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_has_wspace = svc_rdma_has_wspace, 89 .xpo_accept = svc_rdma_accept, 90 .xpo_secure_port = svc_rdma_secure_port, 91 .xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt, 92 }; 93 94 struct svc_xprt_class svc_rdma_class = { 95 .xcl_name = "rdma", 96 .xcl_owner = THIS_MODULE, 97 .xcl_ops = &svc_rdma_ops, 98 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA, 99 .xcl_ident = XPRT_TRANSPORT_RDMA, 100 }; 101 102 /* QP event handler */ 103 static void qp_event_handler(struct ib_event *event, void *context) 104 { 105 struct svc_xprt *xprt = context; 106 107 trace_svcrdma_qp_error(event, (struct sockaddr *)&xprt->xpt_remote); 108 switch (event->event) { 109 /* These are considered benign events */ 110 case IB_EVENT_PATH_MIG: 111 case IB_EVENT_COMM_EST: 112 case IB_EVENT_SQ_DRAINED: 113 case IB_EVENT_QP_LAST_WQE_REACHED: 114 break; 115 116 /* These are considered fatal events */ 117 case IB_EVENT_PATH_MIG_ERR: 118 case IB_EVENT_QP_FATAL: 119 case IB_EVENT_QP_REQ_ERR: 120 case IB_EVENT_QP_ACCESS_ERR: 121 case IB_EVENT_DEVICE_FATAL: 122 default: 123 set_bit(XPT_CLOSE, &xprt->xpt_flags); 124 svc_xprt_enqueue(xprt); 125 break; 126 } 127 } 128 129 static struct svcxprt_rdma *svc_rdma_create_xprt(struct svc_serv *serv, 130 struct net *net) 131 { 132 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); 133 134 if (!cma_xprt) { 135 dprintk("svcrdma: failed to create new transport\n"); 136 return NULL; 137 } 138 svc_xprt_init(net, &svc_rdma_class, &cma_xprt->sc_xprt, serv); 139 INIT_LIST_HEAD(&cma_xprt->sc_accept_q); 140 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); 141 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); 142 INIT_LIST_HEAD(&cma_xprt->sc_send_ctxts); 143 INIT_LIST_HEAD(&cma_xprt->sc_recv_ctxts); 144 INIT_LIST_HEAD(&cma_xprt->sc_rw_ctxts); 145 init_waitqueue_head(&cma_xprt->sc_send_wait); 146 147 spin_lock_init(&cma_xprt->sc_lock); 148 spin_lock_init(&cma_xprt->sc_rq_dto_lock); 149 spin_lock_init(&cma_xprt->sc_send_lock); 150 spin_lock_init(&cma_xprt->sc_recv_lock); 151 spin_lock_init(&cma_xprt->sc_rw_ctxt_lock); 152 153 /* 154 * Note that this implies that the underlying transport support 155 * has some form of congestion control (see RFC 7530 section 3.1 156 * paragraph 2). For now, we assume that all supported RDMA 157 * transports are suitable here. 158 */ 159 set_bit(XPT_CONG_CTRL, &cma_xprt->sc_xprt.xpt_flags); 160 161 return cma_xprt; 162 } 163 164 static void 165 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt, 166 struct rdma_conn_param *param) 167 { 168 const struct rpcrdma_connect_private *pmsg = param->private_data; 169 170 if (pmsg && 171 pmsg->cp_magic == rpcrdma_cmp_magic && 172 pmsg->cp_version == RPCRDMA_CMP_VERSION) { 173 newxprt->sc_snd_w_inv = pmsg->cp_flags & 174 RPCRDMA_CMP_F_SND_W_INV_OK; 175 176 dprintk("svcrdma: client send_size %u, recv_size %u " 177 "remote inv %ssupported\n", 178 rpcrdma_decode_buffer_size(pmsg->cp_send_size), 179 rpcrdma_decode_buffer_size(pmsg->cp_recv_size), 180 newxprt->sc_snd_w_inv ? "" : "un"); 181 } 182 } 183 184 /* 185 * This function handles the CONNECT_REQUEST event on a listening 186 * endpoint. It is passed the cma_id for the _new_ connection. The context in 187 * this cma_id is inherited from the listening cma_id and is the svc_xprt 188 * structure for the listening endpoint. 189 * 190 * This function creates a new xprt for the new connection and enqueues it on 191 * the accept queue for the listent xprt. When the listen thread is kicked, it 192 * will call the recvfrom method on the listen xprt which will accept the new 193 * connection. 