xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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 #include "ipoib.h"
37 
38 #ifdef CONFIG_INFINIBAND_IPOIB_CM
39 
40 #include <netinet/ip.h>
41 #include <netinet/ip_icmp.h>
42 #include <netinet/icmp6.h>
43 
44 #include <rdma/ib_cm.h>
45 #include <rdma/ib_cache.h>
46 #include <linux/delay.h>
47 
48 int ipoib_max_conn_qp = 128;
49 
50 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
51 MODULE_PARM_DESC(max_nonsrq_conn_qp,
52 		 "Max number of connected-mode QPs per interface "
53 		 "(applied only if shared receive queue is not available)");
54 
55 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
56 static int data_debug_level;
57 
58 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
59 MODULE_PARM_DESC(cm_data_debug_level,
60 		 "Enable data path debug tracing for connected mode if > 0");
61 #endif
62 
63 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
64 
65 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
66 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
67 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
68 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
69 
70 static struct ib_qp_attr ipoib_cm_err_attr = {
71 	.qp_state = IB_QPS_ERR
72 };
73 
74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
75 
76 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
77 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
78 	.opcode = IB_WR_SEND,
79 };
80 
81 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
82 			       struct ib_cm_event *event);
83 
84 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
85 {
86 
87 	ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req);
88 
89 }
90 
91 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id)
92 {
93 	const struct ib_recv_wr *bad_wr;
94 	struct ipoib_rx_buf *rx_req;
95 	struct mbuf *m;
96 	int ret;
97 	int i;
98 
99 	rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id];
100 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
101 		priv->cm.rx_sge[i].addr = rx_req->mapping[i];
102 		priv->cm.rx_sge[i].length = m->m_len;
103 	}
104 
105 	priv->cm.rx_wr.num_sge = i;
106 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
107 
108 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
109 	if (unlikely(ret)) {
110 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
111 		ipoib_dma_unmap_rx(priv, rx_req);
112 		m_freem(priv->cm.srq_ring[id].mb);
113 		priv->cm.srq_ring[id].mb = NULL;
114 	}
115 
116 	return ret;
117 }
118 
119 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv,
120 					struct ipoib_cm_rx *rx,
121 					struct ib_recv_wr *wr,
122 					struct ib_sge *sge, int id)
123 {
124 	struct ipoib_rx_buf *rx_req;
125 	const struct ib_recv_wr *bad_wr;
126 	struct mbuf *m;
127 	int ret;
128 	int i;
129 
130 	rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id];
131 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
132 		sge[i].addr = rx_req->mapping[i];
133 		sge[i].length = m->m_len;
134 	}
135 
136 	wr->num_sge = i;
137 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
138 
139 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
140 	if (unlikely(ret)) {
141 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
142 		ipoib_dma_unmap_rx(priv, rx_req);
143 		m_freem(rx->rx_ring[id].mb);
144 		rx->rx_ring[id].mb = NULL;
145 	}
146 
147 	return ret;
148 }
149 
150 static struct mbuf *
151 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
152 {
153 	return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req,
154 	    sizeof(struct ipoib_pseudoheader), priv->cm.max_cm_mtu, IPOIB_CM_RX_SG);
155 }
156 
157 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv,
158 				  struct ipoib_cm_rx_buf *rx_ring)
159 {
160 	int i;
161 
162 	for (i = 0; i < ipoib_recvq_size; ++i)
163 		if (rx_ring[i].mb) {
164 			ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]);
165 			m_freem(rx_ring[i].mb);
166 		}
167 
168 	kfree(rx_ring);
169 }
170 
171 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
172 {
173 	const struct ib_send_wr *bad_wr;
174 	struct ipoib_cm_rx *p;
175 
176 	/* We only reserved 1 extra slot in CQ for drain WRs, so
177 	 * make sure we have at most 1 outstanding WR. */
178 	if (list_empty(&priv->cm.rx_flush_list) ||
179 	    !list_empty(&priv->cm.rx_drain_list))
180 		return;
181 
182 	/*
183 	 * QPs on flush list are error state.  This way, a "flush
184 	 * error" WC will be immediately generated for each WR we post.
