xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision 193d9e768ba63fcfb187cfd17f461f7d41345048)
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 "ipoib.h"
34 
35 #ifdef CONFIG_INFINIBAND_IPOIB_CM
36 
37 #include <netinet/ip.h>
38 #include <netinet/ip_icmp.h>
39 #include <netinet/icmp6.h>
40 
41 #include <rdma/ib_cm.h>
42 #include <rdma/ib_cache.h>
43 #include <linux/delay.h>
44 
45 int ipoib_max_conn_qp = 128;
46 
47 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
48 MODULE_PARM_DESC(max_nonsrq_conn_qp,
49 		 "Max number of connected-mode QPs per interface "
50 		 "(applied only if shared receive queue is not available)");
51 
52 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
53 static int data_debug_level;
54 
55 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
56 MODULE_PARM_DESC(cm_data_debug_level,
57 		 "Enable data path debug tracing for connected mode if > 0");
58 #endif
59 
60 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
61 
62 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
63 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
64 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
65 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
66 
67 static struct ib_qp_attr ipoib_cm_err_attr = {
68 	.qp_state = IB_QPS_ERR
69 };
70 
71 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
72 
73 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
74 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
75 	.opcode = IB_WR_SEND,
76 };
77 
78 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
79 			       struct ib_cm_event *event);
80 
81 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
82 {
83 
84 	ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req);
85 
86 }
87 
88 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id)
89 {
90 	struct ib_recv_wr *bad_wr;
91 	struct ipoib_rx_buf *rx_req;
92 	struct mbuf *m;
93 	int ret;
94 	int i;
95 
96 	rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id];
97 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
98 		priv->cm.rx_sge[i].addr = rx_req->mapping[i];
99 		priv->cm.rx_sge[i].length = m->m_len;
100 	}
101 
102 	priv->cm.rx_wr.num_sge = i;
103 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
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_dma_unmap_rx(priv, rx_req);
109 		m_freem(priv->cm.srq_ring[id].mb);
110 		priv->cm.srq_ring[id].mb = NULL;
111 	}
112 
113 	return ret;
114 }
115 
116 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv,
117 					struct ipoib_cm_rx *rx,
118 					struct ib_recv_wr *wr,
119 					struct ib_sge *sge, int id)
120 {
121 	struct ipoib_rx_buf *rx_req;
122 	struct ib_recv_wr *bad_wr;
123 	struct mbuf *m;
124 	int ret;
125 	int i;
126 
127 	rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id];
128 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
129 		sge[i].addr = rx_req->mapping[i];
130 		sge[i].length = m->m_len;
131 	}
132 
133 	wr->num_sge = i;
134 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
135 
136 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
137 	if (unlikely(ret)) {
138 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
139 		ipoib_dma_unmap_rx(priv, rx_req);
140 		m_freem(rx->rx_ring[id].mb);
141 		rx->rx_ring[id].mb = NULL;
142 	}
143 
144 	return ret;
145 }
146 
147 static struct mbuf *
148 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
149 {
150 	return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req,
151 	    priv->cm.max_cm_mtu);
152 }
153 
154 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv,
155 				  struct ipoib_cm_rx_buf *rx_ring)
156 {
157 	int i;
158 
159 	for (i = 0; i < ipoib_recvq_size; ++i)
160 		if (rx_ring[i].mb) {
161 			ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]);
162 			m_freem(rx_ring[i].mb);
163 		}
164 
165 	kfree(rx_ring);
166 }
167 
168 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
169 {
170 	struct ib_send_wr *bad_wr;
171 	struct ipoib_cm_rx *p;
172 
173 	/* We only reserved 1 extra slot in CQ for drain WRs, so
174 	 * make sure we have at most 1 outstanding WR. */
175 	if (list_empty(&priv->cm.rx_flush_list) ||
176 	    !list_empty(&priv->cm.rx_drain_list))
177 		return;
178 
179 	/*
180 	 * QPs on flush list are error state.  This way, a "flush
181 	 * error" WC will be immediately generated for each WR we post.
