1 /*
2  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
3  */
4 
5 /*
6  * This file contains code imported from the OFED rds source file ib_cm.c
7  * Oracle elects to have and use the contents of ib_cm.c under and governed
8  * by the OpenIB.org BSD license (see below for full license text). However,
9  * the following notice accompanied the original version of this file:
10  */
11 
12 /*
13  * Copyright (c) 2006 Oracle.  All rights reserved.
14  *
15  * This software is available to you under a choice of one of two
16  * licenses.  You may choose to be licensed under the terms of the GNU
17  * General Public License (GPL) Version 2, available from the file
18  * COPYING in the main directory of this source tree, or the
19  * OpenIB.org BSD license below:
20  *
21  *     Redistribution and use in source and binary forms, with or
22  *     without modification, are permitted provided that the following
23  *     conditions are met:
24  *
25  *      - Redistributions of source code must retain the above
26  *        copyright notice, this list of conditions and the following
27  *        disclaimer.
28  *
29  *      - Redistributions in binary form must reproduce the above
30  *        copyright notice, this list of conditions and the following
31  *        disclaimer in the documentation and/or other materials
32  *        provided with the distribution.
33  *
34  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41  * SOFTWARE.
42  *
43  */
44 #include <sys/rds.h>
45 
46 #include <sys/ib/clients/of/ofed_kernel.h>
47 #include <sys/ib/clients/of/rdma/ib_addr.h>
48 #include <sys/ib/clients/of/rdma/rdma_cm.h>
49 
50 #include <sys/ib/clients/rdsv3/rdsv3.h>
51 #include <sys/ib/clients/rdsv3/ib.h>
52 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
53 
54 extern int rdsv3_enable_snd_cq;
55 
56 /*
57  * Set the selected protocol version
58  */
59 static void
60 rdsv3_ib_set_protocol(struct rdsv3_connection *conn, unsigned int version)
61 {
62 	RDSV3_DPRINTF4("rdsv3_ib_set_protocol", "conn: %p version: %d",
63 	    conn, version);
64 	conn->c_version = version;
65 }
66 
67 /*
68  * Set up flow control
69  */
70 static void
71 rdsv3_ib_set_flow_control(struct rdsv3_connection *conn, uint32_t credits)
72 {
73 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
74 
75 	RDSV3_DPRINTF2("rdsv3_ib_set_flow_control",
76 	    "Enter: conn: %p credits: %d", conn, credits);
77 
78 	if (rdsv3_ib_sysctl_flow_control && credits != 0) {
79 		/* We're doing flow control */
80 		ic->i_flowctl = 1;
81 		rdsv3_ib_send_add_credits(conn, credits);
82 	} else {
83 		ic->i_flowctl = 0;
84 	}
85 
86 	RDSV3_DPRINTF2("rdsv3_ib_set_flow_control",
87 	    "Return: conn: %p credits: %d",
88 	    conn, credits);
89 }
90 
91 /*
92  * Tune RNR behavior. Without flow control, we use a rather
93  * low timeout, but not the absolute minimum - this should
94  * be tunable.
95  *
96  * We already set the RNR retry count to 7 (which is the
97  * smallest infinite number :-) above.
98  * If flow control is off, we want to change this back to 0
99  * so that we learn quickly when our credit accounting is
100  * buggy.
101  *
102  * Caller passes in a qp_attr pointer - don't waste stack spacv
103  * by allocation this twice.
104  */
105 static void
106 rdsv3_ib_tune_rnr(struct rdsv3_ib_connection *ic, struct ib_qp_attr *attr)
107 {
108 	int ret;
109 
110 	RDSV3_DPRINTF2("rdsv3_ib_tune_rnr", "Enter ic: %p attr: %p",
111 	    ic, attr);
112 
113 	attr->min_rnr_timer = IB_RNR_TIMER_000_32;
114 	ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
115 	if (ret)
116 		RDSV3_DPRINTF2("rdsv3_ib_tune_rnr",
117 		    "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d", -ret);
118 }
119 
120 /*
121  * Connection established.
122  * We get here for both outgoing and incoming connection.
123  */
124 void
125 rdsv3_ib_cm_connect_complete(struct rdsv3_connection *conn,
126     struct rdma_cm_event *event)
127 {
128 	const struct rdsv3_ib_connect_private *dp = NULL;
129 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
130 	struct rdsv3_ib_device *rds_ibdev =
131 	    ib_get_client_data(ic->i_cm_id->device, &rdsv3_ib_client);
132 	struct ib_qp_attr qp_attr;
133 	int err;
134 
135 	RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
136 	    "Enter conn: %p event: %p", conn, event);
137 
138 	if (event->param.conn.private_data_len >= sizeof (*dp)) {
139 		dp = event->param.conn.private_data;
140 
141 		/* make sure it isn't empty data */
142 		if (dp->dp_protocol_major) {
143 			rdsv3_ib_set_protocol(conn,
144 			    RDS_PROTOCOL(dp->dp_protocol_major,
145 			    dp->dp_protocol_minor));
146 			rdsv3_ib_set_flow_control(conn,
147 			    ntohl(dp->dp_credit));
148 		}
149 	}
150 
151 	if (conn->c_version < RDS_PROTOCOL(3, 1)) {
152 		RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
153 		    "RDS/IB: Connection to %u.%u.%u.%u version %u.%u failed",
154 		    NIPQUAD(conn->c_faddr),
155 		    RDS_PROTOCOL_MAJOR(conn->c_version),
156 		    RDS_PROTOCOL_MINOR(conn->c_version));
157 		rdsv3_conn_destroy(conn);
158 		return;
159 	} else {
160 		RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
161 		    "RDS/IB: connected to %u.%u.%u.%u version %u.%u%s",
162 		    NIPQUAD(conn->c_faddr),
163 		    RDS_PROTOCOL_MAJOR(conn->c_version),
164 		    RDS_PROTOCOL_MINOR(conn->c_version),
165 		    ic->i_flowctl ? ", flow control" : "");
166 	}
167 
168 	ASSERT(ic->i_soft_cq == NULL);
169 	ic->i_soft_cq = rdsv3_af_intr_thr_create(rdsv3_ib_tasklet_fn,
170 	    (void *)ic, SCQ_INTR_BIND_CPU, rds_ibdev->aft_hcagp,
171 	    ic->i_cq->ibt_cq);
172 	if (rdsv3_enable_snd_cq) {
173 		ic->i_snd_soft_cq = rdsv3_af_intr_thr_create(
174 		    rdsv3_ib_snd_tasklet_fn,
175 		    (void *)ic, SCQ_INTR_BIND_CPU, rds_ibdev->aft_hcagp,
176 		    ic->i_snd_cq->ibt_cq);
177 	}
178 	/* rdsv3_ib_refill_fn is expecting i_max_recv_alloc set */
179 	ic->i_max_recv_alloc = rdsv3_ib_sysctl_max_recv_allocation;
180 	ic->i_refill_rq = rdsv3_af_thr_create(rdsv3_ib_refill_fn, (void *)conn,
181 	    SCQ_WRK_BIND_CPU, rds_ibdev->aft_hcagp);
182 	rdsv3_af_grp_draw(rds_ibdev->aft_hcagp);
183 
184 	(void) ib_req_notify_cq(ic->i_cq, IB_CQ_SOLICITED);
185 	if (rdsv3_enable_snd_cq) {
186 		(void) ib_req_notify_cq(ic->i_snd_cq, IB_CQ_NEXT_COMP);
187 	}
188 
189 	/*
190 	 * Init rings and fill recv. this needs to wait until protocol
191 	 * negotiation
192 	 * is complete, since ring layout is different from 3.0 to 3.1.
