1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
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
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
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.
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
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.
45 * Encapsulates the major functions managing:
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>
58 #include <asm-generic/barrier.h>
59 #include <asm/bitops.h>
61 #include <rdma/ib_cm.h>
63 #include "xprt_rdma.h"
64 #include <trace/events/rpcrdma.h>
70 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
71 # define RPCDBG_FACILITY RPCDBG_TRANS
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_destroy(struct rpcrdma_buffer *buf);
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,
86 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb);
87 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb);
89 /* Wait for outstanding transport work to finish. ib_drain_qp
90 * handles the drains in the wrong order for us, so open code
93 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
95 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
97 /* Flush Receives, then wait for deferred Reply work
100 ib_drain_rq(ia->ri_id->qp);
102 /* Deferred Reply processing might have scheduled
103 * local invalidations.
105 ib_drain_sq(ia->ri_id->qp);
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
113 * Called from the RDMA provider (device driver) possibly in an interrupt
117 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
119 struct rpcrdma_ep *ep = context;
120 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
123 trace_xprtrdma_qp_event(r_xprt, event);
127 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
128 * @cq: completion queue
133 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
135 struct ib_cqe *cqe = wc->wr_cqe;
136 struct rpcrdma_sendctx *sc =
137 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
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);
145 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
146 * @cq: completion queue (ignored)
151 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
153 struct ib_cqe *cqe = wc->wr_cqe;
154 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
156 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
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)
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;
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);
173 rpcrdma_reply_handler(rep);
177 rpcrdma_recv_buffer_put(rep);
180 static void rpcrdma_update_cm_private(struct rpcrdma_xprt *r_xprt,
181 struct rdma_conn_param *param)
183 const struct rpcrdma_connect_private *pmsg = param->private_data;
184 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
185 unsigned int rsize, wsize;
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;
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);
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;
205 rpcrdma_set_max_header_sizes(r_xprt);
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
213 * Called with @id's mutex held. Returns 1 if caller should
214 * destroy @id, otherwise 0.
217 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
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;
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:
231 complete(&ia->ri_done);
233 case RDMA_CM_EVENT_ADDR_ERROR:
234 ia->ri_async_rc = -EPROTO;
235 complete(&ia->ri_done);
237 case RDMA_CM_EVENT_ROUTE_ERROR:
238 ia->ri_async_rc = -ENETUNREACH;
239 complete(&ia->ri_done);
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));
247 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
248 ep->rep_connected = -ENODEV;
249 xprt_force_disconnect(xprt);
250 wait_for_completion(&ia->ri_remove_done);
253 /* Return 1 to ensure the core destroys the id. */
255 case RDMA_CM_EVENT_ESTABLISHED:
256 ++xprt->connect_cookie;
257 ep->rep_connected = 1;
258 rpcrdma_update_cm_private(r_xprt, &event->param.conn);
259 trace_xprtrdma_inline_thresh(r_xprt);
260 wake_up_all(&ep->rep_connect_wait);
262 case RDMA_CM_EVENT_CONNECT_ERROR:
263 ep->rep_connected = -ENOTCONN;
265 case RDMA_CM_EVENT_UNREACHABLE:
266 ep->rep_connected = -ENETUNREACH;
268 case RDMA_CM_EVENT_REJECTED:
269 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
270 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
271 rdma_reject_msg(id, event->status));
272 ep->rep_connected = -ECONNREFUSED;
273 if (event->status == IB_CM_REJ_STALE_CONN)
274 ep->rep_connected = -EAGAIN;
276 case RDMA_CM_EVENT_DISCONNECTED:
277 ep->rep_connected = -ECONNABORTED;
279 xprt_force_disconnect(xprt);
280 wake_up_all(&ep->rep_connect_wait);
286 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
287 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
288 ia->ri_id->device->name, rdma_event_msg(event->event));
292 static struct rdma_cm_id *
293 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
295 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
296 struct rdma_cm_id *id;
299 init_completion(&ia->ri_done);
300 init_completion(&ia->ri_remove_done);
302 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
303 xprt, RDMA_PS_TCP, IB_QPT_RC);
307 ia->ri_async_rc = -ETIMEDOUT;
308 rc = rdma_resolve_addr(id, NULL,
309 (struct sockaddr *)&xprt->rx_xprt.addr,
310 RDMA_RESOLVE_TIMEOUT);
313 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
317 rc = ia->ri_async_rc;
321 ia->ri_async_rc = -ETIMEDOUT;
322 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
325 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
328 rc = ia->ri_async_rc;
340 * Exported functions.
