1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
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.
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
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
17 * Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
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.
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
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.
42 * Author: Tom Tucker <tom@opengridcomputing.com>
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>
52 #include <rdma/ib_verbs.h>
53 #include <rdma/rdma_cm.h>
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>
62 #include "xprt_rdma.h"
63 #include <trace/events/rpcrdma.h>
65 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
67 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
69 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
71 struct sockaddr *sa, int salen,
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 *);
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,
95 struct svc_xprt_class svc_rdma_class = {
97 .xcl_owner = THIS_MODULE,
98 .xcl_ops = &svc_rdma_ops,
99 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
100 .xcl_ident = XPRT_TRANSPORT_RDMA,
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 *);
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,
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)
124 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
126 struct sockaddr *sa, int salen,
129 struct svcxprt_rdma *cma_xprt;
130 struct svc_xprt *xprt;
132 cma_xprt = svc_rdma_create_xprt(serv, net);
134 return ERR_PTR(-ENOMEM);
135 xprt = &cma_xprt->sc_xprt;
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;
141 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
145 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
147 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
150 static void svc_rdma_bc_free(struct svc_xprt *xprt)
152 struct svcxprt_rdma *rdma =
153 container_of(xprt, struct svcxprt_rdma, sc_xprt);
155 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
159 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
161 /* QP event handler */
162 static void qp_event_handler(struct ib_event *event, void *context)
164 struct svc_xprt *xprt = context;
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:
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:
182 set_bit(XPT_CLOSE, &xprt->xpt_flags);
183 svc_xprt_enqueue(xprt);
188 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv,
191 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
194 dprintk("svcrdma: failed to create new transport\n");
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);
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);
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.
218 set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags);
224 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt,
225 struct rdma_conn_param *param)
227 const struct rpcrdma_connect_private *pmsg = param->private_data;
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;
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");
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.
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
254 static void handle_connect_req(struct rdma_cm_id *new_cma_id,
255 struct rdma_conn_param *param)
257 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
258 struct svcxprt_rdma *newxprt;
261 /* Create a new transport */
262 newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server,
263 listen_xprt->sc_xprt.xpt_net);
266 newxprt->sc_cm_id = new_cma_id;
267 new_cma_id->context = newxprt;
268 svc_rdma_parse_connect_private(newxprt, param);
270 /* Save client advertised inbound read limit for use later in accept. */
271 newxprt->sc_ord = param->initiator_depth;
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));
280 * Enqueue the new transport on the accept queue of the listening
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);
287 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
288 svc_xprt_enqueue(&listen_xprt->sc_xprt);
292 * Handles events generated on the listening endpoint. These events will be
293 * either be incoming connect requests or adapter removal events.
295 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
296 struct rdma_cm_event *event)
298 struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.src_addr;
300 trace_svcrdma_cm_event(event, sap);
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);
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);
320 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
321 struct rdma_cm_event *event)
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;
327 trace_svcrdma_cm_event(event, sap);
329 switch (event->event) {
330 case RDMA_CM_EVENT_ESTABLISHED:
331 /* Accept complete */
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);
338 case RDMA_CM_EVENT_DISCONNECTED:
339 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
341 set_bit(XPT_CLOSE, &xprt->xpt_flags);
342 svc_xprt_enqueue(xprt);
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);
354 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
355 "event = %s (%d)\n", cma_id,
356 rdma_event_msg(event->event), event->event);
363 * Create a listening RDMA service endpoint.
365 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
367 struct sockaddr *sa, int salen,
370 struct rdma_cm_id *listen_id;
371 struct svcxprt_rdma *cma_xprt;
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);
379 cma_xprt = svc_rdma_create_xprt(serv, net);
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");
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);
393 /* Allow both IPv4 and IPv6 sockets to bind a single port
396 #if IS_ENABLED(CONFIG_IPV6)
397 ret = rdma_set_afonly(listen_id, 1);
399 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret);
403 ret = rdma_bind_addr(listen_id, sa);
405 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
408 cma_xprt->sc_cm_id = listen_id;
410 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
412 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
417 * We need to use the address from the cm_id in case the
418 * caller specified 0 for the port number.
