2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
43 * Encapsulates the major functions managing:
50 #include <linux/interrupt.h>
51 #include <linux/slab.h>
52 #include <linux/prefetch.h>
53 #include <linux/sunrpc/addr.h>
54 #include <asm/bitops.h>
55 #include <linux/module.h> /* try_module_get()/module_put() */
57 #include "xprt_rdma.h"
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY RPCDBG_TRANS
71 static struct workqueue_struct
*rpcrdma_receive_wq
;
74 rpcrdma_alloc_wq(void)
76 struct workqueue_struct
*recv_wq
;
78 recv_wq
= alloc_workqueue("xprtrdma_receive",
79 WQ_MEM_RECLAIM
| WQ_UNBOUND
| WQ_HIGHPRI
,
84 rpcrdma_receive_wq
= recv_wq
;
89 rpcrdma_destroy_wq(void)
91 struct workqueue_struct
*wq
;
93 if (rpcrdma_receive_wq
) {
94 wq
= rpcrdma_receive_wq
;
95 rpcrdma_receive_wq
= NULL
;
96 destroy_workqueue(wq
);
101 rpcrdma_qp_async_error_upcall(struct ib_event
*event
, void *context
)
103 struct rpcrdma_ep
*ep
= context
;
105 pr_err("RPC: %s: %s on device %s ep %p\n",
106 __func__
, ib_event_msg(event
->event
),
107 event
->device
->name
, context
);
108 if (ep
->rep_connected
== 1) {
109 ep
->rep_connected
= -EIO
;
110 rpcrdma_conn_func(ep
);
111 wake_up_all(&ep
->rep_connect_wait
);
116 rpcrdma_cq_async_error_upcall(struct ib_event
*event
, void *context
)
118 struct rpcrdma_ep
*ep
= context
;
120 pr_err("RPC: %s: %s on device %s ep %p\n",
121 __func__
, ib_event_msg(event
->event
),
122 event
->device
->name
, context
);
123 if (ep
->rep_connected
== 1) {
124 ep
->rep_connected
= -EIO
;
125 rpcrdma_conn_func(ep
);
126 wake_up_all(&ep
->rep_connect_wait
);
131 rpcrdma_sendcq_process_wc(struct ib_wc
*wc
)
133 /* WARNING: Only wr_id and status are reliable at this point */
134 if (wc
->wr_id
== RPCRDMA_IGNORE_COMPLETION
) {
135 if (wc
->status
!= IB_WC_SUCCESS
&&
136 wc
->status
!= IB_WC_WR_FLUSH_ERR
)
137 pr_err("RPC: %s: SEND: %s\n",
138 __func__
, ib_wc_status_msg(wc
->status
));
140 struct rpcrdma_mw
*r
;
142 r
= (struct rpcrdma_mw
*)(unsigned long)wc
->wr_id
;
143 r
->mw_sendcompletion(wc
);
147 /* The common case is a single send completion is waiting. By
148 * passing two WC entries to ib_poll_cq, a return code of 1
149 * means there is exactly one WC waiting and no more. We don't
150 * have to invoke ib_poll_cq again to know that the CQ has been
154 rpcrdma_sendcq_poll(struct ib_cq
*cq
)
156 struct ib_wc
*pos
, wcs
[2];
162 rc
= ib_poll_cq(cq
, ARRAY_SIZE(wcs
), pos
);
168 rpcrdma_sendcq_process_wc(pos
++);
169 } while (rc
== ARRAY_SIZE(wcs
));
173 /* Handle provider send completion upcalls.
176 rpcrdma_sendcq_upcall(struct ib_cq
*cq
, void *cq_context
)
179 rpcrdma_sendcq_poll(cq
);
180 } while (ib_req_notify_cq(cq
, IB_CQ_NEXT_COMP
|
181 IB_CQ_REPORT_MISSED_EVENTS
) > 0);
185 rpcrdma_receive_worker(struct work_struct
*work
)
187 struct rpcrdma_rep
*rep
=
188 container_of(work
, struct rpcrdma_rep
, rr_work
);
190 rpcrdma_reply_handler(rep
);
194 rpcrdma_recvcq_process_wc(struct ib_wc
*wc
)
196 struct rpcrdma_rep
*rep
=
197 (struct rpcrdma_rep
*)(unsigned long)wc
->wr_id
;
199 /* WARNING: Only wr_id and status are reliable at this point */
200 if (wc
->status
!= IB_WC_SUCCESS
)
203 /* status == SUCCESS means all fields in wc are trustworthy */
204 if (wc
->opcode
!= IB_WC_RECV
)
207 dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n",
208 __func__
, rep
, wc
->byte_len
);
210 rep
->rr_len
= wc
->byte_len
;
211 ib_dma_sync_single_for_cpu(rep
->rr_device
,
212 rdmab_addr(rep
->rr_rdmabuf
),
213 rep
->rr_len
, DMA_FROM_DEVICE
);
214 prefetch(rdmab_to_msg(rep
->rr_rdmabuf
));
217 queue_work(rpcrdma_receive_wq
, &rep
->rr_work
);
221 if (wc
->status
!= IB_WC_WR_FLUSH_ERR
)
222 pr_err("RPC: %s: rep %p: %s\n",
223 __func__
, rep
, ib_wc_status_msg(wc
->status
));
224 rep
->rr_len
= RPCRDMA_BAD_LEN
;
228 /* The wc array is on stack: automatic memory is always CPU-local.
