1 /* 2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved. 3 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the BSD-type 9 * license below: 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 15 * Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 18 * Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials provided 21 * with the distribution. 22 * 23 * Neither the name of the Network Appliance, Inc. nor the names of 24 * its contributors may be used to endorse or promote products 25 * derived from this software without specific prior written 26 * permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Author: Tom Tucker <tom@opengridcomputing.com> 41 */ 42 43 #include <linux/sunrpc/svc_xprt.h> 44 #include <linux/sunrpc/addr.h> 45 #include <linux/sunrpc/debug.h> 46 #include <linux/sunrpc/rpc_rdma.h> 47 #include <linux/interrupt.h> 48 #include <linux/sched.h> 49 #include <linux/slab.h> 50 #include <linux/spinlock.h> 51 #include <linux/workqueue.h> 52 #include <rdma/ib_verbs.h> 53 #include <rdma/rdma_cm.h> 54 #include <linux/sunrpc/svc_rdma.h> 55 #include <linux/export.h> 56 #include "xprt_rdma.h" 57 58 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 59 60 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int); 61 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 62 struct net *net, 63 struct sockaddr *sa, int salen, 64 int flags); 65 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt); 66 static void svc_rdma_release_rqst(struct svc_rqst *); 67 static void svc_rdma_detach(struct svc_xprt *xprt); 68 static void svc_rdma_free(struct svc_xprt *xprt); 69 static int svc_rdma_has_wspace(struct svc_xprt *xprt); 70 static int svc_rdma_secure_port(struct svc_rqst *); 71 static void svc_rdma_kill_temp_xprt(struct svc_xprt *); 72 73 static struct svc_xprt_ops svc_rdma_ops = { 74 .xpo_create = svc_rdma_create, 75 .xpo_recvfrom = svc_rdma_recvfrom, 76 .xpo_sendto = svc_rdma_sendto, 77 .xpo_release_rqst = svc_rdma_release_rqst, 78 .xpo_detach = svc_rdma_detach, 79 .xpo_free = svc_rdma_free, 80 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, 81 .xpo_has_wspace = svc_rdma_has_wspace, 82 .xpo_accept = svc_rdma_accept, 83 .xpo_secure_port = svc_rdma_secure_port, 84 .xpo_kill_temp_xprt = svc_rdma_kill_temp_xprt, 85 }; 86 87 struct svc_xprt_class svc_rdma_class = { 88 .xcl_name = "rdma", 89 .xcl_owner = THIS_MODULE, 90 .xcl_ops = &svc_rdma_ops, 91 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA, 92 .xcl_ident = XPRT_TRANSPORT_RDMA, 93 }; 94 95 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 96 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *, 97 struct sockaddr *, int, int); 98 static void svc_rdma_bc_detach(struct svc_xprt *); 99 static void svc_rdma_bc_free(struct svc_xprt *); 100 101 static struct svc_xprt_ops svc_rdma_bc_ops = { 102 .xpo_create = svc_rdma_bc_create, 103 .xpo_detach = svc_rdma_bc_detach, 104 .xpo_free = svc_rdma_bc_free, 105 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, 106 .xpo_secure_port = svc_rdma_secure_port, 107 }; 108 109 struct svc_xprt_class svc_rdma_bc_class = { 110 .xcl_name = "rdma-bc", 111 .xcl_owner = THIS_MODULE, 112 .xcl_ops = &svc_rdma_bc_ops, 113 .xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN) 114 }; 115 116 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv, 117 struct net *net, 118 struct sockaddr *sa, int salen, 119 int flags) 120 { 121 struct svcxprt_rdma *cma_xprt; 122 struct svc_xprt *xprt; 123 124 cma_xprt = rdma_create_xprt(serv, 0); 125 if (!cma_xprt) 126 return ERR_PTR(-ENOMEM); 127 xprt = &cma_xprt->sc_xprt; 128 129 svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv); 130 serv->sv_bc_xprt = xprt; 131 132 dprintk("svcrdma: %s(%p)\n", __func__, xprt); 133 return xprt; 134 } 135 136 static void svc_rdma_bc_detach(struct svc_xprt *xprt) 137 { 138 dprintk("svcrdma: %s(%p)\n", __func__, xprt); 139 } 140 141 static void svc_rdma_bc_free(struct svc_xprt *xprt) 142 { 143 struct svcxprt_rdma *rdma = 144 container_of(xprt, struct svcxprt_rdma, sc_xprt); 145 146 dprintk("svcrdma: %s(%p)\n", __func__, xprt); 147 if (xprt) 148 kfree(rdma); 149 } 150 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 151 152 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt, 153 gfp_t flags) 154 { 155 struct svc_rdma_op_ctxt *ctxt; 156 157 ctxt = kmalloc(sizeof(*ctxt), flags); 158 if (ctxt) { 159 ctxt->xprt = xprt; 160 INIT_LIST_HEAD(&ctxt->free); 161 INIT_LIST_HEAD(&ctxt->dto_q); 162 } 163 return ctxt; 164 } 165 166 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt) 167 { 168 unsigned int i; 169 170 /* Each RPC/RDMA credit can consume a number of send 171 * and receive WQEs. One ctxt is allocated for each. 172 */ 173 i = xprt->sc_sq_depth + xprt->sc_rq_depth; 174 175 while (i--) { 176 struct svc_rdma_op_ctxt *ctxt; 177 178 ctxt = alloc_ctxt(xprt, GFP_KERNEL); 179 if (!ctxt) { 180 dprintk("svcrdma: No memory for RDMA ctxt\n"); 181 return false; 182 } 183 list_add(&ctxt->free, &xprt->sc_ctxts); 184 } 185 return true; 186 } 187 188 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt) 189 { 190 struct svc_rdma_op_ctxt *ctxt = NULL; 191 192 spin_lock_bh(&xprt->sc_ctxt_lock); 193 xprt->sc_ctxt_used++; 194 if (list_empty(&xprt->sc_ctxts)) 195 goto out_empty; 196 197 ctxt = list_first_entry(&xprt->sc_ctxts, 198 struct svc_rdma_op_ctxt, free); 199 list_del_init(&ctxt->free); 200 spin_unlock_bh(&xprt->sc_ctxt_lock); 201 202 out: 203 ctxt->count = 0; 204 ctxt->mapped_sges = 0; 205 ctxt->frmr = NULL; 206 return ctxt; 207 208 out_empty: 209 /* Either pre-allocation missed the mark, or send 210 * queue accounting is broken. 