2 * Copyright (c) 2005-2007 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.
39 * Author: Tom Tucker <tom@opengridcomputing.com>
42 #include <linux/sunrpc/svc_xprt.h>
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/rdma_cm.h>
48 #include <linux/sunrpc/svc_rdma.h>
50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52 static struct svc_xprt
*svc_rdma_create(struct svc_serv
*serv
,
53 struct sockaddr
*sa
, int salen
,
55 static struct svc_xprt
*svc_rdma_accept(struct svc_xprt
*xprt
);
56 static void svc_rdma_release_rqst(struct svc_rqst
*);
57 static void dto_tasklet_func(unsigned long data
);
58 static void svc_rdma_detach(struct svc_xprt
*xprt
);
59 static void svc_rdma_free(struct svc_xprt
*xprt
);
60 static int svc_rdma_has_wspace(struct svc_xprt
*xprt
);
61 static void rq_cq_reap(struct svcxprt_rdma
*xprt
);
62 static void sq_cq_reap(struct svcxprt_rdma
*xprt
);
64 DECLARE_TASKLET(dto_tasklet
, dto_tasklet_func
, 0UL);
65 static DEFINE_SPINLOCK(dto_lock
);
66 static LIST_HEAD(dto_xprt_q
);
68 static struct svc_xprt_ops svc_rdma_ops
= {
69 .xpo_create
= svc_rdma_create
,
70 .xpo_recvfrom
= svc_rdma_recvfrom
,
71 .xpo_sendto
= svc_rdma_sendto
,
72 .xpo_release_rqst
= svc_rdma_release_rqst
,
73 .xpo_detach
= svc_rdma_detach
,
74 .xpo_free
= svc_rdma_free
,
75 .xpo_prep_reply_hdr
= svc_rdma_prep_reply_hdr
,
76 .xpo_has_wspace
= svc_rdma_has_wspace
,
77 .xpo_accept
= svc_rdma_accept
,
80 struct svc_xprt_class svc_rdma_class
= {
82 .xcl_owner
= THIS_MODULE
,
83 .xcl_ops
= &svc_rdma_ops
,
84 .xcl_max_payload
= RPCSVC_MAXPAYLOAD_TCP
,
87 static int rdma_bump_context_cache(struct svcxprt_rdma
*xprt
)
91 struct svc_rdma_op_ctxt
*ctxt
;
93 target
= min(xprt
->sc_ctxt_cnt
+ xprt
->sc_ctxt_bump
,
96 spin_lock_bh(&xprt
->sc_ctxt_lock
);
97 while (xprt
->sc_ctxt_cnt
< target
) {
99 spin_unlock_bh(&xprt
->sc_ctxt_lock
);
101 ctxt
= kmalloc(sizeof(*ctxt
), GFP_KERNEL
);
103 spin_lock_bh(&xprt
->sc_ctxt_lock
);
106 ctxt
->next
= xprt
->sc_ctxt_head
;
107 xprt
->sc_ctxt_head
= ctxt
;
109 /* kmalloc failed...give up for now */
114 spin_unlock_bh(&xprt
->sc_ctxt_lock
);
115 dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
116 xprt
->sc_ctxt_max
, xprt
->sc_ctxt_cnt
);
120 struct svc_rdma_op_ctxt
*svc_rdma_get_context(struct svcxprt_rdma
*xprt
)
122 struct svc_rdma_op_ctxt
*ctxt
;
125 spin_lock_bh(&xprt
->sc_ctxt_lock
);
126 if (unlikely(xprt
->sc_ctxt_head
== NULL
)) {
127 /* Try to bump my cache. */
128 spin_unlock_bh(&xprt
->sc_ctxt_lock
);
130 if (rdma_bump_context_cache(xprt
))
133 printk(KERN_INFO
"svcrdma: sleeping waiting for "
134 "context memory on xprt=%p\n",
136 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
139 ctxt
= xprt
->sc_ctxt_head
;
140 xprt
->sc_ctxt_head
= ctxt
->next
;
141 spin_unlock_bh(&xprt
->sc_ctxt_lock
);
143 INIT_LIST_HEAD(&ctxt
->dto_q
);
150 void svc_rdma_put_context(struct svc_rdma_op_ctxt
*ctxt
, int free_pages
)
152 struct svcxprt_rdma
*xprt
;
158 for (i
= 0; i
< ctxt
->count
; i
++)
159 put_page(ctxt
->pages
[i
]);
161 for (i
= 0; i
< ctxt
->count
; i
++)
162 dma_unmap_single(xprt
->sc_cm_id
->device
->dma_device
,
166 spin_lock_bh(&xprt
->sc_ctxt_lock
);
167 ctxt
->next
= xprt
->sc_ctxt_head
;
168 xprt
->sc_ctxt_head
= ctxt
;
169 spin_unlock_bh(&xprt
->sc_ctxt_lock
);
172 /* ib_cq event handler */
173 static void cq_event_handler(struct ib_event
*event
, void *context
)
175 struct svc_xprt
*xprt
= context
;
176 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
177 event
->event
, context
);
178 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
181 /* QP event handler */
