verbs.c

   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/bitops.h>
  53#include <linux/interrupt.h>
  54#include <linux/slab.h>
  55#include <linux/sunrpc/addr.h>
  56#include <linux/sunrpc/svc_rdma.h>
  57#include <linux/log2.h>
  58
  59#include <asm/barrier.h>
  60
  61#include <rdma/ib_cm.h>
  62
  63#include "xprt_rdma.h"
  64#include <trace/events/rpcrdma.h>
  65
  66static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt);
  67static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt);
  68static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
  69				       struct rpcrdma_sendctx *sc);
  70static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt);
  71static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt);
  72static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
  73static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
  74static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
  75static void rpcrdma_ep_get(struct rpcrdma_ep *ep);
  76static int rpcrdma_ep_put(struct rpcrdma_ep *ep);
  77static struct rpcrdma_regbuf *
  78rpcrdma_regbuf_alloc_node(size_t size, enum dma_data_direction direction,
  79			  int node);
  80static struct rpcrdma_regbuf *
  81rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction);
  82static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
  83static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
  84
  85/* Wait for outstanding transport work to finish. ib_drain_qp
  86 * handles the drains in the wrong order for us, so open code
  87 * them here.
  88 */
  89static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
  90{
  91	struct rpcrdma_ep *ep = r_xprt->rx_ep;
  92	struct rdma_cm_id *id = ep->re_id;
  93
  94	/* Wait for rpcrdma_post_recvs() to leave its critical
  95	 * section.
  96	 */
  97	if (atomic_inc_return(&ep->re_receiving) > 1)
  98		wait_for_completion(&ep->re_done);
  99
 100	/* Flush Receives, then wait for deferred Reply work
 101	 * to complete.
 102	 */
 103	ib_drain_rq(id->qp);
 104
 105	/* Deferred Reply processing might have scheduled
 106	 * local invalidations.
 107	 */
 108	ib_drain_sq(id->qp);
 109
 110	rpcrdma_ep_put(ep);
 111}
 112
 113/* Ensure xprt_force_disconnect() is invoked exactly once when a
 114 * connection is closed or lost. (The important thing is it needs
 115 * to be invoked "at least" once).
 116 */
 117void rpcrdma_force_disconnect(struct rpcrdma_ep *ep)
 118{
 119	if (atomic_add_unless(&ep->re_force_disconnect, 1, 1))
 120		xprt_force_disconnect(ep->re_xprt);
 121}
 122
 123/**
 124 * rpcrdma_flush_disconnect - Disconnect on flushed completion
 125 * @r_xprt: transport to disconnect
 126 * @wc: work completion entry
 127 *
 128 * Must be called in process context.
 129 */
 130void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc)
 131{
 132	if (wc->status != IB_WC_SUCCESS)
 133		rpcrdma_force_disconnect(r_xprt->rx_ep);
 134}
 135
 136/**
 137 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 138 * @cq:	completion queue
 139 * @wc:	WCE for a completed Send WR
 140 *
 141 */
 142static void rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 143{
 144	struct ib_cqe *cqe = wc->wr_cqe;
 145	struct rpcrdma_sendctx *sc =
 146		container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
 147	struct rpcrdma_xprt *r_xprt = cq->cq_context;
 148
 149	/* WARNING: Only wr_cqe and status are reliable at this point */
 150	trace_xprtrdma_wc_send(wc, &sc->sc_cid);
 151	rpcrdma_sendctx_put_locked(r_xprt, sc);
 152	rpcrdma_flush_disconnect(r_xprt, wc);
 153}
 154
 155/**
 156 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
 157 * @cq:	completion queue
 158 * @wc:	WCE for a completed Receive WR
 159 *
 160 */
 161static void rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
 162{
 163	struct ib_cqe *cqe = wc->wr_cqe;
 164	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 165					       rr_cqe);
 166	struct rpcrdma_xprt *r_xprt = cq->cq_context;
 167
 168	/* WARNING: Only wr_cqe and status are reliable at this point */
 169	trace_xprtrdma_wc_receive(wc, &rep->rr_cid);
 170	--r_xprt->rx_ep->re_receive_count;
 171	if (wc->status != IB_WC_SUCCESS)
 172		goto out_flushed;
 173
 174	/* status == SUCCESS means all fields in wc are trustworthy */
 175	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
 176	rep->rr_wc_flags = wc->wc_flags;
 177	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
 178
 179	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
 180				   rdmab_addr(rep->rr_rdmabuf),
 181				   wc->byte_len, DMA_FROM_DEVICE);
 182
 183	rpcrdma_reply_handler(rep);
 184	return;
 185
 186out_flushed:
 187	rpcrdma_flush_disconnect(r_xprt, wc);
 188	rpcrdma_rep_put(&r_xprt->rx_buf, rep);
 189}
 190
 191static void rpcrdma_update_cm_private(struct rpcrdma_ep *ep,
 192				      struct rdma_conn_param *param)
 193{
 194	const struct rpcrdma_connect_private *pmsg = param->private_data;
 195	unsigned int rsize, wsize;
 196
 197	/* Default settings for RPC-over-RDMA Version One */
 198	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 199	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 200
 201	if (pmsg &&
 202	    pmsg->cp_magic == rpcrdma_cmp_magic &&
 203	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
 204		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
 205		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
 206	}
 207
 208	if (rsize < ep->re_inline_recv)
 209		ep->re_inline_recv = rsize;
 210	if (wsize < ep->re_inline_send)
 211		ep->re_inline_send = wsize;
 212
 213	rpcrdma_set_max_header_sizes(ep);
 214}
 215
 216/**
 217 * rpcrdma_cm_event_handler - Handle RDMA CM events
 218 * @id: rdma_cm_id on which an event has occurred
 219 * @event: details of the event
 220 *
 221 * Called with @id's mutex held. Returns 1 if caller should
 222 * destroy @id, otherwise 0.
