xref: /openbmc/linux/net/rds/ib_cm.c (revision 52fb57e7)
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 static char *rds_ib_event_type_strings[] = {
43 #define RDS_IB_EVENT_STRING(foo) \
44 		[IB_EVENT_##foo] = __stringify(IB_EVENT_##foo)
45 	RDS_IB_EVENT_STRING(CQ_ERR),
46 	RDS_IB_EVENT_STRING(QP_FATAL),
47 	RDS_IB_EVENT_STRING(QP_REQ_ERR),
48 	RDS_IB_EVENT_STRING(QP_ACCESS_ERR),
49 	RDS_IB_EVENT_STRING(COMM_EST),
50 	RDS_IB_EVENT_STRING(SQ_DRAINED),
51 	RDS_IB_EVENT_STRING(PATH_MIG),
52 	RDS_IB_EVENT_STRING(PATH_MIG_ERR),
53 	RDS_IB_EVENT_STRING(DEVICE_FATAL),
54 	RDS_IB_EVENT_STRING(PORT_ACTIVE),
55 	RDS_IB_EVENT_STRING(PORT_ERR),
56 	RDS_IB_EVENT_STRING(LID_CHANGE),
57 	RDS_IB_EVENT_STRING(PKEY_CHANGE),
58 	RDS_IB_EVENT_STRING(SM_CHANGE),
59 	RDS_IB_EVENT_STRING(SRQ_ERR),
60 	RDS_IB_EVENT_STRING(SRQ_LIMIT_REACHED),
61 	RDS_IB_EVENT_STRING(QP_LAST_WQE_REACHED),
62 	RDS_IB_EVENT_STRING(CLIENT_REREGISTER),
63 #undef RDS_IB_EVENT_STRING
64 };
65 
66 static char *rds_ib_event_str(enum ib_event_type type)
67 {
68 	return rds_str_array(rds_ib_event_type_strings,
69 			     ARRAY_SIZE(rds_ib_event_type_strings), type);
70 };
71 
72 /*
73  * Set the selected protocol version
74  */
75 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
76 {
77 	conn->c_version = version;
78 }
79 
80 /*
81  * Set up flow control
82  */
83 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
84 {
85 	struct rds_ib_connection *ic = conn->c_transport_data;
86 
87 	if (rds_ib_sysctl_flow_control && credits != 0) {
88 		/* We're doing flow control */
89 		ic->i_flowctl = 1;
90 		rds_ib_send_add_credits(conn, credits);
91 	} else {
92 		ic->i_flowctl = 0;
93 	}
94 }
95 
96 /*
97  * Tune RNR behavior. Without flow control, we use a rather
98  * low timeout, but not the absolute minimum - this should
99  * be tunable.
100  *
101  * We already set the RNR retry count to 7 (which is the
102  * smallest infinite number :-) above.
103  * If flow control is off, we want to change this back to 0
104  * so that we learn quickly when our credit accounting is
105  * buggy.
106  *
107  * Caller passes in a qp_attr pointer - don't waste stack spacv
108  * by allocation this twice.
109  */
110 static void
111 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
112 {
113 	int ret;
114 
115 	attr->min_rnr_timer = IB_RNR_TIMER_000_32;
116 	ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
117 	if (ret)
118 		printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
119 }
120 
121 /*
122  * Connection established.
123  * We get here for both outgoing and incoming connection.