211 */ 212 spin_unlock_bh(&xprt->sc_ctxt_lock); 213 214 ctxt = alloc_ctxt(xprt, GFP_NOIO); 215 if (ctxt) 216 goto out; 217 218 spin_lock_bh(&xprt->sc_ctxt_lock); 219 xprt->sc_ctxt_used--; 220 spin_unlock_bh(&xprt->sc_ctxt_lock); 221 WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n"); 222 return NULL; 223 } 224 225 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt) 226 { 227 struct svcxprt_rdma *xprt = ctxt->xprt; 228 struct ib_device *device = xprt->sc_cm_id->device; 229 u32 lkey = xprt->sc_pd->local_dma_lkey; 230 unsigned int i; 231 232 for (i = 0; i < ctxt->mapped_sges; i++) { 233 /* 234 * Unmap the DMA addr in the SGE if the lkey matches 235 * the local_dma_lkey, otherwise, ignore it since it is 236 * an FRMR lkey and will be unmapped later when the 237 * last WR that uses it completes. 238 */ 239 if (ctxt->sge[i].lkey == lkey) 240 ib_dma_unmap_page(device, 241 ctxt->sge[i].addr, 242 ctxt->sge[i].length, 243 ctxt->direction); 244 } 245 ctxt->mapped_sges = 0; 246 } 247 248 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages) 249 { 250 struct svcxprt_rdma *xprt = ctxt->xprt; 251 int i; 252 253 if (free_pages) 254 for (i = 0; i < ctxt->count; i++) 255 put_page(ctxt->pages[i]); 256 257 spin_lock_bh(&xprt->sc_ctxt_lock); 258 xprt->sc_ctxt_used--; 259 list_add(&ctxt->free, &xprt->sc_ctxts); 260 spin_unlock_bh(&xprt->sc_ctxt_lock); 261 } 262 263 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt) 264 { 265 while (!list_empty(&xprt->sc_ctxts)) { 266 struct svc_rdma_op_ctxt *ctxt; 267 268 ctxt = list_first_entry(&xprt->sc_ctxts, 269 struct svc_rdma_op_ctxt, free); 270 list_del(&ctxt->free); 271 kfree(ctxt); 272 } 273 } 274 275 static struct svc_rdma_req_map *alloc_req_map(gfp_t flags) 276 { 277 struct svc_rdma_req_map *map; 278 279 map = kmalloc(sizeof(*map), flags); 280 if (map) 281 INIT_LIST_HEAD(&map->free); 282 return map; 283 } 284 285 static bool svc_rdma_prealloc_maps(struct svcxprt_rdma *xprt) 286 { 287 unsigned int i; 288 289 /* One for each receive buffer on this connection. */ 290 i = xprt->sc_max_requests; 291 292 while (i--) { 293 struct svc_rdma_req_map *map; 294 295 map = alloc_req_map(GFP_KERNEL); 296 if (!map) { 297 dprintk("svcrdma: No memory for request map\n"); 298 return false; 299 } 300 list_add(&map->free, &xprt->sc_maps); 301 } 302 return true; 303 } 304 305 struct svc_rdma_req_map *svc_rdma_get_req_map(struct svcxprt_rdma *xprt) 306 { 307 struct svc_rdma_req_map *map = NULL; 308 309 spin_lock(&xprt->sc_map_lock); 310 if (list_empty(&xprt->sc_maps)) 311 goto out_empty; 312 313 map = list_first_entry(&xprt->sc_maps, 314 struct svc_rdma_req_map, free); 315 list_del_init(&map->free); 316 spin_unlock(&xprt->sc_map_lock); 317 318 out: 319 map->count = 0; 320 return map; 321 322 out_empty: 323 spin_unlock(&xprt->sc_map_lock); 324 325 /* Pre-allocation amount was incorrect */ 326 map = alloc_req_map(GFP_NOIO); 327 if (map) 328 goto out; 329 330 WARN_ONCE(1, "svcrdma: empty request map list?\n"); 331 return NULL; 332 } 333 334 void svc_rdma_put_req_map(struct svcxprt_rdma *xprt, 335 struct svc_rdma_req_map *map) 336 { 337 spin_lock(&xprt->sc_map_lock); 338 list_add(&map->free, &xprt->sc_maps); 339 spin_unlock(&xprt->sc_map_lock); 340 } 341 342 static void svc_rdma_destroy_maps(struct svcxprt_rdma *xprt) 343 { 344 while (!list_empty(&xprt->sc_maps)) { 345 struct svc_rdma_req_map *map; 346 347 map = list_first_entry(&xprt->sc_maps, 348 struct svc_rdma_req_map, free); 349 list_del(&map->free); 350 kfree(map); 351 } 352 } 353 354 /* QP event handler */ 355 static void qp_event_handler(struct ib_event *event, void *context) 356 { 357 struct svc_xprt *xprt = context; 358 359 switch (event->event) { 360 /* These are considered benign events */ 361 case IB_EVENT_PATH_MIG: 362 case IB_EVENT_COMM_EST: 363 case IB_EVENT_SQ_DRAINED: 364 case IB_EVENT_QP_LAST_WQE_REACHED: 365 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n", 366 ib_event_msg(event->event), event->event, 367 event->element.qp); 368 break; 369 /* These are considered fatal events */ 370 case IB_EVENT_PATH_MIG_ERR: 371 case IB_EVENT_QP_FATAL: 372 case IB_EVENT_QP_REQ_ERR: 373 case IB_EVENT_QP_ACCESS_ERR: 374 case IB_EVENT_DEVICE_FATAL: 375 default: 376 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, " 377 "closing transport\n", 378 ib_event_msg(event->event), event->event, 