xref: /linux/net/sunrpc/xprtrdma/svc_rdma_transport.c (revision 0fc8f6200d2313278fbf4539bbab74677c685531)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2015-2018 Oracle. All rights reserved.
4  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
5  * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the BSD-type
11  * license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  *      Redistributions of source code must retain the above copyright
18  *      notice, this list of conditions and the following 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 provided
23  *      with the distribution.
24  *
25  *      Neither the name of the Network Appliance, Inc. nor the names of
26  *      its contributors may be used to endorse or promote products
27  *      derived from this software without specific prior written
28  *      permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
31  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
32  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
33  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
34  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
35  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
36  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
37  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
38  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
40  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41  *
42  * Author: Tom Tucker <tom@opengridcomputing.com>
43  */
44 
45 #include <linux/interrupt.h>
46 #include <linux/sched.h>
47 #include <linux/slab.h>
48 #include <linux/spinlock.h>
49 #include <linux/workqueue.h>
50 #include <linux/export.h>
51 
52 #include <rdma/ib_verbs.h>
53 #include <rdma/rdma_cm.h>
54 #include <rdma/rw.h>
55 
56 #include <linux/sunrpc/addr.h>
57 #include <linux/sunrpc/debug.h>
58 #include <linux/sunrpc/svc_xprt.h>
59 #include <linux/sunrpc/svc_rdma.h>
60 
61 #include "xprt_rdma.h"
62 #include <trace/events/rpcrdma.h>
63 
64 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
65 
66 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
67 						 struct net *net, int node);
68 static int svc_rdma_listen_handler(struct rdma_cm_id *cma_id,
69 				   struct rdma_cm_event *event);
70 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
71 					struct net *net,
72 					struct sockaddr *sa, int salen,
73 					int flags);
74 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
75 static void svc_rdma_detach(struct svc_xprt *xprt);
76 static void svc_rdma_free(struct svc_xprt *xprt);
77 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
78 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
79 
80 static const struct svc_xprt_ops svc_rdma_ops = {
81 	.xpo_create = svc_rdma_create,
82 	.xpo_recvfrom = svc_rdma_recvfrom,
83 	.xpo_sendto = svc_rdma_sendto,
84 	.xpo_result_payload = svc_rdma_result_payload,
85 	.xpo_release_ctxt = svc_rdma_release_ctxt,
86 	.xpo_detach = svc_rdma_detach,
87 	.xpo_free = svc_rdma_free,
88 	.xpo_has_wspace = svc_rdma_has_wspace,
89 	.xpo_accept = svc_rdma_accept,
90 	.xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
91 };
92 
93 struct svc_xprt_class svc_rdma_class = {
94 	.xcl_name = "rdma",
95 	.xcl_owner = THIS_MODULE,
96 	.xcl_ops = &svc_rdma_ops,
97 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
98 	.xcl_ident = XPRT_TRANSPORT_RDMA,
99 };
100 
101 /* QP event handler */
102 static void qp_event_handler(struct ib_event *event, void *context)
103 {
104 	struct svc_xprt *xprt = context;
105 
106 	trace_svcrdma_qp_error(event, (struct sockaddr *)&xprt->xpt_remote);
107 	switch (event->event) {
108 	/* These are considered benign events */
109 	case IB_EVENT_PATH_MIG:
110 	case IB_EVENT_COMM_EST:
111 	case IB_EVENT_SQ_DRAINED:
112 	case IB_EVENT_QP_LAST_WQE_REACHED:
113 		break;
114 
115 	/* These are considered fatal events */
116 	case IB_EVENT_PATH_MIG_ERR:
117 	case IB_EVENT_QP_FATAL:
118 	case IB_EVENT_QP_REQ_ERR:
119 	case IB_EVENT_QP_ACCESS_ERR:
120 	case IB_EVENT_DEVICE_FATAL:
121 	default:
122 		svc_xprt_deferred_close(xprt);
123 		break;
124 	}
125 }
126 
127 static struct rdma_cm_id *
128 svc_rdma_create_listen_id(struct net *net, struct sockaddr *sap,
129 			  void *context)
130 {
131 	struct rdma_cm_id *listen_id;
132 	int ret;
133 
134 	listen_id = rdma_create_id(net, svc_rdma_listen_handler, context,
135 				   RDMA_PS_TCP, IB_QPT_RC);
136 	if (IS_ERR(listen_id))
137 		return listen_id;
138 
139 	/* Allow both IPv4 and IPv6 sockets to bind a single port
140 	 * at the same time.