182 	 */
183 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
184 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
185 		ipoib_warn(priv, "failed to post drain wr\n");
186 
187 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
188 }
189 
190 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
191 {
192 	struct ipoib_cm_rx *p = ctx;
193 	struct ipoib_dev_priv *priv = p->priv;
194 	unsigned long flags;
195 
196 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
197 		return;
198 
199 	spin_lock_irqsave(&priv->lock, flags);
200 	list_move(&p->list, &priv->cm.rx_flush_list);
201 	p->state = IPOIB_CM_RX_FLUSH;
202 	ipoib_cm_start_rx_drain(priv);
203 	spin_unlock_irqrestore(&priv->lock, flags);
204 }
205 
206 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv,
207 					   struct ipoib_cm_rx *p)
208 {
209 	struct ib_qp_init_attr attr = {
210 		.event_handler = ipoib_cm_rx_event_handler,
211 		.send_cq = priv->recv_cq, /* For drain WR */
212 		.recv_cq = priv->recv_cq,
213 		.srq = priv->cm.srq,
214 		.cap.max_send_wr = 1, /* For drain WR */
215 		.cap.max_send_sge = 1,
216 		.sq_sig_type = IB_SIGNAL_ALL_WR,
217 		.qp_type = IB_QPT_RC,
218 		.qp_context = p,
219 	};
220 
221 	if (!ipoib_cm_has_srq(priv)) {
222 		attr.cap.max_recv_wr  = ipoib_recvq_size;
223 		attr.cap.max_recv_sge = priv->cm.num_frags;
224 	}
225 
226 	return ib_create_qp(priv->pd, &attr);
227 }
228 
229 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv,
230 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
231 				 unsigned psn)
232 {
233 	struct ib_qp_attr qp_attr;
234 	int qp_attr_mask, ret;
235 
236 	qp_attr.qp_state = IB_QPS_INIT;
237 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
238 	if (ret) {
239 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
240 		return ret;
241 	}
242 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
243 	if (ret) {
244 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
245 		return ret;
246 	}
247 	qp_attr.qp_state = IB_QPS_RTR;
248 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
249 	if (ret) {
250 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
251 		return ret;
252 	}
253 	qp_attr.rq_psn = psn;
254 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
255 	if (ret) {
256 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
257 		return ret;
258 	}
259 
260 	/*
261 	 * Current Mellanox HCA firmware won't generate completions
262 	 * with error for drain WRs unless the QP has been moved to
263 	 * RTS first. This work-around leaves a window where a QP has
264 	 * moved to error asynchronously, but this will eventually get
265 	 * fixed in firmware, so let's not error out if modify QP
266 	 * fails.
267 	 */
268 	qp_attr.qp_state = IB_QPS_RTS;
269 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
270 	if (ret) {
271 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
272 		return 0;
273 	}
274 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
275 	if (ret) {
276 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
277 		return 0;
278 	}
279 
280 	return 0;
281 }
282 
283 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv,
284 				struct ib_recv_wr *wr,
285 				struct ib_sge *sge)
286 {
287 	int i;
288 
289 	for (i = 0; i < IPOIB_CM_RX_SG; i++)
290 		sge[i].lkey = priv->mr->lkey;
291 
292 	wr->next    = NULL;
293 	wr->sg_list = sge;
294 	wr->num_sge = 1;
295 }
296 
297 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv,
298     struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx)
299 {
300 	struct {
301 		struct ib_recv_wr wr;
302 		struct ib_sge sge[IPOIB_CM_RX_SG];
303 	} *t;
304 	int ret;
305 	int i;
306 
307 	rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL);
308 	if (!rx->rx_ring) {
309 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
310 		       priv->ca->name, ipoib_recvq_size);
311 		return -ENOMEM;
312 	}
313 
314 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
315 
316 	t = kmalloc(sizeof *t, GFP_KERNEL);
317 	if (!t) {
318 		ret = -ENOMEM;
319 		goto err_free;
320 	}
321 
322 	ipoib_cm_init_rx_wr(priv, &t->wr, t->sge);
323 
324 	spin_lock_irq(&priv->lock);
325 
326 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
327 		spin_unlock_irq(&priv->lock);
328 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
329 		ret = -EINVAL;
330 		goto err_free;
331 	} else
332 		++priv->cm.nonsrq_conn_qp;
333 
334 	spin_unlock_irq(&priv->lock);
335 
336 	for (i = 0; i < ipoib_recvq_size; ++i) {
337 		if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) {
338 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
339 				ret = -ENOMEM;
340 				goto err_count;
341 		}
342 		ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i);
343 		if (ret) {
344 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
345 				   "failed for buf %d\n", i);
346 			ret = -EIO;
347 			goto err_count;
348 		}
349 	}
350 
351 	rx->recv_count = ipoib_recvq_size;
352 
353 	kfree(t);
354 
355 	return 0;
356 
357 err_count:
358 	spin_lock_irq(&priv->lock);
359 	--priv->cm.nonsrq_conn_qp;
360 	spin_unlock_irq(&priv->lock);
361 
362 err_free:
363 	kfree(t);
364 	ipoib_cm_free_rx_ring(priv, rx->rx_ring);
365 
366 	return ret;
367 }
368 
369 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id,