379 event->element.qp); 380 set_bit(XPT_CLOSE, &xprt->xpt_flags); 381 break; 382 } 383 } 384 385 /** 386 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC 387 * @cq: completion queue 388 * @wc: completed WR 389 * 390 */ 391 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc) 392 { 393 struct svcxprt_rdma *xprt = cq->cq_context; 394 struct ib_cqe *cqe = wc->wr_cqe; 395 struct svc_rdma_op_ctxt *ctxt; 396 397 /* WARNING: Only wc->wr_cqe and wc->status are reliable */ 398 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe); 399 svc_rdma_unmap_dma(ctxt); 400 401 if (wc->status != IB_WC_SUCCESS) 402 goto flushed; 403 404 /* All wc fields are now known to be valid */ 405 ctxt->byte_len = wc->byte_len; 406 spin_lock(&xprt->sc_rq_dto_lock); 407 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q); 408 spin_unlock(&xprt->sc_rq_dto_lock); 409 410 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 411 if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags)) 412 goto out; 413 svc_xprt_enqueue(&xprt->sc_xprt); 414 goto out; 415 416 flushed: 417 if (wc->status != IB_WC_WR_FLUSH_ERR) 418 pr_warn("svcrdma: receive: %s (%u/0x%x)\n", 419 ib_wc_status_msg(wc->status), 420 wc->status, wc->vendor_err); 421 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 422 svc_rdma_put_context(ctxt, 1); 423 424 out: 425 svc_xprt_put(&xprt->sc_xprt); 426 } 427 428 static void svc_rdma_send_wc_common(struct svcxprt_rdma *xprt, 429 struct ib_wc *wc, 430 const char *opname) 431 { 432 if (wc->status != IB_WC_SUCCESS) 433 goto err; 434 435 out: 436 atomic_inc(&xprt->sc_sq_avail); 437 wake_up(&xprt->sc_send_wait); 438 return; 439 440 err: 441 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 442 if (wc->status != IB_WC_WR_FLUSH_ERR) 443 pr_err("svcrdma: %s: %s (%u/0x%x)\n", 444 opname, ib_wc_status_msg(wc->status), 445 wc->status, wc->vendor_err); 446 goto out; 447 } 448 449 static void svc_rdma_send_wc_common_put(struct ib_cq *cq, struct ib_wc *wc, 450 const char *opname) 451 { 452 struct svcxprt_rdma *xprt = cq->cq_context; 453 454 svc_rdma_send_wc_common(xprt, wc, opname); 455 svc_xprt_put(&xprt->sc_xprt); 456 } 457 458 /** 459 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC 460 * @cq: completion queue 461 * @wc: completed WR 462 * 463 */ 464 void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc) 465 { 466 struct ib_cqe *cqe = wc->wr_cqe; 467 struct svc_rdma_op_ctxt *ctxt; 468 469 svc_rdma_send_wc_common_put(cq, wc, "send"); 470 471 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe); 472 svc_rdma_unmap_dma(ctxt); 473 svc_rdma_put_context(ctxt, 1); 474 } 475 476 /** 477 * svc_rdma_wc_write - Invoked by RDMA provider for each polled Write WC 478 * @cq: completion queue 479 * @wc: completed WR 480 * 481 */ 482 void svc_rdma_wc_write(struct ib_cq *cq, struct ib_wc *wc) 483 { 484 struct ib_cqe *cqe = wc->wr_cqe; 485 struct svc_rdma_op_ctxt *ctxt; 486 487 svc_rdma_send_wc_common_put(cq, wc, "write"); 488 489 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe); 490 svc_rdma_unmap_dma(ctxt); 491 svc_rdma_put_context(ctxt, 0); 492 } 493 494 /** 495 * svc_rdma_wc_reg - Invoked by RDMA provider for each polled FASTREG WC 496 * @cq: completion queue 497 * @wc: completed WR 498 * 499 */ 500 void svc_rdma_wc_reg(struct ib_cq *cq, struct ib_wc *wc) 501 { 502 svc_rdma_send_wc_common_put(cq, wc, "fastreg"); 503 } 504 505 /** 506 * svc_rdma_wc_read - Invoked by RDMA provider for each polled Read WC 507 * @cq: completion queue 508 * @wc: completed WR 509 * 510 */ 511 void svc_rdma_wc_read(struct ib_cq *cq, struct ib_wc *wc) 512 { 513 struct svcxprt_rdma *xprt = cq->cq_context; 514 struct ib_cqe *cqe = wc->wr_cqe; 515 struct svc_rdma_op_ctxt *ctxt; 516 517 svc_rdma_send_wc_common(xprt, wc, "read"); 518 519 ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe); 520 svc_rdma_unmap_dma(ctxt); 521 svc_rdma_put_frmr(xprt, ctxt->frmr); 522 523 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) { 524 struct svc_rdma_op_ctxt *read_hdr; 525 526 read_hdr = ctxt->read_hdr; 527 spin_lock(&xprt->sc_rq_dto_lock); 528 list_add_tail(&read_hdr->dto_q, 529 &xprt->sc_read_complete_q); 530 spin_unlock(&xprt->sc_rq_dto_lock); 531 532 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 533 svc_xprt_enqueue(&xprt->sc_xprt); 534 } 535 536 svc_rdma_put_context(ctxt, 0); 537 svc_xprt_put(&xprt->sc_xprt); 538 } 539 540 /** 541 * svc_rdma_wc_inv - Invoked by RDMA provider for each polled LOCAL_INV WC 542 * @cq: completion queue 543 * @wc: completed WR 544 * 545 */ 546 void svc_rdma_wc_inv(struct ib_cq *cq, struct ib_wc *wc) 547 { 548 svc_rdma_send_wc_common_put(cq, wc, "localInv"); 549 } 550 551 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv, 552 int listener) 553 { 554 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); 555 556 if (!cma_xprt) 557 return NULL; 558 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv); 559 INIT_LIST_HEAD(&cma_xprt->sc_accept_q); 560 