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
;
103 ctxt
= kmem_cache_alloc(svc_rdma_ctxt_cachep
, GFP_KERNEL
);
106 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
109 INIT_LIST_HEAD(&ctxt
->dto_q
);
112 atomic_inc(&xprt
->sc_ctxt_used
);
116 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt
*ctxt
)
118 struct svcxprt_rdma
*xprt
= ctxt
->xprt
;
120 for (i
= 0; i
< ctxt
->count
&& ctxt
->sge
[i
].length
; i
++) {
122 * Unmap the DMA addr in the SGE if the lkey matches
123 * the sc_dma_lkey, otherwise, ignore it since it is
124 * an FRMR lkey and will be unmapped later when the
125 * last WR that uses it completes.
127 if (ctxt
->sge
[i
].lkey
== xprt
->sc_dma_lkey
) {
128 atomic_dec(&xprt
->sc_dma_used
);
129 ib_dma_unmap_page(xprt
->sc_cm_id
->device
,
137 void svc_rdma_put_context(struct svc_rdma_op_ctxt
*ctxt
, int free_pages
)
139 struct svcxprt_rdma
*xprt
;
144 for (i
= 0; i
< ctxt
->count
; i
++)
145 put_page(ctxt
->pages
[i
]);
147 kmem_cache_free(svc_rdma_ctxt_cachep
, ctxt
);
148 atomic_dec(&xprt
->sc_ctxt_used
);
152 * Temporary NFS req mappings are shared across all transport
153 * instances. These are short lived and should be bounded by the number
154 * of concurrent server threads * depth of the SQ.
156 struct svc_rdma_req_map
*svc_rdma_get_req_map(void)
158 struct svc_rdma_req_map
*map
;
160 map
= kmem_cache_alloc(svc_rdma_map_cachep
, GFP_KERNEL
);
163 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
169 void svc_rdma_put_req_map(struct svc_rdma_req_map
*map
)
171 kmem_cache_free(svc_rdma_map_cachep
, map
);
174 /* ib_cq event handler */
175 static void cq_event_handler(struct ib_event
*event
, void *context
)
177 struct svc_xprt
*xprt
= context
;
178 dprintk("svcrdma: received CQ event %s (%d), context=%p\n",
179 ib_event_msg(event
->event
), event
->event
, context
);
180 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
183 /* QP event handler */
184 static void qp_event_handler(struct ib_event
*event
, void *context
)
186 struct svc_xprt
*xprt
= context
;
188 switch (event
->event
) {
189 /* These are considered benign events */
190 case IB_EVENT_PATH_MIG
:
191 case IB_EVENT_COMM_EST
:
192 case IB_EVENT_SQ_DRAINED
:
193 case IB_EVENT_QP_LAST_WQE_REACHED
:
194 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
195 ib_event_msg(event
->event
), event
->event
,
198 /* These are considered fatal events */
199 case IB_EVENT_PATH_MIG_ERR
:
200 case IB_EVENT_QP_FATAL
:
201 case IB_EVENT_QP_REQ_ERR
:
202 case IB_EVENT_QP_ACCESS_ERR
:
203 case IB_EVENT_DEVICE_FATAL
:
205 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
206 "closing transport\n",
207 ib_event_msg(event
->event
), event
->event
,
209 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
215 * Data Transfer Operation Tasklet
217 * Walks a list of transports with I/O pending, removing entries as
218 * they are added to the server's I/O pending list. Two bits indicate
219 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
220 * spinlock that serializes access to the transport list with the RQ
221 * and SQ interrupt handlers.
223 static void dto_tasklet_func(unsigned long data
)
225 struct svcxprt_rdma
*xprt
;
228 spin_lock_irqsave(&dto_lock
, flags
);
229 while (!list_empty(&dto_xprt_q
)) {
230 xprt
= list_entry(dto_xprt_q
.next
,
231 struct svcxprt_rdma
, sc_dto_q
);
232 list_del_init(&xprt
->sc_dto_q
);
233 spin_unlock_irqrestore(&dto_lock
, flags
);
238 svc_xprt_put(&xprt
->sc_xprt
);
239 spin_lock_irqsave(&dto_lock
, flags
);
241 spin_unlock_irqrestore(&dto_lock
, flags
);
245 * Receive Queue Completion Handler
247 * Since an RQ completion handler is called on interrupt context, we
248 * need to defer the handling of the I/O to a tasklet
250 static void rq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
252 struct svcxprt_rdma
*xprt
= cq_context
;
255 /* Guard against unconditional flush call for destroyed QP */
256 if (atomic_read(&xprt
->sc_xprt
.xpt_ref
.refcount
)==0)
260 * Set the bit regardless of whether or not it's on the list
261 * because it may be on the list already due to an SQ
264 set_bit(RDMAXPRT_RQ_PENDING
, &xprt
->sc_flags
);
267 * If this transport is not already on the DTO transport queue,
270 spin_lock_irqsave(&dto_lock
, flags
);
271 if (list_empty(&xprt
->sc_dto_q
)) {
272 svc_xprt_get(&xprt
->sc_xprt
);
273 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
275 spin_unlock_irqrestore(&dto_lock
, flags
);
277 /* Tasklet does all the work to avoid irqsave locks. */
278 tasklet_schedule(&dto_tasklet
);
282 * rq_cq_reap - Process the RQ CQ.
284 * Take all completing WC off the CQE and enqueue the associated DTO
285 * context on the dto_q for the transport.
287 * Note that caller must hold a transport reference.