182 static void qp_event_handler(struct ib_event
*event
, void *context
)
184 struct svc_xprt
*xprt
= context
;
186 switch (event
->event
) {
187 /* These are considered benign events */
188 case IB_EVENT_PATH_MIG
:
189 case IB_EVENT_COMM_EST
:
190 case IB_EVENT_SQ_DRAINED
:
191 case IB_EVENT_QP_LAST_WQE_REACHED
:
192 dprintk("svcrdma: QP event %d received for QP=%p\n",
193 event
->event
, event
->element
.qp
);
195 /* These are considered fatal events */
196 case IB_EVENT_PATH_MIG_ERR
:
197 case IB_EVENT_QP_FATAL
:
198 case IB_EVENT_QP_REQ_ERR
:
199 case IB_EVENT_QP_ACCESS_ERR
:
200 case IB_EVENT_DEVICE_FATAL
:
202 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
203 "closing transport\n",
204 event
->event
, event
->element
.qp
);
205 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
211 * Data Transfer Operation Tasklet
213 * Walks a list of transports with I/O pending, removing entries as
214 * they are added to the server's I/O pending list. Two bits indicate
215 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
216 * spinlock that serializes access to the transport list with the RQ
217 * and SQ interrupt handlers.
219 static void dto_tasklet_func(unsigned long data
)
221 struct svcxprt_rdma
*xprt
;
224 spin_lock_irqsave(&dto_lock
, flags
);
225 while (!list_empty(&dto_xprt_q
)) {
226 xprt
= list_entry(dto_xprt_q
.next
,
227 struct svcxprt_rdma
, sc_dto_q
);
228 list_del_init(&xprt
->sc_dto_q
);
229 spin_unlock_irqrestore(&dto_lock
, flags
);
234 svc_xprt_put(&xprt
->sc_xprt
);
235 spin_lock_irqsave(&dto_lock
, flags
);
237 spin_unlock_irqrestore(&dto_lock
, flags
);
241 * Receive Queue Completion Handler
243 * Since an RQ completion handler is called on interrupt context, we
244 * need to defer the handling of the I/O to a tasklet
246 static void rq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
248 struct svcxprt_rdma
*xprt
= cq_context
;
252 * Set the bit regardless of whether or not it's on the list
253 * because it may be on the list already due to an SQ
256 set_bit(RDMAXPRT_RQ_PENDING
, &xprt
->sc_flags
);
259 * If this transport is not already on the DTO transport queue,
262 spin_lock_irqsave(&dto_lock
, flags
);
263 if (list_empty(&xprt
->sc_dto_q
)) {
264 svc_xprt_get(&xprt
->sc_xprt
);
265 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
267 spin_unlock_irqrestore(&dto_lock
, flags
);
269 /* Tasklet does all the work to avoid irqsave locks. */
270 tasklet_schedule(&dto_tasklet
);
274 * rq_cq_reap - Process the RQ CQ.
276 * Take all completing WC off the CQE and enqueue the associated DTO
277 * context on the dto_q for the transport.
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 spin_lock_bh(&xprt
->sc_rq_dto_lock
);
292 while ((ret
= ib_poll_cq(xprt
->sc_rq_cq
, 1, &wc
)) > 0) {
293 ctxt
= (struct svc_rdma_op_ctxt
*)(unsigned long)wc
.wr_id
;
294 ctxt
->wc_status
= wc
.status
;
295 ctxt
->byte_len
= wc
.byte_len
;
296 if (wc
.status
!= IB_WC_SUCCESS
) {
297 /* Close the transport */
298 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
299 svc_rdma_put_context(ctxt
, 1);
302 list_add_tail(&ctxt
->dto_q
, &xprt
->sc_rq_dto_q
);
304 spin_unlock_bh(&xprt
->sc_rq_dto_lock
);
307 atomic_inc(&rdma_stat_rq_prod
);
309 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
311 * If data arrived before established event,
312 * don't enqueue. This defers RPC I/O until the
313 * RDMA connection is complete.
315 if (!test_bit(RDMAXPRT_CONN_PENDING
, &xprt
->sc_flags
))
316 svc_xprt_enqueue(&xprt
->sc_xprt
);
320 * Send Queue Completion Handler - potentially called on interrupt context.