INIT_LIST_HEAD(&cma_xprt->sc_dto_q); 561 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); 562 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); 563 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q); 564 INIT_LIST_HEAD(&cma_xprt->sc_ctxts); 565 INIT_LIST_HEAD(&cma_xprt->sc_maps); 566 init_waitqueue_head(&cma_xprt->sc_send_wait); 567 568 spin_lock_init(&cma_xprt->sc_lock); 569 spin_lock_init(&cma_xprt->sc_rq_dto_lock); 570 spin_lock_init(&cma_xprt->sc_frmr_q_lock); 571 spin_lock_init(&cma_xprt->sc_ctxt_lock); 572 spin_lock_init(&cma_xprt->sc_map_lock); 573 574 if (listener) 575 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); 576 577 return cma_xprt; 578 } 579 580 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags) 581 { 582 struct ib_recv_wr recv_wr, *bad_recv_wr; 583 struct svc_rdma_op_ctxt *ctxt; 584 struct page *page; 585 dma_addr_t pa; 586 int sge_no; 587 int buflen; 588 int ret; 589 590 ctxt = svc_rdma_get_context(xprt); 591 buflen = 0; 592 ctxt->direction = DMA_FROM_DEVICE; 593 ctxt->cqe.done = svc_rdma_wc_receive; 594 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) { 595 if (sge_no >= xprt->sc_max_sge) { 596 pr_err("svcrdma: Too many sges (%d)\n", sge_no); 597 goto err_put_ctxt; 598 } 599 page = alloc_page(flags); 600 if (!page) 601 goto err_put_ctxt; 602 ctxt->pages[sge_no] = page; 603 pa = ib_dma_map_page(xprt->sc_cm_id->device, 604 page, 0, PAGE_SIZE, 605 DMA_FROM_DEVICE); 606 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa)) 607 goto err_put_ctxt; 608 svc_rdma_count_mappings(xprt, ctxt); 609 ctxt->sge[sge_no].addr = pa; 610 ctxt->sge[sge_no].length = PAGE_SIZE; 611 ctxt->sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey; 612 ctxt->count = sge_no + 1; 613 buflen += PAGE_SIZE; 614 } 615 recv_wr.next = NULL; 616 recv_wr.sg_list = &ctxt->sge[0]; 617 recv_wr.num_sge = ctxt->count; 618 recv_wr.wr_cqe = &ctxt->cqe; 619 620 svc_xprt_get(&xprt->sc_xprt); 621 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr); 622 if (ret) { 623 svc_rdma_unmap_dma(ctxt); 624 svc_rdma_put_context(ctxt, 1); 625 svc_xprt_put(&xprt->sc_xprt); 626 } 627 return ret; 628 629 err_put_ctxt: 630 svc_rdma_unmap_dma(ctxt); 631 svc_rdma_put_context(ctxt, 1); 632 return -ENOMEM; 633 } 634 635 int svc_rdma_repost_recv(struct svcxprt_rdma *xprt, gfp_t flags) 636 { 637 int ret = 0; 638 639 ret = svc_rdma_post_recv(xprt, flags); 640 if (ret) { 641 pr_err("svcrdma: could not post a receive buffer, err=%d.\n", 642 ret); 643 pr_err("svcrdma: closing transport %p.\n", xprt); 644 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 645 ret = -ENOTCONN; 646 } 647 return ret; 648 } 649 650 static void 651 svc_rdma_parse_connect_private(struct svcxprt_rdma *newxprt, 652 struct rdma_conn_param *param) 653 { 654 const struct rpcrdma_connect_private *pmsg = param->private_data; 655 656 if (pmsg && 657 pmsg->cp_magic == rpcrdma_cmp_magic && 658 pmsg->cp_version == RPCRDMA_CMP_VERSION) { 659 newxprt->sc_snd_w_inv = pmsg->cp_flags & 660 RPCRDMA_CMP_F_SND_W_INV_OK; 661 662 dprintk("svcrdma: client send_size %u, recv_size %u " 663 "remote inv %ssupported\n", 664 rpcrdma_decode_buffer_size(pmsg->cp_send_size), 665 rpcrdma_decode_buffer_size(pmsg->cp_recv_size), 666 newxprt->sc_snd_w_inv ? "" : "un"); 667 } 668 } 669 670 /* 671 * This function handles the CONNECT_REQUEST event on a listening 672 * endpoint. It is passed the cma_id for the _new_ connection. The context in 673 * this cma_id is inherited from the listening cma_id and is the svc_xprt 674 * structure for the listening endpoint. 675 * 676 * This function creates a new xprt for the new connection and enqueues it on 677 * the accept queue for the listent xprt. When the listen thread is kicked, it 678 * will call the recvfrom method on the listen xprt which will accept the new 679 * connection. 680 */ 681 static void handle_connect_req(struct rdma_cm_id *new_cma_id, 682 struct rdma_conn_param *param) 683 { 684 struct svcxprt_rdma *listen_xprt = new_cma_id->context; 685 struct svcxprt_rdma *newxprt; 686 struct sockaddr *sa; 687 688 /* Create a new transport */ 689 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); 690 if (!newxprt) { 691 dprintk("svcrdma: failed to create new transport\n"); 692 return; 693 } 694 newxprt->sc_cm_id = new_cma_id; 695 new_cma_id->context = newxprt; 696 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", 697 newxprt, newxprt->sc_cm_id, listen_xprt); 698 svc_rdma_parse_connect_private(newxprt, param); 699 700 /* Save client advertised inbound read limit for use later in accept. */ 701 newxprt->sc_ord = param->initiator_depth; 702 703 /* Set the local and remote addresses in the transport */ 704 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 705 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 706 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 707 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 708 709 /* 710 * Enqueue the new transport on the accept queue of the listening 711 * transport 712 */ 713 spin_lock_bh(&listen_xprt->sc_lock); 714 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); 715 spin_unlock_bh(&listen_xprt->sc_lock); 716 717 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 718 svc_xprt_enqueue(&listen_xprt->sc_xprt); 719 } 720 721 /* 722 * Handles events generated on the listening endpoint. These events will be 723 * either be incoming connect requests or adapter removal events. 