2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2007 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.
40 * Author: Tom Tucker <tom@opengridcomputing.com>
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/debug.h>
45 #include <linux/sunrpc/rpc_rdma.h>
46 #include <linux/interrupt.h>
47 #include <linux/sched.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <rdma/ib_verbs.h>
52 #include <rdma/rdma_cm.h>
53 #include <linux/sunrpc/svc_rdma.h>
54 #include <linux/export.h>
55 #include "xprt_rdma.h"
57 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
59 static struct svc_xprt
*svc_rdma_create(struct svc_serv
*serv
,
61 struct sockaddr
*sa
, int salen
,
63 static struct svc_xprt
*svc_rdma_accept(struct svc_xprt
*xprt
);
64 static void svc_rdma_release_rqst(struct svc_rqst
*);
65 static void dto_tasklet_func(unsigned long data
);
66 static void svc_rdma_detach(struct svc_xprt
*xprt
);
67 static void svc_rdma_free(struct svc_xprt
*xprt
);
68 static int svc_rdma_has_wspace(struct svc_xprt
*xprt
);
69 static int svc_rdma_secure_port(struct svc_rqst
*);
70 static void rq_cq_reap(struct svcxprt_rdma
*xprt
);
71 static void sq_cq_reap(struct svcxprt_rdma
*xprt
);
73 static DECLARE_TASKLET(dto_tasklet
, dto_tasklet_func
, 0UL);
74 static DEFINE_SPINLOCK(dto_lock
);
75 static LIST_HEAD(dto_xprt_q
);
77 static struct svc_xprt_ops svc_rdma_ops
= {
78 .xpo_create
= svc_rdma_create
,
79 .xpo_recvfrom
= svc_rdma_recvfrom
,
80 .xpo_sendto
= svc_rdma_sendto
,
81 .xpo_release_rqst
= svc_rdma_release_rqst
,
82 .xpo_detach
= svc_rdma_detach
,
83 .xpo_free
= svc_rdma_free
,
84 .xpo_prep_reply_hdr
= svc_rdma_prep_reply_hdr
,
85 .xpo_has_wspace
= svc_rdma_has_wspace
,
86 .xpo_accept
= svc_rdma_accept
,
87 .xpo_secure_port
= svc_rdma_secure_port
,
90 struct svc_xprt_class svc_rdma_class
= {
92 .xcl_owner
= THIS_MODULE
,
93 .xcl_ops
= &svc_rdma_ops
,
94 .xcl_max_payload
= RPCSVC_MAXPAYLOAD_RDMA
,
95 .xcl_ident
= XPRT_TRANSPORT_RDMA
,
98 struct svc_rdma_op_ctxt
*svc_rdma_get_context(struct svcxprt_rdma
*xprt
)
100 struct svc_rdma_op_ctxt
*ctxt
;
102 ctxt
= kmem_cache_alloc(svc_rdma_ctxt_cachep
,
103 GFP_KERNEL
| __GFP_NOFAIL
);
105 INIT_LIST_HEAD(&ctxt
->dto_q
);
108 atomic_inc(&xprt
->sc_ctxt_used
);
112 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt
*ctxt
)
114 struct svcxprt_rdma
*xprt
= ctxt
->xprt
;
116 for (i
= 0; i
< ctxt
->count
&& ctxt
->sge
[i
].length
; i
++) {
118 * Unmap the DMA addr in the SGE if the lkey matches
119 * the sc_dma_lkey, otherwise, ignore it since it is
120 * an FRMR lkey and will be unmapped later when the
121 * last WR that uses it completes.
123 if (ctxt
->sge
[i
].lkey
== xprt
->sc_dma_lkey
) {
124 atomic_dec(&xprt
->sc_dma_used
);
125 ib_dma_unmap_page(xprt
->sc_cm_id
->device
,
133 void svc_rdma_put_context(struct svc_rdma_op_ctxt
*ctxt
, int free_pages
)
135 struct svcxprt_rdma
*xprt
;
140 for (i
= 0; i
< ctxt
->count
; i
++)
141 put_page(ctxt
->pages
[i
]);
143 kmem_cache_free(svc_rdma_ctxt_cachep
, ctxt
);
144 atomic_dec(&xprt
->sc_ctxt_used
);
148 * Temporary NFS req mappings are shared across all transport
149 * instances. These are short lived and should be bounded by the number
150 * of concurrent server threads * depth of the SQ.
152 struct svc_rdma_req_map
*svc_rdma_get_req_map(void)
154 struct svc_rdma_req_map
*map
;
155 map
= kmem_cache_alloc(svc_rdma_map_cachep
,
156 GFP_KERNEL
| __GFP_NOFAIL
);
161 void svc_rdma_put_req_map(struct svc_rdma_req_map
*map
)
163 kmem_cache_free(svc_rdma_map_cachep
, map
);
166 /* ib_cq event handler */
167 static void cq_event_handler(struct ib_event
*event
, void *context
)
169 struct svc_xprt
*xprt
= context
;
170 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
171 event
->event
, context
);
172 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
175 /* QP event handler */
176 static void qp_event_handler(struct ib_event
*event
, void *context
)
178 struct svc_xprt
*xprt
= context
;
180 switch (event
->event
) {
181 /* These are considered benign events */
182 case IB_EVENT_PATH_MIG
:
183 case IB_EVENT_COMM_EST
:
184 case IB_EVENT_SQ_DRAINED
:
185 case IB_EVENT_QP_LAST_WQE_REACHED
:
186 dprintk("svcrdma: QP event %d received for QP=%p\n",
187 event
->event
, event
->element
.qp
);
189 /* These are considered fatal events */
190 case IB_EVENT_PATH_MIG_ERR
:
191 case IB_EVENT_QP_FATAL
:
192 case IB_EVENT_QP_REQ_ERR
:
193 case IB_EVENT_QP_ACCESS_ERR
:
194 case IB_EVENT_DEVICE_FATAL
:
196 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
197 "closing transport\n",
198 event
->event
, event
->element
.qp
);
199 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
205 * Data Transfer Operation Tasklet
207 * Walks a list of transports with I/O pending, removing entries as
208 * they are added to the server's I/O pending list. Two bits indicate
209 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
210 * spinlock that serializes access to the transport list with the RQ
211 * and SQ interrupt handlers.