289 static void rq_cq_reap(struct svcxprt_rdma
*xprt
)
293 struct svc_rdma_op_ctxt
*ctxt
= NULL
;
295 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING
, &xprt
->sc_flags
))
298 ib_req_notify_cq(xprt
->sc_rq_cq
, IB_CQ_NEXT_COMP
);
299 atomic_inc(&rdma_stat_rq_poll
);
301 while ((ret
= ib_poll_cq(xprt
->sc_rq_cq
, 1, &wc
)) > 0) {
302 ctxt
= (struct svc_rdma_op_ctxt
*)(unsigned long)wc
.wr_id
;
303 ctxt
->wc_status
= wc
.status
;
304 ctxt
->byte_len
= wc
.byte_len
;
305 svc_rdma_unmap_dma(ctxt
);
306 if (wc
.status
!= IB_WC_SUCCESS
) {
307 /* Close the transport */
308 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt
);
309 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
310 svc_rdma_put_context(ctxt
, 1);
311 svc_xprt_put(&xprt
->sc_xprt
);
314 spin_lock_bh(&xprt
->sc_rq_dto_lock
);
315 list_add_tail(&ctxt
->dto_q
, &xprt
->sc_rq_dto_q
);
316 spin_unlock_bh(&xprt
->sc_rq_dto_lock
);
317 svc_xprt_put(&xprt
->sc_xprt
);
321 atomic_inc(&rdma_stat_rq_prod
);
323 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
325 * If data arrived before established event,
326 * don't enqueue. This defers RPC I/O until the
327 * RDMA connection is complete.
329 if (!test_bit(RDMAXPRT_CONN_PENDING
, &xprt
->sc_flags
))
330 svc_xprt_enqueue(&xprt
->sc_xprt
);
334 * Process a completion context
336 static void process_context(struct svcxprt_rdma
*xprt
,
337 struct svc_rdma_op_ctxt
*ctxt
)
339 svc_rdma_unmap_dma(ctxt
);
341 switch (ctxt
->wr_op
) {
344 pr_err("svcrdma: SEND: ctxt->frmr != NULL\n");
345 svc_rdma_put_context(ctxt
, 1);
348 case IB_WR_RDMA_WRITE
:
350 pr_err("svcrdma: WRITE: ctxt->frmr != NULL\n");
351 svc_rdma_put_context(ctxt
, 0);
354 case IB_WR_RDMA_READ
:
355 case IB_WR_RDMA_READ_WITH_INV
:
356 svc_rdma_put_frmr(xprt
, ctxt
->frmr
);
357 if (test_bit(RDMACTXT_F_LAST_CTXT
, &ctxt
->flags
)) {
358 struct svc_rdma_op_ctxt
*read_hdr
= ctxt
->read_hdr
;
360 spin_lock_bh(&xprt
->sc_rq_dto_lock
);
361 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
362 list_add_tail(&read_hdr
->dto_q
,
363 &xprt
->sc_read_complete_q
);
364 spin_unlock_bh(&xprt
->sc_rq_dto_lock
);
366 pr_err("svcrdma: ctxt->read_hdr == NULL\n");
368 svc_xprt_enqueue(&xprt
->sc_xprt
);
370 svc_rdma_put_context(ctxt
, 0);
374 printk(KERN_ERR
"svcrdma: unexpected completion type, "
382 * Send Queue Completion Handler - potentially called on interrupt context.
384 * Note that caller must hold a transport reference.
386 static void sq_cq_reap(struct svcxprt_rdma
*xprt
)
388 struct svc_rdma_op_ctxt
*ctxt
= NULL
;
389 struct ib_wc wc_a
[6];
391 struct ib_cq
*cq
= xprt
->sc_sq_cq
;
394 memset(wc_a
, 0, sizeof(wc_a
));
396 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
))
399 ib_req_notify_cq(xprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
400 atomic_inc(&rdma_stat_sq_poll
);
401 while ((ret
= ib_poll_cq(cq
, ARRAY_SIZE(wc_a
), wc_a
)) > 0) {
404 for (i
= 0; i
< ret
; i
++) {
406 if (wc
->status
!= IB_WC_SUCCESS
) {
407 dprintk("svcrdma: sq wc err status %s (%d)\n",
408 ib_wc_status_msg(wc
->status
),
411 /* Close the transport */
412 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
415 /* Decrement used SQ WR count */
416 atomic_dec(&xprt
->sc_sq_count
);
417 wake_up(&xprt
->sc_send_wait
);
419 ctxt
= (struct svc_rdma_op_ctxt
*)
420 (unsigned long)wc
->wr_id
;
422 process_context(xprt
, ctxt
);
424 svc_xprt_put(&xprt
->sc_xprt
);
429 atomic_inc(&rdma_stat_sq_prod
);
432 static void sq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
434 struct svcxprt_rdma
*xprt
= cq_context
;
437 /* Guard against unconditional flush call for destroyed QP */
438 if (atomic_read(&xprt
->sc_xprt
.xpt_ref
.refcount
)==0)
442 * Set the bit regardless of whether or not it's on the list
443 * because it may be on the list already due to an RQ
446 set_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
);
449 * If this transport is not already on the DTO transport queue,
452 spin_lock_irqsave(&dto_lock
, flags
);
453 if (list_empty(&xprt
->sc_dto_q
)) {
454 svc_xprt_get(&xprt
->sc_xprt
);
455 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
457 spin_unlock_irqrestore(&dto_lock
, flags
);
459 /* Tasklet does all the work to avoid irqsave locks. */
460 tasklet_schedule(&dto_tasklet
);
463 static struct svcxprt_rdma
*rdma_create_xprt(struct svc_serv
*serv
,
466 struct svcxprt_rdma
*cma_xprt
