1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* 3 * Copyright (c) 2014-2017 Oracle. All rights reserved. 4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the BSD-type 10 * license below: 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 16 * Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 19 * Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials provided 22 * with the distribution. 23 * 24 * Neither the name of the Network Appliance, Inc. nor the names of 25 * its contributors may be used to endorse or promote products 26 * derived from this software without specific prior written 27 * permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 */ 41 42 /* 43 * verbs.c 44 * 45 * Encapsulates the major functions managing: 46 * o adapters 47 * o endpoints 48 * o connections 49 * o buffer memory 50 */ 51 52 #include <linux/interrupt.h> 53 #include <linux/slab.h> 54 #include <linux/sunrpc/addr.h> 55 #include <linux/sunrpc/svc_rdma.h> 56 #include <linux/log2.h> 57 58 #include <asm-generic/barrier.h> 59 #include <asm/bitops.h> 60 61 #include <rdma/ib_cm.h> 62 63 #include "xprt_rdma.h" 64 #include <trace/events/rpcrdma.h> 65 66 /* 67 * Globals/Macros 68 */ 69 70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 71 # define RPCDBG_FACILITY RPCDBG_TRANS 72 #endif 73 74 /* 75 * internal functions 76 */ 77 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt, 78 struct rpcrdma_sendctx *sc); 79 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt); 80 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt); 81 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt); 82 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt); 83 static struct rpcrdma_regbuf * 84 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction, 85 gfp_t flags); 86 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb); 87 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb); 88 89 /* Wait for outstanding transport work to finish. ib_drain_qp 90 * handles the drains in the wrong order for us, so open code 91 * them here. 92 */ 93 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt) 94 { 95 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 96 97 /* Flush Receives, then wait for deferred Reply work 98 * to complete. 99 */ 100 ib_drain_rq(ia->ri_id->qp); 101 102 /* Deferred Reply processing might have scheduled 103 * local invalidations. 104 */ 105 ib_drain_sq(ia->ri_id->qp); 106 } 107 108 /** 109 * rpcrdma_qp_event_handler - Handle one QP event (error notification) 110 * @event: details of the event 111 * @context: ep that owns QP where event occurred 112 * 113 * Called from the RDMA provider (device driver) possibly in an interrupt 114 * context. 115 */ 116 static void 117 rpcrdma_qp_event_handler(struct ib_event *event, void *context) 118 { 119 struct rpcrdma_ep *ep = context; 120 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt, 121 rx_ep); 122 123 trace_xprtrdma_qp_event(r_xprt, event); 124 } 125 126 /** 127 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC 128 * @cq: completion queue 129 * @wc: completed WR 130 * 131 */ 132 static void 133 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc) 134 { 135 struct ib_cqe *cqe = wc->wr_cqe; 136 struct rpcrdma_sendctx *sc = 137 container_of(cqe, struct rpcrdma_sendctx, sc_cqe); 138 139 /* WARNING: Only wr_cqe and status are reliable at this point */ 140 trace_xprtrdma_wc_send(sc, wc); 141 rpcrdma_sendctx_put_locked((struct rpcrdma_xprt *)cq->cq_context, sc); 142 } 143 144 /** 145 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC 146 * @cq: completion queue (ignored) 147 * @wc: completed WR 148 * 149 */ 150 static void 151 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc) 152 { 153 struct ib_cqe *cqe = wc->wr_cqe; 154 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep, 155 rr_cqe); 156 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt; 157 158 /* WARNING: Only wr_cqe and status are reliable at this point */ 159 trace_xprtrdma_wc_receive(wc); 160 --r_xprt->rx_ep.rep_receive_count; 161 if (wc->status != IB_WC_SUCCESS) 162 goto out_flushed; 163 164 /* status == SUCCESS means all fields in wc are trustworthy */ 165 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len); 166 rep->rr_wc_flags = wc->wc_flags; 167 rep->rr_inv_rkey = wc->ex.invalidate_rkey; 168 169 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf), 170 rdmab_addr(rep->rr_rdmabuf), 171 wc->byte_len, DMA_FROM_DEVICE); 172 173 rpcrdma_reply_handler(rep); 174 return; 175 176 out_flushed: 177 rpcrdma_recv_buffer_put(rep); 178 } 179 180 static void rpcrdma_update_cm_private(struct rpcrdma_xprt *r_xprt, 181 struct rdma_conn_param *param) 182 { 183 const struct rpcrdma_connect_private *pmsg = param->private_data; 184 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 185 unsigned int rsize, wsize; 186 187 /* Default settings for RPC-over-RDMA Version One */ 188 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize; 189 rsize = RPCRDMA_V1_DEF_INLINE_SIZE; 190 wsize = RPCRDMA_V1_DEF_INLINE_SIZE; 191 192 if (pmsg && 193 pmsg->cp_magic == rpcrdma_cmp_magic && 194 pmsg->cp_version == RPCRDMA_CMP_VERSION) { 195 r_xprt->rx_ia.ri_implicit_roundup = true; 196 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size); 197 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size); 198 } 199 200 if (rsize < ep->rep_inline_recv) 201 ep->rep_inline_recv = rsize; 202 if (wsize < ep->rep_inline_send) 203 ep->rep_inline_send = wsize; 204 205 rpcrdma_set_max_header_sizes(r_xprt); 206 } 207 208 /** 209 * rpcrdma_cm_event_handler - Handle RDMA CM events 210 * @id: rdma_cm_id on which an event has occurred 211 * @event: details of the event 212 * 213 * Called with @id's mutex held. Returns 1 if caller should 214 * destroy @id, otherwise 0. 