322 static void sq_cq_reap(struct svcxprt_rdma
*xprt
)
324 struct svc_rdma_op_ctxt
*ctxt
= NULL
;
326 struct ib_cq
*cq
= xprt
->sc_sq_cq
;
330 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
))
333 ib_req_notify_cq(xprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
334 atomic_inc(&rdma_stat_sq_poll
);
335 while ((ret
= ib_poll_cq(cq
, 1, &wc
)) > 0) {
336 ctxt
= (struct svc_rdma_op_ctxt
*)(unsigned long)wc
.wr_id
;
339 if (wc
.status
!= IB_WC_SUCCESS
)
340 /* Close the transport */
341 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
343 /* Decrement used SQ WR count */
344 atomic_dec(&xprt
->sc_sq_count
);
345 wake_up(&xprt
->sc_send_wait
);
347 switch (ctxt
->wr_op
) {
349 case IB_WR_RDMA_WRITE
:
350 svc_rdma_put_context(ctxt
, 1);
353 case IB_WR_RDMA_READ
:
354 if (test_bit(RDMACTXT_F_LAST_CTXT
, &ctxt
->flags
)) {
355 set_bit(XPT_DATA
, &xprt
->sc_xprt
.xpt_flags
);
356 set_bit(RDMACTXT_F_READ_DONE
, &ctxt
->flags
);
357 spin_lock_bh(&xprt
->sc_read_complete_lock
);
358 list_add_tail(&ctxt
->dto_q
,
359 &xprt
->sc_read_complete_q
);
360 spin_unlock_bh(&xprt
->sc_read_complete_lock
);
361 svc_xprt_enqueue(&xprt
->sc_xprt
);
366 printk(KERN_ERR
"svcrdma: unexpected completion type, "
367 "opcode=%d, status=%d\n",
368 wc
.opcode
, wc
.status
);
374 atomic_inc(&rdma_stat_sq_prod
);
377 static void sq_comp_handler(struct ib_cq
*cq
, void *cq_context
)
379 struct svcxprt_rdma
*xprt
= cq_context
;
383 * Set the bit regardless of whether or not it's on the list
384 * because it may be on the list already due to an RQ
387 set_bit(RDMAXPRT_SQ_PENDING
, &xprt
->sc_flags
);
390 * If this transport is not already on the DTO transport queue,
393 spin_lock_irqsave(&dto_lock
, flags
);
394 if (list_empty(&xprt
->sc_dto_q
)) {
395 svc_xprt_get(&xprt
->sc_xprt
);
396 list_add_tail(&xprt
->sc_dto_q
, &dto_xprt_q
);
398 spin_unlock_irqrestore(&dto_lock
, flags
);
400 /* Tasklet does all the work to avoid irqsave locks. */
401 tasklet_schedule(&dto_tasklet
);
404 static void create_context_cache(struct svcxprt_rdma
*xprt
,
405 int ctxt_count
, int ctxt_bump
, int ctxt_max
)
407 struct svc_rdma_op_ctxt
*ctxt
;
410 xprt
->sc_ctxt_max
= ctxt_max
;
411 xprt
->sc_ctxt_bump
= ctxt_bump
;
412 xprt
->sc_ctxt_cnt
= 0;
413 xprt
->sc_ctxt_head
= NULL
;
414 for (i
= 0; i
< ctxt_count
; i
++) {
415 ctxt
= kmalloc(sizeof(*ctxt
), GFP_KERNEL
);
417 ctxt
->next
= xprt
->sc_ctxt_head
;
418 xprt
->sc_ctxt_head
= ctxt
;
424 static void destroy_context_cache(struct svc_rdma_op_ctxt
*ctxt
)
426 struct svc_rdma_op_ctxt
*next
;
437 static struct svcxprt_rdma
*rdma_create_xprt(struct svc_serv
*serv
,
440 struct svcxprt_rdma
*cma_xprt
= kzalloc(sizeof *cma_xprt
, GFP_KERNEL
);
444 svc_xprt_init(&svc_rdma_class
, &cma_xprt
->sc_xprt
, serv
);
445 INIT_LIST_HEAD(&cma_xprt
->sc_accept_q
);
446 INIT_LIST_HEAD(&cma_xprt
->sc_dto_q
);
447 INIT_LIST_HEAD(&cma_xprt
->sc_rq_dto_q
);
448 INIT_LIST_HEAD(&cma_xprt
->sc_read_complete_q
);
449 init_waitqueue_head(&cma_xprt
->sc_send_wait
);
451 spin_lock_init(&cma_xprt
->sc_lock
);
452 spin_lock_init(&cma_xprt
->sc_read_complete_lock
);
453 spin_lock_init(&cma_xprt
->sc_ctxt_lock
);
454 spin_lock_init(&cma_xprt
->sc_rq_dto_lock
);
456 cma_xprt
->sc_ord
= svcrdma_ord
;
458 cma_xprt
->sc_max_req_size