344 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
345 * @xprt: transport with IA to (re)initialize
347 * Returns 0 on success, negative errno if an appropriate
348 * Interface Adapter could not be found and opened.
351 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
353 struct rpcrdma_ia *ia = &xprt->rx_ia;
356 ia->ri_id = rpcrdma_create_id(xprt, ia);
357 if (IS_ERR(ia->ri_id)) {
358 rc = PTR_ERR(ia->ri_id);
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);
369 switch (xprt_rdma_memreg_strategy) {
371 if (frwr_is_supported(ia->ri_id->device))
375 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
376 ia->ri_id->device->name, xprt_rdma_memreg_strategy);
384 rpcrdma_ia_close(ia);
389 * rpcrdma_ia_remove - Handle device driver unload
390 * @ia: interface adapter being removed
392 * Divest transport H/W resources associated with this adapter,
393 * but allow it to be restored later.
396 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
398 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
400 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
401 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
402 struct rpcrdma_req *req;
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.
412 rpcrdma_xprt_drain(r_xprt);
413 rdma_destroy_qp(ia->ri_id);
414 ia->ri_id->qp = NULL;
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;
421 /* The ULP is responsible for ensuring all DMA
422 * mappings and MRs are gone.
424 rpcrdma_reps_destroy(buf);
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);
430 rpcrdma_mrs_destroy(r_xprt);
431 ib_dealloc_pd(ia->ri_pd);
434 /* Allow waiters to continue */
435 complete(&ia->ri_remove_done);
437 trace_xprtrdma_remove(r_xprt);
441 * rpcrdma_ia_close - Clean up/close an IA.
442 * @ia: interface adapter to close
446 rpcrdma_ia_close(struct rpcrdma_ia *ia)
448 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
450 rdma_destroy_qp(ia->ri_id);
451 rdma_destroy_id(ia->ri_id);
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);
462 * rpcrdma_ep_create - Create unconnected endpoint
463 * @r_xprt: transport to instantiate
465 * Returns zero on success, or a negative errno.
467 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
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;
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;
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);
486 ia->ri_max_send_sges = max_sge;
488 rc = frwr_open(ia, ep);
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;
502 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
503 "iovs: send %d recv %d\n",
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);
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;
515 sendcq = ib_alloc_cq_any(ia->ri_id->device, r_xprt,
516 ep->rep_attr.cap.max_send_wr + 1,
518 if (IS_ERR(sendcq)) {
519 rc = PTR_ERR(sendcq);
523 recvcq = ib_alloc_cq_any(ia->ri_id->device, NULL,
524 ep->rep_attr.cap.max_recv_wr + 1,
526 if (IS_ERR(recvcq)) {
527 rc = PTR_ERR(recvcq);
531 ep->rep_attr.send_cq = sendcq;
532 ep->rep_attr.recv_cq = recvcq;
534 /* Initialize cma parameters */
535 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
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);
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);
551 /* Limit transport retries so client can detect server
552 * GID changes quickly. RPC layer handles re-establishing
553 * transport connection and retransmission.
555 ep->rep_remote_cma.retry_count = 6;
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).
561 ep->rep_remote_cma.flow_control = 0;
562 ep->rep_remote_cma.rnr_retry_count = 0;
573 * rpcrdma_ep_destroy - Disconnect and destroy endpoint.
574 * @r_xprt: transport instance to shut down
577 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt)
579 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
580 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
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;
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);
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.
598 static int rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
599 struct ib_qp_init_attr *qp_init_attr)
601 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
602 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
605 trace_xprtrdma_reinsert(r_xprt);
608 if (rpcrdma_ia_open(r_xprt))
612 err = rpcrdma_ep_create(r_xprt);
614 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
617 memcpy(qp_init_attr, &ep->rep_attr, sizeof(*qp_init_attr));
620 err = rdma_create_qp(ia->ri_id, ia->ri_pd, qp_init_attr);
622 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
628 rpcrdma_ep_destroy(r_xprt);
630 rpcrdma_ia_close(ia);
635 static int rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt,
636 struct ib_qp_init_attr *qp_init_attr)
638 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
639 struct rdma_cm_id *id, *old;
642 rpcrdma_ep_disconnect(&r_xprt->rx_ep, ia);
645 id = rpcrdma_create_id(r_xprt, ia);
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.
