xref: /openbmc/linux/net/sunrpc/xprtrdma/svc_rdma_transport.c (revision 2e7c04aec86758e0adfcad4a24c86593b45807a3)
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_prep_reply_hdr = svc_rdma_prep_reply_hdr,
89 	.xpo_has_wspace = svc_rdma_has_wspace,
90 	.xpo_accept = svc_rdma_accept,
91 	.xpo_secure_port = svc_rdma_secure_port,
92 	.xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt,
93 };
94 
95 struct svc_xprt_class svc_rdma_class = {
96 	.xcl_name = "rdma",
97 	.xcl_owner = THIS_MODULE,
98 	.xcl_ops = &svc_rdma_ops,
99 	.xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
100 	.xcl_ident = XPRT_TRANSPORT_RDMA,
101 };
102 
103 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
104 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
105 					   struct sockaddr *, int, int);
106 static void svc_rdma_bc_detach(struct svc_xprt *);
107 static void svc_rdma_bc_free(struct svc_xprt *);
108 
109 static const struct svc_xprt_ops svc_rdma_bc_ops = {
110 	.xpo_create = svc_rdma_bc_create,
111 	.xpo_detach = svc_rdma_bc_detach,
112 	.xpo_free = svc_rdma_bc_free,
113 	.xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
114 	.xpo_secure_port = svc_rdma_secure_port,
115 };
116 
117 struct svc_xprt_class svc_rdma_bc_class = {
118 	.xcl_name = "rdma-bc",
119 	.xcl_owner = THIS_MODULE,
120 	.xcl_ops = &svc_rdma_bc_ops,
121 	.xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
122 };
123 
124 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
125 					   struct net *net,
126 					   struct sockaddr *sa, int salen,
127 					   int flags)
128 {
129 	struct svcxprt_rdma *cma_xprt;
130 	struct svc_xprt *xprt;
131 
132 	cma_xprt = svc_rdma_create_xprt(serv, net);
133 	if (!cma_xprt)
134 		return ERR_PTR(-ENOMEM);
135 	xprt = &cma_xprt->sc_xprt;
136 
137 	svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
138 	set_bit(XPT_CONG_CTRL, &xprt->xpt_flags);
139 	serv->sv_bc_xprt = xprt;
140 
141 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
142 	return xprt;
143 }
144 
145 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
146 {
147 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
148 }
149 
150 static void svc_rdma_bc_free(struct svc_xprt *xprt)
151 {
152 	struct svcxprt_rdma *rdma =
153 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
154 
155 	dprintk("svcrdma: %s(%p)\n", __func__, xprt);
156 	if (xprt)
157 		kfree(rdma);
158 }
159 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
160 
161 /* QP event handler */
162 static void qp_event_handler(struct ib_event *event, void *context)
163 {
164 	struct svc_xprt *xprt = context;
165 
166 	trace_svcrdma_qp_error(event, (struct sockaddr *)&xprt->xpt_remote);
167 	switch (event->event) {
168 	/* These are considered benign events */
169 	case IB_EVENT_PATH_MIG:
170 	case IB_EVENT_COMM_EST:
171 	case IB_EVENT_SQ_DRAINED:
172 	case IB_EVENT_QP_LAST_WQE_REACHED:
173 		break;
174 
175 	/* These are considered fatal events */
176 	case IB_EVENT_PATH_MIG_ERR:
177 	case IB_EVENT_QP_FATAL:
178 	case IB_EVENT_QP_REQ_ERR:
179 	case IB_EVENT_QP_ACCESS_ERR:
180 	case IB_EVENT_DEVICE_FATAL:
181 	default:
182 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
183 		svc_xprt_enqueue(xprt);
184 		break;
185 	}
186 }
187 
188 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
189 						 struct net *net)
190 {
191 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
192 
193 	if (!cma_xprt) {
194 		dprintk("svcrdma: failed to create new transport\n");
195 		return NULL;
196 	}
197 	svc_xprt_init(net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
198 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
199 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
200 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
201 	INIT_LIST_HEAD(&cma_xprt->sc_send_ctxts);
202 	INIT_LIST_HEAD(&cma_xprt->sc_recv_ctxts);
203 	INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
204 	init_waitqueue_head(&cma_xprt->sc_send_wait);
205 
206 	spin_lock_init(&cma_xprt->sc_lock);
207 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
208 	spin_lock_init(&cma_xprt->sc_send_lock);
209 	spin_lock_init(&cma_xprt->sc_recv_lock);
210 	spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
211 
212 	/*
213 	 * Note that this implies that the underlying transport support
214 	 * has some form of congestion control (see RFC 7530 section 3.1
215 	 * paragraph 2). For now, we assume that all supported RDMA
216 	 * transports are suitable here.