215 */ 216 static int 217 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event) 218 { 219 struct rpcrdma_xprt *r_xprt = id->context; 220 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 221 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 222 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 223 224 might_sleep(); 225 226 trace_xprtrdma_cm_event(r_xprt, event); 227 switch (event->event) { 228 case RDMA_CM_EVENT_ADDR_RESOLVED: 229 case RDMA_CM_EVENT_ROUTE_RESOLVED: 230 ia->ri_async_rc = 0; 231 complete(&ia->ri_done); 232 return 0; 233 case RDMA_CM_EVENT_ADDR_ERROR: 234 ia->ri_async_rc = -EPROTO; 235 complete(&ia->ri_done); 236 return 0; 237 case RDMA_CM_EVENT_ROUTE_ERROR: 238 ia->ri_async_rc = -ENETUNREACH; 239 complete(&ia->ri_done); 240 return 0; 241 case RDMA_CM_EVENT_DEVICE_REMOVAL: 242 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 243 pr_info("rpcrdma: removing device %s for %s:%s\n", 244 ia->ri_id->device->name, 245 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt)); 246 #endif 247 init_completion(&ia->ri_remove_done); 248 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags); 249 ep->rep_connected = -ENODEV; 250 xprt_force_disconnect(xprt); 251 wait_for_completion(&ia->ri_remove_done); 252 253 ia->ri_id = NULL; 254 /* Return 1 to ensure the core destroys the id. */ 255 return 1; 256 case RDMA_CM_EVENT_ESTABLISHED: 257 ++xprt->connect_cookie; 258 ep->rep_connected = 1; 259 rpcrdma_update_cm_private(r_xprt, &event->param.conn); 260 trace_xprtrdma_inline_thresh(r_xprt); 261 wake_up_all(&ep->rep_connect_wait); 262 break; 263 case RDMA_CM_EVENT_CONNECT_ERROR: 264 ep->rep_connected = -ENOTCONN; 265 goto disconnected; 266 case RDMA_CM_EVENT_UNREACHABLE: 267 ep->rep_connected = -ENETUNREACH; 268 goto disconnected; 269 case RDMA_CM_EVENT_REJECTED: 270 dprintk("rpcrdma: connection to %s:%s rejected: %s\n", 271 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt), 272 rdma_reject_msg(id, event->status)); 273 ep->rep_connected = -ECONNREFUSED; 274 if (event->status == IB_CM_REJ_STALE_CONN) 275 ep->rep_connected = -EAGAIN; 276 goto disconnected; 277 case RDMA_CM_EVENT_DISCONNECTED: 278 ep->rep_connected = -ECONNABORTED; 279 disconnected: 280 xprt_force_disconnect(xprt); 281 wake_up_all(&ep->rep_connect_wait); 282 break; 283 default: 284 break; 285 } 286 287 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__, 288 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt), 289 ia->ri_id->device->name, rdma_event_msg(event->event)); 290 return 0; 291 } 292 293 static struct rdma_cm_id * 294 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia) 295 { 296 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1; 297 struct rdma_cm_id *id; 298 int rc; 299 300 init_completion(&ia->ri_done); 301 302 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler, 303 xprt, RDMA_PS_TCP, IB_QPT_RC); 304 if (IS_ERR(id)) 305 return id; 306 307 ia->ri_async_rc = -ETIMEDOUT; 308 rc = rdma_resolve_addr(id, NULL, 309 (struct sockaddr *)&xprt->rx_xprt.addr, 310 RDMA_RESOLVE_TIMEOUT); 311 if (rc) 312 goto out; 313 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout); 314 if (rc < 0) 315 goto out; 316 317 rc = ia->ri_async_rc; 318 if (rc) 319 goto out; 320 321 ia->ri_async_rc = -ETIMEDOUT; 322 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT); 323 if (rc) 324 goto out; 325 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout); 326 if (rc < 0) 327 goto out; 328 rc = ia->ri_async_rc; 329 if (rc) 330 goto out; 331 332 return id; 333 334 out: 335 rdma_destroy_id(id); 336 return ERR_PTR(rc); 337 } 338 339 /* 340 * Exported functions. 341 */ 342 343 /** 344 * rpcrdma_ia_open - Open and initialize an Interface Adapter. 345 * @xprt: transport with IA to (re)initialize 346 * 347 * Returns 0 on success, negative errno if an appropriate 348 * Interface Adapter could not be found and opened. 349 */ 350 int 351 rpcrdma_ia_open(struct rpcrdma_xprt *xprt) 352 { 353 struct rpcrdma_ia *ia = &xprt->rx_ia; 354 int rc; 355 356 ia->ri_id = rpcrdma_create_id(xprt, ia); 357 if (IS_ERR(ia->ri_id)) { 358 rc = PTR_ERR(ia->ri_id); 359 goto out_err; 360 } 361 362 ia->ri_pd = ib_alloc_pd(ia->ri_id->device, 0); 363 if (IS_ERR(ia->ri_pd)) { 364 rc = PTR_ERR(ia->ri_pd); 365 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc); 366 goto out_err; 367 } 368 369 switch (xprt_rdma_memreg_strategy) { 370 case RPCRDMA_FRWR: 371 if (frwr_is_supported(ia->ri_id->device)) 372 break; 373 /*FALLTHROUGH*/ 374 default: 375 pr_err("rpcrdma: Device %s does not support memreg mode %d\n", 376 ia->ri_id->device->name, xprt_rdma_memreg_strategy); 377 rc = -EINVAL; 378 goto out_err; 379 } 380 381 return 0; 382 383 out_err: 384 rpcrdma_ia_close(ia); 385 return rc; 386 } 387 388 /** 389 * rpcrdma_ia_remove - Handle device driver unload 390 * @ia: interface adapter being removed 391 * 392 * Divest transport H/W resources associated with this adapter, 393 * but allow it to be restored later. 394 */ 395 void 396 rpcrdma_ia_remove(struct rpcrdma_ia *ia) 397 { 398 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt, 399 rx_ia); 400 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 401 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 402 struct rpcrdma_req *req; 403 404 /* This is similar to rpcrdma_ep_destroy, but: 405 * - Don't cancel the connect worker. 