xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c (revision d06955f9bdb1416d9196043ed781f9b36dae9adc)
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 "ipoib.h"
36 
37 #ifdef CONFIG_INFINIBAND_IPOIB_CM
38 
39 #include <netinet/ip.h>
40 #include <netinet/ip_icmp.h>
41 #include <netinet/icmp6.h>
42 
43 #include <rdma/ib_cm.h>
44 #include <rdma/ib_cache.h>
45 #include <linux/delay.h>
46 
47 int ipoib_max_conn_qp = 128;
48 
49 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
50 MODULE_PARM_DESC(max_nonsrq_conn_qp,
51 		 "Max number of connected-mode QPs per interface "
52 		 "(applied only if shared receive queue is not available)");
53 
54 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
55 static int data_debug_level;
56 
57 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
58 MODULE_PARM_DESC(cm_data_debug_level,
59 		 "Enable data path debug tracing for connected mode if > 0");
60 #endif
61 
62 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
63 
64 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
65 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
66 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
67 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
68 
69 static struct ib_qp_attr ipoib_cm_err_attr = {
70 	.qp_state = IB_QPS_ERR
71 };
72 
73 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
74 
75 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
76 	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
77 	.opcode = IB_WR_SEND,
78 };
79 
80 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
81 			       struct ib_cm_event *event);
82 
83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
84 {
85 
86 	ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req);
87 
88 }
89 
90 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id)
91 {
92 	struct ib_recv_wr *bad_wr;
93 	struct ipoib_rx_buf *rx_req;
94 	struct mbuf *m;
95 	int ret;
96 	int i;
97 
98 	rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id];
99 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
100 		priv->cm.rx_sge[i].addr = rx_req->mapping[i];
101 		priv->cm.rx_sge[i].length = m->m_len;
102 	}
103 
104 	priv->cm.rx_wr.num_sge = i;
105 	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
106 
107 	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
108 	if (unlikely(ret)) {
109 		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
110 		ipoib_dma_unmap_rx(priv, rx_req);
111 		m_freem(priv->cm.srq_ring[id].mb);
112 		priv->cm.srq_ring[id].mb = NULL;
113 	}
114 
115 	return ret;
116 }
117 
118 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv,
119 					struct ipoib_cm_rx *rx,
120 					struct ib_recv_wr *wr,
121 					struct ib_sge *sge, int id)
122 {
123 	struct ipoib_rx_buf *rx_req;
124 	struct ib_recv_wr *bad_wr;
125 	struct mbuf *m;
126 	int ret;
127 	int i;
128 
129 	rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id];
130 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
131 		sge[i].addr = rx_req->mapping[i];
132 		sge[i].length = m->m_len;
133 	}
134 
135 	wr->num_sge = i;
136 	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
137 
138 	ret = ib_post_recv(rx->qp, wr, &bad_wr);
139 	if (unlikely(ret)) {
140 		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
141 		ipoib_dma_unmap_rx(priv, rx_req);
142 		m_freem(rx->rx_ring[id].mb);
143 		rx->rx_ring[id].mb = NULL;
144 	}
145 
146 	return ret;
147 }
148 
149 static struct mbuf *
150 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
151 {
152 	return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req,
153 	    priv->cm.max_cm_mtu);
154 }
155 
156 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv,
157 				  struct ipoib_cm_rx_buf *rx_ring)
158 {
159 	int i;
160 
161 	for (i = 0; i < ipoib_recvq_size; ++i)
162 		if (rx_ring[i].mb) {
163 			ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]);
164 			m_freem(rx_ring[i].mb);
165 		}
166 
167 	kfree(rx_ring);
168 }
169 
170 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
171 {
172 	struct ib_send_wr *bad_wr;
173 	struct ipoib_cm_rx *p;
174 
175 	/* We only reserved 1 extra slot in CQ for drain WRs, so
176 	 * make sure we have at most 1 outstanding WR. */
177 	if (list_empty(&priv->cm.rx_flush_list) ||
178 	    !list_empty(&priv->cm.rx_drain_list))
179 		return;
180 
181 	/*
182 	 * QPs on flush list are error state.  This way, a "flush
183 	 * error" WC will be immediately generated for each WR we post.
