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/rpc_rdma.h>
59 #include <linux/sunrpc/svc_xprt.h>
60 #include <linux/sunrpc/svc_rdma.h>
61 
62 #include "xprt_rdma.h"
63 #include <trace/events/rpcrdma.h>
64 
65 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
66 
67 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
68 						 struct net *net);
69 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
70 					struct net *net,
71 					struct sockaddr *sa, int salen,
72 					int flags);
73 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
74 static void svc_rdma_release_rqst(struct svc_rqst *);
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_secure_port(struct svc_rqst *);
79 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
80 
81 static const struct svc_xprt_ops svc_rdma_ops = {
82 	.xpo_create = svc_rdma_create,
83 	.xpo_recvfrom = svc_rdma_recvfrom,
84 	.xpo_sendto = svc_rdma_sendto,
85 	.xpo_release_rqst = svc_rdma_release_rqst,
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_secure_port = svc_rdma_secure_port,
91 	.xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
92 };
93 
94 struct svc_xprt_class svc_rdma_class = {
95 	.xcl_name = "rdma",
96 	.xcl_owner = THIS_MODULE,
97 	.xcl_ops = &svc_rdma_ops,
98 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
99 	.xcl_ident = XPRT_TRANSPORT_RDMA,
100 };
101 
102 /* QP event handler */
103 static void qp_event_handler(struct ib_event *event, void *context)
104 {
105 	struct svc_xprt *xprt = context;
106 
107 	trace_svcrdma_qp_error(event, (struct sockaddr *)&xprt->xpt_remote);
108 	switch (event->event) {
109 	/* These are considered benign events */
110 	case IB_EVENT_PATH_MIG:
111 	case IB_EVENT_COMM_EST:
112 	case IB_EVENT_SQ_DRAINED:
113 	case IB_EVENT_QP_LAST_WQE_REACHED:
114 		break;
115 
116 	/* These are considered fatal events */
117 	case IB_EVENT_PATH_MIG_ERR:
118 	case IB_EVENT_QP_FATAL:
119 	case IB_EVENT_QP_REQ_ERR:
120 	case IB_EVENT_QP_ACCESS_ERR:
121 	case IB_EVENT_DEVICE_FATAL:
122 	default:
123 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
124 		svc_xprt_enqueue(xprt);
125 		break;
126 	}
127 }
128 
129 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
130 						 struct net *net)
131 {
132 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
133 
134 	if (!cma_xprt) {
135 		dprintk("svcrdma: failed to create new transport\n");
136 		return NULL;
137 	}
138 	svc_xprt_init(net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
139 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
140 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
141 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
142 	INIT_LIST_HEAD(&cma_xprt->sc_send_ctxts);
143 	INIT_LIST_HEAD(&cma_xprt->sc_recv_ctxts);
144 	INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
145 	init_waitqueue_head(&cma_xprt->sc_send_wait);
146 
147 	spin_lock_init(&cma_xprt->sc_lock);
148 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
149 	spin_lock_init(&cma_xprt->sc_send_lock);
150 	spin_lock_init(&cma_xprt->sc_recv_lock);
151 	spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
152 
153 	/*
154 	 * Note that this implies that the underlying transport support
155 	 * has some form of congestion control (see RFC 7530 section 3.1
156 	 * paragraph 2). For now, we assume that all supported RDMA
157 	 * transports are suitable here.
158 	 */
159 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
160 
161 	return cma_xprt;
162 }
163 
164 static void
165 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
166 			       struct rdma_conn_param *param)
167 {
168 	const struct rpcrdma_connect_private *pmsg = param->private_data;
169 
170 	if (pmsg &&
171 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
172 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
173 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
174 					RPCRDMA_CMP_F_SND_W_INV_OK;
175 
176 		dprintk("svcrdma: client send_size %u, recv_size %u "
177 			"remote inv %ssupported\n",
178 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
179 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
180 			newxprt->sc_snd_w_inv ? "" : "un");
181 	}
182 }
183 
184 /*
185  * This function handles the CONNECT_REQUEST event on a listening
186  * endpoint. It is passed the cma_id for the _new_ connection. The context in
187  * this cma_id is inherited from the listening cma_id and is the svc_xprt
188  * structure for the listening endpoint.