370 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
371 			     unsigned psn)
372 {
373 	struct ipoib_cm_data data = {};
374 	struct ib_cm_rep_param rep = {};
375 
376 	data.qpn = cpu_to_be32(priv->qp->qp_num);
377 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
378 
379 	rep.private_data = &data;
380 	rep.private_data_len = sizeof data;
381 	rep.flow_control = 0;
382 	rep.rnr_retry_count = req->rnr_retry_count;
383 	rep.srq = ipoib_cm_has_srq(priv);
384 	rep.qp_num = qp->qp_num;
385 	rep.starting_psn = psn;
386 	return ib_send_cm_rep(cm_id, &rep);
387 }
388 
389 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
390 {
391 	struct ipoib_dev_priv *priv = cm_id->context;
392 	struct ipoib_cm_rx *p;
393 	unsigned psn;
394 	int ret;
395 
396 	ipoib_dbg(priv, "REQ arrived\n");
397 	p = kzalloc(sizeof *p, GFP_KERNEL);
398 	if (!p)
399 		return -ENOMEM;
400 	p->priv = priv;
401 	p->id = cm_id;
402 	cm_id->context = p;
403 	p->state = IPOIB_CM_RX_LIVE;
404 	p->jiffies = jiffies;
405 	INIT_LIST_HEAD(&p->list);
406 
407 	p->qp = ipoib_cm_create_rx_qp(priv, p);
408 	if (IS_ERR(p->qp)) {
409 		ret = PTR_ERR(p->qp);
410 		goto err_qp;
411 	}
412 
413 	psn = random() & 0xffffff;
414 	ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn);
415 	if (ret)
416 		goto err_modify;
417 
418 	if (!ipoib_cm_has_srq(priv)) {
419 		ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p);
420 		if (ret)
421 			goto err_modify;
422 	}
423 
424 	spin_lock_irq(&priv->lock);
425 	queue_delayed_work(ipoib_workqueue,
426 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
427 	/* Add this entry to passive ids list head, but do not re-add it
428 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
429 	p->jiffies = jiffies;
430 	if (p->state == IPOIB_CM_RX_LIVE)
431 		list_move(&p->list, &priv->cm.passive_ids);
432 	spin_unlock_irq(&priv->lock);
433 
434 	ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn);
435 	if (ret) {
436 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
437 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
438 			ipoib_warn(priv, "unable to move qp to error state\n");
439 	}
440 	return 0;
441 
442 err_modify:
443 	ib_destroy_qp(p->qp);
444 err_qp:
445 	kfree(p);
446 	return ret;
447 }
448 
449 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
450 			       struct ib_cm_event *event)
451 {
452 	struct ipoib_cm_rx *p;
453 	struct ipoib_dev_priv *priv;
454 
455 	switch (event->event) {
456 	case IB_CM_REQ_RECEIVED:
457 		return ipoib_cm_req_handler(cm_id, event);
458 	case IB_CM_DREQ_RECEIVED:
459 		p = cm_id->context;
460 		ib_send_cm_drep(cm_id, NULL, 0);
461 		/* Fall through */
462 	case IB_CM_REJ_RECEIVED:
463 		p = cm_id->context;
464 		priv = p->priv;
465 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
466 			ipoib_warn(priv, "unable to move qp to error state\n");
467 		/* Fall through */
468 	default:
469 		return 0;
470 	}
471 }
472 
473 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
474 {
475 	struct ipoib_cm_rx_buf saverx;
476 	struct ipoib_cm_rx_buf *rx_ring;
477 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
478 	struct ifnet *dev = priv->dev;
479 	struct mbuf *mb, *newmb;
480 	struct ipoib_cm_rx *p;
481 	int has_srq;
482 	u_short proto;
483 
484 	CURVNET_SET_QUIET(dev->if_vnet);
485 
486 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
487 		       wr_id, wc->status);
488 
489 	if (unlikely(wr_id >= ipoib_recvq_size)) {
490 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
491 			spin_lock(&priv->lock);
492 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
493 			ipoib_cm_start_rx_drain(priv);
494 			if (priv->cm.id != NULL)
495 				queue_work(ipoib_workqueue,
496 				    &priv->cm.rx_reap_task);
497 			spin_unlock(&priv->lock);
498 		} else
499 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
500 				   wr_id, ipoib_recvq_size);
501 		goto done;
502 	}
503 
504 	p = wc->qp->qp_context;
505 
506 	has_srq = ipoib_cm_has_srq(priv);
507 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
508 
509 	mb = rx_ring[wr_id].mb;
510 
511 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
512 		ipoib_dbg(priv, "cm recv error "
513 			   "(status=%d, wrid=%d vend_err %x)\n",
514 			   wc->status, wr_id, wc->vendor_err);
515 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
516 		if (has_srq)
517 			goto repost;
518 		else {
519 			if (!--p->recv_count) {
520 				spin_lock(&priv->lock);
521 				list_move(&p->list, &priv->cm.rx_reap_list);
522 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
523 				spin_unlock(&priv->lock);
524 			}
525 			goto done;
526 		}
527 	}
528 
529 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
530 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
531 			p->jiffies = jiffies;
532 			/* Move this entry to list head, but do not re-add it
533 			 * if it has been moved out of list. */
534 			if (p->state == IPOIB_CM_RX_LIVE)
535 				list_move(&p->list, &priv->cm.passive_ids);
536 		}
537 	}
538 
539 	memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx));
540 	newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]);
541 	if (unlikely(!newmb)) {
542 		/*
543 		 * If we can't allocate a new RX buffer, dump
544 		 * this packet and reuse the old buffer.