184 	 */
185 	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
186 	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
187 		ipoib_warn(priv, "failed to post drain wr\n");
188 
189 	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
190 }
191 
192 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
193 {
194 	struct ipoib_cm_rx *p = ctx;
195 	struct ipoib_dev_priv *priv = p->priv;
196 	unsigned long flags;
197 
198 	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
199 		return;
200 
201 	spin_lock_irqsave(&priv->lock, flags);
202 	list_move(&p->list, &priv->cm.rx_flush_list);
203 	p->state = IPOIB_CM_RX_FLUSH;
204 	ipoib_cm_start_rx_drain(priv);
205 	spin_unlock_irqrestore(&priv->lock, flags);
206 }
207 
208 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv,
209 					   struct ipoib_cm_rx *p)
210 {
211 	struct ib_qp_init_attr attr = {
212 		.event_handler = ipoib_cm_rx_event_handler,
213 		.send_cq = priv->recv_cq, /* For drain WR */
214 		.recv_cq = priv->recv_cq,
215 		.srq = priv->cm.srq,
216 		.cap.max_send_wr = 1, /* For drain WR */
217 		.cap.max_send_sge = 1,
218 		.sq_sig_type = IB_SIGNAL_ALL_WR,
219 		.qp_type = IB_QPT_RC,
220 		.qp_context = p,
221 	};
222 
223 	if (!ipoib_cm_has_srq(priv)) {
224 		attr.cap.max_recv_wr  = ipoib_recvq_size;
225 		attr.cap.max_recv_sge = priv->cm.num_frags;
226 	}
227 
228 	return ib_create_qp(priv->pd, &attr);
229 }
230 
231 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv,
232 				 struct ib_cm_id *cm_id, struct ib_qp *qp,
233 				 unsigned psn)
234 {
235 	struct ib_qp_attr qp_attr;
236 	int qp_attr_mask, ret;
237 
238 	qp_attr.qp_state = IB_QPS_INIT;
239 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
240 	if (ret) {
241 		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
242 		return ret;
243 	}
244 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
245 	if (ret) {
246 		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
247 		return ret;
248 	}
249 	qp_attr.qp_state = IB_QPS_RTR;
250 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
251 	if (ret) {
252 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
253 		return ret;
254 	}
255 	qp_attr.rq_psn = psn;
256 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
257 	if (ret) {
258 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
259 		return ret;
260 	}
261 
262 	/*
263 	 * Current Mellanox HCA firmware won't generate completions
264 	 * with error for drain WRs unless the QP has been moved to
265 	 * RTS first. This work-around leaves a window where a QP has
266 	 * moved to error asynchronously, but this will eventually get
267 	 * fixed in firmware, so let's not error out if modify QP
268 	 * fails.
269 	 */
270 	qp_attr.qp_state = IB_QPS_RTS;
271 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
272 	if (ret) {
273 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
274 		return 0;
275 	}
276 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
277 	if (ret) {
278 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
279 		return 0;
280 	}
281 
282 	return 0;
283 }
284 
285 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv,
286 				struct ib_recv_wr *wr,
287 				struct ib_sge *sge)
288 {
289 	int i;
290 
291 	for (i = 0; i < IPOIB_CM_RX_SG; i++)
292 		sge[i].lkey = priv->pd->local_dma_lkey;
293 
294 	wr->next    = NULL;
295 	wr->sg_list = sge;
296 	wr->num_sge = 1;
297 }
298 
299 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv,
300     struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx)
301 {
302 	struct {
303 		struct ib_recv_wr wr;
304 		struct ib_sge sge[IPOIB_CM_RX_SG];
305 	} *t;
306 	int ret;
307 	int i;
308 
309 	rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL);
310 	if (!rx->rx_ring) {
311 		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
312 		       priv->ca->name, ipoib_recvq_size);
313 		return -ENOMEM;
314 	}
315 
316 	memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
317 
318 	t = kmalloc(sizeof *t, GFP_KERNEL);
319 	if (!t) {
320 		ret = -ENOMEM;
321 		goto err_free;
322 	}
323 
324 	ipoib_cm_init_rx_wr(priv, &t->wr, t->sge);
325 
326 	spin_lock_irq(&priv->lock);
327 
328 	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
329 		spin_unlock_irq(&priv->lock);
330 		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
331 		ret = -EINVAL;
332 		goto err_free;
333 	} else
334 		++priv->cm.nonsrq_conn_qp;
335 
336 	spin_unlock_irq(&priv->lock);
337 
338 	for (i = 0; i < ipoib_recvq_size; ++i) {
339 		if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) {
340 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
341 				ret = -ENOMEM;
342 				goto err_count;
343 		}
344 		ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i);
345 		if (ret) {
346 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
347 				   "failed for buf %d\n", i);
348 			ret = -EIO;
349 			goto err_count;
350 		}
351 	}
352 
353 	rx->recv_count = ipoib_recvq_size;
354 
355 	kfree(t);
356 
357 	return 0;
358 
359 err_count:
360 	spin_lock_irq(&priv->lock);
361 	--priv->cm.nonsrq_conn_qp;
362 	spin_unlock_irq(&priv->lock);
363 
364 err_free:
365 	kfree(t);
366 	ipoib_cm_free_rx_ring(priv, rx->rx_ring);
367 
368 	return ret;
369 }
370 
371 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id,