 223 */
 224static int
 225rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 226{
 227	struct rpcrdma_ep *ep = id->context;
 228
 229	might_sleep();
 230
 231	switch (event->event) {
 232	case RDMA_CM_EVENT_ADDR_RESOLVED:
 233	case RDMA_CM_EVENT_ROUTE_RESOLVED:
 234		ep->re_async_rc = 0;
 235		complete(&ep->re_done);
 236		return 0;
 237	case RDMA_CM_EVENT_ADDR_ERROR:
 238		ep->re_async_rc = -EPROTO;
 239		complete(&ep->re_done);
 240		return 0;
 241	case RDMA_CM_EVENT_ROUTE_ERROR:
 242		ep->re_async_rc = -ENETUNREACH;
 243		complete(&ep->re_done);
 244		return 0;
 245	case RDMA_CM_EVENT_ADDR_CHANGE:
 246		ep->re_connect_status = -ENODEV;
 247		goto disconnected;
 248	case RDMA_CM_EVENT_ESTABLISHED:
 249		rpcrdma_ep_get(ep);
 250		ep->re_connect_status = 1;
 251		rpcrdma_update_cm_private(ep, &event->param.conn);
 252		trace_xprtrdma_inline_thresh(ep);
 253		wake_up_all(&ep->re_connect_wait);
 254		break;
 255	case RDMA_CM_EVENT_CONNECT_ERROR:
 256		ep->re_connect_status = -ENOTCONN;
 257		goto wake_connect_worker;
 258	case RDMA_CM_EVENT_UNREACHABLE:
 259		ep->re_connect_status = -ENETUNREACH;
 260		goto wake_connect_worker;
 261	case RDMA_CM_EVENT_REJECTED:
 262		ep->re_connect_status = -ECONNREFUSED;
 263		if (event->status == IB_CM_REJ_STALE_CONN)
 264			ep->re_connect_status = -ENOTCONN;
 265wake_connect_worker:
 266		wake_up_all(&ep->re_connect_wait);
 267		return 0;
 268	case RDMA_CM_EVENT_DISCONNECTED:
 269		ep->re_connect_status = -ECONNABORTED;
 270disconnected:
 271		rpcrdma_force_disconnect(ep);
 272		return rpcrdma_ep_put(ep);
 273	default:
 274		break;
 275	}
 276
 277	return 0;
 278}
 279
 280static void rpcrdma_ep_removal_done(struct rpcrdma_notification *rn)
 281{
 282	struct rpcrdma_ep *ep = container_of(rn, struct rpcrdma_ep, re_rn);
 283
 284	trace_xprtrdma_device_removal(ep->re_id);
 285	xprt_force_disconnect(ep->re_xprt);
 286}
 287
 288static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
 289					    struct rpcrdma_ep *ep)
 290{
 291	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
 292	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 293	struct rdma_cm_id *id;
 294	int rc;
 295
 296	init_completion(&ep->re_done);
 297
 298	id = rdma_create_id(xprt->xprt_net, rpcrdma_cm_event_handler, ep,
 299			    RDMA_PS_TCP, IB_QPT_RC);
 300	if (IS_ERR(id))
 301		return id;
 302
 303	ep->re_async_rc = -ETIMEDOUT;
 304	rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)&xprt->addr,
 305			       RDMA_RESOLVE_TIMEOUT);
 306	if (rc)
 307		goto out;
 308	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
 309	if (rc < 0)
 310		goto out;
 311
 312	rc = ep->re_async_rc;
 313	if (rc)
 314		goto out;
 315
 316	ep->re_async_rc = -ETIMEDOUT;
 317	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 318	if (rc)
 319		goto out;
 320	rc = wait_for_completion_interruptible_timeout(&ep->re_done, wtimeout);
 321	if (rc < 0)
 322		goto out;
 323	rc = ep->re_async_rc;
 324	if (rc)
 325		goto out;
 326
 327	rc = rpcrdma_rn_register(id->device, &ep->re_rn, rpcrdma_ep_removal_done);
 328	if (rc)
 329		goto out;
 330
 331	return id;
 332
 333out:
 334	rdma_destroy_id(id);
 335	return ERR_PTR(rc);
 336}
 337
 338static void rpcrdma_ep_destroy(struct kref *kref)
 339{
 340	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
 341
 342	if (ep->re_id->qp) {
 343		rdma_destroy_qp(ep->re_id);
 344		ep->re_id->qp = NULL;
 345	}
 346
 347	if (ep->re_attr.recv_cq)
 348		ib_free_cq(ep->re_attr.recv_cq);
 349	ep->re_attr.recv_cq = NULL;
 350	if (ep->re_attr.send_cq)
 351		ib_free_cq(ep->re_attr.send_cq);
 352	ep->re_attr.send_cq = NULL;
 353
 354	if (ep->re_pd)
 355		ib_dealloc_pd(ep->re_pd);
 356	ep->re_pd = NULL;
 357
 358	rpcrdma_rn_unregister(ep->re_id->device, &ep->re_rn);
 359
 360	kfree(ep);
 361	module_put(THIS_MODULE);
 362}
 363
 364static noinline void rpcrdma_ep_get(struct rpcrdma_ep *ep)
 365{
 366	kref_get(&ep->re_kref);
 367}
 368
 369/* Returns:
 370 *     %0 if @ep still has a positive kref count, or
 371 *     %1 if @ep was destroyed successfully.
