xref: /openbmc/linux/net/rds/ib_cm.c (revision e2f1cf25)
1 /*
2  * Copyright (c) 2006 Oracle.  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 <linux/kernel.h>
34 #include <linux/in.h>
35 #include <linux/slab.h>
36 #include <linux/vmalloc.h>
37 #include <linux/ratelimit.h>
38 
39 #include "rds.h"
40 #include "ib.h"
41 
42 /*
43  * Set the selected protocol version
44  */
45 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
46 {
47 	conn->c_version = version;
48 }
49 
50 /*
51  * Set up flow control
52  */
53 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
54 {
55 	struct rds_ib_connection *ic = conn->c_transport_data;
56 
57 	if (rds_ib_sysctl_flow_control && credits != 0) {
58 		/* We're doing flow control */
59 		ic->i_flowctl = 1;
60 		rds_ib_send_add_credits(conn, credits);
61 	} else {
62 		ic->i_flowctl = 0;
63 	}
64 }
65 
66 /*
67  * Tune RNR behavior. Without flow control, we use a rather
68  * low timeout, but not the absolute minimum - this should
69  * be tunable.
70  *
71  * We already set the RNR retry count to 7 (which is the
72  * smallest infinite number :-) above.
73  * If flow control is off, we want to change this back to 0
74  * so that we learn quickly when our credit accounting is
75  * buggy.
76  *
77  * Caller passes in a qp_attr pointer - don't waste stack spacv
78  * by allocation this twice.
79  */
80 static void
81 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
82 {
83 	int ret;
84 
85 	attr->min_rnr_timer = IB_RNR_TIMER_000_32;
86 	ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
87 	if (ret)
88 		printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
89 }
90 
91 /*
92  * Connection established.
93  * We get here for both outgoing and incoming connection.
94  */
95 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
96 {
97 	const struct rds_ib_connect_private *dp = NULL;
98 	struct rds_ib_connection *ic = conn->c_transport_data;
99 	struct ib_qp_attr qp_attr;
100 	int err;
101 
102 	if (event->param.conn.private_data_len >= sizeof(*dp)) {
103 		dp = event->param.conn.private_data;
104 
105 		/* make sure it isn't empty data */
106 		if (dp->dp_protocol_major) {
107 			rds_ib_set_protocol(conn,
108 				RDS_PROTOCOL(dp->dp_protocol_major,
109 				dp->dp_protocol_minor));
110 			rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
111 		}
112 	}
113 
114 	if (conn->c_version < RDS_PROTOCOL(3,1)) {
115 		printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
116 		       " no longer supported\n",
117 		       &conn->c_faddr,
118 		       RDS_PROTOCOL_MAJOR(conn->c_version),
119 		       RDS_PROTOCOL_MINOR(conn->c_version));
120 		rds_conn_destroy(conn);
121 		return;
122 	} else {
123 		printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
124 		       &conn->c_faddr,
125 		       RDS_PROTOCOL_MAJOR(conn->c_version),
126 		       RDS_PROTOCOL_MINOR(conn->c_version),
127 		       ic->i_flowctl ? ", flow control" : "");
128 	}
129 
130 	/*
131 	 * Init rings and fill recv. this needs to wait until protocol negotiation
132 	 * is complete, since ring layout is different from 3.0 to 3.1.
133 	 */
134 	rds_ib_send_init_ring(ic);
135 	rds_ib_recv_init_ring(ic);
136 	/* Post receive buffers - as a side effect, this will update
137 	 * the posted credit count. */
138 	rds_ib_recv_refill(conn, 1);
139 
140 	/* Tune RNR behavior */
141 	rds_ib_tune_rnr(ic, &qp_attr);
142 
143 	qp_attr.qp_state = IB_QPS_RTS;
144 	err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
145 	if (err)
146 		printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
147 
148 	/* update ib_device with this local ipaddr */
149 	err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
150 	if (err)
151 		printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
152 			err);
153 
154 	/* If the peer gave us the last packet it saw, process this as if
155 	 * we had received a regular ACK. */
156 	if (dp) {
157 		/* dp structure start is not guaranteed to be 8 bytes aligned.
158 		 * Since dp_ack_seq is 64-bit extended load operations can be
159 		 * used so go through get_unaligned to avoid unaligned errors.