141 	 */
142 #if IS_ENABLED(CONFIG_IPV6)
143 	ret = rdma_set_afonly(listen_id, 1);
144 	if (ret)
145 		goto out_destroy;
146 #endif
147 	ret = rdma_bind_addr(listen_id, sap);
148 	if (ret)
149 		goto out_destroy;
150 
151 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
152 	if (ret)
153 		goto out_destroy;
154 
155 	return listen_id;
156 
157 out_destroy:
158 	rdma_destroy_id(listen_id);
159 	return ERR_PTR(ret);
160 }
161 
162 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
163 						 struct net *net, int node)
164 {
165 	static struct lock_class_key svcrdma_rwctx_lock;
166 	static struct lock_class_key svcrdma_sctx_lock;
167 	static struct lock_class_key svcrdma_dto_lock;
168 	struct svcxprt_rdma *cma_xprt;
169 
170 	cma_xprt = kzalloc_node(sizeof(*cma_xprt), GFP_KERNEL, node);
171 	if (!cma_xprt)
172 		return NULL;
173 
174 	svc_xprt_init(net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
175 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
176 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
177 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
178 	init_llist_head(&cma_xprt->sc_send_ctxts);
179 	init_llist_head(&cma_xprt->sc_recv_ctxts);
180 	init_llist_head(&cma_xprt->sc_rw_ctxts);
181 	init_waitqueue_head(&cma_xprt->sc_send_wait);
182 	init_waitqueue_head(&cma_xprt->sc_sq_ticket_wait);
183 
184 	spin_lock_init(&cma_xprt->sc_lock);
185 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
186 	lockdep_set_class(&cma_xprt->sc_rq_dto_lock, &svcrdma_dto_lock);
187 	spin_lock_init(&cma_xprt->sc_send_lock);
188 	lockdep_set_class(&cma_xprt->sc_send_lock, &svcrdma_sctx_lock);
189 	spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
190 	lockdep_set_class(&cma_xprt->sc_rw_ctxt_lock, &svcrdma_rwctx_lock);
191 
192 	/*
193 	 * Note that this implies that the underlying transport support
194 	 * has some form of congestion control (see RFC 7530 section 3.1
195 	 * paragraph 2). For now, we assume that all supported RDMA
196 	 * transports are suitable here.
197 	 */
198 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
199 
200 	return cma_xprt;
201 }
202 
203 static void
204 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
205 			       struct rdma_conn_param *param)
206 {
207 	const struct rpcrdma_connect_private *pmsg = param->private_data;
208 
209 	if (pmsg &&
210 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
211 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
212 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
213 					RPCRDMA_CMP_F_SND_W_INV_OK;
214 
215 		dprintk("svcrdma: client send_size %u, recv_size %u "
216 			"remote inv %ssupported\n",
217 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
218 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
219 			newxprt->sc_snd_w_inv ? "" : "un");
220 	}
221 }
222 
223 /*
224  * This function handles the CONNECT_REQUEST event on a listening
225  * endpoint. It is passed the cma_id for the _new_ connection. The context in
226  * this cma_id is inherited from the listening cma_id and is the svc_xprt
227  * structure for the listening endpoint.
228  *
229  * This function creates a new xprt for the new connection and enqueues it on
230  * the accept queue for the listent xprt. When the listen thread is kicked, it
231  * will call the recvfrom method on the listen xprt which will accept the new
232  * connection.