654 * This is a sanity check only. There should be no way these
655 * point to two different devices here.
659 if (ia->ri_id->device != id->device) {
660 pr_err("rpcrdma: can't reconnect on different device!\n");
664 err = rdma_create_qp(id, ia->ri_pd, qp_init_attr);
668 /* Atomically replace the transport's ID and QP. */
672 rdma_destroy_qp(old);
675 rdma_destroy_id(old);
681 * Connect unconnected endpoint.
684 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
686 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
688 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
689 struct ib_qp_init_attr qp_init_attr;
693 memcpy(&qp_init_attr, &ep->rep_attr, sizeof(qp_init_attr));
694 switch (ep->rep_connected) {
696 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &qp_init_attr);
703 rc = rpcrdma_ep_recreate_xprt(r_xprt, &qp_init_attr);
708 rc = rpcrdma_ep_reconnect(r_xprt, &qp_init_attr);
713 ep->rep_connected = 0;
714 xprt_clear_connected(xprt);
716 rpcrdma_reset_cwnd(r_xprt);
717 rpcrdma_post_recvs(r_xprt, true);
719 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
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)
729 rc = ep->rep_connected;
733 rpcrdma_mrs_create(r_xprt);
737 ep->rep_connected = rc;
740 trace_xprtrdma_connect(r_xprt, rc);
745 * rpcrdma_ep_disconnect - Disconnect underlying transport
746 * @ep: endpoint to disconnect
747 * @ia: associated interface adapter
749 * Caller serializes. Either the transport send lock is held,
750 * or we're being called to destroy the transport.
753 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
755 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
759 /* returns without wait if ID is not connected */
760 rc = rdma_disconnect(ia->ri_id);
762 wait_event_interruptible(ep->rep_connect_wait,
763 ep->rep_connected != 1);
765 ep->rep_connected = rc;
766 trace_xprtrdma_disconnect(r_xprt, rc);
768 rpcrdma_xprt_drain(r_xprt);
769 rpcrdma_reqs_reset(r_xprt);
770 rpcrdma_mrs_destroy(r_xprt);
773 /* Fixed-size circular FIFO queue. This implementation is wait-free and
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.
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.
786 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
787 * queue activity, and rpcrdma_xprt_drain has flushed all remaining
790 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
794 for (i = 0; i <= buf->rb_sc_last; i++)
795 kfree(buf->rb_sc_ctxs[i]);
796 kfree(buf->rb_sc_ctxs);
799 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
801 struct rpcrdma_sendctx *sc;
803 sc = kzalloc(struct_size(sc, sc_sges, ia->ri_max_send_sges),
808 sc->sc_cqe.done = rpcrdma_wc_send;
812 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
814 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
815 struct rpcrdma_sendctx *sc;
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
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)
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);
835 buf->rb_sc_ctxs[i] = sc;
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.
845 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
848 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
852 * rpcrdma_sendctx_get_locked - Acquire a send context
853 * @r_xprt: controlling transport instance
855 * Returns pointer to a free send completion context; or NULL if
856 * the queue is empty.
858 * Usage: Called to acquire an SGE array before preparing a Send WR.
860 * The caller serializes calls to this function (per transport), and
861 * provides an effective memory barrier that flushes the new value
864 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt)
866 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
867 struct rpcrdma_sendctx *sc;
868 unsigned long next_head;
870 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
872 if (next_head == READ_ONCE(buf->rb_sc_tail))
875 /* ORDER: item must be accessed _before_ head is updated */
876 sc = buf->rb_sc_ctxs[next_head];
878 /* Releasing the lock in the caller acts as a memory
879 * barrier that flushes rb_sc_head.
881 buf->rb_sc_head = next_head;
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.
890 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
891 r_xprt->rx_stats.empty_sendctx_q++;
896 * rpcrdma_sendctx_put_locked - Release a send context
897 * @r_xprt: controlling transport instance
898 * @sc: send context to release
900 * Usage: Called from Send completion to return a sendctxt
903 * The caller serializes calls to this function (per transport).
905 static void rpcrdma_sendctx_put_locked(struct rpcrdma_xprt *r_xprt,
906 struct rpcrdma_sendctx *sc)
908 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
909 unsigned long next_tail;
911 /* Unmap SGEs of previously completed but unsignaled
912 * Sends by walking up the queue until @sc is found.