372 			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
373 			     unsigned psn)
374 {
375 	struct ipoib_cm_data data = {};
376 	struct ib_cm_rep_param rep = {};
377 
378 	data.qpn = cpu_to_be32(priv->qp->qp_num);
379 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
380 
381 	rep.private_data = &data;
382 	rep.private_data_len = sizeof data;
383 	rep.flow_control = 0;
384 	rep.rnr_retry_count = req->rnr_retry_count;
385 	rep.srq = ipoib_cm_has_srq(priv);
386 	rep.qp_num = qp->qp_num;
387 	rep.starting_psn = psn;
388 	return ib_send_cm_rep(cm_id, &rep);
389 }
390 
391 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
392 {
393 	struct ipoib_dev_priv *priv = cm_id->context;
394 	struct ipoib_cm_rx *p;
395 	unsigned psn;
396 	int ret;
397 
398 	ipoib_dbg(priv, "REQ arrived\n");
399 	p = kzalloc(sizeof *p, GFP_KERNEL);
400 	if (!p)
401 		return -ENOMEM;
402 	p->priv = priv;
403 	p->id = cm_id;
404 	cm_id->context = p;
405 	p->state = IPOIB_CM_RX_LIVE;
406 	p->jiffies = jiffies;
407 	INIT_LIST_HEAD(&p->list);
408 
409 	p->qp = ipoib_cm_create_rx_qp(priv, p);
410 	if (IS_ERR(p->qp)) {
411 		ret = PTR_ERR(p->qp);
412 		goto err_qp;
413 	}
414 
415 	psn = random() & 0xffffff;
416 	ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn);
417 	if (ret)
418 		goto err_modify;
419 
420 	if (!ipoib_cm_has_srq(priv)) {
421 		ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p);
422 		if (ret)
423 			goto err_modify;
424 	}
425 
426 	spin_lock_irq(&priv->lock);
427 	queue_delayed_work(ipoib_workqueue,
428 			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
429 	/* Add this entry to passive ids list head, but do not re-add it
430 	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
431 	p->jiffies = jiffies;
432 	if (p->state == IPOIB_CM_RX_LIVE)
433 		list_move(&p->list, &priv->cm.passive_ids);
434 	spin_unlock_irq(&priv->lock);
435 
436 	ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn);
437 	if (ret) {
438 		ipoib_warn(priv, "failed to send REP: %d\n", ret);
439 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
440 			ipoib_warn(priv, "unable to move qp to error state\n");
441 	}
442 	return 0;
443 
444 err_modify:
445 	ib_destroy_qp(p->qp);
446 err_qp:
447 	kfree(p);
448 	return ret;
449 }
450 
451 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
452 			       struct ib_cm_event *event)
453 {
454 	struct ipoib_cm_rx *p;
455 	struct ipoib_dev_priv *priv;
456 
457 	switch (event->event) {
458 	case IB_CM_REQ_RECEIVED:
459 		return ipoib_cm_req_handler(cm_id, event);
460 	case IB_CM_DREQ_RECEIVED:
461 		p = cm_id->context;
462 		ib_send_cm_drep(cm_id, NULL, 0);
463 		/* Fall through */
464 	case IB_CM_REJ_RECEIVED:
465 		p = cm_id->context;
466 		priv = p->priv;
467 		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
468 			ipoib_warn(priv, "unable to move qp to error state\n");
469 		/* Fall through */
470 	default:
471 		return 0;
472 	}
473 }
474 
475 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
476 {
477 	struct ipoib_cm_rx_buf saverx;
478 	struct ipoib_cm_rx_buf *rx_ring;
479 	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
480 	struct ifnet *dev = priv->dev;
481 	struct mbuf *mb, *newmb;
482 	struct ipoib_cm_rx *p;
483 	int has_srq;
484 	u_short proto;
485 
486 	CURVNET_SET_QUIET(dev->if_vnet);
487 
488 	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
489 		       wr_id, wc->status);
490 
491 	if (unlikely(wr_id >= ipoib_recvq_size)) {
492 		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
493 			spin_lock(&priv->lock);
494 			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
495 			ipoib_cm_start_rx_drain(priv);
496 			if (priv->cm.id != NULL)
497 				queue_work(ipoib_workqueue,
498 				    &priv->cm.rx_reap_task);
499 			spin_unlock(&priv->lock);
500 		} else
501 			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
502 				   wr_id, ipoib_recvq_size);
503 		goto done;
504 	}
505 
506 	p = wc->qp->qp_context;
507 
508 	has_srq = ipoib_cm_has_srq(priv);
509 	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
510 
511 	mb = rx_ring[wr_id].mb;
512 
513 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
514 		ipoib_dbg(priv, "cm recv error "
515 			   "(status=%d, wrid=%d vend_err %x)\n",
516 			   wc->status, wr_id, wc->vendor_err);
517 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
518 		if (has_srq)
519 			goto repost;
520 		else {
521 			if (!--p->recv_count) {
522 				spin_lock(&priv->lock);
523 				list_move(&p->list, &priv->cm.rx_reap_list);
524 				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
525 				spin_unlock(&priv->lock);
526 			}
527 			goto done;
528 		}
529 	}
530 
531 	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
532 		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
533 			p->jiffies = jiffies;
534 			/* Move this entry to list head, but do not re-add it
535 			 * if it has been moved out of list. */
536 			if (p->state == IPOIB_CM_RX_LIVE)
537 				list_move(&p->list, &priv->cm.passive_ids);
538 		}
539 	}
540 
541 	memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx));
542 	newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]);
543 	if (unlikely(!newmb)) {
544 		/*
545 		 * If we can't allocate a new RX buffer, dump
546 		 * this packet and reuse the old buffer.
