xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision 98e0ffaefb0f241cda3a72395d3be04192ae0d47)
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 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
485 		       wr_id, wc->status);
486 
487 	if (unlikely(wr_id >= ipoib_recvq_size)) {
488 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
489 			spin_lock(&priv->lock);
490 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
491 			ipoib_cm_start_rx_drain(priv);
492 			if (priv->cm.id != NULL)
493 				queue_work(ipoib_workqueue,
494 				    &priv->cm.rx_reap_task);
495 			spin_unlock(&priv->lock);
496 		} else
497 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
498 				   wr_id, ipoib_recvq_size);
499 		return;
500 	}
501 
502 	p = wc->qp->qp_context;
503 
504 	has_srq = ipoib_cm_has_srq(priv);
505 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
506 
507 	mb = rx_ring[wr_id].mb;
508 
509 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
510 		ipoib_dbg(priv, "cm recv error "
511 			   "(status=%d, wrid=%d vend_err %x)\n",
512 			   wc->status, wr_id, wc->vendor_err);
513 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
514 		if (has_srq)
515 			goto repost;
516 		else {
517 			if (!--p->recv_count) {
518 				spin_lock(&priv->lock);
519 				list_move(&p->list, &priv->cm.rx_reap_list);
520 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
521 				spin_unlock(&priv->lock);
522 			}
523 			return;
524 		}
525 	}
526 
527 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
528 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
529 			p->jiffies = jiffies;
530 			/* Move this entry to list head, but do not re-add it
531 			 * if it has been moved out of list. */
532 			if (p->state == IPOIB_CM_RX_LIVE)
533 				list_move(&p->list, &priv->cm.passive_ids);
534 		}
535 	}
536 
537 	memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx));
538 	newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]);
539 	if (unlikely(!newmb)) {
540 		/*
541 		 * If we can't allocate a new RX buffer, dump
542 		 * this packet and reuse the old buffer.
543 		 */
544 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
545 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
546 		memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx));
547 		goto repost;
548 	}
549 
550 	ipoib_cm_dma_unmap_rx(priv, &saverx);
551 
552 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
553 		       wc->byte_len, wc->slid);
554 
555 	ipoib_dma_mb(priv, mb, wc->byte_len);
556 
557 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
558 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
559 
560 	mb->m_pkthdr.rcvif = dev;
561 	proto = *mtod(mb, uint16_t *);
562 	m_adj(mb, IPOIB_ENCAP_LEN);
563 
564 	IPOIB_MTAP_PROTO(dev, mb, proto);
565 	ipoib_demux(dev, mb, ntohs(proto));
566 
567 repost:
568 	if (has_srq) {
569 		if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
570 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
571 				   "for buf %d\n", wr_id);
572 	} else {
573 		if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
574 							  &priv->cm.rx_wr,
575 							  priv->cm.rx_sge,
576 							  wr_id))) {
577 			--p->recv_count;
578 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
579 				   "for buf %d\n", wr_id);
580 		}
581 	}
582 }
583 
584 static inline int post_send(struct ipoib_dev_priv *priv,
585 			    struct ipoib_cm_tx *tx,
586 			    struct ipoib_cm_tx_buf *tx_req,
587 			    unsigned int wr_id)
588 {
589 	struct ib_send_wr *bad_wr;
590 	struct mbuf *mb = tx_req->mb;
591 	u64 *mapping = tx_req->mapping;
592 	struct mbuf *m;
593 	int i;
594 
595 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
596 		priv->tx_sge[i].addr = mapping[i];
597 		priv->tx_sge[i].length = m->m_len;
598 	}
599 	priv->tx_wr.num_sge = i;
600 	priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM;
601 	priv->tx_wr.opcode = IB_WR_SEND;
602 
603 	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
604 }
605 
606 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
607 {
608 	struct ipoib_cm_tx_buf *tx_req;
609 	struct ifnet *dev = priv->dev;
610 
611 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
612 		while (ipoib_poll_tx(priv)); /* nothing */
613 
614 	m_adj(mb, sizeof(struct ipoib_pseudoheader));
615 	if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
616 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
617 			   mb->m_pkthdr.len, tx->mtu);
618 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
619 		ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
620 		return;
621 	}
622 
623 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
624 		       tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
625 
626 
627 	/*
628 	 * We put the mb into the tx_ring _before_ we call post_send()
629 	 * because it's entirely possible that the completion handler will
630 	 * run before we execute anything after the post_send().  That
631 	 * means we have to make sure everything is properly recorded and
632 	 * our state is consistent before we call post_send().