545 		 */
546 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
547 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
548 		memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx));
549 		goto repost;
550 	}
551 
552 	ipoib_cm_dma_unmap_rx(priv, &saverx);
553 
554 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
555 		       wc->byte_len, wc->slid);
556 
557 	ipoib_dma_mb(priv, mb, wc->byte_len);
558 
559 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
560 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
561 
562 	mb->m_pkthdr.rcvif = dev;
563 	proto = *mtod(mb, uint16_t *);
564 	m_adj(mb, IPOIB_ENCAP_LEN);
565 
566 	IPOIB_MTAP_PROTO(dev, mb, proto);
567 	ipoib_demux(dev, mb, ntohs(proto));
568 
569 repost:
570 	if (has_srq) {
571 		if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
572 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
573 				   "for buf %d\n", wr_id);
574 	} else {
575 		if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
576 							  &priv->cm.rx_wr,
577 							  priv->cm.rx_sge,
578 							  wr_id))) {
579 			--p->recv_count;
580 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
581 				   "for buf %d\n", wr_id);
582 		}
583 	}
584 done:
585 	CURVNET_RESTORE();
586 	return;
587 }
588 
589 static inline int post_send(struct ipoib_dev_priv *priv,
590 			    struct ipoib_cm_tx *tx,
591 			    struct ipoib_cm_tx_buf *tx_req,
592 			    unsigned int wr_id)
593 {
594 	struct ib_send_wr *bad_wr;
595 	struct mbuf *mb = tx_req->mb;
596 	u64 *mapping = tx_req->mapping;
597 	struct mbuf *m;
598 	int i;
599 
600 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
601 		priv->tx_sge[i].addr = mapping[i];
602 		priv->tx_sge[i].length = m->m_len;
603 	}
604 	priv->tx_wr.num_sge = i;
605 	priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM;
606 	priv->tx_wr.opcode = IB_WR_SEND;
607 
608 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
609 }
610 
611 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
612 {
613 	struct ipoib_cm_tx_buf *tx_req;
614 	struct ifnet *dev = priv->dev;
615 
616 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
617 		while (ipoib_poll_tx(priv)); /* nothing */
618 
619 	m_adj(mb, sizeof(struct ipoib_pseudoheader));
620 	if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
621 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
622 			   mb->m_pkthdr.len, tx->mtu);
623 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
624 		ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
625 		return;
626 	}
627 
628 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
629 		       tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
630 
631 
632 	/*
633 	 * We put the mb into the tx_ring _before_ we call post_send()
634 	 * because it's entirely possible that the completion handler will
635 	 * run before we execute anything after the post_send().  That
636 	 * means we have to make sure everything is properly recorded and
637 	 * our state is consistent before we call post_send().
638 	 */
639 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
640 	tx_req->mb = mb;
641 	if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
642 	    priv->cm.num_frags))) {
643 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
644 		if (tx_req->mb)
645 			m_freem(tx_req->mb);
646 		return;
647 	}
648 
649 	if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
650 		ipoib_warn(priv, "post_send failed\n");
651 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
652 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
653 		m_freem(mb);
654 	} else {
655 		++tx->tx_head;
656 
657 		if (++priv->tx_outstanding == ipoib_sendq_size) {
658 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
659 				  tx->qp->qp_num);
660 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
661 				ipoib_warn(priv, "request notify on send CQ failed\n");
662 			dev->if_drv_flags |= IFF_DRV_OACTIVE;
663 		}
664 	}
665 
666 }
667 
668 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
669 {
670 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
671 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
672 	struct ifnet *dev = priv->dev;
673 	struct ipoib_cm_tx_buf *tx_req;
674 
675 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
676 		       wr_id, wc->status);
677 
678 	if (unlikely(wr_id >= ipoib_sendq_size)) {
679 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
680 			   wr_id, ipoib_sendq_size);
681 		return;
682 	}
683 
684 	tx_req = &tx->tx_ring[wr_id];
685 
686 	ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
687 
688 	/* FIXME: is this right? Shouldn't we only increment on success? */
689 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
690 
691 	m_freem(tx_req->mb);
692 
693 	++tx->tx_tail;
694 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
695 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
696 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
697 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
698 
699 	if (wc->status != IB_WC_SUCCESS &&
700 	    wc->status != IB_WC_WR_FLUSH_ERR) {
701 		struct ipoib_path *path;
702 
703 		ipoib_dbg(priv, "failed cm send event "
704 			   "(status=%d, wrid=%d vend_err %x)\n",
705 			   wc->status, wr_id, wc->vendor_err);
706 
707 		path = tx->path;
708 
709 		if (path) {
710 			path->cm = NULL;
711 			rb_erase(&path->rb_node, &priv->path_tree);
712 			list_del(&path->list);
713 		}
714 
715 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
716 			list_move(&tx->list, &priv->cm.reap_list);
717 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
718 		}
719 
720 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
721 	}
722 
723 }
724 
725 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
726 {
727 	int ret;
728 
729 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)))
730 		return 0;
731 
732 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
733 	if (IS_ERR(priv->cm.id)) {
734 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
735 		ret = PTR_ERR(priv->cm.id);
736 		goto err_cm;
737 	}
738 
739 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
740 			   0, NULL);
