2 * Copyright (c) 2014-2017 Oracle. All rights reserved.
3 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 * Encapsulates the major functions managing:
51 #include <linux/interrupt.h>
52 #include <linux/slab.h>
53 #include <linux/sunrpc/addr.h>
54 #include <linux/sunrpc/svc_rdma.h>
56 #include <asm-generic/barrier.h>
57 #include <asm/bitops.h>
59 #include <rdma/ib_cm.h>
61 #include "xprt_rdma.h"
67 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
68 # define RPCDBG_FACILITY RPCDBG_TRANS
74 static void rpcrdma_create_mrs(struct rpcrdma_xprt
*r_xprt
);
75 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer
*buf
);
76 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf
*rb
);
78 struct workqueue_struct
*rpcrdma_receive_wq __read_mostly
;
81 rpcrdma_alloc_wq(void)
83 struct workqueue_struct
*recv_wq
;
85 recv_wq
= alloc_workqueue("xprtrdma_receive",
86 WQ_MEM_RECLAIM
| WQ_UNBOUND
| WQ_HIGHPRI
,
91 rpcrdma_receive_wq
= recv_wq
;
96 rpcrdma_destroy_wq(void)
98 struct workqueue_struct
*wq
;
100 if (rpcrdma_receive_wq
) {
101 wq
= rpcrdma_receive_wq
;
102 rpcrdma_receive_wq
= NULL
;
103 destroy_workqueue(wq
);
108 rpcrdma_qp_async_error_upcall(struct ib_event
*event
, void *context
)
110 struct rpcrdma_ep
*ep
= context
;
112 pr_err("rpcrdma: %s on device %s ep %p\n",
113 ib_event_msg(event
->event
), event
->device
->name
, context
);
115 if (ep
->rep_connected
== 1) {
116 ep
->rep_connected
= -EIO
;
117 rpcrdma_conn_func(ep
);
118 wake_up_all(&ep
->rep_connect_wait
);
123 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
124 * @cq: completion queue (ignored)
129 rpcrdma_wc_send(struct ib_cq
*cq
, struct ib_wc
*wc
)
131 struct ib_cqe
*cqe
= wc
->wr_cqe
;
132 struct rpcrdma_sendctx
*sc
=
133 container_of(cqe
, struct rpcrdma_sendctx
, sc_cqe
);
135 /* WARNING: Only wr_cqe and status are reliable at this point */
136 if (wc
->status
!= IB_WC_SUCCESS
&& wc
->status
!= IB_WC_WR_FLUSH_ERR
)
137 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
138 ib_wc_status_msg(wc
->status
),
139 wc
->status
, wc
->vendor_err
);
141 rpcrdma_sendctx_put_locked(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
,
157 /* WARNING: Only wr_id and status are reliable at this point */
158 if (wc
->status
!= IB_WC_SUCCESS
)
161 /* status == SUCCESS means all fields in wc are trustworthy */
162 dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n",
163 __func__
, rep
, wc
->byte_len
);
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
);
174 rpcrdma_reply_handler(rep
);
178 if (wc
->status
!= IB_WC_WR_FLUSH_ERR
)
179 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
180 ib_wc_status_msg(wc
->status
),
181 wc
->status
, wc
->vendor_err
);
182 rpcrdma_set_xdrlen(&rep
->rr_hdrbuf
, 0);
187 rpcrdma_update_connect_private(struct rpcrdma_xprt
*r_xprt
,
188 struct rdma_conn_param
*param
)
190 struct rpcrdma_create_data_internal
*cdata
= &r_xprt
->rx_data
;
191 const struct rpcrdma_connect_private
*pmsg
= param
->private_data
;
192 unsigned int rsize
, wsize
;
194 /* Default settings for RPC-over-RDMA Version One */
195 r_xprt
->rx_ia
.ri_reminv_expected
= false;
196 r_xprt
->rx_ia
.ri_implicit_roundup
= xprt_rdma_pad_optimize
;
197 rsize
= RPCRDMA_V1_DEF_INLINE_SIZE
;
198 wsize
= RPCRDMA_V1_DEF_INLINE_SIZE
;
201 pmsg
->cp_magic
== rpcrdma_cmp_magic
&&
202 pmsg
->cp_version
== RPCRDMA_CMP_VERSION
) {
203 r_xprt
->rx_ia
.ri_reminv_expected
= true;
204 r_xprt
->rx_ia
.ri_implicit_roundup
= true;
205 rsize
= rpcrdma_decode_buffer_size(pmsg
->cp_send_size
);
206 wsize
= rpcrdma_decode_buffer_size(pmsg
->cp_recv_size
);
209 if (rsize
< cdata
->inline_rsize
)
210 cdata
->inline_rsize
= rsize
;
211 if (wsize
< cdata
->inline_wsize
)
212 cdata
->inline_wsize
= wsize
;
213 dprintk("RPC: %s: max send %u, max recv %u\n",
214 __func__
, cdata
->inline_wsize
, cdata
->inline_rsize
);
215 rpcrdma_set_max_header_sizes(r_xprt
);
219 rpcrdma_conn_upcall(struct rdma_cm_id
*id
, struct rdma_cm_event
*event
)
221 struct rpcrdma_xprt
*xprt
= id
->context
;
222 struct rpcrdma_ia
*ia
= &xprt
->rx_ia
;
223 struct rpcrdma_ep
*ep
= &xprt
->rx_ep
;
224 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
225 struct sockaddr
*sap
= (struct sockaddr
*)&ep
->rep_remote_addr
;
229 switch (event
->event
) {
230 case RDMA_CM_EVENT_ADDR_RESOLVED
:
231 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
233 complete(&ia
->ri_done
);
235 case RDMA_CM_EVENT_ADDR_ERROR
:
236 ia
->ri_async_rc
= -EHOSTUNREACH
;
237 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
239 complete(&ia
->ri_done
);
241 case RDMA_CM_EVENT_ROUTE_ERROR
:
242 ia
->ri_async_rc
= -ENETUNREACH
;
243 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
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 %pIS:%u\n",