193 	 */
194 	rdsv3_ib_send_init_ring(ic);
195 	rdsv3_ib_recv_init_ring(ic);
196 	/*
197 	 * Post receive buffers - as a side effect, this will update
198 	 * the posted credit count.
199 	 */
200 	(void) rdsv3_ib_recv_refill(conn, 1);
201 
202 	/* Tune RNR behavior */
203 	rdsv3_ib_tune_rnr(ic, &qp_attr);
204 
205 	qp_attr.qp_state = IB_QPS_RTS;
206 	err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
207 	if (err)
208 		RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
209 		    "ib_modify_qp(IB_QP_STATE, RTS): err=%d", err);
210 
211 	/* update ib_device with this local ipaddr & conn */
212 	err = rdsv3_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
213 	if (err)
214 		RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
215 		    "rdsv3_ib_update_ipaddr failed (%d)", err);
216 	rdsv3_ib_add_conn(rds_ibdev, conn);
217 
218 	/*
219 	 * If the peer gave us the last packet it saw, process this as if
220 	 * we had received a regular ACK.
221 	 */
222 	if (dp && dp->dp_ack_seq)
223 		rdsv3_send_drop_acked(conn, ntohll(dp->dp_ack_seq), NULL);
224 
225 	rdsv3_connect_complete(conn);
226 
227 	RDSV3_DPRINTF2("rdsv3_ib_cm_connect_complete",
228 	    "Return conn: %p event: %p",
229 	    conn, event);
230 }
231 
232 static void
233 rdsv3_ib_cm_fill_conn_param(struct rdsv3_connection *conn,
234     struct rdma_conn_param *conn_param,
235     struct rdsv3_ib_connect_private *dp,
236     uint32_t protocol_version,
237     uint32_t max_responder_resources,
238     uint32_t max_initiator_depth)
239 {
240 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
241 	struct rdsv3_ib_device *rds_ibdev;
242 
243 	RDSV3_DPRINTF2("rdsv3_ib_cm_fill_conn_param",
244 	    "Enter conn: %p conn_param: %p private: %p version: %d",
245 	    conn, conn_param, dp, protocol_version);
246 
247 	(void) memset(conn_param, 0, sizeof (struct rdma_conn_param));
248 
249 	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rdsv3_ib_client);
250 
251 	conn_param->responder_resources =
252 	    MIN(rds_ibdev->max_responder_resources, max_responder_resources);
253 	conn_param->initiator_depth =
254 	    MIN(rds_ibdev->max_initiator_depth, max_initiator_depth);
255 	conn_param->retry_count = min(rdsv3_ib_retry_count, 7);
256 	conn_param->rnr_retry_count = 7;
257 
258 	if (dp) {
259 		(void) memset(dp, 0, sizeof (*dp));
260 		dp->dp_saddr = conn->c_laddr;
261 		dp->dp_daddr = conn->c_faddr;
262 		dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
263 		dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
264 		dp->dp_protocol_minor_mask =
265 		    htons(RDSV3_IB_SUPPORTED_PROTOCOLS);
266 		dp->dp_ack_seq = rdsv3_ib_piggyb_ack(ic);
267 
268 		/* Advertise flow control */
269 		if (ic->i_flowctl) {
270 			unsigned int credits;
271 
272 			credits = IB_GET_POST_CREDITS(
273 			    atomic_get(&ic->i_credits));
274 			dp->dp_credit = htonl(credits);
275 			atomic_add_32(&ic->i_credits,
276 			    -IB_SET_POST_CREDITS(credits));
277 		}
278 
279 		conn_param->private_data = dp;
280 		conn_param->private_data_len = sizeof (*dp);
281 	}
282 
283 	RDSV3_DPRINTF2("rdsv3_ib_cm_fill_conn_param",
284 	    "Return conn: %p conn_param: %p private: %p version: %d",
285 	    conn, conn_param, dp, protocol_version);
286 }
287 
288 static void
289 rdsv3_ib_cq_event_handler(struct ib_event *event, void *data)
290 {
291 	RDSV3_DPRINTF3("rdsv3_ib_cq_event_handler", "event %u data %p",
292 	    event->event, data);
293 }
294 
295 static void
296 rdsv3_ib_snd_cq_comp_handler(struct ib_cq *cq, void *context)
297 {
298 	struct rdsv3_connection *conn = context;
299 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
300 
301 	RDSV3_DPRINTF4("rdsv3_ib_snd_cq_comp_handler",
302 	    "Enter(conn: %p ic: %p cq: %p)", conn, ic, cq);
303 
304 	rdsv3_af_thr_fire(ic->i_snd_soft_cq);
305 }
306 
307 void
308 rdsv3_ib_snd_tasklet_fn(void *data)
309 {
310 	struct rdsv3_ib_connection *ic = (struct rdsv3_ib_connection *)data;
311 	struct rdsv3_connection *conn = ic->conn;
312 	struct rdsv3_ib_ack_state ack_state = { 0, };
313 	ibt_wc_t wc;
314 	uint_t polled;
315 
316 	RDSV3_DPRINTF4("rdsv3_ib_snd_tasklet_fn",
317 	    "Enter(conn: %p ic: %p)", conn, ic);
318 
319 	/*
320 	 * Poll in a loop before and after enabling the next event
321 	 */