914 next_tail = buf->rb_sc_tail;
916 next_tail = rpcrdma_sendctx_next(buf, next_tail);
918 /* ORDER: item must be accessed _before_ tail is updated */
919 rpcrdma_sendctx_unmap(buf->rb_sc_ctxs[next_tail]);
921 } while (buf->rb_sc_ctxs[next_tail] != sc);
923 /* Paired with READ_ONCE */
924 smp_store_release(&buf->rb_sc_tail, next_tail);
926 xprt_write_space(&r_xprt->rx_xprt);
930 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
932 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
933 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
936 for (count = 0; count < ia->ri_max_segs; count++) {
937 struct rpcrdma_mr *mr;
940 mr = kzalloc(sizeof(*mr), GFP_NOFS);
944 rc = frwr_init_mr(ia, mr);
950 mr->mr_xprt = r_xprt;
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);
958 r_xprt->rx_stats.mrs_allocated += count;
959 trace_xprtrdma_createmrs(r_xprt, count);
963 rpcrdma_mr_refresh_worker(struct work_struct *work)
965 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
967 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
970 rpcrdma_mrs_create(r_xprt);
971 xprt_write_space(&r_xprt->rx_xprt);
975 * rpcrdma_mrs_refresh - Wake the MR refresh worker
976 * @r_xprt: controlling transport instance
979 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt)
981 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
982 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
984 /* If there is no underlying device, it's no use to
985 * wake the refresh worker.
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.
992 queue_work(xprtiod_workqueue, &buf->rb_refresh_worker);
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
1002 * Returns an allocated and fully initialized rpcrdma_req or NULL.
1004 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
1007 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1008 struct rpcrdma_regbuf *rb;
1009 struct rpcrdma_req *req;
1012 req = kzalloc(sizeof(*req), flags);
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);
1024 req->rl_rdmabuf = rb;
1025 xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
1027 req->rl_sendbuf = rpcrdma_regbuf_alloc(size, DMA_TO_DEVICE, flags);
1028 if (!req->rl_sendbuf)
1031 req->rl_recvbuf = rpcrdma_regbuf_alloc(size, DMA_NONE, flags);
1032 if (!req->rl_recvbuf)
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);
1043 kfree(req->rl_sendbuf);
1045 kfree(req->rl_rdmabuf);
1053 * rpcrdma_reqs_reset - Reset all reqs owned by a transport
1054 * @r_xprt: controlling transport instance
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.
1061 static void rpcrdma_reqs_reset(struct rpcrdma_xprt *r_xprt)
1063 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1064 struct rpcrdma_req *req;
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;
1072 static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
1075 struct rpcrdma_rep *rep;
1077 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
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)
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;
1103 static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
1105 rpcrdma_regbuf_free(rep->rr_rdmabuf);
1109 static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
1111 struct llist_node *node;
1113 /* Calls to llist_del_first are required to be serialized */
1114 node = llist_del_first(&buf->rb_free_reps);
1117 return llist_entry(node, struct rpcrdma_rep, rr_node);
1120 static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
1121 struct rpcrdma_rep *rep)
1124 llist_add(&rep->rr_node, &buf->rb_free_reps);
1126 rpcrdma_rep_destroy(rep);
1129 static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
1131 struct rpcrdma_rep *rep;
1133 while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
1134 rpcrdma_rep_destroy(rep);
1138 * rpcrdma_buffer_create - Create initial set of req/rep objects
1139 * @r_xprt: transport instance to (re)initialize
1141 * Returns zero on success, otherwise a negative errno.
1143 int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1145 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1148 buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
1149 buf->rb_bc_srv_max_requests = 0;
1150 spin_lock_init(&buf->rb_lock);
1151 INIT_LIST_HEAD(&buf->rb_mrs);
1152 INIT_LIST_HEAD(&buf->rb_all_mrs);
1153 INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
1155 INIT_LIST_HEAD(&buf->rb_send_bufs);
1156 INIT_LIST_HEAD(&buf->rb_allreqs);
1159 for (i = 0; i < buf->rb_max_requests; i++) {
1160 struct rpcrdma_req *req;
1162 req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
1166 list_add(&req->rl_list, &buf->rb_send_bufs);
1169 init_llist_head(&buf->rb_free_reps);
1171 rc = rpcrdma_sendctxs_create(r_xprt);
1177 rpcrdma_buffer_destroy(buf);
1182 * rpcrdma_req_destroy - Destroy an rpcrdma_req object
1183 * @req: unused object to be destroyed
1185 * Relies on caller holding the transport send lock to protect
1186 * removing req->rl_all from buf->rb_all_reqs safely.