194 */ 195 static void handle_connect_req(struct rdma_cm_id *new_cma_id, 196 struct rdma_conn_param *param) 197 { 198 struct svcxprt_rdma *listen_xprt = new_cma_id->context; 199 struct svcxprt_rdma *newxprt; 200 struct sockaddr *sa; 201 202 /* Create a new transport */ 203 newxprt = svc_rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 204 listen_xprt->sc_xprt.xpt_net); 205 if (!newxprt) 206 return; 207 newxprt->sc_cm_id = new_cma_id; 208 new_cma_id->context = newxprt; 209 svc_rdma_parse_connect_private(newxprt, param); 210 211 /* Save client advertised inbound read limit for use later in accept. */ 212 newxprt->sc_ord = param->initiator_depth; 213 214 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 215 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 216 /* The remote port is arbitrary and not under the control of the 217 * client ULP. Set it to a fixed value so that the DRC continues 218 * to be effective after a reconnect. 219 */ 220 rpc_set_port((struct sockaddr *)&newxprt->sc_xprt.xpt_remote, 0); 221 222 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 223 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 224 225 /* 226 * Enqueue the new transport on the accept queue of the listening 227 * transport 228 */ 229 spin_lock(&listen_xprt->sc_lock); 230 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); 231 spin_unlock(&listen_xprt->sc_lock); 232 233 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 234 svc_xprt_enqueue(&listen_xprt->sc_xprt); 235 } 236 237 /* 238 * Handles events generated on the listening endpoint. These events will be 239 * either be incoming connect requests or adapter removal events. 240 */ 241 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 242 struct rdma_cm_event *event) 243 { 244 struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.src_addr; 245 246 trace_svcrdma_cm_event(event, sap); 247 248 switch (event->event) { 249 case RDMA_CM_EVENT_CONNECT_REQUEST: 250 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 251 "event = %s (%d)\n", cma_id, cma_id->context, 252 rdma_event_msg(event->event), event->event); 253 handle_connect_req(cma_id, &event->param.conn); 254 break; 255 default: 256 /* NB: No device removal upcall for INADDR_ANY listeners */ 257 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 258 "event = %s (%d)\n", cma_id, 259 rdma_event_msg(event->event), event->event); 260 break; 261 } 262 263 return 0; 264 } 265 266 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 267 struct rdma_cm_event *event) 268 { 269 struct sockaddr *sap = (struct sockaddr *)&cma_id->route.addr.dst_addr; 270 struct svcxprt_rdma *rdma = cma_id->context; 271 struct svc_xprt *xprt = &rdma->sc_xprt; 272 273 trace_svcrdma_cm_event(event, sap); 274 275 switch (event->event) { 276 case RDMA_CM_EVENT_ESTABLISHED: 277 /* Accept complete */ 278 svc_xprt_get(xprt); 279 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 280 "cm_id=%p\n", xprt, cma_id); 281 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 282 svc_xprt_enqueue(xprt); 283 break; 284 case RDMA_CM_EVENT_DISCONNECTED: 285 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 286 xprt, cma_id); 287 set_bit(XPT_CLOSE, &xprt->xpt_flags); 288 svc_xprt_enqueue(xprt); 289 svc_xprt_put(xprt); 290 break; 291 case RDMA_CM_EVENT_DEVICE_REMOVAL: 292 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 293 "event = %s (%d)\n", cma_id, xprt, 294 rdma_event_msg(event->event), event->event); 295 set_bit(XPT_CLOSE, &xprt->xpt_flags); 296 svc_xprt_enqueue(xprt); 297 svc_xprt_put(xprt); 298 break; 299 default: 300 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 301 "event = %s (%d)\n", cma_id, 302 rdma_event_msg(event->event), event->event); 303 break; 304 } 305 return 0; 306 } 307 308 /* 309 * Create a listening RDMA service endpoint. 