160 		 */
161 		__be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
162 
163 		if (dp_ack_seq)
164 			rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
165 					    NULL);
166 	}
167 
168 	rds_connect_complete(conn);
169 }
170 
171 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
172 			struct rdma_conn_param *conn_param,
173 			struct rds_ib_connect_private *dp,
174 			u32 protocol_version,
175 			u32 max_responder_resources,
176 			u32 max_initiator_depth)
177 {
178 	struct rds_ib_connection *ic = conn->c_transport_data;
179 	struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
180 
181 	memset(conn_param, 0, sizeof(struct rdma_conn_param));
182 
183 	conn_param->responder_resources =
184 		min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
185 	conn_param->initiator_depth =
186 		min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
187 	conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
188 	conn_param->rnr_retry_count = 7;
189 
190 	if (dp) {
191 		memset(dp, 0, sizeof(*dp));
192 		dp->dp_saddr = conn->c_laddr;
193 		dp->dp_daddr = conn->c_faddr;
194 		dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
195 		dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
196 		dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
197 		dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
198 
199 		/* Advertise flow control */
200 		if (ic->i_flowctl) {
201 			unsigned int credits;
202 
203 			credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
204 			dp->dp_credit = cpu_to_be32(credits);
205 			atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
206 		}
207 
208 		conn_param->private_data = dp;
209 		conn_param->private_data_len = sizeof(*dp);
210 	}
211 }
212 
213 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
214 {
215 	rdsdebug("event %u (%s) data %p\n",
216 		 event->event, ib_event_msg(event->event), data);
217 }
218 
219 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
220 {
221 	struct rds_connection *conn = data;
222 	struct rds_ib_connection *ic = conn->c_transport_data;
223 
224 	rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
225 		 ib_event_msg(event->event));
226 
227 	switch (event->event) {
228 	case IB_EVENT_COMM_EST:
229 		rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
230 		break;
231 	default:
232 		rdsdebug("Fatal QP Event %u (%s) "
233 			"- connection %pI4->%pI4, reconnecting\n",
234 			event->event, ib_event_msg(event->event),
235 			&conn->c_laddr, &conn->c_faddr);
236 		rds_conn_drop(conn);
237 		break;
238 	}
239 }
240 
241 /*
242  * This needs to be very careful to not leave IS_ERR pointers around for
243  * cleanup to trip over.
244  */
245 static int rds_ib_setup_qp(struct rds_connection *conn)
246 {
247 	struct rds_ib_connection *ic = conn->c_transport_data;
248 	struct ib_device *dev = ic->i_cm_id->device;
249 	struct ib_qp_init_attr attr;
250 	struct ib_cq_init_attr cq_attr = {};
251 	struct rds_ib_device *rds_ibdev;
252 	int ret;
253 
254 	/*
255 	 * It's normal to see a null device if an incoming connection races
256 	 * with device removal, so we don't print a warning.
257 	 */
258 	rds_ibdev = rds_ib_get_client_data(dev);
259 	if (!rds_ibdev)
260 		return -EOPNOTSUPP;
261 
262 	/* add the conn now so that connection establishment has the dev */
263 	rds_ib_add_conn(rds_ibdev, conn);
264 
265 	if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
266 		rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
267 	if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
268 		rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
269 
270 	/* Protection domain and memory range */
271 	ic->i_pd = rds_ibdev->pd;
272 	ic->i_mr = rds_ibdev->mr;
273 
274 	cq_attr.cqe = ic->i_send_ring.w_nr + 1;
275 	ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
276 				     rds_ib_cq_event_handler, conn,
277 				     &cq_attr);
278 	if (IS_ERR(ic->i_send_cq)) {
279 		ret = PTR_ERR(ic->i_send_cq);
280 		ic->i_send_cq = NULL;
281 		rdsdebug("ib_create_cq send failed: %d\n", ret);
282 		goto out;
283 	}
284 
285 	cq_attr.cqe = ic->i_recv_ring.w_nr;
286 	ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
287 				     rds_ib_cq_event_handler, conn,
288 				     &cq_attr);
289 	if (IS_ERR(ic->i_recv_cq)) {
290 		ret = PTR_ERR(ic->i_recv_cq);
291 		ic->i_recv_cq = NULL;
292 		rdsdebug("ib_create_cq recv failed: %d\n", ret);
293 		goto out;
294 	}
295 
296 	ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
297 	if (ret) {
298 		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
299 		goto out;
300 	}
301 
302 	ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
303 	if (ret) {
304 		rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
305 		goto out;
306 	}
307 
308 	/* XXX negotiate max send/recv with remote? */
309 	memset(&attr, 0, sizeof(attr));
310 	attr.event_handler = rds_ib_qp_event_handler;
311 	attr.qp_context = conn;
312 	/* + 1 to allow for the single ack message */
313 	attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
314 	attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
315 	attr.cap.max_send_sge = rds_ibdev->max_sge;
316 	attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
317 	attr.sq_sig_type = IB_SIGNAL_REQ_WR;
318 	attr.qp_type = IB_QPT_RC;
319 	attr.send_cq = ic->i_send_cq;
320 	attr.recv_cq = ic->i_recv_cq;
321 
322 	/*
323 	 * XXX this can fail if max_*_wr is too large?  Are we supposed
324 	 * to back off until we get a value that the hardware can support?