251 sap
, rpc_get_port(sap
));
253 set_bit(RPCRDMA_IAF_REMOVING
, &ia
->ri_flags
);
254 ep
->rep_connected
= -ENODEV
;
255 xprt_force_disconnect(&xprt
->rx_xprt
);
256 wait_for_completion(&ia
->ri_remove_done
);
260 ia
->ri_device
= NULL
;
261 /* Return 1 to ensure the core destroys the id. */
263 case RDMA_CM_EVENT_ESTABLISHED
:
265 rpcrdma_update_connect_private(xprt
, &event
->param
.conn
);
267 case RDMA_CM_EVENT_CONNECT_ERROR
:
268 connstate
= -ENOTCONN
;
270 case RDMA_CM_EVENT_UNREACHABLE
:
271 connstate
= -ENETDOWN
;
273 case RDMA_CM_EVENT_REJECTED
:
274 dprintk("rpcrdma: connection to %pIS:%u rejected: %s\n",
275 sap
, rpc_get_port(sap
),
276 rdma_reject_msg(id
, event
->status
));
277 connstate
= -ECONNREFUSED
;
278 if (event
->status
== IB_CM_REJ_STALE_CONN
)
281 case RDMA_CM_EVENT_DISCONNECTED
:
282 connstate
= -ECONNABORTED
;
284 xprt
->rx_buf
.rb_credits
= 1;
285 ep
->rep_connected
= connstate
;
286 rpcrdma_conn_func(ep
);
287 wake_up_all(&ep
->rep_connect_wait
);
290 dprintk("RPC: %s: %pIS:%u on %s/%s (ep 0x%p): %s\n",
291 __func__
, sap
, rpc_get_port(sap
),
292 ia
->ri_device
->name
, ia
->ri_ops
->ro_displayname
,
293 ep
, rdma_event_msg(event
->event
));
300 static struct rdma_cm_id
*
301 rpcrdma_create_id(struct rpcrdma_xprt
*xprt
,
302 struct rpcrdma_ia
*ia
, struct sockaddr
*addr
)
304 unsigned long wtimeout
= msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT
) + 1;
305 struct rdma_cm_id
*id
;
308 init_completion(&ia
->ri_done
);
309 init_completion(&ia
->ri_remove_done
);
311 id
= rdma_create_id(&init_net
, rpcrdma_conn_upcall
, xprt
, RDMA_PS_TCP
,
315 dprintk("RPC: %s: rdma_create_id() failed %i\n",
320 ia
->ri_async_rc
= -ETIMEDOUT
;
321 rc
= rdma_resolve_addr(id
, NULL
, addr
, RDMA_RESOLVE_TIMEOUT
);
323 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
327 rc
= wait_for_completion_interruptible_timeout(&ia
->ri_done
, wtimeout
);
329 dprintk("RPC: %s: wait() exited: %i\n",
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 dprintk("RPC: %s: wait() exited: %i\n",
351 rc
= ia
->ri_async_rc
;
363 * Exported functions.
367 * rpcrdma_ia_open - Open and initialize an Interface Adapter.
368 * @xprt: controlling transport
369 * @addr: IP address of remote peer
371 * Returns 0 on success, negative errno if an appropriate
372 * Interface Adapter could not be found and opened.
375 rpcrdma_ia_open(struct rpcrdma_xprt
*xprt
, struct sockaddr
*addr
)
377 struct rpcrdma_ia
*ia
= &xprt
->rx_ia
;
380 ia
->ri_id
= rpcrdma_create_id(xprt
, ia
, addr
);
381 if (IS_ERR(ia
->ri_id
)) {
382 rc
= PTR_ERR(ia
->ri_id
);
385 ia
->ri_device
= ia
->ri_id
->device
;
387 ia
->ri_pd
= ib_alloc_pd(ia
->ri_device
, 0);
388 if (IS_ERR(ia
->ri_pd
)) {
389 rc
= PTR_ERR(ia
->ri_pd
);
390 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc
);
394 switch (xprt_rdma_memreg_strategy
) {
396 if (frwr_is_supported(ia
)) {
397 ia
->ri_ops
= &rpcrdma_frwr_memreg_ops
;
401 case RPCRDMA_MTHCAFMR
:
402 if (fmr_is_supported(ia
)) {
403 ia
->ri_ops
= &rpcrdma_fmr_memreg_ops
;
408 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
409 ia
->ri_device
->name
, xprt_rdma_memreg_strategy
);
417 rpcrdma_ia_close(ia
);
422 * rpcrdma_ia_remove - Handle device driver unload
423 * @ia: interface adapter being removed
425 * Divest transport H/W resources associated with this adapter,
426 * but allow it to be restored later.
429 rpcrdma_ia_remove(struct rpcrdma_ia
*ia
)
431 struct rpcrdma_xprt
*r_xprt
= container_of(ia
, struct rpcrdma_xprt
,
433 struct rpcrdma_ep
*ep
= &r_xprt
->rx_ep
;
434 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
435 struct rpcrdma_req
*req
;
436 struct rpcrdma_rep
*rep
;
438 cancel_delayed_work_sync(&buf
->rb_refresh_worker
);
440 /* This is similar to rpcrdma_ep_destroy, but:
441 * - Don't cancel the connect worker.
442 * - Don't call rpcrdma_ep_disconnect, which waits
443 * for another conn upcall, which will deadlock.
444 * - rdma_disconnect is unneeded, the underlying
445 * connection is already gone.
448 ib_drain_qp(ia
->ri_id
->qp
);
449 rdma_destroy_qp(ia
->ri_id
);
450 ia
->ri_id
->qp
= NULL
;
452 ib_free_cq(ep
->rep_attr
.recv_cq
);
453 ib_free_cq(ep
->rep_attr
.send_cq
);
455 /* The ULP is responsible for ensuring all DMA
456 * mappings and MRs are gone.
458 list_for_each_entry(rep
, &buf
->rb_recv_bufs
, rr_list
)
459 rpcrdma_dma_unmap_regbuf(rep
->rr_rdmabuf
);
460 list_for_each_entry(req
, &buf
->rb_allreqs
, rl_all
) {
461 rpcrdma_dma_unmap_regbuf(req
->rl_rdmabuf
);
462 rpcrdma_dma_unmap_regbuf(req
->rl_sendbuf
);
463 rpcrdma_dma_unmap_regbuf(req
->rl_recvbuf
);
465 rpcrdma_destroy_mrs(buf
);
467 /* Allow waiters to continue */
468 complete(&ia
->ri_remove_done
);
472 * rpcrdma_ia_close - Clean up/close an IA.