= svcrdma_max_req_size
;
459 cma_xprt
->sc_max_requests
= svcrdma_max_requests
;
460 cma_xprt
->sc_sq_depth
= svcrdma_max_requests
* RPCRDMA_SQ_DEPTH_MULT
;
461 atomic_set(&cma_xprt
->sc_sq_count
, 0);
464 int reqs
= cma_xprt
->sc_max_requests
;
465 create_context_cache(cma_xprt
,
466 reqs
<< 1, /* starting size */
467 reqs
, /* bump amount */
469 cma_xprt
->sc_sq_depth
+
470 RPCRDMA_MAX_THREADS
+ 1); /* max */
471 if (!cma_xprt
->sc_ctxt_head
) {
475 clear_bit(XPT_LISTENER
, &cma_xprt
->sc_xprt
.xpt_flags
);
477 set_bit(XPT_LISTENER
, &cma_xprt
->sc_xprt
.xpt_flags
);
482 struct page
*svc_rdma_get_page(void)
486 while ((page
= alloc_page(GFP_KERNEL
)) == NULL
) {
487 /* If we can't get memory, wait a bit and try again */
488 printk(KERN_INFO
"svcrdma: out of memory...retrying in 1000 "
490 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
495 int svc_rdma_post_recv(struct svcxprt_rdma
*xprt
)
497 struct ib_recv_wr recv_wr
, *bad_recv_wr
;
498 struct svc_rdma_op_ctxt
*ctxt
;
505 ctxt
= svc_rdma_get_context(xprt
);
507 ctxt
->direction
= DMA_FROM_DEVICE
;
508 for (sge_no
= 0; buflen
< xprt
->sc_max_req_size
; sge_no
++) {
509 BUG_ON(sge_no
>= xprt
->sc_max_sge
);
510 page
= svc_rdma_get_page();
511 ctxt
->pages
[sge_no
] = page
;
512 pa
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
515 ctxt
->sge
[sge_no
].addr
= pa
;
516 ctxt
->sge
[sge_no
].length
= PAGE_SIZE
;
517 ctxt
->sge
[sge_no
].lkey
= xprt
->sc_phys_mr
->lkey
;
520 ctxt
->count
= sge_no
;
522 recv_wr
.sg_list
= &ctxt
->sge
[0];
523 recv_wr
.num_sge
= ctxt
->count
;
524 recv_wr
.wr_id
= (u64
)(unsigned long)ctxt
;
526 ret
= ib_post_recv(xprt
->sc_qp
, &recv_wr
, &bad_recv_wr
);
528 svc_rdma_put_context(ctxt
, 1);
533 * This function handles the CONNECT_REQUEST event on a listening
534 * endpoint. It is passed the cma_id for the _new_ connection. The context in
535 * this cma_id is inherited from the listening cma_id and is the svc_xprt
536 * structure for the listening endpoint.
538 * This function creates a new xprt for the new connection and enqueues it on
539 * the accept queue for the listent xprt. When the listen thread is kicked, it
540 * will call the recvfrom method on the listen xprt which will accept the new
543 static void handle_connect_req(struct rdma_cm_id
*new_cma_id
)
545 struct svcxprt_rdma
*listen_xprt
= new_cma_id
->context
;
546 struct svcxprt_rdma
*newxprt
;
548 /* Create a new transport */
549 newxprt
= rdma_create_xprt(listen_xprt
->sc_xprt
.xpt_server
, 0);
551 dprintk("svcrdma: failed to create new transport\n");
554 newxprt
->sc_cm_id
= new_cma_id
;
555 new_cma_id
->context
= newxprt
;
556 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
557 newxprt
, newxprt
->sc_cm_id
, listen_xprt
);
560 * Enqueue the new transport on the accept queue of the listening
563 spin_lock_bh(&listen_xprt
->sc_lock
);
564 list_add_tail(&newxprt
->sc_accept_q
, &listen_xprt
->sc_accept_q
);
565 spin_unlock_bh(&listen_xprt
->sc_lock
);
568 * Can't use svc_xprt_received here because we are not on a
571 set_bit(XPT_CONN
, &listen_xprt
->sc_xprt
.xpt_flags
);
572 svc_xprt_enqueue(&listen_xprt
->sc_xprt
);
576 * Handles events generated on the listening endpoint. These events will be
577 * either be incoming connect requests or adapter removal events.