322 	while (ibt_poll_cq(RDSV3_CQ2CQHDL(ic->i_snd_cq), &wc, 1, &polled) ==
323 	    IBT_SUCCESS) {
324 		RDSV3_DPRINTF4("rdsv3_ib_snd_tasklet_fn",
325 		    "wc_id 0x%llx type %d status %u byte_len %u imm_data %u\n",
326 		    (unsigned long long)wc.wc_id, wc.wc_type, wc.wc_status,
327 		    wc.wc_bytes_xfer, ntohl(wc.wc_immed_data));
328 
329 		ASSERT(wc.wc_id & RDSV3_IB_SEND_OP);
330 		rdsv3_ib_send_cqe_handler(ic, &wc);
331 	}
332 	(void) ibt_enable_cq_notify(RDSV3_CQ2CQHDL(ic->i_snd_cq),
333 	    IBT_NEXT_COMPLETION);
334 	while (ibt_poll_cq(RDSV3_CQ2CQHDL(ic->i_snd_cq), &wc, 1, &polled) ==
335 	    IBT_SUCCESS) {
336 		RDSV3_DPRINTF4("rdsv3_ib_snd_tasklet_fn",
337 		    "wc_id 0x%llx type %d status %u byte_len %u imm_data %u\n",
338 		    (unsigned long long)wc.wc_id, wc.wc_type, wc.wc_status,
339 		    wc.wc_bytes_xfer, ntohl(wc.wc_immed_data));
340 
341 		ASSERT(wc.wc_id & RDSV3_IB_SEND_OP);
342 		rdsv3_ib_send_cqe_handler(ic, &wc);
343 	}
344 }
345 
346 static void
347 rdsv3_ib_cq_comp_handler(struct ib_cq *cq, void *context)
348 {
349 	struct rdsv3_connection *conn = context;
350 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
351 
352 	RDSV3_DPRINTF4("rdsv3_ib_cq_comp_handler",
353 	    "Enter(conn: %p cq: %p)", conn, cq);
354 
355 	rdsv3_ib_stats_inc(s_ib_evt_handler_call);
356 
357 	rdsv3_af_thr_fire(ic->i_soft_cq);
358 }
359 
360 void
361 rdsv3_ib_refill_fn(void *data)
362 {
363 	struct rdsv3_connection *conn = (struct rdsv3_connection *)data;
364 
365 	(void) rdsv3_ib_recv_refill(conn, 0);
366 }
367 
368 void
369 rdsv3_ib_tasklet_fn(void *data)
370 {
371 	struct rdsv3_ib_connection *ic = (struct rdsv3_ib_connection *)data;
372 	struct rdsv3_connection *conn = ic->conn;
373 	struct rdsv3_ib_ack_state ack_state = { 0, };
374 	ibt_wc_t wc[RDSV3_IB_WC_POLL_SIZE];
375 	uint_t polled;
376 	int i;
377 
378 	RDSV3_DPRINTF4("rdsv3_ib_tasklet_fn",
379 	    "Enter(conn: %p ic: %p)", conn, ic);
380 
381 	rdsv3_ib_stats_inc(s_ib_tasklet_call);
382 
383 	/*
384 	 * Poll in a loop before and after enabling the next event
385 	 */
386 	while (ibt_poll_cq(RDSV3_CQ2CQHDL(ic->i_cq), &wc[0],
387 	    RDSV3_IB_WC_POLL_SIZE, &polled) == IBT_SUCCESS) {
388 		for (i = 0; i < polled; i++) {
389 			RDSV3_DPRINTF4("rdsv3_ib_tasklet_fn",
390 			"wc_id 0x%llx type %d status %u byte_len %u \
391 			    imm_data %u\n",
392 			    (unsigned long long)wc[i].wc_id, wc[i].wc_type,
393 			    wc[i].wc_status, wc[i].wc_bytes_xfer,
394 			    ntohl(wc[i].wc_immed_data));
395 
396 			if (wc[i].wc_id & RDSV3_IB_SEND_OP) {
397 				rdsv3_ib_send_cqe_handler(ic, &wc[i]);
398 			} else {
399 				rdsv3_ib_recv_cqe_handler(ic, &wc[i],
400 				    &ack_state);
401 			}
402 		}
403 	}
404 	(void) ibt_enable_cq_notify(RDSV3_CQ2CQHDL(ic->i_cq),
405 	    IBT_NEXT_SOLICITED);
406 	while (ibt_poll_cq(RDSV3_CQ2CQHDL(ic->i_cq), &wc[0],
407 	    RDSV3_IB_WC_POLL_SIZE, &polled) == IBT_SUCCESS) {
408 		for (i = 0; i < polled; i++) {
409 			RDSV3_DPRINTF4("rdsv3_ib_tasklet_fn",
410 			"wc_id 0x%llx type %d status %u byte_len %u \
411 			    imm_data %u\n",
412 			    (unsigned long long)wc[i].wc_id, wc[i].wc_type,
413 			    wc[i].wc_status, wc[i].wc_bytes_xfer,
414 			    ntohl(wc[i].wc_immed_data));
415 
416 			if (wc[i].wc_id & RDSV3_IB_SEND_OP) {
417 				rdsv3_ib_send_cqe_handler(ic, &wc[i]);
418 			} else {
419 				rdsv3_ib_recv_cqe_handler(ic, &wc[i],
420 				    &ack_state);
421 			}
422 		}
423 	}
424 
425 	if (ack_state.ack_next_valid) {
426 		rdsv3_ib_set_ack(ic, ack_state.ack_next,
427 		    ack_state.ack_required);
428 	}
429 	if (ack_state.ack_recv_valid && ack_state.ack_recv > ic->i_ack_recv) {
430 		rdsv3_send_drop_acked(conn, ack_state.ack_recv, NULL);
431 		ic->i_ack_recv = ack_state.ack_recv;
432 	}
433 	if (rdsv3_conn_up(conn)) {
434 		if (!test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags))
435 			(void) rdsv3_send_xmit(ic->conn);
436 		rdsv3_ib_attempt_ack(ic);
437 	}
438 }
439 
440 static void
441 rdsv3_ib_qp_event_handler(struct ib_event *event, void *data)
442 {
443 	struct rdsv3_connection *conn = data;
444 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
445 
446 	RDSV3_DPRINTF2("rdsv3_ib_qp_event_handler", "conn %p ic %p event %u",
447 	    conn, ic, event->event);
448 
449 	switch (event->event) {
450 	case IB_EVENT_COMM_EST:
