177 	tx->priv = priv;
1178 	list_add(&tx->list, &priv->cm.start_list);
1179 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1180 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1181 	return tx;
1182 }
1183 
1184 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1185 {
1186 	struct ipoib_dev_priv *priv = tx->priv;
1187 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1188 		spin_lock(&priv->lock);
1189 		list_move(&tx->list, &priv->cm.reap_list);
1190 		spin_unlock(&priv->lock);
1191 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1192 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1193 			  tx->path->pathrec.dgid.raw);
1194 		tx->path = NULL;
1195 	}
1196 }
1197 
1198 static void ipoib_cm_tx_start(struct work_struct *work)
1199 {
1200 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1201 						   cm.start_task);
1202 	struct ipoib_path *path;
1203 	struct ipoib_cm_tx *p;
1204 	unsigned long flags;
1205 	int ret;
1206 
1207 	struct ib_sa_path_rec pathrec;
1208 	u32 qpn;
1209 
1210 	ipoib_dbg(priv, "cm start task\n");
1211 	spin_lock_irqsave(&priv->lock, flags);
1212 
1213 	while (!list_empty(&priv->cm.start_list)) {
1214 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1215 		list_del_init(&p->list);
1216 		path = p->path;
1217 		qpn = IPOIB_QPN(path->hwaddr);
1218 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1219 
1220 		spin_unlock_irqrestore(&priv->lock, flags);
1221 
1222 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1223 
1224 		spin_lock_irqsave(&priv->lock, flags);
1225 
1226 		if (ret) {
1227 			path = p->path;
1228 			if (path) {
1229 				path->cm = NULL;
1230 				rb_erase(&path->rb_node, &priv->path_tree);
1231 				list_del(&path->list);
1232 				ipoib_path_free(priv, path);
1233 			}
1234 			list_del(&p->list);
1235 			kfree(p);
1236 		}
1237 	}
1238 
1239 	spin_unlock_irqrestore(&priv->lock, flags);
1240 }
1241 
1242 static void ipoib_cm_tx_reap(struct work_struct *work)
1243 {
1244 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1245 						   cm.reap_task);
1246 	struct ipoib_cm_tx *p;
1247 	unsigned long flags;
1248 
1249 	spin_lock_irqsave(&priv->lock, flags);
1250 
1251 	while (!list_empty(&priv->cm.reap_list)) {
1252 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1253 		list_del(&p->list);
1254 		spin_unlock_irqrestore(&priv->lock, flags);
1255 		ipoib_cm_tx_destroy(p);
1256 		spin_lock_irqsave(&priv->lock, flags);
1257 	}
1258 
1259 	spin_unlock_irqrestore(&priv->lock, flags);
1260 }
1261 
1262 static void ipoib_cm_mb_reap(struct work_struct *work)
1263 {
1264 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1265 						   cm.mb_task);
1266 	struct mbuf *mb;
1267 	unsigned long flags;
1268 #if defined(INET) || defined(INET6)
1269 	unsigned mtu = priv->mcast_mtu;
1270 #endif
1271 	uint16_t proto;
1272 
1273 	spin_lock_irqsave(&priv->lock, flags);
1274 
1275 	CURVNET_SET_QUIET(if_getvnet(priv->dev));
1276 
1277 	for (;;) {
1278 		IF_DEQUEUE(&priv->cm.mb_queue, mb);
1279 		if (mb == NULL)
1280 			break;
1281 		spin_unlock_irqrestore(&priv->lock, flags);
1282 
1283 		proto = htons(*mtod(mb, uint16_t *));
1284 		m_adj(mb, IPOIB_ENCAP_LEN);
1285 		switch (proto) {
1286 #if defined(INET)
1287 		case ETHERTYPE_IP:
1288 			icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1289 			break;
1290 #endif
1291 #if defined(INET6)
1292 		case ETHERTYPE_IPV6:
1293 			icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1294 			break;
1295 #endif
1296 		default:
1297 			m_freem(mb);
1298 		}
1299 
1300 		spin_lock_irqsave(&priv->lock, flags);
1301 	}
1302 
1303 	CURVNET_RESTORE();
1304 
1305 	spin_unlock_irqrestore(&priv->lock, flags);
1306 }
1307 
1308 void
1309 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1310 {
1311 	int e = priv->cm.mb_queue.ifq_len;
1312 
1313 	IF_ENQUEUE(&priv->cm.mb_queue, mb);
1314 	if (e == 0)
1315 		queue_work(ipoib_workqueue, &priv->cm.mb_task);
1316 }
1317 
1318 static void ipoib_cm_rx_reap(struct work_struct *work)
1319 {
1320 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1321 						cm.rx_reap_task));
1322 }
1323 
1324 static void ipoib_cm_stale_task(struct work_struct *work)
1325 {
1326 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1327 						   cm.stale_task.work);
1328 	struct ipoib_cm_rx *p;
1329 	int ret;
1330 
1331 	spin_lock_irq(&priv->lock);
1332 	while (!list_empty(&priv->cm.passive_ids)) {
1333 		/* List is sorted by LRU, start from tail,
1334 		 * stop when we see a recently used entry */