185 	 */
186 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
187 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
188 		ipoib_warn(priv, "failed to post drain wr\n");
189 
190 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
191 }
192 
193 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
194 {
195 	struct ipoib_cm_rx *p = ctx;
196 	struct ipoib_dev_priv *priv = p->priv;
197 	unsigned long flags;
198 
199 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
200 		return;
201 
202 	spin_lock_irqsave(&priv->lock, flags);
203 	list_move(&p->list, &priv->cm.rx_flush_list);
204 	p->state = IPOIB_CM_RX_FLUSH;
205 	ipoib_cm_start_rx_drain(priv);
206 	spin_unlock_irqrestore(&priv->lock, flags);
207 }
208 
209 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv,
210 					   struct ipoib_cm_rx *p)
211 {
212 	struct ib_qp_init_attr attr = {
213 		.event_handler = ipoib_cm_rx_event_handler,
214 		.send_cq = priv->recv_cq, /* For drain WR */
215 		.recv_cq = priv->recv_cq,
216 		.srq = priv->cm.srq,
217 		.cap.max_send_wr = 1, /* For drain WR */
218 		.cap.max_send_sge = 1,
219 		.sq_sig_type = IB_SIGNAL_ALL_WR,
220 		.qp_type = IB_QPT_RC,
221 		.qp_context = p,
222 	};
223 
224 	if (!ipoib_cm_has_srq(priv)) {
225 		attr.cap.max_recv_wr  = ipoib_recvq_size;
226 		attr.cap.max_recv_sge = priv->cm.num_frags;
227 	}
228 
229 	return ib_create_qp(priv->pd, &attr);
230 }
231 
232 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv,
233 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
234 				 unsigned psn)
235 {
236 	struct ib_qp_attr qp_attr;
237 	int qp_attr_mask, ret;
238 
239 	qp_attr.qp_state = IB_QPS_INIT;
240 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
241 	if (ret) {
242 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
243 		return ret;
244 	}
245 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
246 	if (ret) {
247 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
248 		return ret;
249 	}
250 	qp_attr.qp_state = IB_QPS_RTR;
251 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
252 	if (ret) {
253 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
254 		return ret;
255 	}
256 	qp_attr.rq_psn = psn;
257 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
258 	if (ret) {
259 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
260 		return ret;
261 	}
262 
263 	/*
264 	 * Current Mellanox HCA firmware won't generate completions
265 	 * with error for drain WRs unless the QP has been moved to
266 	 * RTS first. This work-around leaves a window where a QP has
267 	 * moved to error asynchronously, but this will eventually get
268 	 * fixed in firmware, so let's not error out if modify QP
269 	 * fails.
270 	 */
271 	qp_attr.qp_state = IB_QPS_RTS;
272 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
273 	if (ret) {
274 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
275 		return 0;
276 	}
277 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
278 	if (ret) {
279 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
280 		return 0;
281 	}
282 
283 	return 0;
284 }
285 
286 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv,
287 				struct ib_recv_wr *wr,
288 				struct ib_sge *sge)
289 {
290 	int i;
291 
292 	for (i = 0; i < IPOIB_CM_RX_SG; i++)
293 		sge[i].lkey = priv->pd->local_dma_lkey;
294 
295 	wr->next    = NULL;
296 	wr->sg_list = sge;
297 	wr->num_sge = 1;
298 }
299 
300 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv,
301     struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx)
302 {
303 	struct {
304 		struct ib_recv_wr wr;
305 		struct ib_sge sge[IPOIB_CM_RX_SG];
306 	} *t;
307 	int ret;
308 	int i;
309 
310 	rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL);
311 	if (!rx->rx_ring) {
312 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
313 		       priv->ca->name, ipoib_recvq_size);
314 		return -ENOMEM;
315 	}
316 
317 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
318 
319 	t = kmalloc(sizeof *t, GFP_KERNEL);
320 	if (!t) {
321 		ret = -ENOMEM;
322 		goto err_free;
323 	}
324 
325 	ipoib_cm_init_rx_wr(priv, &t->wr, t->sge);
326 
327 	spin_lock_irq(&priv->lock);
328 
329 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
330 		spin_unlock_irq(&priv->lock);
331 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
332 		ret = -EINVAL;
333 		goto err_free;
334 	} else
335 		++priv->cm.nonsrq_conn_qp;
336 
337 	spin_unlock_irq(&priv->lock);
338 
339 	for (i = 0; i < ipoib_recvq_size; ++i) {
340 		if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) {
341 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
342 				ret = -ENOMEM;
343 				goto err_count;
344 		}
345 		ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i);
346 		if (ret) {
347 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
348 				   "failed for buf %d\n", i);
349 			ret = -EIO;
350 			goto err_count;
351 		}
352 	}
353 
354 	rx->recv_count = ipoib_recvq_size;
355 
356 	kfree(t);
357 
358 	return 0;
359 
360 err_count:
361 	spin_lock_irq(&priv->lock);
362 	--priv->cm.nonsrq_conn_qp;
363 	spin_unlock_irq(&priv->lock);
364 
365 err_free:
366 	kfree(t);
367 	ipoib_cm_free_rx_ring(priv, rx->rx_ring);
368 
369 	return ret;
370 }
371 
372 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id,
373 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
374 			     unsigned psn)
375 {
376 	struct ipoib_cm_data data = {};
377 	struct ib_cm_rep_param rep = {};
378 
379 	data.qpn = cpu_to_be32(priv->qp->qp_num);
380 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
381 
382 	rep.private_data = &data;
383 	rep.private_data_len = sizeof data;
384 	rep.flow_control = 0;
385 	rep.rnr_retry_count = req->rnr_retry_count;
386 	rep.srq = ipoib_cm_has_srq(priv);