451 		(void) rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
452 		break;
453 	default:
454 		if (conn) {
455 			RDSV3_DPRINTF2("rdsv3_ib_qp_event_handler",
456 			    "RDS/IB: Fatal QP Event %u - "
457 			    "connection %u.%u.%u.%u ->%u.%u.%u.%u "
458 			    "...reconnecting",
459 			    event->event, NIPQUAD(conn->c_laddr),
460 			    NIPQUAD(conn->c_faddr));
461 			rdsv3_conn_drop(conn);
462 		} else {
463 			RDSV3_DPRINTF2("rdsv3_ib_qp_event_handler",
464 			    "RDS/IB: Fatal QP Event %u - connection"
465 			    "...reconnecting", event->event);
466 		}
467 		break;
468 	}
469 
470 	RDSV3_DPRINTF2("rdsv3_ib_qp_event_handler", "Return conn: %p event: %p",
471 	    conn, event);
472 }
473 
474 extern int rdsv3_ib_alloc_hdrs(ib_device_t *dev,
475     struct rdsv3_ib_connection *ic);
476 extern void rdsv3_ib_free_hdrs(ib_device_t *dev,
477     struct rdsv3_ib_connection *ic);
478 
479 /*
480  * This needs to be very careful to not leave IS_ERR pointers around for
481  * cleanup to trip over.
482  */
483 static int
484 rdsv3_ib_setup_qp(struct rdsv3_connection *conn)
485 {
486 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
487 	struct ib_device *dev = ic->i_cm_id->device;
488 	struct ib_qp_init_attr attr;
489 	struct rdsv3_ib_device *rds_ibdev;
490 	ibt_send_wr_t *wrp;
491 	ibt_wr_ds_t *sgl;
492 	int ret, i;
493 
494 	RDSV3_DPRINTF2("rdsv3_ib_setup_qp", "Enter conn: %p", conn);
495 
496 	/*
497 	 * rdsv3_ib_add_one creates a rdsv3_ib_device object per IB device,
498 	 * and allocates a protection domain, memory range and FMR pool
499 	 * for each.  If that fails for any reason, it will not register
500 	 * the rds_ibdev at all.
501 	 */
502 	rds_ibdev = ib_get_client_data(dev, &rdsv3_ib_client);
503 	if (!rds_ibdev) {
504 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
505 		    "RDS/IB: No client_data for device %s", dev->name);
506 		return (-EOPNOTSUPP);
507 	}
508 	ic->rds_ibdev = rds_ibdev;
509 
510 	if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
511 		rdsv3_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
512 	if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
513 		rdsv3_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
514 
515 	/* Protection domain and memory range */
516 	ic->i_pd = rds_ibdev->pd;
517 
518 	/*
519 	 * IB_CQ_VECTOR_LEAST_ATTACHED and/or the corresponding feature is
520 	 * not implmeneted in Hermon yet, but we can pass it to ib_create_cq()
521 	 * anyway.
522 	 */
523 	ic->i_cq = ib_create_cq(dev, rdsv3_ib_cq_comp_handler,
524 	    rdsv3_ib_cq_event_handler, conn,
525 	    ic->i_recv_ring.w_nr + ic->i_send_ring.w_nr + 1,
526 	    rdsv3_af_grp_get_sched(ic->rds_ibdev->aft_hcagp));
527 	if (IS_ERR(ic->i_cq)) {
528 		ret = PTR_ERR(ic->i_cq);
529 		ic->i_cq = NULL;
530 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
531 		    "ib_create_cq failed: %d", ret);
532 		goto out;
533 	}
534 	if (rdsv3_enable_snd_cq) {
535 		ic->i_snd_cq = ib_create_cq(dev, rdsv3_ib_snd_cq_comp_handler,
536 		    rdsv3_ib_cq_event_handler, conn, ic->i_send_ring.w_nr + 1,
537 		    rdsv3_af_grp_get_sched(ic->rds_ibdev->aft_hcagp));
538 		if (IS_ERR(ic->i_snd_cq)) {
539 			ret = PTR_ERR(ic->i_snd_cq);
540 			(void) ib_destroy_cq(ic->i_cq);
541 			ic->i_cq = NULL;
542 			ic->i_snd_cq = NULL;
543 			RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
544 			    "ib_create_cq send cq failed: %d", ret);
545 			goto out;
546 		}
547 	}
548 
549 	/* XXX negotiate max send/recv with remote? */
550 	(void) memset(&attr, 0, sizeof (attr));
551 	attr.event_handler = rdsv3_ib_qp_event_handler;
552 	attr.qp_context = conn;
553 	/* + 1 to allow for the single ack message */
554 	attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
555 	attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
556 	attr.cap.max_send_sge = rds_ibdev->max_sge;
557 	attr.cap.max_recv_sge = RDSV3_IB_RECV_SGE;
558 	attr.sq_sig_type = IB_SIGNAL_REQ_WR;
559 	attr.qp_type = IB_QPT_RC;
560 	if (rdsv3_enable_snd_cq) {
561 		attr.send_cq = ic->i_snd_cq;
562 	} else {
563 		attr.send_cq = ic->i_cq;
564 	}
565 	attr.recv_cq = ic->i_cq;
566 
567 	/*
568 	 * XXX this can fail if max_*_wr is too large?  Are we supposed
569 	 * to back off until we get a value that the hardware can support?