233  */
234 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
235 			       struct rdma_conn_param *param)
236 {
237 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
238 	struct svcxprt_rdma *newxprt;
239 	struct sockaddr *sa;
240 
241 	newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server,
242 				       listen_xprt->sc_xprt.xpt_net,
243 				       ibdev_to_node(new_cma_id->device));
244 	if (!newxprt)
245 		return;
246 	newxprt->sc_cm_id = new_cma_id;
247 	new_cma_id->context = newxprt;
248 	svc_rdma_parse_connect_private(newxprt, param);
249 
250 	/* Save client advertised inbound read limit for use later in accept. */
251 	newxprt->sc_ord = param->initiator_depth;
252 
253 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
254 	newxprt->sc_xprt.xpt_remotelen = svc_addr_len(sa);
255 	memcpy(&newxprt->sc_xprt.xpt_remote, sa,
256 	       newxprt->sc_xprt.xpt_remotelen);
257 	snprintf(newxprt->sc_xprt.xpt_remotebuf,
258 		 sizeof(newxprt->sc_xprt.xpt_remotebuf) - 1, "%pISc", sa);
259 
260 	/* The remote port is arbitrary and not under the control of the
261 	 * client ULP. Set it to a fixed value so that the DRC continues
262 	 * to be effective after a reconnect.
263 	 */
264 	rpc_set_port((struct sockaddr *)&newxprt->sc_xprt.xpt_remote, 0);
265 
266 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
267 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
268 
269 	/*
270 	 * Enqueue the new transport on the accept queue of the listening
271 	 * transport
272 	 */
273 	spin_lock(&listen_xprt->sc_lock);
274 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
275 	spin_unlock(&listen_xprt->sc_lock);
276 
277 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
278 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
279 }
280 
281 /**
282  * svc_rdma_listen_handler - Handle CM events generated on a listening endpoint
283  * @cma_id: the server's listener rdma_cm_id
284  * @event: details of the event
285  *
286  * Return values:
287  *     %0: Do not destroy @cma_id
288  *     %1: Destroy @cma_id
289  *
290  * NB: There is never a DEVICE_REMOVAL event for INADDR_ANY listeners.
291  */
292 static int svc_rdma_listen_handler(struct rdma_cm_id *cma_id,
293 				   struct rdma_cm_event *event)
294 {
295 	struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.src_addr;
296 	struct svcxprt_rdma *cma_xprt = cma_id->context;
297 	struct svc_xprt *cma_rdma = &cma_xprt->sc_xprt;
298 	struct rdma_cm_id *listen_id;
299 
300 	switch (event->event) {
301 	case RDMA_CM_EVENT_CONNECT_REQUEST:
302 		handle_connect_req(cma_id, &event->param.conn);
303 		break;
304 	case RDMA_CM_EVENT_ADDR_CHANGE:
305 		listen_id = svc_rdma_create_listen_id(cma_rdma->xpt_net,
306 						      sap, cma_xprt);
307 		if (IS_ERR(listen_id)) {
308 			pr_err("Listener dead, address change failed for device %s\n",
309 				cma_id->device->name);
310 		} else
311 			cma_xprt->sc_cm_id = listen_id;
312 		return 1;
313 	default:
314 		break;
315 	}
316 	return 0;
317 }
318 
319 /**
320  * svc_rdma_cma_handler - Handle CM events on client connections
321  * @cma_id: the server's listener rdma_cm_id
322  * @event: details of the event
323  *
324  * Return values:
325  *     %0: Do not destroy @cma_id
326  *     %1: Destroy @cma_id (never returned here)
327  */
328 static int svc_rdma_cma_handler(struct rdma_cm_id *cma_id,
329 				struct rdma_cm_event *event)
330 {
331 	struct svcxprt_rdma *rdma = cma_id->context;
332 	struct svc_xprt *xprt = &rdma->sc_xprt;
333 
334 	switch (event->event) {
335 	case RDMA_CM_EVENT_ESTABLISHED:
336 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
337 
338 		/* Handle any requests that were received while
339 		 * CONN_PENDING was set. */
340 		svc_xprt_enqueue(xprt);
341 		break;
342 	case RDMA_CM_EVENT_DISCONNECTED:
343 		svc_xprt_deferred_close(xprt);
344 		break;
345 	default:
346 		break;
347 	}
348 	return 0;
349 }
350 
351 /*
352  * Create a listening RDMA service endpoint.