387 	rep.qp_num = qp->qp_num;
388 	rep.starting_psn = psn;
389 	return ib_send_cm_rep(cm_id, &rep);
390 }
391 
392 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
393 {
394 	struct ipoib_dev_priv *priv = cm_id->context;
395 	struct ipoib_cm_rx *p;
396 	unsigned psn;
397 	int ret;
398 
399 	ipoib_dbg(priv, "REQ arrived\n");
400 	p = kzalloc(sizeof *p, GFP_KERNEL);
401 	if (!p)
402 		return -ENOMEM;
403 	p->priv = priv;
404 	p->id = cm_id;
405 	cm_id->context = p;
406 	p->state = IPOIB_CM_RX_LIVE;
407 	p->jiffies = jiffies;
408 	INIT_LIST_HEAD(&p->list);
409 
410 	p->qp = ipoib_cm_create_rx_qp(priv, p);
411 	if (IS_ERR(p->qp)) {
412 		ret = PTR_ERR(p->qp);
413 		goto err_qp;
414 	}
415 
416 	psn = random() & 0xffffff;
417 	ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn);
418 	if (ret)
419 		goto err_modify;
420 
421 	if (!ipoib_cm_has_srq(priv)) {
422 		ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p);
423 		if (ret)
424 			goto err_modify;
425 	}
426 
427 	spin_lock_irq(&priv->lock);
428 	queue_delayed_work(ipoib_workqueue,
429 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
430 	/* Add this entry to passive ids list head, but do not re-add it
431 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
432 	p->jiffies = jiffies;
433 	if (p->state == IPOIB_CM_RX_LIVE)
434 		list_move(&p->list, &priv->cm.passive_ids);
435 	spin_unlock_irq(&priv->lock);
436 
437 	ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn);
438 	if (ret) {
439 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
440 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
441 			ipoib_warn(priv, "unable to move qp to error state\n");
442 	}
443 	return 0;
444 
445 err_modify:
446 	ib_destroy_qp(p->qp);
447 err_qp:
448 	kfree(p);
449 	return ret;
450 }
451 
452 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
453 			       struct ib_cm_event *event)
454 {
455 	struct ipoib_cm_rx *p;
456 	struct ipoib_dev_priv *priv;
457 
458 	switch (event->event) {
459 	case IB_CM_REQ_RECEIVED:
460 		return ipoib_cm_req_handler(cm_id, event);
461 	case IB_CM_DREQ_RECEIVED:
462 		p = cm_id->context;
463 		ib_send_cm_drep(cm_id, NULL, 0);
464 		/* Fall through */
465 	case IB_CM_REJ_RECEIVED:
466 		p = cm_id->context;
467 		priv = p->priv;
468 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
469 			ipoib_warn(priv, "unable to move qp to error state\n");
470 		/* Fall through */
471 	default:
472 		return 0;
473 	}
474 }
475 
476 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
477 {
478 	struct ipoib_cm_rx_buf saverx;
479 	struct ipoib_cm_rx_buf *rx_ring;
480 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
481 	if_t dev = priv->dev;
482 	struct mbuf *mb, *newmb;
483 	struct ipoib_cm_rx *p;
484 	int has_srq;
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 	mb->m_pkthdr.rcvif = dev;
560 
561 	M_PREPEND(mb, sizeof(struct ipoib_pseudoheader), M_NOWAIT);
562 	if (likely(mb != NULL)) {
563 		struct ipoib_header *ibh;
564 
565 		if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
566 		if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
567 
568 		/* fixup destination infiniband address */
569 		ibh = mtod(mb, struct ipoib_header *);
570 		memset(ibh->hwaddr, 0, 4);
571 		memcpy(ibh->hwaddr + 4, priv->local_gid.raw, sizeof(union ib_gid));
572 
573 		if_input(dev, mb);
574 	} else {
575 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
576 	}
577 repost:
578 	if (has_srq) {
579 		if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
580 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
581 				   "for buf %d\n", wr_id);
582 	} else {
583 		if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
584 							  &priv->cm.rx_wr,
585 							  priv->cm.rx_sge,
586 							  wr_id))) {
587 			--p->recv_count;
588 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
589 				   "for buf %d\n", wr_id);
590 		}
591 	}
592 done:
593 	return;
594 }
595 
596 static inline int post_send(struct ipoib_dev_priv *priv,
597 			    struct ipoib_cm_tx *tx,
598 			    struct ipoib_cm_tx_buf *tx_req,
599 			    unsigned int wr_id)
600 {
601 	const struct ib_send_wr *bad_wr;
602 	struct mbuf *mb = tx_req->mb;
603 	u64 *mapping = tx_req->mapping;
604 	struct mbuf *m;
605 	int i;
606 
607 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
608 		priv->tx_sge[i].addr = mapping[i];
609 		priv->tx_sge[i].length = m->m_len;
610 	}
611 	priv->tx_wr.wr.num_sge = i;
612 	priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
613 	priv->tx_wr.wr.opcode = IB_WR_SEND;
614 
615 	return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
616 }
617 
618 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
619 {
620 	struct ipoib_cm_tx_buf *tx_req;
621 	if_t dev = priv->dev;
622 
623 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) {
624 		while (ipoib_poll_tx(priv, false))
625 			;	/* nothing */
626 	}
627 
628 	m_adj(mb, sizeof(struct ipoib_pseudoheader));
629 	if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
630 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
631 			   mb->m_pkthdr.len, tx->mtu);
632 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
633 		ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
634 		return;
635 	}
636 
637 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
638 		       tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
639 
640 
641 	/*
642 	 * We put the mb into the tx_ring _before_ we call post_send()
643 	 * because it's entirely possible that the completion handler will
644 	 * run before we execute anything after the post_send().  That
645 	 * means we have to make sure everything is properly recorded and
646 	 * our state is consistent before we call post_send().
647 	 */
648 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
649 	tx_req->mb = mb;
650 	if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
651 	    priv->cm.num_frags))) {
652 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
653 		if (tx_req->mb)
654 			m_freem(tx_req->mb);
655 		return;
656 	}
657 
658 	if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
659 		ipoib_warn(priv, "post_send failed\n");
660 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
661 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
662 		m_freem(mb);
663 	} else {
664 		++tx->tx_head;
665 
666 		if (++priv->tx_outstanding == ipoib_sendq_size) {
667 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
668 				  tx->qp->qp_num);
669 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
670 				ipoib_warn(priv, "request notify on send CQ failed\n");
671 			if_setdrvflagbits(dev, IFF_DRV_OACTIVE, 0);
672 		}
673 	}
674 
675 }
676 
677 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
678 {
679 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
680 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
681 	if_t dev = priv->dev;
682 	struct ipoib_cm_tx_buf *tx_req;
683 
684 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
685 		       wr_id, wc->status);
686 
687 	if (unlikely(wr_id >= ipoib_sendq_size)) {
688 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
689 			   wr_id, ipoib_sendq_size);
690 		return;
691 	}
692 
693 	tx_req = &tx->tx_ring[wr_id];
694 
695 	ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
696 
697 	/* FIXME: is this right? Shouldn't we only increment on success? */
698 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
699 
700 	m_freem(tx_req->mb);
701 
702 	++tx->tx_tail;
703 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
704 	    (if_getdrvflags(dev) & IFF_DRV_OACTIVE) != 0 &&
705 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
706 		if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
707 
708 	if (wc->status != IB_WC_SUCCESS &&
709 	    wc->status != IB_WC_WR_FLUSH_ERR) {
710 		struct ipoib_path *path;
711 
712 		ipoib_dbg(priv, "failed cm send event "
713 			   "(status=%d, wrid=%d vend_err %x)\n",
714 			   wc->status, wr_id, wc->vendor_err);
715 
716 		path = tx->path;
717 
718 		if (path) {
719 			path->cm = NULL;
720 			rb_erase(&path->rb_node, &priv->path_tree);
721 			list_del(&path->list);
722 		}
723 
724 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
725 			list_move(&tx->list, &priv->cm.reap_list);
726 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
727 		}
728 
729 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
730 	}
731 
732 }
733 
734 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
735 {
736 	int ret;
737 
738 	if (!IPOIB_CM_SUPPORTED(if_getlladdr(priv->dev)))
739 		return 0;
740 
741 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
742 	if (IS_ERR(priv->cm.id)) {
743 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
744 		ret = PTR_ERR(priv->cm.id);
745 		goto err_cm;
746 	}
747 
748 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 0);
749 	if (ret) {
750 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
751 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
752 		goto err_listen;
753 	}
754 
755 	return 0;
756 
757 err_listen:
758 	ib_destroy_cm_id(priv->cm.id);
759 err_cm:
760 	priv->cm.id = NULL;
761 	return ret;
762 }
763 
764 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
765 {
766 	struct ipoib_cm_rx *rx, *n;
767 	LIST_HEAD(list);
768 
769 	spin_lock_irq(&priv->lock);
770 	list_splice_init(&priv->cm.rx_reap_list, &list);
771 	spin_unlock_irq(&priv->lock);
772 
773 	list_for_each_entry_safe(rx, n, &list, list) {
774 		ib_destroy_cm_id(rx->id);
775 		ib_destroy_qp(rx->qp);
776 		if (!ipoib_cm_has_srq(priv)) {
777 			ipoib_cm_free_rx_ring(priv, rx->rx_ring);
778 			spin_lock_irq(&priv->lock);
779 			--priv->cm.nonsrq_conn_qp;
780 			spin_unlock_irq(&priv->lock);
781 		}
782 		kfree(rx);
783 	}
784 }
785 
786 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
787 {
788 	struct ipoib_cm_rx *p;
789 	unsigned long begin;
790 	int ret;
791 
792 	if (!IPOIB_CM_SUPPORTED(if_getlladdr(priv->dev)) || !priv->cm.id)
793 		return;
794 
795 	ib_destroy_cm_id(priv->cm.id);
796 	priv->cm.id = NULL;
797 
798 	cancel_work_sync(&priv->cm.rx_reap_task);
799 
800 	spin_lock_irq(&priv->lock);
801 	while (!list_empty(&priv->cm.passive_ids)) {
802 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
803 		list_move(&p->list, &priv->cm.rx_error_list);
804 		p->state = IPOIB_CM_RX_ERROR;
805 		spin_unlock_irq(&priv->lock);
806 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
807 		if (ret)
808 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
809 		spin_lock_irq(&priv->lock);
810 	}
811 
812 	/* Wait for all RX to be drained */
813 	begin = jiffies;
814 
815 	while (!list_empty(&priv->cm.rx_error_list) ||
816 	       !list_empty(&priv->cm.rx_flush_list) ||
817 	       !list_empty(&priv->cm.rx_drain_list)) {
818 		if (time_after(jiffies, begin + 5 * HZ)) {
819 			ipoib_warn(priv, "RX drain timing out\n");
820 
821 			/*
822 			 * assume the HW is wedged and just free up everything.