213 static void dto_tasklet_func(unsigned long data
)
215 struct svcxprt_rdma
*xprt
;
218 spin_lock_irqsave(&dto_lock
, flags
);
219 while (!list_empty(&dto_xprt_q
)) {
220 xprt
= list_entry(dto_xprt_q
.next
,
221 struct svcxprt_rdma
, sc_dto_q
);
222 list_del_init(&xprt
->sc_dto_q
);
223 spin_unlock_irqrestore(&dto_lock
, flags
);
228 svc_xprt_put(&xprt
->sc_xprt
);
229 spin_lock_irqsave(&dto_lock
, flags
);
231 spin_unlock_irqrestore(&dto_lock
, flags
);
235 * Receive Queue Completion Handler
237 * Since an RQ completion handler is called on interrupt context, we
238 * need to defer the handling of the I/O to a tasklet
240 static void rq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
242 struct svcxprt_rdma
*xprt
= cq_context
;
245 /* Guard against unconditional flush call for destroyed QP */
246 if (atomic_read(&xprt
->sc_xprt
.xpt_ref
.refcount
)==0)
250 * Set the bit regardless of whether or not it's on the list
251 * because it may be on the list already due to an SQ
254 set_bit(RDMAXPRT_RQ_PENDING
, &xprt
->sc_flags
);
257 * If this transport is not already on the DTO transport queue,
260 spin_lock_irqsave(&dto_lock
, flags
);
261 if (list_empty(&xprt
->sc_dto_q
)) {
262 svc_xprt_get(&xprt
->sc_xprt
);
263 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
265 spin_unlock_irqrestore(&dto_lock
, flags
);
267 /* Tasklet does all the work to avoid irqsave locks. */
268 tasklet_schedule(&dto_tasklet
);
272 * rq_cq_reap - Process the RQ CQ.
274 * Take all completing WC off the CQE and enqueue the associated DTO
275 * context on the dto_q for the transport.
277 * Note that caller must hold a transport reference.
279 static void rq_cq_reap(struct svcxprt_rdma
*xprt
)
283 struct svc_rdma_op_ctxt
*ctxt
= NULL
;
285 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING
, &xprt
->sc_flags
))
288 ib_req_notify_cq(xprt
->sc_rq_cq
, IB_CQ_NEXT_COMP
);
289 atomic_inc(&rdma_stat_rq_poll
);
291 while ((ret
= ib_poll_cq(xprt
->sc_rq_cq
, 1, &wc
)) > 0) {
292 ctxt
= (struct svc_rdma_op_ctxt
*)(unsigned long)wc
.wr_id
;
293 ctxt
->wc_status
= wc
.status
;
294 ctxt
->byte_len
= wc
.byte_len
;
295 svc_rdma_unmap_dma(ctxt
);
296 if (wc
.status
!= IB_WC_SUCCESS
) {
297 /* Close the transport */
298 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt
);
299 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
300 svc_rdma_put_context(ctxt
, 1);
301 svc_xprt_put(&xprt
->sc_xprt
);
304 spin_lock_bh(&xprt
->sc_rq_dto_lock
);
305 list_add_tail(&ctxt
->dto_q
, &xprt
->sc_rq_dto_q
);
306 spin_unlock_bh(&xprt
->sc_rq_dto_lock
);
307 svc_xprt_put(&xprt
->sc_xprt
);
311 atomic_inc(&rdma_stat_rq_prod
);
313 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
315 * If data arrived before established event,
316 * don't enqueue. This defers RPC I/O until the
317 * RDMA connection is complete.
319 if (!test_bit(RDMAXPRT_CONN_PENDING
, &xprt
->sc_flags
))
320 svc_xprt_enqueue(&xprt
->sc_xprt
);
324 * Process a completion context
326 static void process_context(struct svcxprt_rdma
*xprt
,
327 struct svc_rdma_op_ctxt
*ctxt
)
329 svc_rdma_unmap_dma(ctxt
);
331 switch (ctxt
->wr_op
) {
334 pr_err("svcrdma: SEND: ctxt->frmr != NULL\n");
335 svc_rdma_put_context(ctxt
, 1);
338 case IB_WR_RDMA_WRITE
:
340 pr_err("svcrdma: WRITE: ctxt->frmr != NULL\n");
341 svc_rdma_put_context(ctxt
, 0);
344 case IB_WR_RDMA_READ
:
345 case IB_WR_RDMA_READ_WITH_INV
:
346 svc_rdma_put_frmr(xprt
, ctxt
->frmr
);
347 if (test_bit(RDMACTXT_F_LAST_CTXT
, &ctxt
->flags
)) {
348 struct svc_rdma_op_ctxt
*read_hdr
= ctxt
->read_hdr
;
350 spin_lock_bh(&xprt
->sc_rq_dto_lock
);
351 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
352 list_add_tail(&read_hdr
->dto_q
,
353 &xprt
->sc_read_complete_q
);
354 spin_unlock_bh(&xprt
->sc_rq_dto_lock
);
356 pr_err("svcrdma: ctxt->read_hdr == NULL\n");
358 svc_xprt_enqueue(&xprt
->sc_xprt
);
360 svc_rdma_put_context(ctxt
, 0);
364 printk(KERN_ERR
"svcrdma: unexpected completion type, "
372 * Send Queue Completion Handler - potentially called on interrupt context.