473 * @ia: interface adapter to close
477 rpcrdma_ia_close(struct rpcrdma_ia
*ia
)
479 dprintk("RPC: %s: entering\n", __func__
);
480 if (ia
->ri_id
!= NULL
&& !IS_ERR(ia
->ri_id
)) {
482 rdma_destroy_qp(ia
->ri_id
);
483 rdma_destroy_id(ia
->ri_id
);
486 ia
->ri_device
= NULL
;
488 /* If the pd is still busy, xprtrdma missed freeing a resource */
489 if (ia
->ri_pd
&& !IS_ERR(ia
->ri_pd
))
490 ib_dealloc_pd(ia
->ri_pd
);
495 * Create unconnected endpoint.
498 rpcrdma_ep_create(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
,
499 struct rpcrdma_create_data_internal
*cdata
)
501 struct rpcrdma_connect_private
*pmsg
= &ep
->rep_cm_private
;
502 unsigned int max_qp_wr
, max_sge
;
503 struct ib_cq
*sendcq
, *recvcq
;
506 max_sge
= min_t(unsigned int, ia
->ri_device
->attrs
.max_sge
,
507 RPCRDMA_MAX_SEND_SGES
);
508 if (max_sge
< RPCRDMA_MIN_SEND_SGES
) {
509 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge
);
512 ia
->ri_max_send_sges
= max_sge
- RPCRDMA_MIN_SEND_SGES
;
514 if (ia
->ri_device
->attrs
.max_qp_wr
<= RPCRDMA_BACKWARD_WRS
) {
515 dprintk("RPC: %s: insufficient wqe's available\n",
519 max_qp_wr
= ia
->ri_device
->attrs
.max_qp_wr
- RPCRDMA_BACKWARD_WRS
- 1;
521 /* check provider's send/recv wr limits */
522 if (cdata
->max_requests
> max_qp_wr
)
523 cdata
->max_requests
= max_qp_wr
;
525 ep
->rep_attr
.event_handler
= rpcrdma_qp_async_error_upcall
;
526 ep
->rep_attr
.qp_context
= ep
;
527 ep
->rep_attr
.srq
= NULL
;
528 ep
->rep_attr
.cap
.max_send_wr
= cdata
->max_requests
;
529 ep
->rep_attr
.cap
.max_send_wr
+= RPCRDMA_BACKWARD_WRS
;
530 ep
->rep_attr
.cap
.max_send_wr
+= 1; /* drain cqe */
531 rc
= ia
->ri_ops
->ro_open(ia
, ep
, cdata
);
534 ep
->rep_attr
.cap
.max_recv_wr
= cdata
->max_requests
;
535 ep
->rep_attr
.cap
.max_recv_wr
+= RPCRDMA_BACKWARD_WRS
;
536 ep
->rep_attr
.cap
.max_recv_wr
+= 1; /* drain cqe */
537 ep
->rep_attr
.cap
.max_send_sge
= max_sge
;
538 ep
->rep_attr
.cap
.max_recv_sge
= 1;
539 ep
->rep_attr
.cap
.max_inline_data
= 0;
540 ep
->rep_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
541 ep
->rep_attr
.qp_type
= IB_QPT_RC
;
542 ep
->rep_attr
.port_num
= ~0;
544 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
545 "iovs: send %d recv %d\n",
547 ep
->rep_attr
.cap
.max_send_wr
,
548 ep
->rep_attr
.cap
.max_recv_wr
,
549 ep
->rep_attr
.cap
.max_send_sge
,
550 ep
->rep_attr
.cap
.max_recv_sge
);
552 /* set trigger for requesting send completion */
553 ep
->rep_send_batch
= min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH
,
554 cdata
->max_requests
>> 2);
555 ep
->rep_send_count
= ep
->rep_send_batch
;
556 init_waitqueue_head(&ep
->rep_connect_wait
);
557 INIT_DELAYED_WORK(&ep
->rep_connect_worker
, rpcrdma_connect_worker
);
559 sendcq
= ib_alloc_cq(ia
->ri_device
, NULL
,
560 ep
->rep_attr
.cap
.max_send_wr
+ 1,
561 1, IB_POLL_WORKQUEUE
);
562 if (IS_ERR(sendcq
)) {
563 rc
= PTR_ERR(sendcq
);
564 dprintk("RPC: %s: failed to create send CQ: %i\n",
569 recvcq
= ib_alloc_cq(ia
->ri_device
, NULL
,
570 ep
->rep_attr
.cap
.max_recv_wr
+ 1,
571 0, IB_POLL_WORKQUEUE
);
572 if (IS_ERR(recvcq
)) {
573 rc
= PTR_ERR(recvcq
);
574 dprintk("RPC: %s: failed to create recv CQ: %i\n",
579 ep
->rep_attr
.send_cq
= sendcq
;
580 ep
->rep_attr
.recv_cq
= recvcq
;
582 /* Initialize cma parameters */
583 memset(&ep
->rep_remote_cma
, 0, sizeof(ep
->rep_remote_cma
));
585 /* Prepare RDMA-CM private message */
586 pmsg
->cp_magic
= rpcrdma_cmp_magic
;
587 pmsg
->cp_version
= RPCRDMA_CMP_VERSION
;
588 pmsg
->cp_flags
|= ia
->ri_ops
->ro_send_w_inv_ok
;
589 pmsg
->cp_send_size
= rpcrdma_encode_buffer_size(cdata
->inline_wsize
);
590 pmsg
->cp_recv_size
= rpcrdma_encode_buffer_size(cdata
->inline_rsize
);
591 ep
->rep_remote_cma
.private_data
= pmsg
;
592 ep
->rep_remote_cma
.private_data_len
= sizeof(*pmsg
);
594 /* Client offers RDMA Read but does not initiate */
595 ep
->rep_remote_cma
.initiator_depth
= 0;
596 if (ia
->ri_device
->attrs
.max_qp_rd_atom
> 32) /* arbitrary but <= 255 */
597 ep
->rep_remote_cma
.responder_resources
= 32;
599 ep
->rep_remote_cma
.responder_resources
=
600 ia
->ri_device
->attrs
.max_qp_rd_atom
;
602 /* Limit transport retries so client can detect server
603 * GID changes quickly. RPC layer handles re-establishing
604 * transport connection and retransmission.