823 			 */
824 			list_splice_init(&priv->cm.rx_flush_list,
825 					 &priv->cm.rx_reap_list);
826 			list_splice_init(&priv->cm.rx_error_list,
827 					 &priv->cm.rx_reap_list);
828 			list_splice_init(&priv->cm.rx_drain_list,
829 					 &priv->cm.rx_reap_list);
830 			break;
831 		}
832 		spin_unlock_irq(&priv->lock);
833 		msleep(1);
834 		ipoib_drain_cq(priv);
835 		spin_lock_irq(&priv->lock);
836 	}
837 
838 	spin_unlock_irq(&priv->lock);
839 
840 	ipoib_cm_free_rx_reap_list(priv);
841 
842 	cancel_delayed_work_sync(&priv->cm.stale_task);
843 }
844 
845 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
846 {
847 	struct ipoib_cm_tx *p = cm_id->context;
848 	struct ipoib_dev_priv *priv = p->priv;
849 	struct ipoib_cm_data *data = event->private_data;
850 	struct epoch_tracker et;
851 	struct ifqueue mbqueue;
852 	struct ib_qp_attr qp_attr;
853 	int qp_attr_mask, ret;
854 	struct mbuf *mb;
855 
856 	ipoib_dbg(priv, "cm rep handler\n");
857 	p->mtu = be32_to_cpu(data->mtu);
858 
859 	if (p->mtu <= IPOIB_ENCAP_LEN) {
860 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
861 			   p->mtu, IPOIB_ENCAP_LEN);
862 		return -EINVAL;
863 	}
864 
865 	qp_attr.qp_state = IB_QPS_RTR;
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 RTR: %d\n", ret);
869 		return ret;
870 	}
871 
872 	qp_attr.rq_psn = 0 /* FIXME */;
873 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
874 	if (ret) {
875 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
876 		return ret;
877 	}
878 
879 	qp_attr.qp_state = IB_QPS_RTS;
880 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
881 	if (ret) {
882 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
883 		return ret;
884 	}
885 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
886 	if (ret) {
887 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
888 		return ret;
889 	}
890 
891 	bzero(&mbqueue, sizeof(mbqueue));
892 
893 	spin_lock_irq(&priv->lock);
894 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
895 	if (p->path)
896 		for (;;) {
897 			_IF_DEQUEUE(&p->path->queue, mb);
898 			if (mb == NULL)
899 				break;
900 			_IF_ENQUEUE(&mbqueue, mb);
901 		}
902 	spin_unlock_irq(&priv->lock);
903 
904 	NET_EPOCH_ENTER(et);
905 	for (;;) {
906 		if_t dev = p->priv->dev;
907 		_IF_DEQUEUE(&mbqueue, mb);
908 		if (mb == NULL)
909 			break;
910 		mb->m_pkthdr.rcvif = dev;
911 		if (if_transmit(dev, mb))
912 			ipoib_warn(priv, "dev_queue_xmit failed "
913 				   "to requeue packet\n");
914 	}
915 	NET_EPOCH_EXIT(et);
916 
917 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
918 	if (ret) {
919 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
920 		return ret;
921 	}
922 	return 0;
923 }
924 
925 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
926     struct ipoib_cm_tx *tx)
927 {
928 	struct ib_qp_init_attr attr = {
929 		.send_cq		= priv->send_cq,
930 		.recv_cq		= priv->recv_cq,
931 		.srq			= priv->cm.srq,
932 		.cap.max_send_wr	= ipoib_sendq_size,
933 		.cap.max_send_sge	= priv->cm.num_frags,
934 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
935 		.qp_type		= IB_QPT_RC,
936 		.qp_context		= tx
937 	};
938 
939 	return ib_create_qp(priv->pd, &attr);
940 }
941 
942 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
943 			     struct ib_cm_id *id, struct ib_qp *qp,
944 			     u32 qpn,
945 			     struct ib_sa_path_rec *pathrec)
946 {
947 	struct ipoib_cm_data data = {};
948 	struct ib_cm_req_param req = {};
949 
950 	ipoib_dbg(priv, "cm send req\n");
951 
952 	data.qpn = cpu_to_be32(priv->qp->qp_num);
953 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
954 
955 	req.primary_path		= pathrec;
956 	req.alternate_path		= NULL;
957 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
958 	req.qp_num			= qp->qp_num;
959 	req.qp_type			= qp->qp_type;
960 	req.private_data		= &data;
961 	req.private_data_len		= sizeof data;
962 	req.flow_control		= 0;
963 
964 	req.starting_psn		= 0; /* FIXME */
965 
966 	/*
967 	 * Pick some arbitrary defaults here; we could make these
968 	 * module parameters if anyone cared about setting them.