353  */
354 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
355 					struct net *net,
356 					struct sockaddr *sa, int salen,
357 					int flags)
358 {
359 	struct rdma_cm_id *listen_id;
360 	struct svcxprt_rdma *cma_xprt;
361 
362 	if (sa->sa_family != AF_INET && sa->sa_family != AF_INET6)
363 		return ERR_PTR(-EAFNOSUPPORT);
364 	cma_xprt = svc_rdma_create_xprt(serv, net, NUMA_NO_NODE);
365 	if (!cma_xprt)
366 		return ERR_PTR(-ENOMEM);
367 	set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
368 	strcpy(cma_xprt->sc_xprt.xpt_remotebuf, "listener");
369 
370 	listen_id = svc_rdma_create_listen_id(net, sa, cma_xprt);
371 	if (IS_ERR(listen_id)) {
372 		kfree(cma_xprt);
373 		return ERR_CAST(listen_id);
374 	}
375 	cma_xprt->sc_cm_id = listen_id;
376 
377 	/*
378 	 * We need to use the address from the cm_id in case the
379 	 * caller specified 0 for the port number.
380 	 */
381 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
382 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
383 
384 	return &cma_xprt->sc_xprt;
385 }
386 
387 static void svc_rdma_xprt_done(struct rpcrdma_notification *rn)
388 {
389 	struct svcxprt_rdma *rdma = container_of(rn, struct svcxprt_rdma,
390 						 sc_rn);
391 	struct rdma_cm_id *id = rdma->sc_cm_id;
392 
393 	trace_svcrdma_device_removal(id);
394 	svc_xprt_close(&rdma->sc_xprt);
395 }
396 
397 /*
398  * This is the xpo_recvfrom function for listening endpoints. Its
399  * purpose is to accept incoming connections. The CMA callback handler
400  * has already created a new transport and attached it to the new CMA
401  * ID.
402  *
403  * There is a queue of pending connections hung on the listening
404  * transport. This queue contains the new svc_xprt structure. This
405  * function takes svc_xprt structures off the accept_q and completes
406  * the connection.
407  */
408 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
409 {
410 	unsigned int ctxts, rq_depth, maxpayload;
411 	struct svcxprt_rdma *listen_rdma;
412 	struct svcxprt_rdma *newxprt = NULL;
413 	struct rdma_conn_param conn_param;
414 	struct rpcrdma_connect_private pmsg;
415 	struct ib_qp_init_attr qp_attr;
416 	struct ib_device *dev;
417 	int ret = 0;
418 
419 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
420 	clear_bit(XPT_CONN, &xprt->xpt_flags);
421 	/* Get the next entry off the accept list */
422 	spin_lock(&listen_rdma->sc_lock);
423 	if (!list_empty(&listen_rdma->sc_accept_q)) {
424 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
425 				     struct svcxprt_rdma, sc_accept_q);
426 		list_del_init(&newxprt->sc_accept_q);
427 	}
428 	if (!list_empty(&listen_rdma->sc_accept_q))
429 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
430 	spin_unlock(&listen_rdma->sc_lock);
431 	if (!newxprt)
432 		return NULL;
433 
434 	dev = newxprt->sc_cm_id->device;
435 	newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
436 
437 	if (rpcrdma_rn_register(dev, &newxprt->sc_rn, svc_rdma_xprt_done))
438 		goto errout;
439 
440 	newxprt->sc_max_req_size = svcrdma_max_req_size;
441 	newxprt->sc_max_requests = svcrdma_max_requests;
442 	newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
443 	newxprt->sc_recv_batch = RPCRDMA_MAX_RECV_BATCH;
444 	newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
445 
446 	/* Qualify the transport's resource defaults with the
447 	 * capabilities of this particular device.
448 	 */
449 
450 	/* Transport header, head iovec, tail iovec */
451 	newxprt->sc_max_send_sges = 3;
452 	/* Add one SGE per page list entry */
453 	newxprt->sc_max_send_sges += (svcrdma_max_req_size / PAGE_SIZE) + 1;
454 	if (newxprt->sc_max_send_sges > dev->attrs.max_send_sge)
455 		newxprt->sc_max_send_sges = dev->attrs.max_send_sge;
456 	rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests +
457 		   newxprt->sc_recv_batch + 1 /* drain */;
458 	if (rq_depth > dev->attrs.max_qp_wr) {
459 		rq_depth = dev->attrs.max_qp_wr;
460 		newxprt->sc_recv_batch = 1;
461 		newxprt->sc_max_requests = rq_depth - 2;
462 		newxprt->sc_max_bc_requests = 2;
463 	}
464 
465 	/* Estimate the needed number of rdma_rw contexts. The maximum
466 	 * Read and Write chunks have one segment each. Each request
467 	 * can involve one Read chunk and either a Write chunk or Reply
468 	 * chunk; thus a factor of three.