406 * - Don't call rpcrdma_ep_disconnect, which waits 407 * for another conn upcall, which will deadlock. 408 * - rdma_disconnect is unneeded, the underlying 409 * connection is already gone. 410 */ 411 if (ia->ri_id->qp) { 412 rpcrdma_xprt_drain(r_xprt); 413 rdma_destroy_qp(ia->ri_id); 414 ia->ri_id->qp = NULL; 415 } 416 ib_free_cq(ep->rep_attr.recv_cq); 417 ep->rep_attr.recv_cq = NULL; 418 ib_free_cq(ep->rep_attr.send_cq); 419 ep->rep_attr.send_cq = NULL; 420 421 /* The ULP is responsible for ensuring all DMA 422 * mappings and MRs are gone. 423 */ 424 rpcrdma_reps_unmap(r_xprt); 425 list_for_each_entry(req, &buf->rb_allreqs, rl_all) { 426 rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf); 427 rpcrdma_regbuf_dma_unmap(req->rl_sendbuf); 428 rpcrdma_regbuf_dma_unmap(req->rl_recvbuf); 429 } 430 rpcrdma_mrs_destroy(r_xprt); 431 ib_dealloc_pd(ia->ri_pd); 432 ia->ri_pd = NULL; 433 434 /* Allow waiters to continue */ 435 complete(&ia->ri_remove_done); 436 437 trace_xprtrdma_remove(r_xprt); 438 } 439 440 /** 441 * rpcrdma_ia_close - Clean up/close an IA. 442 * @ia: interface adapter to close 443 * 444 */ 445 void 446 rpcrdma_ia_close(struct rpcrdma_ia *ia) 447 { 448 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) { 449 if (ia->ri_id->qp) 450 rdma_destroy_qp(ia->ri_id); 451 rdma_destroy_id(ia->ri_id); 452 } 453 ia->ri_id = NULL; 454 455 /* If the pd is still busy, xprtrdma missed freeing a resource */ 456 if (ia->ri_pd && !IS_ERR(ia->ri_pd)) 457 ib_dealloc_pd(ia->ri_pd); 458 ia->ri_pd = NULL; 459 } 460 461 /** 462 * rpcrdma_ep_create - Create unconnected endpoint 463 * @r_xprt: transport to instantiate 464 * 465 * Returns zero on success, or a negative errno. 466 */ 467 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt) 468 { 469 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 470 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 471 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private; 472 struct ib_cq *sendcq, *recvcq; 473 unsigned int max_sge; 474 int rc; 475 476 ep->rep_max_requests = xprt_rdma_slot_table_entries; 477 ep->rep_inline_send = xprt_rdma_max_inline_write; 478 ep->rep_inline_recv = xprt_rdma_max_inline_read; 479 480 max_sge = min_t(unsigned int, ia->ri_id->device->attrs.max_send_sge, 481 RPCRDMA_MAX_SEND_SGES); 482 if (max_sge < RPCRDMA_MIN_SEND_SGES) { 483 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge); 484 return -ENOMEM; 485 } 486 ia->ri_max_send_sges = max_sge; 487 488 rc = frwr_open(ia, ep); 489 if (rc) 490 return rc; 491 492 ep->rep_attr.event_handler = rpcrdma_qp_event_handler; 493 ep->rep_attr.qp_context = ep; 494 ep->rep_attr.srq = NULL; 495 ep->rep_attr.cap.max_send_sge = max_sge; 496 ep->rep_attr.cap.max_recv_sge = 1; 497 ep->rep_attr.cap.max_inline_data = 0; 498 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 499 ep->rep_attr.qp_type = IB_QPT_RC; 500 ep->rep_attr.port_num = ~0; 501 502 dprintk("RPC: %s: requested max: dtos: send %d recv %d; " 503 "iovs: send %d recv %d\n", 504 __func__, 505 ep->rep_attr.cap.max_send_wr, 506 ep->rep_attr.cap.max_recv_wr, 507 ep->rep_attr.cap.max_send_sge, 508 ep->rep_attr.cap.max_recv_sge); 509 510 ep->rep_send_batch = ep->rep_max_requests >> 3; 511 ep->rep_send_count = ep->rep_send_batch; 512 init_waitqueue_head(&ep->rep_connect_wait); 513 ep->rep_receive_count = 0; 514 515 sendcq = ib_alloc_cq_any(ia->ri_id->device, r_xprt, 516 ep->rep_attr.cap.max_send_wr + 1, 517 IB_POLL_WORKQUEUE); 518 if (IS_ERR(sendcq)) { 519 rc = PTR_ERR(sendcq); 520 goto out1; 521 } 522 523 recvcq = ib_alloc_cq_any(ia->ri_id->device, NULL, 524 ep->rep_attr.cap.max_recv_wr + 1, 525 IB_POLL_WORKQUEUE); 526 if (IS_ERR(recvcq)) { 527 rc = PTR_ERR(recvcq); 528 goto out2; 529 } 530 531 ep->rep_attr.send_cq = sendcq; 532 ep->rep_attr.recv_cq = recvcq; 533 534 /* Initialize cma parameters */ 535 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma)); 536 537 /* Prepare RDMA-CM private message */ 538 pmsg->cp_magic = rpcrdma_cmp_magic; 539 pmsg->cp_version = RPCRDMA_CMP_VERSION; 540 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK; 541 pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->rep_inline_send); 542 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->rep_inline_recv); 543 ep->rep_remote_cma.private_data = pmsg; 544 ep->rep_remote_cma.private_data_len = sizeof(*pmsg); 545 546 /* Client offers RDMA Read but does not initiate */ 547 ep->rep_remote_cma.initiator_depth = 0; 548 ep->rep_remote_cma.responder_resources = 549 min_t(int, U8_MAX, ia->ri_id->device->attrs.max_qp_rd_atom); 550 551 /* Limit transport retries so client can detect server 552 * GID changes quickly. RPC layer handles re-establishing 553 * transport connection and retransmission. 554 */ 555 ep->rep_remote_cma.retry_count = 6; 556 557 /* RPC-over-RDMA handles its own flow control. In addition, 558 * make all RNR NAKs visible so we know that RPC-over-RDMA 559 * flow control is working correctly (no NAKs should be seen). 560 */ 561 ep->rep_remote_cma.flow_control = 0; 562 ep->rep_remote_cma.rnr_retry_count = 0; 563 564 return 0; 565 566 out2: 567 ib_free_cq(sendcq); 568 out1: 569 return rc; 570 } 571 572 /** 573 * rpcrdma_ep_destroy - Disconnect and destroy endpoint. 