633 	 */
634 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
635 	tx_req->mb = mb;
636 	if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
637 	    priv->cm.num_frags))) {
638 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
639 		if (tx_req->mb)
640 			m_freem(tx_req->mb);
641 		return;
642 	}
643 
644 	if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
645 		ipoib_warn(priv, "post_send failed\n");
646 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
647 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
648 		m_freem(mb);
649 	} else {
650 		++tx->tx_head;
651 
652 		if (++priv->tx_outstanding == ipoib_sendq_size) {
653 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
654 				  tx->qp->qp_num);
655 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
656 				ipoib_warn(priv, "request notify on send CQ failed\n");
657 			dev->if_drv_flags |= IFF_DRV_OACTIVE;
658 		}
659 	}
660 
661 }
662 
663 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
664 {
665 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
666 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
667 	struct ifnet *dev = priv->dev;
668 	struct ipoib_cm_tx_buf *tx_req;
669 
670 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
671 		       wr_id, wc->status);
672 
673 	if (unlikely(wr_id >= ipoib_sendq_size)) {
674 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
675 			   wr_id, ipoib_sendq_size);
676 		return;
677 	}
678 
679 	tx_req = &tx->tx_ring[wr_id];
680 
681 	ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
682 
683 	/* FIXME: is this right? Shouldn't we only increment on success? */
684 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
685 
686 	m_freem(tx_req->mb);
687 
688 	++tx->tx_tail;
689 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
690 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
691 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
692 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
693 
694 	if (wc->status != IB_WC_SUCCESS &&
695 	    wc->status != IB_WC_WR_FLUSH_ERR) {
696 		struct ipoib_path *path;
697 
698 		ipoib_dbg(priv, "failed cm send event "
699 			   "(status=%d, wrid=%d vend_err %x)\n",
700 			   wc->status, wr_id, wc->vendor_err);
701 
702 		path = tx->path;
703 
704 		if (path) {
705 			path->cm = NULL;
706 			rb_erase(&path->rb_node, &priv->path_tree);
707 			list_del(&path->list);
708 		}
709 
710 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
711 			list_move(&tx->list, &priv->cm.reap_list);
712 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
713 		}
714 
715 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
716 	}
717 
718 }
719 
720 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
721 {
722 	int ret;
723 
724 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)))
725 		return 0;
726 
727 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
728 	if (IS_ERR(priv->cm.id)) {
729 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
730 		ret = PTR_ERR(priv->cm.id);
731 		goto err_cm;
732 	}
733 
734 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
735 			   0, NULL);
736 	if (ret) {
737 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
738 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
739 		goto err_listen;
740 	}
741 
742 	return 0;
743 
744 err_listen:
745 	ib_destroy_cm_id(priv->cm.id);
746 err_cm:
747 	priv->cm.id = NULL;
748 	return ret;
749 }
750 
751 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
752 {
753 	struct ipoib_cm_rx *rx, *n;
754 	LIST_HEAD(list);
755 
756 	spin_lock_irq(&priv->lock);
757 	list_splice_init(&priv->cm.rx_reap_list, &list);
758 	spin_unlock_irq(&priv->lock);
759 
760 	list_for_each_entry_safe(rx, n, &list, list) {
761 		ib_destroy_cm_id(rx->id);
762 		ib_destroy_qp(rx->qp);
763 		if (!ipoib_cm_has_srq(priv)) {
764 			ipoib_cm_free_rx_ring(priv, rx->rx_ring);
765 			spin_lock_irq(&priv->lock);
766 			--priv->cm.nonsrq_conn_qp;
767 			spin_unlock_irq(&priv->lock);
768 		}
769 		kfree(rx);
770 	}
771 }