469 	 */
470 	maxpayload = min(xprt->xpt_server->sv_max_payload,
471 			 RPCSVC_MAXPAYLOAD_RDMA);
472 	ctxts = newxprt->sc_max_requests * 3 *
473 		rdma_rw_mr_factor(dev, newxprt->sc_port_num,
474 				  maxpayload >> PAGE_SHIFT);
475 
476 	newxprt->sc_sq_depth = rq_depth +
477 		rdma_rw_max_send_wr(dev, newxprt->sc_port_num, ctxts, 0);
478 	if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr)
479 		newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
480 	atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
481 	atomic_set(&newxprt->sc_sq_ticket_head, 0);
482 	atomic_set(&newxprt->sc_sq_ticket_tail, 0);
483 
484 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
485 	if (IS_ERR(newxprt->sc_pd)) {
486 		trace_svcrdma_pd_err(newxprt, PTR_ERR(newxprt->sc_pd));
487 		goto errout;
488 	}
489 	newxprt->sc_sq_cq = ib_alloc_cq_any(dev, newxprt, newxprt->sc_sq_depth,
490 					    IB_POLL_WORKQUEUE);
491 	if (IS_ERR(newxprt->sc_sq_cq))
492 		goto errout;
493 	newxprt->sc_rq_cq =
494 		ib_alloc_cq_any(dev, newxprt, rq_depth, IB_POLL_WORKQUEUE);
495 	if (IS_ERR(newxprt->sc_rq_cq))
496 		goto errout;
497 
498 	memset(&qp_attr, 0, sizeof qp_attr);
499 	qp_attr.event_handler = qp_event_handler;
500 	qp_attr.qp_context = &newxprt->sc_xprt;
501 	qp_attr.port_num = newxprt->sc_port_num;
502 	qp_attr.cap.max_rdma_ctxs = ctxts;
503 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
504 	qp_attr.cap.max_recv_wr = rq_depth;
505 	qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges;
506 	qp_attr.cap.max_recv_sge = 1;
507 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
508 	qp_attr.qp_type = IB_QPT_RC;
509 	qp_attr.send_cq = newxprt->sc_sq_cq;
510 	qp_attr.recv_cq = newxprt->sc_rq_cq;
511 	dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
512 		qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
513 	dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
514 		qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
515 	dprintk("    send CQ depth = %u, recv CQ depth = %u\n",
516 		newxprt->sc_sq_depth, rq_depth);
517 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
518 	if (ret) {
519 		trace_svcrdma_qp_err(newxprt, ret);
520 		goto errout;
521 	}
522 	newxprt->sc_max_send_sges = qp_attr.cap.max_send_sge;
523 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
524 
525 	if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
526 		newxprt->sc_snd_w_inv = false;
527 	if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
528 	    !rdma_ib_or_roce(dev, newxprt->sc_port_num)) {
529 		trace_svcrdma_fabric_err(newxprt, -EINVAL);
530 		goto errout;
531 	}
532 
533 	if (!svc_rdma_post_recvs(newxprt))
534 		goto errout;
535 
536 	/* Construct RDMA-CM private message */
537 	pmsg.cp_magic = rpcrdma_cmp_magic;
538 	pmsg.cp_version = RPCRDMA_CMP_VERSION;
539 	pmsg.cp_flags = 0;
540 	pmsg.cp_send_size = pmsg.cp_recv_size =
541 		rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
542 
543 	/* Accept Connection */
544 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
545 	memset(&conn_param, 0, sizeof conn_param);
546 	conn_param.responder_resources = 0;
547 	conn_param.initiator_depth = min_t(int, newxprt->sc_ord,
548 					   dev->attrs.max_qp_init_rd_atom);
549 	if (!conn_param.initiator_depth) {
550 		ret = -EINVAL;
551 		trace_svcrdma_initdepth_err(newxprt, ret);
552 		goto errout;
553 	}
554 	conn_param.private_data = &pmsg;
555 	conn_param.private_data_len = sizeof(pmsg);
556 	rdma_lock_handler(newxprt->sc_cm_id);