574 * @r_xprt: transport instance to shut down 575 * 576 */ 577 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt) 578 { 579 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 580 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 581 582 if (ia->ri_id && ia->ri_id->qp) { 583 rpcrdma_ep_disconnect(ep, ia); 584 rdma_destroy_qp(ia->ri_id); 585 ia->ri_id->qp = NULL; 586 } 587 588 if (ep->rep_attr.recv_cq) 589 ib_free_cq(ep->rep_attr.recv_cq); 590 if (ep->rep_attr.send_cq) 591 ib_free_cq(ep->rep_attr.send_cq); 592 } 593 594 /* Re-establish a connection after a device removal event. 595 * Unlike a normal reconnection, a fresh PD and a new set 596 * of MRs and buffers is needed. 597 */ 598 static int rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt, 599 struct ib_qp_init_attr *qp_init_attr) 600 { 601 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 602 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 603 int rc, err; 604 605 trace_xprtrdma_reinsert(r_xprt); 606 607 rc = -EHOSTUNREACH; 608 if (rpcrdma_ia_open(r_xprt)) 609 goto out1; 610 611 rc = -ENOMEM; 612 err = rpcrdma_ep_create(r_xprt); 613 if (err) { 614 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err); 615 goto out2; 616 } 617 memcpy(qp_init_attr, &ep->rep_attr, sizeof(*qp_init_attr)); 618 619 rc = -ENETUNREACH; 620 err = rdma_create_qp(ia->ri_id, ia->ri_pd, qp_init_attr); 621 if (err) { 622 pr_err("rpcrdma: rdma_create_qp returned %d\n", err); 623 goto out3; 624 } 625 return 0; 626 627 out3: 628 rpcrdma_ep_destroy(r_xprt); 629 out2: 630 rpcrdma_ia_close(ia); 631 out1: 632 return rc; 633 } 634 635 static int rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, 636 struct ib_qp_init_attr *qp_init_attr) 637 { 638 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 639 struct rdma_cm_id *id, *old; 640 int err, rc; 641 642 rpcrdma_ep_disconnect(&r_xprt->rx_ep, ia); 643 644 rc = -EHOSTUNREACH; 645 id = rpcrdma_create_id(r_xprt, ia); 646 if (IS_ERR(id)) 647 goto out; 648 649 /* As long as the new ID points to the same device as the 650 * old ID, we can reuse the transport's existing PD and all 651 * previously allocated MRs. Also, the same device means 652 * the transport's previous DMA mappings are still valid. 653 * 654 * This is a sanity check only. There should be no way these 655 * point to two different devices here. 656 */ 657 old = id; 658 rc = -ENETUNREACH; 659 if (ia->ri_id->device != id->device) { 660 pr_err("rpcrdma: can't reconnect on different device!\n"); 661 goto out_destroy; 662 } 663 664 err = rdma_create_qp(id, ia->ri_pd, qp_init_attr); 665 if (err) 666 goto out_destroy; 667 668 /* Atomically replace the transport's ID and QP. */ 669 rc = 0; 670 old = ia->ri_id; 671 ia->ri_id = id; 672 rdma_destroy_qp(old); 673 674 out_destroy: 675 rdma_destroy_id(old); 676 out: 677 return rc; 678 } 679 680 /* 681 * Connect unconnected endpoint. 682 */ 683 int 684 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) 685 { 686 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt, 687 rx_ia); 688 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 689 struct ib_qp_init_attr qp_init_attr; 690 int rc; 691 692 retry: 693 memcpy(&qp_init_attr, &ep->rep_attr, sizeof(qp_init_attr)); 694 switch (ep->rep_connected) { 695 case 0: 696 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &qp_init_attr); 697 if (rc) { 698 rc = -ENETUNREACH; 699 goto out_noupdate; 700 } 701 break; 702 case -ENODEV: 703 rc = rpcrdma_ep_recreate_xprt(r_xprt, &qp_init_attr); 704 if (rc) 705 goto out_noupdate; 706 break; 707 default: 708 rc = rpcrdma_ep_reconnect(r_xprt, &qp_init_attr); 709 if (rc) 710 goto out; 711 } 712 713 ep->rep_connected = 0; 714 xprt_clear_connected(xprt); 715 716 rpcrdma_reset_cwnd(r_xprt); 717 rpcrdma_post_recvs(r_xprt, true); 718 719 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma); 720 if (rc) 721 goto out; 722 723 if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO) 724 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO; 725 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0); 726 if (ep->rep_connected <= 0) { 727 if (ep->rep_connected == -EAGAIN) 728 goto retry; 729 rc = ep->rep_connected; 730 goto out; 731 } 732 733 rpcrdma_mrs_create(r_xprt); 734 735 out: 736 if (rc) 737 ep->rep_connected = rc; 738 739 out_noupdate: 740 trace_xprtrdma_connect(r_xprt, rc); 741 return rc; 742 } 743 744 /** 745 * rpcrdma_ep_disconnect - Disconnect underlying transport 746 * @ep: endpoint to disconnect 747 * @ia: associated interface adapter 748 * 749 * Caller serializes. Either the transport send lock is held, 750 * or we're being called to destroy the transport. 751 */ 752 void 753 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) 754 { 755 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt, 756 rx_ep); 757 int rc; 758 759 /* returns without wait if ID is not connected */ 760 rc = rdma_disconnect(ia->ri_id); 761 if (!rc) 762 wait_event_interruptible(ep->rep_connect_wait, 763 ep->rep_connected != 1); 764 else 765 ep->rep_connected = rc; 766 trace_xprtrdma_disconnect(r_xprt, rc); 767 768 rpcrdma_xprt_drain(r_xprt); 769 rpcrdma_reqs_reset(r_xprt); 770 rpcrdma_mrs_destroy(r_xprt); 771 } 772 773 /* Fixed-size circular FIFO queue. This implementation is wait-free and 774 * lock-free. 775 * 776 * Consumer is the code path that posts Sends. This path dequeues a 777 * sendctx for use by a Send operation. Multiple consumer threads 778 * are serialized by the RPC transport lock, which allows only one 779 * ->send_request call at a time. 780 * 781 * Producer is the code path that handles Send completions. This path 782 * enqueues a sendctx that has been completed. Multiple producer 783 * threads are serialized by the ib_poll_cq() function. 