 372 */
 373static noinline int rpcrdma_ep_put(struct rpcrdma_ep *ep)
 374{
 375	return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
 376}
 377
 378static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
 379{
 380	struct rpcrdma_connect_private *pmsg;
 381	struct ib_device *device;
 382	struct rdma_cm_id *id;
 383	struct rpcrdma_ep *ep;
 384	int rc;
 385
 386	ep = kzalloc(sizeof(*ep), XPRTRDMA_GFP_FLAGS);
 387	if (!ep)
 388		return -ENOTCONN;
 389	ep->re_xprt = &r_xprt->rx_xprt;
 390	kref_init(&ep->re_kref);
 391
 392	id = rpcrdma_create_id(r_xprt, ep);
 393	if (IS_ERR(id)) {
 394		kfree(ep);
 395		return PTR_ERR(id);
 396	}
 397	__module_get(THIS_MODULE);
 398	device = id->device;
 399	ep->re_id = id;
 400	reinit_completion(&ep->re_done);
 401
 402	ep->re_max_requests = r_xprt->rx_xprt.max_reqs;
 403	ep->re_inline_send = xprt_rdma_max_inline_write;
 404	ep->re_inline_recv = xprt_rdma_max_inline_read;
 405	rc = frwr_query_device(ep, device);
 406	if (rc)
 407		goto out_destroy;
 408
 409	r_xprt->rx_buf.rb_max_requests = cpu_to_be32(ep->re_max_requests);
 410
 411	ep->re_attr.srq = NULL;
 412	ep->re_attr.cap.max_inline_data = 0;
 413	ep->re_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 414	ep->re_attr.qp_type = IB_QPT_RC;
 415	ep->re_attr.port_num = ~0;
 416
 417	ep->re_send_batch = ep->re_max_requests >> 3;
 418	ep->re_send_count = ep->re_send_batch;
 419	init_waitqueue_head(&ep->re_connect_wait);
 420
 421	ep->re_attr.send_cq = ib_alloc_cq_any(device, r_xprt,
 422					      ep->re_attr.cap.max_send_wr,
 423					      IB_POLL_WORKQUEUE);
 424	if (IS_ERR(ep->re_attr.send_cq)) {
 425		rc = PTR_ERR(ep->re_attr.send_cq);
 426		ep->re_attr.send_cq = NULL;
 427		goto out_destroy;
 428	}
 429
 430	ep->re_attr.recv_cq = ib_alloc_cq_any(device, r_xprt,
 431					      ep->re_attr.cap.max_recv_wr,
 432					      IB_POLL_WORKQUEUE);
 433	if (IS_ERR(ep->re_attr.recv_cq)) {
 434		rc = PTR_ERR(ep->re_attr.recv_cq);
 435		ep->re_attr.recv_cq = NULL;
 436		goto out_destroy;
 437	}
 438	ep->re_receive_count = 0;
 439
 440	/* Initialize cma parameters */
 441	memset(&ep->re_remote_cma, 0, sizeof(ep->re_remote_cma));
 442
 443	/* Prepare RDMA-CM private message */
 444	pmsg = &ep->re_cm_private;
 445	pmsg->cp_magic = rpcrdma_cmp_magic;
 446	pmsg->cp_version = RPCRDMA_CMP_VERSION;
 447	pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
 448	pmsg->cp_send_size = rpcrdma_encode_buffer_size(ep->re_inline_send);
 449	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(ep->re_inline_recv);
 450	ep->re_remote_cma.private_data = pmsg;
 451	ep->re_remote_cma.private_data_len = sizeof(*pmsg);
 452
 453	/* Client offers RDMA Read but does not initiate */
 454	ep->re_remote_cma.initiator_depth = 0;
 455	ep->re_remote_cma.responder_resources =
 456		min_t(int, U8_MAX, device->attrs.max_qp_rd_atom);
 457
 458	/* Limit transport retries so client can detect server
 459	 * GID changes quickly. RPC layer handles re-establishing
 460	 * transport connection and retransmission.
 461	 */
 462	ep->re_remote_cma.retry_count = 6;
 463
 464	/* RPC-over-RDMA handles its own flow control. In addition,
 465	 * make all RNR NAKs visible so we know that RPC-over-RDMA
 466	 * flow control is working correctly (no NAKs should be seen).
 467	 */
 468	ep->re_remote_cma.flow_control = 0;
 469	ep->re_remote_cma.rnr_retry_count = 0;
 470
 471	ep->re_pd = ib_alloc_pd(device, 0);
 472	if (IS_ERR(ep->re_pd)) {
 473		rc = PTR_ERR(ep->re_pd);
 474		ep->re_pd = NULL;
 475		goto out_destroy;
 476	}
 477
 478	rc = rdma_create_qp(id, ep->re_pd, &ep->re_attr);
 479	if (rc)
 480		goto out_destroy;
 481
 482	r_xprt->rx_ep = ep;
 483	return 0;
 484
 485out_destroy:
 486	rpcrdma_ep_put(ep);
 487	rdma_destroy_id(id);
 488	return rc;
 489}
 490
 491/**
 492 * rpcrdma_xprt_connect - Connect an unconnected transport
 493 * @r_xprt: controlling transport instance
 494 *
 495 * Returns 0 on success or a negative errno.
 496 */
 497int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 498{
 499	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 500	struct rpcrdma_ep *ep;
 501	int rc;
 502
 503	rc = rpcrdma_ep_create(r_xprt);
 504	if (rc)
 505		return rc;
 506	ep = r_xprt->rx_ep;
 507
 508	xprt_clear_connected(xprt);
 509	rpcrdma_reset_cwnd(r_xprt);
 510
 511	/* Bump the ep's reference count while there are
 512	 * outstanding Receives.