1188 void rpcrdma_req_destroy(struct rpcrdma_req *req)
1190 struct rpcrdma_mr *mr;
1192 list_del(&req->rl_all);
1194 while ((mr = rpcrdma_mr_pop(&req->rl_free_mrs))) {
1195 struct rpcrdma_buffer *buf = &mr->mr_xprt->rx_buf;
1197 spin_lock(&buf->rb_lock);
1198 list_del(&mr->mr_all);
1199 spin_unlock(&buf->rb_lock);
1201 frwr_release_mr(mr);
1204 rpcrdma_regbuf_free(req->rl_recvbuf);
1205 rpcrdma_regbuf_free(req->rl_sendbuf);
1206 rpcrdma_regbuf_free(req->rl_rdmabuf);
1211 * rpcrdma_mrs_destroy - Release all of a transport's MRs
1212 * @r_xprt: controlling transport instance
1214 * Relies on caller holding the transport send lock to protect
1215 * removing mr->mr_list from req->rl_free_mrs safely.
1217 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt)
1219 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1220 struct rpcrdma_mr *mr;
1222 cancel_work_sync(&buf->rb_refresh_worker);
1224 spin_lock(&buf->rb_lock);
1225 while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
1228 list_del(&mr->mr_list);
1229 list_del(&mr->mr_all);
1230 spin_unlock(&buf->rb_lock);
1232 frwr_release_mr(mr);
1234 spin_lock(&buf->rb_lock);
1236 spin_unlock(&buf->rb_lock);
1240 * rpcrdma_buffer_destroy - Release all hw resources
1241 * @buf: root control block for resources
1243 * ORDERING: relies on a prior rpcrdma_xprt_drain :
1244 * - No more Send or Receive completions can occur
1245 * - All MRs, reps, and reqs are returned to their free lists
1248 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1250 rpcrdma_sendctxs_destroy(buf);
1251 rpcrdma_reps_destroy(buf);
1253 while (!list_empty(&buf->rb_send_bufs)) {
1254 struct rpcrdma_req *req;
1256 req = list_first_entry(&buf->rb_send_bufs,
1257 struct rpcrdma_req, rl_list);
1258 list_del(&req->rl_list);
1259 rpcrdma_req_destroy(req);
1264 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1265 * @r_xprt: controlling transport
1267 * Returns an initialized rpcrdma_mr or NULL if no free
1268 * rpcrdma_mr objects are available.
1271 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1273 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1274 struct rpcrdma_mr *mr;
1276 spin_lock(&buf->rb_lock);
1277 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1278 spin_unlock(&buf->rb_lock);
1283 * rpcrdma_mr_put - DMA unmap an MR and release it
1284 * @mr: MR to release
1287 void rpcrdma_mr_put(struct rpcrdma_mr *mr)
1289 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1291 if (mr->mr_dir != DMA_NONE) {
1292 trace_xprtrdma_mr_unmap(mr);
1293 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
1294 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1295 mr->mr_dir = DMA_NONE;
1298 rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
1302 * rpcrdma_buffer_get - Get a request buffer
1303 * @buffers: Buffer pool from which to obtain a buffer
1305 * Returns a fresh rpcrdma_req, or NULL if none are available.
1307 struct rpcrdma_req *
1308 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1310 struct rpcrdma_req *req;
1312 spin_lock(&buffers->rb_lock);
1313 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1314 struct rpcrdma_req, rl_list);
1316 list_del_init(&req->rl_list);
1317 spin_unlock(&buffers->rb_lock);
1322 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1323 * @buffers: buffer pool
1324 * @req: object to return
1327 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
1330 rpcrdma_rep_put(buffers, req->rl_reply);
1331 req->rl_reply = NULL;
1333 spin_lock(&buffers->rb_lock);
1334 list_add(&req->rl_list, &buffers->rb_send_bufs);
1335 spin_unlock(&buffers->rb_lock);
1339 * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
1340 * @rep: rep to release
1342 * Used after error conditions.
1344 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1346 rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
1349 /* Returns a pointer to a rpcrdma_regbuf object, or NULL.