189  *
190  * This function creates a new xprt for the new connection and enqueues it on
191  * the accept queue for the listent xprt. When the listen thread is kicked, it
192  * will call the recvfrom method on the listen xprt which will accept the new
193  * connection.
194  */
195 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
196 			       struct rdma_conn_param *param)
197 {
198 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
199 	struct svcxprt_rdma *newxprt;
200 	struct sockaddr *sa;
201 
202 	/* Create a new transport */
203 	newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server,
204 				       listen_xprt->sc_xprt.xpt_net);
205 	if (!newxprt)
206 		return;
207 	newxprt->sc_cm_id = new_cma_id;
208 	new_cma_id->context = newxprt;
209 	svc_rdma_parse_connect_private(newxprt, param);
210 
211 	/* Save client advertised inbound read limit for use later in accept. */
212 	newxprt->sc_ord = param->initiator_depth;
213 
214 	/* Set the local and remote addresses in the transport */
215 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
216 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
217 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
218 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
219 
220 	/*
221 	 * Enqueue the new transport on the accept queue of the listening
222 	 * transport
223 	 */
224 	spin_lock_bh(&listen_xprt->sc_lock);
225 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
226 	spin_unlock_bh(&listen_xprt->sc_lock);
227 
228 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
229 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
230 }
231 
232 /*
233  * Handles events generated on the listening endpoint. These events will be
234  * either be incoming connect requests or adapter removal  events.
235  */
236 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
237 			       struct rdma_cm_event *event)
238 {
239 	struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.src_addr;
240 
241 	trace_svcrdma_cm_event(event, sap);
242 
243 	switch (event->event) {
244 	case RDMA_CM_EVENT_CONNECT_REQUEST:
245 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
246 			"event = %s (%d)\n", cma_id, cma_id->context,
247 			rdma_event_msg(event->event), event->event);
248 		handle_connect_req(cma_id, &event->param.conn);
249 		break;
250 	default:
251 		/* NB: No device removal upcall for INADDR_ANY listeners */
252 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
253 			"event = %s (%d)\n", cma_id,
254 			rdma_event_msg(event->event), event->event);
255 		break;
256 	}
257 
258 	return 0;
259 }
260 
261 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
262 			    struct rdma_cm_event *event)
263 {
264 	struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.dst_addr;
265 	struct svcxprt_rdma *rdma = cma_id->context;
266 	struct svc_xprt *xprt = &rdma->sc_xprt;
267 
268 	trace_svcrdma_cm_event(event, sap);
269 
270 	switch (event->event) {
271 	case RDMA_CM_EVENT_ESTABLISHED:
272 		/* Accept complete */
273 		svc_xprt_get(xprt);
274 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
275 			"cm_id=%p\n", xprt, cma_id);
276 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
277 		svc_xprt_enqueue(xprt);
278 		break;
279 	case RDMA_CM_EVENT_DISCONNECTED:
280 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
281 			xprt, cma_id);
282 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
283 		svc_xprt_enqueue(xprt);
284 		svc_xprt_put(xprt);
285 		break;
286 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
287 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
288 			"event = %s (%d)\n", cma_id, xprt,
289 			rdma_event_msg(event->event), event->event);
290 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
291 		svc_xprt_enqueue(xprt);
292 		svc_xprt_put(xprt);
293 		break;
294 	default:
295 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
296 			"event = %s (%d)\n", cma_id,
297 			rdma_event_msg(event->event), event->event);
298 		break;
299 	}
300 	return 0;
301 }
302 
303 /*
304  * Create a listening RDMA service endpoint.