325 	 */
326 	ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
327 	if (ret) {
328 		rdsdebug("rdma_create_qp failed: %d\n", ret);
329 		goto out;
330 	}
331 
332 	ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
333 					   ic->i_send_ring.w_nr *
334 						sizeof(struct rds_header),
335 					   &ic->i_send_hdrs_dma, GFP_KERNEL);
336 	if (!ic->i_send_hdrs) {
337 		ret = -ENOMEM;
338 		rdsdebug("ib_dma_alloc_coherent send failed\n");
339 		goto out;
340 	}
341 
342 	ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
343 					   ic->i_recv_ring.w_nr *
344 						sizeof(struct rds_header),
345 					   &ic->i_recv_hdrs_dma, GFP_KERNEL);
346 	if (!ic->i_recv_hdrs) {
347 		ret = -ENOMEM;
348 		rdsdebug("ib_dma_alloc_coherent recv failed\n");
349 		goto out;
350 	}
351 
352 	ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
353 				       &ic->i_ack_dma, GFP_KERNEL);
354 	if (!ic->i_ack) {
355 		ret = -ENOMEM;
356 		rdsdebug("ib_dma_alloc_coherent ack failed\n");
357 		goto out;
358 	}
359 
360 	ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
361 				   ibdev_to_node(dev));
362 	if (!ic->i_sends) {
363 		ret = -ENOMEM;
364 		rdsdebug("send allocation failed\n");
365 		goto out;
366 	}
367 
368 	ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
369 				   ibdev_to_node(dev));
370 	if (!ic->i_recvs) {
371 		ret = -ENOMEM;
372 		rdsdebug("recv allocation failed\n");
373 		goto out;
374 	}
375 
376 	rds_ib_recv_init_ack(ic);
377 
378 	rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
379 		 ic->i_send_cq, ic->i_recv_cq);
380 
381 out:
382 	rds_ib_dev_put(rds_ibdev);
383 	return ret;
384 }
385 
386 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
387 {
388 	const struct rds_ib_connect_private *dp = event->param.conn.private_data;
389 	u16 common;
390 	u32 version = 0;
391 
392 	/*
393 	 * rdma_cm private data is odd - when there is any private data in the
394 	 * request, we will be given a pretty large buffer without telling us the
395 	 * original size. The only way to tell the difference is by looking at
396 	 * the contents, which are initialized to zero.
397 	 * If the protocol version fields aren't set, this is a connection attempt
398 	 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
399 	 * We really should have changed this for OFED 1.3 :-(
400 	 */
401 
402 	/* Be paranoid. RDS always has privdata */
403 	if (!event->param.conn.private_data_len) {
404 		printk(KERN_NOTICE "RDS incoming connection has no private data, "
405 			"rejecting\n");
406 		return 0;
407 	}
408 
409 	/* Even if len is crap *now* I still want to check it. -ASG */
410 	if (event->param.conn.private_data_len < sizeof (*dp) ||
411 	    dp->dp_protocol_major == 0)
412 		return RDS_PROTOCOL_3_0;
413 
414 	common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
415 	if (dp->dp_protocol_major == 3 && common) {
416 		version = RDS_PROTOCOL_3_0;
417 		while ((common >>= 1) != 0)
418 			version++;
419 	} else
420 		printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
421 				&dp->dp_saddr,
422 				dp->dp_protocol_major,
423 				dp->dp_protocol_minor);
424 	return version;
425 }
426 
427 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
428 				    struct rdma_cm_event *event)
429 {
430 	__be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
431 	__be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
432 	const struct rds_ib_connect_private *dp = event->param.conn.private_data;
433 	struct rds_ib_connect_private dp_rep;
434 	struct rds_connection *conn = NULL;
435 	struct rds_ib_connection *ic = NULL;
436 	struct rdma_conn_param conn_param;
437 	u32 version;
438 	int err = 1, destroy = 1;
439 
440 	/* Check whether the remote protocol version matches ours. */
441 	version = rds_ib_protocol_compatible(event);
442 	if (!version)
443 		goto out;
444 
445 	rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
446 		 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
447 		 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
448 		 (unsigned long long)be64_to_cpu(lguid),
449 		 (unsigned long long)be64_to_cpu(fguid));
450 
451 	conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
452 			       GFP_KERNEL);
453 	if (IS_ERR(conn)) {
454 		rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
455 		conn = NULL;
456 		goto out;
457 	}
458 
459 	/*
460 	 * The connection request may occur while the
461 	 * previous connection exist, e.g. in case of failover.
462 	 * But as connections may be initiated simultaneously
463 	 * by both hosts, we have a random backoff mechanism -
464 	 * see the comment above rds_queue_reconnect()
465 	 */
466 	mutex_lock(&conn->c_cm_lock);
467 	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
468 		if (rds_conn_state(conn) == RDS_CONN_UP) {
469 			rdsdebug("incoming connect while connecting\n");
470 			rds_conn_drop(conn);
471 			rds_ib_stats_inc(s_ib_listen_closed_stale);
472 		} else
473 		if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
474 			/* Wait and see - our connect may still be succeeding */
475 			rds_ib_stats_inc(s_ib_connect_raced);
476 		}
477 		goto out;
478 	}
479 
480 	ic = conn->c_transport_data;
481 
482 	rds_ib_set_protocol(conn, version);
483 	rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
484 
485 	/* If the peer gave us the last packet it saw, process this as if
486 	 * we had received a regular ACK. */
487 	if (dp->dp_ack_seq)
488 		rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
489 
490 	BUG_ON(cm_id->context);
491 	BUG_ON(ic->i_cm_id);
492 
493 	ic->i_cm_id = cm_id;
494 	cm_id->context = conn;
495 
496 	/* We got halfway through setting up the ib_connection, if we
497 	 * fail now, we have to take the long route out of this mess. */
498 	destroy = 0;
499 
500 	err = rds_ib_setup_qp(conn);
501 	if (err) {
502 		rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
503 		goto out;
504 	}
505 
506 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
507 		event->param.conn.responder_resources,
508 		event->param.conn.initiator_depth);
509 
510 	/* rdma_accept() calls rdma_reject() internally if it fails */
511 	err = rdma_accept(cm_id, &conn_param);
512 	if (err)
513 		rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
514 
515 out:
516 	if (conn)
517 		mutex_unlock(&conn->c_cm_lock);
518 	if (err)
519 		rdma_reject(cm_id, NULL, 0);
520 	return destroy;
521 }
522 
523 
524 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
525 {
526 	struct rds_connection *conn = cm_id->context;
527 	struct rds_ib_connection *ic = conn->c_transport_data;
528 	struct rdma_conn_param conn_param;
529 	struct rds_ib_connect_private dp;
530 	int ret;
531 
532 	/* If the peer doesn't do protocol negotiation, we must