374 * Note that caller must hold a transport reference.
376 static void sq_cq_reap(struct svcxprt_rdma
*xprt
)
378 struct svc_rdma_op_ctxt
*ctxt
= NULL
;
379 struct ib_wc wc_a
[6];
381 struct ib_cq
*cq
= xprt
->sc_sq_cq
;
384 memset(wc_a
, 0, sizeof(wc_a
));
386 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
))
389 ib_req_notify_cq(xprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
390 atomic_inc(&rdma_stat_sq_poll
);
391 while ((ret
= ib_poll_cq(cq
, ARRAY_SIZE(wc_a
), wc_a
)) > 0) {
394 for (i
= 0; i
< ret
; i
++) {
396 if (wc
->status
!= IB_WC_SUCCESS
) {
397 dprintk("svcrdma: sq wc err status %d\n",
400 /* Close the transport */
401 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
404 /* Decrement used SQ WR count */
405 atomic_dec(&xprt
->sc_sq_count
);
406 wake_up(&xprt
->sc_send_wait
);
408 ctxt
= (struct svc_rdma_op_ctxt
*)
409 (unsigned long)wc
->wr_id
;
411 process_context(xprt
, ctxt
);
413 svc_xprt_put(&xprt
->sc_xprt
);
418 atomic_inc(&rdma_stat_sq_prod
);
421 static void sq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
423 struct svcxprt_rdma
*xprt
= cq_context
;
426 /* Guard against unconditional flush call for destroyed QP */
427 if (atomic_read(&xprt
->sc_xprt
.xpt_ref
.refcount
)==0)
431 * Set the bit regardless of whether or not it's on the list
432 * because it may be on the list already due to an RQ
435 set_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
);
438 * If this transport is not already on the DTO transport queue,
441 spin_lock_irqsave(&dto_lock
, flags
);
442 if (list_empty(&xprt
->sc_dto_q
)) {
443 svc_xprt_get(&xprt
->sc_xprt
);
444 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
446 spin_unlock_irqrestore(&dto_lock
, flags
);
448 /* Tasklet does all the work to avoid irqsave locks. */
449 tasklet_schedule(&dto_tasklet
);
452 static struct svcxprt_rdma
*rdma_create_xprt(struct svc_serv
*serv
,
455 struct svcxprt_rdma
*cma_xprt
= kzalloc(sizeof *cma_xprt
, GFP_KERNEL
);
459 svc_xprt_init(&init_net
, &svc_rdma_class
, &cma_xprt
->sc_xprt
, serv
);
460 INIT_LIST_HEAD(&cma_xprt
->sc_accept_q
);
461 INIT_LIST_HEAD(&cma_xprt
->sc_dto_q
);
462 INIT_LIST_HEAD(&cma_xprt
->sc_rq_dto_q
);
463 INIT_LIST_HEAD(&cma_xprt
->sc_read_complete_q
);
464 INIT_LIST_HEAD(&cma_xprt
->sc_frmr_q
);
465 init_waitqueue_head(&cma_xprt
->sc_send_wait
);
467 spin_lock_init(&cma_xprt
->sc_lock
);
468 spin_lock_init(&cma_xprt
->sc_rq_dto_lock
);
469 spin_lock_init(&cma_xprt
->sc_frmr_q_lock
);
471 cma_xprt
->sc_ord
= svcrdma_ord
;
473 cma_xprt
->sc_max_req_size
= svcrdma_max_req_size
;
474 cma_xprt
->sc_max_requests
= svcrdma_max_requests
;
475 cma_xprt
->sc_sq_depth
= svcrdma_max_requests
* RPCRDMA_SQ_DEPTH_MULT
;
476 atomic_set(&cma_xprt
->sc_sq_count
, 0);
477 atomic_set(&cma_xprt
->sc_ctxt_used
, 0);
480 set_bit(XPT_LISTENER
, &cma_xprt
->sc_xprt
.xpt_flags
);
485 int svc_rdma_post_recv(struct svcxprt_rdma
*xprt
)
487 struct ib_recv_wr recv_wr
, *bad_recv_wr
;
488 struct svc_rdma_op_ctxt
*ctxt
;
495 ctxt
= svc_rdma_get_context(xprt
);
497 ctxt
->direction
= DMA_FROM_DEVICE
;
498 for (sge_no
= 0; buflen
< xprt
->sc_max_req_size
; sge_no
++) {
499 if (sge_no
>= xprt
->sc_max_sge
) {
500 pr_err("svcrdma: Too many sges (%d)\n", sge_no
);
503 page
= alloc_page(GFP_KERNEL
| __GFP_NOFAIL
);
504 ctxt
->pages
[sge_no
] = page
;
505 pa
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
508 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
, pa
))
510 atomic_inc(&xprt
->sc_dma_used
);
511 ctxt
->sge
[sge_no
].addr
= pa
;
512 ctxt
->sge
[sge_no
].length
= PAGE_SIZE
;
513 ctxt
->sge
[sge_no
].lkey
= xprt
->sc_dma_lkey
;
514 ctxt
->count
= sge_no
+ 1;
518 recv_wr
.sg_list
= &ctxt
->sge
[0];
519 recv_wr
.num_sge
= ctxt
->count
;
520 recv_wr
.wr_id
= (u64
)(unsigned long)ctxt
;
522 svc_xprt_get(&xprt
->sc_xprt
);
523 ret
= ib_post_recv(xprt
->sc_qp
, &recv_wr
, &bad_recv_wr
);
525 svc_rdma_unmap_dma(ctxt
);
526 svc_rdma_put_context(ctxt
, 1);
527 svc_xprt_put(&xprt
->sc_xprt
);
532 svc_rdma_unmap_dma(ctxt
);
533 svc_rdma_put_context(ctxt
, 1);
538 * This function handles the CONNECT_REQUEST event on a listening
539 * endpoint. It is passed the cma_id for the _new_ connection. The context in
540 * this cma_id is inherited from the listening cma_id and is the svc_xprt
541 * structure for the listening endpoint.