230 * struct ib_wc is 64 bytes, making the poll array potentially
231 * large. But this is at the bottom of the call chain. Further
232 * substantial work is done in another thread.
235 rpcrdma_recvcq_poll(struct ib_cq
*cq
)
237 struct ib_wc
*pos
, wcs
[4];
243 rc
= ib_poll_cq(cq
, ARRAY_SIZE(wcs
), pos
);
249 rpcrdma_recvcq_process_wc(pos
++);
250 } while (rc
== ARRAY_SIZE(wcs
));
253 /* Handle provider receive completion upcalls.
256 rpcrdma_recvcq_upcall(struct ib_cq
*cq
, void *cq_context
)
259 rpcrdma_recvcq_poll(cq
);
260 } while (ib_req_notify_cq(cq
, IB_CQ_NEXT_COMP
|
261 IB_CQ_REPORT_MISSED_EVENTS
) > 0);
265 rpcrdma_flush_cqs(struct rpcrdma_ep
*ep
)
269 while (ib_poll_cq(ep
->rep_attr
.recv_cq
, 1, &wc
) > 0)
270 rpcrdma_recvcq_process_wc(&wc
);
271 while (ib_poll_cq(ep
->rep_attr
.send_cq
, 1, &wc
) > 0)
272 rpcrdma_sendcq_process_wc(&wc
);
276 rpcrdma_conn_upcall(struct rdma_cm_id
*id
, struct rdma_cm_event
*event
)
278 struct rpcrdma_xprt
*xprt
= id
->context
;
279 struct rpcrdma_ia
*ia
= &xprt
->rx_ia
;
280 struct rpcrdma_ep
*ep
= &xprt
->rx_ep
;
281 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
282 struct sockaddr
*sap
= (struct sockaddr
*)&ep
->rep_remote_addr
;
284 struct ib_qp_attr
*attr
= &ia
->ri_qp_attr
;
285 struct ib_qp_init_attr
*iattr
= &ia
->ri_qp_init_attr
;
288 switch (event
->event
) {
289 case RDMA_CM_EVENT_ADDR_RESOLVED
:
290 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
292 complete(&ia
->ri_done
);
294 case RDMA_CM_EVENT_ADDR_ERROR
:
295 ia
->ri_async_rc
= -EHOSTUNREACH
;
296 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
298 complete(&ia
->ri_done
);
300 case RDMA_CM_EVENT_ROUTE_ERROR
:
301 ia
->ri_async_rc
= -ENETUNREACH
;
302 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
304 complete(&ia
->ri_done
);
306 case RDMA_CM_EVENT_ESTABLISHED
:
308 ib_query_qp(ia
->ri_id
->qp
, attr
,
309 IB_QP_MAX_QP_RD_ATOMIC
| IB_QP_MAX_DEST_RD_ATOMIC
,
311 dprintk("RPC: %s: %d responder resources"
313 __func__
, attr
->max_dest_rd_atomic
,
314 attr
->max_rd_atomic
);
316 case RDMA_CM_EVENT_CONNECT_ERROR
:
317 connstate
= -ENOTCONN
;
319 case RDMA_CM_EVENT_UNREACHABLE
:
320 connstate
= -ENETDOWN
;
322 case RDMA_CM_EVENT_REJECTED
:
323 connstate
= -ECONNREFUSED
;
325 case RDMA_CM_EVENT_DISCONNECTED
:
326 connstate
= -ECONNABORTED
;
328 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
331 dprintk("RPC: %s: %sconnected\n",
332 __func__
, connstate
> 0 ? "" : "dis");
333 ep
->rep_connected
= connstate
;
334 rpcrdma_conn_func(ep
);
335 wake_up_all(&ep
->rep_connect_wait
);
338 dprintk("RPC: %s: %pIS:%u (ep 0x%p): %s\n",
339 __func__
, sap
, rpc_get_port(sap
), ep
,
340 rdma_event_msg(event
->event
));
344 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
345 if (connstate
== 1) {
346 int ird
= attr
->max_dest_rd_atomic
;
347 int tird
= ep
->rep_remote_cma
.responder_resources
;
349 pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
350 sap
, rpc_get_port(sap
),
352 ia
->ri_ops
->ro_displayname
,
353 xprt
->rx_buf
.rb_max_requests
,
354 ird
, ird
< 4 && ird
< tird
/ 2 ? " (low!)" : "");
355 } else if (connstate
< 0) {
356 pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
357 sap
, rpc_get_port(sap
), connstate
);
364 static void rpcrdma_destroy_id(struct rdma_cm_id
*id
)
367 module_put(id
->device
->owner
);
372 static struct rdma_cm_id
*
373 rpcrdma_create_id(struct rpcrdma_xprt
*xprt
,
374 struct rpcrdma_ia
*ia
, struct sockaddr
*addr
)
376 struct rdma_cm_id
*id
;
379 init_completion(&ia
->ri_done
);
381 id
= rdma_create_id(&init_net
, rpcrdma_conn_upcall
, xprt
, RDMA_PS_TCP
,
385 dprintk("RPC: %s: rdma_create_id() failed %i\n",
390 ia
->ri_async_rc
= -ETIMEDOUT
;
391 rc
= rdma_resolve_addr(id
, NULL
, addr
, RDMA_RESOLVE_TIMEOUT
);
393 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
397 wait_for_completion_interruptible_timeout(&ia
->ri_done
,
398 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT
) + 1);
401 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
402 * be pinned while there are active NFS/RDMA mounts to prevent
403 * hangs and crashes at umount time.