969 	 */
970 	req.responder_resources		= 4;
971 	req.remote_cm_response_timeout	= 20;
972 	req.local_cm_response_timeout	= 20;
973 	req.retry_count			= 0; /* RFC draft warns against retries */
974 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
975 	req.max_cm_retries		= 15;
976 	req.srq				= ipoib_cm_has_srq(priv);
977 	return ib_send_cm_req(id, &req);
978 }
979 
980 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
981 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
982 {
983 	struct ib_qp_attr qp_attr;
984 	int qp_attr_mask, ret;
985 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
986 	if (ret) {
987 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
988 		return ret;
989 	}
990 
991 	qp_attr.qp_state = IB_QPS_INIT;
992 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
993 	qp_attr.port_num = priv->port;
994 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
995 
996 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
997 	if (ret) {
998 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
999 		return ret;
1000 	}
1001 	return 0;
1002 }
1003 
1004 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1005 			    struct ib_sa_path_rec *pathrec)
1006 {
1007 	struct ipoib_dev_priv *priv = p->priv;
1008 	int ret;
1009 
1010 	p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
1011 	if (!p->tx_ring) {
1012 		ipoib_warn(priv, "failed to allocate tx ring\n");
1013 		ret = -ENOMEM;
1014 		goto err_tx;
1015 	}
1016 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1017 
1018 	p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1019 	if (IS_ERR(p->qp)) {
1020 		ret = PTR_ERR(p->qp);
1021 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1022 		goto err_qp;
1023 	}
1024 
1025 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1026 	if (IS_ERR(p->id)) {
1027 		ret = PTR_ERR(p->id);
1028 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1029 		goto err_id;
1030 	}
1031 
1032 	ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1033 	if (ret) {
1034 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1035 		goto err_modify;
1036 	}
1037 
1038 	ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1039 	if (ret) {
1040 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1041 		goto err_send_cm;
1042 	}
1043 
1044 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1045 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1046 
1047 	return 0;
1048 
1049 err_send_cm:
1050 err_modify:
1051 	ib_destroy_cm_id(p->id);
1052 err_id:
1053 	p->id = NULL;
1054 	ib_destroy_qp(p->qp);
1055 err_qp:
1056 	p->qp = NULL;
1057 	kfree(p->tx_ring);
1058 err_tx:
1059 	return ret;
1060 }
1061 
1062 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1063 {
1064 	struct ipoib_dev_priv *priv = p->priv;
1065 	if_t dev = priv->dev;
1066 	struct ipoib_cm_tx_buf *tx_req;
1067 	unsigned long begin;
1068 
1069 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1070 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1071 
1072 	if (p->path)
1073 		ipoib_path_free(priv, p->path);
1074 
1075 	if (p->id)
1076 		ib_destroy_cm_id(p->id);
1077 
1078 	if (p->tx_ring) {
1079 		/* Wait for all sends to complete */
1080 		begin = jiffies;
1081 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1082 			if (time_after(jiffies, begin + 5 * HZ)) {
1083 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1084 					   p->tx_head - p->tx_tail);
1085 				goto timeout;
1086 			}
1087 
1088 			msleep(1);
1089 		}
1090 	}
1091 
1092 timeout:
1093 
1094 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1095 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1096 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1097 		m_freem(tx_req->mb);
1098 		++p->tx_tail;
1099 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1100 		    (if_getdrvflags(dev) & IFF_DRV_OACTIVE) != 0 &&
1101 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1102 			if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
1103 	}
1104 
1105 	if (p->qp)
1106 		ib_destroy_qp(p->qp);
1107 
1108 	kfree(p->tx_ring);
1109 	kfree(p);
1110 }
1111 
1112 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1113 			       struct ib_cm_event *event)
1114 {
1115 	struct ipoib_cm_tx *tx = cm_id->context;
1116 	struct ipoib_dev_priv *priv = tx->priv;
1117 	struct ipoib_path *path;
1118 	unsigned long flags;
1119 	int ret;
1120 
1121 	switch (event->event) {
1122 	case IB_CM_DREQ_RECEIVED:
1123 		ipoib_dbg(priv, "DREQ received.\n");
1124 		ib_send_cm_drep(cm_id, NULL, 0);
1125 		break;
1126 	case IB_CM_REP_RECEIVED:
1127 		ipoib_dbg(priv, "REP received.\n");
1128 		ret = ipoib_cm_rep_handler(cm_id, event);
1129 		if (ret)
1130 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1131 				       NULL, 0, NULL, 0);
1132 		break;
1133 	case IB_CM_REQ_ERROR:
1134 	case IB_CM_REJ_RECEIVED:
1135 	case IB_CM_TIMEWAIT_EXIT:
1136 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1137 		spin_lock_irqsave(&priv->lock, flags);