543 * This function creates a new xprt for the new connection and enqueues it on
544 * the accept queue for the listent xprt. When the listen thread is kicked, it
545 * will call the recvfrom method on the listen xprt which will accept the new
548 static void handle_connect_req(struct rdma_cm_id
*new_cma_id
, size_t client_ird
)
550 struct svcxprt_rdma
*listen_xprt
= new_cma_id
->context
;
551 struct svcxprt_rdma
*newxprt
;
554 /* Create a new transport */
555 newxprt
= rdma_create_xprt(listen_xprt
->sc_xprt
.xpt_server
, 0);
557 dprintk("svcrdma: failed to create new transport\n");
560 newxprt
->sc_cm_id
= new_cma_id
;
561 new_cma_id
->context
= newxprt
;
562 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
563 newxprt
, newxprt
->sc_cm_id
, listen_xprt
);
565 /* Save client advertised inbound read limit for use later in accept. */
566 newxprt
->sc_ord
= client_ird
;
568 /* Set the local and remote addresses in the transport */
569 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.dst_addr
;
570 svc_xprt_set_remote(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
571 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.src_addr
;
572 svc_xprt_set_local(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
575 * Enqueue the new transport on the accept queue of the listening
578 spin_lock_bh(&listen_xprt
->sc_lock
);
579 list_add_tail(&newxprt
->sc_accept_q
, &listen_xprt
->sc_accept_q
);
580 spin_unlock_bh(&listen_xprt
->sc_lock
);
582 set_bit(XPT_CONN
, &listen_xprt
->sc_xprt
.xpt_flags
);
583 svc_xprt_enqueue(&listen_xprt
->sc_xprt
);
587 * Handles events generated on the listening endpoint. These events will be
588 * either be incoming connect requests or adapter removal events.
590 static int rdma_listen_handler(struct rdma_cm_id
*cma_id
,
591 struct rdma_cm_event
*event
)
593 struct svcxprt_rdma
*xprt
= cma_id
->context
;
596 switch (event
->event
) {
597 case RDMA_CM_EVENT_CONNECT_REQUEST
:
598 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
599 "event=%d\n", cma_id
, cma_id
->context
, event
->event
);
600 handle_connect_req(cma_id
,
601 event
->param
.conn
.initiator_depth
);
604 case RDMA_CM_EVENT_ESTABLISHED
:
605 /* Accept complete */
606 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
607 "cm_id=%p\n", xprt
, cma_id
);
610 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
611 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
614 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
618 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
619 "event=%d\n", cma_id
, event
->event
);
626 static int rdma_cma_handler(struct rdma_cm_id
*cma_id
,
627 struct rdma_cm_event
*event
)
629 struct svc_xprt
*xprt
= cma_id
->context
;
630 struct svcxprt_rdma
*rdma
=
631 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
632 switch (event
->event
) {
633 case RDMA_CM_EVENT_ESTABLISHED
:
634 /* Accept complete */
636 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
637 "cm_id=%p\n", xprt
, cma_id
);
638 clear_bit(RDMAXPRT_CONN_PENDING
, &rdma
->sc_flags
);
639 svc_xprt_enqueue(xprt
);
641 case RDMA_CM_EVENT_DISCONNECTED
:
642 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
645 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
646 svc_xprt_enqueue(xprt
);
650 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
651 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
652 "event=%d\n", cma_id
, xprt
, event
->event
);
654 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
655 svc_xprt_enqueue(xprt
);
659 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
660 "event=%d\n", cma_id
, event
->event
);
667 * Create a listening RDMA service endpoint.
669 static struct svc_xprt
*svc_rdma_create(struct svc_serv
*serv
,
671 struct sockaddr
*sa
, int salen
,
674 struct rdma_cm_id
*listen_id
;
675 struct svcxprt_rdma
*cma_xprt
;
678 dprintk("svcrdma: Creating RDMA socket\n");
679 if (sa
->sa_family
!= AF_INET
) {
680 dprintk("svcrdma: Address family %d is not supported.\n", sa
->sa_family
);
681 return ERR_PTR(-EAFNOSUPPORT
);
683 cma_xprt
= rdma_create_xprt(serv
, 1);
685 return ERR_PTR(-ENOMEM
);
687 listen_id
= rdma_create_id(rdma_listen_handler
, cma_xprt
, RDMA_PS_TCP
,
689 if (IS_ERR(listen_id
)) {
690 ret
= PTR_ERR(listen_id
);
691 dprintk("svcrdma: rdma_create_id failed = %d\n", ret
);
695 ret
= rdma_bind_addr(listen_id
, sa
);
697 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret
);
700 cma_xprt
->sc_cm_id
= listen_id
;
702 ret
= rdma_listen(listen_id
, RPCRDMA_LISTEN_BACKLOG
);
704 dprintk("svcrdma: rdma_listen failed = %d\n", ret
);
709 * We need to use the address from the cm_id in case the
710 * caller specified 0 for the port number.