420 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
421 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
423 return &cma_xprt->sc_xprt;
426 rdma_destroy_id(listen_id);
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
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
443 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
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;
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);
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);
470 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
471 newxprt, newxprt->sc_cm_id);
473 dev = newxprt->sc_cm_id->device;
474 newxprt->sc_port_num = newxprt->sc_cm_id->port_num;
476 /* Qualify the transport resource defaults with the
477 * capabilities of this particular device */
478 /* Transport header, head iovec, tail iovec */
479 newxprt->sc_max_send_sges = 3;
480 /* Add one SGE per page list entry */
481 newxprt->sc_max_send_sges += (svcrdma_max_req_size / PAGE_SIZE) + 1;
482 if (newxprt->sc_max_send_sges > dev->attrs.max_send_sge)
483 newxprt->sc_max_send_sges = dev->attrs.max_send_sge;
484 newxprt->sc_max_req_size = svcrdma_max_req_size;
485 newxprt->sc_max_requests = svcrdma_max_requests;
486 newxprt->sc_max_bc_requests = svcrdma_max_bc_requests;
487 rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests;
488 if (rq_depth > dev->attrs.max_qp_wr) {
489 pr_warn("svcrdma: reducing receive depth to %d\n",
490 dev->attrs.max_qp_wr);
491 rq_depth = dev->attrs.max_qp_wr;
492 newxprt->sc_max_requests = rq_depth - 2;
493 newxprt->sc_max_bc_requests = 2;
495 newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests);
496 ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES);
497 ctxts *= newxprt->sc_max_requests;
498 newxprt->sc_sq_depth = rq_depth + ctxts;
499 if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) {
500 pr_warn("svcrdma: reducing send depth to %d\n",
501 dev->attrs.max_qp_wr);
502 newxprt->sc_sq_depth = dev->attrs.max_qp_wr;
504 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth);
506 newxprt->sc_pd = ib_alloc_pd(dev, 0);
507 if (IS_ERR(newxprt->sc_pd)) {
508 dprintk("svcrdma: error creating PD for connect request\n");
511 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
512 0, IB_POLL_WORKQUEUE);
513 if (IS_ERR(newxprt->sc_sq_cq)) {
514 dprintk("svcrdma: error creating SQ CQ for connect request\n");
517 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, rq_depth,
518 0, IB_POLL_WORKQUEUE);
519 if (IS_ERR(newxprt->sc_rq_cq)) {
520 dprintk("svcrdma: error creating RQ CQ for connect request\n");
524 memset(&qp_attr, 0, sizeof qp_attr);
525 qp_attr.event_handler = qp_event_handler;
526 qp_attr.qp_context = &newxprt->sc_xprt;
527 qp_attr.port_num = newxprt->sc_port_num;
528 qp_attr.cap.max_rdma_ctxs = ctxts;
529 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
530 qp_attr.cap.max_recv_wr = rq_depth;
531 qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges;
532 qp_attr.cap.max_recv_sge = 1;
533 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
534 qp_attr.qp_type = IB_QPT_RC;
535 qp_attr.send_cq = newxprt->sc_sq_cq;
536 qp_attr.recv_cq = newxprt->sc_rq_cq;
537 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n",
538 newxprt->sc_cm_id, newxprt->sc_pd);
539 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n",
540 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr);
541 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n",
542 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge);
544 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
546 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
549 newxprt->sc_qp = newxprt->sc_cm_id->qp;
551 if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
552 newxprt->sc_snd_w_inv = false;
553 if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) &&
554 !rdma_ib_or_roce(dev, newxprt->sc_port_num))
557 if (!svc_rdma_post_recvs(newxprt))
560 /* Swap out the handler */
561 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
563 /* Construct RDMA-CM private message */
564 pmsg.cp_magic = rpcrdma_cmp_magic;
565 pmsg.cp_version = RPCRDMA_CMP_VERSION;
567 pmsg.cp_send_size = pmsg.cp_recv_size =
568 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size);