= kzalloc(sizeof *cma_xprt
, GFP_KERNEL
);
470 svc_xprt_init(&init_net
, &svc_rdma_class
, &cma_xprt
->sc_xprt
, serv
);
471 INIT_LIST_HEAD(&cma_xprt
->sc_accept_q
);
472 INIT_LIST_HEAD(&cma_xprt
->sc_dto_q
);
473 INIT_LIST_HEAD(&cma_xprt
->sc_rq_dto_q
);
474 INIT_LIST_HEAD(&cma_xprt
->sc_read_complete_q
);
475 INIT_LIST_HEAD(&cma_xprt
->sc_frmr_q
);
476 init_waitqueue_head(&cma_xprt
->sc_send_wait
);
478 spin_lock_init(&cma_xprt
->sc_lock
);
479 spin_lock_init(&cma_xprt
->sc_rq_dto_lock
);
480 spin_lock_init(&cma_xprt
->sc_frmr_q_lock
);
482 cma_xprt
->sc_ord
= svcrdma_ord
;
484 cma_xprt
->sc_max_req_size
= svcrdma_max_req_size
;
485 cma_xprt
->sc_max_requests
= svcrdma_max_requests
;
486 cma_xprt
->sc_sq_depth
= svcrdma_max_requests
* RPCRDMA_SQ_DEPTH_MULT
;
487 atomic_set(&cma_xprt
->sc_sq_count
, 0);
488 atomic_set(&cma_xprt
->sc_ctxt_used
, 0);
491 set_bit(XPT_LISTENER
, &cma_xprt
->sc_xprt
.xpt_flags
);
496 struct page
*svc_rdma_get_page(void)
500 while ((page
= alloc_page(GFP_KERNEL
)) == NULL
) {
501 /* If we can't get memory, wait a bit and try again */
502 printk(KERN_INFO
"svcrdma: out of memory...retrying in 1s\n");
503 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
508 int svc_rdma_post_recv(struct svcxprt_rdma
*xprt
)
510 struct ib_recv_wr recv_wr
, *bad_recv_wr
;
511 struct svc_rdma_op_ctxt
*ctxt
;
518 ctxt
= svc_rdma_get_context(xprt
);
520 ctxt
->direction
= DMA_FROM_DEVICE
;
521 for (sge_no
= 0; buflen
< xprt
->sc_max_req_size
; sge_no
++) {
522 if (sge_no
>= xprt
->sc_max_sge
) {
523 pr_err("svcrdma: Too many sges (%d)\n", sge_no
);
526 page
= svc_rdma_get_page();
527 ctxt
->pages
[sge_no
] = page
;
528 pa
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
531 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
, pa
))
533 atomic_inc(&xprt
->sc_dma_used
);
534 ctxt
->sge
[sge_no
].addr
= pa
;
535 ctxt
->sge
[sge_no
].length
= PAGE_SIZE
;
536 ctxt
->sge
[sge_no
].lkey
= xprt
->sc_dma_lkey
;
537 ctxt
->count
= sge_no
+ 1;
541 recv_wr
.sg_list
= &ctxt
->sge
[0];
542 recv_wr
.num_sge
= ctxt
->count
;
543 recv_wr
.wr_id
= (u64
)(unsigned long)ctxt
;
545 svc_xprt_get(&xprt
->sc_xprt
);
546 ret
= ib_post_recv(xprt
->sc_qp
, &recv_wr
, &bad_recv_wr
);
548 svc_rdma_unmap_dma(ctxt
);
549 svc_rdma_put_context(ctxt
, 1);
550 svc_xprt_put(&xprt
->sc_xprt
);
555 svc_rdma_unmap_dma(ctxt
);
556 svc_rdma_put_context(ctxt
, 1);
561 * This function handles the CONNECT_REQUEST event on a listening
562 * endpoint. It is passed the cma_id for the _new_ connection. The context in
563 * this cma_id is inherited from the listening cma_id and is the svc_xprt
564 * structure for the listening endpoint.
566 * This function creates a new xprt for the new connection and enqueues it on
567 * the accept queue for the listent xprt. When the listen thread is kicked, it
568 * will call the recvfrom method on the listen xprt which will accept the new
571 static void handle_connect_req(struct rdma_cm_id
*new_cma_id
, size_t client_ird
)
573 struct svcxprt_rdma
*listen_xprt
= new_cma_id
->context
;
574 struct svcxprt_rdma
*newxprt
;
577 /* Create a new transport */
578 newxprt
= rdma_create_xprt(listen_xprt
->sc_xprt
.xpt_server
, 0);
580 dprintk("svcrdma: failed to create new transport\n");
583 newxprt
->sc_cm_id
= new_cma_id
;
584 new_cma_id
->context
= newxprt
;
585 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
586 newxprt
, newxprt
->sc_cm_id
, listen_xprt
);
588 /* Save client advertised inbound read limit for use later in accept. */
589 newxprt
->sc_ord
= client_ird
;
591 /* Set the local and remote addresses in the transport */
592 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.dst_addr
;
593 svc_xprt_set_remote(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
594 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.src_addr
;
595 svc_xprt_set_local(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
598 * Enqueue the new transport on the accept queue of the listening
601 spin_lock_bh(&listen_xprt
->sc_lock
);
602 list_add_tail(&newxprt
->sc_accept_q
, &listen_xprt
->sc_accept_q
);
603 spin_unlock_bh(&listen_xprt
->sc_lock
);
605 set_bit(XPT_CONN
, &listen_xprt
->sc_xprt
.xpt_flags
);
606 svc_xprt_enqueue(&listen_xprt
->sc_xprt
);
610 * Handles events generated on the listening endpoint. These events will be
611 * either be incoming connect requests or adapter removal events.