712 sa
= (struct sockaddr
*)&cma_xprt
->sc_cm_id
->route
.addr
.src_addr
;
713 svc_xprt_set_local(&cma_xprt
->sc_xprt
, sa
, salen
);
715 return &cma_xprt
->sc_xprt
;
718 rdma_destroy_id(listen_id
);
724 static struct svc_rdma_fastreg_mr
*rdma_alloc_frmr(struct svcxprt_rdma
*xprt
)
727 struct ib_fast_reg_page_list
*pl
;
728 struct svc_rdma_fastreg_mr
*frmr
;
730 frmr
= kmalloc(sizeof(*frmr
), GFP_KERNEL
);
734 mr
= ib_alloc_fast_reg_mr(xprt
->sc_pd
, RPCSVC_MAXPAGES
);
738 pl
= ib_alloc_fast_reg_page_list(xprt
->sc_cm_id
->device
,
744 frmr
->page_list
= pl
;
745 INIT_LIST_HEAD(&frmr
->frmr_list
);
753 return ERR_PTR(-ENOMEM
);
756 static void rdma_dealloc_frmr_q(struct svcxprt_rdma
*xprt
)
758 struct svc_rdma_fastreg_mr
*frmr
;
760 while (!list_empty(&xprt
->sc_frmr_q
)) {
761 frmr
= list_entry(xprt
->sc_frmr_q
.next
,
762 struct svc_rdma_fastreg_mr
, frmr_list
);
763 list_del_init(&frmr
->frmr_list
);
764 ib_dereg_mr(frmr
->mr
);
765 ib_free_fast_reg_page_list(frmr
->page_list
);
770 struct svc_rdma_fastreg_mr
*svc_rdma_get_frmr(struct svcxprt_rdma
*rdma
)
772 struct svc_rdma_fastreg_mr
*frmr
= NULL
;
774 spin_lock_bh(&rdma
->sc_frmr_q_lock
);
775 if (!list_empty(&rdma
->sc_frmr_q
)) {
776 frmr
= list_entry(rdma
->sc_frmr_q
.next
,
777 struct svc_rdma_fastreg_mr
, frmr_list
);
778 list_del_init(&frmr
->frmr_list
);
780 frmr
->page_list_len
= 0;
782 spin_unlock_bh(&rdma
->sc_frmr_q_lock
);
786 return rdma_alloc_frmr(rdma
);
789 static void frmr_unmap_dma(struct svcxprt_rdma
*xprt
,
790 struct svc_rdma_fastreg_mr
*frmr
)
793 for (page_no
= 0; page_no
< frmr
->page_list_len
; page_no
++) {
794 dma_addr_t addr
= frmr
->page_list
->page_list
[page_no
];
795 if (ib_dma_mapping_error(frmr
->mr
->device
, addr
))
797 atomic_dec(&xprt
->sc_dma_used
);
798 ib_dma_unmap_page(frmr
->mr
->device
, addr
, PAGE_SIZE
,
803 void svc_rdma_put_frmr(struct svcxprt_rdma
*rdma
,
804 struct svc_rdma_fastreg_mr
*frmr
)
807 frmr_unmap_dma(rdma
, frmr
);
808 spin_lock_bh(&rdma
->sc_frmr_q_lock
);
809 WARN_ON_ONCE(!list_empty(&frmr
->frmr_list
));
810 list_add(&frmr
->frmr_list
, &rdma
->sc_frmr_q
);
811 spin_unlock_bh(&rdma
->sc_frmr_q_lock
);
816 * This is the xpo_recvfrom function for listening endpoints. Its
817 * purpose is to accept incoming connections. The CMA callback handler
818 * has already created a new transport and attached it to the new CMA
821 * There is a queue of pending connections hung on the listening
822 * transport. This queue contains the new svc_xprt structure. This
823 * function takes svc_xprt structures off the accept_q and completes
826 static struct svc_xprt
*svc_rdma_accept(struct svc_xprt
*xprt
)
828 struct svcxprt_rdma
*listen_rdma
;
829 struct svcxprt_rdma
*newxprt
= NULL
;
830 struct rdma_conn_param conn_param
;
831 struct ib_qp_init_attr qp_attr
;
832 struct ib_device_attr devattr
;
833 int uninitialized_var(dma_mr_acc
);
838 listen_rdma
= container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
839 clear_bit(XPT_CONN
, &xprt
->xpt_flags
);
840 /* Get the next entry off the accept list */
841 spin_lock_bh(&listen_rdma
->sc_lock
);
842 if (!list_empty(&listen_rdma
->sc_accept_q
)) {
843 newxprt
= list_entry(listen_rdma
->sc_accept_q
.next
,
844 struct svcxprt_rdma
, sc_accept_q
);
845 list_del_init(&newxprt
->sc_accept_q
);
847 if (!list_empty(&listen_rdma
->sc_accept_q
))
848 set_bit(XPT_CONN
, &listen_rdma
->sc_xprt
.xpt_flags
);
849 spin_unlock_bh(&listen_rdma
->sc_lock
);
853 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
854 newxprt
, newxprt
->sc_cm_id
);
856 ret
= ib_query_device(newxprt
->sc_cm_id
->device
, &devattr
);
858 dprintk("svcrdma: could not query device attributes on "
859 "device %p, rc=%d\n", newxprt
->sc_cm_id
->device
, ret
);
863 /* Qualify the transport resource defaults with the
864 * capabilities of this particular device */
865 newxprt
->sc_max_sge
= min((size_t)devattr
.max_sge
,
866 (size_t)RPCSVC_MAXPAGES
);
867 newxprt
->sc_max_requests
= min((size_t)devattr
.max_qp_wr
,
868 (size_t)svcrdma_max_requests
);
869 newxprt
->sc_sq_depth
= RPCRDMA_SQ_DEPTH_MULT
* newxprt
->sc_max_requests
;
872 * Limit ORD based on client limit, local device limit, and
873 * configured svcrdma limit.