784 */ 785 786 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced 787 * queue activity, and rpcrdma_xprt_drain has flushed all remaining 788 * Send requests. 789 */ 790 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf) 791 { 792 unsigned long i; 793 794 for (i = 0; i <= buf->rb_sc_last; i++) 795 kfree(buf->rb_sc_ctxs[i]); 796 kfree(buf->rb_sc_ctxs); 797 } 798 799 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia) 800 { 801 struct rpcrdma_sendctx *sc; 802 803 sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges), 804 GFP_KERNEL); 805 if (!sc) 806 return NULL; 807 808 sc->sc_cqe.done = rpcrdma_wc_send; 809 return sc; 810 } 811 812 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt) 813 { 814 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 815 struct rpcrdma_sendctx *sc; 816 unsigned long i; 817 818 /* Maximum number of concurrent outstanding Send WRs. Capping 819 * the circular queue size stops Send Queue overflow by causing 820 * the ->send_request call to fail temporarily before too many 821 * Sends are posted. 822 */ 823 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; 824 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i); 825 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL); 826 if (!buf->rb_sc_ctxs) 827 return -ENOMEM; 828 829 buf->rb_sc_last = i - 1; 830 for (i = 0; i <= buf->rb_sc_last; i++) { 831 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia); 832 if (!sc) 833 return -ENOMEM; 834 835 buf->rb_sc_ctxs[i] = sc; 836 } 837 838 return 0; 839 } 840 841 /* The sendctx queue is not guaranteed to have a size that is a 842 * power of two, thus the helpers in circ_buf.h cannot be used. 843 * The other option is to use modulus (%), which can be expensive. 844 */ 845 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf, 846 unsigned long item) 847 { 848 return likely(item < buf->rb_sc_last) ? item + 1 : 0; 849 } 850 851 /** 852 * rpcrdma_sendctx_get_locked - Acquire a send context 853 * @r_xprt: controlling transport instance 854 * 855 * Returns pointer to a free send completion context; or NULL if 856 * the queue is empty. 857 * 858 * Usage: Called to acquire an SGE array before preparing a Send WR. 859 * 860 * The caller serializes calls to this function (per transport), and 861 * provides an effective memory barrier that flushes the new value 862 * of rb_sc_head. 863 */ 864 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt) 865 { 866 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 867 struct rpcrdma_sendctx *sc; 868 unsigned long next_head; 869 870 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head); 871 872 if (next_head == READ_ONCE(buf->rb_sc_tail)) 873 goto out_emptyq; 874 875 /* ORDER: item must be accessed _before_ head is updated */ 876 sc = buf->rb_sc_ctxs[next_head]; 877 878 /* Releasing the lock in the caller acts as a memory 879 * barrier that flushes rb_sc_head. 880 */ 881 buf->rb_sc_head = next_head; 882 883 return sc; 884 885 out_emptyq: 886 /* The queue is "empty" if there have not been enough Send 887 * completions recently. This is a sign the Send Queue is 888 * backing up. Cause the caller to pause and try again. 889 */ 890 xprt_wait_for_buffer_space(&r_xprt->rx_xprt); 891 r_xprt->rx_stats.empty_sendctx_q++; 892 return NULL; 893 } 894 895 /** 896 * rpcrdma_sendctx_put_locked - Release a send context 897 * @r_xprt: controlling transport instance 898 * @sc: send context to release 899 * 900 * Usage: Called from Send completion to return a sendctxt 901 * to the queue. 902 * 903 * The caller serializes calls to this function (per transport). 904 */ 905 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt, 906 struct rpcrdma_sendctx *sc) 907 { 908 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 909 unsigned long next_tail; 910 911 /* Unmap SGEs of previously completed but unsignaled 912 * Sends by walking up the queue until @sc is found. 913 */ 914 next_tail = buf->rb_sc_tail; 915 do { 916 next_tail = rpcrdma_sendctx_next(buf, next_tail); 917 918 /* ORDER: item must be accessed _before_ tail is updated */ 919 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]); 920 921 } while (buf->rb_sc_ctxs[next_tail] != sc); 922 923 /* Paired with READ_ONCE */ 924 smp_store_release(&buf->rb_sc_tail, next_tail); 925 926 xprt_write_space(&r_xprt->rx_xprt); 927 } 928 929 static void 930 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt) 931 { 932 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 933 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 934 unsigned int count; 935 936 for (count = 0; count < ia->ri_max_segs; count++) { 937 struct rpcrdma_mr *mr; 938 int rc; 939 940 mr = kzalloc(sizeof(*mr), GFP_NOFS); 941 if (!mr) 942 break; 943 944 rc = frwr_init_mr(ia, mr); 945 if (rc) { 946 kfree(mr); 947 break; 948 } 949 950 mr->mr_xprt = r_xprt; 951 952 spin_lock(&buf->rb_lock); 953 rpcrdma_mr_push(mr, &buf->rb_mrs); 954 list_add(&mr->mr_all, &buf->rb_all_mrs); 955 spin_unlock(&buf->rb_lock); 956 } 957 958 r_xprt->rx_stats.mrs_allocated += count; 959 trace_xprtrdma_createmrs(r_xprt, count); 960 } 961 962 static void 963 rpcrdma_mr_refresh_worker(struct work_struct *work) 964 { 965 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer, 966 rb_refresh_worker); 967 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt, 968 rx_buf); 969 970 rpcrdma_mrs_create(r_xprt); 971 xprt_write_space(&r_xprt->rx_xprt); 972 } 973 974 /** 975 * rpcrdma_mrs_refresh - Wake the MR refresh worker 976 * @r_xprt: controlling transport instance 977 * 978 */ 979 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt) 980 { 981 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 982 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 983 984 /* If there is no underlying device, it's no use to 985 * wake the refresh worker. 