570 	 */
571 	ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
572 	if (ret) {
573 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
574 		    "rdma_create_qp failed: %d", ret);
575 		goto out;
576 	}
577 
578 	ret = rdsv3_ib_alloc_hdrs(dev, ic);
579 	if (ret != 0) {
580 		ret = -ENOMEM;
581 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
582 		    "rdsv3_ib_alloc_hdrs failed: %d", ret);
583 		goto out;
584 	}
585 
586 	ic->i_sends = kmem_alloc(ic->i_send_ring.w_nr *
587 	    sizeof (struct rdsv3_ib_send_work), KM_NOSLEEP);
588 	if (ic->i_sends == NULL) {
589 		ret = -ENOMEM;
590 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
591 		    "send allocation failed: %d", ret);
592 		goto out;
593 	}
594 	(void) memset(ic->i_sends, 0, ic->i_send_ring.w_nr *
595 	    sizeof (struct rdsv3_ib_send_work));
596 
597 	ic->i_send_wrs =
598 	    kmem_alloc(ic->i_send_ring.w_nr * (sizeof (ibt_send_wr_t) +
599 	    RDSV3_IB_MAX_SGE * sizeof (ibt_wr_ds_t)), KM_NOSLEEP);
600 	if (ic->i_send_wrs == NULL) {
601 		ret = -ENOMEM;
602 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
603 		    "Send WR allocation failed: %d", ret);
604 		goto out;
605 	}
606 	sgl = (ibt_wr_ds_t *)((uint8_t *)ic->i_send_wrs +
607 	    (ic->i_send_ring.w_nr * sizeof (ibt_send_wr_t)));
608 	for (i = 0; i < ic->i_send_ring.w_nr; i++) {
609 		wrp = &ic->i_send_wrs[i];
610 		wrp->wr_sgl = &sgl[i * RDSV3_IB_MAX_SGE];
611 	}
612 
613 	ic->i_recvs = kmem_alloc(ic->i_recv_ring.w_nr *
614 	    sizeof (struct rdsv3_ib_recv_work), KM_NOSLEEP);
615 	if (ic->i_recvs == NULL) {
616 		ret = -ENOMEM;
617 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
618 		    "recv allocation failed: %d", ret);
619 		goto out;
620 	}
621 	(void) memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr *
622 	    sizeof (struct rdsv3_ib_recv_work));
623 
624 	ic->i_recv_wrs =
625 	    kmem_alloc(ic->i_recv_ring.w_nr * sizeof (ibt_recv_wr_t),
626 	    KM_NOSLEEP);
627 	if (ic->i_recv_wrs == NULL) {
628 		ret = -ENOMEM;
629 		RDSV3_DPRINTF2("rdsv3_ib_setup_qp",
630 		    "Recv WR allocation failed: %d", ret);
631 		goto out;
632 	}
633 
634 	rdsv3_ib_recv_init_ack(ic);
635 
636 	RDSV3_DPRINTF2("rdsv3_ib_setup_qp", "conn %p pd %p mr %p cq %p",
637 	    conn, ic->i_pd, ic->i_mr, ic->i_cq);
638 
639 out:
640 	return (ret);
641 }
642 
643 static uint32_t
644 rdsv3_ib_protocol_compatible(struct rdma_cm_event *event)
645 {
646 	const struct rdsv3_ib_connect_private *dp =
647 	    event->param.conn.private_data;
648 	uint16_t common;
649 	uint32_t version = 0;
650 
651 	RDSV3_DPRINTF2("rdsv3_ib_protocol_compatible", "Enter event: %p",
652 	    event);
653 
654 	/*
655 	 * rdma_cm private data is odd - when there is any private data in the
656 	 * request, we will be given a pretty large buffer without telling us
657 	 * the
658 	 * original size. The only way to tell the difference is by looking at
659 	 * the contents, which are initialized to zero.
660 	 * If the protocol version fields aren't set,
661 	 * this is a connection attempt
662 	 * from an older version. This could could be 3.0 or 2.0 -
663 	 * we can't tell.
664 	 * We really should have changed this for OFED 1.3 :-(
665 	 */
666 
667 	/* Be paranoid. RDS always has privdata */
668 	if (!event->param.conn.private_data_len) {
669 		RDSV3_DPRINTF2("rdsv3_ib_protocol_compatible",
670 		    "RDS incoming connection has no private data, rejecting");
671 		return (0);
672 	}
673 
674 	/* Even if len is crap *now* I still want to check it. -ASG */
675 	if (event->param.conn.private_data_len < sizeof (*dp) ||
676 	    dp->dp_protocol_major == 0)
677 		return (RDS_PROTOCOL_3_0);
678 
679 	common = ntohs(dp->dp_protocol_minor_mask) &
680 	    RDSV3_IB_SUPPORTED_PROTOCOLS;
681 	if (dp->dp_protocol_major == 3 && common) {
682 		version = RDS_PROTOCOL_3_0;
683 		while ((common >>= 1) != 0)
684 			version++;
685 	} else {
686 		RDSV3_DPRINTF2("rdsv3_ib_protocol_compatible",
687 		    "RDS: Connection from %u.%u.%u.%u using "
688 		    "incompatible protocol version %u.%u\n",
689 		    NIPQUAD(dp->dp_saddr),
690 		    dp->dp_protocol_major,
691 		    dp->dp_protocol_minor);
692 	}
693 
694 	RDSV3_DPRINTF2("rdsv3_ib_protocol_compatible", "Return event: %p",
695 	    event);
696 
697 	return (version);
698 }
699 
700 int
701 rdsv3_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
702     struct rdma_cm_event *event)
703 {
704 	uint64_be_t lguid = cm_id->route.path_rec->sgid.global.interface_id;
705 	uint64_be_t fguid = cm_id->route.path_rec->dgid.global.interface_id;
706 	const struct rdsv3_ib_connect_private *dp =
707 	    event->param.conn.private_data;
708 	struct rdsv3_ib_connect_private dp_rep;
709 	struct rdsv3_connection *conn = NULL;
710 	struct rdsv3_ib_connection *ic = NULL;
711 	struct rdma_conn_param conn_param;
712 	uint32_t version;
713 	int err, destroy = 1;
714 	boolean_t conn_created = B_FALSE;
715 
716 	RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
717 	    "Enter cm_id: %p event: %p", cm_id, event);
718 
719 	/* Check whether the remote protocol version matches ours. */
720 	version = rdsv3_ib_protocol_compatible(event);
721 	if (!version) {
722 		RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
723 		    "version mismatch");
724 		goto out;
725 	}
726 
727 	RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
728 	    "saddr %u.%u.%u.%u daddr %u.%u.%u.%u RDSv%d.%d lguid 0x%llx fguid "
729 	    "0x%llx", NIPQUAD(dp->dp_saddr), NIPQUAD(dp->dp_daddr),
730 	    RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
731 	    (unsigned long long)ntohll(lguid),
732 	    (unsigned long long)ntohll(fguid));
733 
734 	conn = rdsv3_conn_create(dp->dp_daddr, dp->dp_saddr,
735 	    &rdsv3_ib_transport, KM_NOSLEEP);
736 	if (IS_ERR(conn)) {
737 		RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
738 		    "rdsv3_conn_create failed (%ld)", PTR_ERR(conn));
739 		conn = NULL;
740 		goto out;
741 	}
742 
743 	/*
744 	 * The connection request may occur while the
745 	 * previous connection exist, e.g. in case of failover.