606 ep
->rep_remote_cma
.retry_count
= 6;
608 /* RPC-over-RDMA handles its own flow control. In addition,
609 * make all RNR NAKs visible so we know that RPC-over-RDMA
610 * flow control is working correctly (no NAKs should be seen).
612 ep
->rep_remote_cma
.flow_control
= 0;
613 ep
->rep_remote_cma
.rnr_retry_count
= 0;
626 * Disconnect and destroy endpoint. After this, the only
627 * valid operations on the ep are to free it (if dynamically
628 * allocated) or re-create it.
631 rpcrdma_ep_destroy(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
633 dprintk("RPC: %s: entering, connected is %d\n",
634 __func__
, ep
->rep_connected
);
636 cancel_delayed_work_sync(&ep
->rep_connect_worker
);
639 rpcrdma_ep_disconnect(ep
, ia
);
640 rdma_destroy_qp(ia
->ri_id
);
641 ia
->ri_id
->qp
= NULL
;
644 ib_free_cq(ep
->rep_attr
.recv_cq
);
645 ib_free_cq(ep
->rep_attr
.send_cq
);
648 /* Re-establish a connection after a device removal event.
649 * Unlike a normal reconnection, a fresh PD and a new set
650 * of MRs and buffers is needed.
653 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt
*r_xprt
,
654 struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
656 struct sockaddr
*sap
= (struct sockaddr
*)&r_xprt
->rx_data
.addr
;
659 pr_info("%s: r_xprt = %p\n", __func__
, r_xprt
);
662 if (rpcrdma_ia_open(r_xprt
, sap
))
666 err
= rpcrdma_ep_create(ep
, ia
, &r_xprt
->rx_data
);
668 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err
);
673 err
= rdma_create_qp(ia
->ri_id
, ia
->ri_pd
, &ep
->rep_attr
);
675 pr_err("rpcrdma: rdma_create_qp returned %d\n", err
);
679 rpcrdma_create_mrs(r_xprt
);
683 rpcrdma_ep_destroy(ep
, ia
);
685 rpcrdma_ia_close(ia
);
691 rpcrdma_ep_reconnect(struct rpcrdma_xprt
*r_xprt
, struct rpcrdma_ep
*ep
,
692 struct rpcrdma_ia
*ia
)
694 struct sockaddr
*sap
= (struct sockaddr
*)&r_xprt
->rx_data
.addr
;
695 struct rdma_cm_id
*id
, *old
;
698 dprintk("RPC: %s: reconnecting...\n", __func__
);
700 rpcrdma_ep_disconnect(ep
, ia
);
703 id
= rpcrdma_create_id(r_xprt
, ia
, sap
);
707 /* As long as the new ID points to the same device as the
708 * old ID, we can reuse the transport's existing PD and all
709 * previously allocated MRs. Also, the same device means
710 * the transport's previous DMA mappings are still valid.
712 * This is a sanity check only. There should be no way these
713 * point to two different devices here.
717 if (ia
->ri_device
!= id
->device
) {
718 pr_err("rpcrdma: can't reconnect on different device!\n");
722 err
= rdma_create_qp(id
, ia
->ri_pd
, &ep
->rep_attr
);
724 dprintk("RPC: %s: rdma_create_qp returned %d\n",
729 /* Atomically replace the transport's ID and QP. */
733 rdma_destroy_qp(old
);
736 rdma_destroy_id(old
);
742 * Connect unconnected endpoint.
745 rpcrdma_ep_connect(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
747 struct rpcrdma_xprt
*r_xprt
= container_of(ia
, struct rpcrdma_xprt
,
753 switch (ep
->rep_connected
) {
755 dprintk("RPC: %s: connecting...\n", __func__
);
756 rc
= rdma_create_qp(ia
->ri_id
, ia
->ri_pd
, &ep
->rep_attr
);
758 dprintk("RPC: %s: rdma_create_qp failed %i\n",
765 rc
= rpcrdma_ep_recreate_xprt(r_xprt
, ep
, ia
);
770 rc
= rpcrdma_ep_reconnect(r_xprt
, ep
, ia
);
775 ep
->rep_connected
= 0;
777 rc
= rdma_connect(ia
->ri_id
, &ep
->rep_remote_cma
);
779 dprintk("RPC: %s: rdma_connect() failed with %i\n",
784 wait_event_interruptible(ep
->rep_connect_wait
, ep
->rep_connected
!= 0);
785 if (ep
->rep_connected
<= 0) {
786 if (ep
->rep_connected
== -EAGAIN
)
788 rc
= ep
->rep_connected
;
792 dprintk("RPC: %s: connected\n", __func__
);
793 extras
= r_xprt
->rx_buf
.rb_bc_srv_max_requests
;
795 rpcrdma_ep_post_extra_recv(r_xprt
, extras
);
799 ep
->rep_connected
= rc
;
806 * rpcrdma_ep_disconnect
808 * This is separate from destroy to facilitate the ability
809 * to reconnect without recreating the endpoint.
811 * This call is not reentrant, and must not be made in parallel
812 * on the same endpoint.
815 rpcrdma_ep_disconnect(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
819 rc
= rdma_disconnect(ia
->ri_id
);
821 /* returns without wait if not connected */
822 wait_event_interruptible(ep
->rep_connect_wait
,
823 ep
->rep_connected
!= 1);
824 dprintk("RPC: %s: after wait, %sconnected\n", __func__
,
825 (ep
->rep_connected
== 1) ? "still " : "dis");
827 dprintk("RPC: %s: rdma_disconnect %i\n", __func__
, rc
);
828 ep
->rep_connected
= rc
;
831 ib_drain_qp(ia
->ri_id
->qp
);
834 /* Fixed-size circular FIFO queue. This implementation is wait-free and
837 * Consumer is the code path that posts Sends. This path dequeues a
838 * sendctx for use by a Send operation. Multiple consumer threads
839 * are serialized by the RPC transport lock, which allows only one
840 * ->send_request call at a time.
842 * Producer is the code path that handles Send completions. This path
843 * enqueues a sendctx that has been completed. Multiple producer
844 * threads are serialized by the ib_poll_cq() function.