579 static int rdma_listen_handler(struct rdma_cm_id
*cma_id
,
580 struct rdma_cm_event
*event
)
582 struct svcxprt_rdma
*xprt
= cma_id
->context
;
585 switch (event
->event
) {
586 case RDMA_CM_EVENT_CONNECT_REQUEST
:
587 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
588 "event=%d\n", cma_id
, cma_id
->context
, event
->event
);
589 handle_connect_req(cma_id
);
592 case RDMA_CM_EVENT_ESTABLISHED
:
593 /* Accept complete */
594 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
595 "cm_id=%p\n", xprt
, cma_id
);
598 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
599 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
602 set_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
);
606 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
607 "event=%d\n", cma_id
, event
->event
);
614 static int rdma_cma_handler(struct rdma_cm_id
*cma_id
,
615 struct rdma_cm_event
*event
)
617 struct svc_xprt
*xprt
= cma_id
->context
;
618 struct svcxprt_rdma
*rdma
=
619 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
620 switch (event
->event
) {
621 case RDMA_CM_EVENT_ESTABLISHED
:
622 /* Accept complete */
624 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
625 "cm_id=%p\n", xprt
, cma_id
);
626 clear_bit(RDMAXPRT_CONN_PENDING
, &rdma
->sc_flags
);
627 svc_xprt_enqueue(xprt
);
629 case RDMA_CM_EVENT_DISCONNECTED
:
630 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
633 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
634 svc_xprt_enqueue(xprt
);
638 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
639 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
640 "event=%d\n", cma_id
, xprt
, event
->event
);
642 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
643 svc_xprt_enqueue(xprt
);
647 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
648 "event=%d\n", cma_id
, event
->event
);
655 * Create a listening RDMA service endpoint.
657 static struct svc_xprt
*svc_rdma_create(struct svc_serv
*serv
,
658 struct sockaddr
*sa
, int salen
,
661 struct rdma_cm_id
*listen_id
;
662 struct svcxprt_rdma
*cma_xprt
;
663 struct svc_xprt
*xprt
;
666 dprintk("svcrdma: Creating RDMA socket\n");
668 cma_xprt
= rdma_create_xprt(serv
, 1);
670 return ERR_PTR(-ENOMEM
);
671 xprt
= &cma_xprt
->sc_xprt
;
673 listen_id
= rdma_create_id(rdma_listen_handler
, cma_xprt
, RDMA_PS_TCP
);
674 if (IS_ERR(listen_id
)) {
675 ret
= PTR_ERR(listen_id
);
676 dprintk("svcrdma: rdma_create_id failed = %d\n", ret
);
680 ret
= rdma_bind_addr(listen_id
, sa
);
682 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret
);
685 cma_xprt
->sc_cm_id
= listen_id
;
687 ret
= rdma_listen(listen_id
, RPCRDMA_LISTEN_BACKLOG
);
689 dprintk("svcrdma: rdma_listen failed = %d\n", ret
);
694 * We need to use the address from the cm_id in case the
695 * caller specified 0 for the port number.
697 sa
= (struct sockaddr
*)&cma_xprt
->sc_cm_id
->route
.addr
.src_addr
;
698 svc_xprt_set_local(&cma_xprt
->sc_xprt
, sa
, salen
);
700 return &cma_xprt
->sc_xprt
;
703 rdma_destroy_id(listen_id
);
710 * This is the xpo_recvfrom function for listening endpoints. Its
711 * purpose is to accept incoming connections. The CMA callback handler
712 * has already created a new transport and attached it to the new CMA
715 * There is a queue of pending connections hung on the listening
716 * transport. This queue contains the new svc_xprt structure. This
717 * function takes svc_xprt structures off the accept_q and completes
720 static struct svc_xprt