613 static int rdma_listen_handler(struct rdma_cm_id
*cma_id
,
614 struct rdma_cm_event
*event
)
616 struct svcxprt_rdma
*xprt
= cma_id
->context
;
619 switch (event
->event
) {
620 case RDMA_CM_EVENT_CONNECT_REQUEST
:
621 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
622 "event = %s (%d)\n", cma_id
, cma_id
->context
,
623 rdma_event_msg(event
->event
), event
->event
);
624 handle_connect_req(cma_id
,
625 event
->param
.conn
.initiator_depth
);
628 case RDMA_CM_EVENT_ESTABLISHED
:
629 /* Accept complete */
630 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
631 "cm_id=%p\n", xprt
, cma_id
);
634 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
635 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
638 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
642 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
643 "event = %s (%d)\n", cma_id
,
644 rdma_event_msg(event
->event
), event
->event
);
651 static int rdma_cma_handler(struct rdma_cm_id
*cma_id
,
652 struct rdma_cm_event
*event
)
654 struct svc_xprt
*xprt
= cma_id
->context
;
655 struct svcxprt_rdma
*rdma
=
656 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
657 switch (event
->event
) {
658 case RDMA_CM_EVENT_ESTABLISHED
:
659 /* Accept complete */
661 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
662 "cm_id=%p\n", xprt
, cma_id
);
663 clear_bit(RDMAXPRT_CONN_PENDING
, &rdma
->sc_flags
);
664 svc_xprt_enqueue(xprt
);
666 case RDMA_CM_EVENT_DISCONNECTED
:
667 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
670 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
671 svc_xprt_enqueue(xprt
);
675 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
676 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
677 "event = %s (%d)\n", cma_id
, xprt
,
678 rdma_event_msg(event
->event
), event
->event
);
680 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
681 svc_xprt_enqueue(xprt
);
685 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
686 "event = %s (%d)\n", cma_id
,
687 rdma_event_msg(event
->event
), event
->event
);
694 * Create a listening RDMA service endpoint.
696 static struct svc_xprt
*svc_rdma_create(struct svc_serv
*serv
,
698 struct sockaddr
*sa
, int salen
,
701 struct rdma_cm_id
*listen_id
;
702 struct svcxprt_rdma
*cma_xprt
;
705 dprintk("svcrdma: Creating RDMA socket\n");
706 if (sa
->sa_family
!= AF_INET
) {
707 dprintk("svcrdma: Address family %d is not supported.\n", sa
->sa_family
);
708 return ERR_PTR(-EAFNOSUPPORT
);
710 cma_xprt
= rdma_create_xprt(serv
, 1);
712 return ERR_PTR(-ENOMEM
);
714 listen_id
= rdma_create_id(rdma_listen_handler
, cma_xprt
, RDMA_PS_TCP
,
716 if (IS_ERR(listen_id
)) {
717 ret
= PTR_ERR(listen_id
);
718 dprintk("svcrdma: rdma_create_id failed = %d\n", ret
);
722 ret
= rdma_bind_addr(listen_id
, sa
);
724 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret
);
727 cma_xprt
->sc_cm_id
= listen_id
;
729 ret
= rdma_listen(listen_id
, RPCRDMA_LISTEN_BACKLOG
);
731 dprintk("svcrdma: rdma_listen failed = %d\n", ret
);
736 * We need to use the address from the cm_id in case the
737 * caller specified 0 for the port number.
739 sa
= (struct sockaddr
*)&cma_xprt
->sc_cm_id
->route
.addr
.src_addr
;
740 svc_xprt_set_local(&cma_xprt
->sc_xprt
, sa
, salen
);
742 return &cma_xprt
->sc_xprt
;
745 rdma_destroy_id(listen_id
);
751 static struct svc_rdma_fastreg_mr
*rdma_alloc_frmr(struct svcxprt_rdma
*xprt
)
754 struct ib_fast_reg_page_list
*pl
;
755 struct svc_rdma_fastreg_mr
*frmr
;
757 frmr
= kmalloc(sizeof(*frmr
), GFP_KERNEL
);
761 mr
= ib_alloc_fast_reg_mr(xprt
->sc_pd
, RPCSVC_MAXPAGES
);
765 pl
= ib_alloc_fast_reg_page_list(xprt
->sc_cm_id
->device
,
771 frmr
->page_list