217 	 */
218 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
219 
220 	return cma_xprt;
221 }
222 
223 static void
224 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
225 			       struct rdma_conn_param *param)
226 {
227 	const struct rpcrdma_connect_private *pmsg = param->private_data;
228 
229 	if (pmsg &&
230 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
231 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
232 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
233 					RPCRDMA_CMP_F_SND_W_INV_OK;
234 
235 		dprintk("svcrdma: client send_size %u, recv_size %u "
236 			"remote inv %ssupported\n",
237 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
238 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
239 			newxprt->sc_snd_w_inv ? "" : "un");
240 	}
241 }
242 
243 /*
244  * This function handles the CONNECT_REQUEST event on a listening
245  * endpoint. It is passed the cma_id for the _new_ connection. The context in
246  * this cma_id is inherited from the listening cma_id and is the svc_xprt
247  * structure for the listening endpoint.
248  *
249  * This function creates a new xprt for the new connection and enqueues it on
250  * the accept queue for the listent xprt. When the listen thread is kicked, it
251  * will call the recvfrom method on the listen xprt which will accept the new
252  * connection.
253  */
254 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
255 			       struct rdma_conn_param *param)
256 {
257 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
258 	struct svcxprt_rdma *newxprt;
259 	struct sockaddr *sa;
260 
261 	/* Create a new transport */
262 	newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server,
263 				       listen_xprt->sc_xprt.xpt_net);
264 	if (!newxprt)
265 		return;
266 	newxprt->sc_cm_id = new_cma_id;
267 	new_cma_id->context = newxprt;
268 	svc_rdma_parse_connect_private(newxprt, param);
269 
270 	/* Save client advertised inbound read limit for use later in accept. */
271 	newxprt->sc_ord = param->initiator_depth;
272 
273 	/* Set the local and remote addresses in the transport */
274 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
275 	svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
276 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
277 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
278 
279 	/*
280 	 * Enqueue the new transport on the accept queue of the listening
281 	 * transport
282 	 */
283 	spin_lock_bh(&listen_xprt->sc_lock);
284 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
285 	spin_unlock_bh(&listen_xprt->sc_lock);
286 
287 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
288 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
289 }
290 
291 /*
292  * Handles events generated on the listening endpoint. These events will be
293  * either be incoming connect requests or adapter removal  events.
294  */
295 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
296 			       struct rdma_cm_event *event)
297 {
298 	struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.src_addr;
299 
300 	trace_svcrdma_cm_event(event, sap);
301 
302 	switch (event->event) {
303 	case RDMA_CM_EVENT_CONNECT_REQUEST:
304 		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
305 			"event = %s (%d)\n", cma_id, cma_id->context,
306 			rdma_event_msg(event->event), event->event);
307 		handle_connect_req(cma_id, &event->param.conn);
308 		break;
309 	default:
310 		/* NB: No device removal upcall for INADDR_ANY listeners */
311 		dprintk("svcrdma: Unexpected event on listening endpoint %p, "
312 			"event = %s (%d)\n", cma_id,
313 			rdma_event_msg(event->event), event->event);
314 		break;
315 	}
316 
317 	return 0;
318 }
319 
320 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
321 			    struct rdma_cm_event *event)
322 {
323 	struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.dst_addr;
324 	struct svcxprt_rdma *rdma = cma_id->context;
325 	struct svc_xprt *xprt = &rdma->sc_xprt;
326 
327 	trace_svcrdma_cm_event(event, sap);
328 
329 	switch (event->event) {
330 	case RDMA_CM_EVENT_ESTABLISHED:
331 		/* Accept complete */
332 		svc_xprt_get(xprt);
333 		dprintk("svcrdma: Connection completed on DTO xprt=%p, "
334 			"cm_id=%p\n", xprt, cma_id);
335 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
336 		svc_xprt_enqueue(xprt);
337 		break;
338 	case RDMA_CM_EVENT_DISCONNECTED:
339 		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
340 			xprt, cma_id);
341 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
342 		svc_xprt_enqueue(xprt);
343 		svc_xprt_put(xprt);
344 		break;
345 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
346 		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
347 			"event = %s (%d)\n", cma_id, xprt,
348 			rdma_event_msg(event->event), event->event);
349 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
350 		svc_xprt_enqueue(xprt);
351 		svc_xprt_put(xprt);
352 		break;
353 	default:
354 		dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
355 			"event = %s (%d)\n", cma_id,
356 			rdma_event_msg(event->event), event->event);
357 		break;
358 	}
359 	return 0;
360 }
361 
362 /*
363  * Create a listening RDMA service endpoint.