741 	if (ret) {
742 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
743 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
744 		goto err_listen;
745 	}
746 
747 	return 0;
748 
749 err_listen:
750 	ib_destroy_cm_id(priv->cm.id);
751 err_cm:
752 	priv->cm.id = NULL;
753 	return ret;
754 }
755 
756 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
757 {
758 	struct ipoib_cm_rx *rx, *n;
759 	LIST_HEAD(list);
760 
761 	spin_lock_irq(&priv->lock);
762 	list_splice_init(&priv->cm.rx_reap_list, &list);
763 	spin_unlock_irq(&priv->lock);
764 
765 	list_for_each_entry_safe(rx, n, &list, list) {
766 		ib_destroy_cm_id(rx->id);
767 		ib_destroy_qp(rx->qp);
768 		if (!ipoib_cm_has_srq(priv)) {
769 			ipoib_cm_free_rx_ring(priv, rx->rx_ring);
770 			spin_lock_irq(&priv->lock);
771 			--priv->cm.nonsrq_conn_qp;
772 			spin_unlock_irq(&priv->lock);
773 		}
774 		kfree(rx);
775 	}
776 }
777 
778 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
779 {
780 	struct ipoib_cm_rx *p;
781 	unsigned long begin;
782 	int ret;
783 
784 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)) || !priv->cm.id)
785 		return;
786 
787 	ib_destroy_cm_id(priv->cm.id);
788 	priv->cm.id = NULL;
789 
790 	cancel_work_sync(&priv->cm.rx_reap_task);
791 
792 	spin_lock_irq(&priv->lock);
793 	while (!list_empty(&priv->cm.passive_ids)) {
794 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
795 		list_move(&p->list, &priv->cm.rx_error_list);
796 		p->state = IPOIB_CM_RX_ERROR;
797 		spin_unlock_irq(&priv->lock);
798 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
799 		if (ret)
800 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
801 		spin_lock_irq(&priv->lock);
802 	}
803 
804 	/* Wait for all RX to be drained */
805 	begin = jiffies;
806 
807 	while (!list_empty(&priv->cm.rx_error_list) ||
808 	       !list_empty(&priv->cm.rx_flush_list) ||
809 	       !list_empty(&priv->cm.rx_drain_list)) {
810 		if (time_after(jiffies, begin + 5 * HZ)) {
811 			ipoib_warn(priv, "RX drain timing out\n");
812 
813 			/*
814 			 * assume the HW is wedged and just free up everything.
815 			 */
816 			list_splice_init(&priv->cm.rx_flush_list,
817 					 &priv->cm.rx_reap_list);
818 			list_splice_init(&priv->cm.rx_error_list,
819 					 &priv->cm.rx_reap_list);
820 			list_splice_init(&priv->cm.rx_drain_list,
821 					 &priv->cm.rx_reap_list);
822 			break;
823 		}
824 		spin_unlock_irq(&priv->lock);
825 		msleep(1);
826 		ipoib_drain_cq(priv);
827 		spin_lock_irq(&priv->lock);
828 	}
829 
830 	spin_unlock_irq(&priv->lock);
831 
832 	ipoib_cm_free_rx_reap_list(priv);
833 
834 	cancel_delayed_work(&priv->cm.stale_task);
835 }
836 
837 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
838 {
839 	struct ipoib_cm_tx *p = cm_id->context;
840 	struct ipoib_dev_priv *priv = p->priv;
841 	struct ipoib_cm_data *data = event->private_data;
842 	struct ifqueue mbqueue;
843 	struct ib_qp_attr qp_attr;
844 	int qp_attr_mask, ret;
845 	struct mbuf *mb;
846 
847 	ipoib_dbg(priv, "cm rep handler\n");
848 	p->mtu = be32_to_cpu(data->mtu);
849 
850 	if (p->mtu <= IPOIB_ENCAP_LEN) {
851 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
852 			   p->mtu, IPOIB_ENCAP_LEN);
853 		return -EINVAL;
854 	}
855 
856 	qp_attr.qp_state = IB_QPS_RTR;
857 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
858 	if (ret) {
859 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
860 		return ret;
861 	}
862 
863 	qp_attr.rq_psn = 0 /* FIXME */;
864 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
865 	if (ret) {
866 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
867 		return ret;
868 	}
869 
870 	qp_attr.qp_state = IB_QPS_RTS;
871 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
872 	if (ret) {
873 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
874 		return ret;
875 	}
876 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
877 	if (ret) {
878 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
879 		return ret;
880 	}
881 
882 	bzero(&mbqueue, sizeof(mbqueue));
883 
884 	spin_lock_irq(&priv->lock);
885 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
886 	if (p->path)
887 		for (;;) {
888 			_IF_DEQUEUE(&p->path->queue, mb);
889 			if (mb == NULL)
890 				break;
891 			_IF_ENQUEUE(&mbqueue, mb);
892 		}
893 	spin_unlock_irq(&priv->lock);
894 
895 	for (;;) {
896 		struct ifnet *dev = p->priv->dev;
897 		_IF_DEQUEUE(&mbqueue, mb);
898 		if (mb == NULL)
899 			break;
900 		mb->m_pkthdr.rcvif = dev;
901 		if (dev->if_transmit(dev, mb))
902 			ipoib_warn(priv, "dev_queue_xmit failed "
903 				   "to requeue packet\n");
904 	}
905 
906 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
907 	if (ret) {
908 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
909 		return ret;
910 	}
911 	return 0;
912 }
913 
914 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
915     struct ipoib_cm_tx *tx)
916 {
917 	struct ib_qp_init_attr attr = {
918 		.send_cq		= priv->send_cq,
919 		.recv_cq		= priv->recv_cq,
920 		.srq			= priv->cm.srq,
921 		.cap.max_send_wr	= ipoib_sendq_size,
922 		.cap.max_send_sge	= priv->cm.num_frags,
923 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
924 		.qp_type		= IB_QPT_RC,
925 		.qp_context		= tx
926 	};
927 
928 	return ib_create_qp(priv->pd, &attr);
929 }
930 
931 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
932 			     struct ib_cm_id *id, struct ib_qp *qp,
933 			     u32 qpn,
934 			     struct ib_sa_path_rec *pathrec)
935 {
936 	struct ipoib_cm_data data = {};
937 	struct ib_cm_req_param req = {};
938 
939 	ipoib_dbg(priv, "cm send req\n");
940 
941 	data.qpn = cpu_to_be32(priv->qp->qp_num);
942 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
943 
944 	req.primary_path		= pathrec;
945 	req.alternate_path		= NULL;
946 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
947 	req.qp_num			= qp->qp_num;
948 	req.qp_type			= qp->qp_type;
949 	req.private_data		= &data;
950 	req.private_data_len		= sizeof data;
951 	req.flow_control		= 0;
952 
953 	req.starting_psn		= 0; /* FIXME */
954 
955 	/*
956 	 * Pick some arbitrary defaults here; we could make these
957 	 * module parameters if anyone cared about setting them.