305  */
306 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
307 					struct net *net,
308 					struct sockaddr *sa, int salen,
309 					int flags)
310 {
311 	struct rdma_cm_id *listen_id;
312 	struct svcxprt_rdma *cma_xprt;
313 	int ret;
314 
315 	dprintk("svcrdma: Creating RDMA listener\n");
316 	if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
317 		dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
318 		return ERR_PTR(-EAFNOSUPPORT);
319 	}
320 	cma_xprt = svc_rdma_create_xprt(serv, net);
321 	if (!cma_xprt)
322 		return ERR_PTR(-ENOMEM);
323 	set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
324 	strcpy(cma_xprt->sc_xprt.xpt_remotebuf, "listener");
325 
326 	listen_id = rdma_create_id(net, rdma_listen_handler, cma_xprt,
327 				   RDMA_PS_TCP, IB_QPT_RC);
328 	if (IS_ERR(listen_id)) {
329 		ret = PTR_ERR(listen_id);
330 		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
331 		goto err0;
332 	}
333 
334 	/* Allow both IPv4 and IPv6 sockets to bind a single port
335 	 * at the same time.
336 	 */
337 #if IS_ENABLED(CONFIG_IPV6)
338 	ret = rdma_set_afonly(listen_id, 1);
339 	if (ret) {
340 		dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
341 		goto err1;
342 	}
343 #endif
344 	ret = rdma_bind_addr(listen_id, sa);
345 	if (ret) {
346 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
347 		goto err1;
348 	}
349 	cma_xprt->sc_cm_id = listen_id;
350 
351 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
352 	if (ret) {
353 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
354 		goto err1;
355 	}
356 
357 	/*
358 	 * We need to use the address from the cm_id in case the
359 	 * caller specified 0 for the port number.
360 	 */
361 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
362 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
363 
364 	return &cma_xprt->sc_xprt;
365 
366  err1:
367 	rdma_destroy_id(listen_id);
368  err0:
369 	kfree(cma_xprt);
370 	return ERR_PTR(ret);
371 }
372 
373 /*
374  * This is the xpo_recvfrom function for listening endpoints. Its
375  * purpose is to accept incoming connections. The CMA callback handler
376  * has already created a new transport and attached it to the new CMA
377  * ID.
378  *
379  * There is a queue of pending connections hung on the listening
380  * transport. This queue contains the new svc_xprt structure. This
381  * function takes svc_xprt structures off the accept_q and completes
382  * the connection.
383  */
384 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
385 {
386 	struct svcxprt_rdma *listen_rdma;
387 	struct svcxprt_rdma *newxprt = NULL;
388 	struct rdma_conn_param conn_param;
389 	struct rpcrdma_connect_private pmsg;
390 	struct ib_qp_init_attr qp_attr;
391 	unsigned int ctxts, rq_depth;
392 	struct ib_device *dev;
393 	struct sockaddr *sap;
394 	int ret = 0;
395 
396 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
397 	clear_bit(XPT_CONN, &xprt->xpt_flags);
398 	/* Get the next entry off the accept list */
399 	spin_lock_bh(&listen_rdma->sc_lock);
400 	if (!list_empty(&listen_rdma->sc_accept_q)) {
401 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
402 				     struct svcxprt_rdma, sc_accept_q);
403 		list_del_init(&newxprt->sc_accept_q);
404 	}
405 	if (!list_empty(&listen_rdma->sc_accept_q))
406 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
407 	spin_unlock_bh(&listen_rdma->sc_lock);
408 	if (!newxprt)
409 		return NULL;
410 
411 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
412 		newxprt, newxprt->sc_cm_id);
413 
414 	dev = newxprt->sc_cm_id->device;
415 	newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
416 
417 	/* Qualify the transport resource defaults with the
418 	 * capabilities of this particular device */
419 	/* Transport header, head iovec, tail iovec */
420 	newxprt->sc_max_send_sges = 3;
421 	/* Add one SGE per page list entry */
422 	newxprt->sc_max_send_sges += svcrdma_max_req_size / PAGE_SIZE;
423 	if (newxprt->sc_max_send_sges > dev->attrs.max_send_sge) {
424 		pr_err("svcrdma: too few Send SGEs available (%d needed)\n",
425 		       newxprt->sc_max_send_sges);
426 		goto errout;
427 	}
428 	newxprt->sc_max_req_size = svcrdma_max_req_size;
429 	newxprt->sc_max_requests = svcrdma_max_requests;
430 	newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
431 	rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests;
432 	if (rq_depth > dev->attrs.max_qp_wr) {
433 		pr_warn("svcrdma: reducing receive depth to %d\n",
434 			dev->attrs.max_qp_wr);
435 		rq_depth = dev->attrs.max_qp_wr;
436 		newxprt->sc_max_requests = rq_depth - 2;
437 		newxprt->sc_max_bc_requests = 2;
438 	}
439 	newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
440 	ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
441 	ctxts *= newxprt->sc_max_requests;
442 	newxprt->sc_sq_depth = rq_depth + ctxts;