*svc_rdma_accept(struct svc_xprt
*xprt
)
722 struct svcxprt_rdma
*listen_rdma
;
723 struct svcxprt_rdma
*newxprt
= NULL
;
724 struct rdma_conn_param conn_param
;
725 struct ib_qp_init_attr qp_attr
;
726 struct ib_device_attr devattr
;
731 listen_rdma
= container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
732 clear_bit(XPT_CONN
, &xprt
->xpt_flags
);
733 /* Get the next entry off the accept list */
734 spin_lock_bh(&listen_rdma
->sc_lock
);
735 if (!list_empty(&listen_rdma
->sc_accept_q
)) {
736 newxprt
= list_entry(listen_rdma
->sc_accept_q
.next
,
737 struct svcxprt_rdma
, sc_accept_q
);
738 list_del_init(&newxprt
->sc_accept_q
);
740 if (!list_empty(&listen_rdma
->sc_accept_q
))
741 set_bit(XPT_CONN
, &listen_rdma
->sc_xprt
.xpt_flags
);
742 spin_unlock_bh(&listen_rdma
->sc_lock
);
746 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
747 newxprt
, newxprt
->sc_cm_id
);
749 ret
= ib_query_device(newxprt
->sc_cm_id
->device
, &devattr
);
751 dprintk("svcrdma: could not query device attributes on "
752 "device %p, rc=%d\n", newxprt
->sc_cm_id
->device
, ret
);
756 /* Qualify the transport resource defaults with the
757 * capabilities of this particular device */
758 newxprt
->sc_max_sge
= min((size_t)devattr
.max_sge
,
759 (size_t)RPCSVC_MAXPAGES
);
760 newxprt
->sc_max_requests
= min((size_t)devattr
.max_qp_wr
,
761 (size_t)svcrdma_max_requests
);
762 newxprt
->sc_sq_depth
= RPCRDMA_SQ_DEPTH_MULT
* newxprt
->sc_max_requests
;
764 newxprt
->sc_ord
= min((size_t)devattr
.max_qp_rd_atom
,
765 (size_t)svcrdma_ord
);
767 newxprt
->sc_pd
= ib_alloc_pd(newxprt
->sc_cm_id
->device
);
768 if (IS_ERR(newxprt
->sc_pd
)) {
769 dprintk("svcrdma: error creating PD for connect request\n");
772 newxprt
->sc_sq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
776 newxprt
->sc_sq_depth
,
778 if (IS_ERR(newxprt
->sc_sq_cq
)) {
779 dprintk("svcrdma: error creating SQ CQ for connect request\n");
782 newxprt
->sc_rq_cq
= ib_create_cq(newxprt
->sc_cm_id
->device
,
786 newxprt
->sc_max_requests
,
788 if (IS_ERR(newxprt
->sc_rq_cq
)) {
789 dprintk("svcrdma: error creating RQ CQ for connect request\n");
793 memset(&qp_attr
, 0, sizeof qp_attr
);
794 qp_attr
.event_handler
= qp_event_handler
;
795 qp_attr
.qp_context
= &newxprt
->sc_xprt
;
796 qp_attr
.cap
.max_send_wr
= newxprt
->sc_sq_depth
;
797 qp_attr
.cap
.max_recv_wr
= newxprt
->sc_max_requests
;
798 qp_attr
.cap
.max_send_sge
= newxprt
->sc_max_sge
;
799 qp_attr
.cap
.max_recv_sge
= newxprt
->sc_max_sge
;
800 qp_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
801 qp_attr
.qp_type
= IB_QPT_RC
;
802 qp_attr
.send_cq
= newxprt
->sc_sq_cq
;
803 qp_attr
.recv_cq
= newxprt
->sc_rq_cq
;
804 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
805 " cm_id->device=%p, sc_pd->device=%p\n"
806 " cap.max_send_wr = %d\n"
807 " cap.max_recv_wr = %d\n"
808 " cap.max_send_sge = %d\n"
809 " cap.max_recv_sge = %d\n",
810 newxprt
->sc_cm_id
, newxprt
->sc_pd
,
811 newxprt
->sc_cm_id
->device
, newxprt
->sc_pd
->device
,
812 qp_attr
.cap
.max_send_wr
,
813 qp_attr
.cap
.max_recv_wr
,
814 qp_attr
.cap
.max_send_sge
,
815 qp_attr
.cap
.max_recv_sge
);
817 ret
= rdma_create_qp(newxprt
->sc_cm_id
, newxprt
->sc_pd
, &qp_attr
);
820 * XXX: This is a hack. We need a xx_request_qp interface
821 * that will adjust the qp_attr's with a best-effort
824 qp_attr
.cap
.max_send_sge
-= 2;
825 qp_attr
.cap
.max_recv_sge
-= 2;
826 ret
= rdma_create_qp(newxprt
->sc_cm_id