746 	 * But as connections may be initiated simultaneously
747 	 * by both hosts, we have a random backoff mechanism -
748 	 * see the comment above rdsv3_queue_reconnect()
749 	 */
750 	mutex_enter(&conn->c_cm_lock);
751 	if (!rdsv3_conn_transition(conn, RDSV3_CONN_DOWN,
752 	    RDSV3_CONN_CONNECTING)) {
753 		if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
754 			RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
755 			    "incoming connect when connected: %p",
756 			    conn);
757 			rdsv3_conn_drop(conn);
758 			rdsv3_ib_stats_inc(s_ib_listen_closed_stale);
759 			mutex_exit(&conn->c_cm_lock);
760 			goto out;
761 		} else if (rdsv3_conn_state(conn) == RDSV3_CONN_CONNECTING) {
762 			/* Wait and see - our connect may still be succeeding */
763 			RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
764 			    "peer-to-peer connection request: %p, "
765 			    "lguid: 0x%llx fguid: 0x%llx",
766 			    conn, lguid, fguid);
767 			rdsv3_ib_stats_inc(s_ib_connect_raced);
768 		}
769 		mutex_exit(&conn->c_cm_lock);
770 		goto out;
771 	}
772 
773 	ic = conn->c_transport_data;
774 
775 	rdsv3_ib_set_protocol(conn, version);
776 	rdsv3_ib_set_flow_control(conn, ntohl(dp->dp_credit));
777 
778 	/*
779 	 * If the peer gave us the last packet it saw, process this as if
780 	 * we had received a regular ACK.
781 	 */
782 	if (dp->dp_ack_seq)
783 		rdsv3_send_drop_acked(conn, ntohll(dp->dp_ack_seq), NULL);
784 
785 	ASSERT(!cm_id->context);
786 	ASSERT(!ic->i_cm_id);
787 
788 	if (ic->i_cm_id != NULL)
789 		RDSV3_PANIC();
790 
791 	ic->i_cm_id = cm_id;
792 	cm_id->context = conn;
793 
794 	/*
795 	 * We got halfway through setting up the ib_connection, if we
796 	 * fail now, we have to take the long route out of this mess.
797 	 */
798 	destroy = 0;
799 
800 	err = rdsv3_ib_setup_qp(conn);
801 	if (err) {
802 		RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
803 		    "rdsv3_ib_setup_qp failed (%d)", err);
804 		mutex_exit(&conn->c_cm_lock);
805 		rdsv3_conn_drop(conn);
806 		goto out;
807 	}
808 
809 	rdsv3_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
810 	    event->param.conn.responder_resources,
811 	    event->param.conn.initiator_depth);
812 
813 	/* rdma_accept() calls rdma_reject() internally if it fails */
814 	err = rdma_accept(cm_id, &conn_param);
815 	mutex_exit(&conn->c_cm_lock);
816 	if (err) {
817 		RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
818 		    "rdma_accept failed (%d)", err);
819 		rdsv3_conn_drop(conn);
820 		goto out;
821 	}
822 
823 	RDSV3_DPRINTF2("rdsv3_ib_cm_handle_connect",
824 	    "Return cm_id: %p event: %p", cm_id, event);
825 
826 	return (0);
827 
828 out:
829 	(void) rdma_reject(cm_id, NULL, 0);
830 	return (destroy);
831 }
832 
833 
834 int
835 rdsv3_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
836 {
837 	struct rdsv3_connection *conn = cm_id->context;
838 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
839 	struct rdma_conn_param conn_param;
840 	struct rdsv3_ib_connect_private dp;
841 	int ret;
842 
843 	RDSV3_DPRINTF2("rdsv3_ib_cm_initiate_connect", "Enter: cm_id: %p",
844 	    cm_id);
845 
846 	/*
847 	 * If the peer doesn't do protocol negotiation, we must
848 	 * default to RDSv3.0
849 	 */
850 	rdsv3_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
851 	ic->i_flowctl =
852 	    rdsv3_ib_sysctl_flow_control;	/* advertise flow control */
853 
854 	ret = rdsv3_ib_setup_qp(conn);
855 	if (ret) {
856 		RDSV3_DPRINTF2("rdsv3_ib_cm_initiate_connect",
857 		    "rdsv3_ib_setup_qp failed (%d)", ret);
858 		rdsv3_conn_drop(conn);
859 		goto out;
860 	}
861 
862 	rdsv3_ib_cm_fill_conn_param(conn, &conn_param, &dp,
863 	    RDS_PROTOCOL_VERSION, UINT_MAX, UINT_MAX);
864 
865 	ret = rdma_connect(cm_id, &conn_param);
866 	if (ret) {
867 		RDSV3_DPRINTF2("rdsv3_ib_cm_initiate_connect",
868 		    "rdma_connect failed (%d)", ret);
869 		rdsv3_conn_drop(conn);
870 	}
871 
872 	RDSV3_DPRINTF2("rdsv3_ib_cm_initiate_connect",
873 	    "Return: cm_id: %p", cm_id);
874 
875 out:
876 	/*
877 	 * Beware - returning non-zero tells the rdma_cm to destroy
878 	 * the cm_id. We should certainly not do it as long as we still
879 	 * "own" the cm_id.