772 
773 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
774 {
775 	struct ipoib_cm_rx *p;
776 	unsigned long begin;
777 	int ret;
778 
779 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)) || !priv->cm.id)
780 		return;
781 
782 	ib_destroy_cm_id(priv->cm.id);
783 	priv->cm.id = NULL;
784 
785 	cancel_work_sync(&priv->cm.rx_reap_task);
786 
787 	spin_lock_irq(&priv->lock);
788 	while (!list_empty(&priv->cm.passive_ids)) {
789 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
790 		list_move(&p->list, &priv->cm.rx_error_list);
791 		p->state = IPOIB_CM_RX_ERROR;
792 		spin_unlock_irq(&priv->lock);
793 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
794 		if (ret)
795 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
796 		spin_lock_irq(&priv->lock);
797 	}
798 
799 	/* Wait for all RX to be drained */
800 	begin = jiffies;
801 
802 	while (!list_empty(&priv->cm.rx_error_list) ||
803 	       !list_empty(&priv->cm.rx_flush_list) ||
804 	       !list_empty(&priv->cm.rx_drain_list)) {
805 		if (time_after(jiffies, begin + 5 * HZ)) {
806 			ipoib_warn(priv, "RX drain timing out\n");
807 
808 			/*
809 			 * assume the HW is wedged and just free up everything.
810 			 */
811 			list_splice_init(&priv->cm.rx_flush_list,
812 					 &priv->cm.rx_reap_list);
813 			list_splice_init(&priv->cm.rx_error_list,
814 					 &priv->cm.rx_reap_list);
815 			list_splice_init(&priv->cm.rx_drain_list,
816 					 &priv->cm.rx_reap_list);
817 			break;
818 		}
819 		spin_unlock_irq(&priv->lock);
820 		msleep(1);
821 		ipoib_drain_cq(priv);
822 		spin_lock_irq(&priv->lock);
823 	}
824 
825 	spin_unlock_irq(&priv->lock);
826 
827 	ipoib_cm_free_rx_reap_list(priv);
828 
829 	cancel_delayed_work(&priv->cm.stale_task);
830 }
831 
832 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
833 {
834 	struct ipoib_cm_tx *p = cm_id->context;
835 	struct ipoib_dev_priv *priv = p->priv;
836 	struct ipoib_cm_data *data = event->private_data;
837 	struct ifqueue mbqueue;
838 	struct ib_qp_attr qp_attr;
839 	int qp_attr_mask, ret;
840 	struct mbuf *mb;
841 
842 	ipoib_dbg(priv, "cm rep handler\n");
843 	p->mtu = be32_to_cpu(data->mtu);
844 
845 	if (p->mtu <= IPOIB_ENCAP_LEN) {
846 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
847 			   p->mtu, IPOIB_ENCAP_LEN);
848 		return -EINVAL;
849 	}
850 
851 	qp_attr.qp_state = IB_QPS_RTR;
852 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
853 	if (ret) {
854 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
855 		return ret;
856 	}
857 
858 	qp_attr.rq_psn = 0 /* FIXME */;
859 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
860 	if (ret) {
861 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
862 		return ret;
863 	}
864 
865 	qp_attr.qp_state = IB_QPS_RTS;
866 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
867 	if (ret) {
868 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
869 		return ret;
870 	}
871 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
872 	if (ret) {
873 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
874 		return ret;
875 	}
876 
877 	bzero(&mbqueue, sizeof(mbqueue));
878 
879 	spin_lock_irq(&priv->lock);
880 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
881 	if (p->path)
882 		for (;;) {
883 			_IF_DEQUEUE(&p->path->queue, mb);
884 			if (mb == NULL)
885 				break;
886 			_IF_ENQUEUE(&mbqueue, mb);
887 		}
888 	spin_unlock_irq(&priv->lock);
889 
890 	for (;;) {
891 		struct ifnet *dev = p->priv->dev;
892 		_IF_DEQUEUE(&mbqueue, mb);
893 		if (mb == NULL)
894 			break;
895 		mb->m_pkthdr.rcvif = dev;
896 		if (dev->if_transmit(dev, mb))
897 			ipoib_warn(priv, "dev_queue_xmit failed "
898 				   "to requeue packet\n");
899 	}
900 
901 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
902 	if (ret) {
903 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
904 		return ret;
905 	}
906 	return 0;
907 }
908 
909 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
910     struct ipoib_cm_tx *tx)
911 {
912 	struct ib_qp_init_attr attr = {