= pl
;
772 INIT_LIST_HEAD(&frmr
->frmr_list
);
780 return ERR_PTR(-ENOMEM
);
783 static void rdma_dealloc_frmr_q(struct svcxprt_rdma
*xprt
)
785 struct svc_rdma_fastreg_mr
*frmr
;
787 while (!list_empty(&xprt
->sc_frmr_q
)) {
788 frmr
= list_entry(xprt
->sc_frmr_q
.next
,
789 struct svc_rdma_fastreg_mr
, frmr_list
);
790 list_del_init(&frmr
->frmr_list
);
791 ib_dereg_mr(frmr
->mr
);
792 ib_free_fast_reg_page_list(frmr
->page_list
);
797 struct svc_rdma_fastreg_mr
*svc_rdma_get_frmr(struct svcxprt_rdma
*rdma
)
799 struct svc_rdma_fastreg_mr
*frmr
= NULL
;
801 spin_lock_bh(&rdma
->sc_frmr_q_lock
);
802 if (!list_empty(&rdma
->sc_frmr_q
)) {
803 frmr
= list_entry(rdma
->sc_frmr_q
.next
,
804 struct svc_rdma_fastreg_mr
, frmr_list
);
805 list_del_init(&frmr
->frmr_list
);
807 frmr
->page_list_len
= 0;
809 spin_unlock_bh(&rdma
->sc_frmr_q_lock
);
813 return rdma_alloc_frmr(rdma
);
816 static void frmr_unmap_dma(struct svcxprt_rdma
*xprt
,
817 struct svc_rdma_fastreg_mr
*frmr
)
820 for (page_no
= 0; page_no
< frmr
->page_list_len
; page_no
++) {
821 dma_addr_t addr
= frmr
->page_list
->page_list
[page_no
];
822 if (ib_dma_mapping_error(frmr
->mr
->device
, addr
))
824 atomic_dec(&xprt
->sc_dma_used
);
825 ib_dma_unmap_page(frmr
->mr
->device
, addr
, PAGE_SIZE
,
830 void svc_rdma_put_frmr(struct svcxprt_rdma
*rdma
,
831 struct svc_rdma_fastreg_mr
*frmr
)
834 frmr_unmap_dma(rdma
, frmr
);
835 spin_lock_bh(&rdma
->sc_frmr_q_lock
);
836 WARN_ON_ONCE(!list_empty(&frmr
->frmr_list
));
837 list_add(&frmr
->frmr_list
, &rdma
->sc_frmr_q
);
838 spin_unlock_bh(&rdma
->sc_frmr_q_lock
);
843 * This is the xpo_recvfrom function for listening endpoints. Its
844 * purpose is to accept incoming connections. The CMA callback handler
845 * has already created a new transport and attached it to the new CMA
848 * There is a queue of pending connections hung on the listening
849 * transport. This queue contains the new svc_xprt structure. This
850 * function takes svc_xprt structures off the accept_q and completes
853 static struct svc_xprt
*svc_rdma_accept(struct svc_xprt
*xprt
)
855 struct svcxprt_rdma
*listen_rdma
;
856 struct svcxprt_rdma
*newxprt
= NULL
;
857 struct rdma_conn_param conn_param
;
858 struct ib_cq_init_attr cq_attr
= {};
859 struct ib_qp_init_attr qp_attr
;
860 struct ib_device_attr devattr
;
861 int uninitialized_var(dma_mr_acc
);
866 listen_rdma
= container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
867 clear_bit(XPT_CONN
, &xprt
->xpt_flags
);
868 /* Get the next entry off the accept list */
869 spin_lock_bh(&listen_rdma
->sc_lock
);
870 if (!list_empty(&listen_rdma
->sc_accept_q
)) {
871 newxprt
= list_entry(listen_rdma
->sc_accept_q
.next
,
872 struct svcxprt_rdma
, sc_accept_q
);
873 list_del_init(&newxprt
->sc_accept_q
);
875 if (!list_empty(&listen_rdma
->sc_accept_q
))
876 set_bit(XPT_CONN
, &listen_rdma
->sc_xprt
.xpt_flags
);
877 spin_unlock_bh(&listen_rdma
->sc_lock
);
881 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
882 newxprt
, newxprt
->sc_cm_id
);
884 ret
= ib_query_device(newxprt
->sc_cm_id
->device
, &devattr
);
886 dprintk("svcrdma: could not query device attributes on "
887 "device %p, rc=%d\n", newxprt
->sc_cm_id
->device
, ret
);
891 /* Qualify the transport resource defaults with the
892 * capabilities of this particular device */
893 newxprt
->sc_max_sge
= min((size_t)devattr
.max_sge
,
894 (size_t)RPCSVC_MAXPAGES
);
895 newxprt
->sc_max_requests
= min((size_t)devattr
.max_qp_wr
,
896 (size_t)svcrdma_max_requests
);
897 newxprt
->sc_sq_depth
= RPCRDMA_SQ_DEPTH_MULT
* newxprt
->sc_max_requests
;
900 * Limit ORD based on client limit, local device limit, and
901 * configured svcrdma limit.