310 */ 311 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 312 struct net *net, 313 struct sockaddr *sa, int salen, 314 int flags) 315 { 316 struct rdma_cm_id *listen_id; 317 struct svcxprt_rdma *cma_xprt; 318 int ret; 319 320 dprintk("svcrdma: Creating RDMA listener\n"); 321 if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) { 322 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); 323 return ERR_PTR(-EAFNOSUPPORT); 324 } 325 cma_xprt = svc_rdma_create_xprt(serv, net); 326 if (!cma_xprt) 327 return ERR_PTR(-ENOMEM); 328 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); 329 strcpy(cma_xprt->sc_xprt.xpt_remotebuf, "listener"); 330 331 listen_id = rdma_create_id(net, rdma_listen_handler, cma_xprt, 332 RDMA_PS_TCP, IB_QPT_RC); 333 if (IS_ERR(listen_id)) { 334 ret = PTR_ERR(listen_id); 335 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 336 goto err0; 337 } 338 339 /* Allow both IPv4 and IPv6 sockets to bind a single port 340 * at the same time. 341 */ 342 #if IS_ENABLED(CONFIG_IPV6) 343 ret = rdma_set_afonly(listen_id, 1); 344 if (ret) { 345 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret); 346 goto err1; 347 } 348 #endif 349 ret = rdma_bind_addr(listen_id, sa); 350 if (ret) { 351 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 352 goto err1; 353 } 354 cma_xprt->sc_cm_id = listen_id; 355 356 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 357 if (ret) { 358 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 359 goto err1; 360 } 361 362 /* 363 * We need to use the address from the cm_id in case the 364 * caller specified 0 for the port number. 365 */ 366 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 367 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 368 369 return &cma_xprt->sc_xprt; 370 371 err1: 372 rdma_destroy_id(listen_id); 373 err0: 374 kfree(cma_xprt); 375 return ERR_PTR(ret); 376 } 377 378 /* 379 * This is the xpo_recvfrom function for listening endpoints. Its 380 * purpose is to accept incoming connections. The CMA callback handler 381 * has already created a new transport and attached it to the new CMA 382 * ID. 383 * 384 * There is a queue of pending connections hung on the listening 385 * transport. This queue contains the new svc_xprt structure. This 386 * function takes svc_xprt structures off the accept_q and completes 387 * the connection. 388 */ 389 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 390 { 391 struct svcxprt_rdma *listen_rdma; 392 struct svcxprt_rdma *newxprt = NULL; 393 struct rdma_conn_param conn_param; 394 struct rpcrdma_connect_private pmsg; 395 struct ib_qp_init_attr qp_attr; 396 unsigned int ctxts, rq_depth; 397 struct ib_device *dev; 398 int ret = 0; 399 RPC_IFDEBUG(struct sockaddr *sap); 400 401 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 402 clear_bit(XPT_CONN, &xprt->xpt_flags); 403 /* Get the next entry off the accept list */ 404 spin_lock(&listen_rdma->sc_lock); 405 if (!list_empty(&listen_rdma->sc_accept_q)) { 406 newxprt = list_entry(listen_rdma->sc_accept_q.next, 407 struct svcxprt_rdma, sc_accept_q); 408 list_del_init(&newxprt->sc_accept_q); 409 } 410 if (!list_empty(&listen_rdma->sc_accept_q)) 411 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 412 spin_unlock(&listen_rdma->sc_lock); 413 if (!newxprt) 414 return NULL; 415 416 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 417 newxprt, newxprt->sc_cm_id); 418 419 dev = newxprt->sc_cm_id->device; 420 newxprt->sc_port_num = newxprt->sc_cm_id->port_num; 421 422 /* Qualify the transport resource defaults with the 423 * capabilities of this particular device */ 424 /* Transport header, head iovec, tail iovec */ 425 newxprt->sc_max_send_sges = 3; 426 /* Add one SGE per page list entry */ 427 newxprt->sc_max_send_sges += (svcrdma_max_req_size / PAGE_SIZE) + 1; 428 if (newxprt->sc_max_send_sges > dev->attrs.max_send_sge) 429 newxprt->sc_max_send_sges = dev->attrs.max_send_sge; 430 newxprt->sc_max_req_size = svcrdma_max_req_size; 431 newxprt->sc_max_requests = svcrdma_max_requests; 432 newxprt->sc_max_bc_requests = svcrdma_max_bc_requests; 433 rq_depth = newxprt->sc_max_requests + newxprt->sc_max_bc_requests; 434 if (rq_depth > dev->attrs.max_qp_wr) { 435 pr_warn("svcrdma: reducing receive depth to %d\n", 436 dev->attrs.max_qp_wr); 437 rq_depth = dev->attrs.max_qp_wr; 438 newxprt->sc_max_requests = rq_depth - 2; 439 newxprt->sc_max_bc_requests = 2; 440 } 441 newxprt->sc_fc_credits = cpu_to_be32(newxprt->sc_max_requests); 442 ctxts = rdma_rw_mr_factor(dev, newxprt->sc_port_num, RPCSVC_MAXPAGES); 443 ctxts *= newxprt->sc_max_requests; 444 newxprt->sc_sq_depth = rq_depth + ctxts; 445 if (newxprt->sc_sq_depth > dev->attrs.max_qp_wr) { 446 pr_warn("svcrdma: reducing send depth to %d\n", 447 dev->attrs.max_qp_wr); 448 newxprt->sc_sq_depth = dev->attrs.max_qp_wr; 449 } 450 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth); 451 452 newxprt->sc_pd = ib_alloc_pd(dev, 0); 453 if (IS_ERR(newxprt->sc_pd)) { 454 dprintk("svcrdma: error creating PD for connect request\n"); 455 goto errout; 456 } 457 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth, 458 0, IB_POLL_WORKQUEUE); 459 if (IS_ERR(newxprt->sc_sq_cq)) { 460 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 461 goto errout; 462 } 463 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, rq_depth, 464 0, IB_POLL_WORKQUEUE); 465 if (IS_ERR(newxprt->sc_rq_cq)) { 466 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 467 goto errout; 468 } 469 470 memset(&qp_attr, 0, sizeof qp_attr); 471 qp_attr.event_handler = qp_event_handler; 472 qp_attr.qp_context = &newxprt->sc_xprt; 473 qp_attr.port_num = newxprt->sc_port_num; 474 qp_attr.cap.max_rdma_ctxs = ctxts; 475 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts; 476 qp_attr.cap.max_recv_wr = rq_depth; 477 qp_attr.cap.max_send_sge = newxprt->sc_max_send_sges; 478 qp_attr.cap.max_recv_sge = 1; 479 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 480 qp_attr.qp_type = IB_QPT_RC; 481 qp_attr.send_cq = newxprt->sc_sq_cq; 482 qp_attr.recv_cq = newxprt->sc_rq_cq; 483 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n", 484 newxprt->sc_cm_id, newxprt->sc_pd); 485 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n", 486 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr); 487 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n", 488 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge); 489 490 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 491 if (ret) { 492 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 493 goto errout; 494 } 495 newxprt->sc_qp = newxprt->sc_cm_id->qp; 496 497 if (!(dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)) 498 newxprt->sc_snd_w_inv = false; 499 if (!rdma_protocol_iwarp(dev, newxprt->sc_port_num) && 500 !rdma_ib_or_roce(dev, newxprt->sc_port_num)) 501 goto errout; 502 503 if (!svc_rdma_post_recvs(newxprt)) 504 goto errout; 505 506 /* Swap out the handler */ 507 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 508 509 /* Construct RDMA-CM private message */ 510 pmsg.cp_magic = rpcrdma_cmp_magic; 511 pmsg.cp_version = RPCRDMA_CMP_VERSION; 512 pmsg.cp_flags = 0; 513 pmsg.cp_send_size = pmsg.cp_recv_size = 514 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size); 515 516 /* Accept Connection */ 517 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 518 memset(&conn_param, 0, sizeof conn_param); 519 conn_param.responder_resources = 0; 520 conn_param.initiator_depth = min_t(int, newxprt->sc_ord, 521 dev->attrs.max_qp_init_rd_atom); 522 if (!conn_param.initiator_depth) { 523 dprintk("svcrdma: invalid ORD setting\n"); 524 ret = -EINVAL; 525 goto errout; 526 } 527 conn_param.private_data = &pmsg; 528 conn_param.private_data_len = sizeof(pmsg); 529 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 530 if (ret) 531 goto errout; 532 533 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 534 dprintk("svcrdma: new connection %p accepted:\n", newxprt); 535 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 536 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap)); 537 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 538 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap)); 539 dprintk(" max_sge : %d\n", newxprt->sc_max_send_sges); 540 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth); 541 dprintk(" rdma_rw_ctxs : %d\n", ctxts); 542 dprintk(" max_requests : %d\n", newxprt->sc_max_requests); 543 dprintk(" ord : %d\n", conn_param.initiator_depth); 544 #endif 545 546 trace_svcrdma_xprt_accept(&newxprt->sc_xprt); 547 return &newxprt->sc_xprt; 548 549 errout: 550 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 551 trace_svcrdma_xprt_fail(&newxprt->sc_xprt); 552 /* Take a reference in case the DTO handler runs */ 553 svc_xprt_get(&newxprt->sc_xprt); 554 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 555 ib_destroy_qp(newxprt->sc_qp); 556 rdma_destroy_id(newxprt->sc_cm_id); 557 /* This call to put will destroy the transport */ 558 svc_xprt_put(&newxprt->sc_xprt); 559 return NULL; 560 } 561 562 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 563 { 564 } 565 566 /* 567 * When connected, an svc_xprt has at least two references: 568 * 569 * - A reference held by the cm_id between the ESTABLISHED and 570 * DISCONNECTED events. If the remote peer disconnected first, this 571 * reference could be gone. 572 * 573 * - A reference held by the svc_recv code that called this function 574 * as part of close processing. 575 * 576 * At a minimum one references should still be held. 577 */ 578 static void svc_rdma_detach(struct svc_xprt *xprt) 579 { 580 struct svcxprt_rdma *rdma = 581 container_of(xprt, struct svcxprt_rdma, sc_xprt); 582 583 /* Disconnect and flush posted WQE */ 584 rdma_disconnect(rdma->sc_cm_id); 585 } 586 587 static void __svc_rdma_free(struct work_struct *work) 588 { 589 struct svcxprt_rdma *rdma = 590 container_of(work, struct svcxprt_rdma, sc_work); 591 struct svc_xprt *xprt = &rdma->sc_xprt; 592 593 trace_svcrdma_xprt_free(xprt); 594 595 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 596 ib_drain_qp(rdma->sc_qp); 597 598 svc_rdma_flush_recv_queues(rdma); 599 600 /* Final put of backchannel client transport */ 601 if (xprt->xpt_bc_xprt) { 602 xprt_put(xprt->xpt_bc_xprt); 603 xprt->xpt_bc_xprt = NULL; 604 } 605 606 svc_rdma_destroy_rw_ctxts(rdma); 607 svc_rdma_send_ctxts_destroy(rdma); 608 svc_rdma_recv_ctxts_destroy(rdma); 609 610 /* Destroy the QP if present (not a listener) */ 611 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 612 ib_destroy_qp(rdma->sc_qp); 613 614 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 615 ib_free_cq(rdma->sc_sq_cq); 616 617 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 618 ib_free_cq(rdma->sc_rq_cq); 619 620 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 621 ib_dealloc_pd(rdma->sc_pd); 622 623 /* Destroy the CM ID */ 624 rdma_destroy_id(rdma->sc_cm_id); 625 626 kfree(rdma); 627 } 628 629 static void svc_rdma_free(struct svc_xprt *xprt) 630 { 631 struct svcxprt_rdma *rdma = 632 container_of(xprt, struct svcxprt_rdma, sc_xprt); 633 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 634 queue_work(svc_rdma_wq, &rdma->sc_work); 635 } 636 637 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 638 { 639 struct svcxprt_rdma *rdma = 640 container_of(xprt, struct svcxprt_rdma, sc_xprt); 641 642 /* 643 * If there are already waiters on the SQ, 644 * return false. 645 */ 646 if (waitqueue_active(&rdma->sc_send_wait)) 647 return 0; 648 649 /* Otherwise return true. */ 650 return 1; 651 } 652 653 static void svc_rdma_secure_port(struct svc_rqst *rqstp) 654 { 655 set_bit(RQ_SECURE, &rqstp->rq_flags); 656 } 657 658 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt) 659 { 660 } 661