124  */
125 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
126 {
127 	const struct rds_ib_connect_private *dp = NULL;
128 	struct rds_ib_connection *ic = conn->c_transport_data;
129 	struct ib_qp_attr qp_attr;
130 	int err;
131 
132 	if (event->param.conn.private_data_len >= sizeof(*dp)) {
133 		dp = event->param.conn.private_data;
134 
135 		/* make sure it isn't empty data */
136 		if (dp->dp_protocol_major) {
137 			rds_ib_set_protocol(conn,
138 				RDS_PROTOCOL(dp->dp_protocol_major,
139 				dp->dp_protocol_minor));
140 			rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
141 		}
142 	}
143 
144 	if (conn->c_version < RDS_PROTOCOL(3,1)) {
145 		printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
146 		       " no longer supported\n",
147 		       &conn->c_faddr,
148 		       RDS_PROTOCOL_MAJOR(conn->c_version),
149 		       RDS_PROTOCOL_MINOR(conn->c_version));
150 		rds_conn_destroy(conn);
151 		return;
152 	} else {
153 		printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
154 		       &conn->c_faddr,
155 		       RDS_PROTOCOL_MAJOR(conn->c_version),
156 		       RDS_PROTOCOL_MINOR(conn->c_version),
157 		       ic->i_flowctl ? ", flow control" : "");
158 	}
159 
160 	/*
161 	 * Init rings and fill recv. this needs to wait until protocol negotiation
162 	 * is complete, since ring layout is different from 3.0 to 3.1.
163 	 */
164 	rds_ib_send_init_ring(ic);
165 	rds_ib_recv_init_ring(ic);
166 	/* Post receive buffers - as a side effect, this will update
167 	 * the posted credit count. */
168 	rds_ib_recv_refill(conn, 1);
169 
170 	/* Tune RNR behavior */
171 	rds_ib_tune_rnr(ic, &qp_attr);
172 
173 	qp_attr.qp_state = IB_QPS_RTS;
174 	err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
175 	if (err)
176 		printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
177 
178 	/* update ib_device with this local ipaddr */
179 	err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
180 	if (err)
181 		printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
182 			err);
183 
184 	/* If the peer gave us the last packet it saw, process this as if
185 	 * we had received a regular ACK. */
186 	if (dp) {
187 		/* dp structure start is not guaranteed to be 8 bytes aligned.
188 		 * Since dp_ack_seq is 64-bit extended load operations can be
189 		 * used so go through get_unaligned to avoid unaligned errors.
190 		 */
191 		__be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
192 
193 		if (dp_ack_seq)
194 			rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
195 					    NULL);
196 	}
197 
198 	rds_connect_complete(conn);
199 }
200 
201 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
202 			struct rdma_conn_param *conn_param,
203 			struct rds_ib_connect_private *dp,
204 			u32 protocol_version,
205 			u32 max_responder_resources,
206 			u32 max_initiator_depth)
207 {
208 	struct rds_ib_connection *ic = conn->c_transport_data;
209 	struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
210 
211 	memset(conn_param, 0, sizeof(struct rdma_conn_param));
212 
213 	conn_param->responder_resources =
214 		min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
215 	conn_param->initiator_depth =
216 		min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
217 	conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
218 	conn_param->rnr_retry_count = 7;
219 
220 	if (dp) {
221 		memset(dp, 0, sizeof(*dp));
222 		dp->dp_saddr = conn->c_laddr;
223 		dp->dp_daddr = conn->c_faddr;
224 		dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
225 		dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
226 		dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
227 		dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
228 
229 		/* Advertise flow control */
230 		if (ic->i_flowctl) {
231 			unsigned int credits;
232 
233 			credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
234 			dp->dp_credit = cpu_to_be32(credits);
235 			atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
236 		}
237 
238 		conn_param->private_data = dp;
239 		conn_param->private_data_len = sizeof(*dp);
240 	}
241 }
242 
243 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
244 {
245 	rdsdebug("event %u (%s) data %p\n",
246 		 event->event, rds_ib_event_str(event->event), data);
247 }
248 
249 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
250 {
251 	struct rds_connection *conn = data;
252 	struct rds_ib_connection *ic = conn->c_transport_data;
253 
254 	rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
255 		 rds_ib_event_str(event->event));
256 
257 	switch (event->event) {
258 	case IB_EVENT_COMM_EST:
259 		rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
260 		break;
261 	default:
262 		rdsdebug("Fatal QP Event %u (%s) "
263 			"- connection %pI4->%pI4, reconnecting\n",
264 			event->event, rds_ib_event_str(event->event),
265 			&conn->c_laddr, &conn->c_faddr);
266 		rds_conn_drop(conn);
267 		break;
268 	}
269 }
270 
271 /*
272  * This needs to be very careful to not leave IS_ERR pointers around for
273  * cleanup to trip over.
274  */
275 static int rds_ib_setup_qp(struct rds_connection *conn)
276 {
277 	struct rds_ib_connection *ic = conn->c_transport_data;
278 	struct ib_device *dev = ic->i_cm_id->device;
279 	struct ib_qp_init_attr attr;
280 	struct rds_ib_device *rds_ibdev;
281 	int ret;
282 
283 	/*
284 	 * It's normal to see a null device if an incoming connection races
285 	 * with device removal, so we don't print a warning.
286 	 */
287 	rds_ibdev = rds_ib_get_client_data(dev);
288 	if (!rds_ibdev)
289 		return -EOPNOTSUPP;
290 
291 	/* add the conn now so that connection establishment has the dev */
292 	rds_ib_add_conn(rds_ibdev, conn);
293 
294 	if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
295 		rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
296 	if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
297 		rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
298 
299 	/* Protection domain and memory range */
300 	ic->i_pd = rds_ibdev->pd;
301 	ic->i_mr = rds_ibdev->mr;
302 
303 	ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
304 				     rds_ib_cq_event_handler, conn,
305 				     ic->i_send_ring.w_nr + 1, 0);
306 	if (IS_ERR(ic->i_send_cq)) {
307 		ret = PTR_ERR(ic->i_send_cq);
308 		ic->i_send_cq = NULL;
309 		rdsdebug("ib_create_cq send failed: %d\n", ret);
310 		goto out;
311 	}
312 
313 	ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
314 				     rds_ib_cq_event_handler, conn,
315 				     ic->i_recv_ring.w_nr, 0);
316 	if (IS_ERR(ic->i_recv_cq)) {
317 		ret = PTR_ERR(ic->i_recv_cq);
318 		ic->i_recv_cq = NULL;
319 		rdsdebug("ib_create_cq recv failed: %d\n", ret);
320 		goto out;
321 	}
322 
323 	ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
324 	if (ret) {
325 		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
326 		goto out;
327 	}
328 
329 	ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
330 	if (ret) {
331 		rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
332 		goto out;
333 	}
334 
335 	/* XXX negotiate max send/recv with remote? */
336 	memset(&attr, 0, sizeof(attr));
337 	attr.event_handler = rds_ib_qp_event_handler;
338 	attr.qp_context = conn;
339 	/* + 1 to allow for the single ack message */
340 	attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
341 	attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
342 	attr.cap.max_send_sge = rds_ibdev->max_sge;
343 	attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
344 	attr.sq_sig_type = IB_SIGNAL_REQ_WR;
345 	attr.qp_type = IB_QPT_RC;
346 	attr.send_cq = ic->i_send_cq;
347 	attr.recv_cq = ic->i_recv_cq;
348 
349 	/*
350 	 * XXX this can fail if max_*_wr is too large?  Are we supposed
351 	 * to back off until we get a value that the hardware can support?
352 	 */
353 	ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
354 	if (ret) {
355 		rdsdebug("rdma_create_qp failed: %d\n", ret);
356 		goto out;
357 	}
358 
359 	ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
360 					   ic->i_send_ring.w_nr *
361 						sizeof(struct rds_header),
362 					   &ic->i_send_hdrs_dma, GFP_KERNEL);
363 	if (!ic->i_send_hdrs) {
364 		ret = -ENOMEM;
365 		rdsdebug("ib_dma_alloc_coherent send failed\n");
366 		goto out;
367 	}
368 
369 	ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
370 					   ic->i_recv_ring.w_nr *
371 						sizeof(struct rds_header),
372 					   &ic->i_recv_hdrs_dma, GFP_KERNEL);
373 	if (!ic->i_recv_hdrs) {
374 		ret = -ENOMEM;
375 		rdsdebug("ib_dma_alloc_coherent recv failed\n");
376 		goto out;
377 	}
378 
379 	ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
380 				       &ic->i_ack_dma, GFP_KERNEL);
381 	if (!ic->i_ack) {
382 		ret = -ENOMEM;
383 		rdsdebug("ib_dma_alloc_coherent ack failed\n");
384 		goto out;
385 	}
386 
387 	ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
388 				   ibdev_to_node(dev));
389 	if (!ic->i_sends) {
390 		ret = -ENOMEM;
391 		rdsdebug("send allocation failed\n");
392 		goto out;
393 	}
394 
395 	ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
396 				   ibdev_to_node(dev));
397 	if (!ic->i_recvs) {
398 		ret = -ENOMEM;
399 		rdsdebug("recv allocation failed\n");
400 		goto out;
401 	}
402 
403 	rds_ib_recv_init_ack(ic);
404 
405 	rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
406 		 ic->i_send_cq, ic->i_recv_cq);
407 
408 out:
409 	rds_ib_dev_put(rds_ibdev);
410 	return ret;
411 }
412 
413 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
414 {
415 	const struct rds_ib_connect_private *dp = event->param.conn.private_data;
416 	u16 common;
417 	u32 version = 0;
418 
419 	/*
420 	 * rdma_cm private data is odd - when there is any private data in the
421 	 * request, we will be given a pretty large buffer without telling us the
422 	 * original size. The only way to tell the difference is by looking at
423 	 * the contents, which are initialized to zero.
424 	 * If the protocol version fields aren't set, this is a connection attempt
425 	 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
426 	 * We really should have changed this for OFED 1.3 :-(
427 	 */
428 
429 	/* Be paranoid. RDS always has privdata */
430 	if (!event->param.conn.private_data_len) {
431 		printk(KERN_NOTICE "RDS incoming connection has no private data, "
432 			"rejecting\n");
433 		return 0;
434 	}
435 
436 	/* Even if len is crap *now* I still want to check it. -ASG */
437 	if (event->param.conn.private_data_len < sizeof (*dp) ||
438 	    dp->dp_protocol_major == 0)
439 		return RDS_PROTOCOL_3_0;
440 
441 	common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
442 	if (dp->dp_protocol_major == 3 && common) {
443 		version = RDS_PROTOCOL_3_0;
444 		while ((common >>= 1) != 0)
445 			version++;
446 	} else
447 		printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
448 				&dp->dp_saddr,
449 				dp->dp_protocol_major,
450 				dp->dp_protocol_minor);
451 	return version;
452 }
453 
454 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
455 				    struct rdma_cm_event *event)
456 {
457 	__be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
458 	__be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
459 	const struct rds_ib_connect_private *dp = event->param.conn.private_data;
460 	struct rds_ib_connect_private dp_rep;
461 	struct rds_connection *conn = NULL;
462 	struct rds_ib_connection *ic = NULL;
463 	struct rdma_conn_param conn_param;
464 	u32 version;
465 	int err = 1, destroy = 1;
466 
467 	/* Check whether the remote protocol version matches ours. */
468 	version = rds_ib_protocol_compatible(event);
469 	if (!version)
470 		goto out;
471 
472 	rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
473 		 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
474 		 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
475 		 (unsigned long long)be64_to_cpu(lguid),
476 		 (unsigned long long)be64_to_cpu(fguid));
477 
478 	conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
479 			       GFP_KERNEL);
480 	if (IS_ERR(conn)) {
481 		rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
482 		conn = NULL;
483 		goto out;
484 	}
485 
486 	/*
487 	 * The connection request may occur while the
488 	 * previous connection exist, e.g. in case of failover.