724 */ 725 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 726 struct rdma_cm_event *event) 727 { 728 struct svcxprt_rdma *xprt = cma_id->context; 729 int ret = 0; 730 731 switch (event->event) { 732 case RDMA_CM_EVENT_CONNECT_REQUEST: 733 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 734 "event = %s (%d)\n", cma_id, cma_id->context, 735 rdma_event_msg(event->event), event->event); 736 handle_connect_req(cma_id, &event->param.conn); 737 break; 738 739 case RDMA_CM_EVENT_ESTABLISHED: 740 /* Accept complete */ 741 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " 742 "cm_id=%p\n", xprt, cma_id); 743 break; 744 745 case RDMA_CM_EVENT_DEVICE_REMOVAL: 746 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", 747 xprt, cma_id); 748 if (xprt) 749 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 750 break; 751 752 default: 753 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 754 "event = %s (%d)\n", cma_id, 755 rdma_event_msg(event->event), event->event); 756 break; 757 } 758 759 return ret; 760 } 761 762 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 763 struct rdma_cm_event *event) 764 { 765 struct svc_xprt *xprt = cma_id->context; 766 struct svcxprt_rdma *rdma = 767 container_of(xprt, struct svcxprt_rdma, sc_xprt); 768 switch (event->event) { 769 case RDMA_CM_EVENT_ESTABLISHED: 770 /* Accept complete */ 771 svc_xprt_get(xprt); 772 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 773 "cm_id=%p\n", xprt, cma_id); 774 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 775 svc_xprt_enqueue(xprt); 776 break; 777 case RDMA_CM_EVENT_DISCONNECTED: 778 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 779 xprt, cma_id); 780 if (xprt) { 781 set_bit(XPT_CLOSE, &xprt->xpt_flags); 782 svc_xprt_enqueue(xprt); 783 svc_xprt_put(xprt); 784 } 785 break; 786 case RDMA_CM_EVENT_DEVICE_REMOVAL: 787 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 788 "event = %s (%d)\n", cma_id, xprt, 789 rdma_event_msg(event->event), event->event); 790 if (xprt) { 791 set_bit(XPT_CLOSE, &xprt->xpt_flags); 792 svc_xprt_enqueue(xprt); 793 svc_xprt_put(xprt); 794 } 795 break; 796 default: 797 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 798 "event = %s (%d)\n", cma_id, 799 rdma_event_msg(event->event), event->event); 800 break; 801 } 802 return 0; 803 } 804 805 /* 806 * Create a listening RDMA service endpoint. 807 */ 808 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 809 struct net *net, 810 struct sockaddr *sa, int salen, 811 int flags) 812 { 813 struct rdma_cm_id *listen_id; 814 struct svcxprt_rdma *cma_xprt; 815 int ret; 816 817 dprintk("svcrdma: Creating RDMA socket\n"); 818 if ((sa->sa_family != AF_INET) && (sa->sa_family != AF_INET6)) { 819 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); 820 return ERR_PTR(-EAFNOSUPPORT); 821 } 822 cma_xprt = rdma_create_xprt(serv, 1); 823 if (!cma_xprt) 824 return ERR_PTR(-ENOMEM); 825 826 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt, 827 RDMA_PS_TCP, IB_QPT_RC); 828 if (IS_ERR(listen_id)) { 829 ret = PTR_ERR(listen_id); 830 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 831 goto err0; 832 } 833 834 /* Allow both IPv4 and IPv6 sockets to bind a single port 835 * at the same time. 836 */ 837 #if IS_ENABLED(CONFIG_IPV6) 838 ret = rdma_set_afonly(listen_id, 1); 839 if (ret) { 840 dprintk("svcrdma: rdma_set_afonly failed = %d\n", ret); 841 goto err1; 842 } 843 #endif 844 ret = rdma_bind_addr(listen_id, sa); 845 if (ret) { 846 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 847 goto err1; 848 } 849 cma_xprt->sc_cm_id = listen_id; 850 851 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 852 if (ret) { 853 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 854 goto err1; 855 } 856 857 /* 858 * We need to use the address from the cm_id in case the 859 * caller specified 0 for the port number. 