913 		.send_cq		= priv->send_cq,
914 		.recv_cq		= priv->recv_cq,
915 		.srq			= priv->cm.srq,
916 		.cap.max_send_wr	= ipoib_sendq_size,
917 		.cap.max_send_sge	= priv->cm.num_frags,
918 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
919 		.qp_type		= IB_QPT_RC,
920 		.qp_context		= tx
921 	};
922 
923 	return ib_create_qp(priv->pd, &attr);
924 }
925 
926 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
927 			     struct ib_cm_id *id, struct ib_qp *qp,
928 			     u32 qpn,
929 			     struct ib_sa_path_rec *pathrec)
930 {
931 	struct ipoib_cm_data data = {};
932 	struct ib_cm_req_param req = {};
933 
934 	ipoib_dbg(priv, "cm send req\n");
935 
936 	data.qpn = cpu_to_be32(priv->qp->qp_num);
937 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
938 
939 	req.primary_path		= pathrec;
940 	req.alternate_path		= NULL;
941 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
942 	req.qp_num			= qp->qp_num;
943 	req.qp_type			= qp->qp_type;
944 	req.private_data		= &data;
945 	req.private_data_len		= sizeof data;
946 	req.flow_control		= 0;
947 
948 	req.starting_psn		= 0; /* FIXME */
949 
950 	/*
951 	 * Pick some arbitrary defaults here; we could make these
952 	 * module parameters if anyone cared about setting them.
953 	 */
954 	req.responder_resources		= 4;
955 	req.remote_cm_response_timeout	= 20;
956 	req.local_cm_response_timeout	= 20;
957 	req.retry_count			= 0; /* RFC draft warns against retries */
958 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
959 	req.max_cm_retries		= 15;
960 	req.srq				= ipoib_cm_has_srq(priv);
961 	return ib_send_cm_req(id, &req);
962 }
963 
964 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
965 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
966 {
967 	struct ib_qp_attr qp_attr;
968 	int qp_attr_mask, ret;
969 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
970 	if (ret) {
971 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
972 		return ret;
973 	}
974 
975 	qp_attr.qp_state = IB_QPS_INIT;
976 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
977 	qp_attr.port_num = priv->port;
978 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
979 
980 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
981 	if (ret) {
982 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
983 		return ret;
984 	}
985 	return 0;
986 }
987 
988 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
989 			    struct ib_sa_path_rec *pathrec)
990 {
991 	struct ipoib_dev_priv *priv = p->priv;
992 	int ret;
993 
994 	p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
995 	if (!p->tx_ring) {
996 		ipoib_warn(priv, "failed to allocate tx ring\n");
997 		ret = -ENOMEM;
998 		goto err_tx;
999 	}
1000 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1001 
1002 	p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1003 	if (IS_ERR(p->qp)) {
1004 		ret = PTR_ERR(p->qp);
1005 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1006 		goto err_qp;
1007 	}
1008 
1009 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1010 	if (IS_ERR(p->id)) {
1011 		ret = PTR_ERR(p->id);
1012 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1013 		goto err_id;
1014 	}
1015 
1016 	ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1017 	if (ret) {
1018 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1019 		goto err_modify;
1020 	}
1021 
1022 	ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1023 	if (ret) {
1024 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1025 		goto err_send_cm;
1026 	}
1027 
1028 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1029 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1030 
1031 	return 0;
1032 
1033 err_send_cm:
1034 err_modify:
1035 	ib_destroy_cm_id(p->id);
1036 err_id:
1037 	p->id = NULL;
1038 	ib_destroy_qp(p->qp);
1039 err_qp:
1040 	p->qp = NULL;
1041 	kfree(p->tx_ring);
1042 err_tx:
1043 	return ret;
1044 }
1045 
1046 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1047 {
1048 	struct ipoib_dev_priv *priv = p->priv;
1049 	struct ifnet *dev = priv->dev;