489 	 * But as connections may be initiated simultaneously
490 	 * by both hosts, we have a random backoff mechanism -
491 	 * see the comment above rds_queue_reconnect()
492 	 */
493 	mutex_lock(&conn->c_cm_lock);
494 	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
495 		if (rds_conn_state(conn) == RDS_CONN_UP) {
496 			rdsdebug("incoming connect while connecting\n");
497 			rds_conn_drop(conn);
498 			rds_ib_stats_inc(s_ib_listen_closed_stale);
499 		} else
500 		if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
501 			/* Wait and see - our connect may still be succeeding */
502 			rds_ib_stats_inc(s_ib_connect_raced);
503 		}
504 		goto out;
505 	}
506 
507 	ic = conn->c_transport_data;
508 
509 	rds_ib_set_protocol(conn, version);
510 	rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
511 
512 	/* If the peer gave us the last packet it saw, process this as if
513 	 * we had received a regular ACK. */
514 	if (dp->dp_ack_seq)
515 		rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
516 
517 	BUG_ON(cm_id->context);
518 	BUG_ON(ic->i_cm_id);
519 
520 	ic->i_cm_id = cm_id;
521 	cm_id->context = conn;
522 
523 	/* We got halfway through setting up the ib_connection, if we
524 	 * fail now, we have to take the long route out of this mess. */
525 	destroy = 0;
526 
527 	err = rds_ib_setup_qp(conn);
528 	if (err) {
529 		rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
530 		goto out;
531 	}
532 
533 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
534 		event->param.conn.responder_resources,
535 		event->param.conn.initiator_depth);
536 
537 	/* rdma_accept() calls rdma_reject() internally if it fails */
538 	err = rdma_accept(cm_id, &conn_param);
539 	if (err)
540 		rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
541 
542 out:
543 	if (conn)
544 		mutex_unlock(&conn->c_cm_lock);
545 	if (err)
546 		rdma_reject(cm_id, NULL, 0);
547 	return destroy;
548 }
549 
550 
551 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
552 {
553 	struct rds_connection *conn = cm_id->context;
554 	struct rds_ib_connection *ic = conn->c_transport_data;
555 	struct rdma_conn_param conn_param;
556 	struct rds_ib_connect_private dp;
557 	int ret;
558 
559 	/* If the peer doesn't do protocol negotiation, we must
560 	 * default to RDSv3.0 */
561 	rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
562 	ic->i_flowctl = rds_ib_sysctl_flow_control;	/* advertise flow control */
563 
564 	ret = rds_ib_setup_qp(conn);
565 	if (ret) {
566 		rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
567 		goto out;
568 	}
569 
570 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
571 		UINT_MAX, UINT_MAX);
572 	ret = rdma_connect(cm_id, &conn_param);
573 	if (ret)
574 		rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
575 
576 out:
577 	/* Beware - returning non-zero tells the rdma_cm to destroy
578 	 * the cm_id. We should certainly not do it as long as we still
579 	 * "own" the cm_id. */
580 	if (ret) {
581 		if (ic->i_cm_id == cm_id)
582 			ret = 0;
583 	}
584 	return ret;
585 }
586 
587 int rds_ib_conn_connect(struct rds_connection *conn)
588 {
589 	struct rds_ib_connection *ic = conn->c_transport_data;
590 	struct sockaddr_in src, dest;
591 	int ret;
592 
593 	/* XXX I wonder what affect the port space has */
594 	/* delegate cm event handler to rdma_transport */
595 	ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
596 				     RDMA_PS_TCP, IB_QPT_RC);
597 	if (IS_ERR(ic->i_cm_id)) {
598 		ret = PTR_ERR(ic->i_cm_id);
599 		ic->i_cm_id = NULL;
600 		rdsdebug("rdma_create_id() failed: %d\n", ret);
601 		goto out;
602 	}
603 
604 	rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
605 
606 	src.sin_family = AF_INET;
607 	src.sin_addr.s_addr = (__force u32)conn->c_laddr;
608 	src.sin_port = (__force u16)htons(0);
609 
610 	dest.sin_family = AF_INET;
611 	dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
612 	dest.sin_port = (__force u16)htons(RDS_PORT);
613 
614 	ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
615 				(struct sockaddr *)&dest,
616 				RDS_RDMA_RESOLVE_TIMEOUT_MS);
617 	if (ret) {
618 		rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
619 			 ret);
620 		rdma_destroy_id(ic->i_cm_id);
621 		ic->i_cm_id = NULL;
622 	}
623 
624 out:
625 	return ret;
626 }
627 
628 /*
629  * This is so careful about only cleaning up resources that were built up
630  * so that it can be called at any point during startup.  In fact it
631  * can be called multiple times for a given connection.