, newxprt
->sc_pd
,
829 dprintk("svcrdma: failed to create QP, ret=%d\n", ret
);
832 newxprt
->sc_max_sge
= qp_attr
.cap
.max_send_sge
;
833 newxprt
->sc_max_sge
= qp_attr
.cap
.max_recv_sge
;
834 newxprt
->sc_sq_depth
= qp_attr
.cap
.max_send_wr
;
835 newxprt
->sc_max_requests
= qp_attr
.cap
.max_recv_wr
;
837 svc_xprt_get(&newxprt
->sc_xprt
);
838 newxprt
->sc_qp
= newxprt
->sc_cm_id
->qp
;
840 /* Register all of physical memory */
841 newxprt
->sc_phys_mr
= ib_get_dma_mr(newxprt
->sc_pd
,
842 IB_ACCESS_LOCAL_WRITE
|
843 IB_ACCESS_REMOTE_WRITE
);
844 if (IS_ERR(newxprt
->sc_phys_mr
)) {
845 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret
);
849 /* Post receive buffers */
850 for (i
= 0; i
< newxprt
->sc_max_requests
; i
++) {
851 ret
= svc_rdma_post_recv(newxprt
);
853 dprintk("svcrdma: failure posting receive buffers\n");
858 /* Swap out the handler */
859 newxprt
->sc_cm_id
->event_handler
= rdma_cma_handler
;
861 /* Accept Connection */
862 set_bit(RDMAXPRT_CONN_PENDING
, &newxprt
->sc_flags
);
863 memset(&conn_param
, 0, sizeof conn_param
);
864 conn_param
.responder_resources
= 0;
865 conn_param
.initiator_depth
= newxprt
->sc_ord
;
866 ret
= rdma_accept(newxprt
->sc_cm_id
, &conn_param
);
868 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
873 dprintk("svcrdma: new connection %p accepted with the following "
875 " local_ip : %d.%d.%d.%d\n"
877 " remote_ip : %d.%d.%d.%d\n"
878 " remote_port : %d\n"
881 " max_requests : %d\n"
884 NIPQUAD(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
885 route
.addr
.src_addr
)->sin_addr
.s_addr
),
886 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
887 route
.addr
.src_addr
)->sin_port
),
888 NIPQUAD(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
889 route
.addr
.dst_addr
)->sin_addr
.s_addr
),
890 ntohs(((struct sockaddr_in
*)&newxprt
->sc_cm_id
->
891 route
.addr
.dst_addr
)->sin_port
),
893 newxprt
->sc_sq_depth
,
894 newxprt
->sc_max_requests
,
897 /* Set the local and remote addresses in the transport */
898 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.dst_addr
;
899 svc_xprt_set_remote(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
900 sa
= (struct sockaddr
*)&newxprt
->sc_cm_id
->route
.addr
.src_addr
;
901 svc_xprt_set_local(&newxprt
->sc_xprt
, sa
, svc_addr_len(sa
));
903 ib_req_notify_cq(newxprt
->sc_sq_cq
, IB_CQ_NEXT_COMP
);
904 ib_req_notify_cq(newxprt
->sc_rq_cq
, IB_CQ_NEXT_COMP
);
905 return &newxprt
->sc_xprt
;
908 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret
);
909 /* Take a reference in case the DTO handler runs */
910 svc_xprt_get(&newxprt
->sc_xprt
);
911 if (newxprt
->sc_qp
&& !IS_ERR(newxprt
->sc_qp
)) {
912 ib_destroy_qp(newxprt
->sc_qp
);
913 svc_xprt_put(&newxprt
->sc_xprt
);
915 rdma_destroy_id(newxprt
->sc_cm_id
);
916 /* This call to put will destroy the transport */
917 svc_xprt_put(&newxprt
->sc_xprt
);
921 static void svc_rdma_release_rqst(struct svc_rqst
*rqstp
)
926 * When connected, an svc_xprt has at least three references:
928 * - A reference held by the QP. We still hold that here because this
929 * code deletes the QP and puts the reference.
931 * - A reference held by the cm_id between the ESTABLISHED and
932 * DISCONNECTED events. If the remote peer disconnected first, this
933 * reference could be gone.
935 * - A reference held by the svc_recv code that called this function
936 * as part of close processing.
938 * At a minimum two references should still be held.