 513	 */
 514	rpcrdma_ep_get(ep);
 515	rpcrdma_post_recvs(r_xprt, 1);
 516
 517	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
 518	if (rc)
 519		goto out;
 520
 521	if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 522		xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 523	wait_event_interruptible(ep->re_connect_wait,
 524				 ep->re_connect_status != 0);
 525	if (ep->re_connect_status <= 0) {
 526		rc = ep->re_connect_status;
 527		goto out;
 528	}
 529
 530	rc = rpcrdma_sendctxs_create(r_xprt);
 531	if (rc) {
 532		rc = -ENOTCONN;
 533		goto out;
 534	}
 535
 536	rc = rpcrdma_reqs_setup(r_xprt);
 537	if (rc) {
 538		rc = -ENOTCONN;
 539		goto out;
 540	}
 541	rpcrdma_mrs_create(r_xprt);
 542	frwr_wp_create(r_xprt);
 543
 544out:
 545	trace_xprtrdma_connect(r_xprt, rc);
 546	return rc;
 547}
 548
 549/**
 550 * rpcrdma_xprt_disconnect - Disconnect underlying transport
 551 * @r_xprt: controlling transport instance
 552 *
 553 * Caller serializes. Either the transport send lock is held,
 554 * or we're being called to destroy the transport.
 555 *
 556 * On return, @r_xprt is completely divested of all hardware
 557 * resources and prepared for the next ->connect operation.
 558 */
 559void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
 560{
 561	struct rpcrdma_ep *ep = r_xprt->rx_ep;
 562	struct rdma_cm_id *id;
 563	int rc;
 564
 565	if (!ep)
 566		return;
 567
 568	id = ep->re_id;
 569	rc = rdma_disconnect(id);
 570	trace_xprtrdma_disconnect(r_xprt, rc);
 571
 572	rpcrdma_xprt_drain(r_xprt);
 573	rpcrdma_reps_unmap(r_xprt);
 574	rpcrdma_reqs_reset(r_xprt);
 575	rpcrdma_mrs_destroy(r_xprt);
 576	rpcrdma_sendctxs_destroy(r_xprt);
 577
 578	if (rpcrdma_ep_put(ep))
 579		rdma_destroy_id(id);
 580
 581	r_xprt->rx_ep = NULL;
 582}
 583
 584/* Fixed-size circular FIFO queue. This implementation is wait-free and
 585 * lock-free.
 586 *
 587 * Consumer is the code path that posts Sends. This path dequeues a
 588 * sendctx for use by a Send operation. Multiple consumer threads
 589 * are serialized by the RPC transport lock, which allows only one
 590 * ->send_request call at a time.
 591 *
 592 * Producer is the code path that handles Send completions. This path
 593 * enqueues a sendctx that has been completed. Multiple producer
 594 * threads are serialized by the ib_poll_cq() function.
 595 */
 596
 597/* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
 598 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
 599 * Send requests.
 600 */
 601static void rpcrdma_sendctxs_destroy(struct rpcrdma_xprt *r_xprt)
 602{
 603	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 604	unsigned long i;
 605
 606	if (!buf->rb_sc_ctxs)
 607		return;
 608	for (i = 0; i <= buf->rb_sc_last; i++)
 609		kfree(buf->rb_sc_ctxs[i]);
 610	kfree(buf->rb_sc_ctxs);
 611	buf->rb_sc_ctxs = NULL;
 612}
 613
 614static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ep *ep)
 615{
 616	struct rpcrdma_sendctx *sc;
 617
 618	sc = kzalloc(struct_size(sc, sc_sges, ep->re_attr.cap.max_send_sge),
 619		     XPRTRDMA_GFP_FLAGS);
 620	if (!sc)
 621		return NULL;
 622
 623	sc->sc_cqe.done = rpcrdma_wc_send;
 624	sc->sc_cid.ci_queue_id = ep->re_attr.send_cq->res.id;
 625	sc->sc_cid.ci_completion_id =
 626		atomic_inc_return(&ep->re_completion_ids);
 627	return sc;
 628}
 629
 630static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
 631{
 632	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 633	struct rpcrdma_sendctx *sc;
 634	unsigned long i;
 635
 636	/* Maximum number of concurrent outstanding Send WRs. Capping
 637	 * the circular queue size stops Send Queue overflow by causing
 638	 * the ->send_request call to fail temporarily before too many
 639	 * Sends are posted.
 640	 */
 641	i = r_xprt->rx_ep->re_max_requests + RPCRDMA_MAX_BC_REQUESTS;
 642	buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), XPRTRDMA_GFP_FLAGS);
 643	if (!buf->rb_sc_ctxs)
 644		return -ENOMEM;
 645
 646	buf->rb_sc_last = i - 1;
 647	for (i = 0; i <= buf->rb_sc_last; i++) {
 648		sc = rpcrdma_sendctx_create(r_xprt->rx_ep);
 649		if (!sc)
 650			return -ENOMEM;
 651
 652		buf->rb_sc_ctxs[i] = sc;
 653	}
 654
 655	buf->rb_sc_head = 0;
 656	buf->rb_sc_tail = 0;
 657	return 0;
 658}
 659
 660/* The sendctx queue is not guaranteed to have a size that is a
 661 * power of two, thus the helpers in circ_buf.h cannot be used.
 662 * The other option is to use modulus (%), which can be expensive.
 663 */
 664static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
 665					  unsigned long item)
 666{
 667	return likely(item < buf->rb_sc_last) ? item + 1 : 0;
 668}
 669
 670/**
 671 * rpcrdma_sendctx_get_locked - Acquire a send context
 672 * @r_xprt: controlling transport instance
 673 *
 674 * Returns pointer to a free send completion context; or NULL if
 675 * the queue is empty.
 676 *
 677 * Usage: Called to acquire an SGE array before preparing a Send WR.
 678 *
 679 * The caller serializes calls to this function (per transport), and
 680 * provides an effective memory barrier that flushes the new value
 681 * of rb_sc_head.