986 */ 987 if (ep->rep_connected != -ENODEV) { 988 /* The work is scheduled on a WQ_MEM_RECLAIM 989 * workqueue in order to prevent MR allocation 990 * from recursing into NFS during direct reclaim. 991 */ 992 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker); 993 } 994 } 995 996 /** 997 * rpcrdma_req_create - Allocate an rpcrdma_req object 998 * @r_xprt: controlling r_xprt 999 * @size: initial size, in bytes, of send and receive buffers 1000 * @flags: GFP flags passed to memory allocators 1001 * 1002 * Returns an allocated and fully initialized rpcrdma_req or NULL. 1003 */ 1004 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size, 1005 gfp_t flags) 1006 { 1007 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; 1008 struct rpcrdma_regbuf *rb; 1009 struct rpcrdma_req *req; 1010 size_t maxhdrsize; 1011 1012 req = kzalloc(sizeof(*req), flags); 1013 if (req == NULL) 1014 goto out1; 1015 1016 /* Compute maximum header buffer size in bytes */ 1017 maxhdrsize = rpcrdma_fixed_maxsz + 3 + 1018 r_xprt->rx_ia.ri_max_segs * rpcrdma_readchunk_maxsz; 1019 maxhdrsize *= sizeof(__be32); 1020 rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize), 1021 DMA_TO_DEVICE, flags); 1022 if (!rb) 1023 goto out2; 1024 req->rl_rdmabuf = rb; 1025 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb)); 1026 1027 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags); 1028 if (!req->rl_sendbuf) 1029 goto out3; 1030 1031 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags); 1032 if (!req->rl_recvbuf) 1033 goto out4; 1034 1035 INIT_LIST_HEAD(&req->rl_free_mrs); 1036 INIT_LIST_HEAD(&req->rl_registered); 1037 spin_lock(&buffer->rb_lock); 1038 list_add(&req->rl_all, &buffer->rb_allreqs); 1039 spin_unlock(&buffer->rb_lock); 1040 return req; 1041 1042 out4: 1043 kfree(req->rl_sendbuf); 1044 out3: 1045 kfree(req->rl_rdmabuf); 1046 out2: 1047 kfree(req); 1048 out1: 1049 return NULL; 1050 } 1051 1052 /** 1053 * rpcrdma_reqs_reset - Reset all reqs owned by a transport 1054 * @r_xprt: controlling transport instance 1055 * 1056 * ASSUMPTION: the rb_allreqs list is stable for the duration, 1057 * and thus can be walked without holding rb_lock. Eg. the 1058 * caller is holding the transport send lock to exclude 1059 * device removal or disconnection. 1060 */ 1061 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt) 1062 { 1063 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1064 struct rpcrdma_req *req; 1065 1066 list_for_each_entry(req, &buf->rb_allreqs, rl_all) { 1067 /* Credits are valid only for one connection */ 1068 req->rl_slot.rq_cong = 0; 1069 } 1070 } 1071 1072 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt, 1073 bool temp) 1074 { 1075 struct rpcrdma_rep *rep; 1076 1077 rep = kzalloc(sizeof(*rep), GFP_KERNEL); 1078 if (rep == NULL) 1079 goto out; 1080 1081 rep->rr_rdmabuf = rpcrdma_regbuf_alloc(r_xprt->rx_ep.rep_inline_recv, 1082 DMA_FROM_DEVICE, GFP_KERNEL); 1083 if (!rep->rr_rdmabuf) 1084 goto out_free; 1085 1086 xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf), 1087 rdmab_length(rep->rr_rdmabuf)); 1088 rep->rr_cqe.done = rpcrdma_wc_receive; 1089 rep->rr_rxprt = r_xprt; 1090 rep->rr_recv_wr.next = NULL; 1091 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe; 1092 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov; 1093 rep->rr_recv_wr.num_sge = 1; 1094 rep->rr_temp = temp; 1095 list_add(&rep->rr_all, &r_xprt->rx_buf.rb_all_reps); 1096 return rep; 1097 1098 out_free: 1099 kfree(rep); 1100 out: 1101 return NULL; 1102 } 1103 1104 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep) 1105 { 1106 list_del(&rep->rr_all); 1107 rpcrdma_regbuf_free(rep->rr_rdmabuf); 1108 kfree(rep); 1109 } 1110 1111 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf) 1112 { 1113 struct llist_node *node; 1114 1115 /* Calls to llist_del_first are required to be serialized */ 1116 node = llist_del_first(&buf->rb_free_reps); 1117 if (!node) 1118 return NULL; 1119 return llist_entry(node, struct rpcrdma_rep, rr_node); 1120 } 1121 1122 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf, 1123 struct rpcrdma_rep *rep) 1124 { 1125 llist_add(&rep->rr_node, &buf->rb_free_reps); 1126 } 1127 1128 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt) 1129 { 1130 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1131 struct rpcrdma_rep *rep; 1132 1133 list_for_each_entry(rep, &buf->rb_all_reps, rr_all) 1134 rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf); 1135 } 1136 1137 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf) 1138 { 1139 struct rpcrdma_rep *rep; 1140 1141 while ((rep = rpcrdma_rep_get_locked(buf)) != NULL) 1142 rpcrdma_rep_destroy(rep); 1143 } 1144 1145 /** 1146 * rpcrdma_buffer_create - Create initial set of req/rep objects 1147 * @r_xprt: transport instance to (re)initialize 1148 * 1149 * Returns zero on success, otherwise a negative errno. 1150 */ 1151 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt) 1152 { 1153 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1154 int i, rc; 1155 1156 buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests; 1157 buf->rb_bc_srv_max_requests = 0; 1158 spin_lock_init(&buf->rb_lock); 1159 INIT_LIST_HEAD(&buf->rb_mrs); 1160 INIT_LIST_HEAD(&buf->rb_all_mrs); 1161 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker); 1162 1163 INIT_LIST_HEAD(&buf->rb_send_bufs); 1164 INIT_LIST_HEAD(&buf->rb_allreqs); 1165 INIT_LIST_HEAD(&buf->rb_all_reps); 1166 1167 rc = -ENOMEM; 1168 for (i = 0; i < buf->rb_max_requests; i++) { 1169 struct rpcrdma_req *req; 1170 1171 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2, 1172 GFP_KERNEL); 1173 if (!req) 1174 goto out; 1175 list_add(&req->rl_list, &buf->rb_send_bufs); 1176 } 1177 1178 init_llist_head(&buf->rb_free_reps); 1179 1180 rc = rpcrdma_sendctxs_create(r_xprt); 1181 if (rc) 1182 goto out; 1183 1184 return 0; 1185 out: 1186 rpcrdma_buffer_destroy(buf); 1187 return rc; 1188 } 1189 1190 /** 1191 * rpcrdma_req_destroy - Destroy an rpcrdma_req object 1192 * @req: unused object to be destroyed 1193 * 1194 * Relies on caller holding the transport send lock to protect 1195 * removing req->rl_all from buf->rb_all_reqs safely. 1196 */ 1197 void rpcrdma_req_destroy(struct rpcrdma_req *req) 1198 { 1199 struct rpcrdma_mr *mr; 1200 1201 list_del(&req->rl_all); 1202 1203 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) { 1204 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf; 1205 1206 spin_lock(&buf->rb_lock); 1207 list_del(&mr->mr_all); 1208 spin_unlock(&buf->rb_lock); 1209 1210 frwr_release_mr(mr); 1211 } 1212 1213 rpcrdma_regbuf_free(req->rl_recvbuf); 1214 rpcrdma_regbuf_free(req->rl_sendbuf); 1215 rpcrdma_regbuf_free(req->rl_rdmabuf); 1216 kfree(req); 1217 } 1218 1219 /** 1220 * rpcrdma_mrs_destroy - Release all of a transport's MRs 1221 * @r_xprt: controlling transport instance 1222 * 1223 * Relies on caller holding the transport send lock to protect 1224 * removing mr->mr_list from req->rl_free_mrs safely. 1225 */ 1226 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt) 1227 { 1228 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1229 struct rpcrdma_mr *mr; 1230 1231 cancel_work_sync(&buf->rb_refresh_worker); 1232 1233 spin_lock(&buf->rb_lock); 1234 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs, 1235 struct rpcrdma_mr, 1236 mr_all)) != NULL) { 1237 list_del(&mr->mr_list); 1238 list_del(&mr->mr_all); 1239 spin_unlock(&buf->rb_lock); 1240 1241 frwr_release_mr(mr); 1242 1243 spin_lock(&buf->rb_lock); 1244 } 1245 spin_unlock(&buf->rb_lock); 1246 } 1247 1248 /** 1249 * rpcrdma_buffer_destroy - Release all hw resources 1250 * @buf: root control block for resources 1251 * 1252 * ORDERING: relies on a prior rpcrdma_xprt_drain : 1253 * - No more Send or Receive completions can occur 1254 * - All MRs, reps, and reqs are returned to their free lists 1255 */ 1256 void 1257 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf) 1258 { 1259 rpcrdma_sendctxs_destroy(buf); 1260 rpcrdma_reps_destroy(buf); 1261 1262 while (!list_empty(&buf->rb_send_bufs)) { 1263 struct rpcrdma_req *req; 1264 1265 req = list_first_entry(&buf->rb_send_bufs, 1266 struct rpcrdma_req, rl_list); 1267 list_del(&req->rl_list); 1268 rpcrdma_req_destroy(req); 1269 } 1270 } 1271 1272 /** 1273 * rpcrdma_mr_get - Allocate an rpcrdma_mr object 1274 * @r_xprt: controlling transport 1275 * 1276 * Returns an initialized rpcrdma_mr or NULL if no free 1277 * rpcrdma_mr objects are available. 1278 */ 1279 struct rpcrdma_mr * 1280 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt) 1281 { 1282 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1283 struct rpcrdma_mr *mr; 1284 1285 spin_lock(&buf->rb_lock); 1286 mr = rpcrdma_mr_pop(&buf->rb_mrs); 1287 spin_unlock(&buf->rb_lock); 1288 return mr; 1289 } 1290 1291 /** 1292 * rpcrdma_mr_put - DMA unmap an MR and release it 1293 * @mr: MR to release 1294 * 1295 */ 1296 void rpcrdma_mr_put(struct rpcrdma_mr *mr) 1297 { 1298 struct rpcrdma_xprt *r_xprt = mr->mr_xprt; 1299 1300 if (mr->mr_dir != DMA_NONE) { 1301 trace_xprtrdma_mr_unmap(mr); 1302 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device, 1303 mr->mr_sg, mr->mr_nents, mr->mr_dir); 1304 mr->mr_dir = DMA_NONE; 1305 } 1306 1307 rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs); 1308 } 1309 1310 /** 1311 * rpcrdma_buffer_get - Get a request buffer 1312 * @buffers: Buffer pool from which to obtain a buffer 1313 * 1314 * Returns a fresh rpcrdma_req, or NULL if none are available. 1315 */ 1316 struct rpcrdma_req * 1317 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers) 1318 { 1319 struct rpcrdma_req *req; 1320 1321 spin_lock(&buffers->rb_lock); 1322 req = list_first_entry_or_null(&buffers->rb_send_bufs, 1323 struct rpcrdma_req, rl_list); 1324 if (req) 1325 list_del_init(&req->rl_list); 1326 spin_unlock(&buffers->rb_lock); 1327 return req; 1328 } 1329 1330 /** 1331 * rpcrdma_buffer_put - Put request/reply buffers back into pool 1332 * @buffers: buffer pool 1333 * @req: object to return 1334 * 1335 */ 1336 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req) 1337 { 1338 if (req->rl_reply) 1339 rpcrdma_rep_put(buffers, req->rl_reply); 1340 req->rl_reply = NULL; 1341 1342 spin_lock(&buffers->rb_lock); 1343 list_add(&req->rl_list, &buffers->rb_send_bufs); 1344 spin_unlock(&buffers->rb_lock); 1345 } 1346 1347 /** 1348 * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list 1349 * @rep: rep to release 1350 * 1351 * Used after error conditions. 