• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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);
68 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
69 					struct net *net,
70 					struct sockaddr *sa, int salen,
71 					int flags);
72 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
73 static void svc_rdma_detach(struct svc_xprt *xprt);
74 static void svc_rdma_free(struct svc_xprt *xprt);
75 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
76 static void svc_rdma_secure_port(struct svc_rqst *);
77 static void svc_rdma_kill_temp_xprt(struct svc_xprt *);
78 
79 static const struct svc_xprt_ops svc_rdma_ops = {
80 	.xpo_create = svc_rdma_create,
81 	.xpo_recvfrom = svc_rdma_recvfrom,
82 	.xpo_sendto = svc_rdma_sendto,
83 	.xpo_read_payload = svc_rdma_read_payload,
84 	.xpo_release_rqst = svc_rdma_release_rqst,
85 	.xpo_detach = svc_rdma_detach,
86 	.xpo_free = svc_rdma_free,
87 	.xpo_has_wspace = svc_rdma_has_wspace,
88 	.xpo_accept = svc_rdma_accept,
89 	.xpo_secure_port = svc_rdma_secure_port,
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 */
qp_event_handler(struct ib_event * event,void * context)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 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
123 		svc_xprt_enqueue(xprt);
124 		break;
125 	}
126 }
127 
svc_rdma_create_xprt(struct svc_serv * serv,struct net * net)128 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
129 						 struct net *net)
130 {
131 	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
132 
133 	if (!cma_xprt) {
134 		dprintk("svcrdma: failed to create new transport\n");
135 		return NULL;
136 	}
137 	svc_xprt_init(net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
138 	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
139 	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
140 	INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
141 	INIT_LIST_HEAD(&cma_xprt->sc_send_ctxts);
142 	init_llist_head(&cma_xprt->sc_recv_ctxts);
143 	INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts);
144 	init_waitqueue_head(&cma_xprt->sc_send_wait);
145 
146 	spin_lock_init(&cma_xprt->sc_lock);
147 	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
148 	spin_lock_init(&cma_xprt->sc_send_lock);
149 	spin_lock_init(&cma_xprt->sc_rw_ctxt_lock);
150 
151 	/*
152 	 * Note that this implies that the underlying transport support
153 	 * has some form of congestion control (see RFC 7530 section 3.1
154 	 * paragraph 2). For now, we assume that all supported RDMA
155 	 * transports are suitable here.
156 	 */
157 	set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
158 
159 	return cma_xprt;
160 }
161 
162 static void
svc_rdma_parse_connect_private(struct svcxprt_rdma * newxprt,struct rdma_conn_param * param)163 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
164 			       struct rdma_conn_param *param)
165 {
166 	const struct rpcrdma_connect_private *pmsg = param->private_data;
167 
168 	if (pmsg &&
169 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
170 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
171 		newxprt->sc_snd_w_inv = pmsg->cp_flags &
172 					RPCRDMA_CMP_F_SND_W_INV_OK;
173 
174 		dprintk("svcrdma: client send_size %u, recv_size %u "
175 			"remote inv %ssupported\n",
176 			rpcrdma_decode_buffer_size(pmsg->cp_send_size),
177 			rpcrdma_decode_buffer_size(pmsg->cp_recv_size),
178 			newxprt->sc_snd_w_inv ? "" : "un");
179 	}
180 }
181 
182 /*
183  * This function handles the CONNECT_REQUEST event on a listening
184  * endpoint. It is passed the cma_id for the _new_ connection. The context in
185  * this cma_id is inherited from the listening cma_id and is the svc_xprt
186  * structure for the listening endpoint.
187  *
188  * This function creates a new xprt for the new connection and enqueues it on
189  * the accept queue for the listent xprt. When the listen thread is kicked, it
190  * will call the recvfrom method on the listen xprt which will accept the new
191  * connection.