570 /* Accept Connection */
571 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
572 memset(&conn_param, 0, sizeof conn_param);
573 conn_param.responder_resources = 0;
574 conn_param.initiator_depth = min_t(int, newxprt->sc_ord,
575 dev->attrs.max_qp_init_rd_atom);
576 if (!conn_param.initiator_depth) {
577 dprintk("svcrdma: invalid ORD setting\n");
581 conn_param.private_data = &pmsg;
582 conn_param.private_data_len = sizeof(pmsg);
583 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
587 dprintk("svcrdma: new connection %p accepted:\n", newxprt);
588 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
589 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap));
590 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
591 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap));
592 dprintk(" max_sge : %d\n", newxprt->sc_max_send_sges);
593 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth);
594 dprintk(" rdma_rw_ctxs : %d\n", ctxts);
595 dprintk(" max_requests : %d\n", newxprt->sc_max_requests);
596 dprintk(" ord : %d\n", conn_param.initiator_depth);
598 trace_svcrdma_xprt_accept(&newxprt->sc_xprt);
599 return &newxprt->sc_xprt;
602 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
603 trace_svcrdma_xprt_fail(&newxprt->sc_xprt);
604 /* Take a reference in case the DTO handler runs */
605 svc_xprt_get(&newxprt->sc_xprt);
606 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
607 ib_destroy_qp(newxprt->sc_qp);
608 rdma_destroy_id(newxprt->sc_cm_id);
609 /* This call to put will destroy the transport */
610 svc_xprt_put(&newxprt->sc_xprt);
614 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
619 * When connected, an svc_xprt has at least two references:
621 * - A reference held by the cm_id between the ESTABLISHED and
622 * DISCONNECTED events. If the remote peer disconnected first, this
623 * reference could be gone.
625 * - A reference held by the svc_recv code that called this function
626 * as part of close processing.
628 * At a minimum one references should still be held.
630 static void svc_rdma_detach(struct svc_xprt *xprt)
632 struct svcxprt_rdma *rdma =
633 container_of(xprt, struct svcxprt_rdma, sc_xprt);
635 /* Disconnect and flush posted WQE */
636 rdma_disconnect(rdma->sc_cm_id);
639 static void __svc_rdma_free(struct work_struct *work)
641 struct svcxprt_rdma *rdma =
642 container_of(work, struct svcxprt_rdma, sc_work);
643 struct svc_xprt *xprt = &rdma->sc_xprt;
645 trace_svcrdma_xprt_free(xprt);
647 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
648 ib_drain_qp(rdma->sc_qp);
650 /* We should only be called from kref_put */
651 if (kref_read(&xprt->xpt_ref) != 0)
652 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
653 kref_read(&xprt->xpt_ref));
655 svc_rdma_flush_recv_queues(rdma);
657 /* Final put of backchannel client transport */
658 if (xprt->xpt_bc_xprt) {
659 xprt_put(xprt->xpt_bc_xprt);
660 xprt->xpt_bc_xprt = NULL;
663 svc_rdma_destroy_rw_ctxts(rdma);
664 svc_rdma_send_ctxts_destroy(rdma);
665 svc_rdma_recv_ctxts_destroy(rdma);
667 /* Destroy the QP if present (not a listener) */
668 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
669 ib_destroy_qp(rdma->sc_qp);
671 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
672 ib_free_cq(rdma->sc_sq_cq);
674 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
675 ib_free_cq(rdma->sc_rq_cq);
677 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
678 ib_dealloc_pd(rdma->sc_pd);
680 /* Destroy the CM ID */
681 rdma_destroy_id(rdma->sc_cm_id);
686 static void svc_rdma_free(struct svc_xprt *xprt)
688 struct svcxprt_rdma *rdma =
689 container_of(xprt, struct svcxprt_rdma, sc_xprt);
690 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
691 queue_work(svc_rdma_wq, &rdma->sc_work);
694 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
696 struct svcxprt_rdma *rdma =
697 container_of(xprt, struct svcxprt_rdma, sc_xprt);
700 * If there are already waiters on the SQ,
703 if (waitqueue_active(&rdma->sc_send_wait))
706 /* Otherwise return true. */
710 static void svc_rdma_secure_port(struct svc_rqst *rqstp)
712 set_bit(RQ_SECURE, &rqstp->rq_flags);
715 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt)