2 * Copyright (c) 2005-2006 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/debug.h>
43 #include <linux/sunrpc/rpc_rdma.h>
44 #include <linux/spinlock.h>
45 #include <asm/unaligned.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 /* Encode an XDR as an array of IB SGE
55 * - head[0] is physically contiguous.
56 * - tail[0] is physically contiguous.
57 * - pages[] is not physically or virtually contiguous and consists of
61 * SGE[0] reserved for RCPRDMA header
62 * SGE[1] data from xdr->head[]
63 * SGE[2..sge_count-2] data from xdr->pages[]
64 * SGE[sge_count-1] data from xdr->tail.
66 * The max SGE we need is the length of the XDR / pagesize + one for
67 * head + one for tail + one for RPCRDMA header. Since RPCSVC_MAXPAGES
68 * reserves a page for both the request and the reply header, and this
69 * array is only concerned with the reply we are assured that we have
70 * on extra page for the RPCRMDA header.
72 static int fast_reg_xdr(struct svcxprt_rdma
*xprt
,
74 struct svc_rdma_req_map
*vec
)
82 struct svc_rdma_fastreg_mr
*frmr
;
84 frmr
= svc_rdma_get_frmr(xprt
);
89 /* Skip the RPCRDMA header */
93 frva
= (void *)((unsigned long)(xdr
->head
[0].iov_base
) & PAGE_MASK
);
94 vec
->sge
[sge_no
].iov_base
= xdr
->head
[0].iov_base
;
95 vec
->sge
[sge_no
].iov_len
= xdr
->head
[0].iov_len
;
101 frmr
->direction
= DMA_TO_DEVICE
;
102 frmr
->access_flags
= 0;
103 frmr
->map_len
= PAGE_SIZE
;
104 frmr
->page_list_len
= 1;
105 frmr
->page_list
->page_list
[page_no
] =
106 ib_dma_map_single(xprt
->sc_cm_id
->device
,
107 (void *)xdr
->head
[0].iov_base
,
108 PAGE_SIZE
, DMA_TO_DEVICE
);
109 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
110 frmr
->page_list
->page_list
[page_no
]))
112 atomic_inc(&xprt
->sc_dma_used
);
114 page_off
= xdr
->page_base
;
115 page_bytes
= xdr
->page_len
+ page_off
;
120 vec
->sge
[sge_no
].iov_base
= frva
+ frmr
->map_len
+ page_off
;
121 vec
->sge
[sge_no
].iov_len
= page_bytes
;
126 page
= xdr
->pages
[page_no
++];
127 sge_bytes
= min_t(u32
, page_bytes
, (PAGE_SIZE
- page_off
));
128 page_bytes
-= sge_bytes
;
130 frmr
->page_list
->page_list
[page_no
] =
131 ib_dma_map_single(xprt
->sc_cm_id
->device
,
133 PAGE_SIZE
, DMA_TO_DEVICE
);
134 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
135 frmr
->page_list
->page_list
[page_no
]))
138 atomic_inc(&xprt
->sc_dma_used
);
139 page_off
= 0; /* reset for next time through loop */
140 frmr
->map_len
+= PAGE_SIZE
;
141 frmr
->page_list_len
++;
147 if (0 == xdr
->tail
[0].iov_len
)
151 vec
->sge
[sge_no
].iov_len
= xdr
->tail
[0].iov_len
;
153 if (((unsigned long)xdr
->tail
[0].iov_base
& PAGE_MASK
) ==
154 ((unsigned long)xdr
->head
[0].iov_base
& PAGE_MASK
)) {
156 * If head and tail use the same page, we don't need
159 vec
->sge
[sge_no
].iov_base
= xdr
->tail
[0].iov_base
;
163 /* Map another page for the tail */
164 page_off
= (unsigned long)xdr
->tail
[0].iov_base
& ~PAGE_MASK
;
165 va
= (void *)((unsigned long)xdr
->tail
[0].iov_base
& PAGE_MASK
);
166 vec
->sge
[sge_no
].iov_base
= frva
+ frmr
->map_len
+ page_off
;
168 frmr
->page_list
->page_list
[page_no
] =
169 ib_dma_map_single(xprt
->sc_cm_id
->device
, va
, PAGE_SIZE
,
171 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
172 frmr
->page_list
->page_list