192  */
handle_connect_req(struct rdma_cm_id * new_cma_id,struct rdma_conn_param * param)193 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
194 			       struct rdma_conn_param *param)
195 {
196 	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
197 	struct svcxprt_rdma *newxprt;
198 	struct sockaddr *sa;
199 
200 	/* Create a new transport */
201 	newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server,
202 				       listen_xprt->sc_xprt.xpt_net);
203 	if (!newxprt)
204 		return;
205 	newxprt->sc_cm_id = new_cma_id;
206 	new_cma_id->context = newxprt;
207 	svc_rdma_parse_connect_private(newxprt, param);
208 
209 	/* Save client advertised inbound read limit for use later in accept. */
210 	newxprt->sc_ord = param->initiator_depth;
211 
212 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
213 	newxprt->sc_xprt.xpt_remotelen = svc_addr_len(sa);
214 	memcpy(&newxprt->sc_xprt.xpt_remote, sa,
215 	       newxprt->sc_xprt.xpt_remotelen);
216 	snprintf(newxprt->sc_xprt.xpt_remotebuf,
217 		 sizeof(newxprt->sc_xprt.xpt_remotebuf) - 1, "%pISc", sa);
218 
219 	/* The remote port is arbitrary and not under the control of the
220 	 * client ULP. Set it to a fixed value so that the DRC continues
221 	 * to be effective after a reconnect.
222 	 */
223 	rpc_set_port((struct sockaddr *)&newxprt->sc_xprt.xpt_remote, 0);
224 
225 	sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
226 	svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
227 
228 	/*
229 	 * Enqueue the new transport on the accept queue of the listening
230 	 * transport
231 	 */
232 	spin_lock(&listen_xprt->sc_lock);
233 	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
234 	spin_unlock(&listen_xprt->sc_lock);
235 
236 	set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
237 	svc_xprt_enqueue(&listen_xprt->sc_xprt);
238 }
239 
240 /**
241  * svc_rdma_listen_handler - Handle CM events generated on a listening endpoint
242  * @cma_id: the server's listener rdma_cm_id
243  * @event: details of the event
244  *
245  * Return values:
246  *     %0: Do not destroy @cma_id
247  *     %1: Destroy @cma_id (never returned here)
248  *
249  * NB: There is never a DEVICE_REMOVAL event for INADDR_ANY listeners.
250  */
svc_rdma_listen_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)251 static int svc_rdma_listen_handler(struct rdma_cm_id *cma_id,
252 				   struct rdma_cm_event *event)
253 {
254 	switch (event->event) {
255 	case RDMA_CM_EVENT_CONNECT_REQUEST:
256 		handle_connect_req(cma_id, &event->param.conn);
257 		break;
258 	default:
259 		break;
260 	}
261 	return 0;
262 }
263 
264 /**
265  * svc_rdma_cma_handler - Handle CM events on client connections
266  * @cma_id: the server's listener rdma_cm_id
267  * @event: details of the event
268  *
269  * Return values:
270  *     %0: Do not destroy @cma_id
271  *     %1: Destroy @cma_id (never returned here)
272  */
svc_rdma_cma_handler(struct rdma_cm_id * cma_id,struct rdma_cm_event * event)273 static int svc_rdma_cma_handler(struct rdma_cm_id *cma_id,
274 				struct rdma_cm_event *event)
275 {
276 	struct svcxprt_rdma *rdma = cma_id->context;
277 	struct svc_xprt *xprt = &rdma->sc_xprt;
278 
279 	switch (event->event) {
280 	case RDMA_CM_EVENT_ESTABLISHED:
281 		clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
282 		svc_xprt_enqueue(xprt);
283 		break;
284 	case RDMA_CM_EVENT_DISCONNECTED:
285 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
286 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
287 		svc_xprt_enqueue(xprt);
288 		break;
289 	default:
290 		break;
291 	}
292 	return 0;
293 }
294 
295 /*
296  * Create a listening RDMA service endpoint.
