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>
57 #include <asm-generic/barrier.h>
58 #include <asm/bitops.h>
60 #include <rdma/ib_cm.h>
62 #include "xprt_rdma.h"
63 #include <trace/events/rpcrdma.h>
69 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
70 # define RPCDBG_FACILITY RPCDBG_TRANS
76 static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
77 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
78 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
79 static int rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp);
80 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
81 static void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
83 /* Wait for outstanding transport work to finish.
85 static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
87 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
88 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
90 /* Flush Receives, then wait for deferred Reply work
93 ib_drain_qp(ia->ri_id->qp);
94 drain_workqueue(buf->rb_completion_wq);
96 /* Deferred Reply processing might have scheduled
97 * local invalidations.
99 ib_drain_sq(ia->ri_id->qp);
103 * rpcrdma_qp_event_handler - Handle one QP event (error notification)
104 * @event: details of the event
105 * @context: ep that owns QP where event occurred
107 * Called from the RDMA provider (device driver) possibly in an interrupt
111 rpcrdma_qp_event_handler(struct ib_event *event, void *context)
113 struct rpcrdma_ep *ep = context;
114 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
117 trace_xprtrdma_qp_event(r_xprt, event);
121 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
122 * @cq: completion queue (ignored)
127 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
129 struct ib_cqe *cqe = wc->wr_cqe;
130 struct rpcrdma_sendctx *sc =
131 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
133 /* WARNING: Only wr_cqe and status are reliable at this point */
134 trace_xprtrdma_wc_send(sc, wc);
135 if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
136 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
137 ib_wc_status_msg(wc->status),
138 wc->status, wc->vendor_err);
140 rpcrdma_sendctx_put_locked(sc);
144 * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
145 * @cq: completion queue (ignored)
150 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
152 struct ib_cqe *cqe = wc->wr_cqe;
153 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
155 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
157 /* WARNING: Only wr_cqe and status are reliable at this point */
158 trace_xprtrdma_wc_receive(wc);
159 --r_xprt->rx_ep.rep_receive_count;
160 if (wc->status != IB_WC_SUCCESS)
163 /* status == SUCCESS means all fields in wc are trustworthy */
164 rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
165 rep->rr_wc_flags = wc->wc_flags;
166 rep->rr_inv_rkey = wc->ex.invalidate_rkey;
168 ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
169 rdmab_addr(rep->rr_rdmabuf),
170 wc->byte_len, DMA_FROM_DEVICE);
172 rpcrdma_post_recvs(r_xprt, false);
173 rpcrdma_reply_handler(rep);
177 if (wc->status != IB_WC_WR_FLUSH_ERR)
178 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
179 ib_wc_status_msg(wc->status),
180 wc->status, wc->vendor_err);
181 rpcrdma_recv_buffer_put(rep);
185 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
186 struct rdma_conn_param *param)
188 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
189 const struct rpcrdma_connect_private *pmsg = param->private_data;
190 unsigned int rsize, wsize;
192 /* Default settings for RPC-over-RDMA Version One */
193 r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
194 rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
195 wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
198 pmsg->cp_magic == rpcrdma_cmp_magic &&
199 pmsg->cp_version == RPCRDMA_CMP_VERSION) {
200 r_xprt->rx_ia.ri_implicit_roundup = true;
201 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
202 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
205 if (rsize < cdata->inline_rsize)
206 cdata->inline_rsize = rsize;
207 if (wsize < cdata->inline_wsize)
208 cdata->inline_wsize = wsize;
209 dprintk("RPC: %s: max send %u, max recv %u\n",
210 __func__, cdata->inline_wsize, cdata->inline_rsize);
211 rpcrdma_set_max_header_sizes(r_xprt);
215 * rpcrdma_cm_event_handler - Handle RDMA CM events
216 * @id: rdma_cm_id on which an event has occurred
217 * @event: details of the event
219 * Called with @id's mutex held. Returns 1 if caller should
220 * destroy @id, otherwise 0.