958 	 */
959 	req.responder_resources		= 4;
960 	req.remote_cm_response_timeout	= 20;
961 	req.local_cm_response_timeout	= 20;
962 	req.retry_count			= 0; /* RFC draft warns against retries */
963 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
964 	req.max_cm_retries		= 15;
965 	req.srq				= ipoib_cm_has_srq(priv);
966 	return ib_send_cm_req(id, &req);
967 }
968 
969 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
970 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
971 {
972 	struct ib_qp_attr qp_attr;
973 	int qp_attr_mask, ret;
974 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
975 	if (ret) {
976 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
977 		return ret;
978 	}
979 
980 	qp_attr.qp_state = IB_QPS_INIT;
981 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
982 	qp_attr.port_num = priv->port;
983 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
984 
985 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
986 	if (ret) {
987 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
988 		return ret;
989 	}
990 	return 0;
991 }
992 
993 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
994 			    struct ib_sa_path_rec *pathrec)
995 {
996 	struct ipoib_dev_priv *priv = p->priv;
997 	int ret;
998 
999 	p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
1000 	if (!p->tx_ring) {
1001 		ipoib_warn(priv, "failed to allocate tx ring\n");
1002 		ret = -ENOMEM;
1003 		goto err_tx;
1004 	}
1005 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1006 
1007 	p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1008 	if (IS_ERR(p->qp)) {
1009 		ret = PTR_ERR(p->qp);
1010 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1011 		goto err_qp;
1012 	}
1013 
1014 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1015 	if (IS_ERR(p->id)) {
1016 		ret = PTR_ERR(p->id);
1017 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1018 		goto err_id;
1019 	}
1020 
1021 	ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1022 	if (ret) {
1023 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1024 		goto err_modify;
1025 	}
1026 
1027 	ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1028 	if (ret) {
1029 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1030 		goto err_send_cm;
1031 	}
1032 
1033 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1034 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1035 
1036 	return 0;
1037 
1038 err_send_cm:
1039 err_modify:
1040 	ib_destroy_cm_id(p->id);
1041 err_id:
1042 	p->id = NULL;
1043 	ib_destroy_qp(p->qp);
1044 err_qp:
1045 	p->qp = NULL;
1046 	kfree(p->tx_ring);
1047 err_tx:
1048 	return ret;
1049 }
1050 
1051 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1052 {
1053 	struct ipoib_dev_priv *priv = p->priv;
1054 	struct ifnet *dev = priv->dev;
1055 	struct ipoib_cm_tx_buf *tx_req;
1056 	unsigned long begin;
1057 
1058 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1059 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1060 
1061 	if (p->path)
1062 		ipoib_path_free(priv, p->path);
1063 
1064 	if (p->id)
1065 		ib_destroy_cm_id(p->id);
1066 
1067 	if (p->tx_ring) {
1068 		/* Wait for all sends to complete */
1069 		begin = jiffies;
1070 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1071 			if (time_after(jiffies, begin + 5 * HZ)) {
1072 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1073 					   p->tx_head - p->tx_tail);
1074 				goto timeout;
1075 			}
1076 
1077 			msleep(1);
1078 		}
1079 	}
1080 
1081 timeout:
1082 
1083 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1084 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1085 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1086 		m_freem(tx_req->mb);
1087 		++p->tx_tail;
1088 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1089 		    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
1090 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1091 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
1092 	}
1093 
1094 	if (p->qp)
1095 		ib_destroy_qp(p->qp);
1096 
1097 	kfree(p->tx_ring);
1098 	kfree(p);
1099 }
1100 
1101 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1102 			       struct ib_cm_event *event)
1103 {
1104 	struct ipoib_cm_tx *tx = cm_id->context;
1105 	struct ipoib_dev_priv *priv = tx->priv;
1106 	struct ipoib_path *path;
1107 	unsigned long flags;
1108 	int ret;
1109 
1110 	switch (event->event) {
1111 	case IB_CM_DREQ_RECEIVED:
1112 		ipoib_dbg(priv, "DREQ received.\n");
1113 		ib_send_cm_drep(cm_id, NULL, 0);
1114 		break;
1115 	case IB_CM_REP_RECEIVED:
1116 		ipoib_dbg(priv, "REP received.\n");
1117 		ret = ipoib_cm_rep_handler(cm_id, event);
1118 		if (ret)
1119 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1120 				       NULL, 0, NULL, 0);
1121 		break;
1122 	case IB_CM_REQ_ERROR:
1123 	case IB_CM_REJ_RECEIVED:
1124 	case IB_CM_TIMEWAIT_EXIT:
1125 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1126 		spin_lock_irqsave(&priv->lock, flags);