297  */
svc_rdma_create(struct svc_serv * serv,struct net * net,struct sockaddr * sa,int salen,int flags)298 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
299 					struct net *net,
300 					struct sockaddr *sa, int salen,
301 					int flags)
302 {
303 	struct rdma_cm_id *listen_id;
304 	struct svcxprt_rdma *cma_xprt;
305 	int ret;
306 
307 	if (sa->sa_family != AF_INET && sa->sa_family != AF_INET6)
308 		return ERR_PTR(-EAFNOSUPPORT);
309 	cma_xprt = svc_rdma_create_xprt(serv, net);
310 	if (!cma_xprt)
311 		return ERR_PTR(-ENOMEM);
312 	set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
313 	strcpy(cma_xprt->sc_xprt.xpt_remotebuf, "listener");
314 
315 	listen_id = rdma_create_id(net, svc_rdma_listen_handler, cma_xprt,
316 				   RDMA_PS_TCP, IB_QPT_RC);
317 	if (IS_ERR(listen_id)) {
318 		ret = PTR_ERR(listen_id);
319 		goto err0;
320 	}
321 
322 	/* Allow both IPv4 and IPv6 sockets to bind a single port
323 	 * at the same time.
324 	 */
325 #if IS_ENABLED(CONFIG_IPV6)
326 	ret = rdma_set_afonly(listen_id, 1);
327 	if (ret)
328 		goto err1;
329 #endif
330 	ret = rdma_bind_addr(listen_id, sa);
331 	if (ret)
332 		goto err1;
333 	cma_xprt->sc_cm_id = listen_id;
334 
335 	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
336 	if (ret)
337 		goto err1;
338 
339 	/*
340 	 * We need to use the address from the cm_id in case the
341 	 * caller specified 0 for the port number.
342 	 */
343 	sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
344 	svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
345 
346 	return &cma_xprt->sc_xprt;
347 
348  err1:
349 	rdma_destroy_id(listen_id);
350  err0:
351 	kfree(cma_xprt);
352 	return ERR_PTR(ret);
353 }
354 
355 /*
356  * This is the xpo_recvfrom function for listening endpoints. Its
357  * purpose is to accept incoming connections. The CMA callback handler
358  * has already created a new transport and attached it to the new CMA
359  * ID.
360  *
361  * There is a queue of pending connections hung on the listening
362  * transport. This queue contains the new svc_xprt structure. This
363  * function takes svc_xprt structures off the accept_q and completes
364  * the connection.
365  */
svc_rdma_accept(struct svc_xprt * xprt)366 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
367 {
368 	struct svcxprt_rdma *listen_rdma;
369 	struct svcxprt_rdma *newxprt = NULL;
370 	struct rdma_conn_param conn_param;
371 	struct rpcrdma_connect_private pmsg;
372 	struct ib_qp_init_attr qp_attr;
373 	unsigned int ctxts, rq_depth;
374 	struct ib_device *dev;
375 	int ret = 0;
376 	RPC_IFDEBUG(struct sockaddr *sap);
377 
378 	listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
379 	clear_bit(XPT_CONN, &xprt->xpt_flags);
380 	/* Get the next entry off the accept list */
381 	spin_lock(&listen_rdma->sc_lock);
382 	if (!list_empty(&listen_rdma->sc_accept_q)) {
383 		newxprt = list_entry(listen_rdma->sc_accept_q.next,
384 				     struct svcxprt_rdma, sc_accept_q);
385 		list_del_init(&newxprt->sc_accept_q);
386 	}
387 	if (!list_empty(&listen_rdma->sc_accept_q))
388 		set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
389 	spin_unlock(&listen_rdma->sc_lock);
390 	if (!newxprt)
391 		return NULL;
392 
393 	dev = newxprt->sc_cm_id->device;
394 	newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
395 
396 	/* Qualify the transport resource defaults with the
397 	 * capabilities of this particular device */
398 	/* Transport header, head iovec, tail iovec */
399 	newxprt->sc_max_send_sges = 3;
400 	/* Add one SGE per page list entry */
401 	newxprt->sc_max_send_sges += (svcrdma_max_req_size / PAGE_SIZE) + 1;
402 	if (newxprt->sc_max_send_sges > dev->attrs.max_send_sge)
403 		newxprt->sc_max_send_sges = dev->attrs.max_send_sge;
404 	newxprt->sc_max_req_size = svcrdma_max_req_size;