223 rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
225 struct rpcrdma_xprt *r_xprt = id->context;
226 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
227 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
228 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
232 trace_xprtrdma_cm_event(r_xprt, event);
233 switch (event->event) {
234 case RDMA_CM_EVENT_ADDR_RESOLVED:
235 case RDMA_CM_EVENT_ROUTE_RESOLVED:
237 complete(&ia->ri_done);
239 case RDMA_CM_EVENT_ADDR_ERROR:
240 ia->ri_async_rc = -EPROTO;
241 complete(&ia->ri_done);
243 case RDMA_CM_EVENT_ROUTE_ERROR:
244 ia->ri_async_rc = -ENETUNREACH;
245 complete(&ia->ri_done);
247 case RDMA_CM_EVENT_DEVICE_REMOVAL:
248 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
249 pr_info("rpcrdma: removing device %s for %s:%s\n",
251 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt));
253 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
254 ep->rep_connected = -ENODEV;
255 xprt_force_disconnect(xprt);
256 wait_for_completion(&ia->ri_remove_done);
259 ia->ri_device = NULL;
260 /* Return 1 to ensure the core destroys the id. */
262 case RDMA_CM_EVENT_ESTABLISHED:
263 ++xprt->connect_cookie;
264 ep->rep_connected = 1;
265 rpcrdma_update_connect_private(r_xprt, &event->param.conn);
266 wake_up_all(&ep->rep_connect_wait);
268 case RDMA_CM_EVENT_CONNECT_ERROR:
269 ep->rep_connected = -ENOTCONN;
271 case RDMA_CM_EVENT_UNREACHABLE:
272 ep->rep_connected = -ENETUNREACH;
274 case RDMA_CM_EVENT_REJECTED:
275 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
276 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
277 rdma_reject_msg(id, event->status));
278 ep->rep_connected = -ECONNREFUSED;
279 if (event->status == IB_CM_REJ_STALE_CONN)
280 ep->rep_connected = -EAGAIN;
282 case RDMA_CM_EVENT_DISCONNECTED:
283 ep->rep_connected = -ECONNABORTED;
285 xprt_force_disconnect(xprt);
286 wake_up_all(&ep->rep_connect_wait);
292 dprintk("RPC: %s: %s:%s on %s/frwr: %s\n", __func__,
293 rpcrdma_addrstr(r_xprt), rpcrdma_portstr(r_xprt),
294 ia->ri_device->name, rdma_event_msg(event->event));
298 static struct rdma_cm_id *
299 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
301 unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
302 struct rdma_cm_id *id;
305 trace_xprtrdma_conn_start(xprt);
307 init_completion(&ia->ri_done);
308 init_completion(&ia->ri_remove_done);
310 id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_cm_event_handler,
311 xprt, RDMA_PS_TCP, IB_QPT_RC);
314 dprintk("RPC: %s: rdma_create_id() failed %i\n",
319 ia->ri_async_rc = -ETIMEDOUT;
320 rc = rdma_resolve_addr(id, NULL,
321 (struct sockaddr *)&xprt->rx_xprt.addr,
322 RDMA_RESOLVE_TIMEOUT);
324 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
328 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
330 trace_xprtrdma_conn_tout(xprt);
334 rc = ia->ri_async_rc;
338 ia->ri_async_rc = -ETIMEDOUT;
339 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
341 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
345 rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
347 trace_xprtrdma_conn_tout(xprt);
350 rc = ia->ri_async_rc;
362 * Exported functions.
366 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
367 * @xprt: transport with IA to (re)initialize
369 * Returns 0 on success, negative errno if an appropriate
370 * Interface Adapter could not be found and opened.
373 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
375 struct rpcrdma_ia *ia = &xprt->rx_ia;
378 ia->ri_id = rpcrdma_create_id(xprt, ia);
379 if (IS_ERR(ia->ri_id)) {
380 rc = PTR_ERR(ia->ri_id);
383 ia->ri_device = ia->ri_id->device;
385 ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
386 if (IS_ERR(ia->ri_pd)) {
387 rc = PTR_ERR(ia->ri_pd);
388 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
392 switch (xprt_rdma_memreg_strategy) {
394 if (frwr_is_supported(ia))
398 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
399 ia->ri_device->name, xprt_rdma_memreg_strategy);
407 rpcrdma_ia_close(ia);
412 * rpcrdma_ia_remove - Handle device driver unload
413 * @ia: interface adapter being removed
415 * Divest transport H/W resources associated with this adapter,
416 * but allow it to be restored later.