405 if (!ia
->ri_async_rc
&& !try_module_get(id
->device
->owner
)) {
406 dprintk("RPC: %s: Failed to get device module\n",
408 ia
->ri_async_rc
= -ENODEV
;
410 rc
= ia
->ri_async_rc
;
414 ia
->ri_async_rc
= -ETIMEDOUT
;
415 rc
= rdma_resolve_route(id
, RDMA_RESOLVE_TIMEOUT
);
417 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
421 wait_for_completion_interruptible_timeout(&ia
->ri_done
,
422 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT
) + 1);
423 rc
= ia
->ri_async_rc
;
429 module_put(id
->device
->owner
);
436 * Drain any cq, prior to teardown.
439 rpcrdma_clean_cq(struct ib_cq
*cq
)
444 while (1 == ib_poll_cq(cq
, 1, &wc
))
448 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
449 __func__
, count
, wc
.opcode
);
453 * Exported functions.
457 * Open and initialize an Interface Adapter.
458 * o initializes fields of struct rpcrdma_ia, including
459 * interface and provider attributes and protection zone.
462 rpcrdma_ia_open(struct rpcrdma_xprt
*xprt
, struct sockaddr
*addr
, int memreg
)
464 struct rpcrdma_ia
*ia
= &xprt
->rx_ia
;
467 ia
->ri_dma_mr
= NULL
;
469 ia
->ri_id
= rpcrdma_create_id(xprt
, ia
, addr
);
470 if (IS_ERR(ia
->ri_id
)) {
471 rc
= PTR_ERR(ia
->ri_id
);
474 ia
->ri_device
= ia
->ri_id
->device
;
476 ia
->ri_pd
= ib_alloc_pd(ia
->ri_device
);
477 if (IS_ERR(ia
->ri_pd
)) {
478 rc
= PTR_ERR(ia
->ri_pd
);
479 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
484 if (memreg
== RPCRDMA_FRMR
) {
485 if (!(ia
->ri_device
->attrs
.device_cap_flags
&
486 IB_DEVICE_MEM_MGT_EXTENSIONS
) ||
487 (ia
->ri_device
->attrs
.max_fast_reg_page_list_len
== 0)) {
488 dprintk("RPC: %s: FRMR registration "
489 "not supported by HCA\n", __func__
);
490 memreg
= RPCRDMA_MTHCAFMR
;
493 if (memreg
== RPCRDMA_MTHCAFMR
) {
494 if (!ia
->ri_device
->alloc_fmr
) {
495 dprintk("RPC: %s: MTHCAFMR registration "
496 "not supported by HCA\n", __func__
);
504 ia
->ri_ops
= &rpcrdma_frwr_memreg_ops
;
506 case RPCRDMA_ALLPHYSICAL
:
507 ia
->ri_ops
= &rpcrdma_physical_memreg_ops
;
509 case RPCRDMA_MTHCAFMR
:
510 ia
->ri_ops
= &rpcrdma_fmr_memreg_ops
;
513 printk(KERN_ERR
"RPC: Unsupported memory "
514 "registration mode: %d\n", memreg
);
518 dprintk("RPC: %s: memory registration strategy is '%s'\n",
519 __func__
, ia
->ri_ops
->ro_displayname
);
521 rwlock_init(&ia
->ri_qplock
);
525 ib_dealloc_pd(ia
->ri_pd
);
528 rpcrdma_destroy_id(ia
->ri_id
);
535 * Clean up/close an IA.
536 * o if event handles and PD have been initialized, free them.
540 rpcrdma_ia_close(struct rpcrdma_ia
*ia
)
542 dprintk("RPC: %s: entering\n", __func__
);
543 if (ia
->ri_id
!= NULL
&& !IS_ERR(ia
->ri_id
)) {
545 rdma_destroy_qp(ia
->ri_id
);
546 rpcrdma_destroy_id(ia
->ri_id
);
550 /* If the pd is still busy, xprtrdma missed freeing a resource */
551 if (ia
->ri_pd
&& !IS_ERR(ia
->ri_pd
))
552 ib_dealloc_pd(ia
->ri_pd
);
556 * Create unconnected endpoint.