875 newxprt
->sc_ord
= min_t(size_t, devattr
.max_qp_rd_atom
, newxprt
->sc_ord
);
876 newxprt
->sc_ord
= min_t(size_t, svcrdma_ord
, newxprt
->sc_ord
);
878 newxprt
->sc_pd
= ib_alloc_pd(newxprt
->sc_cm_id
->device
);
879 if (IS_ERR(newxprt
->sc_pd
)) {
880 dprintk("svcrdma: error creating PD for connect request\n");
883 newxprt
->sc_sq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
887 newxprt
->sc_sq_depth
,
889 if (IS_ERR(newxprt
->sc_sq_cq
)) {
890 dprintk("svcrdma: error creating SQ CQ for connect request\n");
893 newxprt
->sc_rq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
897 newxprt
->sc_max_requests
,
899 if (IS_ERR(newxprt
->sc_rq_cq
)) {
900 dprintk("svcrdma: error creating RQ CQ for connect request\n");
904 memset(&qp_attr
, 0, sizeof qp_attr
);
905 qp_attr
.event_handler
= qp_event_handler
;
906 qp_attr
.qp_context
= &newxprt
->sc_xprt
;
907 qp_attr
.cap
.max_send_wr
= newxprt
->sc_sq_depth
;
908 qp_attr
.cap
.max_recv_wr
= newxprt
->sc_max_requests
;
909 qp_attr
.cap
.max_send_sge
= newxprt
->sc_max_sge
;
910 qp_attr
.cap
.max_recv_sge
= newxprt
->sc_max_sge
;
911 qp_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
912 qp_attr
.qp_type
= IB_QPT_RC
;
913 qp_attr
.send_cq
= newxprt
->sc_sq_cq
;
914 qp_attr
.recv_cq
= newxprt
->sc_rq_cq
;
915 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
916 " cm_id->device=%p, sc_pd->device=%p\n"
917 " cap.max_send_wr = %d\n"
918 " cap.max_recv_wr = %d\n"
919 " cap.max_send_sge = %d\n"
920 " cap.max_recv_sge = %d\n",
921 newxprt
->sc_cm_id
, newxprt
->sc_pd
,
922 newxprt
->sc_cm_id
->device
, newxprt
->sc_pd
->device
,
923 qp_attr
.cap
.max_send_wr
,
924 qp_attr
.cap
.max_recv_wr
,
925 qp_attr
.cap
.max_send_sge
,
926 qp_attr
.cap
.max_recv_sge
);
928 ret
= rdma_create_qp(newxprt
->sc_cm_id
, newxprt
->sc_pd
, &qp_attr
);
930 dprintk("svcrdma: failed to create QP, ret=%d\n", ret
);
933 newxprt
->sc_qp
= newxprt
->sc_cm_id
->qp
;
936 * Use the most secure set of MR resources based on the
937 * transport type and available memory management features in
938 * the device. Here's the table implemented below:
940 * Fast Global DMA Remote WR
942 * Sup'd Sup'd Needed Needed
954 * NB: iWARP requires remote write access for the data sink
955 * of an RDMA_READ. IB does not.
957 newxprt
->sc_reader
= rdma_read_chunk_lcl
;
958 if (devattr
.device_cap_flags
& IB_DEVICE_MEM_MGT_EXTENSIONS
) {
959 newxprt
->sc_frmr_pg_list_len
=
960 devattr
.max_fast_reg_page_list_len
;
961 newxprt
->sc_dev_caps
|= SVCRDMA_DEVCAP_FAST_REG
;
962 newxprt
->sc_reader
= rdma_read_chunk_frmr
;
966 * Determine if a DMA MR is required and if so, what privs are required
968 switch (rdma_node_get_transport(newxprt
->sc_cm_id
->device
->node_type
)) {
969 case RDMA_TRANSPORT_IWARP
:
970 newxprt
->sc_dev_caps
|= SVCRDMA_DEVCAP_READ_W_INV
;
971 if (!(newxprt
->sc_dev_caps
& SVCRDMA_DEVCAP_FAST_REG
)) {
974 (IB_ACCESS_LOCAL_WRITE
|
975 IB_ACCESS_REMOTE_WRITE
);
976 } else if (!(devattr
.device_cap_flags
& IB_DEVICE_LOCAL_DMA_LKEY
)) {
978 dma_mr_acc
= IB_ACCESS_LOCAL_WRITE
;
982 case RDMA_TRANSPORT_IB
:
983 if (!(newxprt
->sc_dev_caps
& SVCRDMA_DEVCAP_FAST_REG
)) {
985 dma_mr_acc
= IB_ACCESS_LOCAL_WRITE
;
986 } else if (!(devattr
.device_cap_flags
&
987 IB_DEVICE_LOCAL_DMA_LKEY
)) {
989 dma_mr_acc
= IB_ACCESS_LOCAL_WRITE
;
997 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */
999 /* Register all of physical memory */
1000 newxprt
->sc_phys_mr
=
1001 ib_get_dma_mr(newxprt
->sc_pd
, dma_mr_acc
);
1002 if (IS_ERR(newxprt
->sc_phys_mr
)) {
1003 dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
1007 newxprt
->sc_dma_lkey
= newxprt
->sc_phys_mr
->lkey
;
1009 newxprt
->sc_dma_lkey
=
1010 newxprt
->sc_cm_id
->device
->local_dma_lkey
;
1012 /* Post receive buffers */
1013 for (i
= 0; i
< newxprt
->sc_max_requests
; i
++) {
1014 ret
= svc_rdma_post_recv(newxprt
);
1016 dprintk("svcrdma: failure posting receive buffers\n");
1021 /* Swap out the handler */
1022 newxprt
->sc_cm_id
->event_handler
= rdma_cma_handler
;
1025 * Arm the CQs for the SQ and RQ before accepting so we can't
1026 * miss the first message
1028 ib_req_notify_cq(newxprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
1029 ib_req_notify_cq(newxprt
->sc_rq_cq
, IB_CQ_NEXT_COMP
);
1031 /* Accept Connection */
1032 set_bit(RDMAXPRT_CONN_PENDING
, &newxprt
->sc_flags
);
1033 memset(&conn_param
, 0, sizeof conn_param
);
1034 conn_param
.responder_resources
= 0;
1035 conn_param
.initiator_depth
= newxprt
->sc_ord
;
1036 ret
= rdma_accept(newxprt
->sc_cm_id
, &conn_param
);
1038 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1043 dprintk("svcrdma: new connection %p accepted with the following "
1045 " local_ip : %pI4\n"
1046 " local_port : %d\n"
1047 " remote_ip : %pI4\n"