419 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
421 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
423 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
424 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
425 struct rpcrdma_req *req;
426 struct rpcrdma_rep *rep;
428 cancel_delayed_work_sync(&buf->rb_refresh_worker);
430 /* This is similar to rpcrdma_ep_destroy, but:
431 * - Don't cancel the connect worker.
432 * - Don't call rpcrdma_ep_disconnect, which waits
433 * for another conn upcall, which will deadlock.
434 * - rdma_disconnect is unneeded, the underlying
435 * connection is already gone.
438 rpcrdma_xprt_drain(r_xprt);
439 rdma_destroy_qp(ia->ri_id);
440 ia->ri_id->qp = NULL;
442 ib_free_cq(ep->rep_attr.recv_cq);
443 ep->rep_attr.recv_cq = NULL;
444 ib_free_cq(ep->rep_attr.send_cq);
445 ep->rep_attr.send_cq = NULL;
447 /* The ULP is responsible for ensuring all DMA
448 * mappings and MRs are gone.
450 list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
451 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
452 list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
453 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
454 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
455 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
457 rpcrdma_mrs_destroy(buf);
458 ib_dealloc_pd(ia->ri_pd);
461 /* Allow waiters to continue */
462 complete(&ia->ri_remove_done);
464 trace_xprtrdma_remove(r_xprt);
468 * rpcrdma_ia_close - Clean up/close an IA.
469 * @ia: interface adapter to close
473 rpcrdma_ia_close(struct rpcrdma_ia *ia)
475 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
477 rdma_destroy_qp(ia->ri_id);
478 rdma_destroy_id(ia->ri_id);
481 ia->ri_device = NULL;
483 /* If the pd is still busy, xprtrdma missed freeing a resource */
484 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
485 ib_dealloc_pd(ia->ri_pd);
490 * Create unconnected endpoint.
493 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
494 struct rpcrdma_create_data_internal *cdata)
496 struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
497 struct ib_cq *sendcq, *recvcq;
498 unsigned int max_sge;
501 max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
502 RPCRDMA_MAX_SEND_SGES);
503 if (max_sge < RPCRDMA_MIN_SEND_SGES) {
504 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
507 ia->ri_max_send_sges = max_sge;
509 rc = frwr_open(ia, ep, cdata);
513 ep->rep_attr.event_handler = rpcrdma_qp_event_handler;
514 ep->rep_attr.qp_context = ep;
515 ep->rep_attr.srq = NULL;
516 ep->rep_attr.cap.max_send_sge = max_sge;
517 ep->rep_attr.cap.max_recv_sge = 1;
518 ep->rep_attr.cap.max_inline_data = 0;
519 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
520 ep->rep_attr.qp_type = IB_QPT_RC;
521 ep->rep_attr.port_num = ~0;
523 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
524 "iovs: send %d recv %d\n",
526 ep->rep_attr.cap.max_send_wr,
527 ep->rep_attr.cap.max_recv_wr,
528 ep->rep_attr.cap.max_send_sge,
529 ep->rep_attr.cap.max_recv_sge);
531 /* set trigger for requesting send completion */
532 ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
533 cdata->max_requests >> 2);
534 ep->rep_send_count = ep->rep_send_batch;
535 init_waitqueue_head(&ep->rep_connect_wait);
536 ep->rep_receive_count = 0;
538 sendcq = ib_alloc_cq(ia->ri_device, NULL,
539 ep->rep_attr.cap.max_send_wr + 1,
540 1, IB_POLL_WORKQUEUE);
541 if (IS_ERR(sendcq)) {
542 rc = PTR_ERR(sendcq);
543 dprintk("RPC: %s: failed to create send CQ: %i\n",
548 recvcq = ib_alloc_cq(ia->ri_device, NULL,
549 ep->rep_attr.cap.max_recv_wr + 1,
550 0, IB_POLL_WORKQUEUE);
551 if (IS_ERR(recvcq)) {
552 rc = PTR_ERR(recvcq);
553 dprintk("RPC: %s: failed to create recv CQ: %i\n",
558 ep->rep_attr.send_cq = sendcq;
559 ep->rep_attr.recv_cq = recvcq;
561 /* Initialize cma parameters */
562 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
564 /* Prepare RDMA-CM private message */
565 pmsg->cp_magic = rpcrdma_cmp_magic;
566 pmsg->cp_version = RPCRDMA_CMP_VERSION;
567 pmsg->cp_flags |= RPCRDMA_CMP_F_SND_W_INV_OK;
568 pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
569 pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
570 ep->rep_remote_cma.private_data = pmsg;
571 ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
573 /* Client offers RDMA Read but does not initiate */
574 ep->rep_remote_cma.initiator_depth = 0;
575 ep->rep_remote_cma.responder_resources =
576 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
578 /* Limit transport retries so client can detect server
579 * GID changes quickly. RPC layer handles re-establishing
580 * transport connection and retransmission.