847 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
848 * queue activity, and ib_drain_qp has flushed all remaining Send
851 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer
*buf
)
855 for (i
= 0; i
<= buf
->rb_sc_last
; i
++)
856 kfree(buf
->rb_sc_ctxs
[i
]);
857 kfree(buf
->rb_sc_ctxs
);
860 static struct rpcrdma_sendctx
*rpcrdma_sendctx_create(struct rpcrdma_ia
*ia
)
862 struct rpcrdma_sendctx
*sc
;
864 sc
= kzalloc(sizeof(*sc
) +
865 ia
->ri_max_send_sges
* sizeof(struct ib_sge
),
870 sc
->sc_wr
.wr_cqe
= &sc
->sc_cqe
;
871 sc
->sc_wr
.sg_list
= sc
->sc_sges
;
872 sc
->sc_wr
.opcode
= IB_WR_SEND
;
873 sc
->sc_cqe
.done
= rpcrdma_wc_send
;
877 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt
*r_xprt
)
879 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
880 struct rpcrdma_sendctx
*sc
;
883 /* Maximum number of concurrent outstanding Send WRs. Capping
884 * the circular queue size stops Send Queue overflow by causing
885 * the ->send_request call to fail temporarily before too many
888 i
= buf
->rb_max_requests
+ RPCRDMA_MAX_BC_REQUESTS
;
889 dprintk("RPC: %s: allocating %lu send_ctxs\n", __func__
, i
);
890 buf
->rb_sc_ctxs
= kcalloc(i
, sizeof(sc
), GFP_KERNEL
);
891 if (!buf
->rb_sc_ctxs
)
894 buf
->rb_sc_last
= i
- 1;
895 for (i
= 0; i
<= buf
->rb_sc_last
; i
++) {
896 sc
= rpcrdma_sendctx_create(&r_xprt
->rx_ia
);
900 sc
->sc_xprt
= r_xprt
;
901 buf
->rb_sc_ctxs
[i
] = sc
;
907 rpcrdma_sendctxs_destroy(buf
);
911 /* The sendctx queue is not guaranteed to have a size that is a
912 * power of two, thus the helpers in circ_buf.h cannot be used.
913 * The other option is to use modulus (%), which can be expensive.
915 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer
*buf
,
918 return likely(item
< buf
->rb_sc_last
) ? item
+ 1 : 0;
922 * rpcrdma_sendctx_get_locked - Acquire a send context
923 * @buf: transport buffers from which to acquire an unused context
925 * Returns pointer to a free send completion context; or NULL if
926 * the queue is empty.
928 * Usage: Called to acquire an SGE array before preparing a Send WR.
930 * The caller serializes calls to this function (per rpcrdma_buffer),
931 * and provides an effective memory barrier that flushes the new value
934 struct rpcrdma_sendctx
*rpcrdma_sendctx_get_locked(struct rpcrdma_buffer
*buf
)
936 struct rpcrdma_xprt
*r_xprt
;
937 struct rpcrdma_sendctx
*sc
;
938 unsigned long next_head
;
940 next_head
= rpcrdma_sendctx_next(buf
, buf
->rb_sc_head
);
942 if (next_head
== READ_ONCE(buf
->rb_sc_tail
))
945 /* ORDER: item must be accessed _before_ head is updated */
946 sc
= buf
->rb_sc_ctxs
[next_head
];
948 /* Releasing the lock in the caller acts as a memory
949 * barrier that flushes rb_sc_head.
951 buf
->rb_sc_head
= next_head
;
956 /* The queue is "empty" if there have not been enough Send
957 * completions recently. This is a sign the Send Queue is
958 * backing up. Cause the caller to pause and try again.
960 dprintk("RPC: %s: empty sendctx queue\n", __func__
);
961 r_xprt
= container_of(buf
, struct rpcrdma_xprt
, rx_buf
);
962 r_xprt
->rx_stats
.empty_sendctx_q
++;
967 * rpcrdma_sendctx_put_locked - Release a send context
968 * @sc: send context to release
970 * Usage: Called from Send completion to return a sendctxt
973 * The caller serializes calls to this function (per rpcrdma_buffer).
975 void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx
*sc
)
977 struct rpcrdma_buffer
*buf
= &sc
->sc_xprt
->rx_buf
;
978 unsigned long next_tail
;
980 /* Unmap SGEs of previously completed by unsignaled
981 * Sends by walking up the queue until @sc is found.
983 next_tail
= buf
->rb_sc_tail
;
985 next_tail
= rpcrdma_sendctx_next(buf
, next_tail
);
987 /* ORDER: item must be accessed _before_ tail is updated */
988 rpcrdma_unmap_sendctx(buf
->rb_sc_ctxs
[next_tail
]);
990 } while (buf
->rb_sc_ctxs
[next_tail
] != sc
);
992 /* Paired with READ_ONCE */
993 smp_store_release(&buf
->rb_sc_tail
, next_tail
);
997 rpcrdma_mr_recovery_worker(struct work_struct
*work
)
999 struct rpcrdma_buffer
*buf
= container_of(work
, struct rpcrdma_buffer
,
1000 rb_recovery_worker
.work
);
1001 struct rpcrdma_mw
*mw
;
1003 spin_lock(&buf
->rb_recovery_lock
);
1004 while (!list_empty(&buf
->rb_stale_mrs
)) {
1005 mw
= rpcrdma_pop_mw(&buf
->rb_stale_mrs
);
1006 spin_unlock(&buf
->rb_recovery_lock
);
1008 dprintk("RPC: %s: recovering MR %p\n", __func__
, mw
);
1009 mw
->mw_xprt
->rx_ia
.ri_ops
->ro_recover_mr(mw
);