860 */ 861 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 862 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 863 864 return &cma_xprt->sc_xprt; 865 866 err1: 867 rdma_destroy_id(listen_id); 868 err0: 869 kfree(cma_xprt); 870 return ERR_PTR(ret); 871 } 872 873 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) 874 { 875 struct ib_mr *mr; 876 struct scatterlist *sg; 877 struct svc_rdma_fastreg_mr *frmr; 878 u32 num_sg; 879 880 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); 881 if (!frmr) 882 goto err; 883 884 num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len); 885 mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg); 886 if (IS_ERR(mr)) 887 goto err_free_frmr; 888 889 sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL); 890 if (!sg) 891 goto err_free_mr; 892 893 sg_init_table(sg, RPCSVC_MAXPAGES); 894 895 frmr->mr = mr; 896 frmr->sg = sg; 897 INIT_LIST_HEAD(&frmr->frmr_list); 898 return frmr; 899 900 err_free_mr: 901 ib_dereg_mr(mr); 902 err_free_frmr: 903 kfree(frmr); 904 err: 905 return ERR_PTR(-ENOMEM); 906 } 907 908 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) 909 { 910 struct svc_rdma_fastreg_mr *frmr; 911 912 while (!list_empty(&xprt->sc_frmr_q)) { 913 frmr = list_entry(xprt->sc_frmr_q.next, 914 struct svc_rdma_fastreg_mr, frmr_list); 915 list_del_init(&frmr->frmr_list); 916 kfree(frmr->sg); 917 ib_dereg_mr(frmr->mr); 918 kfree(frmr); 919 } 920 } 921 922 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) 923 { 924 struct svc_rdma_fastreg_mr *frmr = NULL; 925 926 spin_lock_bh(&rdma->sc_frmr_q_lock); 927 if (!list_empty(&rdma->sc_frmr_q)) { 928 frmr = list_entry(rdma->sc_frmr_q.next, 929 struct svc_rdma_fastreg_mr, frmr_list); 930 list_del_init(&frmr->frmr_list); 931 frmr->sg_nents = 0; 932 } 933 spin_unlock_bh(&rdma->sc_frmr_q_lock); 934 if (frmr) 935 return frmr; 936 937 return rdma_alloc_frmr(rdma); 938 } 939 940 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, 941 struct svc_rdma_fastreg_mr *frmr) 942 { 943 if (frmr) { 944 ib_dma_unmap_sg(rdma->sc_cm_id->device, 945 frmr->sg, frmr->sg_nents, frmr->direction); 946 spin_lock_bh(&rdma->sc_frmr_q_lock); 947 WARN_ON_ONCE(!list_empty(&frmr->frmr_list)); 948 list_add(&frmr->frmr_list, &rdma->sc_frmr_q); 949 spin_unlock_bh(&rdma->sc_frmr_q_lock); 950 } 951 } 952 953 /* 954 * This is the xpo_recvfrom function for listening endpoints. Its 955 * purpose is to accept incoming connections. The CMA callback handler 956 * has already created a new transport and attached it to the new CMA 957 * ID. 958 * 959 * There is a queue of pending connections hung on the listening 960 * transport. This queue contains the new svc_xprt structure. This 961 * function takes svc_xprt structures off the accept_q and completes 962 * the connection. 963 */ 964 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 965 { 966 struct svcxprt_rdma *listen_rdma; 967 struct svcxprt_rdma *newxprt = NULL; 968 struct rdma_conn_param conn_param; 969 struct rpcrdma_connect_private pmsg; 970 struct ib_qp_init_attr qp_attr; 971 struct ib_device *dev; 972 struct sockaddr *sap; 973 unsigned int i; 974 int ret = 0; 975 976 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 977 clear_bit(XPT_CONN, &xprt->xpt_flags); 978 /* Get the next entry off the accept list */ 979 spin_lock_bh(&listen_rdma->sc_lock); 980 if (!list_empty(&listen_rdma->sc_accept_q)) { 981 newxprt = list_entry(listen_rdma->sc_accept_q.next, 982 struct svcxprt_rdma, sc_accept_q); 983 list_del_init(&newxprt->sc_accept_q); 984 } 985 if (!list_empty(&listen_rdma->sc_accept_q)) 986 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 987 spin_unlock_bh(&listen_rdma->sc_lock); 988 if (!newxprt) 989 return NULL; 990 991 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 992 newxprt, newxprt->sc_cm_id); 993 994 dev = newxprt->sc_cm_id->device; 995 996 /* Qualify the transport resource defaults with the 997 * capabilities of this particular device */ 998 newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge, 999 (size_t)RPCSVC_MAXPAGES); 1000 newxprt->sc_max_sge_rd = min_t(size_t, dev->attrs.max_sge_rd, 1001 RPCSVC_MAXPAGES); 1002 newxprt->sc_max_req_size = svcrdma_max_req_size; 1003 newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr, 1004 svcrdma_max_requests); 1005 newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr, 1006 svcrdma_max_bc_requests); 1007 newxprt->sc_rq_depth = newxprt->sc_max_requests + 1008 newxprt->sc_max_bc_requests; 1009 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_rq_depth; 1010 atomic_set(&newxprt->sc_sq_avail, newxprt->sc_sq_depth); 1011 1012 if (!svc_rdma_prealloc_ctxts(newxprt)) 1013 goto errout; 1014 if (!svc_rdma_prealloc_maps(newxprt)) 1015 goto errout; 1016 1017 /* 1018 * Limit ORD based on client limit, local device limit, and 1019 * configured svcrdma limit. 1020 */ 1021 newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord); 1022 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); 1023 1024 newxprt->sc_pd = ib_alloc_pd(dev, 0); 1025 if (IS_ERR(newxprt->sc_pd)) { 1026 dprintk("svcrdma: error creating PD for connect request\n"); 1027 goto errout; 1028 } 1029 newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth, 1030 0, IB_POLL_SOFTIRQ); 1031 if (IS_ERR(newxprt->sc_sq_cq)) { 1032 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 1033 goto errout; 1034 } 1035 newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth, 1036 0, IB_POLL_SOFTIRQ); 1037 if (IS_ERR(newxprt->sc_rq_cq)) { 1038 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 1039 goto errout; 1040 } 1041 1042 memset(&qp_attr, 0, sizeof qp_attr); 1043 