1138 		path = tx->path;
1139 
1140 		if (path) {
1141 			path->cm = NULL;
1142 			tx->path = NULL;
1143 			rb_erase(&path->rb_node, &priv->path_tree);
1144 			list_del(&path->list);
1145 		}
1146 
1147 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1148 			list_move(&tx->list, &priv->cm.reap_list);
1149 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1150 		}
1151 
1152 		spin_unlock_irqrestore(&priv->lock, flags);
1153 		if (path)
1154 			ipoib_path_free(tx->priv, path);
1155 		break;
1156 	default:
1157 		break;
1158 	}
1159 
1160 	return 0;
1161 }
1162 
1163 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1164     struct ipoib_path *path)
1165 {
1166 	struct ipoib_cm_tx *tx;
1167 
1168 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1169 	if (!tx)
1170 		return NULL;
1171 
1172 	ipoib_dbg(priv, "Creating cm tx\n");
1173 	path->cm = tx;
1174 	tx->path = path;
1175 	tx->priv = priv;
1176 	list_add(&tx->list, &priv->cm.start_list);
1177 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1178 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1179 	return tx;
1180 }
1181 
1182 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1183 {
1184 	struct ipoib_dev_priv *priv = tx->priv;
1185 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1186 		spin_lock(&priv->lock);
1187 		list_move(&tx->list, &priv->cm.reap_list);
1188 		spin_unlock(&priv->lock);
1189 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1190 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1191 			  tx->path->pathrec.dgid.raw);
1192 		tx->path = NULL;
1193 	}
1194 }
1195 
1196 static void ipoib_cm_tx_start(struct work_struct *work)
1197 {
1198 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1199 						   cm.start_task);
1200 	struct ipoib_path *path;
1201 	struct ipoib_cm_tx *p;
1202 	unsigned long flags;
1203 	int ret;
1204 
1205 	struct ib_sa_path_rec pathrec;
1206 	u32 qpn;
1207 
1208 	ipoib_dbg(priv, "cm start task\n");
1209 	spin_lock_irqsave(&priv->lock, flags);
1210 
1211 	while (!list_empty(&priv->cm.start_list)) {
1212 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1213 		list_del_init(&p->list);
1214 		path = p->path;
1215 		qpn = IPOIB_QPN(path->hwaddr);
1216 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1217 
1218 		spin_unlock_irqrestore(&priv->lock, flags);
1219 
1220 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1221 
1222 		spin_lock_irqsave(&priv->lock, flags);
1223 
1224 		if (ret) {
1225 			path = p->path;
1226 			if (path) {
1227 				path->cm = NULL;
1228 				rb_erase(&path->rb_node, &priv->path_tree);
1229 				list_del(&path->list);
1230 				ipoib_path_free(priv, path);
1231 			}
1232 			list_del(&p->list);
1233 			kfree(p);
1234 		}
1235 	}
1236 
1237 	spin_unlock_irqrestore(&priv->lock, flags);
1238 }
1239 
1240 static void ipoib_cm_tx_reap(struct work_struct *work)
1241 {
1242 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1243 						   cm.reap_task);
1244 	struct ipoib_cm_tx *p;
1245 	unsigned long flags;
1246 
1247 	spin_lock_irqsave(&priv->lock, flags);
1248 
1249 	while (!list_empty(&priv->cm.reap_list)) {
1250 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1251 		list_del(&p->list);
1252 		spin_unlock_irqrestore(&priv->lock, flags);
1253 		ipoib_cm_tx_destroy(p);
1254 		spin_lock_irqsave(&priv->lock, flags);
1255 	}
1256 
1257 	spin_unlock_irqrestore(&priv->lock, flags);
1258 }
1259 
1260 static void ipoib_cm_mb_reap(struct work_struct *work)
1261 {
1262 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1263 						   cm.mb_task);
1264 	struct mbuf *mb;
1265 	unsigned long flags;
1266 #if defined(INET) || defined(INET6)
1267 	unsigned mtu = priv->mcast_mtu;
1268 #endif
1269 	uint16_t proto;
1270 
1271 	spin_lock_irqsave(&priv->lock, flags);
1272 
1273 	CURVNET_SET_QUIET(if_getvnet(priv->dev));
1274 
1275 	for (;;) {
1276 		IF_DEQUEUE(&priv->cm.mb_queue, mb);
1277 		if (mb == NULL)
1278 			break;
1279 		spin_unlock_irqrestore(&priv->lock, flags);
1280 
1281 		proto = htons(*mtod(mb, uint16_t *));
1282 		m_adj(mb, IPOIB_ENCAP_LEN);
1283 		switch (proto) {
1284 #if defined(INET)
1285 		case ETHERTYPE_IP:
1286 			icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1287 			break;
1288 #endif
1289 #if defined(INET6)
1290 		case ETHERTYPE_IPV6:
1291 			icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1292 			break;
1293 #endif
1294 		default:
1295 			m_freem(mb);
1296 		}
1297 
1298 		spin_lock_irqsave(&priv->lock, flags);
1299 	}
1300 
1301 	CURVNET_RESTORE();
1302 
1303 	spin_unlock_irqrestore(&priv->lock, flags);
1304 }
1305 
1306 void
1307 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1308 {
1309 	int e = priv->cm.mb_queue.ifq_len;
1310 
1311 	IF_ENQUEUE(&priv->cm.mb_queue, mb);
1312 	if (e == 0)
1313 		queue_work(ipoib_workqueue, &priv->cm.mb_task);
1314 }
1315 
1316 static void ipoib_cm_rx_reap(struct work_struct *work)
1317 {