903 newxprt
->sc_ord
= min_t(size_t, devattr
.max_qp_rd_atom
, newxprt
->sc_ord
);
904 newxprt
->sc_ord
= min_t(size_t, svcrdma_ord
, newxprt
->sc_ord
);
906 newxprt
->sc_pd
= ib_alloc_pd(newxprt
->sc_cm_id
->device
);
907 if (IS_ERR(newxprt
->sc_pd
)) {
908 dprintk("svcrdma: error creating PD for connect request\n");
911 cq_attr
.cqe
= newxprt
->sc_sq_depth
;
912 newxprt
->sc_sq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
917 if (IS_ERR(newxprt
->sc_sq_cq
)) {
918 dprintk("svcrdma: error creating SQ CQ for connect request\n");
921 cq_attr
.cqe
= newxprt
->sc_max_requests
;
922 newxprt
->sc_rq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
927 if (IS_ERR(newxprt
->sc_rq_cq
)) {
928 dprintk("svcrdma: error creating RQ CQ for connect request\n");
932 memset(&qp_attr
, 0, sizeof qp_attr
);
933 qp_attr
.event_handler
= qp_event_handler
;
934 qp_attr
.qp_context
= &newxprt
->sc_xprt
;
935 qp_attr
.cap
.max_send_wr
= newxprt
->sc_sq_depth
;
936 qp_attr
.cap
.max_recv_wr
= newxprt
->sc_max_requests
;
937 qp_attr
.cap
.max_send_sge
= newxprt
->sc_max_sge
;
938 qp_attr
.cap
.max_recv_sge
= newxprt
->sc_max_sge
;
939 qp_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
940 qp_attr
.qp_type
= IB_QPT_RC
;
941 qp_attr
.send_cq
= newxprt
->sc_sq_cq
;
942 qp_attr
.recv_cq
= newxprt
->sc_rq_cq
;
943 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
944 " cm_id->device=%p, sc_pd->device=%p\n"
945 " cap.max_send_wr = %d\n"
946 " cap.max_recv_wr = %d\n"
947 " cap.max_send_sge = %d\n"
948 " cap.max_recv_sge = %d\n",
949 newxprt
->sc_cm_id
, newxprt
->sc_pd
,
950 newxprt
->sc_cm_id
->device
, newxprt
->sc_pd
->device
,
951 qp_attr
.cap
.max_send_wr
,
952 qp_attr
.cap
.max_recv_wr
,
953 qp_attr
.cap
.max_send_sge
,
954 qp_attr
.cap
.max_recv_sge
);
956 ret
= rdma_create_qp(newxprt
->sc_cm_id
, newxprt
->sc_pd
, &qp_attr
);
958 dprintk("svcrdma: failed to create QP, ret=%d\n", ret
);
961 newxprt
->sc_qp
= newxprt
->sc_cm_id
->qp
;
964 * Use the most secure set of MR resources based on the
965 * transport type and available memory management features in
966 * the device. Here's the table implemented below:
968 * Fast Global DMA Remote WR
970 * Sup'd Sup'd Needed Needed
982 * NB: iWARP requires remote write access for the data sink
983 * of an RDMA_READ. IB does not.
985 newxprt
->sc_reader
= rdma_read_chunk_lcl
;
986 if (devattr
.device_cap_flags
& IB_DEVICE_MEM_MGT_EXTENSIONS
) {
987 newxprt
->sc_frmr_pg_list_len
=
988 devattr
.max_fast_reg_page_list_len
;
989 newxprt
->sc_dev_caps
|= SVCRDMA_DEVCAP_FAST_REG
;
990 newxprt
->sc_reader
= rdma_read_chunk_frmr
;
994 * Determine if a DMA MR is required and if so, what privs are required
996 if (!rdma_protocol_iwarp(newxprt
->sc_cm_id
->device
,
997 newxprt
->sc_cm_id
->port_num
) &&
998 !rdma_ib_or_roce(newxprt
->sc_cm_id
->device
,
999 newxprt
->sc_cm_id
->port_num
))
1002 if (!(newxprt
->sc_dev_caps
& SVCRDMA_DEVCAP_FAST_REG
) ||
1003 !(devattr
.device_cap_flags
& IB_DEVICE_LOCAL_DMA_LKEY
)) {
1005 dma_mr_acc
= IB_ACCESS_LOCAL_WRITE
;
1006 if (rdma_protocol_iwarp(newxprt
->sc_cm_id
->device
,
1007 newxprt
->sc_cm_id
->port_num
) &&
1008 !(newxprt
->sc_dev_caps
& SVCRDMA_DEVCAP_FAST_REG
))
1009 dma_mr_acc
|= IB_ACCESS_REMOTE_WRITE
;
1012 if (rdma_protocol_iwarp(newxprt
->sc_cm_id
->device
,
1013 newxprt
->sc_cm_id
->port_num
))
1014 newxprt
->sc_dev_caps
|= SVCRDMA_DEVCAP_READ_W_INV
;
1016 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */
1018 /* Register all of physical memory */
1019 newxprt
->sc_phys_mr
=
1020 ib_get_dma_mr(newxprt
->sc_pd
, dma_mr_acc
);
1021 if (IS_ERR(newxprt
->sc_phys_mr
)) {
1022 dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
1026 newxprt
->sc_dma_lkey
= newxprt
->sc_phys_mr
->lkey
;
1028 newxprt
->sc_dma_lkey
=
1029 newxprt
->sc_cm_id
->device
->local_dma_lkey
;
1031 /* Post receive buffers */
1032 for (i
= 0; i
< newxprt
->sc_max_requests
; i
++) {
1033 ret
= svc_rdma_post_recv(newxprt
);
1035 dprintk("svcrdma: failure posting receive buffers\n");
1040 /* Swap out the handler */
1041 newxprt
->sc_cm_id
->event_handler
= rdma_cma_handler
;
1044 * Arm the CQs for the SQ and RQ before accepting so we can't
1045 * miss the first message
1047 ib_req_notify_cq(newxprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
1048 ib_req_notify_cq(newxprt
->sc_rq_cq
, IB_CQ_NEXT_COMP
);
1050 /* Accept Connection */
1051 set_bit(RDMAXPRT_CONN_PENDING
, &newxprt
->sc_flags
);
1052 memset(&conn_param
, 0, sizeof conn_param
);
1053 conn_param
.responder_resources
= 0;
1054 conn_param
.initiator_depth
= newxprt
->sc_ord
;
1055 ret
= rdma_accept(newxprt
->sc_cm_id
, &conn_param
);
1057 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1062 dprintk("svcrdma: new connection %p accepted with the following "
1064 " local_ip : %pI4\n"
1065 " local_port : %d\n"
1066 " remote_ip : %pI4\n"
1067 " remote_port : %d\n"
1070 " max_requests : %d\n"
1073 &((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1074 route
.addr
.src_addr
)->sin_addr
.s_addr
,
1075 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1076 route
.addr
.src_addr
)->sin_port
),
1077 &((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1078 route
.addr
.dst_addr
)->sin_addr
.s_addr
,
1079 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
1080 route
.addr
.dst_addr
)->sin_port
),
1081 newxprt
->sc_max_sge
,
1082 newxprt
->sc_sq_depth
,
1083 newxprt
->sc_max_requests
,
1086 return &newxprt
->sc_xprt
;
1089 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret
);
1090 /* Take a reference in case the DTO handler runs */
1091 svc_xprt_get(&newxprt
->sc_xprt
);
1092 if (newxprt
->sc_qp
&& !IS_ERR(newxprt
->sc_qp
))
1093 ib_destroy_qp(newxprt
->sc_qp
);
1094 rdma_destroy_id(newxprt
->sc_cm_id
);
1095 /* This call to put will destroy the transport */
1096 svc_xprt_put(&newxprt
->sc_xprt
);
1100 static void svc_rdma_release_rqst(struct svc_rqst
*rqstp
)
1105 * When connected, an svc_xprt has at least two references:
1107 * - A reference held by the cm_id between the ESTABLISHED and
1108 * DISCONNECTED events. If the remote peer disconnected first, this
1109 * reference could be gone.