582 ep->rep_remote_cma.retry_count = 6;
584 /* RPC-over-RDMA handles its own flow control. In addition,
585 * make all RNR NAKs visible so we know that RPC-over-RDMA
586 * flow control is working correctly (no NAKs should be seen).
588 ep->rep_remote_cma.flow_control = 0;
589 ep->rep_remote_cma.rnr_retry_count = 0;
602 * Disconnect and destroy endpoint. After this, the only
603 * valid operations on the ep are to free it (if dynamically
604 * allocated) or re-create it.
607 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
609 if (ia->ri_id && ia->ri_id->qp) {
610 rpcrdma_ep_disconnect(ep, ia);
611 rdma_destroy_qp(ia->ri_id);
612 ia->ri_id->qp = NULL;
615 if (ep->rep_attr.recv_cq)
616 ib_free_cq(ep->rep_attr.recv_cq);
617 if (ep->rep_attr.send_cq)
618 ib_free_cq(ep->rep_attr.send_cq);
621 /* Re-establish a connection after a device removal event.
622 * Unlike a normal reconnection, a fresh PD and a new set
623 * of MRs and buffers is needed.
626 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
627 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
631 trace_xprtrdma_reinsert(r_xprt);
634 if (rpcrdma_ia_open(r_xprt))
638 err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
640 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
645 err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
647 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
651 rpcrdma_mrs_create(r_xprt);
655 rpcrdma_ep_destroy(ep, ia);
657 rpcrdma_ia_close(ia);
663 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
664 struct rpcrdma_ia *ia)
666 struct rdma_cm_id *id, *old;
669 trace_xprtrdma_reconnect(r_xprt);
671 rpcrdma_ep_disconnect(ep, ia);
674 id = rpcrdma_create_id(r_xprt, ia);
678 /* As long as the new ID points to the same device as the
679 * old ID, we can reuse the transport's existing PD and all
680 * previously allocated MRs. Also, the same device means
681 * the transport's previous DMA mappings are still valid.
683 * This is a sanity check only. There should be no way these
684 * point to two different devices here.
688 if (ia->ri_device != id->device) {
689 pr_err("rpcrdma: can't reconnect on different device!\n");
693 err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
695 dprintk("RPC: %s: rdma_create_qp returned %d\n",
700 /* Atomically replace the transport's ID and QP. */
704 rdma_destroy_qp(old);
707 rdma_destroy_id(old);
713 * Connect unconnected endpoint.
716 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
718 struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
720 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
724 switch (ep->rep_connected) {
726 dprintk("RPC: %s: connecting...\n", __func__);
727 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
729 dprintk("RPC: %s: rdma_create_qp failed %i\n",
736 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
741 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
746 ep->rep_connected = 0;
747 xprt_clear_connected(xprt);
749 rpcrdma_post_recvs(r_xprt, true);
751 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
753 dprintk("RPC: %s: rdma_connect() failed with %i\n",
758 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
759 if (ep->rep_connected <= 0) {
760 if (ep->rep_connected == -EAGAIN)
762 rc = ep->rep_connected;
766 dprintk("RPC: %s: connected\n", __func__);
770 ep->rep_connected = rc;
777 * rpcrdma_ep_disconnect - Disconnect underlying transport
778 * @ep: endpoint to disconnect
779 * @ia: associated interface adapter
781 * This is separate from destroy to facilitate the ability
782 * to reconnect without recreating the endpoint.
784 * This call is not reentrant, and must not be made in parallel
785 * on the same endpoint.