 682 */
 683struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
 684{
 685	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 686	struct rpcrdma_sendctx *sc;
 687	unsigned long next_head;
 688
 689	next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
 690
 691	if (next_head == READ_ONCE(buf->rb_sc_tail))
 692		goto out_emptyq;
 693
 694	/* ORDER: item must be accessed _before_ head is updated */
 695	sc = buf->rb_sc_ctxs[next_head];
 696
 697	/* Releasing the lock in the caller acts as a memory
 698	 * barrier that flushes rb_sc_head.
 699	 */
 700	buf->rb_sc_head = next_head;
 701
 702	return sc;
 703
 704out_emptyq:
 705	/* The queue is "empty" if there have not been enough Send
 706	 * completions recently. This is a sign the Send Queue is
 707	 * backing up. Cause the caller to pause and try again.
 708	 */
 709	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
 710	r_xprt->rx_stats.empty_sendctx_q++;
 711	return NULL;
 712}
 713
 714/**
 715 * rpcrdma_sendctx_put_locked - Release a send context
 716 * @r_xprt: controlling transport instance
 717 * @sc: send context to release
 718 *
 719 * Usage: Called from Send completion to return a sendctxt
 720 * to the queue.
 721 *
 722 * The caller serializes calls to this function (per transport).
 723 */
 724static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
 725				       struct rpcrdma_sendctx *sc)
 726{
 727	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 728	unsigned long next_tail;
 729
 730	/* Unmap SGEs of previously completed but unsignaled
 731	 * Sends by walking up the queue until @sc is found.
 732	 */
 733	next_tail = buf->rb_sc_tail;
 734	do {
 735		next_tail = rpcrdma_sendctx_next(buf, next_tail);
 736
 737		/* ORDER: item must be accessed _before_ tail is updated */
 738		rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
 739
 740	} while (buf->rb_sc_ctxs[next_tail] != sc);
 741
 742	/* Paired with READ_ONCE */
 743	smp_store_release(&buf->rb_sc_tail, next_tail);
 744
 745	xprt_write_space(&r_xprt->rx_xprt);
 746}
 747
 748static void
 749rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
 750{
 751	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 752	struct rpcrdma_ep *ep = r_xprt->rx_ep;
 753	struct ib_device *device = ep->re_id->device;
 754	unsigned int count;
 755
 756	/* Try to allocate enough to perform one full-sized I/O */
 757	for (count = 0; count < ep->re_max_rdma_segs; count++) {
 758		struct rpcrdma_mr *mr;
 759		int rc;
 760
 761		mr = kzalloc_node(sizeof(*mr), XPRTRDMA_GFP_FLAGS,
 762				  ibdev_to_node(device));
 763		if (!mr)
 764			break;
 765
 766		rc = frwr_mr_init(r_xprt, mr);
 767		if (rc) {
 768			kfree(mr);
 769			break;
 770		}
 771
 772		spin_lock(&buf->rb_lock);
 773		rpcrdma_mr_push(mr, &buf->rb_mrs);
 774		list_add(&mr->mr_all, &buf->rb_all_mrs);
 775		spin_unlock(&buf->rb_lock);
 776	}
 777
 778	r_xprt->rx_stats.mrs_allocated += count;
 779	trace_xprtrdma_createmrs(r_xprt, count);
 780}
 781
 782static void
 783rpcrdma_mr_refresh_worker(struct work_struct *work)
 784{
 785	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 786						  rb_refresh_worker);
 787	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
 788						   rx_buf);
 789
 790	rpcrdma_mrs_create(r_xprt);
 791	xprt_write_space(&r_xprt->rx_xprt);
 792}
 793
 794/**
 795 * rpcrdma_mrs_refresh - Wake the MR refresh worker
 796 * @r_xprt: controlling transport instance
 797 *
 798 */
 799void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
 800{
 801	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 802	struct rpcrdma_ep *ep = r_xprt->rx_ep;
 803
 804	/* If there is no underlying connection, it's no use
 805	 * to wake the refresh worker.
 806	 */
 807	if (ep->re_connect_status != 1)
 808		return;
 809	queue_work(system_highpri_wq, &buf->rb_refresh_worker);
 810}
 811
 812/**
 813 * rpcrdma_req_create - Allocate an rpcrdma_req object
 814 * @r_xprt: controlling r_xprt
 815 * @size: initial size, in bytes, of send and receive buffers
 816 *
 817 * Returns an allocated and fully initialized rpcrdma_req or NULL.
 818 */
 819struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt,
 820				       size_t size)
 821{
 822	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
 823	struct rpcrdma_req *req;
 824
 825	req = kzalloc(sizeof(*req), XPRTRDMA_GFP_FLAGS);
 826	if (req == NULL)
 827		goto out1;
 828
 829	req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE);
 830	if (!req->rl_sendbuf)
 831		goto out2;
 832
 833	req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE);
 834	if (!req->rl_recvbuf)
 835		goto out3;
 836
 837	INIT_LIST_HEAD(&req->rl_free_mrs);
 838	INIT_LIST_HEAD(&req->rl_registered);
 839	spin_lock(&buffer->rb_lock);
 840	list_add(&req->rl_all, &buffer->rb_allreqs);
 841	spin_unlock(&buffer->rb_lock);
 842	return req;
 843
 844out3:
 845	rpcrdma_regbuf_free(req->rl_sendbuf);
 846out2:
 847	kfree(req);
 848out1:
 849	return NULL;
 850}
 851
 852/**
 853 * rpcrdma_req_setup - Per-connection instance setup of an rpcrdma_req object
 854 * @r_xprt: controlling transport instance
 855 * @req: rpcrdma_req object to set up
 856 *
 857 * Returns zero on success, and a negative errno on failure.
