svc_rdma_transport.c

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