443 	if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
444 		pr_warn("svcrdma: reducing send depth to %d\n",
445 			dev->attrs.max_qp_wr);
446 		newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
447 	}
448 	atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
449 
450 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
451 	if (IS_ERR(newxprt->sc_pd)) {
452 		dprintk("svcrdma: error creating PD for connect request\n");
453 		goto errout;
454 	}
455 	newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
456 					0, IB_POLL_WORKQUEUE);
457 	if (IS_ERR(newxprt->sc_sq_cq)) {
458 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
459 		goto errout;
460 	}
461 	newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, rq_depth,
462 					0, IB_POLL_WORKQUEUE);
463 	if (IS_ERR(newxprt->sc_rq_cq)) {
464 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
465 		goto errout;
466 	}
467 
468 	memset(&qp_attr, 0, sizeof qp_attr);
469 	qp_attr.event_handler = qp_event_handler;
470 	qp_attr.qp_context = &newxprt->sc_xprt;
471 	qp_attr.port_num = newxprt->sc_port_num;
472 	qp_attr.cap.max_rdma_ctxs = ctxts;
473 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
474 	qp_attr.cap.max_recv_wr = rq_depth;
475 	qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges;
476 	qp_attr.cap.max_recv_sge = 1;
477 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
478 	qp_attr.qp_type = IB_QPT_RC;
479 	qp_attr.send_cq = newxprt->sc_sq_cq;
480 	qp_attr.recv_cq = newxprt->sc_rq_cq;
481 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
482 		newxprt->sc_cm_id, newxprt->sc_pd);
483 	dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
484 		qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
485 	dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
486 		qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
487 
488 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
489 	if (ret) {
490 		dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
491 		goto errout;
492 	}
493 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
494 
495 	if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
496 		newxprt->sc_snd_w_inv = false;
497 	if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
498 	    !rdma_ib_or_roce(dev, newxprt->sc_port_num))
499 		goto errout;
500 
501 	if (!svc_rdma_post_recvs(newxprt))
502 		goto errout;
503 
504 	/* Swap out the handler */
505 	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
506 
507 	/* Construct RDMA-CM private message */
508 	pmsg.cp_magic = rpcrdma_cmp_magic;
509 	pmsg.cp_version = RPCRDMA_CMP_VERSION;
510 	pmsg.cp_flags = 0;
511 	pmsg.cp_send_size = pmsg.cp_recv_size =
512 		rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
513 
514 	/* Accept Connection */
515 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
516 	memset(&conn_param, 0, sizeof conn_param);
517 	conn_param.responder_resources = 0;
518 	conn_param.initiator_depth = min_t(int, newxprt->sc_ord,
519 					   dev->attrs.max_qp_init_rd_atom);
520 	if (!conn_param.initiator_depth) {
521 		dprintk("svcrdma: invalid ORD setting\n");
522 		ret = -EINVAL;
523 		goto errout;
524 	}
525 	conn_param.private_data = &pmsg;
526 	conn_param.private_data_len = sizeof(pmsg);
527 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
528 	if (ret)
529 		goto errout;
530 
531 	dprintk("svcrdma: new connection %p accepted:\n", newxprt);
532 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
533 	dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
534 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
535 	dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
536 	dprintk("    max_sge         : %d\n", newxprt->sc_max_send_sges);
537 	dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
538 	dprintk("    rdma_rw_ctxs    : %d\n", ctxts);
539 	dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
540 	dprintk("    ord             : %d\n", conn_param.initiator_depth);
541 
542 	trace_svcrdma_xprt_accept(&newxprt->sc_xprt);
543 	return &newxprt->sc_xprt;
544 
545  errout:
546 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
547 	trace_svcrdma_xprt_fail(&newxprt->sc_xprt);
548 	/* Take a reference in case the DTO handler runs */
549 	svc_xprt_get(&newxprt->sc_xprt);
550 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
551 		ib_destroy_qp(newxprt->sc_qp);
552 	rdma_destroy_id(newxprt->sc_cm_id);
553 	/* This call to put will destroy the transport */
554 	svc_xprt_put(&newxprt->sc_xprt);
555 	return NULL;
556 }
557 
558 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
559 {
560 }
561 
562 /*
563  * When connected, an svc_xprt has at least two references:
564  *
565  * - A reference held by the cm_id between the ESTABLISHED and
566  *   DISCONNECTED events. If the remote peer disconnected first, this
567  *   reference could be gone.