 858 */
 859int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 860{
 861	struct rpcrdma_regbuf *rb;
 862	size_t maxhdrsize;
 863
 864	/* Compute maximum header buffer size in bytes */
 865	maxhdrsize = rpcrdma_fixed_maxsz + 3 +
 866		     r_xprt->rx_ep->re_max_rdma_segs * rpcrdma_readchunk_maxsz;
 867	maxhdrsize *= sizeof(__be32);
 868	rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
 869				  DMA_TO_DEVICE);
 870	if (!rb)
 871		goto out;
 872
 873	if (!__rpcrdma_regbuf_dma_map(r_xprt, rb))
 874		goto out_free;
 875
 876	req->rl_rdmabuf = rb;
 877	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
 878	return 0;
 879
 880out_free:
 881	rpcrdma_regbuf_free(rb);
 882out:
 883	return -ENOMEM;
 884}
 885
 886/* ASSUMPTION: the rb_allreqs list is stable for the duration,
 887 * and thus can be walked without holding rb_lock. Eg. the
 888 * caller is holding the transport send lock to exclude
 889 * device removal or disconnection.
 890 */
 891static int rpcrdma_reqs_setup(struct rpcrdma_xprt *r_xprt)
 892{
 893	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 894	struct rpcrdma_req *req;
 895	int rc;
 896
 897	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
 898		rc = rpcrdma_req_setup(r_xprt, req);
 899		if (rc)
 900			return rc;
 901	}
 902	return 0;
 903}
 904
 905static void rpcrdma_req_reset(struct rpcrdma_req *req)
 906{
 907	struct rpcrdma_mr *mr;
 908
 909	/* Credits are valid for only one connection */
 910	req->rl_slot.rq_cong = 0;
 911
 912	rpcrdma_regbuf_free(req->rl_rdmabuf);
 913	req->rl_rdmabuf = NULL;
 914
 915	rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
 916	rpcrdma_regbuf_dma_unmap(req->rl_recvbuf);
 917
 918	/* The verbs consumer can't know the state of an MR on the
 919	 * req->rl_registered list unless a successful completion
 920	 * has occurred, so they cannot be re-used.
 921	 */
 922	while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
 923		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
 924
 925		spin_lock(&buf->rb_lock);
 926		list_del(&mr->mr_all);
 927		spin_unlock(&buf->rb_lock);
 928
 929		frwr_mr_release(mr);
 930	}
 931}
 932
 933/* ASSUMPTION: the rb_allreqs list is stable for the duration,
 934 * and thus can be walked without holding rb_lock. Eg. the
 935 * caller is holding the transport send lock to exclude
 936 * device removal or disconnection.
 937 */
 938static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
 939{
 940	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 941	struct rpcrdma_req *req;
 942
 943	list_for_each_entry(req, &buf->rb_allreqs, rl_all)
 944		rpcrdma_req_reset(req);
 945}
 946
 947static noinline
 948struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt)
 949{
 950	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 951	struct rpcrdma_ep *ep = r_xprt->rx_ep;
 952	struct ib_device *device = ep->re_id->device;
 953	struct rpcrdma_rep *rep;
 954
 955	rep = kzalloc(sizeof(*rep), XPRTRDMA_GFP_FLAGS);
 956	if (rep == NULL)
 957		goto out;
 958
 959	rep->rr_rdmabuf = rpcrdma_regbuf_alloc_node(ep->re_inline_recv,
 960						    DMA_FROM_DEVICE,
 961						    ibdev_to_node(device));
 962	if (!rep->rr_rdmabuf)
 963		goto out_free;
 964
 965	rep->rr_cid.ci_completion_id =
 966		atomic_inc_return(&r_xprt->rx_ep->re_completion_ids);
 967
 968	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
 969		     rdmab_length(rep->rr_rdmabuf));
 970	rep->rr_cqe.done = rpcrdma_wc_receive;
 971	rep->rr_rxprt = r_xprt;
 972	rep->rr_recv_wr.next = NULL;
 973	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
 974	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
 975	rep->rr_recv_wr.num_sge = 1;
 976
 977	spin_lock(&buf->rb_lock);
 978	list_add(&rep->rr_all, &buf->rb_all_reps);
 979	spin_unlock(&buf->rb_lock);
 980	return rep;
 981
 982out_free:
 983	kfree(rep);
 984out:
 985	return NULL;
 986}
 987
 988static void rpcrdma_rep_free(struct rpcrdma_rep *rep)
 989{
 990	rpcrdma_regbuf_free(rep->rr_rdmabuf);
 991	kfree(rep);
 992}
 993
 994static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
 995{
 996	struct llist_node *node;
 997
 998	/* Calls to llist_del_first are required to be serialized */
 999	node = llist_del_first(&buf->rb_free_reps);
1000	if (!node)
1001		return NULL;
1002	return llist_entry(node, struct rpcrdma_rep, rr_node);
1003}
1004
1005/**
1006 * rpcrdma_rep_put - Release rpcrdma_rep back to free list
1007 * @buf: buffer pool
1008 * @rep: rep to release
1009 *
1010 */
1011void rpcrdma_rep_put(struct rpcrdma_buffer *buf, struct rpcrdma_rep *rep)
1012{
1013	llist_add(&rep->rr_node, &buf->rb_free_reps);
1014}
1015
1016/* Caller must ensure the QP is quiescent (RQ is drained) before
1017 * invoking this function, to guarantee rb_all_reps is not
1018 * changing.