533 	 * default to RDSv3.0 */
534 	rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
535 	ic->i_flowctl = rds_ib_sysctl_flow_control;	/* advertise flow control */
536 
537 	ret = rds_ib_setup_qp(conn);
538 	if (ret) {
539 		rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
540 		goto out;
541 	}
542 
543 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
544 		UINT_MAX, UINT_MAX);
545 	ret = rdma_connect(cm_id, &conn_param);
546 	if (ret)
547 		rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
548 
549 out:
550 	/* Beware - returning non-zero tells the rdma_cm to destroy
551 	 * the cm_id. We should certainly not do it as long as we still
552 	 * "own" the cm_id. */
553 	if (ret) {
554 		if (ic->i_cm_id == cm_id)
555 			ret = 0;
556 	}
557 	return ret;
558 }
559 
560 int rds_ib_conn_connect(struct rds_connection *conn)
561 {
562 	struct rds_ib_connection *ic = conn->c_transport_data;
563 	struct sockaddr_in src, dest;
564 	int ret;
565 
566 	/* XXX I wonder what affect the port space has */
567 	/* delegate cm event handler to rdma_transport */
568 	ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
569 				     RDMA_PS_TCP, IB_QPT_RC);
570 	if (IS_ERR(ic->i_cm_id)) {
571 		ret = PTR_ERR(ic->i_cm_id);
572 		ic->i_cm_id = NULL;
573 		rdsdebug("rdma_create_id() failed: %d\n", ret);
574 		goto out;
575 	}
576 
577 	rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
578 
579 	src.sin_family = AF_INET;
580 	src.sin_addr.s_addr = (__force u32)conn->c_laddr;
581 	src.sin_port = (__force u16)htons(0);
582 
583 	dest.sin_family = AF_INET;
584 	dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
585 	dest.sin_port = (__force u16)htons(RDS_PORT);
586 
587 	ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
588 				(struct sockaddr *)&dest,
589 				RDS_RDMA_RESOLVE_TIMEOUT_MS);
590 	if (ret) {
591 		rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
592 			 ret);
593 		rdma_destroy_id(ic->i_cm_id);
594 		ic->i_cm_id = NULL;
595 	}
596 
597 out:
598 	return ret;
599 }
600 
601 /*
602  * This is so careful about only cleaning up resources that were built up
603  * so that it can be called at any point during startup.  In fact it
604  * can be called multiple times for a given connection.
605  */
606 void rds_ib_conn_shutdown(struct rds_connection *conn)
607 {
608 	struct rds_ib_connection *ic = conn->c_transport_data;
609 	int err = 0;
610 
611 	rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
612 		 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
613 		 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
614 
615 	if (ic->i_cm_id) {
616 		struct ib_device *dev = ic->i_cm_id->device;
617 
618 		rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
619 		err = rdma_disconnect(ic->i_cm_id);
620 		if (err) {
621 			/* Actually this may happen quite frequently, when
622 			 * an outgoing connect raced with an incoming connect.
623 			 */
624 			rdsdebug("failed to disconnect, cm: %p err %d\n",
625 				ic->i_cm_id, err);
626 		}
627 
628 		/*
629 		 * We want to wait for tx and rx completion to finish
630 		 * before we tear down the connection, but we have to be
631 		 * careful not to get stuck waiting on a send ring that
632 		 * only has unsignaled sends in it.  We've shutdown new
633 		 * sends before getting here so by waiting for signaled
634 		 * sends to complete we're ensured that there will be no
635 		 * more tx processing.
636 		 */
637 		wait_event(rds_ib_ring_empty_wait,
638 			   rds_ib_ring_empty(&ic->i_recv_ring) &&
639 			   (atomic_read(&ic->i_signaled_sends) == 0));
640 		tasklet_kill(&ic->i_recv_tasklet);
641 
642 		if (ic->i_send_hdrs)
643 			ib_dma_free_coherent(dev,
644 					   ic->i_send_ring.w_nr *
645 						sizeof(struct rds_header),
646 					   ic->i_send_hdrs,
647 					   ic->i_send_hdrs_dma);
648 
649 		if (ic->i_recv_hdrs)
650 			ib_dma_free_coherent(dev,
651 					   ic->i_recv_ring.w_nr *
652 						sizeof(struct rds_header),
653 					   ic->i_recv_hdrs,
654 					   ic->i_recv_hdrs_dma);
655 
656 		if (ic->i_ack)
657 			ib_dma_free_coherent(dev, sizeof(struct rds_header),
658 					     ic->i_ack, ic->i_ack_dma);
659 
660 		if (ic->i_sends)
661 			rds_ib_send_clear_ring(ic);
662 		if (ic->i_recvs)
663 			rds_ib_recv_clear_ring(ic);
664 
665 		if (ic->i_cm_id->qp)
666 			rdma_destroy_qp(ic->i_cm_id);
667 		if (ic->i_send_cq)
668 			ib_destroy_cq(ic->i_send_cq);
669 		if (ic->i_recv_cq)
670 			ib_destroy_cq(ic->i_recv_cq);
671 		rdma_destroy_id(ic->i_cm_id);
672 
673 		/*
674 		 * Move connection back to the nodev list.