[page_no
]))
174 atomic_inc(&xprt
->sc_dma_used
);
175 frmr
->map_len
+= PAGE_SIZE
;
176 frmr
->page_list_len
++;
180 if (svc_rdma_fastreg(xprt
, frmr
))
186 printk("svcrdma: Error fast registering memory for xprt %p\n", xprt
);
188 svc_rdma_put_frmr(xprt
, frmr
);
192 static int map_xdr(struct svcxprt_rdma
*xprt
,
194 struct svc_rdma_req_map
*vec
)
203 (xdr
->head
[0].iov_len
+ xdr
->page_len
+ xdr
->tail
[0].iov_len
));
205 if (xprt
->sc_frmr_pg_list_len
)
206 return fast_reg_xdr(xprt
, xdr
, vec
);
208 /* Skip the first sge, this is for the RPCRDMA header */
212 vec
->sge
[sge_no
].iov_base
= xdr
->head
[0].iov_base
;
213 vec
->sge
[sge_no
].iov_len
= xdr
->head
[0].iov_len
;
218 page_bytes
= xdr
->page_len
;
219 page_off
= xdr
->page_base
;
221 vec
->sge
[sge_no
].iov_base
=
222 page_address(xdr
->pages
[page_no
]) + page_off
;
223 sge_bytes
= min_t(u32
, page_bytes
, (PAGE_SIZE
- page_off
));
224 page_bytes
-= sge_bytes
;
225 vec
->sge
[sge_no
].iov_len
= sge_bytes
;
229 page_off
= 0; /* reset for next time through loop */
233 if (xdr
->tail
[0].iov_len
) {
234 vec
->sge
[sge_no
].iov_base
= xdr
->tail
[0].iov_base
;
235 vec
->sge
[sge_no
].iov_len
= xdr
->tail
[0].iov_len
;
239 dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
240 "page_base %u page_len %u head_len %zu tail_len %zu\n",
241 sge_no
, page_no
, xdr
->page_base
, xdr
->page_len
,
242 xdr
->head
[0].iov_len
, xdr
->tail
[0].iov_len
);
249 * - We are using FRMR
251 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
253 static int send_write(struct svcxprt_rdma
*xprt
, struct svc_rqst
*rqstp
,
255 u32 xdr_off
, int write_len
,
256 struct svc_rdma_req_map
*vec
)
258 struct ib_send_wr write_wr
;
265 struct svc_rdma_op_ctxt
*ctxt
;
267 BUG_ON(vec
->count
> RPCSVC_MAXPAGES
);
268 dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
269 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
270 rmr
, (unsigned long long)to
, xdr_off
,
271 write_len
, vec
->sge
, vec
->count
);
273 ctxt
= svc_rdma_get_context(xprt
);
274 ctxt
->direction
= DMA_TO_DEVICE
;
277 /* Find the SGE associated with xdr_off */
278 for (bc
= xdr_off
, xdr_sge_no
= 1; bc
&& xdr_sge_no
< vec
->count
;
280 if (vec
->sge
[xdr_sge_no
].iov_len
> bc
)
282 bc
-= vec
->sge
[xdr_sge_no
].iov_len
;
289 /* Copy the remaining SGE */
291 sge_bytes
= min_t(size_t,
292 bc
, vec
->sge
[xdr_sge_no
].iov_len
-sge_off
);
293 sge
[sge_no
].length
= sge_bytes
;
296 ib_dma_map_single(xprt
->sc_cm_id
->device
,
298 vec
->sge
[xdr_sge_no
].iov_base
+ sge_off
,
299 sge_bytes
, DMA_TO_DEVICE
);
300 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
303 atomic_inc(&xprt
->sc_dma_used
);
304 sge
[sge_no
].lkey
= xprt
->sc_dma_lkey
;
306 sge
[sge_no
].addr
= (unsigned long)
307 vec
->sge
[xdr_sge_no
].iov_base
+ sge_off
;
308 sge
[sge_no
].lkey
= vec
->frmr
->mr
->lkey
;
311 ctxt
->frmr
= vec
->frmr
;
315 BUG_ON(xdr_sge_no
> vec
->count
);
319 /* Prepare WRITE WR */
320 memset(&write_wr
, 0, sizeof write_wr
);
321 ctxt
->wr_op
= IB_WR_RDMA_WRITE
;
322 write_wr
.wr_id
= (unsigned long)ctxt
;
323 write_wr
.sg_list
= &sge
[0];
324 write_wr
.num_sge
= sge_no
;
325 write_wr
.opcode
= IB_WR_RDMA_WRITE
;
326 write_wr
.send_flags
= IB_SEND_SIGNALED
;
327 write_wr
.wr
.rdma
.rkey
= rmr
;
328 write_wr
.wr