632  */
633 void rds_ib_conn_shutdown(struct rds_connection *conn)
634 {
635 	struct rds_ib_connection *ic = conn->c_transport_data;
636 	int err = 0;
637 
638 	rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
639 		 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
640 		 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
641 
642 	if (ic->i_cm_id) {
643 		struct ib_device *dev = ic->i_cm_id->device;
644 
645 		rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
646 		err = rdma_disconnect(ic->i_cm_id);
647 		if (err) {
648 			/* Actually this may happen quite frequently, when
649 			 * an outgoing connect raced with an incoming connect.
650 			 */
651 			rdsdebug("failed to disconnect, cm: %p err %d\n",
652 				ic->i_cm_id, err);
653 		}
654 
655 		/*
656 		 * We want to wait for tx and rx completion to finish
657 		 * before we tear down the connection, but we have to be
658 		 * careful not to get stuck waiting on a send ring that
659 		 * only has unsignaled sends in it.  We've shutdown new
660 		 * sends before getting here so by waiting for signaled
661 		 * sends to complete we're ensured that there will be no
662 		 * more tx processing.
663 		 */
664 		wait_event(rds_ib_ring_empty_wait,
665 			   rds_ib_ring_empty(&ic->i_recv_ring) &&
666 			   (atomic_read(&ic->i_signaled_sends) == 0));
667 		tasklet_kill(&ic->i_recv_tasklet);
668 
669 		if (ic->i_send_hdrs)
670 			ib_dma_free_coherent(dev,
671 					   ic->i_send_ring.w_nr *
672 						sizeof(struct rds_header),
673 					   ic->i_send_hdrs,
674 					   ic->i_send_hdrs_dma);
675 
676 		if (ic->i_recv_hdrs)
677 			ib_dma_free_coherent(dev,
678 					   ic->i_recv_ring.w_nr *
679 						sizeof(struct rds_header),
680 					   ic->i_recv_hdrs,
681 					   ic->i_recv_hdrs_dma);
682 
683 		if (ic->i_ack)
684 			ib_dma_free_coherent(dev, sizeof(struct rds_header),
685 					     ic->i_ack, ic->i_ack_dma);
686 
687 		if (ic->i_sends)
688 			rds_ib_send_clear_ring(ic);
689 		if (ic->i_recvs)
690 			rds_ib_recv_clear_ring(ic);
691 
692 		if (ic->i_cm_id->qp)
693 			rdma_destroy_qp(ic->i_cm_id);
694 		if (ic->i_send_cq)
695 			ib_destroy_cq(ic->i_send_cq);
696 		if (ic->i_recv_cq)
697 			ib_destroy_cq(ic->i_recv_cq);
698 		rdma_destroy_id(ic->i_cm_id);
699 
700 		/*
701 		 * Move connection back to the nodev list.