559 rpcrdma_ep_create(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
,
560 struct rpcrdma_create_data_internal
*cdata
)
562 struct ib_cq
*sendcq
, *recvcq
;
563 struct ib_cq_init_attr cq_attr
= {};
564 unsigned int max_qp_wr
;
567 if (ia
->ri_device
->attrs
.max_sge
< RPCRDMA_MAX_IOVS
) {
568 dprintk("RPC: %s: insufficient sge's available\n",
573 if (ia
->ri_device
->attrs
.max_qp_wr
<= RPCRDMA_BACKWARD_WRS
) {
574 dprintk("RPC: %s: insufficient wqe's available\n",
578 max_qp_wr
= ia
->ri_device
->attrs
.max_qp_wr
- RPCRDMA_BACKWARD_WRS
;
580 /* check provider's send/recv wr limits */
581 if (cdata
->max_requests
> max_qp_wr
)
582 cdata
->max_requests
= max_qp_wr
;
584 ep
->rep_attr
.event_handler
= rpcrdma_qp_async_error_upcall
;
585 ep
->rep_attr
.qp_context
= ep
;
586 ep
->rep_attr
.srq
= NULL
;
587 ep
->rep_attr
.cap
.max_send_wr
= cdata
->max_requests
;
588 ep
->rep_attr
.cap
.max_send_wr
+= RPCRDMA_BACKWARD_WRS
;
589 rc
= ia
->ri_ops
->ro_open(ia
, ep
, cdata
);
592 ep
->rep_attr
.cap
.max_recv_wr
= cdata
->max_requests
;
593 ep
->rep_attr
.cap
.max_recv_wr
+= RPCRDMA_BACKWARD_WRS
;
594 ep
->rep_attr
.cap
.max_send_sge
= RPCRDMA_MAX_IOVS
;
595 ep
->rep_attr
.cap
.max_recv_sge
= 1;
596 ep
->rep_attr
.cap
.max_inline_data
= 0;
597 ep
->rep_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
598 ep
->rep_attr
.qp_type
= IB_QPT_RC
;
599 ep
->rep_attr
.port_num
= ~0;
601 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
602 "iovs: send %d recv %d\n",
604 ep
->rep_attr
.cap
.max_send_wr
,
605 ep
->rep_attr
.cap
.max_recv_wr
,
606 ep
->rep_attr
.cap
.max_send_sge
,
607 ep
->rep_attr
.cap
.max_recv_sge
);
609 /* set trigger for requesting send completion */
610 ep
->rep_cqinit
= ep
->rep_attr
.cap
.max_send_wr
/2 - 1;
611 if (ep
->rep_cqinit
> RPCRDMA_MAX_UNSIGNALED_SENDS
)
612 ep
->rep_cqinit
= RPCRDMA_MAX_UNSIGNALED_SENDS
;
613 else if (ep
->rep_cqinit
<= 2)
616 init_waitqueue_head(&ep
->rep_connect_wait
);
617 INIT_DELAYED_WORK(&ep
->rep_connect_worker
, rpcrdma_connect_worker
);
619 cq_attr
.cqe
= ep
->rep_attr
.cap
.max_send_wr
+ 1;
620 sendcq
= ib_create_cq(ia
->ri_device
, rpcrdma_sendcq_upcall
,
621 rpcrdma_cq_async_error_upcall
, NULL
, &cq_attr
);
622 if (IS_ERR(sendcq
)) {
623 rc
= PTR_ERR(sendcq
);
624 dprintk("RPC: %s: failed to create send CQ: %i\n",
629 rc
= ib_req_notify_cq(sendcq
, IB_CQ_NEXT_COMP
);
631 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
636 cq_attr
.cqe
= ep
->rep_attr
.cap
.max_recv_wr
+ 1;
637 recvcq
= ib_create_cq(ia
->ri_device
, rpcrdma_recvcq_upcall
,
638 rpcrdma_cq_async_error_upcall
, NULL
, &cq_attr
);
639 if (IS_ERR(recvcq
)) {
640 rc
= PTR_ERR(recvcq
);
641 dprintk("RPC: %s: failed to create recv CQ: %i\n",
646 rc
= ib_req_notify_cq(recvcq
, IB_CQ_NEXT_COMP
);
648 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
650 ib_destroy_cq(recvcq
);
654 ep
->rep_attr
.send_cq
= sendcq
;
655 ep
->rep_attr
.recv_cq
= recvcq
;
657 /* Initialize cma parameters */
659 /* RPC/RDMA does not use private data */
660 ep
->rep_remote_cma
.private_data
= NULL
;
661 ep
->rep_remote_cma
.private_data_len
= 0;
663 /* Client offers RDMA Read but does not initiate */
664 ep
->rep_remote_cma
.initiator_depth
= 0;
665 if (ia
->ri_device
->attrs
.max_qp_rd_atom
> 32) /* arbitrary but <= 255 */
666 ep
->rep_remote_cma
.responder_resources
= 32;
668 ep
->rep_remote_cma
.responder_resources
=
669 ia
->ri_device
->attrs
.max_qp_rd_atom
;
671 ep
->rep_remote_cma
.retry_count
= 7;
672 ep
->rep_remote_cma
.flow_control
= 0;
673 ep
->rep_remote_cma
.rnr_retry_count
= 0;
678 err
= ib_destroy_cq(sendcq
);
680 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
684 ib_dereg_mr(ia
->ri_dma_mr
);
691 * Disconnect and destroy endpoint. After this, the only
692 * valid operations on the ep are to free it (if dynamically
693 * allocated) or re-create it.
696 rpcrdma_ep_destroy(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
700 dprintk("RPC: %s: entering, connected is %d\n",
701 __func__
, ep
->rep_connected
);
703 cancel_delayed_work_sync(&ep
->rep_connect_worker
);
706 rpcrdma_ep_disconnect(ep
, ia
);
708 rpcrdma_clean_cq(ep
->rep_attr
.recv_cq
);
709 rpcrdma_clean_cq(ep
->rep_attr
.send_cq
);
712 rdma_destroy_qp(ia
->ri_id
);
713 ia
->ri_id
->qp
= NULL
;
716 rc
= ib_destroy_cq(ep
->rep_attr
.recv_cq
);
718 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
721 rc
= ib_destroy_cq(ep
->rep_attr
.send_cq
);
723 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
727 rc
= ib_dereg_mr(ia
->ri_dma_mr
);
728 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
734 * Connect unconnected endpoint.