1048 " remote_port : %d\n"
1051 " max_requests : %d\n"
1054 &((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1055 route
.addr
.src_addr
)->sin_addr
.s_addr
,
1056 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1057 route
.addr
.src_addr
)->sin_port
),
1058 &((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1059 route
.addr
.dst_addr
)->sin_addr
.s_addr
,
1060 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1061 route
.addr
.dst_addr
)->sin_port
),
1062 newxprt
->sc_max_sge
,
1063 newxprt
->sc_sq_depth
,
1064 newxprt
->sc_max_requests
,
1067 return &newxprt
->sc_xprt
;
1070 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret
);
1071 /* Take a reference in case the DTO handler runs */
1072 svc_xprt_get(&newxprt
->sc_xprt
);
1073 if (newxprt
->sc_qp
&& !IS_ERR(newxprt
->sc_qp
))
1074 ib_destroy_qp(newxprt
->sc_qp
);
1075 rdma_destroy_id(newxprt
->sc_cm_id
);
1076 /* This call to put will destroy the transport */
1077 svc_xprt_put(&newxprt
->sc_xprt
);
1081 static void svc_rdma_release_rqst(struct svc_rqst
*rqstp
)
1086 * When connected, an svc_xprt has at least two references:
1088 * - A reference held by the cm_id between the ESTABLISHED and
1089 * DISCONNECTED events. If the remote peer disconnected first, this
1090 * reference could be gone.
1092 * - A reference held by the svc_recv code that called this function
1093 * as part of close processing.
1095 * At a minimum one references should still be held.
1097 static void svc_rdma_detach(struct svc_xprt
*xprt
)
1099 struct svcxprt_rdma
*rdma
=
1100 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1101 dprintk("svc: svc_rdma_detach(%p)\n", xprt
);
1103 /* Disconnect and flush posted WQE */
1104 rdma_disconnect(rdma
->sc_cm_id
);
1107 static void __svc_rdma_free(struct work_struct
*work
)
1109 struct svcxprt_rdma
*rdma
=
1110 container_of(work
, struct svcxprt_rdma
, sc_work
);
1111 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma
);
1113 /* We should only be called from kref_put */
1114 if (atomic_read(&rdma
->sc_xprt
.xpt_ref
.refcount
) != 0)
1115 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1116 atomic_read(&rdma
->sc_xprt
.xpt_ref
.refcount
));
1119 * Destroy queued, but not processed read completions. Note
1120 * that this cleanup has to be done before destroying the
1121 * cm_id because the device ptr is needed to unmap the dma in
1122 * svc_rdma_put_context.
1124 while (!list_empty(&rdma
->sc_read_complete_q
)) {
1125 struct svc_rdma_op_ctxt
*ctxt
;
1126 ctxt
= list_entry(rdma
->sc_read_complete_q
.next
,
1127 struct svc_rdma_op_ctxt
,
1129 list_del_init(&ctxt
->dto_q
);
1130 svc_rdma_put_context(ctxt
, 1);
1133 /* Destroy queued, but not processed recv completions */
1134 while (!list_empty(&rdma
->sc_rq_dto_q
)) {
1135 struct svc_rdma_op_ctxt
*ctxt
;
1136 ctxt
= list_entry(rdma
->sc_rq_dto_q
.next
,
1137 struct svc_rdma_op_ctxt
,
1139 list_del_init(&ctxt
->dto_q
);
1140 svc_rdma_put_context(ctxt
, 1);
1143 /* Warn if we leaked a resource or under-referenced */
1144 if (atomic_read(&rdma
->sc_ctxt_used
) != 0)
1145 pr_err("svcrdma: ctxt still in use? (%d)\n",
1146 atomic_read(&rdma
->sc_ctxt_used
));
1147 if (atomic_read(&rdma
->sc_dma_used
) != 0)
1148 pr_err("svcrdma: dma still in use? (%d)\n",
1149 atomic_read(&rdma
->sc_dma_used
));
1151 /* De-allocate fastreg mr */
1152 rdma_dealloc_frmr_q(rdma
);
1154 /* Destroy the QP if present (not a listener) */
1155 if (rdma
->sc_qp
&& !IS_ERR(rdma
->sc_qp
))
1156 ib_destroy_qp(rdma
->sc_qp
);
1158 if (rdma
->sc_sq_cq
&& !IS_ERR(rdma
->sc_sq_cq
))
1159 ib_destroy_cq(rdma
->sc_sq_cq
);
1161 if (rdma
->sc_rq_cq
&& !IS_ERR(rdma
->sc_rq_cq
))
1162 ib_destroy_cq(rdma
->sc_rq_cq
);
1164 if (rdma
->sc_phys_mr
&& !IS_ERR(rdma
->sc_phys_mr
))
1165 ib_dereg_mr(rdma
->sc_phys_mr
);
1167 if (rdma
->sc_pd
&& !IS_ERR(rdma
->sc_pd
))
1168 ib_dealloc_pd(rdma
->sc_pd
);
1170 /* Destroy the CM ID */
1171 rdma_destroy_id(rdma
->sc_cm_id
);
1176 static void svc_rdma_free(struct svc_xprt
*xprt
)
1178 struct svcxprt_rdma
*rdma
=
1179 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1180 INIT_WORK(&rdma
->sc_work
, __svc_rdma_free
);
1181 queue_work(svc_rdma_wq
, &rdma
->sc_work
);
1184 static int svc_rdma_has_wspace(struct svc_xprt
*xprt
)
1186 struct svcxprt_rdma
*rdma
=
1187 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1190 * If there are already waiters on the SQ,
1193 if (waitqueue_active(&rdma
->sc_send_wait
))
1196 /* Otherwise return true. */
1200 static int svc_rdma_secure_port(struct svc_rqst
*rqstp
)
1206 * Attempt to register the kvec representing the RPC memory with the
1210 * NULL : The device does not support fastreg or there were no more
1212 * frmr : The kvec register request was successfully posted.