547 		 */
548 		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
549 		if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
550 		memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx));
551 		goto repost;
552 	}
553 
554 	ipoib_cm_dma_unmap_rx(priv, &saverx);
555 
556 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
557 		       wc->byte_len, wc->slid);
558 
559 	ipoib_dma_mb(priv, mb, wc->byte_len);
560 
561 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
562 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
563 
564 	mb->m_pkthdr.rcvif = dev;
565 	proto = *mtod(mb, uint16_t *);
566 	m_adj(mb, IPOIB_ENCAP_LEN);
567 
568 	IPOIB_MTAP_PROTO(dev, mb, proto);
569 	ipoib_demux(dev, mb, ntohs(proto));
570 
571 repost:
572 	if (has_srq) {
573 		if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
574 			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
575 				   "for buf %d\n", wr_id);
576 	} else {
577 		if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
578 							  &priv->cm.rx_wr,
579 							  priv->cm.rx_sge,
580 							  wr_id))) {
581 			--p->recv_count;
582 			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
583 				   "for buf %d\n", wr_id);
584 		}
585 	}
586 done:
587 	CURVNET_RESTORE();
588 	return;
589 }
590 
591 static inline int post_send(struct ipoib_dev_priv *priv,
592 			    struct ipoib_cm_tx *tx,
593 			    struct ipoib_cm_tx_buf *tx_req,
594 			    unsigned int wr_id)
595 {
596 	struct ib_send_wr *bad_wr;
597 	struct mbuf *mb = tx_req->mb;
598 	u64 *mapping = tx_req->mapping;
599 	struct mbuf *m;
600 	int i;
601 
602 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
603 		priv->tx_sge[i].addr = mapping[i];
604 		priv->tx_sge[i].length = m->m_len;
605 	}
606 	priv->tx_wr.wr.num_sge = i;
607 	priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
608 	priv->tx_wr.wr.opcode = IB_WR_SEND;
609 
610 	return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
611 }
612 
613 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
614 {
615 	struct ipoib_cm_tx_buf *tx_req;
616 	struct ifnet *dev = priv->dev;
617 
618 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
619 		while (ipoib_poll_tx(priv)); /* nothing */
620 
621 	m_adj(mb, sizeof(struct ipoib_pseudoheader));
622 	if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
623 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
624 			   mb->m_pkthdr.len, tx->mtu);
625 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
626 		ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
627 		return;
628 	}
629 
630 	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
631 		       tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
632 
633 
634 	/*
635 	 * We put the mb into the tx_ring _before_ we call post_send()
636 	 * because it's entirely possible that the completion handler will
637 	 * run before we execute anything after the post_send().  That
638 	 * means we have to make sure everything is properly recorded and
639 	 * our state is consistent before we call post_send().
640 	 */
641 	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
642 	tx_req->mb = mb;
643 	if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
644 	    priv->cm.num_frags))) {
645 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
646 		if (tx_req->mb)
647 			m_freem(tx_req->mb);
648 		return;
649 	}
650 
651 	if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
652 		ipoib_warn(priv, "post_send failed\n");
653 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
654 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
655 		m_freem(mb);
656 	} else {
657 		++tx->tx_head;
658 
659 		if (++priv->tx_outstanding == ipoib_sendq_size) {
660 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
661 				  tx->qp->qp_num);
662 			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
663 				ipoib_warn(priv, "request notify on send CQ failed\n");
664 			dev->if_drv_flags |= IFF_DRV_OACTIVE;
665 		}
666 	}
667 
668 }
669 
670 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
671 {
672 	struct ipoib_cm_tx *tx = wc->qp->qp_context;
673 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
674 	struct ifnet *dev = priv->dev;
675 	struct ipoib_cm_tx_buf *tx_req;
676 
677 	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
678 		       wr_id, wc->status);
679 
680 	if (unlikely(wr_id >= ipoib_sendq_size)) {
681 		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
682 			   wr_id, ipoib_sendq_size);
683 		return;
684 	}
685 
686 	tx_req = &tx->tx_ring[wr_id];
687 
688 	ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
689 
690 	/* FIXME: is this right? Shouldn't we only increment on success? */
691 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
692 
693 	m_freem(tx_req->mb);
694 
695 	++tx->tx_tail;
696 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
697 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
698 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
699 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
700 
701 	if (wc->status != IB_WC_SUCCESS &&
702 	    wc->status != IB_WC_WR_FLUSH_ERR) {
703 		struct ipoib_path *path;
704 
705 		ipoib_dbg(priv, "failed cm send event "
706 			   "(status=%d, wrid=%d vend_err %x)\n",
707 			   wc->status, wr_id, wc->vendor_err);
708 
709 		path = tx->path;
710 
711 		if (path) {
712 			path->cm = NULL;
713 			rb_erase(&path->rb_node, &priv->path_tree);
714 			list_del(&path->list);
715 		}
716 
717 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
718 			list_move(&tx->list, &priv->cm.reap_list);
719 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
720 		}
721 
722 		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
723 	}
724 
725 }
726 
727 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
728 {
729 	int ret;
730 
731 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)))
732 		return 0;
733 
734 	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
735 	if (IS_ERR(priv->cm.id)) {
736 		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
737 		ret = PTR_ERR(priv->cm.id);
738 		goto err_cm;
739 	}
740 
741 	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 0);
742 	if (ret) {
743 		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
744 		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
745 		goto err_listen;
746 	}
747 
748 	return 0;
749 
750 err_listen:
751 	ib_destroy_cm_id(priv->cm.id);
752 err_cm:
753 	priv->cm.id = NULL;
754 	return ret;
755 }
756 
757 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
758 {
759 	struct ipoib_cm_rx *rx, *n;
760 	LIST_HEAD(list);
761 
762 	spin_lock_irq(&priv->lock);
763 	list_splice_init(&priv->cm.rx_reap_list, &list);
764 	spin_unlock_irq(&priv->lock);
765 
766 	list_for_each_entry_safe(rx, n, &list, list) {
767 		ib_destroy_cm_id(rx->id);
768 		ib_destroy_qp(rx->qp);
769 		if (!ipoib_cm_has_srq(priv)) {
770 			ipoib_cm_free_rx_ring(priv, rx->rx_ring);
771 			spin_lock_irq(&priv->lock);
772 			--priv->cm.nonsrq_conn_qp;
773 			spin_unlock_irq(&priv->lock);
774 		}
775 		kfree(rx);
776 	}
777 }
778 
779 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
780 {
781 	struct ipoib_cm_rx *p;
782 	unsigned long begin;
783 	int ret;
784 
785 	if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)) || !priv->cm.id)
786 		return;
787 
788 	ib_destroy_cm_id(priv->cm.id);
789 	priv->cm.id = NULL;
790 
791 	cancel_work_sync(&priv->cm.rx_reap_task);
792 
793 	spin_lock_irq(&priv->lock);
794 	while (!list_empty(&priv->cm.passive_ids)) {
795 		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
796 		list_move(&p->list, &priv->cm.rx_error_list);
797 		p->state = IPOIB_CM_RX_ERROR;
798 		spin_unlock_irq(&priv->lock);
799 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
800 		if (ret)
801 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
802 		spin_lock_irq(&priv->lock);
803 	}
804 
805 	/* Wait for all RX to be drained */
806 	begin = jiffies;
807 
808 	while (!list_empty(&priv->cm.rx_error_list) ||
809 	       !list_empty(&priv->cm.rx_flush_list) ||
810 	       !list_empty(&priv->cm.rx_drain_list)) {
811 		if (time_after(jiffies, begin + 5 * HZ)) {
812 			ipoib_warn(priv, "RX drain timing out\n");
813 
814 			/*
815 			 * assume the HW is wedged and just free up everything.