1050 	struct ipoib_cm_tx_buf *tx_req;
1051 	unsigned long begin;
1052 
1053 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1054 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1055 
1056 	if (p->path)
1057 		ipoib_path_free(priv, p->path);
1058 
1059 	if (p->id)
1060 		ib_destroy_cm_id(p->id);
1061 
1062 	if (p->tx_ring) {
1063 		/* Wait for all sends to complete */
1064 		begin = jiffies;
1065 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1066 			if (time_after(jiffies, begin + 5 * HZ)) {
1067 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1068 					   p->tx_head - p->tx_tail);
1069 				goto timeout;
1070 			}
1071 
1072 			msleep(1);
1073 		}
1074 	}
1075 
1076 timeout:
1077 
1078 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1079 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1080 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1081 		m_freem(tx_req->mb);
1082 		++p->tx_tail;
1083 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1084 		    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
1085 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1086 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
1087 	}
1088 
1089 	if (p->qp)
1090 		ib_destroy_qp(p->qp);
1091 
1092 	kfree(p->tx_ring);
1093 	kfree(p);
1094 }
1095 
1096 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1097 			       struct ib_cm_event *event)
1098 {
1099 	struct ipoib_cm_tx *tx = cm_id->context;
1100 	struct ipoib_dev_priv *priv = tx->priv;
1101 	struct ipoib_path *path;
1102 	unsigned long flags;
1103 	int ret;
1104 
1105 	switch (event->event) {
1106 	case IB_CM_DREQ_RECEIVED:
1107 		ipoib_dbg(priv, "DREQ received.\n");
1108 		ib_send_cm_drep(cm_id, NULL, 0);
1109 		break;
1110 	case IB_CM_REP_RECEIVED:
1111 		ipoib_dbg(priv, "REP received.\n");
1112 		ret = ipoib_cm_rep_handler(cm_id, event);
1113 		if (ret)
1114 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1115 				       NULL, 0, NULL, 0);
1116 		break;
1117 	case IB_CM_REQ_ERROR:
1118 	case IB_CM_REJ_RECEIVED:
1119 	case IB_CM_TIMEWAIT_EXIT:
1120 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1121 		spin_lock_irqsave(&priv->lock, flags);
1122 		path = tx->path;
1123 
1124 		if (path) {
1125 			path->cm = NULL;
1126 			tx->path = NULL;
1127 			rb_erase(&path->rb_node, &priv->path_tree);
1128 			list_del(&path->list);
1129 		}
1130 
1131 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1132 			list_move(&tx->list, &priv->cm.reap_list);
1133 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1134 		}
1135 
1136 		spin_unlock_irqrestore(&priv->lock, flags);
1137 		if (path)
1138 			ipoib_path_free(tx->priv, path);
1139 		break;
1140 	default:
1141 		break;
1142 	}
1143 
1144 	return 0;
1145 }
1146 
1147 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1148     struct ipoib_path *path)
1149 {
1150 	struct ipoib_cm_tx *tx;
1151 
1152 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1153 	if (!tx)
1154 		return NULL;
1155 
1156 	ipoib_dbg(priv, "Creating cm tx\n");
1157 	path->cm = tx;
1158 	tx->path = path;
1159 	tx->priv = priv;
1160 	list_add(&tx->list, &priv->cm.start_list);
1161 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1162 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1163 	return tx;
1164 }
1165 
1166 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1167 {
1168 	struct ipoib_dev_priv *priv = tx->priv;
1169 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1170 		spin_lock(&priv->lock);
1171 		list_move(&tx->list, &priv->cm.reap_list);
1172 		spin_unlock(&priv->lock);
1173 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1174 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1175 			  tx->path->pathrec.dgid.raw);
1176 		tx->path = NULL;
1177 	}
1178 }
1179 
1180 static void ipoib_cm_tx_start(struct work_struct *work)
1181 {
1182 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1183 						   cm.start_task);
1184 	struct ipoib_path *path;
1185 	struct ipoib_cm_tx *p;
1186 	unsigned long flags;
1187 	int ret;