737 rpcrdma_ep_connect(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
739 struct rdma_cm_id
*id
, *old
;
743 if (ep
->rep_connected
!= 0) {
744 struct rpcrdma_xprt
*xprt
;
746 dprintk("RPC: %s: reconnecting...\n", __func__
);
748 rpcrdma_ep_disconnect(ep
, ia
);
749 rpcrdma_flush_cqs(ep
);
751 xprt
= container_of(ia
, struct rpcrdma_xprt
, rx_ia
);
752 id
= rpcrdma_create_id(xprt
, ia
,
753 (struct sockaddr
*)&xprt
->rx_data
.addr
);
758 /* TEMP TEMP TEMP - fail if new device:
759 * Deregister/remarshal *all* requests!
760 * Close and recreate adapter, pd, etc!
761 * Re-determine all attributes still sane!
762 * More stuff I haven't thought of!
765 if (ia
->ri_device
!= id
->device
) {
766 printk("RPC: %s: can't reconnect on "
767 "different device!\n", __func__
);
768 rpcrdma_destroy_id(id
);
773 rc
= rdma_create_qp(id
, ia
->ri_pd
, &ep
->rep_attr
);
775 dprintk("RPC: %s: rdma_create_qp failed %i\n",
777 rpcrdma_destroy_id(id
);
782 write_lock(&ia
->ri_qplock
);
785 write_unlock(&ia
->ri_qplock
);
787 rdma_destroy_qp(old
);
788 rpcrdma_destroy_id(old
);
790 dprintk("RPC: %s: connecting...\n", __func__
);
791 rc
= rdma_create_qp(ia
->ri_id
, ia
->ri_pd
, &ep
->rep_attr
);
793 dprintk("RPC: %s: rdma_create_qp failed %i\n",
795 /* do not update ep->rep_connected */
800 ep
->rep_connected
= 0;
802 rc
= rdma_connect(ia
->ri_id
, &ep
->rep_remote_cma
);
804 dprintk("RPC: %s: rdma_connect() failed with %i\n",
809 wait_event_interruptible(ep
->rep_connect_wait
, ep
->rep_connected
!= 0);
812 * Check state. A non-peer reject indicates no listener
813 * (ECONNREFUSED), which may be a transient state. All
814 * others indicate a transport condition which has already
815 * undergone a best-effort.
817 if (ep
->rep_connected
== -ECONNREFUSED
&&
818 ++retry_count
<= RDMA_CONNECT_RETRY_MAX
) {
819 dprintk("RPC: %s: non-peer_reject, retry\n", __func__
);
822 if (ep
->rep_connected
<= 0) {
823 /* Sometimes, the only way to reliably connect to remote
824 * CMs is to use same nonzero values for ORD and IRD. */
825 if (retry_count
++ <= RDMA_CONNECT_RETRY_MAX
+ 1 &&
826 (ep
->rep_remote_cma
.responder_resources
== 0 ||
827 ep
->rep_remote_cma
.initiator_depth
!=
828 ep
->rep_remote_cma
.responder_resources
)) {
829 if (ep
->rep_remote_cma
.responder_resources
== 0)
830 ep
->rep_remote_cma
.responder_resources
= 1;
831 ep
->rep_remote_cma
.initiator_depth
=
832 ep
->rep_remote_cma
.responder_resources
;
835 rc
= ep
->rep_connected
;
837 struct rpcrdma_xprt
*r_xprt
;
840 dprintk("RPC: %s: connected\n", __func__
);
842 r_xprt
= container_of(ia
, struct rpcrdma_xprt
, rx_ia
);
843 extras
= r_xprt
->rx_buf
.rb_bc_srv_max_requests
;
846 rc
= rpcrdma_ep_post_extra_recv(r_xprt
, extras
);
848 pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
856 ep
->rep_connected
= rc
;
861 * rpcrdma_ep_disconnect
863 * This is separate from destroy to facilitate the ability
864 * to reconnect without recreating the endpoint.
866 * This call is not reentrant, and must not be made in parallel
867 * on the same endpoint.