1127 		path = tx->path;
1128 
1129 		if (path) {
1130 			path->cm = NULL;
1131 			tx->path = NULL;
1132 			rb_erase(&path->rb_node, &priv->path_tree);
1133 			list_del(&path->list);
1134 		}
1135 
1136 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1137 			list_move(&tx->list, &priv->cm.reap_list);
1138 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1139 		}
1140 
1141 		spin_unlock_irqrestore(&priv->lock, flags);
1142 		if (path)
1143 			ipoib_path_free(tx->priv, path);
1144 		break;
1145 	default:
1146 		break;
1147 	}
1148 
1149 	return 0;
1150 }
1151 
1152 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1153     struct ipoib_path *path)
1154 {
1155 	struct ipoib_cm_tx *tx;
1156 
1157 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1158 	if (!tx)
1159 		return NULL;
1160 
1161 	ipoib_dbg(priv, "Creating cm tx\n");
1162 	path->cm = tx;
1163 	tx->path = path;
1164 	tx->priv = priv;
1165 	list_add(&tx->list, &priv->cm.start_list);
1166 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1167 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1168 	return tx;
1169 }
1170 
1171 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1172 {
1173 	struct ipoib_dev_priv *priv = tx->priv;
1174 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1175 		spin_lock(&priv->lock);
1176 		list_move(&tx->list, &priv->cm.reap_list);
1177 		spin_unlock(&priv->lock);
1178 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1179 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1180 			  tx->path->pathrec.dgid.raw);
1181 		tx->path = NULL;
1182 	}
1183 }
1184 
1185 static void ipoib_cm_tx_start(struct work_struct *work)
1186 {
1187 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1188 						   cm.start_task);
1189 	struct ipoib_path *path;
1190 	struct ipoib_cm_tx *p;
1191 	unsigned long flags;
1192 	int ret;
1193 
1194 	struct ib_sa_path_rec pathrec;
1195 	u32 qpn;
1196 
1197 	ipoib_dbg(priv, "cm start task\n");
1198 	spin_lock_irqsave(&priv->lock, flags);
1199 
1200 	while (!list_empty(&priv->cm.start_list)) {
1201 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1202 		list_del_init(&p->list);
1203 		path = p->path;
1204 		qpn = IPOIB_QPN(path->hwaddr);
1205 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1206 
1207 		spin_unlock_irqrestore(&priv->lock, flags);
1208 
1209 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1210 
1211 		spin_lock_irqsave(&priv->lock, flags);
1212 
1213 		if (ret) {
1214 			path = p->path;
1215 			if (path) {
1216 				path->cm = NULL;
1217 				rb_erase(&path->rb_node, &priv->path_tree);
1218 				list_del(&path->list);
1219 				ipoib_path_free(priv, path);
1220 			}
1221 			list_del(&p->list);
1222 			kfree(p);
1223 		}
1224 	}
1225 
1226 	spin_unlock_irqrestore(&priv->lock, flags);
1227 }
1228 
1229 static void ipoib_cm_tx_reap(struct work_struct *work)
1230 {
1231 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1232 						   cm.reap_task);
1233 	struct ipoib_cm_tx *p;
1234 	unsigned long flags;
1235 
1236 	spin_lock_irqsave(&priv->lock, flags);
1237 
1238 	while (!list_empty(&priv->cm.reap_list)) {
1239 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1240 		list_del(&p->list);
1241 		spin_unlock_irqrestore(&priv->lock, flags);
1242 		ipoib_cm_tx_destroy(p);
1243 		spin_lock_irqsave(&priv->lock, flags);
1244 	}
1245 
1246 	spin_unlock_irqrestore(&priv->lock, flags);
1247 }
1248 
1249 static void ipoib_cm_mb_reap(struct work_struct *work)
1250 {
1251 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1252 						   cm.mb_task);
1253 	struct mbuf *mb;
1254 	unsigned long flags;
1255 #if defined(INET) || defined(INET6)
1256 	unsigned mtu = priv->mcast_mtu;
1257 #endif
1258 	uint16_t proto;
1259 
1260 	spin_lock_irqsave(&priv->lock, flags);
1261 
1262 	for (;;) {
1263 		IF_DEQUEUE(&priv->cm.mb_queue, mb);
1264 		if (mb == NULL)
1265 			break;
1266 		spin_unlock_irqrestore(&priv->lock, flags);
1267 
1268 		proto = htons(*mtod(mb, uint16_t *));
1269 		m_adj(mb, IPOIB_ENCAP_LEN);
1270 		switch (proto) {
1271 #if defined(INET)
1272 		case ETHERTYPE_IP:
1273 			icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1274 			break;
1275 #endif
1276 #if defined(INET6)
1277 		case ETHERTYPE_IPV6:
1278 			icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1279 			break;
1280 #endif
1281 		default:
1282 			m_freem(mb);
1283 		}
1284 
1285 		spin_lock_irqsave(&priv->lock, flags);
1286 	}
1287 
1288 	spin_unlock_irqrestore(&priv->lock, flags);
1289 }
1290 
1291 void
1292 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1293 {
1294 	int e = priv->cm.mb_queue.ifq_len;
1295 
1296 	IF_ENQUEUE(&priv->cm.mb_queue, mb);
1297 	if (e == 0)
1298 		queue_work(ipoib_workqueue, &priv->cm.mb_task);
1299 }
1300 
1301 static void ipoib_cm_rx_reap(struct work_struct *work)
1302 {
1303 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1304 						cm.rx_reap_task));
1305 }
1306 
1307 static void ipoib_cm_stale_task(struct work_struct *work)