675 		 */
676 		if (ic->rds_ibdev)
677 			rds_ib_remove_conn(ic->rds_ibdev, conn);
678 
679 		ic->i_cm_id = NULL;
680 		ic->i_pd = NULL;
681 		ic->i_mr = NULL;
682 		ic->i_send_cq = NULL;
683 		ic->i_recv_cq = NULL;
684 		ic->i_send_hdrs = NULL;
685 		ic->i_recv_hdrs = NULL;
686 		ic->i_ack = NULL;
687 	}
688 	BUG_ON(ic->rds_ibdev);
689 
690 	/* Clear pending transmit */
691 	if (ic->i_data_op) {
692 		struct rds_message *rm;
693 
694 		rm = container_of(ic->i_data_op, struct rds_message, data);
695 		rds_message_put(rm);
696 		ic->i_data_op = NULL;
697 	}
698 
699 	/* Clear the ACK state */
700 	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
701 #ifdef KERNEL_HAS_ATOMIC64
702 	atomic64_set(&ic->i_ack_next, 0);
703 #else
704 	ic->i_ack_next = 0;
705 #endif
706 	ic->i_ack_recv = 0;
707 
708 	/* Clear flow control state */
709 	ic->i_flowctl = 0;
710 	atomic_set(&ic->i_credits, 0);
711 
712 	rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
713 	rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
714 
715 	if (ic->i_ibinc) {
716 		rds_inc_put(&ic->i_ibinc->ii_inc);
717 		ic->i_ibinc = NULL;
718 	}
719 
720 	vfree(ic->i_sends);
721 	ic->i_sends = NULL;
722 	vfree(ic->i_recvs);
723 	ic->i_recvs = NULL;
724 }
725 
726 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
727 {
728 	struct rds_ib_connection *ic;
729 	unsigned long flags;
730 	int ret;
731 
732 	/* XXX too lazy? */
733 	ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
734 	if (!ic)
735 		return -ENOMEM;
736 
737 	ret = rds_ib_recv_alloc_caches(ic);
738 	if (ret) {
739 		kfree(ic);
740 		return ret;
741 	}
742 
743 	INIT_LIST_HEAD(&ic->ib_node);
744 	tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
745 		     (unsigned long) ic);
746 	mutex_init(&ic->i_recv_mutex);
747 #ifndef KERNEL_HAS_ATOMIC64
748 	spin_lock_init(&ic->i_ack_lock);
749 #endif
750 	atomic_set(&ic->i_signaled_sends, 0);
751 
752 	/*
753 	 * rds_ib_conn_shutdown() waits for these to be emptied so they
754 	 * must be initialized before it can be called.
755 	 */
756 	rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
757 	rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
758 
759 	ic->conn = conn;
760 	conn->c_transport_data = ic;
761 
762 	spin_lock_irqsave(&ib_nodev_conns_lock, flags);
763 	list_add_tail(&ic->ib_node, &ib_nodev_conns);
764 	spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
765 
766 
767 	rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
768 	return 0;
769 }
770 
771 /*
772  * Free a connection. Connection must be shut down and not set for reconnect.
773  */
774 void rds_ib_conn_free(void *arg)
775 {
776 	struct rds_ib_connection *ic = arg;
777 	spinlock_t	*lock_ptr;
778 
779 	rdsdebug("ic %p\n", ic);
780 
781 	/*
782 	 * Conn is either on a dev's list or on the nodev list.
783 	 * A race with shutdown() or connect() would cause problems
784 	 * (since rds_ibdev would change) but that should never happen.
785 	 */
786 	lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
787 
788 	spin_lock_irq(lock_ptr);
789 	list_del(&ic->ib_node);
790 	spin_unlock_irq(lock_ptr);
791 
792 	rds_ib_recv_free_caches(ic);
793 
794 	kfree(ic);
795 }
796 
797 
798 /*
799  * An error occurred on the connection
800  */
801 void
802 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
803 {
804 	va_list ap;
805 
806 	rds_conn_drop(conn);
807 
808 	va_start(ap, fmt);
809 	vprintk(fmt, ap);
810 	va_end(ap);
811 }
812