870 rpcrdma_ep_disconnect(struct rpcrdma_ep
*ep
, struct rpcrdma_ia
*ia
)
874 rpcrdma_flush_cqs(ep
);
875 rc
= rdma_disconnect(ia
->ri_id
);
877 /* returns without wait if not connected */
878 wait_event_interruptible(ep
->rep_connect_wait
,
879 ep
->rep_connected
!= 1);
880 dprintk("RPC: %s: after wait, %sconnected\n", __func__
,
881 (ep
->rep_connected
== 1) ? "still " : "dis");
883 dprintk("RPC: %s: rdma_disconnect %i\n", __func__
, rc
);
884 ep
->rep_connected
= rc
;
889 rpcrdma_create_req(struct rpcrdma_xprt
*r_xprt
)
891 struct rpcrdma_buffer
*buffer
= &r_xprt
->rx_buf
;
892 struct rpcrdma_req
*req
;
894 req
= kzalloc(sizeof(*req
), GFP_KERNEL
);
896 return ERR_PTR(-ENOMEM
);
898 INIT_LIST_HEAD(&req
->rl_free
);
899 spin_lock(&buffer
->rb_reqslock
);
900 list_add(&req
->rl_all
, &buffer
->rb_allreqs
);
901 spin_unlock(&buffer
->rb_reqslock
);
902 req
->rl_buffer
= &r_xprt
->rx_buf
;
907 rpcrdma_create_rep(struct rpcrdma_xprt
*r_xprt
)
909 struct rpcrdma_create_data_internal
*cdata
= &r_xprt
->rx_data
;
910 struct rpcrdma_ia
*ia
= &r_xprt
->rx_ia
;
911 struct rpcrdma_rep
*rep
;
915 rep
= kzalloc(sizeof(*rep
), GFP_KERNEL
);
919 rep
->rr_rdmabuf
= rpcrdma_alloc_regbuf(ia
, cdata
->inline_rsize
,
921 if (IS_ERR(rep
->rr_rdmabuf
)) {
922 rc
= PTR_ERR(rep
->rr_rdmabuf
);
926 rep
->rr_device
= ia
->ri_device
;
927 rep
->rr_rxprt
= r_xprt
;
928 INIT_WORK(&rep
->rr_work
, rpcrdma_receive_worker
);
938 rpcrdma_buffer_create(struct rpcrdma_xprt
*r_xprt
)
940 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
941 struct rpcrdma_ia
*ia
= &r_xprt
->rx_ia
;
944 buf
->rb_max_requests
= r_xprt
->rx_data
.max_requests
;
945 buf
->rb_bc_srv_max_requests
= 0;
946 spin_lock_init(&buf
->rb_lock
);
948 rc
= ia
->ri_ops
->ro_init(r_xprt
);
952 INIT_LIST_HEAD(&buf
->rb_send_bufs
);
953 INIT_LIST_HEAD(&buf
->rb_allreqs
);
954 spin_lock_init(&buf
->rb_reqslock
);
955 for (i
= 0; i
< buf
->rb_max_requests
; i
++) {
956 struct rpcrdma_req
*req
;
958 req
= rpcrdma_create_req(r_xprt
);
960 dprintk("RPC: %s: request buffer %d alloc"
961 " failed\n", __func__
, i
);
965 req
->rl_backchannel
= false;
966 list_add(&req
->rl_free
, &buf
->rb_send_bufs
);
969 INIT_LIST_HEAD(&buf
->rb_recv_bufs
);
970 for (i
= 0; i
< buf
->rb_max_requests
+ 2; i
++) {
971 struct rpcrdma_rep
*rep
;
973 rep
= rpcrdma_create_rep(r_xprt
);
975 dprintk("RPC: %s: reply buffer %d alloc failed\n",
980 list_add(&rep
->rr_list
, &buf
->rb_recv_bufs
);
985 rpcrdma_buffer_destroy(buf
);
989 static struct rpcrdma_req
*
990 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer
*buf
)
992 struct rpcrdma_req
*req
;
994 req
= list_first_entry(&buf
->rb_send_bufs
,
995 struct rpcrdma_req
, rl_free
);
996 list_del(&req
->rl_free
);
1000 static struct rpcrdma_rep
*
1001 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer
*buf
)
1003 struct rpcrdma_rep
*rep
;
1005 rep
= list_first_entry(&buf
->rb_recv_bufs
,
1006 struct rpcrdma_rep
, rr_list
);
1007 list_del(&rep
->rr_list
);
1012 rpcrdma_destroy_rep(struct rpcrdma_ia
*ia
, struct rpcrdma_rep
*rep
)
1014 rpcrdma_free_regbuf(ia
, rep
->rr_rdmabuf
);
1019 rpcrdma_destroy_req(struct rpcrdma_ia
*ia
, struct rpcrdma_req
*req
)
1021 rpcrdma_free_regbuf(ia
, req
->rl_sendbuf
);
1022 rpcrdma_free_regbuf(ia
, req
->rl_rdmabuf
);
1027 rpcrdma_buffer_destroy(struct rpcrdma_buffer
*buf
)
1029 struct rpcrdma_ia
*ia
= rdmab_to_ia(buf
);
1031 while (!list_empty(&buf
->rb_recv_bufs
)) {
1032 struct rpcrdma_rep
*rep
;
1034 rep
= rpcrdma_buffer_get_rep_locked(buf
);
1035 rpcrdma_destroy_rep(ia
, rep
);
1038 spin_lock(&buf
->rb_reqslock
);
1039 while (!list_empty(&buf
->rb_allreqs
)) {
1040 struct rpcrdma_req
*req
;
1042 req
= list_first_entry(&buf
->rb_allreqs
,
1043 struct rpcrdma_req
, rl_all
);
1044 list_del(&req
->rl_all
);
1046 spin_unlock(&buf
->rb_reqslock
);
1047 rpcrdma_destroy_req(ia
, req
);
1048 spin_lock(&buf
->rb_reqslock
);
1050 spin_unlock(&buf
->rb_reqslock
);
1052 ia
->ri_ops
->ro_destroy(buf
);
1056 rpcrdma_get_mw(struct rpcrdma_xprt
*r_xprt
)
1058 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1059 struct rpcrdma_mw
*mw
= NULL
;
1061 spin_lock(&buf
->rb_mwlock
);
1062 if (!list_empty(&buf
->rb_mws
)) {
1063 mw
= list_first_entry(&buf
->rb_mws
,
1064 struct rpcrdma_mw
, mw_list
);
1065 list_del_init(&mw
->mw_list
);
1067 spin_unlock(&buf
->rb_mwlock
);
1070 pr_err("RPC: %s: no MWs available\n", __func__
);
1075 rpcrdma_put_mw(struct rpcrdma_xprt
*r_xprt
, struct rpcrdma_mw
*mw
)
1077 struct rpcrdma_buffer
*buf
= &r_xprt
->rx_buf
;
1079 spin_lock(&buf
->rb_mwlock
);
1080 list_add_tail(&mw
->mw_list
, &buf
->rb_mws
);
1081 spin_unlock(&buf
->rb_mwlock
);
1085 * Get a set of request/reply buffers.
