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