405 	newxprt->sc_max_requests = svcrdma_max_requests;
406 	newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
407 	rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests;
408 	if (rq_depth > dev->attrs.max_qp_wr) {
409 		pr_warn("svcrdma: reducing receive depth to %d\n",
410 			dev->attrs.max_qp_wr);
411 		rq_depth = dev->attrs.max_qp_wr;
412 		newxprt->sc_max_requests = rq_depth - 2;
413 		newxprt->sc_max_bc_requests = 2;
414 	}
415 	newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
416 	ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
417 	ctxts *= newxprt->sc_max_requests;
418 	newxprt->sc_sq_depth = rq_depth + ctxts;
419 	if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
420 		pr_warn("svcrdma: reducing send depth to %d\n",
421 			dev->attrs.max_qp_wr);
422 		newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
423 	}
424 	atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
425 
426 	newxprt->sc_pd = ib_alloc_pd(dev, 0);
427 	if (IS_ERR(newxprt->sc_pd)) {
428 		trace_svcrdma_pd_err(newxprt, PTR_ERR(newxprt->sc_pd));
429 		goto errout;
430 	}
431 	newxprt->sc_sq_cq = ib_alloc_cq_any(dev, newxprt, newxprt->sc_sq_depth,
432 					    IB_POLL_WORKQUEUE);
433 	if (IS_ERR(newxprt->sc_sq_cq))
434 		goto errout;
435 	newxprt->sc_rq_cq =
436 		ib_alloc_cq_any(dev, newxprt, rq_depth, IB_POLL_WORKQUEUE);
437 	if (IS_ERR(newxprt->sc_rq_cq))
438 		goto errout;
439 
440 	memset(&qp_attr, 0, sizeof qp_attr);
441 	qp_attr.event_handler = qp_event_handler;
442 	qp_attr.qp_context = &newxprt->sc_xprt;
443 	qp_attr.port_num = newxprt->sc_port_num;
444 	qp_attr.cap.max_rdma_ctxs = ctxts;
445 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
446 	qp_attr.cap.max_recv_wr = rq_depth;
447 	qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges;
448 	qp_attr.cap.max_recv_sge = 1;
449 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
450 	qp_attr.qp_type = IB_QPT_RC;
451 	qp_attr.send_cq = newxprt->sc_sq_cq;
452 	qp_attr.recv_cq = newxprt->sc_rq_cq;
453 	dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
454 		newxprt->sc_cm_id, newxprt->sc_pd);
455 	dprintk("    cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
456 		qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
457 	dprintk("    cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
458 		qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
459 
460 	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
461 	if (ret) {
462 		trace_svcrdma_qp_err(newxprt, ret);
463 		goto errout;
464 	}
465 	newxprt->sc_qp = newxprt->sc_cm_id->qp;
466 
467 	if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
468 		newxprt->sc_snd_w_inv = false;
469 	if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
470 	    !rdma_ib_or_roce(dev, newxprt->sc_port_num)) {
471 		trace_svcrdma_fabric_err(newxprt, -EINVAL);
472 		goto errout;
473 	}
474 
475 	if (!svc_rdma_post_recvs(newxprt))
476 		goto errout;
477 
478 	/* Construct RDMA-CM private message */
479 	pmsg.cp_magic = rpcrdma_cmp_magic;
480 	pmsg.cp_version = RPCRDMA_CMP_VERSION;
481 	pmsg.cp_flags = 0;
482 	pmsg.cp_send_size = pmsg.cp_recv_size =
483 		rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
484 
485 	/* Accept Connection */
486 	set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
487 	memset(&conn_param, 0, sizeof conn_param);
488 	conn_param.responder_resources = 0;
489 	conn_param.initiator_depth = min_t(int, newxprt->sc_ord,
490 					   dev->attrs.max_qp_init_rd_atom);
491 	if (!conn_param.initiator_depth) {
492 		ret = -EINVAL;
493 		trace_svcrdma_initdepth_err(newxprt, ret);
494 		goto errout;
495 	}
496 	conn_param.private_data = &pmsg;
497 	conn_param.private_data_len = sizeof(pmsg);
498 	rdma_lock_handler(newxprt->sc_cm_id);
499 	newxprt->sc_cm_id->event_handler = svc_rdma_cma_handler;
500 	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