880 	 */
881 	if (ret) {
882 		if (ic->i_cm_id == cm_id)
883 			ret = 0;
884 	}
885 	return (ret);
886 }
887 
888 int
889 rdsv3_ib_conn_connect(struct rdsv3_connection *conn)
890 {
891 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
892 	struct sockaddr_in src, dest;
893 	ipaddr_t	laddr, faddr;
894 	int ret;
895 
896 	RDSV3_DPRINTF2("rdsv3_ib_conn_connect", "Enter: conn: %p", conn);
897 
898 	/*
899 	 * XXX I wonder what affect the port space has
900 	 */
901 	/* delegate cm event handler to rdma_transport */
902 	ic->i_cm_id = rdma_create_id(rdsv3_rdma_cm_event_handler, conn,
903 	    RDMA_PS_TCP);
904 	if (IS_ERR(ic->i_cm_id)) {
905 		ret = PTR_ERR(ic->i_cm_id);
906 		ic->i_cm_id = NULL;
907 		RDSV3_DPRINTF2("rdsv3_ib_conn_connect",
908 		    "rdma_create_id() failed: %d", ret);
909 		goto out;
910 	}
911 
912 	RDSV3_DPRINTF3("rdsv3_ib_conn_connect",
913 	    "created cm id %p for conn %p", ic->i_cm_id, conn);
914 
915 	/* The ipaddr should be in the network order */
916 	laddr = conn->c_laddr;
917 	faddr = conn->c_faddr;
918 	ret = rdsv3_sc_path_lookup(&laddr, &faddr);
919 	if (ret == 0) {
920 		RDSV3_DPRINTF2(LABEL, "Path not found (0x%x 0x%x)",
921 		    ntohl(laddr), ntohl(faddr));
922 	}
923 
924 	src.sin_family = AF_INET;
925 	src.sin_addr.s_addr = (uint32_t)laddr;
926 	src.sin_port = (uint16_t)htons(0);
927 
928 	dest.sin_family = AF_INET;
929 	dest.sin_addr.s_addr = (uint32_t)faddr;
930 	dest.sin_port = (uint16_t)htons(RDSV3_PORT);
931 
932 	ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
933 	    (struct sockaddr *)&dest,
934 	    RDSV3_RDMA_RESOLVE_TIMEOUT_MS);
935 	if (ret) {
936 		RDSV3_DPRINTF2("rdsv3_ib_conn_connect",
937 		    "addr resolve failed for cm id %p: %d", ic->i_cm_id, ret);
938 		rdma_destroy_id(ic->i_cm_id);
939 		ic->i_cm_id = NULL;
940 	}
941 
942 	RDSV3_DPRINTF2("rdsv3_ib_conn_connect", "Return: conn: %p", conn);
943 
944 out:
945 	return (ret);
946 }
947 
948 /*
949  * This is so careful about only cleaning up resources that were built up
950  * so that it can be called at any point during startup.  In fact it
951  * can be called multiple times for a given connection.
952  */
953 void
954 rdsv3_ib_conn_shutdown(struct rdsv3_connection *conn)
955 {
956 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
957 	int err = 0;
958 
959 	RDSV3_DPRINTF2("rdsv3_ib_conn_shutdown",
960 	    "cm %p pd %p cq %p qp %p", ic->i_cm_id,
961 	    ic->i_pd, ic->i_cq, ic->i_cm_id ? ic->i_cm_id->qp : NULL);
962 
963 	if (ic->i_cm_id) {
964 		struct ib_device *dev = ic->i_cm_id->device;
965 
966 		RDSV3_DPRINTF2("rdsv3_ib_conn_shutdown",
967 		    "disconnecting cm %p", ic->i_cm_id);
968 		err = rdma_disconnect(ic->i_cm_id);
969 		if (err) {
970 			/*
971 			 * Actually this may happen quite frequently, when
972 			 * an outgoing connect raced with an incoming connect.
973 			 */
974 			RDSV3_DPRINTF2("rdsv3_ib_conn_shutdown",
975 			    "failed to disconnect, cm: %p err %d",
976 			    ic->i_cm_id, err);
977 		}
978 
979 		if (ic->i_cm_id->qp) {
980 			(void) ibt_flush_qp(
981 			    ib_get_ibt_channel_hdl(ic->i_cm_id));
982 			/*
983 			 * Don't wait for the send ring to be empty -- there
984 			 * may be completed non-signaled entries sitting on
985 			 * there. We unmap these below.
986 			 */
987 			rdsv3_wait_event(&ic->i_recv_ring.w_empty_wait,
988 			    rdsv3_ib_ring_empty(&ic->i_recv_ring));
989 			/*
990 			 * Note that Linux original code calls
991 			 * rdma_destroy_qp() after rdsv3_ib_recv_clear_ring(ic).
992 			 */
993 			rdma_destroy_qp(ic->i_cm_id);
994 		}
995 
996 		if (rdsv3_enable_snd_cq) {
997 			if (ic->i_snd_soft_cq) {
998 				rdsv3_af_thr_destroy(ic->i_snd_soft_cq);
999 				ic->i_snd_soft_cq = NULL;
1000 			}
1001 			if (ic->i_snd_cq)
1002 				(void) ib_destroy_cq(ic->i_snd_cq);
1003 		}
1004 		if (ic->i_soft_cq) {
1005 			rdsv3_af_thr_destroy(ic->i_soft_cq);
1006 			ic->i_soft_cq = NULL;
1007 		}
1008 		if (ic->i_refill_rq) {
1009 			rdsv3_af_thr_destroy(ic->i_refill_rq);
1010 			ic->i_refill_rq = NULL;
1011 		}
1012 		if (ic->i_cq)
1013 			(void) ib_destroy_cq(ic->i_cq);
1014 
1015 		if (ic->i_mr)
1016 			rdsv3_ib_free_hdrs(dev, ic);
1017 
1018 		if (ic->i_sends)
1019 			rdsv3_ib_send_clear_ring(ic);
1020 		if (ic->i_recvs)
1021 			rdsv3_ib_recv_clear_ring(ic);
1022 
1023 		rdma_destroy_id(ic->i_cm_id);
1024 
1025 		/*
1026 		 * Move connection back to the nodev list.