1308 {
1309 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1310 						   cm.stale_task.work);
1311 	struct ipoib_cm_rx *p;
1312 	int ret;
1313 
1314 	spin_lock_irq(&priv->lock);
1315 	while (!list_empty(&priv->cm.passive_ids)) {
1316 		/* List is sorted by LRU, start from tail,
1317 		 * stop when we see a recently used entry */
1318 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1319 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1320 			break;
1321 		list_move(&p->list, &priv->cm.rx_error_list);
1322 		p->state = IPOIB_CM_RX_ERROR;
1323 		spin_unlock_irq(&priv->lock);
1324 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1325 		if (ret)
1326 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1327 		spin_lock_irq(&priv->lock);
1328 	}
1329 
1330 	if (!list_empty(&priv->cm.passive_ids))
1331 		queue_delayed_work(ipoib_workqueue,
1332 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1333 	spin_unlock_irq(&priv->lock);
1334 }
1335 
1336 
1337 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1338 {
1339 	struct ib_srq_init_attr srq_init_attr = {
1340 		.attr = {
1341 			.max_wr  = ipoib_recvq_size,
1342 			.max_sge = max_sge
1343 		}
1344 	};
1345 
1346 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1347 	if (IS_ERR(priv->cm.srq)) {
1348 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1349 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1350 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1351 		priv->cm.srq = NULL;
1352 		return;
1353 	}
1354 
1355 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1356 	if (!priv->cm.srq_ring) {
1357 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1358 		       priv->ca->name, ipoib_recvq_size);
1359 		ib_destroy_srq(priv->cm.srq);
1360 		priv->cm.srq = NULL;
1361 		return;
1362 	}
1363 
1364 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1365 }
1366 
1367 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1368 {
1369 	struct ifnet *dev = priv->dev;
1370 	int i, ret;
1371 	struct ib_device_attr attr;
1372 
1373 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1374 	INIT_LIST_HEAD(&priv->cm.reap_list);
1375 	INIT_LIST_HEAD(&priv->cm.start_list);
1376 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1377 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1378 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1379 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1380 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1381 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1382 	INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1383 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1384 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1385 
1386 	bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1387 	mtx_init(&priv->cm.mb_queue.ifq_mtx,
1388 	    dev->if_xname, "if send queue", MTX_DEF);
1389 
1390 	ret = ib_query_device(priv->ca, &attr);
1391 	if (ret) {
1392 		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1393 		return ret;
1394 	}
1395 
1396 	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1397 
1398 	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1399 	ipoib_cm_create_srq(priv, attr.max_srq_sge);
1400 	if (ipoib_cm_has_srq(priv)) {
1401 		priv->cm.max_cm_mtu = attr.max_srq_sge * MJUMPAGESIZE;
1402 		priv->cm.num_frags  = attr.max_srq_sge;
1403 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1404 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1405 	} else {
1406 		priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1407 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1408 	}
1409 
1410 	ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1411 
1412 	if (ipoib_cm_has_srq(priv)) {
1413 		for (i = 0; i < ipoib_recvq_size; ++i) {
1414 			if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1415 				ipoib_warn(priv, "failed to allocate "
1416 					   "receive buffer %d\n", i);
1417 				ipoib_cm_dev_cleanup(priv);
1418 				return -ENOMEM;
1419 			}
1420 
1421 			if (ipoib_cm_post_receive_srq(priv, i)) {
1422 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1423 					   "failed for buf %d\n", i);
1424 				ipoib_cm_dev_cleanup(priv);
1425 				return -EIO;
1426 			}
1427 		}
1428 	}
1429 
1430 	IF_LLADDR(priv->dev)[0] = IPOIB_FLAGS_RC;
1431 	return 0;
1432 }
1433 
1434 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1435 {
1436 	int ret;
1437 
1438 	if (!priv->cm.srq)
1439 		return;
1440 
1441 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1442 
1443 	ret = ib_destroy_srq(priv->cm.srq);
1444 	if (ret)
1445 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1446 
1447 	priv->cm.srq = NULL;
1448 	if (!priv->cm.srq_ring)
1449 		return;
1450 
1451 	ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1452 	priv->cm.srq_ring = NULL;
1453 
1454 	mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1455 }
1456 
1457 #endif /* CONFIG_INFINIBAND_IPOIB_CM */
1458