501 	rdma_unlock_handler(newxprt->sc_cm_id);
502 	if (ret) {
503 		trace_svcrdma_accept_err(newxprt, ret);
504 		goto errout;
505 	}
506 
507 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
508 	dprintk("svcrdma: new connection %p accepted:\n", newxprt);
509 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
510 	dprintk("    local address   : %pIS:%u\n", sap, rpc_get_port(sap));
511 	sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
512 	dprintk("    remote address  : %pIS:%u\n", sap, rpc_get_port(sap));
513 	dprintk("    max_sge         : %d\n", newxprt->sc_max_send_sges);
514 	dprintk("    sq_depth        : %d\n", newxprt->sc_sq_depth);
515 	dprintk("    rdma_rw_ctxs    : %d\n", ctxts);
516 	dprintk("    max_requests    : %d\n", newxprt->sc_max_requests);
517 	dprintk("    ord             : %d\n", conn_param.initiator_depth);
518 #endif
519 
520 	return &newxprt->sc_xprt;
521 
522  errout:
523 	/* Take a reference in case the DTO handler runs */
524 	svc_xprt_get(&newxprt->sc_xprt);
525 	if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
526 		ib_destroy_qp(newxprt->sc_qp);
527 	rdma_destroy_id(newxprt->sc_cm_id);
528 	/* This call to put will destroy the transport */
529 	svc_xprt_put(&newxprt->sc_xprt);
530 	return NULL;
531 }
532 
svc_rdma_detach(struct svc_xprt * xprt)533 static void svc_rdma_detach(struct svc_xprt *xprt)
534 {
535 	struct svcxprt_rdma *rdma =
536 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
537 
538 	rdma_disconnect(rdma->sc_cm_id);
539 }
540 
__svc_rdma_free(struct work_struct * work)541 static void __svc_rdma_free(struct work_struct *work)
542 {
543 	struct svcxprt_rdma *rdma =
544 		container_of(work, struct svcxprt_rdma, sc_work);
545 	struct svc_xprt *xprt = &rdma->sc_xprt;
546 
547 	/* This blocks until the Completion Queues are empty */
548 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
549 		ib_drain_qp(rdma->sc_qp);
550 
551 	svc_rdma_flush_recv_queues(rdma);
552 
553 	/* Final put of backchannel client transport */
554 	if (xprt->xpt_bc_xprt) {
555 		xprt_put(xprt->xpt_bc_xprt);
556 		xprt->xpt_bc_xprt = NULL;
557 	}
558 
559 	svc_rdma_destroy_rw_ctxts(rdma);
560 	svc_rdma_send_ctxts_destroy(rdma);
561 	svc_rdma_recv_ctxts_destroy(rdma);
562 
563 	/* Destroy the QP if present (not a listener) */
564 	if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
565 		ib_destroy_qp(rdma->sc_qp);
566 
567 	if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
568 		ib_free_cq(rdma->sc_sq_cq);
569 
570 	if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
571 		ib_free_cq(rdma->sc_rq_cq);
572 
573 	if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
574 		ib_dealloc_pd(rdma->sc_pd);
575 
576 	/* Destroy the CM ID */
577 	rdma_destroy_id(rdma->sc_cm_id);
578 
579 	kfree(rdma);
580 }
581 
svc_rdma_free(struct svc_xprt * xprt)582 static void svc_rdma_free(struct svc_xprt *xprt)
583 {
584 	struct svcxprt_rdma *rdma =
585 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
586 
587 	INIT_WORK(&rdma->sc_work, __svc_rdma_free);
588 	schedule_work(&rdma->sc_work);
589 }
590 
svc_rdma_has_wspace(struct svc_xprt * xprt)591 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
592 {
593 	struct svcxprt_rdma *rdma =
594 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
595 
596 	/*
597 	 * If there are already waiters on the SQ,
598 	 * return false.
599 	 */
600 	if (waitqueue_active(&rdma->sc_send_wait))
601 		return 0;
602 
603 	/* Otherwise return true. */
604 	return 1;
605 }
606 
svc_rdma_secure_port(struct svc_rqst * rqstp)607 static void svc_rdma_secure_port(struct svc_rqst *rqstp)
608 {
609 	set_bit(RQ_SECURE, &rqstp->rq_flags);
610 }
611 
svc_rdma_kill_temp_xprt(struct svc_xprt * xprt)612 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)
613 {
614 }
615