364  */
365 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
366 					struct net *net,
367 					struct sockaddr *sa, int salen,
368 					int flags)
369 {
370 	struct rdma_cm_id *listen_id;
371 	struct svcxprt_rdma *cma_xprt;
372 	int ret;
373 
374 	dprintk("svcrdma: Creating RDMA listener\n");
375 	if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) {
376 		dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
377 		return ERR_PTR(-EAFNOSUPPORT);
378 	}
379 	cma_xprt = svc_rdma_create_xprt(serv, net);
380 	if (!cma_xprt)
381 		return ERR_PTR(-ENOMEM);
382 	set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
383 	strcpy(cma_xprt->sc_xprt.xpt_remotebuf, "listener");
384 
385 	listen_id = rdma_create_id(net, rdma_listen_handler, cma_xprt,
386 				   RDMA_PS_TCP, IB_QPT_RC);
387 	if (IS_ERR(listen_id)) {
388 		ret = PTR_ERR(listen_id);
389 		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
390 		goto err0;
391 	}
392 
393 	/* Allow both IPv4 and IPv6 sockets to bind a single port
394 	 * at the same time.
395 	 */
396 #if IS_ENABLED(CONFIG_IPV6)
397 	ret = rdma_set_afonly(listen_id, 1);
398 	if (ret) {
399 		dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
400 		goto err1;
401 	}
402 #endif
403 	ret = rdma_bind_addr(listen_id, sa);
404 	if (ret) {
405 		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
406 		goto err1;
407 	}
408 	cma_xprt->sc_cm_id = listen_id;
409 
410 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
411 	if (ret) {
412 		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
413 		goto err1;
414 	}
415 
416 	/*
417 	 * We need to use the address from the cm_id in case the
418 	 * caller specified 0 for the port number.
419 	 */
420 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
421 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
422 
423 	return &cma_xprt->sc_xprt;
424 
425  err1:
426 	rdma_destroy_id(listen_id);
427  err0:
428 	kfree(cma_xprt);
429 	return ERR_PTR(ret);
430 }
431 
432 /*
433  * This is the xpo_recvfrom function for listening endpoints. Its
434  * purpose is to accept incoming connections. The CMA callback handler
435  * has already created a new transport and attached it to the new CMA
436  * ID.
437  *
438  * There is a queue of pending connections hung on the listening
439  * transport. This queue contains the new svc_xprt structure. This
440  * function takes svc_xprt structures off the accept_q and completes
441  * the connection.