557 	newxprt->sc_cm_id->event_handler = svc_rdma_cma_handler;
558 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
559 	rdma_unlock_handler(newxprt->sc_cm_id);
560 	if (ret) {
561 		trace_svcrdma_accept_err(newxprt, ret);
562 		goto errout;
563 	}
564 
565 	if (IS_ENABLED(CONFIG_SUNRPC_DEBUG)) {
566 		struct sockaddr *sap;
567 
568 		dprintk("svcrdma: new connection accepted on device %s:\n", dev->name);
569 		sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
570 		dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
571 		sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
572 		dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
573 		dprintk("    max_sge         : %d\n", newxprt->sc_max_send_sges);
574 		dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
575 		dprintk("    rdma_rw_ctxs    : %d\n", ctxts);
576 		dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
577 		dprintk("    ord             : %d\n", conn_param.initiator_depth);
578 	}
579 
580 	return &newxprt->sc_xprt;
581 
582  errout:
583 	/* Take a reference in case the DTO handler runs */
584 	svc_xprt_get(&newxprt->sc_xprt);
585 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
586 		ib_destroy_qp(newxprt->sc_qp);
587 	rdma_destroy_id(newxprt->sc_cm_id);
588 	rpcrdma_rn_unregister(dev, &newxprt->sc_rn);
589 	/* This call to put will destroy the transport */
590 	svc_xprt_put(&newxprt->sc_xprt);
591 	return NULL;
592 }
593 
594 static void svc_rdma_detach(struct svc_xprt *xprt)
595 {
596 	struct svcxprt_rdma *rdma =
597 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
598 
599 	rdma_disconnect(rdma->sc_cm_id);
600 }
601 
602 /**
603  * svc_rdma_free - Release class-specific transport resources
604  * @xprt: Generic svc transport object
605  */
606 static void svc_rdma_free(struct svc_xprt *xprt)
607 {
608 	struct svcxprt_rdma *rdma =
609 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
610 	struct ib_device *device = rdma->sc_cm_id->device;
611 
612 	might_sleep();
613 
614 	/* This blocks until the Completion Queues are empty */
615 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
616 		ib_drain_qp(rdma->sc_qp);
617 	flush_workqueue(svcrdma_wq);
618 
619 	svc_rdma_flush_recv_queues(rdma);
620 
621 	svc_rdma_destroy_rw_ctxts(rdma);
622 	svc_rdma_send_ctxts_destroy(rdma);
623 	svc_rdma_recv_ctxts_destroy(rdma);
624 
625 	/* Destroy the QP if present (not a listener) */
626 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
627 		ib_destroy_qp(rdma->sc_qp);
628 
629 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
630 		ib_free_cq(rdma->sc_sq_cq);
631 
632 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
633 		ib_free_cq(rdma->sc_rq_cq);
634 
635 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
636 		ib_dealloc_pd(rdma->sc_pd);
637 
638 	/* Destroy the CM ID */
639 	rdma_destroy_id(rdma->sc_cm_id);
640 
641 	if (!test_bit(XPT_LISTENER, &rdma->sc_xprt.xpt_flags))
642 		rpcrdma_rn_unregister(device, &rdma->sc_rn);
643 	kfree(rdma);
644 }
645 
646 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
647 {
648 	struct svcxprt_rdma *rdma =
649 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
650 
651 	/*
652 	 * If there are already waiters on the SQ,
653 	 * return false.
654 	 */
655 	if (waitqueue_active(&rdma->sc_send_wait) ||
656 	    waitqueue_active(&rdma->sc_sq_ticket_wait))
657 		return 0;
658 
659 	/* Otherwise return true. */
660 	return 1;
661 }
662 
663 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
664 {
665 }
666