1019 */
1020static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt)
1021{
1022	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1023	struct rpcrdma_rep *rep;
1024
1025	list_for_each_entry(rep, &buf->rb_all_reps, rr_all)
1026		rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
1027}
1028
1029static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1030{
1031	struct rpcrdma_rep *rep;
1032
1033	spin_lock(&buf->rb_lock);
1034	while ((rep = list_first_entry_or_null(&buf->rb_all_reps,
1035					       struct rpcrdma_rep,
1036					       rr_all)) != NULL) {
1037		list_del(&rep->rr_all);
1038		spin_unlock(&buf->rb_lock);
1039
1040		rpcrdma_rep_free(rep);
1041
1042		spin_lock(&buf->rb_lock);
1043	}
1044	spin_unlock(&buf->rb_lock);
1045}
1046
1047/**
1048 * rpcrdma_buffer_create - Create initial set of req/rep objects
1049 * @r_xprt: transport instance to (re)initialize
1050 *
1051 * Returns zero on success, otherwise a negative errno.
1052 */
1053int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1054{
1055	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1056	int i, rc;
1057
1058	buf->rb_bc_srv_max_requests = 0;
1059	spin_lock_init(&buf->rb_lock);
1060	INIT_LIST_HEAD(&buf->rb_mrs);
1061	INIT_LIST_HEAD(&buf->rb_all_mrs);
1062	INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1063
1064	INIT_LIST_HEAD(&buf->rb_send_bufs);
1065	INIT_LIST_HEAD(&buf->rb_allreqs);
1066	INIT_LIST_HEAD(&buf->rb_all_reps);
1067
1068	rc = -ENOMEM;
1069	for (i = 0; i < r_xprt->rx_xprt.max_reqs; i++) {
1070		struct rpcrdma_req *req;
1071
1072		req = rpcrdma_req_create(r_xprt,
1073					 RPCRDMA_V1_DEF_INLINE_SIZE * 2);
1074		if (!req)
1075			goto out;
1076		list_add(&req->rl_list, &buf->rb_send_bufs);
1077	}
1078
1079	init_llist_head(&buf->rb_free_reps);
1080
1081	return 0;
1082out:
1083	rpcrdma_buffer_destroy(buf);
1084	return rc;
1085}
1086
1087/**
1088 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1089 * @req: unused object to be destroyed
1090 *
1091 * Relies on caller holding the transport send lock to protect
1092 * removing req->rl_all from buf->rb_all_reqs safely.
1093 */
1094void rpcrdma_req_destroy(struct rpcrdma_req *req)
1095{
1096	struct rpcrdma_mr *mr;
1097
1098	list_del(&req->rl_all);
1099
1100	while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1101		struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1102
1103		spin_lock(&buf->rb_lock);
1104		list_del(&mr->mr_all);
1105		spin_unlock(&buf->rb_lock);
1106
1107		frwr_mr_release(mr);
1108	}
1109
1110	rpcrdma_regbuf_free(req->rl_recvbuf);
1111	rpcrdma_regbuf_free(req->rl_sendbuf);
1112	rpcrdma_regbuf_free(req->rl_rdmabuf);
1113	kfree(req);
1114}
1115
1116/**
1117 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1118 * @r_xprt: controlling transport instance
1119 *
1120 * Relies on caller holding the transport send lock to protect
1121 * removing mr->mr_list from req->rl_free_mrs safely.
1122 */
1123static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1124{
1125	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1126	struct rpcrdma_mr *mr;
1127
1128	cancel_work_sync(&buf->rb_refresh_worker);
1129
1130	spin_lock(&buf->rb_lock);
1131	while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1132					      struct rpcrdma_mr,
1133					      mr_all)) != NULL) {
1134		list_del(&mr->mr_list);
1135		list_del(&mr->mr_all);
1136		spin_unlock(&buf->rb_lock);
1137
1138		frwr_mr_release(mr);
1139
1140		spin_lock(&buf->rb_lock);
1141	}
1142	spin_unlock(&buf->rb_lock);
1143}
1144
1145/**
1146 * rpcrdma_buffer_destroy - Release all hw resources
1147 * @buf: root control block for resources
1148 *
1149 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1150 * - No more Send or Receive completions can occur
1151 * - All MRs, reps, and reqs are returned to their free lists
1152 */
1153void
1154rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1155{
1156	rpcrdma_reps_destroy(buf);
1157
1158	while (!list_empty(&buf->rb_send_bufs)) {
1159		struct rpcrdma_req *req;
1160
1161		req = list_first_entry(&buf->rb_send_bufs,
1162				       struct rpcrdma_req, rl_list);
1163		list_del(&req->rl_list);
1164		rpcrdma_req_destroy(req);
1165	}
1166}
1167
1168/**
1169 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1170 * @r_xprt: controlling transport
1171 *
1172 * Returns an initialized rpcrdma_mr or NULL if no free
1173 * rpcrdma_mr objects are available.
1174 */
1175struct rpcrdma_mr *
1176rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1177{
1178	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1179	struct rpcrdma_mr *mr;
1180
1181	spin_lock(&buf->rb_lock);
1182	mr = rpcrdma_mr_pop(&buf->rb_mrs);
1183	spin_unlock(&buf->rb_lock);
1184	return mr;
1185}
1186
1187/**
1188 * rpcrdma_reply_put - Put reply buffers back into pool
1189 * @buffers: buffer pool
1190 * @req: object to return
1191 *
1192 */
1193void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1194{
1195	if (req->rl_reply) {
1196		rpcrdma_rep_put(buffers, req->rl_reply);
1197		req->rl_reply = NULL;
1198	}
1199}
1200
1201/**
1202 * rpcrdma_buffer_get - Get a request buffer
1203 * @buffers: Buffer pool from which to obtain a buffer
1204 *
1205 * Returns a fresh rpcrdma_req, or NULL if none are available.