442  */
443 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
444 {
445 	struct svcxprt_rdma *listen_rdma;
446 	struct svcxprt_rdma *newxprt = NULL;
447 	struct rdma_conn_param conn_param;
448 	struct rpcrdma_connect_private pmsg;
449 	struct ib_qp_init_attr qp_attr;
450 	unsigned int ctxts, rq_depth;
451 	struct ib_device *dev;
452 	struct sockaddr *sap;
453 	int ret = 0;
454 
455 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
456 	clear_bit(XPT_CONN, &xprt->xpt_flags);
457 	/* Get the next entry off the accept list */
458 	spin_lock_bh(&listen_rdma->sc_lock);
459 	if (!list_empty(&listen_rdma->sc_accept_q)) {
460 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
461 				     struct svcxprt_rdma, sc_accept_q);
462 		list_del_init(&newxprt->sc_accept_q);
463 	}
464 	if (!list_empty(&listen_rdma->sc_accept_q))
465 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
466 	spin_unlock_bh(&listen_rdma->sc_lock);
467 	if (!newxprt)
468 		return NULL;
469 
470 	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
471 		newxprt, newxprt->sc_cm_id);
472 
473 	dev = newxprt->sc_cm_id->device;
474 	newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
475 
476 	/* Qualify the transport resource defaults with the
477 	 * capabilities of this particular device */
478 	newxprt->sc_max_send_sges = dev->attrs.max_send_sge;
479 	/* transport hdr, head iovec, one page list entry, tail iovec */
480 	if (newxprt->sc_max_send_sges < 4) {
481 		pr_err("svcrdma: too few Send SGEs available (%d)\n",
482 		       newxprt->sc_max_send_sges);
483 		goto errout;
484 	}
485 	newxprt->sc_max_req_size = svcrdma_max_req_size;
486 	newxprt->sc_max_requests = svcrdma_max_requests;
487 	newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
488 	rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests;
489 	if (rq_depth > dev->attrs.max_qp_wr) {
490 		pr_warn("svcrdma: reducing receive depth to %d\n",
491 			dev->attrs.max_qp_wr);
492 		rq_depth = dev->attrs.max_qp_wr;
493 		newxprt->sc_max_requests = rq_depth - 2;
494 		newxprt->sc_max_bc_requests = 2;
495 	}
496 	newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
497 	ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
498 	ctxts *= newxprt->sc_max_requests;
499 	newxprt->sc_sq_depth = rq_depth + ctxts;
500 	if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
501 		pr_warn("svcrdma: reducing send depth to %d\n",
502 			dev->attrs.max_qp_wr);
503 		newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
504 	}
505 	atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
506 
507 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
508 	if (IS_ERR(newxprt->sc_pd)) {
509 		dprintk("svcrdma: error creating PD for connect request\n");
510 		goto errout;
511 	}
512 	newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
513 					0, IB_POLL_WORKQUEUE);
514 	if (IS_ERR(newxprt->sc_sq_cq)) {
515 		dprintk("svcrdma: error creating SQ CQ for connect request\n");
516 		goto errout;
517 	}
518 	newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, rq_depth,
519 					0, IB_POLL_WORKQUEUE);
520 	if (IS_ERR(newxprt->sc_rq_cq)) {
521 		dprintk("svcrdma: error creating RQ CQ for connect request\n");
522 		goto errout;
523 	}
524 
525 	memset(&qp_attr, 0, sizeof qp_attr);
526 	qp_attr.event_handler = qp_event_handler;
527 	qp_attr.qp_context = &newxprt->sc_xprt;
528 	qp_attr.port_num = newxprt->sc_port_num;
529 	qp_attr.cap.max_rdma_ctxs = ctxts;
530 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
531 	qp_attr.cap.max_recv_wr = rq_depth;
532 	qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges;
533 	qp_attr.cap.max_recv_sge = 1;
534 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
535 	qp_attr.qp_type = IB_QPT_RC;
536 	qp_attr.send_cq = newxprt->sc_sq_cq;
537 	qp_attr.recv_cq = newxprt->sc_rq_cq;
538 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
539 		newxprt->sc_cm_id, newxprt->sc_pd);
540 	dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
541 		qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
542 	dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
543 		qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
544 
545 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
546 	if (ret) {
547 		dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
548 		goto errout;
549 	}
550 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
551 
552 	if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
553 		newxprt->sc_snd_w_inv = false;
554 	if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
555 	    !rdma_ib_or_roce(dev, newxprt->sc_port_num))
556 		goto errout;
557 
558 	if (!svc_rdma_post_recvs(newxprt))
559 		goto errout;
560 
561 	/* Swap out the handler */
562 	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
563 
564 	/* Construct RDMA-CM private message */
565 	pmsg.cp_magic = rpcrdma_cmp_magic;
566 	pmsg.cp_version = RPCRDMA_CMP_VERSION;