1111 * - A reference held by the svc_recv code that called this function
1112 * as part of close processing.
1114 * At a minimum one references should still be held.
1116 static void svc_rdma_detach(struct svc_xprt
*xprt
)
1118 struct svcxprt_rdma
*rdma
=
1119 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1120 dprintk("svc: svc_rdma_detach(%p)\n", xprt
);
1122 /* Disconnect and flush posted WQE */
1123 rdma_disconnect(rdma
->sc_cm_id
);
1126 static void __svc_rdma_free(struct work_struct
*work
)
1128 struct svcxprt_rdma
*rdma
=
1129 container_of(work
, struct svcxprt_rdma
, sc_work
);
1130 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma
);
1132 /* We should only be called from kref_put */
1133 if (atomic_read(&rdma
->sc_xprt
.xpt_ref
.refcount
) != 0)
1134 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1135 atomic_read(&rdma
->sc_xprt
.xpt_ref
.refcount
));
1138 * Destroy queued, but not processed read completions. Note
1139 * that this cleanup has to be done before destroying the
1140 * cm_id because the device ptr is needed to unmap the dma in
1141 * svc_rdma_put_context.
1143 while (!list_empty(&rdma
->sc_read_complete_q
)) {
1144 struct svc_rdma_op_ctxt
*ctxt
;
1145 ctxt
= list_entry(rdma
->sc_read_complete_q
.next
,
1146 struct svc_rdma_op_ctxt
,
1148 list_del_init(&ctxt
->dto_q
);
1149 svc_rdma_put_context(ctxt
, 1);
1152 /* Destroy queued, but not processed recv completions */
1153 while (!list_empty(&rdma
->sc_rq_dto_q
)) {
1154 struct svc_rdma_op_ctxt
*ctxt
;
1155 ctxt
= list_entry(rdma
->sc_rq_dto_q
.next
,
1156 struct svc_rdma_op_ctxt
,
1158 list_del_init(&ctxt
->dto_q
);
1159 svc_rdma_put_context(ctxt
, 1);
1162 /* Warn if we leaked a resource or under-referenced */
1163 if (atomic_read(&rdma
->sc_ctxt_used
) != 0)
1164 pr_err("svcrdma: ctxt still in use? (%d)\n",
1165 atomic_read(&rdma
->sc_ctxt_used
));
1166 if (atomic_read(&rdma
->sc_dma_used
) != 0)
1167 pr_err("svcrdma: dma still in use? (%d)\n",
1168 atomic_read(&rdma
->sc_dma_used
));
1170 /* De-allocate fastreg mr */
1171 rdma_dealloc_frmr_q(rdma
);
1173 /* Destroy the QP if present (not a listener) */
1174 if (rdma
->sc_qp
&& !IS_ERR(rdma
->sc_qp
))
1175 ib_destroy_qp(rdma
->sc_qp
);
1177 if (rdma
->sc_sq_cq
&& !IS_ERR(rdma
->sc_sq_cq
))
1178 ib_destroy_cq(rdma
->sc_sq_cq
);
1180 if (rdma
->sc_rq_cq
&& !IS_ERR(rdma
->sc_rq_cq
))
1181 ib_destroy_cq(rdma
->sc_rq_cq
);
1183 if (rdma
->sc_phys_mr
&& !IS_ERR(rdma
->sc_phys_mr
))
1184 ib_dereg_mr(rdma
->sc_phys_mr
);
1186 if (rdma
->sc_pd
&& !IS_ERR(rdma
->sc_pd
))
1187 ib_dealloc_pd(rdma
->sc_pd
);
1189 /* Destroy the CM ID */
1190 rdma_destroy_id(rdma
->sc_cm_id
);
1195 static void svc_rdma_free(struct svc_xprt
*xprt
)
1197 struct svcxprt_rdma
*rdma
=
1198 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1199 INIT_WORK(&rdma
->sc_work
, __svc_rdma_free
);
1200 queue_work(svc_rdma_wq
, &rdma
->sc_work
);
1203 static int svc_rdma_has_wspace(struct svc_xprt
*xprt
)
1205 struct svcxprt_rdma
*rdma
=
1206 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
1209 * If there are already waiters on the SQ,
1212 if (waitqueue_active(&rdma
->sc_send_wait
))
1215 /* Otherwise return true. */
1219 static int svc_rdma_secure_port(struct svc_rqst
*rqstp
)
1225 * Attempt to register the kvec representing the RPC memory with the
1229 * NULL : The device does not support fastreg or there were no more
1231 * frmr : The kvec register request was successfully posted.