788 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
790 struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
794 /* returns without wait if ID is not connected */
795 rc = rdma_disconnect(ia->ri_id);
797 wait_event_interruptible(ep->rep_connect_wait,
798 ep->rep_connected != 1);
800 ep->rep_connected = rc;
801 trace_xprtrdma_disconnect(r_xprt, rc);
803 rpcrdma_xprt_drain(r_xprt);
806 /* Fixed-size circular FIFO queue. This implementation is wait-free and
809 * Consumer is the code path that posts Sends. This path dequeues a
810 * sendctx for use by a Send operation. Multiple consumer threads
811 * are serialized by the RPC transport lock, which allows only one
812 * ->send_request call at a time.
814 * Producer is the code path that handles Send completions. This path
815 * enqueues a sendctx that has been completed. Multiple producer
816 * threads are serialized by the ib_poll_cq() function.
819 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
820 * queue activity, and ib_drain_qp has flushed all remaining Send
823 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
827 for (i = 0; i <= buf->rb_sc_last; i++)
828 kfree(buf->rb_sc_ctxs[i]);
829 kfree(buf->rb_sc_ctxs);
832 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
834 struct rpcrdma_sendctx *sc;
836 sc = kzalloc(sizeof(*sc) +
837 ia->ri_max_send_sges * sizeof(struct ib_sge),
842 sc->sc_wr.wr_cqe = &sc->sc_cqe;
843 sc->sc_wr.sg_list = sc->sc_sges;
844 sc->sc_wr.opcode = IB_WR_SEND;
845 sc->sc_cqe.done = rpcrdma_wc_send;
849 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
851 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
852 struct rpcrdma_sendctx *sc;
855 /* Maximum number of concurrent outstanding Send WRs. Capping
856 * the circular queue size stops Send Queue overflow by causing
857 * the ->send_request call to fail temporarily before too many
860 i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
861 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__, i);
862 buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
863 if (!buf->rb_sc_ctxs)
866 buf->rb_sc_last = i - 1;
867 for (i = 0; i <= buf->rb_sc_last; i++) {
868 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
872 sc->sc_xprt = r_xprt;
873 buf->rb_sc_ctxs[i] = sc;
879 rpcrdma_sendctxs_destroy(buf);
883 /* The sendctx queue is not guaranteed to have a size that is a
884 * power of two, thus the helpers in circ_buf.h cannot be used.
885 * The other option is to use modulus (%), which can be expensive.
887 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
890 return likely(item < buf->rb_sc_last) ? item + 1 : 0;
894 * rpcrdma_sendctx_get_locked - Acquire a send context
895 * @buf: transport buffers from which to acquire an unused context
897 * Returns pointer to a free send completion context; or NULL if
898 * the queue is empty.
900 * Usage: Called to acquire an SGE array before preparing a Send WR.
902 * The caller serializes calls to this function (per rpcrdma_buffer),
903 * and provides an effective memory barrier that flushes the new value
906 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
908 struct rpcrdma_xprt *r_xprt;
909 struct rpcrdma_sendctx *sc;
910 unsigned long next_head;
912 next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
914 if (next_head == READ_ONCE(buf->rb_sc_tail))
917 /* ORDER: item must be accessed _before_ head is updated */
918 sc = buf->rb_sc_ctxs[next_head];
920 /* Releasing the lock in the caller acts as a memory
921 * barrier that flushes rb_sc_head.
923 buf->rb_sc_head = next_head;
928 /* The queue is "empty" if there have not been enough Send
929 * completions recently. This is a sign the Send Queue is
930 * backing up. Cause the caller to pause and try again.
932 set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
933 r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
934 r_xprt->rx_stats.empty_sendctx_q++;
939 * rpcrdma_sendctx_put_locked - Release a send context
940 * @sc: send context to release
942 * Usage: Called from Send completion to return a sendctxt
945 * The caller serializes calls to this function (per rpcrdma_buffer).
948 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
950 struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
951 unsigned long next_tail;
953 /* Unmap SGEs of previously completed by unsignaled
954 * Sends by walking up the queue until @sc is found.