816 			 */
817 			list_splice_init(&priv->cm.rx_flush_list,
818 					 &priv->cm.rx_reap_list);
819 			list_splice_init(&priv->cm.rx_error_list,
820 					 &priv->cm.rx_reap_list);
821 			list_splice_init(&priv->cm.rx_drain_list,
822 					 &priv->cm.rx_reap_list);
823 			break;
824 		}
825 		spin_unlock_irq(&priv->lock);
826 		msleep(1);
827 		ipoib_drain_cq(priv);
828 		spin_lock_irq(&priv->lock);
829 	}
830 
831 	spin_unlock_irq(&priv->lock);
832 
833 	ipoib_cm_free_rx_reap_list(priv);
834 
835 	cancel_delayed_work_sync(&priv->cm.stale_task);
836 }
837 
838 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
839 {
840 	struct ipoib_cm_tx *p = cm_id->context;
841 	struct ipoib_dev_priv *priv = p->priv;
842 	struct ipoib_cm_data *data = event->private_data;
843 	struct ifqueue mbqueue;
844 	struct ib_qp_attr qp_attr;
845 	int qp_attr_mask, ret;
846 	struct mbuf *mb;
847 
848 	ipoib_dbg(priv, "cm rep handler\n");
849 	p->mtu = be32_to_cpu(data->mtu);
850 
851 	if (p->mtu <= IPOIB_ENCAP_LEN) {
852 		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
853 			   p->mtu, IPOIB_ENCAP_LEN);
854 		return -EINVAL;
855 	}
856 
857 	qp_attr.qp_state = IB_QPS_RTR;
858 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
859 	if (ret) {
860 		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
861 		return ret;
862 	}
863 
864 	qp_attr.rq_psn = 0 /* FIXME */;
865 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
866 	if (ret) {
867 		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
868 		return ret;
869 	}
870 
871 	qp_attr.qp_state = IB_QPS_RTS;
872 	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
873 	if (ret) {
874 		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
875 		return ret;
876 	}
877 	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
878 	if (ret) {
879 		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
880 		return ret;
881 	}
882 
883 	bzero(&mbqueue, sizeof(mbqueue));
884 
885 	spin_lock_irq(&priv->lock);
886 	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
887 	if (p->path)
888 		for (;;) {
889 			_IF_DEQUEUE(&p->path->queue, mb);
890 			if (mb == NULL)
891 				break;
892 			_IF_ENQUEUE(&mbqueue, mb);
893 		}
894 	spin_unlock_irq(&priv->lock);
895 
896 	for (;;) {
897 		struct ifnet *dev = p->priv->dev;
898 		_IF_DEQUEUE(&mbqueue, mb);
899 		if (mb == NULL)
900 			break;
901 		mb->m_pkthdr.rcvif = dev;
902 		if (dev->if_transmit(dev, mb))
903 			ipoib_warn(priv, "dev_queue_xmit failed "
904 				   "to requeue packet\n");
905 	}
906 
907 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
908 	if (ret) {
909 		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
910 		return ret;
911 	}
912 	return 0;
913 }
914 
915 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
916     struct ipoib_cm_tx *tx)
917 {
918 	struct ib_qp_init_attr attr = {
919 		.send_cq		= priv->send_cq,
920 		.recv_cq		= priv->recv_cq,
921 		.srq			= priv->cm.srq,
922 		.cap.max_send_wr	= ipoib_sendq_size,
923 		.cap.max_send_sge	= priv->cm.num_frags,
924 		.sq_sig_type		= IB_SIGNAL_ALL_WR,
925 		.qp_type		= IB_QPT_RC,
926 		.qp_context		= tx
927 	};
928 
929 	return ib_create_qp(priv->pd, &attr);
930 }
931 
932 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
933 			     struct ib_cm_id *id, struct ib_qp *qp,
934 			     u32 qpn,
935 			     struct ib_sa_path_rec *pathrec)
936 {
937 	struct ipoib_cm_data data = {};
938 	struct ib_cm_req_param req = {};
939 
940 	ipoib_dbg(priv, "cm send req\n");
941 
942 	data.qpn = cpu_to_be32(priv->qp->qp_num);
943 	data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
944 
945 	req.primary_path		= pathrec;
946 	req.alternate_path		= NULL;
947 	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
948 	req.qp_num			= qp->qp_num;
949 	req.qp_type			= qp->qp_type;
950 	req.private_data		= &data;
951 	req.private_data_len		= sizeof data;
952 	req.flow_control		= 0;
953 
954 	req.starting_psn		= 0; /* FIXME */
955 
956 	/*
957 	 * Pick some arbitrary defaults here; we could make these
958 	 * module parameters if anyone cared about setting them.