1352 */ 1353 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep) 1354 { 1355 rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep); 1356 } 1357 1358 /* Returns a pointer to a rpcrdma_regbuf object, or NULL. 1359 * 1360 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for 1361 * receiving the payload of RDMA RECV operations. During Long Calls 1362 * or Replies they may be registered externally via frwr_map. 1363 */ 1364 static struct rpcrdma_regbuf * 1365 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction, 1366 gfp_t flags) 1367 { 1368 struct rpcrdma_regbuf *rb; 1369 1370 rb = kmalloc(sizeof(*rb), flags); 1371 if (!rb) 1372 return NULL; 1373 rb->rg_data = kmalloc(size, flags); 1374 if (!rb->rg_data) { 1375 kfree(rb); 1376 return NULL; 1377 } 1378 1379 rb->rg_device = NULL; 1380 rb->rg_direction = direction; 1381 rb->rg_iov.length = size; 1382 return rb; 1383 } 1384 1385 /** 1386 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer 1387 * @rb: regbuf to reallocate 1388 * @size: size of buffer to be allocated, in bytes 1389 * @flags: GFP flags 1390 * 1391 * Returns true if reallocation was successful. If false is 1392 * returned, @rb is left untouched. 1393 */ 1394 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags) 1395 { 1396 void *buf; 1397 1398 buf = kmalloc(size, flags); 1399 if (!buf) 1400 return false; 1401 1402 rpcrdma_regbuf_dma_unmap(rb); 1403 kfree(rb->rg_data); 1404 1405 rb->rg_data = buf; 1406 rb->rg_iov.length = size; 1407 return true; 1408 } 1409 1410 /** 1411 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf 1412 * @r_xprt: controlling transport instance 1413 * @rb: regbuf to be mapped 1414 * 1415 * Returns true if the buffer is now DMA mapped to @r_xprt's device 1416 */ 1417 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt, 1418 struct rpcrdma_regbuf *rb) 1419 { 1420 struct ib_device *device = r_xprt->rx_ia.ri_id->device; 1421 1422 if (rb->rg_direction == DMA_NONE) 1423 return false; 1424 1425 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb), 1426 rdmab_length(rb), rb->rg_direction); 1427 if (ib_dma_mapping_error(device, rdmab_addr(rb))) { 1428 trace_xprtrdma_dma_maperr(rdmab_addr(rb)); 1429 return false; 1430 } 1431 1432 rb->rg_device = device; 1433 rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey; 1434 return true; 1435 } 1436 1437 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb) 1438 { 1439 if (!rb) 1440 return; 1441 1442 if (!rpcrdma_regbuf_is_mapped(rb)) 1443 return; 1444 1445 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb), 1446 rb->rg_direction); 1447 rb->rg_device = NULL; 1448 } 1449 1450 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb) 1451 { 1452 rpcrdma_regbuf_dma_unmap(rb); 1453 if (rb) 1454 kfree(rb->rg_data); 1455 kfree(rb); 1456 } 1457 1458 /** 1459 * rpcrdma_ep_post - Post WRs to a transport's Send Queue 1460 * @ia: transport's device information 1461 * @ep: transport's RDMA endpoint information 1462 * @req: rpcrdma_req containing the Send WR to post 1463 * 1464 * Returns 0 if the post was successful, otherwise -ENOTCONN 1465 * is returned. 1466 */ 1467 int 1468 rpcrdma_ep_post(struct rpcrdma_ia *ia, 1469 struct rpcrdma_ep *ep, 1470 struct rpcrdma_req *req) 1471 { 1472 struct ib_send_wr *send_wr = &req->rl_wr; 1473 int rc; 1474 1475 if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) { 1476 send_wr->send_flags |= IB_SEND_SIGNALED; 1477 ep->rep_send_count = ep->rep_send_batch; 1478 } else { 1479 send_wr->send_flags &= ~IB_SEND_SIGNALED; 1480 --ep->rep_send_count; 1481 } 1482 1483 rc = frwr_send(ia, req); 1484 trace_xprtrdma_post_send(req, rc); 1485 if (rc) 1486 return -ENOTCONN; 1487 return 0; 1488 } 1489 1490 /** 1491 * rpcrdma_post_recvs - Refill the Receive Queue 1492 * @r_xprt: controlling transport instance 1493 * @temp: mark Receive buffers to be deleted after use 1494 * 1495 */ 1496 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp) 1497 { 1498 struct rpcrdma_buffer *buf = &r_xprt->rx_buf; 1499 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 1500 struct ib_recv_wr *i, *wr, *bad_wr; 1501 struct rpcrdma_rep *rep; 1502 int needed, count, rc; 1503 1504 rc = 0; 1505 count = 0; 1506 1507 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1); 1508 if (likely(ep->rep_receive_count > needed)) 1509 goto out; 1510 needed -= ep->rep_receive_count; 1511 if (!temp) 1512 needed += RPCRDMA_MAX_RECV_BATCH; 1513 1514 /* fast path: all needed reps can be found on the free list */ 1515 wr = NULL; 1516 while (needed) { 1517 rep = rpcrdma_rep_get_locked(buf); 1518 if (rep && rep->rr_temp) { 1519 rpcrdma_rep_destroy(rep); 1520 continue; 1521 } 1522 if (!rep) 1523 rep = rpcrdma_rep_create(r_xprt, temp); 1524 if (!rep) 1525 break; 1526 1527 rep->rr_recv_wr.next = wr; 1528 wr = &rep->rr_recv_wr; 1529 --needed; 1530 } 1531 if (!wr) 1532 goto out; 1533 1534 for (i = wr; i; i = i->next) { 1535 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr); 1536 1537 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf)) 1538 goto release_wrs; 1539 1540 trace_xprtrdma_post_recv(rep); 1541 ++count; 1542 } 1543 1544 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr, 1545 (const struct ib_recv_wr **)&bad_wr); 1546 out: 1547 trace_xprtrdma_post_recvs(r_xprt, count, rc); 1548 if (rc) { 1549 for (wr = bad_wr; wr;) { 1550 struct rpcrdma_rep *rep; 1551 1552 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr); 1553 wr = wr->next; 1554 rpcrdma_recv_buffer_put(rep); 1555 --count; 1556 } 1557 } 1558 ep->rep_receive_count += count; 1559 return; 1560 1561 release_wrs: 1562 for (i = wr; i;) { 1563 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr); 1564 i = i->next; 1565 rpcrdma_recv_buffer_put(rep); 1566 } 1567 } 1568