568  *
569  * - A reference held by the svc_recv code that called this function
570  *   as part of close processing.
571  *
572  * At a minimum one references should still be held.
573  */
574 static void svc_rdma_detach(struct svc_xprt *xprt)
575 {
576 	struct svcxprt_rdma *rdma =
577 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
578 
579 	/* Disconnect and flush posted WQE */
580 	rdma_disconnect(rdma->sc_cm_id);
581 }
582 
583 static void __svc_rdma_free(struct work_struct *work)
584 {
585 	struct svcxprt_rdma *rdma =
586 		container_of(work, struct svcxprt_rdma, sc_work);
587 	struct svc_xprt *xprt = &rdma->sc_xprt;
588 
589 	trace_svcrdma_xprt_free(xprt);
590 
591 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
592 		ib_drain_qp(rdma->sc_qp);
593 
594 	/* We should only be called from kref_put */
595 	if (kref_read(&xprt->xpt_ref) != 0)
596 		pr_err("svcrdma: sc_xprt still in use? (%d)\n",
597 		       kref_read(&xprt->xpt_ref));
598 
599 	svc_rdma_flush_recv_queues(rdma);
600 
601 	/* Final put of backchannel client transport */
602 	if (xprt->xpt_bc_xprt) {
603 		xprt_put(xprt->xpt_bc_xprt);
604 		xprt->xpt_bc_xprt = NULL;
605 	}
606 
607 	svc_rdma_destroy_rw_ctxts(rdma);
608 	svc_rdma_send_ctxts_destroy(rdma);
609 	svc_rdma_recv_ctxts_destroy(rdma);
610 
611 	/* Destroy the QP if present (not a listener) */
612 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
613 		ib_destroy_qp(rdma->sc_qp);
614 
615 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
616 		ib_free_cq(rdma->sc_sq_cq);
617 
618 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
619 		ib_free_cq(rdma->sc_rq_cq);
620 
621 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
622 		ib_dealloc_pd(rdma->sc_pd);
623 
624 	/* Destroy the CM ID */
625 	rdma_destroy_id(rdma->sc_cm_id);
626 
627 	kfree(rdma);
628 }
629 
630 static void svc_rdma_free(struct svc_xprt *xprt)
631 {
632 	struct svcxprt_rdma *rdma =
633 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
634 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
635 	queue_work(svc_rdma_wq, &rdma->sc_work);
636 }
637 
638 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
639 {
640 	struct svcxprt_rdma *rdma =
641 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
642 
643 	/*
644 	 * If there are already waiters on the SQ,
645 	 * return false.
646 	 */
647 	if (waitqueue_active(&rdma->sc_send_wait))
648 		return 0;
649 
650 	/* Otherwise return true. */
651 	return 1;
652 }
653 
654 static void svc_rdma_secure_port(struct svc_rqst *rqstp)
655 {
656 	set_bit(RQ_SECURE, &rqstp->rq_flags);
657 }
658 
659 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
660 {
661 }
662