1087 * Reply buffer (if available) is attached to send buffer upon return.
1089 struct rpcrdma_req
*
1090 rpcrdma_buffer_get(struct rpcrdma_buffer
*buffers
)
1092 struct rpcrdma_req
*req
;
1094 spin_lock(&buffers
->rb_lock
);
1095 if (list_empty(&buffers
->rb_send_bufs
))
1097 req
= rpcrdma_buffer_get_req_locked(buffers
);
1098 if (list_empty(&buffers
->rb_recv_bufs
))
1100 req
->rl_reply
= rpcrdma_buffer_get_rep_locked(buffers
);
1101 spin_unlock(&buffers
->rb_lock
);
1105 spin_unlock(&buffers
->rb_lock
);
1106 pr_warn("RPC: %s: out of request buffers\n", __func__
);
1109 spin_unlock(&buffers
->rb_lock
);
1110 pr_warn("RPC: %s: out of reply buffers\n", __func__
);
1111 req
->rl_reply
= NULL
;
1116 * Put request/reply buffers back into pool.
1117 * Pre-decrement counter/array index.
1120 rpcrdma_buffer_put(struct rpcrdma_req
*req
)
1122 struct rpcrdma_buffer
*buffers
= req
->rl_buffer
;
1123 struct rpcrdma_rep
*rep
= req
->rl_reply
;
1126 req
->rl_reply
= NULL
;
1128 spin_lock(&buffers
->rb_lock
);
1129 list_add_tail(&req
->rl_free
, &buffers
->rb_send_bufs
);
1131 list_add_tail(&rep
->rr_list
, &buffers
->rb_recv_bufs
);
1132 spin_unlock(&buffers
->rb_lock
);
1136 * Recover reply buffers from pool.
1137 * This happens when recovering from disconnect.
1140 rpcrdma_recv_buffer_get(struct rpcrdma_req
*req
)
1142 struct rpcrdma_buffer
*buffers
= req
->rl_buffer
;
1144 spin_lock(&buffers
->rb_lock
);
1145 if (!list_empty(&buffers
->rb_recv_bufs
))
1146 req
->rl_reply
= rpcrdma_buffer_get_rep_locked(buffers
);
1147 spin_unlock(&buffers
->rb_lock
);
1151 * Put reply buffers back into pool when not attached to
1152 * request. This happens in error conditions.
1155 rpcrdma_recv_buffer_put(struct rpcrdma_rep
*rep
)
1157 struct rpcrdma_buffer
*buffers
= &rep
->rr_rxprt
->rx_buf
;
1159 spin_lock(&buffers
->rb_lock
);
1160 list_add_tail(&rep
->rr_list
, &buffers
->rb_recv_bufs
);
1161 spin_unlock(&buffers
->rb_lock
);
1165 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1169 rpcrdma_mapping_error(struct rpcrdma_mr_seg
*seg
)
1171 dprintk("RPC: map_one: offset %p iova %llx len %zu\n",
1173 (unsigned long long)seg
->mr_dma
, seg
->mr_dmalen
);
1177 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1178 * @ia: controlling rpcrdma_ia
1179 * @size: size of buffer to be allocated, in bytes
1182 * Returns pointer to private header of an area of internally
1183 * registered memory, or an ERR_PTR. The registered buffer follows
1184 * the end of the private header.
1186 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1187 * receiving the payload of RDMA RECV operations. regbufs are not
1188 * used for RDMA READ/WRITE operations, thus are registered only for
1191 struct rpcrdma_regbuf
*
1192 rpcrdma_alloc_regbuf(struct rpcrdma_ia
*ia
, size_t size
, gfp_t flags
)
1194 struct rpcrdma_regbuf
*rb
;
1197 rb
= kmalloc(sizeof(*rb
) + size
, flags
);
1202 iov
->addr
= ib_dma_map_single(ia
->ri_device
,
1203 (void *)rb
->rg_base
, size
,
1205 if (ib_dma_mapping_error(ia
->ri_device
, iov
->addr
))
1209 iov
->lkey
= ia
->ri_pd
->local_dma_lkey
;
1211 rb
->rg_owner
= NULL
;
1217 return ERR_PTR(-ENOMEM
);
1221 * rpcrdma_free_regbuf - deregister and free registered buffer
1222 * @ia: controlling rpcrdma_ia
1223 * @rb: regbuf to be deregistered and freed
1226 rpcrdma_free_regbuf(struct rpcrdma_ia
*ia
, struct rpcrdma_regbuf
*rb
)
1234 ib_dma_unmap_single(ia
->ri_device
,
1235 iov
->addr
, iov
->length
, DMA_BIDIRECTIONAL
);
1240 * Prepost any receive buffer, then post send.