956 next_tail = buf->rb_sc_tail;
958 next_tail = rpcrdma_sendctx_next(buf, next_tail);
960 /* ORDER: item must be accessed _before_ tail is updated */
961 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
963 } while (buf->rb_sc_ctxs[next_tail] != sc);
965 /* Paired with READ_ONCE */
966 smp_store_release(&buf->rb_sc_tail, next_tail);
968 if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
969 smp_mb__after_atomic();
970 xprt_write_space(&sc->sc_xprt->rx_xprt);
975 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
977 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
978 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
983 for (count = 0; count < ia->ri_max_segs; count++) {
984 struct rpcrdma_mr *mr;
987 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
991 rc = frwr_init_mr(ia, mr);
997 mr->mr_xprt = r_xprt;
999 list_add(&mr->mr_list, &free);
1000 list_add(&mr->mr_all, &all);
1003 spin_lock(&buf->rb_mrlock);
1004 list_splice(&free, &buf->rb_mrs);
1005 list_splice(&all, &buf->rb_all);
1006 r_xprt->rx_stats.mrs_allocated += count;
1007 spin_unlock(&buf->rb_mrlock);
1008 trace_xprtrdma_createmrs(r_xprt, count);
1010 xprt_write_space(&r_xprt->rx_xprt);
1014 rpcrdma_mr_refresh_worker(struct work_struct *work)
1016 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1017 rb_refresh_worker.work);
1018 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1021 rpcrdma_mrs_create(r_xprt);
1024 struct rpcrdma_req *
1025 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1027 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1028 struct rpcrdma_regbuf *rb;
1029 struct rpcrdma_req *req;
1031 req = kzalloc(sizeof(*req), GFP_KERNEL);
1033 return ERR_PTR(-ENOMEM);
1035 rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1036 DMA_TO_DEVICE, GFP_KERNEL);
1039 return ERR_PTR(-ENOMEM);
1041 req->rl_rdmabuf = rb;
1042 xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1043 req->rl_buffer = buffer;
1044 INIT_LIST_HEAD(&req->rl_registered);
1046 spin_lock(&buffer->rb_reqslock);
1047 list_add(&req->rl_all, &buffer->rb_allreqs);
1048 spin_unlock(&buffer->rb_reqslock);
1053 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1055 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1056 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1057 struct rpcrdma_rep *rep;
1061 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1065 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1066 DMA_FROM_DEVICE, GFP_KERNEL);
1067 if (IS_ERR(rep->rr_rdmabuf)) {
1068 rc = PTR_ERR(rep->rr_rdmabuf);
1071 xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1072 rdmab_length(rep->rr_rdmabuf));
1074 rep->rr_cqe.done = rpcrdma_wc_receive;
1075 rep->rr_rxprt = r_xprt;
1076 INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1077 rep->rr_recv_wr.next = NULL;
1078 rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1079 rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1080 rep->rr_recv_wr.num_sge = 1;
1081 rep->rr_temp = temp;
1083 spin_lock(&buf->rb_lock);
1084 list_add(&rep->rr_list, &buf->rb_recv_bufs);
1085 spin_unlock(&buf->rb_lock);
1091 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1097 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1099 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1103 buf->rb_max_requests = r_xprt->rx_data.max_requests;
1104 buf->rb_bc_srv_max_requests = 0;
1105 spin_lock_init(&buf->rb_mrlock);
1106 spin_lock_init(&buf->rb_lock);
1107 INIT_LIST_HEAD(&buf->rb_mrs);
1108 INIT_LIST_HEAD(&buf->rb_all);
1109 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1110 rpcrdma_mr_refresh_worker);
1112 rpcrdma_mrs_create(r_xprt);
1114 INIT_LIST_HEAD(&buf->rb_send_bufs);
1115 INIT_LIST_HEAD(&buf->rb_allreqs);
1116 spin_lock_init(&buf->rb_reqslock);
1117 for (i = 0; i < buf->rb_max_requests; i++) {
1118 struct rpcrdma_req *req;
1120 req = rpcrdma_create_req(r_xprt);
1122 dprintk("RPC: %s: request buffer %d alloc"