qp_attr.event_handler = qp_event_handler; 1044 qp_attr.qp_context = &newxprt->sc_xprt; 1045 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; 1046 qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth; 1047 qp_attr.cap.max_send_sge = newxprt->sc_max_sge; 1048 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; 1049 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 1050 qp_attr.qp_type = IB_QPT_RC; 1051 qp_attr.send_cq = newxprt->sc_sq_cq; 1052 qp_attr.recv_cq = newxprt->sc_rq_cq; 1053 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n", 1054 newxprt->sc_cm_id, newxprt->sc_pd); 1055 dprintk(" cap.max_send_wr = %d, cap.max_recv_wr = %d\n", 1056 qp_attr.cap.max_send_wr, qp_attr.cap.max_recv_wr); 1057 dprintk(" cap.max_send_sge = %d, cap.max_recv_sge = %d\n", 1058 qp_attr.cap.max_send_sge, qp_attr.cap.max_recv_sge); 1059 1060 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 1061 if (ret) { 1062 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 1063 goto errout; 1064 } 1065 newxprt->sc_qp = newxprt->sc_cm_id->qp; 1066 1067 /* 1068 * Use the most secure set of MR resources based on the 1069 * transport type and available memory management features in 1070 * the device. Here's the table implemented below: 1071 * 1072 * Fast Global DMA Remote WR 1073 * Reg LKEY MR Access 1074 * Sup'd Sup'd Needed Needed 1075 * 1076 * IWARP N N Y Y 1077 * N Y Y Y 1078 * Y N Y N 1079 * Y Y N - 1080 * 1081 * IB N N Y N 1082 * N Y N - 1083 * Y N Y N 1084 * Y Y N - 1085 * 1086 * NB: iWARP requires remote write access for the data sink 1087 * of an RDMA_READ. IB does not. 1088 */ 1089 newxprt->sc_reader = rdma_read_chunk_lcl; 1090 if (dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 1091 newxprt->sc_frmr_pg_list_len = 1092 dev->attrs.max_fast_reg_page_list_len; 1093 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; 1094 newxprt->sc_reader = rdma_read_chunk_frmr; 1095 } else 1096 newxprt->sc_snd_w_inv = false; 1097 1098 /* 1099 * Determine if a DMA MR is required and if so, what privs are required 1100 */ 1101 if (!rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) && 1102 !rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num)) 1103 goto errout; 1104 1105 if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num)) 1106 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; 1107 1108 /* Post receive buffers */ 1109 for (i = 0; i < newxprt->sc_max_requests; i++) { 1110 ret = svc_rdma_post_recv(newxprt, GFP_KERNEL); 1111 if (ret) { 1112 dprintk("svcrdma: failure posting receive buffers\n"); 1113 goto errout; 1114 } 1115 } 1116 1117 /* Swap out the handler */ 1118 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 1119 1120 /* Construct RDMA-CM private message */ 1121 pmsg.cp_magic = rpcrdma_cmp_magic; 1122 pmsg.cp_version = RPCRDMA_CMP_VERSION; 1123 pmsg.cp_flags = 0; 1124 pmsg.cp_send_size = pmsg.cp_recv_size = 1125 rpcrdma_encode_buffer_size(newxprt->sc_max_req_size); 1126 1127 /* Accept Connection */ 1128 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 1129 memset(&conn_param, 0, sizeof conn_param); 1130 conn_param.responder_resources = 0; 1131 conn_param.initiator_depth = newxprt->sc_ord; 1132 conn_param.private_data = &pmsg; 1133 conn_param.private_data_len = sizeof(pmsg); 1134 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 1135 if (ret) { 1136 dprintk("svcrdma: failed to accept new connection, ret=%d\n", 1137 ret); 1138 goto errout; 1139 } 1140 1141 dprintk("svcrdma: new connection %p accepted:\n", newxprt); 1142 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 1143 dprintk(" local address : %pIS:%u\n", sap, rpc_get_port(sap)); 1144 sap = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 1145 dprintk(" remote address : %pIS:%u\n", sap, rpc_get_port(sap)); 1146 dprintk(" max_sge : %d\n", newxprt->sc_max_sge); 1147 dprintk(" max_sge_rd : %d\n", newxprt->sc_max_sge_rd); 1148 dprintk(" sq_depth : %d\n", newxprt->sc_sq_depth); 1149 dprintk(" max_requests : %d\n", newxprt->sc_max_requests); 1150 dprintk(" ord : %d\n", newxprt->sc_ord); 1151 1152 return &newxprt->sc_xprt; 1153 1154 errout: 1155 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 1156 /* Take a reference in case the DTO handler runs */ 1157 svc_xprt_get(&newxprt->sc_xprt); 1158 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 1159 ib_destroy_qp(newxprt->sc_qp); 1160 rdma_destroy_id(newxprt->sc_cm_id); 1161 /* This call to put will destroy the transport */ 1162 svc_xprt_put(&newxprt->sc_xprt); 1163 return NULL; 1164 } 1165 1166 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 1167 { 1168 } 1169 1170 /* 1171 * When connected, an svc_xprt has at least two references: 1172 * 1173 * - A reference held by the cm_id between the ESTABLISHED and 1174 * DISCONNECTED events. If the remote peer disconnected first, this 1175 * reference could be gone. 1176 * 1177 * - A reference held by the svc_recv code that called this function 1178 * as part of close processing. 