1188 
1189 	struct ib_sa_path_rec pathrec;
1190 	u32 qpn;
1191 
1192 	ipoib_dbg(priv, "cm start task\n");
1193 	spin_lock_irqsave(&priv->lock, flags);
1194 
1195 	while (!list_empty(&priv->cm.start_list)) {
1196 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1197 		list_del_init(&p->list);
1198 		path = p->path;
1199 		qpn = IPOIB_QPN(path->hwaddr);
1200 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1201 
1202 		spin_unlock_irqrestore(&priv->lock, flags);
1203 
1204 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1205 
1206 		spin_lock_irqsave(&priv->lock, flags);
1207 
1208 		if (ret) {
1209 			path = p->path;
1210 			if (path) {
1211 				path->cm = NULL;
1212 				rb_erase(&path->rb_node, &priv->path_tree);
1213 				list_del(&path->list);
1214 				ipoib_path_free(priv, path);
1215 			}
1216 			list_del(&p->list);
1217 			kfree(p);
1218 		}
1219 	}
1220 
1221 	spin_unlock_irqrestore(&priv->lock, flags);
1222 }
1223 
1224 static void ipoib_cm_tx_reap(struct work_struct *work)
1225 {
1226 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1227 						   cm.reap_task);
1228 	struct ipoib_cm_tx *p;
1229 	unsigned long flags;
1230 
1231 	spin_lock_irqsave(&priv->lock, flags);
1232 
1233 	while (!list_empty(&priv->cm.reap_list)) {
1234 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1235 		list_del(&p->list);
1236 		spin_unlock_irqrestore(&priv->lock, flags);
1237 		ipoib_cm_tx_destroy(p);
1238 		spin_lock_irqsave(&priv->lock, flags);
1239 	}
1240 
1241 	spin_unlock_irqrestore(&priv->lock, flags);
1242 }
1243 
1244 static void ipoib_cm_mb_reap(struct work_struct *work)
1245 {
1246 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1247 						   cm.mb_task);
1248 	struct mbuf *mb;
1249 	unsigned long flags;
1250 #if defined(INET) || defined(INET6)
1251 	unsigned mtu = priv->mcast_mtu;
1252 #endif
1253 	uint16_t proto;
1254 
1255 	spin_lock_irqsave(&priv->lock, flags);
1256 
1257 	for (;;) {
1258 		IF_DEQUEUE(&priv->cm.mb_queue, mb);
1259 		if (mb == NULL)
1260 			break;
1261 		spin_unlock_irqrestore(&priv->lock, flags);
1262 
1263 		proto = htons(*mtod(mb, uint16_t *));
1264 		m_adj(mb, IPOIB_ENCAP_LEN);
1265 		switch (proto) {
1266 #if defined(INET)
1267 		case ETHERTYPE_IP:
1268 			icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1269 			break;
1270 #endif
1271 #if defined(INET6)
1272 		case ETHERTYPE_IPV6:
1273 			icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1274 			break;
1275 #endif
1276 		default:
1277 			m_freem(mb);
1278 		}
1279 
1280 		spin_lock_irqsave(&priv->lock, flags);
1281 	}
1282 
1283 	spin_unlock_irqrestore(&priv->lock, flags);
1284 }
1285 
1286 void
1287 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1288 {
1289 	int e = priv->cm.mb_queue.ifq_len;
1290 
1291 	IF_ENQUEUE(&priv->cm.mb_queue, mb);
1292 	if (e == 0)
1293 		queue_work(ipoib_workqueue, &priv->cm.mb_task);
1294 }
1295 
1296 static void ipoib_cm_rx_reap(struct work_struct *work)
1297 {
1298 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1299 						cm.rx_reap_task));
1300 }
1301 
1302 static void ipoib_cm_stale_task(struct work_struct *work)
1303 {
1304 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1305 						   cm.stale_task.work);
1306 	struct ipoib_cm_rx *p;
1307 	int ret;
1308 
1309 	spin_lock_irq(&priv->lock);
1310 	while (!list_empty(&priv->cm.passive_ids)) {
1311 		/* List is sorted by LRU, start from tail,
1312 		 * stop when we see a recently used entry */
1313 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1314 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1315 			break;
1316 		list_move(&p->list, &priv->cm.rx_error_list);
1317 		p->state = IPOIB_CM_RX_ERROR;
1318 		spin_unlock_irq(&priv->lock);
1319 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1320 		if (ret)
1321 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1322 		spin_lock_irq(&priv->lock);
1323 	}
1324 
1325 	if (!list_empty(&priv->cm.passive_ids))
1326 		queue_delayed_work(ipoib_workqueue,