567 	pmsg.cp_flags = 0;
568 	pmsg.cp_send_size = pmsg.cp_recv_size =
569 		rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
570 
571 	/* Accept Connection */
572 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
573 	memset(&conn_param, 0, sizeof conn_param);
574 	conn_param.responder_resources = 0;
575 	conn_param.initiator_depth = min_t(int, newxprt->sc_ord,
576 					   dev->attrs.max_qp_init_rd_atom);
577 	if (!conn_param.initiator_depth) {
578 		dprintk("svcrdma: invalid ORD setting\n");
579 		ret = -EINVAL;
580 		goto errout;
581 	}
582 	conn_param.private_data = &pmsg;
583 	conn_param.private_data_len = sizeof(pmsg);
584 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
585 	if (ret)
586 		goto errout;
587 
588 	dprintk("svcrdma: new connection %p accepted:\n", newxprt);
589 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
590 	dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
591 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
592 	dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
593 	dprintk("    max_sge         : %d\n", newxprt->sc_max_send_sges);
594 	dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
595 	dprintk("    rdma_rw_ctxs    : %d\n", ctxts);
596 	dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
597 	dprintk("    ord             : %d\n", conn_param.initiator_depth);
598 
599 	trace_svcrdma_xprt_accept(&newxprt->sc_xprt);
600 	return &newxprt->sc_xprt;
601 
602  errout:
603 	dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
604 	trace_svcrdma_xprt_fail(&newxprt->sc_xprt);
605 	/* Take a reference in case the DTO handler runs */
606 	svc_xprt_get(&newxprt->sc_xprt);
607 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
608 		ib_destroy_qp(newxprt->sc_qp);
609 	rdma_destroy_id(newxprt->sc_cm_id);
610 	/* This call to put will destroy the transport */
611 	svc_xprt_put(&newxprt->sc_xprt);
612 	return NULL;
613 }
614 
615 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
616 {
617 }
618 
619 /*
620  * When connected, an svc_xprt has at least two references:
621  *
622  * - A reference held by the cm_id between the ESTABLISHED and
623  *   DISCONNECTED events. If the remote peer disconnected first, this
624  *   reference could be gone.
625  *
626  * - A reference held by the svc_recv code that called this function
627  *   as part of close processing.
628  *
629  * At a minimum one references should still be held.
630  */
631 static void svc_rdma_detach(struct svc_xprt *xprt)
632 {
633 	struct svcxprt_rdma *rdma =
634 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
635 
636 	/* Disconnect and flush posted WQE */
637 	rdma_disconnect(rdma->sc_cm_id);
638 }
639 
640 static void __svc_rdma_free(struct work_struct *work)
641 {
642 	struct svcxprt_rdma *rdma =
643 		container_of(work, struct svcxprt_rdma, sc_work);
644 	struct svc_xprt *xprt = &rdma->sc_xprt;
645 
646 	trace_svcrdma_xprt_free(xprt);
647 
648 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
649 		ib_drain_qp(rdma->sc_qp);
650 
651 	/* We should only be called from kref_put */
652 	if (kref_read(&xprt->xpt_ref) != 0)
653 		pr_err("svcrdma: sc_xprt still in use? (%d)\n",
654 		       kref_read(&xprt->xpt_ref));
655 
656 	svc_rdma_flush_recv_queues(rdma);
657 
658 	/* Final put of backchannel client transport */
659 	if (xprt->xpt_bc_xprt) {
660 		xprt_put(xprt->xpt_bc_xprt);
661 		xprt->xpt_bc_xprt = NULL;
662 	}
663 
664 	svc_rdma_destroy_rw_ctxts(rdma);
665 	svc_rdma_send_ctxts_destroy(rdma);
666 	svc_rdma_recv_ctxts_destroy(rdma);
667 
668 	/* Destroy the QP if present (not a listener) */
669 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
670 		ib_destroy_qp(rdma->sc_qp);
671 
672 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
673 		ib_free_cq(rdma->sc_sq_cq);
674 
675 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
676 		ib_free_cq(rdma->sc_rq_cq);
677 
678 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
679 		ib_dealloc_pd(rdma->sc_pd);
680 
681 	/* Destroy the CM ID */
682 	rdma_destroy_id(rdma->sc_cm_id);
683 
684 	kfree(rdma);
685 }
686 
687 static void svc_rdma_free(struct svc_xprt *xprt)
688 {
689 	struct svcxprt_rdma *rdma =
690 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
691 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
692 	queue_work(svc_rdma_wq, &rdma->sc_work);
693 }
694 
695 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
696 {
697 	struct svcxprt_rdma *rdma =
698 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
699 
700 	/*
701 	 * If there are already waiters on the SQ,
702 	 * return false.
703 	 */
704 	if (waitqueue_active(&rdma->sc_send_wait))
705 		return 0;
706 
707 	/* Otherwise return true. */
708 	return 1;
709 }
710 
711 static void svc_rdma_secure_port(struct svc_rqst *rqstp)
712 {
713 	set_bit(RQ_SECURE, &rqstp->rq_flags);
714 }
715 
716 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
717 {
718 }
719