1318 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1319 						cm.rx_reap_task));
1320 }
1321 
1322 static void ipoib_cm_stale_task(struct work_struct *work)
1323 {
1324 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1325 						   cm.stale_task.work);
1326 	struct ipoib_cm_rx *p;
1327 	int ret;
1328 
1329 	spin_lock_irq(&priv->lock);
1330 	while (!list_empty(&priv->cm.passive_ids)) {
1331 		/* List is sorted by LRU, start from tail,
1332 		 * stop when we see a recently used entry */
1333 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1334 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1335 			break;
1336 		list_move(&p->list, &priv->cm.rx_error_list);
1337 		p->state = IPOIB_CM_RX_ERROR;
1338 		spin_unlock_irq(&priv->lock);
1339 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1340 		if (ret)
1341 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1342 		spin_lock_irq(&priv->lock);
1343 	}
1344 
1345 	if (!list_empty(&priv->cm.passive_ids))
1346 		queue_delayed_work(ipoib_workqueue,
1347 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1348 	spin_unlock_irq(&priv->lock);
1349 }
1350 
1351 
1352 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1353 {
1354 	struct ib_srq_init_attr srq_init_attr = {
1355 		.attr = {
1356 			.max_wr  = ipoib_recvq_size,
1357 			.max_sge = max_sge
1358 		}
1359 	};
1360 
1361 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1362 	if (IS_ERR(priv->cm.srq)) {
1363 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1364 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1365 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1366 		priv->cm.srq = NULL;
1367 		return;
1368 	}
1369 
1370 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1371 	if (!priv->cm.srq_ring) {
1372 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1373 		       priv->ca->name, ipoib_recvq_size);
1374 		ib_destroy_srq(priv->cm.srq);
1375 		priv->cm.srq = NULL;
1376 		return;
1377 	}
1378 
1379 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1380 }
1381 
1382 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1383 {
1384 	if_t dev = priv->dev;
1385 	int i;
1386 	int max_srq_sge;
1387 
1388 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1389 	INIT_LIST_HEAD(&priv->cm.reap_list);
1390 	INIT_LIST_HEAD(&priv->cm.start_list);
1391 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1392 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1393 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1394 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1395 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1396 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1397 	INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1398 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1399 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1400 
1401 	bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1402 	mtx_init(&priv->cm.mb_queue.ifq_mtx,
1403 	    if_name(dev), "if send queue", MTX_DEF);
1404 
1405 	max_srq_sge = priv->ca->attrs.max_srq_sge;
1406 
1407 	ipoib_dbg(priv, "max_srq_sge=%d\n", max_srq_sge);
1408 
1409 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, max_srq_sge);
1410 	ipoib_cm_create_srq(priv, max_srq_sge);
1411 	if (ipoib_cm_has_srq(priv)) {
1412 		priv->cm.max_cm_mtu = max_srq_sge * MJUMPAGESIZE;
1413 		priv->cm.num_frags  = max_srq_sge;
1414 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1415 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1416 	} else {
1417 		priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1418 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1419 	}
1420 
1421 	ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1422 
1423 	if (ipoib_cm_has_srq(priv)) {
1424 		for (i = 0; i < ipoib_recvq_size; ++i) {
1425 			if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1426 				ipoib_warn(priv, "failed to allocate "
1427 					   "receive buffer %d\n", i);
1428 				ipoib_cm_dev_cleanup(priv);
1429 				return -ENOMEM;
1430 			}
1431 
1432 			if (ipoib_cm_post_receive_srq(priv, i)) {
1433 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1434 					   "failed for buf %d\n", i);
1435 				ipoib_cm_dev_cleanup(priv);
1436 				return -EIO;
1437 			}
1438 		}
1439 	}
1440 
1441 	if_getlladdr(priv->dev)[0] = IPOIB_FLAGS_RC;
1442 	return 0;
1443 }
1444 
1445 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1446 {
1447 	int ret;
1448 
1449 	if (!priv->cm.srq)
1450 		return;
1451 
1452 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1453 
1454 	ret = ib_destroy_srq(priv->cm.srq);
1455 	if (ret)
1456 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1457 
1458 	priv->cm.srq = NULL;
1459 	if (!priv->cm.srq_ring)
1460 		return;
1461 
1462 	ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1463 	priv->cm.srq_ring = NULL;
1464 
1465 	mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1466 }
1467 
1468 #endif /* CONFIG_INFINIBAND_IPOIB_CM */
1469