1027 		 */
1028 		if (ic->i_on_dev_list)
1029 			rdsv3_ib_remove_conn(ic->rds_ibdev, conn);
1030 
1031 		ic->i_cm_id = NULL;
1032 		ic->i_pd = NULL;
1033 		ic->i_mr = NULL;
1034 		ic->i_cq = NULL;
1035 		ic->i_snd_cq = NULL;
1036 		ic->i_send_hdrs = NULL;
1037 		ic->i_recv_hdrs = NULL;
1038 		ic->i_ack = NULL;
1039 	}
1040 	ASSERT(!ic->i_on_dev_list);
1041 
1042 	/* Clear pending transmit */
1043 	if (ic->i_rm) {
1044 		rdsv3_message_put(ic->i_rm);
1045 		ic->i_rm = NULL;
1046 	}
1047 
1048 	/* Clear the ACK state */
1049 	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1050 	ic->i_ack_next = 0;
1051 	ic->i_ack_recv = 0;
1052 
1053 	/* Clear flow control state */
1054 	ic->i_flowctl = 0;
1055 	ic->i_credits = 0;
1056 
1057 	rdsv3_ib_ring_init(&ic->i_send_ring, rdsv3_ib_sysctl_max_send_wr);
1058 	rdsv3_ib_ring_init(&ic->i_recv_ring, rdsv3_ib_sysctl_max_recv_wr);
1059 
1060 	if (ic->i_ibinc) {
1061 		rdsv3_inc_put(&ic->i_ibinc->ii_inc);
1062 		ic->i_ibinc = NULL;
1063 	}
1064 
1065 	if (ic->i_sends) {
1066 		kmem_free(ic->i_sends,
1067 		    ic->i_send_ring.w_nr * sizeof (struct rdsv3_ib_send_work));
1068 		ic->i_sends = NULL;
1069 	}
1070 	if (ic->i_send_wrs) {
1071 		kmem_free(ic->i_send_wrs, ic->i_send_ring.w_nr *
1072 		    (sizeof (ibt_send_wr_t) +
1073 		    RDSV3_IB_MAX_SGE * sizeof (ibt_wr_ds_t)));
1074 		ic->i_send_wrs = NULL;
1075 	}
1076 	if (ic->i_recvs) {
1077 		kmem_free(ic->i_recvs,
1078 		    ic->i_recv_ring.w_nr * sizeof (struct rdsv3_ib_recv_work));
1079 		ic->i_recvs = NULL;
1080 	}
1081 	if (ic->i_recv_wrs) {
1082 		kmem_free(ic->i_recv_wrs, ic->i_recv_ring.w_nr *
1083 		    (sizeof (ibt_recv_wr_t)));
1084 		ic->i_recv_wrs = NULL;
1085 	}
1086 
1087 	RDSV3_DPRINTF2("rdsv3_ib_conn_shutdown", "Return conn: %p", conn);
1088 }
1089 
1090 /* ARGSUSED */
1091 int
1092 rdsv3_ib_conn_alloc(struct rdsv3_connection *conn, int gfp)
1093 {
1094 	struct rdsv3_ib_connection *ic;
1095 
1096 	RDSV3_DPRINTF2("rdsv3_ib_conn_alloc", "conn: %p", conn);
1097 
1098 	/* XXX too lazy? */
1099 	ic = kmem_zalloc(sizeof (struct rdsv3_ib_connection), gfp);
1100 	if (!ic)
1101 		return (-ENOMEM);
1102 
1103 	list_link_init(&ic->ib_node);
1104 
1105 	mutex_init(&ic->i_recv_mutex, NULL, MUTEX_DRIVER, NULL);
1106 	mutex_init(&ic->i_ack_lock, NULL, MUTEX_DRIVER, NULL);
1107 
1108 	/*
1109 	 * rdsv3_ib_conn_shutdown() waits for these to be emptied so they
1110 	 * must be initialized before it can be called.
1111 	 */
1112 	rdsv3_ib_ring_init(&ic->i_send_ring, rdsv3_ib_sysctl_max_send_wr);
1113 	rdsv3_ib_ring_init(&ic->i_recv_ring, rdsv3_ib_sysctl_max_recv_wr);
1114 
1115 	ic->conn = conn;
1116 	conn->c_transport_data = ic;
1117 
1118 	mutex_enter(&ib_nodev_conns_lock);
1119 	list_insert_tail(&ib_nodev_conns, ic);
1120 	mutex_exit(&ib_nodev_conns_lock);
1121 
1122 	RDSV3_DPRINTF2("rdsv3_ib_conn_alloc", "conn %p conn ic %p",
1123 	    conn, conn->c_transport_data);
1124 	return (0);
1125 }
1126 
1127 /*
1128  * Free a connection. Connection must be shut down and not set for reconnect.
1129  */
1130 void
1131 rdsv3_ib_conn_free(void *arg)
1132 {
1133 	struct rdsv3_ib_connection *ic = arg;
1134 	kmutex_t	*lock_ptr;
1135 
1136 	RDSV3_DPRINTF2("rdsv3_ib_conn_free", "ic %p\n", ic);
1137 
1138 #ifndef __lock_lint
1139 	/*
1140 	 * Conn is either on a dev's list or on the nodev list.
1141 	 * A race with shutdown() or connect() would cause problems
1142 	 * (since rds_ibdev would change) but that should never happen.
1143 	 */
1144 	lock_ptr = ic->i_on_dev_list ?
1145 	    &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1146 
1147 	mutex_enter(lock_ptr);
1148 	list_remove_node(&ic->ib_node);
1149 	mutex_exit(lock_ptr);
1150 #endif
1151 	kmem_free(ic, sizeof (*ic));
1152 }
1153 
1154 /*
1155  * An error occurred on the connection
1156  */
1157 void
1158 __rdsv3_ib_conn_error(struct rdsv3_connection *conn)
1159 {
1160 	rdsv3_conn_drop(conn);
1161 }
1162