1351 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1352 * receiving the payload of RDMA RECV operations. During Long Calls
1353 * or Replies they may be registered externally via frwr_map.
1355 static struct rpcrdma_regbuf *
1356 rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
1359 struct rpcrdma_regbuf *rb;
1361 rb = kmalloc(sizeof(*rb), flags);
1364 rb->rg_data = kmalloc(size, flags);
1370 rb->rg_device = NULL;
1371 rb->rg_direction = direction;
1372 rb->rg_iov.length = size;
1377 * rpcrdma_regbuf_realloc - re-allocate a SEND/RECV buffer
1378 * @rb: regbuf to reallocate
1379 * @size: size of buffer to be allocated, in bytes
1382 * Returns true if reallocation was successful. If false is
1383 * returned, @rb is left untouched.
1385 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size, gfp_t flags)
1389 buf = kmalloc(size, flags);
1393 rpcrdma_regbuf_dma_unmap(rb);
1397 rb->rg_iov.length = size;
1402 * __rpcrdma_regbuf_dma_map - DMA-map a regbuf
1403 * @r_xprt: controlling transport instance
1404 * @rb: regbuf to be mapped
1406 * Returns true if the buffer is now DMA mapped to @r_xprt's device
1408 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
1409 struct rpcrdma_regbuf *rb)
1411 struct ib_device *device = r_xprt->rx_ia.ri_id->device;
1413 if (rb->rg_direction == DMA_NONE)
1416 rb->rg_iov.addr = ib_dma_map_single(device, rdmab_data(rb),
1417 rdmab_length(rb), rb->rg_direction);
1418 if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
1419 trace_xprtrdma_dma_maperr(rdmab_addr(rb));
1423 rb->rg_device = device;
1424 rb->rg_iov.lkey = r_xprt->rx_ia.ri_pd->local_dma_lkey;
1428 static void rpcrdma_regbuf_dma_unmap(struct rpcrdma_regbuf *rb)
1433 if (!rpcrdma_regbuf_is_mapped(rb))
1436 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), rdmab_length(rb),
1438 rb->rg_device = NULL;
1441 static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
1443 rpcrdma_regbuf_dma_unmap(rb);
1450 * rpcrdma_ep_post - Post WRs to a transport's Send Queue
1451 * @ia: transport's device information
1452 * @ep: transport's RDMA endpoint information
1453 * @req: rpcrdma_req containing the Send WR to post
1455 * Returns 0 if the post was successful, otherwise -ENOTCONN
1459 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1460 struct rpcrdma_ep *ep,
1461 struct rpcrdma_req *req)
1463 struct ib_send_wr *send_wr = &req->rl_wr;
1466 if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
1467 send_wr->send_flags |= IB_SEND_SIGNALED;
1468 ep->rep_send_count = ep->rep_send_batch;
1470 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1471 --ep->rep_send_count;
1474 rc = frwr_send(ia, req);
1475 trace_xprtrdma_post_send(req, rc);
1482 * rpcrdma_post_recvs - Refill the Receive Queue
1483 * @r_xprt: controlling transport instance
1484 * @temp: mark Receive buffers to be deleted after use
1487 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1489 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1490 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1491 struct ib_recv_wr *i, *wr, *bad_wr;
1492 struct rpcrdma_rep *rep;
1493 int needed, count, rc;
1498 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1499 if (likely(ep->rep_receive_count > needed))
1501 needed -= ep->rep_receive_count;
1503 needed += RPCRDMA_MAX_RECV_BATCH;
1505 /* fast path: all needed reps can be found on the free list */
1508 rep = rpcrdma_rep_get_locked(buf);
1510 rep = rpcrdma_rep_create(r_xprt, temp);
1514 rep->rr_recv_wr.next = wr;
1515 wr = &rep->rr_recv_wr;
1521 for (i = wr; i; i = i->next) {
1522 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1524 if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
1527 trace_xprtrdma_post_recv(rep);
1531 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1532 (const struct ib_recv_wr **)&bad_wr);
1534 trace_xprtrdma_post_recvs(r_xprt, count, rc);
1536 for (wr = bad_wr; wr;) {
1537 struct rpcrdma_rep *rep;
1539 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1541 rpcrdma_recv_buffer_put(rep);
1545 ep->rep_receive_count += count;
1550 rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
1552 rpcrdma_recv_buffer_put(rep);