1011 spin_lock(&buf
->rb_recovery_lock
);
1013 spin_unlock(&buf
->rb_recovery_lock
);
1017 rpcrdma_defer_mr_recovery(struct rpcrdma_mw
*mw
)
1019 struct rpcrdma_xprt
*r_xprt
= mw
->mw_xprt
;
1020 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1022 spin_lock(&buf
->rb_recovery_lock
);
1023 rpcrdma_push_mw(mw
, &buf
->rb_stale_mrs
);
1024 spin_unlock(&buf
->rb_recovery_lock
);
1026 schedule_delayed_work(&buf
->rb_recovery_worker
, 0);
1030 rpcrdma_create_mrs(struct rpcrdma_xprt
*r_xprt
)
1032 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1033 struct rpcrdma_ia
*ia
= &r_xprt
->rx_ia
;
1038 for (count
= 0; count
< 32; count
++) {
1039 struct rpcrdma_mw
*mw
;
1042 mw
= kzalloc(sizeof(*mw
), GFP_KERNEL
);
1046 rc
= ia
->ri_ops
->ro_init_mr(ia
, mw
);
1052 mw
->mw_xprt
= r_xprt
;
1054 list_add(&mw
->mw_list
, &free
);
1055 list_add(&mw
->mw_all
, &all
);
1058 spin_lock(&buf
->rb_mwlock
);
1059 list_splice(&free
, &buf
->rb_mws
);
1060 list_splice(&all
, &buf
->rb_all
);
1061 r_xprt
->rx_stats
.mrs_allocated
+= count
;
1062 spin_unlock(&buf
->rb_mwlock
);
1064 dprintk("RPC: %s: created %u MRs\n", __func__
, count
);
1068 rpcrdma_mr_refresh_worker(struct work_struct
*work
)
1070 struct rpcrdma_buffer
*buf
= container_of(work
, struct rpcrdma_buffer
,
1071 rb_refresh_worker
.work
);
1072 struct rpcrdma_xprt
*r_xprt
= container_of(buf
, struct rpcrdma_xprt
,
1075 rpcrdma_create_mrs(r_xprt
);
1078 struct rpcrdma_req
*
1079 rpcrdma_create_req(struct rpcrdma_xprt
*r_xprt
)
1081 struct rpcrdma_buffer
*buffer
= &r_xprt
->rx_buf
;
1082 struct rpcrdma_req
*req
;
1084 req
= kzalloc(sizeof(*req
), GFP_KERNEL
);
1086 return ERR_PTR(-ENOMEM
);
1088 spin_lock(&buffer
->rb_reqslock
);
1089 list_add(&req
->rl_all
, &buffer
->rb_allreqs
);
1090 spin_unlock(&buffer
->rb_reqslock
);
1091 req
->rl_buffer
= &r_xprt
->rx_buf
;
1092 INIT_LIST_HEAD(&req
->rl_registered
);
1096 struct rpcrdma_rep
*
1097 rpcrdma_create_rep(struct rpcrdma_xprt
*r_xprt
)
1099 struct rpcrdma_create_data_internal
*cdata
= &r_xprt
->rx_data
;
1100 struct rpcrdma_rep
*rep
;
1104 rep
= kzalloc(sizeof(*rep
), GFP_KERNEL
);
1108 rep
->rr_rdmabuf
= rpcrdma_alloc_regbuf(cdata
->inline_rsize
,
1109 DMA_FROM_DEVICE
, GFP_KERNEL
);
1110 if (IS_ERR(rep
->rr_rdmabuf
)) {
1111 rc
= PTR_ERR(rep
->rr_rdmabuf
);
1114 xdr_buf_init(&rep
->rr_hdrbuf
, rep
->rr_rdmabuf
->rg_base
,
1115 rdmab_length(rep
->rr_rdmabuf
));
1117 rep
->rr_cqe
.done
= rpcrdma_wc_receive
;
1118 rep
->rr_rxprt
= r_xprt
;
1119 INIT_WORK(&rep
->rr_work
, rpcrdma_deferred_completion
);
1120 rep
->rr_recv_wr
.next
= NULL
;
1121 rep
->rr_recv_wr
.wr_cqe
= &rep
->rr_cqe
;
1122 rep
->rr_recv_wr
.sg_list
= &rep
->rr_rdmabuf
->rg_iov
;
1123 rep
->rr_recv_wr
.num_sge
= 1;
1133 rpcrdma_buffer_create(struct rpcrdma_xprt
*r_xprt
)
1135 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1138 buf
->rb_max_requests
= r_xprt
->rx_data
.max_requests
;
1139 buf
->rb_bc_srv_max_requests
= 0;
1140 spin_lock_init(&buf
->rb_mwlock
);
1141 spin_lock_init(&buf
->rb_lock
);
1142 spin_lock_init(&buf
->rb_recovery_lock
);
1143 INIT_LIST_HEAD(&buf
->rb_mws
);
1144 INIT_LIST_HEAD(&buf
->rb_all
);
1145 INIT_LIST_HEAD(&buf
->rb_stale_mrs
);
1146 INIT_DELAYED_WORK(&buf
->rb_refresh_worker
,
1147 rpcrdma_mr_refresh_worker
);
1148 INIT_DELAYED_WORK(&buf
->rb_recovery_worker
,
1149 rpcrdma_mr_recovery_worker
);
1151 rpcrdma_create_mrs(r_xprt
);
1153 INIT_LIST_HEAD(&buf
->rb_send_bufs
);
1154 INIT_LIST_HEAD(&buf
->rb_allreqs
);
1155 spin_lock_init(&buf
->rb_reqslock
);
1156 for (i
= 0; i
< buf
->rb_max_requests
; i
++) {
1157 struct rpcrdma_req
*req
;
1159 req
= rpcrdma_create_req(r_xprt
);
1161 dprintk("RPC: %s: request buffer %d alloc"
1162 " failed\n", __func__
, i
);
1166 list_add(&req
->rl_list
, &buf
->rb_send_bufs
);
1169 INIT_LIST_HEAD(&buf
->rb_recv_bufs
);
1170 for (i
= 0; i
< buf
->rb_max_requests
+ RPCRDMA_MAX_BC_REQUESTS
; i
++) {
1171 struct rpcrdma_rep
*rep
;
1173 rep
= rpcrdma_create_rep(r_xprt
);
1175 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1180 list_add(&rep
->rr_list
, &buf
->rb_recv_bufs
);
1183 rc
= rpcrdma_sendctxs_create(r_xprt
);
1189 rpcrdma_buffer_destroy(buf
);
1193 static struct rpcrdma_req
*
1194 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer
*buf
)
1196 struct rpcrdma_req
*req
;
1198 req
= list_first_entry(&buf
->rb_send_bufs
,
1199 struct rpcrdma_req
, rl_list
);
1200 list_del_init(&req
->rl_list
);
1204 static struct rpcrdma_rep
*