1242 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1245 rpcrdma_ep_post(struct rpcrdma_ia
*ia
,
1246 struct rpcrdma_ep
*ep
,
1247 struct rpcrdma_req
*req
)
1249 struct ib_device
*device
= ia
->ri_device
;
1250 struct ib_send_wr send_wr
, *send_wr_fail
;
1251 struct rpcrdma_rep
*rep
= req
->rl_reply
;
1252 struct ib_sge
*iov
= req
->rl_send_iov
;
1256 rc
= rpcrdma_ep_post_recv(ia
, ep
, rep
);
1259 req
->rl_reply
= NULL
;
1262 send_wr
.next
= NULL
;
1263 send_wr
.wr_id
= RPCRDMA_IGNORE_COMPLETION
;
1264 send_wr
.sg_list
= iov
;
1265 send_wr
.num_sge
= req
->rl_niovs
;
1266 send_wr
.opcode
= IB_WR_SEND
;
1268 for (i
= 0; i
< send_wr
.num_sge
; i
++)
1269 ib_dma_sync_single_for_device(device
, iov
[i
].addr
,
1270 iov
[i
].length
, DMA_TO_DEVICE
);
1271 dprintk("RPC: %s: posting %d s/g entries\n",
1272 __func__
, send_wr
.num_sge
);
1274 if (DECR_CQCOUNT(ep
) > 0)
1275 send_wr
.send_flags
= 0;
1276 else { /* Provider must take a send completion every now and then */
1278 send_wr
.send_flags
= IB_SEND_SIGNALED
;
1281 rc
= ib_post_send(ia
->ri_id
->qp
, &send_wr
, &send_wr_fail
);
1283 dprintk("RPC: %s: ib_post_send returned %i\n", __func__
,
1290 * (Re)post a receive buffer.
1293 rpcrdma_ep_post_recv(struct rpcrdma_ia
*ia
,
1294 struct rpcrdma_ep
*ep
,
1295 struct rpcrdma_rep
*rep
)
1297 struct ib_recv_wr recv_wr
, *recv_wr_fail
;
1300 recv_wr
.next
= NULL
;
1301 recv_wr
.wr_id
= (u64
) (unsigned long) rep
;
1302 recv_wr
.sg_list
= &rep
->rr_rdmabuf
->rg_iov
;
1303 recv_wr
.num_sge
= 1;
1305 ib_dma_sync_single_for_cpu(ia
->ri_device
,
1306 rdmab_addr(rep
->rr_rdmabuf
),
1307 rdmab_length(rep
->rr_rdmabuf
),
1310 rc
= ib_post_recv(ia
->ri_id
->qp
, &recv_wr
, &recv_wr_fail
);
1313 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__
,
1319 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1320 * @r_xprt: transport associated with these backchannel resources
1321 * @min_reqs: minimum number of incoming requests expected
1323 * Returns zero if all requested buffers were posted, or a negative errno.
1326 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*r_xprt
, unsigned int count
)
1328 struct rpcrdma_buffer
*buffers
= &r_xprt
->rx_buf
;
1329 struct rpcrdma_ia
*ia
= &r_xprt
->rx_ia
;
1330 struct rpcrdma_ep
*ep
= &r_xprt
->rx_ep
;
1331 struct rpcrdma_rep
*rep
;
1332 unsigned long flags
;
1336 spin_lock_irqsave(&buffers
->rb_lock
, flags
);
1337 if (list_empty(&buffers
->rb_recv_bufs
))
1339 rep
= rpcrdma_buffer_get_rep_locked(buffers
);
1340 spin_unlock_irqrestore(&buffers
->rb_lock
, flags
);
1342 rc
= rpcrdma_ep_post_recv(ia
, ep
, rep
);
1350 spin_unlock_irqrestore(&buffers
->rb_lock
, flags
);
1351 pr_warn("%s: no extra receive buffers\n", __func__
);
1355 rpcrdma_recv_buffer_put(rep
);
1359 /* How many chunk list items fit within our inline buffers?
1362 rpcrdma_max_segments(struct rpcrdma_xprt
*r_xprt
)
1364 struct rpcrdma_create_data_internal
*cdata
= &r_xprt
->rx_data
;
1365 int bytes
, segments
;
1367 bytes
= min_t(unsigned int, cdata
->inline_wsize
, cdata
->inline_rsize
);
1368 bytes
-= RPCRDMA_HDRLEN_MIN
;
1369 if (bytes
< sizeof(struct rpcrdma_segment
) * 2) {
1370 pr_warn("RPC: %s: inline threshold too small\n",
1375 segments
= 1 << (fls(bytes
/ sizeof(struct rpcrdma_segment
)) - 1);
1376 dprintk("RPC: %s: max chunk list size = %d segments\n",
1377 __func__
, segments
);