1206 */
1207struct rpcrdma_req *
1208rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1209{
1210	struct rpcrdma_req *req;
1211
1212	spin_lock(&buffers->rb_lock);
1213	req = list_first_entry_or_null(&buffers->rb_send_bufs,
1214				       struct rpcrdma_req, rl_list);
1215	if (req)
1216		list_del_init(&req->rl_list);
1217	spin_unlock(&buffers->rb_lock);
1218	return req;
1219}
1220
1221/**
1222 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1223 * @buffers: buffer pool
1224 * @req: object to return
1225 *
1226 */
1227void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1228{
1229	rpcrdma_reply_put(buffers, req);
1230
1231	spin_lock(&buffers->rb_lock);
1232	list_add(&req->rl_list, &buffers->rb_send_bufs);
1233	spin_unlock(&buffers->rb_lock);
1234}
1235
1236/* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1237 *
1238 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1239 * receiving the payload of RDMA RECV operations. During Long Calls
1240 * or Replies they may be registered externally via frwr_map.
1241 */
1242static struct rpcrdma_regbuf *
1243rpcrdma_regbuf_alloc_node(size_t size, enum dma_data_direction direction,
1244			  int node)
1245{
1246	struct rpcrdma_regbuf *rb;
1247
1248	rb = kmalloc_node(sizeof(*rb), XPRTRDMA_GFP_FLAGS, node);
1249	if (!rb)
1250		return NULL;
1251	rb->rg_data = kmalloc_node(size, XPRTRDMA_GFP_FLAGS, node);
1252	if (!rb->rg_data) {
1253		kfree(rb);
1254		return NULL;
1255	}
1256
1257	rb->rg_device = NULL;
1258	rb->rg_direction = direction;
1259	rb->rg_iov.length = size;
1260	return rb;
1261}
1262
1263static struct rpcrdma_regbuf *
1264rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction)
1265{
1266	return rpcrdma_regbuf_alloc_node(size, direction, NUMA_NO_NODE);
1267}
1268
1269/**
1270 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1271 * @rb: regbuf to reallocate
1272 * @size: size of buffer to be allocated, in bytes
1273 * @flags: GFP flags
1274 *
1275 * Returns true if reallocation was successful. If false is
1276 * returned, @rb is left untouched.
1277 */
1278bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1279{
1280	void *buf;
1281
1282	buf = kmalloc(size, flags);
1283	if (!buf)
1284		return false;
1285
1286	rpcrdma_regbuf_dma_unmap(rb);
1287	kfree(rb->rg_data);
1288
1289	rb->rg_data = buf;
1290	rb->rg_iov.length = size;
1291	return true;
1292}
1293
1294/**
1295 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1296 * @r_xprt: controlling transport instance
1297 * @rb: regbuf to be mapped
1298 *
1299 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1300 */
1301bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1302			      struct rpcrdma_regbuf *rb)
1303{
1304	struct ib_device *device = r_xprt->rx_ep->re_id->device;
1305
1306	if (rb->rg_direction == DMA_NONE)
1307		return false;
1308
1309	rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1310					    rdmab_length(rb), rb->rg_direction);
1311	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1312		trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1313		return false;
1314	}
1315
1316	rb->rg_device = device;
1317	rb->rg_iov.lkey = r_xprt->rx_ep->re_pd->local_dma_lkey;
1318	return true;
1319}
1320
1321static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1322{
1323	if (!rb)
1324		return;
1325
1326	if (!rpcrdma_regbuf_is_mapped(rb))
1327		return;
1328
1329	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1330			    rb->rg_direction);
1331	rb->rg_device = NULL;
1332}
1333
1334static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1335{
1336	rpcrdma_regbuf_dma_unmap(rb);
1337	if (rb)
1338		kfree(rb->rg_data);
1339	kfree(rb);
1340}
1341
1342/**
1343 * rpcrdma_post_recvs - Refill the Receive Queue
1344 * @r_xprt: controlling transport instance
1345 * @needed: current credit grant
1346 *
1347 */
1348void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed)
1349{
1350	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1351	struct rpcrdma_ep *ep = r_xprt->rx_ep;
1352	struct ib_recv_wr *wr, *bad_wr;
1353	struct rpcrdma_rep *rep;
1354	int count, rc;
1355
1356	rc = 0;
1357	count = 0;
1358
1359	if (likely(ep->re_receive_count > needed))
1360		goto out;
1361	needed -= ep->re_receive_count;
1362	needed += RPCRDMA_MAX_RECV_BATCH;
1363
1364	if (atomic_inc_return(&ep->re_receiving) > 1)
1365		goto out_dec;
1366
1367	/* fast path: all needed reps can be found on the free list */
1368	wr = NULL;
1369	while (needed) {
1370		rep = rpcrdma_rep_get_locked(buf);
1371		if (!rep)
1372			rep = rpcrdma_rep_create(r_xprt);
1373		if (!rep)
1374			break;
1375		if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf)) {
1376			rpcrdma_rep_put(buf, rep);
1377			break;
1378		}
1379
1380		rep->rr_cid.ci_queue_id = ep->re_attr.recv_cq->res.id;
1381		trace_xprtrdma_post_recv(&rep->rr_cid);
1382		rep->rr_recv_wr.next = wr;
1383		wr = &rep->rr_recv_wr;
1384		--needed;
1385		++count;
1386	}
1387	if (!wr)
1388		goto out_dec;
1389
1390	rc = ib_post_recv(ep->re_id->qp, wr,
1391			  (const struct ib_recv_wr **)&bad_wr);
1392	if (rc) {
1393		trace_xprtrdma_post_recvs_err(r_xprt, rc);
1394		for (wr = bad_wr; wr;) {
1395			struct rpcrdma_rep *rep;
1396
1397			rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1398			wr = wr->next;
1399			rpcrdma_rep_put(buf, rep);
1400			--count;
1401		}
1402	}
1403
1404out_dec:
1405	if (atomic_dec_return(&ep->re_receiving) > 0)
1406		complete(&ep->re_done);
1407out:
1408	trace_xprtrdma_post_recvs(r_xprt, count);
1409	ep->re_receive_count += count;
1410	return;
1411}