959 	 */
960 	req.responder_resources		= 4;
961 	req.remote_cm_response_timeout	= 20;
962 	req.local_cm_response_timeout	= 20;
963 	req.retry_count			= 0; /* RFC draft warns against retries */
964 	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
965 	req.max_cm_retries		= 15;
966 	req.srq				= ipoib_cm_has_srq(priv);
967 	return ib_send_cm_req(id, &req);
968 }
969 
970 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
971 				  struct ib_cm_id *cm_id, struct ib_qp *qp)
972 {
973 	struct ib_qp_attr qp_attr;
974 	int qp_attr_mask, ret;
975 	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
976 	if (ret) {
977 		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
978 		return ret;
979 	}
980 
981 	qp_attr.qp_state = IB_QPS_INIT;
982 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
983 	qp_attr.port_num = priv->port;
984 	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
985 
986 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
987 	if (ret) {
988 		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
989 		return ret;
990 	}
991 	return 0;
992 }
993 
994 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
995 			    struct ib_sa_path_rec *pathrec)
996 {
997 	struct ipoib_dev_priv *priv = p->priv;
998 	int ret;
999 
1000 	p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
1001 	if (!p->tx_ring) {
1002 		ipoib_warn(priv, "failed to allocate tx ring\n");
1003 		ret = -ENOMEM;
1004 		goto err_tx;
1005 	}
1006 	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1007 
1008 	p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1009 	if (IS_ERR(p->qp)) {
1010 		ret = PTR_ERR(p->qp);
1011 		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1012 		goto err_qp;
1013 	}
1014 
1015 	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1016 	if (IS_ERR(p->id)) {
1017 		ret = PTR_ERR(p->id);
1018 		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1019 		goto err_id;
1020 	}
1021 
1022 	ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1023 	if (ret) {
1024 		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1025 		goto err_modify;
1026 	}
1027 
1028 	ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1029 	if (ret) {
1030 		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1031 		goto err_send_cm;
1032 	}
1033 
1034 	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1035 		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1036 
1037 	return 0;
1038 
1039 err_send_cm:
1040 err_modify:
1041 	ib_destroy_cm_id(p->id);
1042 err_id:
1043 	p->id = NULL;
1044 	ib_destroy_qp(p->qp);
1045 err_qp:
1046 	p->qp = NULL;
1047 	kfree(p->tx_ring);
1048 err_tx:
1049 	return ret;
1050 }
1051 
1052 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1053 {
1054 	struct ipoib_dev_priv *priv = p->priv;
1055 	struct ifnet *dev = priv->dev;
1056 	struct ipoib_cm_tx_buf *tx_req;
1057 	unsigned long begin;
1058 
1059 	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1060 		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1061 
1062 	if (p->path)
1063 		ipoib_path_free(priv, p->path);
1064 
1065 	if (p->id)
1066 		ib_destroy_cm_id(p->id);
1067 
1068 	if (p->tx_ring) {
1069 		/* Wait for all sends to complete */
1070 		begin = jiffies;
1071 		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1072 			if (time_after(jiffies, begin + 5 * HZ)) {
1073 				ipoib_warn(priv, "timing out; %d sends not completed\n",
1074 					   p->tx_head - p->tx_tail);
1075 				goto timeout;
1076 			}
1077 
1078 			msleep(1);
1079 		}
1080 	}
1081 
1082 timeout:
1083 
1084 	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1085 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1086 		ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1087 		m_freem(tx_req->mb);
1088 		++p->tx_tail;
1089 		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1090 		    (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
1091 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1092 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
1093 	}
1094 
1095 	if (p->qp)
1096 		ib_destroy_qp(p->qp);
1097 
1098 	kfree(p->tx_ring);
1099 	kfree(p);
1100 }
1101 
1102 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1103 			       struct ib_cm_event *event)
1104 {
1105 	struct ipoib_cm_tx *tx = cm_id->context;
1106 	struct ipoib_dev_priv *priv = tx->priv;
1107 	struct ipoib_path *path;
1108 	unsigned long flags;
1109 	int ret;
1110 
1111 	switch (event->event) {
1112 	case IB_CM_DREQ_RECEIVED:
1113 		ipoib_dbg(priv, "DREQ received.\n");
1114 		ib_send_cm_drep(cm_id, NULL, 0);
1115 		break;
1116 	case IB_CM_REP_RECEIVED:
1117 		ipoib_dbg(priv, "REP received.\n");
1118 		ret = ipoib_cm_rep_handler(cm_id, event);
1119 		if (ret)
1120 			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1121 				       NULL, 0, NULL, 0);
1122 		break;
1123 	case IB_CM_REQ_ERROR:
1124 	case IB_CM_REJ_RECEIVED:
1125 	case IB_CM_TIMEWAIT_EXIT:
1126 		ipoib_dbg(priv, "CM error %d.\n", event->event);