1179 * 1180 * At a minimum one references should still be held. 1181 */ 1182 static void svc_rdma_detach(struct svc_xprt *xprt) 1183 { 1184 struct svcxprt_rdma *rdma = 1185 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1186 dprintk("svc: svc_rdma_detach(%p)\n", xprt); 1187 1188 /* Disconnect and flush posted WQE */ 1189 rdma_disconnect(rdma->sc_cm_id); 1190 } 1191 1192 static void __svc_rdma_free(struct work_struct *work) 1193 { 1194 struct svcxprt_rdma *rdma = 1195 container_of(work, struct svcxprt_rdma, sc_work); 1196 struct svc_xprt *xprt = &rdma->sc_xprt; 1197 1198 dprintk("svcrdma: %s(%p)\n", __func__, rdma); 1199 1200 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1201 ib_drain_qp(rdma->sc_qp); 1202 1203 /* We should only be called from kref_put */ 1204 if (kref_read(&xprt->xpt_ref) != 0) 1205 pr_err("svcrdma: sc_xprt still in use? (%d)\n", 1206 kref_read(&xprt->xpt_ref)); 1207 1208 /* 1209 * Destroy queued, but not processed read completions. Note 1210 * that this cleanup has to be done before destroying the 1211 * cm_id because the device ptr is needed to unmap the dma in 1212 * svc_rdma_put_context. 1213 */ 1214 while (!list_empty(&rdma->sc_read_complete_q)) { 1215 struct svc_rdma_op_ctxt *ctxt; 1216 ctxt = list_entry(rdma->sc_read_complete_q.next, 1217 struct svc_rdma_op_ctxt, 1218 dto_q); 1219 list_del_init(&ctxt->dto_q); 1220 svc_rdma_put_context(ctxt, 1); 1221 } 1222 1223 /* Destroy queued, but not processed recv completions */ 1224 while (!list_empty(&rdma->sc_rq_dto_q)) { 1225 struct svc_rdma_op_ctxt *ctxt; 1226 ctxt = list_entry(rdma->sc_rq_dto_q.next, 1227 struct svc_rdma_op_ctxt, 1228 dto_q); 1229 list_del_init(&ctxt->dto_q); 1230 svc_rdma_put_context(ctxt, 1); 1231 } 1232 1233 /* Warn if we leaked a resource or under-referenced */ 1234 if (rdma->sc_ctxt_used != 0) 1235 pr_err("svcrdma: ctxt still in use? (%d)\n", 1236 rdma->sc_ctxt_used); 1237 1238 /* Final put of backchannel client transport */ 1239 if (xprt->xpt_bc_xprt) { 1240 xprt_put(xprt->xpt_bc_xprt); 1241 xprt->xpt_bc_xprt = NULL; 1242 } 1243 1244 rdma_dealloc_frmr_q(rdma); 1245 svc_rdma_destroy_ctxts(rdma); 1246 svc_rdma_destroy_maps(rdma); 1247 1248 /* Destroy the QP if present (not a listener) */ 1249 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1250 ib_destroy_qp(rdma->sc_qp); 1251 1252 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 1253 ib_free_cq(rdma->sc_sq_cq); 1254 1255 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 1256 ib_free_cq(rdma->sc_rq_cq); 1257 1258 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 1259 ib_dealloc_pd(rdma->sc_pd); 1260 1261 /* Destroy the CM ID */ 1262 rdma_destroy_id(rdma->sc_cm_id); 1263 1264 kfree(rdma); 1265 } 1266 1267 static void svc_rdma_free(struct svc_xprt *xprt) 1268 { 1269 struct svcxprt_rdma *rdma = 1270 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1271 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 1272 queue_work(svc_rdma_wq, &rdma->sc_work); 1273 } 1274 1275 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 1276 { 1277 struct svcxprt_rdma *rdma = 1278 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1279 1280 /* 1281 * If there are already waiters on the SQ, 1282 * return false. 1283 */ 1284 if (waitqueue_active(&rdma->sc_send_wait)) 1285 return 0; 1286 1287 /* Otherwise return true. */ 1288 return 1; 1289 } 1290 1291 static int svc_rdma_secure_port(struct svc_rqst *rqstp) 1292 { 1293 return 1; 1294 } 1295 1296 static void svc_rdma_kill_temp_xprt(struct svc_xprt *xprt) 1297 { 1298 } 1299 1300 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) 1301 { 1302 struct ib_send_wr *bad_wr, *n_wr; 1303 int wr_count; 1304 int i; 1305 int ret; 1306 1307 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1308 return -ENOTCONN; 1309 1310 wr_count = 1; 1311 for (n_wr = wr->next; n_wr; n_wr = n_wr->next) 1312 wr_count++; 1313 1314 /* If the SQ is full, wait until an SQ entry is available */ 1315 while (1) { 1316 if ((atomic_sub_return(wr_count, &xprt->sc_sq_avail) < 0)) { 1317 atomic_inc(&rdma_stat_sq_starve); 1318 1319 /* Wait until SQ WR available if SQ still full */ 1320 atomic_add(wr_count, &xprt->sc_sq_avail); 1321 wait_event(xprt->sc_send_wait, 1322 atomic_read(&xprt->sc_sq_avail) > wr_count); 1323 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1324 return -ENOTCONN; 1325 continue; 1326 } 1327 /* Take a transport ref for each WR posted */ 1328 for (i = 0; i < wr_count; i++) 1329 svc_xprt_get(&xprt->sc_xprt); 1330 1331 /* Bump used SQ WR count and post */ 1332 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); 1333 if (ret) { 1334 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 1335 for (i = 0; i < wr_count; i ++) 1336 svc_xprt_put(&xprt->sc_xprt); 1337 dprintk("svcrdma: failed to post SQ WR rc=%d\n", ret); 1338 dprintk(" sc_sq_avail=%d, sc_sq_depth=%d\n", 1339 atomic_read(&xprt->sc_sq_avail), 1340 xprt->sc_sq_depth); 1341 wake_up(&xprt->sc_send_wait); 1342 } 1343 break; 1344 } 1345 return ret; 1346 } 1347