1232 * <0 : An error was encountered attempting to register the kvec.
1234 int svc_rdma_fastreg(struct svcxprt_rdma
*xprt
,
1235 struct svc_rdma_fastreg_mr
*frmr
)
1237 struct ib_send_wr fastreg_wr
;
1241 key
= (u8
)(frmr
->mr
->lkey
& 0x000000FF);
1242 ib_update_fast_reg_key(frmr
->mr
, ++key
);
1244 /* Prepare FASTREG WR */
1245 memset(&fastreg_wr
, 0, sizeof fastreg_wr
);
1246 fastreg_wr
.opcode
= IB_WR_FAST_REG_MR
;
1247 fastreg_wr
.send_flags
= IB_SEND_SIGNALED
;
1248 fastreg_wr
.wr
.fast_reg
.iova_start
= (unsigned long)frmr
->kva
;
1249 fastreg_wr
.wr
.fast_reg
.page_list
= frmr
->page_list
;
1250 fastreg_wr
.wr
.fast_reg
.page_list_len
= frmr
->page_list_len
;
1251 fastreg_wr
.wr
.fast_reg
.page_shift
= PAGE_SHIFT
;
1252 fastreg_wr
.wr
.fast_reg
.length
= frmr
->map_len
;
1253 fastreg_wr
.wr
.fast_reg
.access_flags
= frmr
->access_flags
;
1254 fastreg_wr
.wr
.fast_reg
.rkey
= frmr
->mr
->lkey
;
1255 return svc_rdma_send(xprt
, &fastreg_wr
);
1258 int svc_rdma_send(struct svcxprt_rdma
*xprt
, struct ib_send_wr
*wr
)
1260 struct ib_send_wr
*bad_wr
, *n_wr
;
1265 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1269 for (n_wr
= wr
->next
; n_wr
; n_wr
= n_wr
->next
)
1272 /* If the SQ is full, wait until an SQ entry is available */
1274 spin_lock_bh(&xprt
->sc_lock
);
1275 if (xprt
->sc_sq_depth
< atomic_read(&xprt
->sc_sq_count
) + wr_count
) {
1276 spin_unlock_bh(&xprt
->sc_lock
);
1277 atomic_inc(&rdma_stat_sq_starve
);
1279 /* See if we can opportunistically reap SQ WR to make room */
1282 /* Wait until SQ WR available if SQ still full */
1283 wait_event(xprt
->sc_send_wait
,
1284 atomic_read(&xprt
->sc_sq_count
) <
1286 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1290 /* Take a transport ref for each WR posted */
1291 for (i
= 0; i
< wr_count
; i
++)
1292 svc_xprt_get(&xprt
->sc_xprt
);
1294 /* Bump used SQ WR count and post */
1295 atomic_add(wr_count
, &xprt
->sc_sq_count
);
1296 ret
= ib_post_send(xprt
->sc_qp
, wr
, &bad_wr
);
1298 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
1299 atomic_sub(wr_count
, &xprt
->sc_sq_count
);
1300 for (i
= 0; i
< wr_count
; i
++)
1301 svc_xprt_put(&xprt
->sc_xprt
);
1302 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1303 "sc_sq_count=%d, sc_sq_depth=%d\n",
1304 ret
, atomic_read(&xprt
->sc_sq_count
),
1307 spin_unlock_bh(&xprt
->sc_lock
);
1309 wake_up(&xprt
->sc_send_wait
);
1315 void svc_rdma_send_error(struct svcxprt_rdma
*xprt
, struct rpcrdma_msg
*rmsgp
,
1316 enum rpcrdma_errcode err
)
1318 struct ib_send_wr err_wr
;
1320 struct svc_rdma_op_ctxt
*ctxt
;
1325 p
= svc_rdma_get_page();
1326 va
= page_address(p
);
1328 /* XDR encode error */
1329 length
= svc_rdma_xdr_encode_error(xprt
, rmsgp
, err
, va
);
1331 ctxt
= svc_rdma_get_context(xprt
);
1332 ctxt
->direction
= DMA_FROM_DEVICE
;
1336 /* Prepare SGE for local address */
1337 ctxt
->sge
[0].addr
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
1338 p
, 0, length
, DMA_FROM_DEVICE
);
1339 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
, ctxt
->sge
[0].addr
)) {
1341 svc_rdma_put_context(ctxt
, 1);
1344 atomic_inc(&xprt
->sc_dma_used
);
1345 ctxt
->sge
[0].lkey
= xprt
->sc_dma_lkey
;
1346 ctxt
->sge
[0].length
= length
;
1348 /* Prepare SEND WR */
1349 memset(&err_wr
, 0, sizeof err_wr
);
1350 ctxt
->wr_op
= IB_WR_SEND
;
1351 err_wr
.wr_id
= (unsigned long)ctxt
;
1352 err_wr
.sg_list
= ctxt
->sge
;
1354 err_wr
.opcode
= IB_WR_SEND
;
1355 err_wr
.send_flags
= IB_SEND_SIGNALED
;
1358 ret
= svc_rdma_send(xprt
, &err_wr
);
1360 dprintk("svcrdma: Error %d posting send for protocol error\n",
1362 svc_rdma_unmap_dma(ctxt
);
1363 svc_rdma_put_context(ctxt
, 1);