1123 " failed\n", __func__, i);
1127 list_add(&req->rl_list, &buf->rb_send_bufs);
1130 buf->rb_credits = 1;
1131 INIT_LIST_HEAD(&buf->rb_recv_bufs);
1133 rc = rpcrdma_sendctxs_create(r_xprt);
1137 buf->rb_completion_wq = alloc_workqueue("rpcrdma-%s",
1138 WQ_MEM_RECLAIM | WQ_HIGHPRI,
1140 r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR]);
1141 if (!buf->rb_completion_wq)
1146 rpcrdma_buffer_destroy(buf);
1151 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1153 rpcrdma_free_regbuf(rep->rr_rdmabuf);
1158 rpcrdma_destroy_req(struct rpcrdma_req *req)
1160 rpcrdma_free_regbuf(req->rl_recvbuf);
1161 rpcrdma_free_regbuf(req->rl_sendbuf);
1162 rpcrdma_free_regbuf(req->rl_rdmabuf);
1167 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1169 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1171 struct rpcrdma_mr *mr;
1175 spin_lock(&buf->rb_mrlock);
1176 while (!list_empty(&buf->rb_all)) {
1177 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1178 list_del(&mr->mr_all);
1180 spin_unlock(&buf->rb_mrlock);
1182 /* Ensure MW is not on any rl_registered list */
1183 if (!list_empty(&mr->mr_list))
1184 list_del(&mr->mr_list);
1186 frwr_release_mr(mr);
1188 spin_lock(&buf->rb_mrlock);
1190 spin_unlock(&buf->rb_mrlock);
1191 r_xprt->rx_stats.mrs_allocated = 0;
1193 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1197 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1199 cancel_delayed_work_sync(&buf->rb_refresh_worker);
1201 if (buf->rb_completion_wq) {
1202 destroy_workqueue(buf->rb_completion_wq);
1203 buf->rb_completion_wq = NULL;
1206 rpcrdma_sendctxs_destroy(buf);
1208 while (!list_empty(&buf->rb_recv_bufs)) {
1209 struct rpcrdma_rep *rep;
1211 rep = list_first_entry(&buf->rb_recv_bufs,
1212 struct rpcrdma_rep, rr_list);
1213 list_del(&rep->rr_list);
1214 rpcrdma_destroy_rep(rep);
1217 spin_lock(&buf->rb_reqslock);
1218 while (!list_empty(&buf->rb_allreqs)) {
1219 struct rpcrdma_req *req;
1221 req = list_first_entry(&buf->rb_allreqs,
1222 struct rpcrdma_req, rl_all);
1223 list_del(&req->rl_all);
1225 spin_unlock(&buf->rb_reqslock);
1226 rpcrdma_destroy_req(req);
1227 spin_lock(&buf->rb_reqslock);
1229 spin_unlock(&buf->rb_reqslock);
1231 rpcrdma_mrs_destroy(buf);
1235 * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1236 * @r_xprt: controlling transport
1238 * Returns an initialized rpcrdma_mr or NULL if no free
1239 * rpcrdma_mr objects are available.
1242 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1244 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1245 struct rpcrdma_mr *mr = NULL;
1247 spin_lock(&buf->rb_mrlock);
1248 if (!list_empty(&buf->rb_mrs))
1249 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1250 spin_unlock(&buf->rb_mrlock);
1257 trace_xprtrdma_nomrs(r_xprt);
1258 if (r_xprt->rx_ep.rep_connected != -ENODEV)
1259 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1261 /* Allow the reply handler and refresh worker to run */
1268 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1270 spin_lock(&buf->rb_mrlock);
1271 rpcrdma_mr_push(mr, &buf->rb_mrs);
1272 spin_unlock(&buf->rb_mrlock);
1276 * rpcrdma_mr_put - Release an rpcrdma_mr object
1277 * @mr: object to release
1281 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1283 __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1287 * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1288 * @mr: object to release
1292 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1294 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1296 if (mr->mr_dir != DMA_NONE) {
1297 trace_xprtrdma_mr_unmap(mr);
1298 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1299 mr->mr_sg, mr->mr_nents, mr->mr_dir);
1300 mr->mr_dir = DMA_NONE;
1302 __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1306 * rpcrdma_buffer_get - Get a request buffer
1307 * @buffers: Buffer pool from which to obtain a buffer
1309 * Returns a fresh rpcrdma_req, or NULL if none are available.