1205 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer
*buf
)
1207 struct rpcrdma_rep
*rep
;
1209 rep
= list_first_entry(&buf
->rb_recv_bufs
,
1210 struct rpcrdma_rep
, rr_list
);
1211 list_del(&rep
->rr_list
);
1216 rpcrdma_destroy_rep(struct rpcrdma_rep
*rep
)
1218 rpcrdma_free_regbuf(rep
->rr_rdmabuf
);
1223 rpcrdma_destroy_req(struct rpcrdma_req
*req
)
1225 rpcrdma_free_regbuf(req
->rl_recvbuf
);
1226 rpcrdma_free_regbuf(req
->rl_sendbuf
);
1227 rpcrdma_free_regbuf(req
->rl_rdmabuf
);
1232 rpcrdma_destroy_mrs(struct rpcrdma_buffer
*buf
)
1234 struct rpcrdma_xprt
*r_xprt
= container_of(buf
, struct rpcrdma_xprt
,
1236 struct rpcrdma_ia
*ia
= rdmab_to_ia(buf
);
1237 struct rpcrdma_mw
*mw
;
1241 spin_lock(&buf
->rb_mwlock
);
1242 while (!list_empty(&buf
->rb_all
)) {
1243 mw
= list_entry(buf
->rb_all
.next
, struct rpcrdma_mw
, mw_all
);
1244 list_del(&mw
->mw_all
);
1246 spin_unlock(&buf
->rb_mwlock
);
1247 ia
->ri_ops
->ro_release_mr(mw
);
1249 spin_lock(&buf
->rb_mwlock
);
1251 spin_unlock(&buf
->rb_mwlock
);
1252 r_xprt
->rx_stats
.mrs_allocated
= 0;
1254 dprintk("RPC: %s: released %u MRs\n", __func__
, count
);
1258 rpcrdma_buffer_destroy(struct rpcrdma_buffer
*buf
)
1260 cancel_delayed_work_sync(&buf
->rb_recovery_worker
);
1261 cancel_delayed_work_sync(&buf
->rb_refresh_worker
);
1263 rpcrdma_sendctxs_destroy(buf
);
1265 while (!list_empty(&buf
->rb_recv_bufs
)) {
1266 struct rpcrdma_rep
*rep
;
1268 rep
= rpcrdma_buffer_get_rep_locked(buf
);
1269 rpcrdma_destroy_rep(rep
);
1271 buf
->rb_send_count
= 0;
1273 spin_lock(&buf
->rb_reqslock
);
1274 while (!list_empty(&buf
->rb_allreqs
)) {
1275 struct rpcrdma_req
*req
;
1277 req
= list_first_entry(&buf
->rb_allreqs
,
1278 struct rpcrdma_req
, rl_all
);
1279 list_del(&req
->rl_all
);
1281 spin_unlock(&buf
->rb_reqslock
);
1282 rpcrdma_destroy_req(req
);
1283 spin_lock(&buf
->rb_reqslock
);
1285 spin_unlock(&buf
->rb_reqslock
);
1286 buf
->rb_recv_count
= 0;
1288 rpcrdma_destroy_mrs(buf
);
1292 rpcrdma_get_mw(struct rpcrdma_xprt
*r_xprt
)
1294 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1295 struct rpcrdma_mw
*mw
= NULL
;
1297 spin_lock(&buf
->rb_mwlock
);
1298 if (!list_empty(&buf
->rb_mws
))
1299 mw
= rpcrdma_pop_mw(&buf
->rb_mws
);
1300 spin_unlock(&buf
->rb_mwlock
);
1308 dprintk("RPC: %s: no MWs available\n", __func__
);
1309 if (r_xprt
->rx_ep
.rep_connected
!= -ENODEV
)
1310 schedule_delayed_work(&buf
->rb_refresh_worker
, 0);
1312 /* Allow the reply handler and refresh worker to run */
1319 rpcrdma_put_mw(struct rpcrdma_xprt
*r_xprt
, struct rpcrdma_mw
*mw
)
1321 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1323 spin_lock(&buf
->rb_mwlock
);
1324 rpcrdma_push_mw(mw
, &buf
->rb_mws
);
1325 spin_unlock(&buf
->rb_mwlock
);
1328 static struct rpcrdma_rep
*
1329 rpcrdma_buffer_get_rep(struct rpcrdma_buffer
*buffers
)
1331 /* If an RPC previously completed without a reply (say, a
1332 * credential problem or a soft timeout occurs) then hold off
1333 * on supplying more Receive buffers until the number of new
1334 * pending RPCs catches up to the number of posted Receives.
1336 if (unlikely(buffers
->rb_send_count
< buffers
->rb_recv_count
))
1339 if (unlikely(list_empty(&buffers
->rb_recv_bufs
)))
1341 buffers
->rb_recv_count
++;
1342 return rpcrdma_buffer_get_rep_locked(buffers
);
1346 * Get a set of request/reply buffers.
1348 * Reply buffer (if available) is attached to send buffer upon return.
1350 struct rpcrdma_req
*
1351 rpcrdma_buffer_get(struct rpcrdma_buffer
*buffers
)
1353 struct rpcrdma_req
*req
;
1355 spin_lock(&buffers
->rb_lock
);
1356 if (list_empty(&buffers
->rb_send_bufs
))
1358 buffers
->rb_send_count
++;
1359 req
= rpcrdma_buffer_get_req_locked(buffers
);
1360 req
->rl_reply
= rpcrdma_buffer_get_rep(buffers
);
1361 spin_unlock(&buffers
->rb_lock
);
1365 spin_unlock(&buffers
->rb_lock
);
1366 pr_warn("RPC: %s: out of request buffers\n", __func__
);
1371 * Put request/reply buffers back into pool.
1372 * Pre-decrement counter/array index.
1375 rpcrdma_buffer_put(struct rpcrdma_req
*req
)
1377 struct rpcrdma_buffer
*buffers
= req
->rl_buffer
;
1378 struct rpcrdma_rep
*rep
= req
->rl_reply
;
1380 req
->rl_reply
= NULL
;
1382 spin_lock(&buffers
->rb_lock
);
1383 buffers
->rb_send_count
--;
1384 list_add_tail(&req
->rl_list
, &buffers
->rb_send_bufs
);
1386 buffers
->rb_recv_count
--;
1387 list_add_tail(&rep
->rr_list
, &buffers
->rb_recv_bufs
);
1389 spin_unlock(&buffers
->rb_lock
);
1393 * Recover reply buffers from pool.