702 		 */
703 		if (ic->rds_ibdev)
704 			rds_ib_remove_conn(ic->rds_ibdev, conn);
705 
706 		ic->i_cm_id = NULL;
707 		ic->i_pd = NULL;
708 		ic->i_mr = NULL;
709 		ic->i_send_cq = NULL;
710 		ic->i_recv_cq = NULL;
711 		ic->i_send_hdrs = NULL;
712 		ic->i_recv_hdrs = NULL;
713 		ic->i_ack = NULL;
714 	}
715 	BUG_ON(ic->rds_ibdev);
716 
717 	/* Clear pending transmit */
718 	if (ic->i_data_op) {
719 		struct rds_message *rm;
720 
721 		rm = container_of(ic->i_data_op, struct rds_message, data);
722 		rds_message_put(rm);
723 		ic->i_data_op = NULL;
724 	}
725 
726 	/* Clear the ACK state */
727 	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
728 #ifdef KERNEL_HAS_ATOMIC64
729 	atomic64_set(&ic->i_ack_next, 0);
730 #else
731 	ic->i_ack_next = 0;
732 #endif
733 	ic->i_ack_recv = 0;
734 
735 	/* Clear flow control state */
736 	ic->i_flowctl = 0;
737 	atomic_set(&ic->i_credits, 0);
738 
739 	rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
740 	rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
741 
742 	if (ic->i_ibinc) {
743 		rds_inc_put(&ic->i_ibinc->ii_inc);
744 		ic->i_ibinc = NULL;
745 	}
746 
747 	vfree(ic->i_sends);
748 	ic->i_sends = NULL;
749 	vfree(ic->i_recvs);
750 	ic->i_recvs = NULL;
751 }
752 
753 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
754 {
755 	struct rds_ib_connection *ic;
756 	unsigned long flags;
757 	int ret;
758 
759 	/* XXX too lazy? */
760 	ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
761 	if (!ic)
762 		return -ENOMEM;
763 
764 	ret = rds_ib_recv_alloc_caches(ic);
765 	if (ret) {
766 		kfree(ic);
767 		return ret;
768 	}
769 
770 	INIT_LIST_HEAD(&ic->ib_node);
771 	tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
772 		     (unsigned long) ic);
773 	mutex_init(&ic->i_recv_mutex);
774 #ifndef KERNEL_HAS_ATOMIC64
775 	spin_lock_init(&ic->i_ack_lock);
776 #endif
777 	atomic_set(&ic->i_signaled_sends, 0);
778 
779 	/*
780 	 * rds_ib_conn_shutdown() waits for these to be emptied so they
781 	 * must be initialized before it can be called.
782 	 */
783 	rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
784 	rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
785 
786 	ic->conn = conn;
787 	conn->c_transport_data = ic;
788 
789 	spin_lock_irqsave(&ib_nodev_conns_lock, flags);
790 	list_add_tail(&ic->ib_node, &ib_nodev_conns);
791 	spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
792 
793 
794 	rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
795 	return 0;
796 }
797 
798 /*
799  * Free a connection. Connection must be shut down and not set for reconnect.
800  */
801 void rds_ib_conn_free(void *arg)
802 {
803 	struct rds_ib_connection *ic = arg;
804 	spinlock_t	*lock_ptr;
805 
806 	rdsdebug("ic %p\n", ic);
807 
808 	/*
809 	 * Conn is either on a dev's list or on the nodev list.
810 	 * A race with shutdown() or connect() would cause problems
811 	 * (since rds_ibdev would change) but that should never happen.
812 	 */
813 	lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
814 
815 	spin_lock_irq(lock_ptr);
816 	list_del(&ic->ib_node);
817 	spin_unlock_irq(lock_ptr);
818 
819 	rds_ib_recv_free_caches(ic);
820 
821 	kfree(ic);
822 }
823 
824 
825 /*
826  * An error occurred on the connection
827  */
828 void
829 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
830 {
831 	va_list ap;
832 
833 	rds_conn_drop(conn);
834 
835 	va_start(ap, fmt);
836 	vprintk(fmt, ap);
837 	va_end(ap);
838 }
839