1127 		spin_lock_irqsave(&priv->lock, flags);
1128 		path = tx->path;
1129 
1130 		if (path) {
1131 			path->cm = NULL;
1132 			tx->path = NULL;
1133 			rb_erase(&path->rb_node, &priv->path_tree);
1134 			list_del(&path->list);
1135 		}
1136 
1137 		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1138 			list_move(&tx->list, &priv->cm.reap_list);
1139 			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1140 		}
1141 
1142 		spin_unlock_irqrestore(&priv->lock, flags);
1143 		if (path)
1144 			ipoib_path_free(tx->priv, path);
1145 		break;
1146 	default:
1147 		break;
1148 	}
1149 
1150 	return 0;
1151 }
1152 
1153 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1154     struct ipoib_path *path)
1155 {
1156 	struct ipoib_cm_tx *tx;
1157 
1158 	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1159 	if (!tx)
1160 		return NULL;
1161 
1162 	ipoib_dbg(priv, "Creating cm tx\n");
1163 	path->cm = tx;
1164 	tx->path = path;
1165 	tx->priv = priv;
1166 	list_add(&tx->list, &priv->cm.start_list);
1167 	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1168 	queue_work(ipoib_workqueue, &priv->cm.start_task);
1169 	return tx;
1170 }
1171 
1172 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1173 {
1174 	struct ipoib_dev_priv *priv = tx->priv;
1175 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1176 		spin_lock(&priv->lock);
1177 		list_move(&tx->list, &priv->cm.reap_list);
1178 		spin_unlock(&priv->lock);
1179 		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1180 		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1181 			  tx->path->pathrec.dgid.raw);
1182 		tx->path = NULL;
1183 	}
1184 }
1185 
1186 static void ipoib_cm_tx_start(struct work_struct *work)
1187 {
1188 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1189 						   cm.start_task);
1190 	struct ipoib_path *path;
1191 	struct ipoib_cm_tx *p;
1192 	unsigned long flags;
1193 	int ret;
1194 
1195 	struct ib_sa_path_rec pathrec;
1196 	u32 qpn;
1197 
1198 	ipoib_dbg(priv, "cm start task\n");
1199 	spin_lock_irqsave(&priv->lock, flags);
1200 
1201 	while (!list_empty(&priv->cm.start_list)) {
1202 		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1203 		list_del_init(&p->list);
1204 		path = p->path;
1205 		qpn = IPOIB_QPN(path->hwaddr);
1206 		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1207 
1208 		spin_unlock_irqrestore(&priv->lock, flags);
1209 
1210 		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1211 
1212 		spin_lock_irqsave(&priv->lock, flags);
1213 
1214 		if (ret) {
1215 			path = p->path;
1216 			if (path) {
1217 				path->cm = NULL;
1218 				rb_erase(&path->rb_node, &priv->path_tree);
1219 				list_del(&path->list);
1220 				ipoib_path_free(priv, path);
1221 			}
1222 			list_del(&p->list);
1223 			kfree(p);
1224 		}
1225 	}
1226 
1227 	spin_unlock_irqrestore(&priv->lock, flags);
1228 }
1229 
1230 static void ipoib_cm_tx_reap(struct work_struct *work)
1231 {
1232 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1233 						   cm.reap_task);
1234 	struct ipoib_cm_tx *p;
1235 	unsigned long flags;
1236 
1237 	spin_lock_irqsave(&priv->lock, flags);
1238 
1239 	while (!list_empty(&priv->cm.reap_list)) {
1240 		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1241 		list_del(&p->list);
1242 		spin_unlock_irqrestore(&priv->lock, flags);
1243 		ipoib_cm_tx_destroy(p);
1244 		spin_lock_irqsave(&priv->lock, flags);
1245 	}
1246 
1247 	spin_unlock_irqrestore(&priv->lock, flags);
1248 }
1249 
1250 static void ipoib_cm_mb_reap(struct work_struct *work)
1251 {
1252 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1253 						   cm.mb_task);
1254 	struct mbuf *mb;
1255 	unsigned long flags;
1256 #if defined(INET) || defined(INET6)
1257 	unsigned mtu = priv->mcast_mtu;
1258 #endif
1259 	uint16_t proto;
1260 
1261 	spin_lock_irqsave(&priv->lock, flags);
1262 
1263 	for (;;) {
1264 		IF_DEQUEUE(&priv->cm.mb_queue, mb);
1265 		if (mb == NULL)
1266 			break;
1267 		spin_unlock_irqrestore(&priv->lock, flags);
1268 
1269 		proto = htons(*mtod(mb, uint16_t *));
1270 		m_adj(mb, IPOIB_ENCAP_LEN);
1271 		switch (proto) {
1272 #if defined(INET)
1273 		case ETHERTYPE_IP:
1274 			icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1275 			break;
1276 #endif
1277 #if defined(INET6)
1278 		case ETHERTYPE_IPV6:
1279 			icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1280 			break;
1281 #endif
1282 		default:
1283 			m_freem(mb);
1284 		}
1285 
1286 		spin_lock_irqsave(&priv->lock, flags);
1287 	}
1288 
1289 	spin_unlock_irqrestore(&priv->lock, flags);
1290 }
1291 
1292 void
1293 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1294 {
1295 	int e = priv->cm.mb_queue.ifq_len;
1296 
1297 	IF_ENQUEUE(&priv->cm.mb_queue, mb);
1298 	if (e == 0)
1299 		queue_work(ipoib_workqueue, &priv->cm.mb_task);
1300 }
1301 
1302 static void ipoib_cm_rx_reap(struct work_struct *work)
1303 {
1304 	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1305 						cm.rx_reap_task));