1327 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1328 	spin_unlock_irq(&priv->lock);
1329 }
1330 
1331 
1332 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1333 {
1334 	struct ib_srq_init_attr srq_init_attr = {
1335 		.attr = {
1336 			.max_wr  = ipoib_recvq_size,
1337 			.max_sge = max_sge
1338 		}
1339 	};
1340 
1341 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1342 	if (IS_ERR(priv->cm.srq)) {
1343 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1344 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1345 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1346 		priv->cm.srq = NULL;
1347 		return;
1348 	}
1349 
1350 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1351 	if (!priv->cm.srq_ring) {
1352 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1353 		       priv->ca->name, ipoib_recvq_size);
1354 		ib_destroy_srq(priv->cm.srq);
1355 		priv->cm.srq = NULL;
1356 		return;
1357 	}
1358 
1359 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1360 }
1361 
1362 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1363 {
1364 	struct ifnet *dev = priv->dev;
1365 	int i, ret;
1366 	struct ib_device_attr attr;
1367 
1368 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1369 	INIT_LIST_HEAD(&priv->cm.reap_list);
1370 	INIT_LIST_HEAD(&priv->cm.start_list);
1371 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1372 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1373 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1374 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1375 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1376 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1377 	INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1378 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1379 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1380 
1381 	bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1382 	mtx_init(&priv->cm.mb_queue.ifq_mtx,
1383 	    dev->if_xname, "if send queue", MTX_DEF);
1384 
1385 	ret = ib_query_device(priv->ca, &attr);
1386 	if (ret) {
1387 		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1388 		return ret;
1389 	}
1390 
1391 	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1392 
1393 	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1394 	ipoib_cm_create_srq(priv, attr.max_srq_sge);
1395 	if (ipoib_cm_has_srq(priv)) {
1396 		priv->cm.max_cm_mtu = attr.max_srq_sge * MJUMPAGESIZE;
1397 		priv->cm.num_frags  = attr.max_srq_sge;
1398 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1399 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1400 	} else {
1401 		priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1402 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1403 	}
1404 
1405 	ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1406 
1407 	if (ipoib_cm_has_srq(priv)) {
1408 		for (i = 0; i < ipoib_recvq_size; ++i) {
1409 			if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1410 				ipoib_warn(priv, "failed to allocate "
1411 					   "receive buffer %d\n", i);
1412 				ipoib_cm_dev_cleanup(priv);
1413 				return -ENOMEM;
1414 			}
1415 
1416 			if (ipoib_cm_post_receive_srq(priv, i)) {
1417 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1418 					   "failed for buf %d\n", i);
1419 				ipoib_cm_dev_cleanup(priv);
1420 				return -EIO;
1421 			}
1422 		}
1423 	}
1424 
1425 	IF_LLADDR(priv->dev)[0] = IPOIB_FLAGS_RC;
1426 	return 0;
1427 }
1428 
1429 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1430 {
1431 	int ret;
1432 
1433 	if (!priv->cm.srq)
1434 		return;
1435 
1436 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1437 
1438 	ret = ib_destroy_srq(priv->cm.srq);
1439 	if (ret)
1440 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1441 
1442 	priv->cm.srq = NULL;
1443 	if (!priv->cm.srq_ring)
1444 		return;
1445 
1446 	ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1447 	priv->cm.srq_ring = NULL;
1448 
1449 	mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1450 }
1451 
1452 #endif /* CONFIG_INFINIBAND_IPOIB_CM */
1453