1311 struct rpcrdma_req *
1312 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1314 struct rpcrdma_req *req;
1316 spin_lock(&buffers->rb_lock);
1317 req = list_first_entry_or_null(&buffers->rb_send_bufs,
1318 struct rpcrdma_req, rl_list);
1320 list_del_init(&req->rl_list);
1321 spin_unlock(&buffers->rb_lock);
1326 * rpcrdma_buffer_put - Put request/reply buffers back into pool
1327 * @req: object to return
1331 rpcrdma_buffer_put(struct rpcrdma_req *req)
1333 struct rpcrdma_buffer *buffers = req->rl_buffer;
1334 struct rpcrdma_rep *rep = req->rl_reply;
1336 req->rl_reply = NULL;
1338 spin_lock(&buffers->rb_lock);
1339 list_add(&req->rl_list, &buffers->rb_send_bufs);
1341 if (!rep->rr_temp) {
1342 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1346 spin_unlock(&buffers->rb_lock);
1348 rpcrdma_destroy_rep(rep);
1352 * Put reply buffers back into pool when not attached to
1353 * request. This happens in error conditions.
1356 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1358 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1360 if (!rep->rr_temp) {
1361 spin_lock(&buffers->rb_lock);
1362 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1363 spin_unlock(&buffers->rb_lock);
1365 rpcrdma_destroy_rep(rep);
1370 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1371 * @size: size of buffer to be allocated, in bytes
1372 * @direction: direction of data movement
1375 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1376 * can be persistently DMA-mapped for I/O.
1378 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1379 * receiving the payload of RDMA RECV operations. During Long Calls
1380 * or Replies they may be registered externally via frwr_map.
1382 struct rpcrdma_regbuf *
1383 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1386 struct rpcrdma_regbuf *rb;
1388 rb = kmalloc(sizeof(*rb) + size, flags);
1390 return ERR_PTR(-ENOMEM);
1392 rb->rg_device = NULL;
1393 rb->rg_direction = direction;
1394 rb->rg_iov.length = size;
1400 * __rpcrdma_map_regbuf - DMA-map a regbuf
1401 * @ia: controlling rpcrdma_ia
1402 * @rb: regbuf to be mapped
1405 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1407 struct ib_device *device = ia->ri_device;
1409 if (rb->rg_direction == DMA_NONE)
1412 rb->rg_iov.addr = ib_dma_map_single(device,
1413 (void *)rb->rg_base,
1416 if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1419 rb->rg_device = device;
1420 rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1425 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1430 if (!rpcrdma_regbuf_is_mapped(rb))
1433 ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1434 rdmab_length(rb), rb->rg_direction);
1435 rb->rg_device = NULL;
1439 * rpcrdma_free_regbuf - deregister and free registered buffer
1440 * @rb: regbuf to be deregistered and freed
1443 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1445 rpcrdma_dma_unmap_regbuf(rb);
1450 * Prepost any receive buffer, then post send.
1452 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1455 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1456 struct rpcrdma_ep *ep,
1457 struct rpcrdma_req *req)
1459 struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1462 if (!ep->rep_send_count ||
1463 test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1464 send_wr->send_flags |= IB_SEND_SIGNALED;
1465 ep->rep_send_count = ep->rep_send_batch;
1467 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1468 --ep->rep_send_count;
1471 rc = frwr_send(ia, req);
1472 trace_xprtrdma_post_send(req, rc);
1479 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1481 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1482 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1483 struct ib_recv_wr *wr, *bad_wr;
1484 int needed, count, rc;
1488 needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1489 if (ep->rep_receive_count > needed)
1491 needed -= ep->rep_receive_count;
1496 struct rpcrdma_regbuf *rb;
1497 struct rpcrdma_rep *rep;
1499 spin_lock(&buf->rb_lock);
1500 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1501 struct rpcrdma_rep, rr_list);
1503 list_del(&rep->rr_list);
1504 spin_unlock(&buf->rb_lock);
1506 if (rpcrdma_create_rep(r_xprt, temp))
1511 rb = rep->rr_rdmabuf;
1512 if (!rpcrdma_regbuf_is_mapped(rb)) {
1513 if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1514 rpcrdma_recv_buffer_put(rep);
1519 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1520 rep->rr_recv_wr.next = wr;
1521 wr = &rep->rr_recv_wr;
1528 rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1529 (const struct ib_recv_wr **)&bad_wr);
1531 for (wr = bad_wr; wr; wr = wr->next) {
1532 struct rpcrdma_rep *rep;
1534 rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1535 rpcrdma_recv_buffer_put(rep);
1539 ep->rep_receive_count += count;
1541 trace_xprtrdma_post_recvs(r_xprt, count, rc);