1213 * <0 : An error was encountered attempting to register the kvec.
1215 int svc_rdma_fastreg(struct svcxprt_rdma
*xprt
,
1216 struct svc_rdma_fastreg_mr
*frmr
)
1218 struct ib_send_wr fastreg_wr
;
1222 key
= (u8
)(frmr
->mr
->lkey
& 0x000000FF);
1223 ib_update_fast_reg_key(frmr
->mr
, ++key
);
1225 /* Prepare FASTREG WR */
1226 memset(&fastreg_wr
, 0, sizeof fastreg_wr
);
1227 fastreg_wr
.opcode
= IB_WR_FAST_REG_MR
;
1228 fastreg_wr
.send_flags
= IB_SEND_SIGNALED
;
1229 fastreg_wr
.wr
.fast_reg
.iova_start
= (unsigned long)frmr
->kva
;
1230 fastreg_wr
.wr
.fast_reg
.page_list
= frmr
->page_list
;
1231 fastreg_wr
.wr
.fast_reg
.page_list_len
= frmr
->page_list_len
;
1232 fastreg_wr
.wr
.fast_reg
.page_shift
= PAGE_SHIFT
;
1233 fastreg_wr
.wr
.fast_reg
.length
= frmr
->map_len
;
1234 fastreg_wr
.wr
.fast_reg
.access_flags
= frmr
->access_flags
;
1235 fastreg_wr
.wr
.fast_reg
.rkey
= frmr
->mr
->lkey
;
1236 return svc_rdma_send(xprt
, &fastreg_wr
);
1239 int svc_rdma_send(struct svcxprt_rdma
*xprt
, struct ib_send_wr
*wr
)
1241 struct ib_send_wr
*bad_wr
, *n_wr
;
1246 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1250 for (n_wr
= wr
->next
; n_wr
; n_wr
= n_wr
->next
)
1253 /* If the SQ is full, wait until an SQ entry is available */
1255 spin_lock_bh(&xprt
->sc_lock
);
1256 if (xprt
->sc_sq_depth
< atomic_read(&xprt
->sc_sq_count
) + wr_count
) {
1257 spin_unlock_bh(&xprt
->sc_lock
);
1258 atomic_inc(&rdma_stat_sq_starve
);
1260 /* See if we can opportunistically reap SQ WR to make room */
1263 /* Wait until SQ WR available if SQ still full */
1264 wait_event(xprt
->sc_send_wait
,
1265 atomic_read(&xprt
->sc_sq_count
) <
1267 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1271 /* Take a transport ref for each WR posted */
1272 for (i
= 0; i
< wr_count
; i
++)
1273 svc_xprt_get(&xprt
->sc_xprt
);
1275 /* Bump used SQ WR count and post */
1276 atomic_add(wr_count
, &xprt
->sc_sq_count
);
1277 ret
= ib_post_send(xprt
->sc_qp
, wr
, &bad_wr
);
1279 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
1280 atomic_sub(wr_count
, &xprt
->sc_sq_count
);
1281 for (i
= 0; i
< wr_count
; i
++)
1282 svc_xprt_put(&xprt
->sc_xprt
);
1283 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1284 "sc_sq_count=%d, sc_sq_depth=%d\n",
1285 ret
, atomic_read(&xprt
->sc_sq_count
),
1288 spin_unlock_bh(&xprt
->sc_lock
);
1290 wake_up(&xprt
->sc_send_wait
);
1296 void svc_rdma_send_error(struct svcxprt_rdma
*xprt
, struct rpcrdma_msg
*rmsgp
,
1297 enum rpcrdma_errcode err
)
1299 struct ib_send_wr err_wr
;
1301 struct svc_rdma_op_ctxt
*ctxt
;
1306 p
= alloc_page(GFP_KERNEL
| __GFP_NOFAIL
);
1307 va
= page_address(p
);
1309 /* XDR encode error */
1310 length
= svc_rdma_xdr_encode_error(xprt
, rmsgp
, err
, va
);
1312 ctxt
= svc_rdma_get_context(xprt
);
1313 ctxt
->direction
= DMA_FROM_DEVICE
;
1317 /* Prepare SGE for local address */
1318 ctxt
->sge
[0].addr
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
1319 p
, 0, length
, DMA_FROM_DEVICE
);
1320 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
, ctxt
->sge
[0].addr
)) {
1322 svc_rdma_put_context(ctxt
, 1);
1325 atomic_inc(&xprt
->sc_dma_used
);
1326 ctxt
->sge
[0].lkey
= xprt
->sc_dma_lkey
;
1327 ctxt
->sge
[0].length
= length
;
1329 /* Prepare SEND WR */
1330 memset(&err_wr
, 0, sizeof err_wr
);
1331 ctxt
->wr_op
= IB_WR_SEND
;
1332 err_wr
.wr_id
= (unsigned long)ctxt
;
1333 err_wr
.sg_list
= ctxt
->sge
;
1335 err_wr
.opcode
= IB_WR_SEND
;
1336 err_wr
.send_flags
= IB_SEND_SIGNALED
;
1339 ret
= svc_rdma_send(xprt
, &err_wr
);
1341 dprintk("svcrdma: Error %d posting send for protocol error\n",
1343 svc_rdma_unmap_dma(ctxt
);
1344 svc_rdma_put_context(ctxt
, 1);