1394 * This happens when recovering from disconnect.
1397 rpcrdma_recv_buffer_get(struct rpcrdma_req
*req
)
1399 struct rpcrdma_buffer
*buffers
= req
->rl_buffer
;
1401 spin_lock(&buffers
->rb_lock
);
1402 req
->rl_reply
= rpcrdma_buffer_get_rep(buffers
);
1403 spin_unlock(&buffers
->rb_lock
);
1407 * Put reply buffers back into pool when not attached to
1408 * request. This happens in error conditions.
1411 rpcrdma_recv_buffer_put(struct rpcrdma_rep
*rep
)
1413 struct rpcrdma_buffer
*buffers
= &rep
->rr_rxprt
->rx_buf
;
1415 spin_lock(&buffers
->rb_lock
);
1416 buffers
->rb_recv_count
--;
1417 list_add_tail(&rep
->rr_list
, &buffers
->rb_recv_bufs
);
1418 spin_unlock(&buffers
->rb_lock
);
1422 * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1423 * @size: size of buffer to be allocated, in bytes
1424 * @direction: direction of data movement
1427 * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1428 * can be persistently DMA-mapped for I/O.
1430 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1431 * receiving the payload of RDMA RECV operations. During Long Calls
1432 * or Replies they may be registered externally via ro_map.
1434 struct rpcrdma_regbuf
*
1435 rpcrdma_alloc_regbuf(size_t size
, enum dma_data_direction direction
,
1438 struct rpcrdma_regbuf
*rb
;
1440 rb
= kmalloc(sizeof(*rb
) + size
, flags
);
1442 return ERR_PTR(-ENOMEM
);
1444 rb
->rg_device
= NULL
;
1445 rb
->rg_direction
= direction
;
1446 rb
->rg_iov
.length
= size
;
1452 * __rpcrdma_map_regbuf - DMA-map a regbuf
1453 * @ia: controlling rpcrdma_ia
1454 * @rb: regbuf to be mapped
1457 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia
*ia
, struct rpcrdma_regbuf
*rb
)
1459 struct ib_device
*device
= ia
->ri_device
;
1461 if (rb
->rg_direction
== DMA_NONE
)
1464 rb
->rg_iov
.addr
= ib_dma_map_single(device
,
1465 (void *)rb
->rg_base
,
1468 if (ib_dma_mapping_error(device
, rdmab_addr(rb
)))
1471 rb
->rg_device
= device
;
1472 rb
->rg_iov
.lkey
= ia
->ri_pd
->local_dma_lkey
;
1477 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf
*rb
)
1479 if (!rpcrdma_regbuf_is_mapped(rb
))
1482 ib_dma_unmap_single(rb
->rg_device
, rdmab_addr(rb
),
1483 rdmab_length(rb
), rb
->rg_direction
);
1484 rb
->rg_device
= NULL
;
1488 * rpcrdma_free_regbuf - deregister and free registered buffer
1489 * @rb: regbuf to be deregistered and freed
1492 rpcrdma_free_regbuf(struct rpcrdma_regbuf
*rb
)
1497 rpcrdma_dma_unmap_regbuf(rb
);
1502 * Prepost any receive buffer, then post send.
1504 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1507 rpcrdma_ep_post(struct rpcrdma_ia
*ia
,
1508 struct rpcrdma_ep
*ep
,
1509 struct rpcrdma_req
*req
)
1511 struct ib_send_wr
*send_wr
= &req
->rl_sendctx
->sc_wr
;
1512 struct ib_send_wr
*send_wr_fail
;
1515 if (req
->rl_reply
) {
1516 rc
= rpcrdma_ep_post_recv(ia
, req
->rl_reply
);
1519 req
->rl_reply
= NULL
;
1522 dprintk("RPC: %s: posting %d s/g entries\n",
1523 __func__
, send_wr
->num_sge
);
1525 if (!ep
->rep_send_count
||
1526 test_bit(RPCRDMA_REQ_F_TX_RESOURCES
, &req
->rl_flags
)) {
1527 send_wr
->send_flags
|= IB_SEND_SIGNALED
;
1528 ep
->rep_send_count
= ep
->rep_send_batch
;
1530 send_wr
->send_flags
&= ~IB_SEND_SIGNALED
;
1531 --ep
->rep_send_count
;
1533 rc
= ib_post_send(ia
->ri_id
->qp
, send_wr
, &send_wr_fail
);
1535 goto out_postsend_err
;
1539 pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc
);
1544 rpcrdma_ep_post_recv(struct rpcrdma_ia
*ia
,
1545 struct rpcrdma_rep
*rep
)
1547 struct ib_recv_wr
*recv_wr_fail
;
1550 if (!rpcrdma_dma_map_regbuf(ia
, rep
->rr_rdmabuf
))
1552 rc
= ib_post_recv(ia
->ri_id
->qp
, &rep
->rr_recv_wr
, &recv_wr_fail
);
1558 pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1562 pr_err("rpcrdma: ib_post_recv returned %i\n", rc
);
1567 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1568 * @r_xprt: transport associated with these backchannel resources
1569 * @min_reqs: minimum number of incoming requests expected
1571 * Returns zero if all requested buffers were posted, or a negative errno.
1574 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*r_xprt
, unsigned int count
)
1576 struct rpcrdma_buffer
*buffers
= &r_xprt
->rx_buf
;
1577 struct rpcrdma_ia
*ia
= &r_xprt
->rx_ia
;
1578 struct rpcrdma_rep
*rep
;
1582 spin_lock(&buffers
->rb_lock
);
1583 if (list_empty(&buffers
->rb_recv_bufs
))
1585 rep
= rpcrdma_buffer_get_rep_locked(buffers
);
1586 spin_unlock(&buffers
->rb_lock
);
1588 rc
= rpcrdma_ep_post_recv(ia
, rep
);
1596 spin_unlock(&buffers
->rb_lock
);
1597 pr_warn("%s: no extra receive buffers\n", __func__
);
1601 rpcrdma_recv_buffer_put(rep
);