1306 }
1307 
1308 static void ipoib_cm_stale_task(struct work_struct *work)
1309 {
1310 	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1311 						   cm.stale_task.work);
1312 	struct ipoib_cm_rx *p;
1313 	int ret;
1314 
1315 	spin_lock_irq(&priv->lock);
1316 	while (!list_empty(&priv->cm.passive_ids)) {
1317 		/* List is sorted by LRU, start from tail,
1318 		 * stop when we see a recently used entry */
1319 		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1320 		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1321 			break;
1322 		list_move(&p->list, &priv->cm.rx_error_list);
1323 		p->state = IPOIB_CM_RX_ERROR;
1324 		spin_unlock_irq(&priv->lock);
1325 		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1326 		if (ret)
1327 			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1328 		spin_lock_irq(&priv->lock);
1329 	}
1330 
1331 	if (!list_empty(&priv->cm.passive_ids))
1332 		queue_delayed_work(ipoib_workqueue,
1333 				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1334 	spin_unlock_irq(&priv->lock);
1335 }
1336 
1337 
1338 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1339 {
1340 	struct ib_srq_init_attr srq_init_attr = {
1341 		.attr = {
1342 			.max_wr  = ipoib_recvq_size,
1343 			.max_sge = max_sge
1344 		}
1345 	};
1346 
1347 	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1348 	if (IS_ERR(priv->cm.srq)) {
1349 		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1350 			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1351 			       priv->ca->name, PTR_ERR(priv->cm.srq));
1352 		priv->cm.srq = NULL;
1353 		return;
1354 	}
1355 
1356 	priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1357 	if (!priv->cm.srq_ring) {
1358 		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1359 		       priv->ca->name, ipoib_recvq_size);
1360 		ib_destroy_srq(priv->cm.srq);
1361 		priv->cm.srq = NULL;
1362 		return;
1363 	}
1364 
1365 	memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1366 }
1367 
1368 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1369 {
1370 	struct ifnet *dev = priv->dev;
1371 	int i;
1372 	int max_srq_sge;
1373 
1374 	INIT_LIST_HEAD(&priv->cm.passive_ids);
1375 	INIT_LIST_HEAD(&priv->cm.reap_list);
1376 	INIT_LIST_HEAD(&priv->cm.start_list);
1377 	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1378 	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1379 	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1380 	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1381 	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1382 	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1383 	INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1384 	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1385 	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1386 
1387 	bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1388 	mtx_init(&priv->cm.mb_queue.ifq_mtx,
1389 	    dev->if_xname, "if send queue", MTX_DEF);
1390 
1391 	max_srq_sge = priv->ca->attrs.max_srq_sge;
1392 
1393 	ipoib_dbg(priv, "max_srq_sge=%d\n", max_srq_sge);
1394 
1395 	max_srq_sge = min_t(int, IPOIB_CM_RX_SG, max_srq_sge);
1396 	ipoib_cm_create_srq(priv, max_srq_sge);
1397 	if (ipoib_cm_has_srq(priv)) {
1398 		priv->cm.max_cm_mtu = max_srq_sge * MJUMPAGESIZE;
1399 		priv->cm.num_frags  = max_srq_sge;
1400 		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1401 			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1402 	} else {
1403 		priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1404 		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1405 	}
1406 
1407 	ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1408 
1409 	if (ipoib_cm_has_srq(priv)) {
1410 		for (i = 0; i < ipoib_recvq_size; ++i) {
1411 			if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1412 				ipoib_warn(priv, "failed to allocate "
1413 					   "receive buffer %d\n", i);
1414 				ipoib_cm_dev_cleanup(priv);
1415 				return -ENOMEM;
1416 			}
1417 
1418 			if (ipoib_cm_post_receive_srq(priv, i)) {
1419 				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1420 					   "failed for buf %d\n", i);
1421 				ipoib_cm_dev_cleanup(priv);
1422 				return -EIO;
1423 			}
1424 		}
1425 	}
1426 
1427 	IF_LLADDR(priv->dev)[0] = IPOIB_FLAGS_RC;
1428 	return 0;
1429 }
1430 
1431 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1432 {
1433 	int ret;
1434 
1435 	if (!priv->cm.srq)
1436 		return;
1437 
1438 	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1439 
1440 	ret = ib_destroy_srq(priv->cm.srq);
1441 	if (ret)
1442 		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1443 
1444 	priv->cm.srq = NULL;
1445 	if (!priv->cm.srq_ring)
1446 		return;
1447 
1448 	ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1449 	priv->cm.srq_ring = NULL;
1450 
1451 	mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1452 }
1453 
1454 #endif /* CONFIG_INFINIBAND_IPOIB_CM */
1455