940 static void svc_rdma_detach(struct svc_xprt
*xprt
)
942 struct svcxprt_rdma
*rdma
=
943 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
944 dprintk("svc: svc_rdma_detach(%p)\n", xprt
);
946 /* Disconnect and flush posted WQE */
947 rdma_disconnect(rdma
->sc_cm_id
);
949 /* Destroy the QP if present (not a listener) */
950 if (rdma
->sc_qp
&& !IS_ERR(rdma
->sc_qp
)) {
951 ib_destroy_qp(rdma
->sc_qp
);
955 /* Destroy the CM ID */
956 rdma_destroy_id(rdma
->sc_cm_id
);
959 static void svc_rdma_free(struct svc_xprt
*xprt
)
961 struct svcxprt_rdma
*rdma
= (struct svcxprt_rdma
*)xprt
;
962 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma
);
963 /* We should only be called from kref_put */
964 BUG_ON(atomic_read(&xprt
->xpt_ref
.refcount
) != 0);
965 if (rdma
->sc_sq_cq
&& !IS_ERR(rdma
->sc_sq_cq
))
966 ib_destroy_cq(rdma
->sc_sq_cq
);
968 if (rdma
->sc_rq_cq
&& !IS_ERR(rdma
->sc_rq_cq
))
969 ib_destroy_cq(rdma
->sc_rq_cq
);
971 if (rdma
->sc_phys_mr
&& !IS_ERR(rdma
->sc_phys_mr
))
972 ib_dereg_mr(rdma
->sc_phys_mr
);
974 if (rdma
->sc_pd
&& !IS_ERR(rdma
->sc_pd
))
975 ib_dealloc_pd(rdma
->sc_pd
);
977 destroy_context_cache(rdma
->sc_ctxt_head
);
981 static int svc_rdma_has_wspace(struct svc_xprt
*xprt
)
983 struct svcxprt_rdma
*rdma
=
984 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
987 * If there are fewer SQ WR available than required to send a
988 * simple response, return false.
990 if ((rdma
->sc_sq_depth
- atomic_read(&rdma
->sc_sq_count
) < 3))
994 * ...or there are already waiters on the SQ,
997 if (waitqueue_active(&rdma
->sc_send_wait
))
1000 /* Otherwise return true. */
1004 int svc_rdma_send(struct svcxprt_rdma
*xprt
, struct ib_send_wr
*wr
)
1006 struct ib_send_wr
*bad_wr
;
1009 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1012 BUG_ON(wr
->send_flags
!= IB_SEND_SIGNALED
);
1013 BUG_ON(((struct svc_rdma_op_ctxt
*)(unsigned long)wr
->wr_id
)->wr_op
!=
1015 /* If the SQ is full, wait until an SQ entry is available */
1017 spin_lock_bh(&xprt
->sc_lock
);
1018 if (xprt
->sc_sq_depth
== atomic_read(&xprt
->sc_sq_count
)) {
1019 spin_unlock_bh(&xprt
->sc_lock
);
1020 atomic_inc(&rdma_stat_sq_starve
);
1022 /* See if we can opportunistically reap SQ WR to make room */
1025 /* Wait until SQ WR available if SQ still full */
1026 wait_event(xprt
->sc_send_wait
,
1027 atomic_read(&xprt
->sc_sq_count
) <
1029 if (test_bit(XPT_CLOSE
, &xprt
->sc_xprt
.xpt_flags
))
1033 /* Bumped used SQ WR count and post */
1034 ret
= ib_post_send(xprt
->sc_qp
, wr
, &bad_wr
);
1036 atomic_inc(&xprt
->sc_sq_count
);
1038 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1039 "sc_sq_count=%d, sc_sq_depth=%d\n",
1040 ret
, atomic_read(&xprt
->sc_sq_count
),
1042 spin_unlock_bh(&xprt
->sc_lock
);
1048 int svc_rdma_send_error(struct svcxprt_rdma
*xprt
, struct rpcrdma_msg
*rmsgp
,
1049 enum rpcrdma_errcode err
)
1051 struct ib_send_wr err_wr
;
1054 struct svc_rdma_op_ctxt
*ctxt
;
1059 p
= svc_rdma_get_page();
1060 va
= page_address(p
);
1062 /* XDR encode error */
1063 length
= svc_rdma_xdr_encode_error(xprt
, rmsgp
, err
, va
);
1065 /* Prepare SGE for local address */
1066 sge
.addr
= ib_dma_map_page(xprt
->sc_cm_id
->device
,
1067 p
, 0, PAGE_SIZE
, DMA_FROM_DEVICE
);
1068 sge
.lkey
= xprt
->sc_phys_mr
->lkey
;
1069 sge
.length
= length
;
1071 ctxt
= svc_rdma_get_context(xprt
);
1075 /* Prepare SEND WR */
1076 memset(&err_wr
, 0, sizeof err_wr
);
1077 ctxt
->wr_op
= IB_WR_SEND
;
1078 err_wr
.wr_id
= (unsigned long)ctxt
;
1079 err_wr
.sg_list
= &sge
;
1081 err_wr
.opcode
= IB_WR_SEND
;
1082 err_wr
.send_flags
= IB_SEND_SIGNALED
;
1085 ret
= svc_rdma_send(xprt
, &err_wr
);
1087 dprintk("svcrdma: Error posting send = %d\n", ret
);
1088 svc_rdma_put_context(ctxt
, 1);