1335 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1336 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1337 			break;
1338 		list_move(&p->list, &priv->cm.rx_error_list);
1339 		p->state = IPOIB_CM_RX_ERROR;
1340 		spin_unlock_irq(&priv->lock);
1341 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1342 		if (ret)
1343 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1344 		spin_lock_irq(&priv->lock);
1345 	}
1346 
1347 	if (!list_empty(&priv->cm.passive_ids))
1348 		queue_delayed_work(ipoib_workqueue,
1349 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1350 	spin_unlock_irq(&priv->lock);
1351 }
1352 
1353 
1354 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1355 {
1356 	struct ib_srq_init_attr srq_init_attr = {
1357 		.attr = {
1358 			.max_wr  = ipoib_recvq_size,
1359 			.max_sge = max_sge
1360 		}
1361 	};
1362 
1363 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1364 	if (IS_ERR(priv->cm.srq)) {
1365 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1366 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1367 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1368 		priv->cm.srq = NULL;
1369 		return;
1370 	}
1371 
1372 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1373 	if (!priv->cm.srq_ring) {
1374 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1375 		       priv->ca->name, ipoib_recvq_size);
1376 		ib_destroy_srq(priv->cm.srq);
1377 		priv->cm.srq = NULL;
1378 		return;
1379 	}
1380 
1381 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1382 }
1383 
1384 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1385 {
1386 	if_t dev = priv->dev;
1387 	int i;
1388 	int max_srq_sge;
1389 
1390 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1391 	INIT_LIST_HEAD(&priv->cm.reap_list);
1392 	INIT_LIST_HEAD(&priv->cm.start_list);
1393 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1394 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1395 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1396 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1397 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1398 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1399 	INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1400 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1401 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1402 
1403 	bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1404 	mtx_init(&priv->cm.mb_queue.ifq_mtx,
1405 	    if_name(dev), "if send queue", MTX_DEF);
1406 
1407 	max_srq_sge = priv->ca->attrs.max_srq_sge;
1408 
1409 	ipoib_dbg(priv, "max_srq_sge=%d\n", max_srq_sge);
1410 
1411 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, max_srq_sge);
1412 	ipoib_cm_create_srq(priv, max_srq_sge);
1413 	if (ipoib_cm_has_srq(priv)) {
1414 		priv->cm.max_cm_mtu = max_srq_sge * MJUMPAGESIZE;
1415 		priv->cm.num_frags  = max_srq_sge;
1416 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1417 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1418 	} else {
1419 		priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1420 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1421 	}
1422 
1423 	ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1424 
1425 	if (ipoib_cm_has_srq(priv)) {
1426 		for (i = 0; i < ipoib_recvq_size; ++i) {
1427 			if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1428 				ipoib_warn(priv, "failed to allocate "
1429 					   "receive buffer %d\n", i);
1430 				ipoib_cm_dev_cleanup(priv);
1431 				return -ENOMEM;
1432 			}
1433 
1434 			if (ipoib_cm_post_receive_srq(priv, i)) {
1435 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1436 					   "failed for buf %d\n", i);
1437 				ipoib_cm_dev_cleanup(priv);
1438 				return -EIO;
1439 			}
1440 		}
1441 	}
1442 
1443 	if_getlladdr(priv->dev)[0] = IPOIB_FLAGS_RC;
1444 	return 0;
1445 }
1446 
1447 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1448 {
1449 	int ret;
1450 
1451 	if (!priv->cm.srq)
1452 		return;
1453 
1454 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1455 
1456 	ret = ib_destroy_srq(priv->cm.srq);
1457 	if (ret)
1458 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1459 
1460 	priv->cm.srq = NULL;
1461 	if (!priv->cm.srq_ring)
1462 		return;
1463 
1464 	ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1465 	priv->cm.srq_ring = NULL;
1466 
1467 	mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1468 }
1469 
1470 #endif /* CONFIG_INFINIBAND_IPOIB_CM */
1471