.rdma
.remote_addr
= to
;
331 atomic_inc(&rdma_stat_write
);
332 if (svc_rdma_send(xprt
, &write_wr
))
336 svc_rdma_put_context(ctxt
, 0);
337 /* Fatal error, close transport */
341 static int send_write_chunks(struct svcxprt_rdma
*xprt
,
342 struct rpcrdma_msg
*rdma_argp
,
343 struct rpcrdma_msg
*rdma_resp
,
344 struct svc_rqst
*rqstp
,
345 struct svc_rdma_req_map
*vec
)
347 u32 xfer_len
= rqstp
->rq_res
.page_len
+ rqstp
->rq_res
.tail
[0].iov_len
;
353 struct rpcrdma_write_array
*arg_ary
;
354 struct rpcrdma_write_array
*res_ary
;
357 arg_ary
= svc_rdma_get_write_array(rdma_argp
);
360 res_ary
= (struct rpcrdma_write_array
*)
361 &rdma_resp
->rm_body
.rm_chunks
[1];
364 max_write
= vec
->frmr
->map_len
;
366 max_write
= xprt
->sc_max_sge
* PAGE_SIZE
;
368 /* Write chunks start at the pagelist */
369 for (xdr_off
= rqstp
->rq_res
.head
[0].iov_len
, chunk_no
= 0;
370 xfer_len
&& chunk_no
< arg_ary
->wc_nchunks
;
372 struct rpcrdma_segment
*arg_ch
;
375 arg_ch
= &arg_ary
->wc_array
[chunk_no
].wc_target
;
376 write_len
= min(xfer_len
, arg_ch
->rs_length
);
378 /* Prepare the response chunk given the length actually
380 rs_offset
= get_unaligned(&(arg_ch
->rs_offset
));
381 svc_rdma_xdr_encode_array_chunk(res_ary
, chunk_no
,
388 this_write
= min(write_len
, max_write
);
389 ret
= send_write(xprt
, rqstp
,
391 rs_offset
+ chunk_off
,
396 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
400 chunk_off
+= this_write
;
401 xdr_off
+= this_write
;
402 xfer_len
-= this_write
;
403 write_len
-= this_write
;
406 /* Update the req with the number of chunks actually used */
407 svc_rdma_xdr_encode_write_list(rdma_resp
, chunk_no
);
409 return rqstp
->rq_res
.page_len
+ rqstp
->rq_res
.tail
[0].iov_len
;
412 static int send_reply_chunks(struct svcxprt_rdma
*xprt
,
413 struct rpcrdma_msg
*rdma_argp
,
414 struct rpcrdma_msg
*rdma_resp
,
415 struct svc_rqst
*rqstp
,
416 struct svc_rdma_req_map
*vec
)
418 u32 xfer_len
= rqstp
->rq_res
.len
;
424 struct rpcrdma_segment
*ch
;
425 struct rpcrdma_write_array
*arg_ary
;
426 struct rpcrdma_write_array
*res_ary
;
429 arg_ary
= svc_rdma_get_reply_array(rdma_argp
);
432 res_ary
= (struct rpcrdma_write_array
*)
433 &rdma_resp
->rm_body
.rm_chunks
[2];
436 max_write
= vec
->frmr
->map_len
;
438 max_write
= xprt
->sc_max_sge
* PAGE_SIZE
;
440 /* xdr offset starts at RPC message */
441 for (xdr_off
= 0, chunk_no
= 0;
442 xfer_len
&& chunk_no
< arg_ary
->wc_nchunks
;
445 ch
= &arg_ary
->wc_array
[chunk_no
].wc_target
;
446 write_len
= min(xfer_len
, ch
->rs_length
);
448 /* Prepare the reply chunk given the length actually
450 rs_offset
= get_unaligned(&(ch
->rs_offset
));
451 svc_rdma_xdr_encode_array_chunk(res_ary
, chunk_no
,
452 ch
->rs_handle
, rs_offset
,
458 this_write
= min(write_len
, max_write
);
459 ret
= send_write(xprt
, rqstp
,
461 rs_offset
+ chunk_off
,
466 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
470 chunk_off
+= this_write
;
471 xdr_off
+= this_write
;
472 xfer_len
-= this_write
;
473 write_len
-= this_write
;
476 /* Update the req with the number of chunks actually used */
477 svc_rdma_xdr_encode_reply_array(res_ary
, chunk_no
);
479 return rqstp
->rq_res
.len
;
482 /* This function prepares the portion of the RPCRDMA message to be
483 * sent in the RDMA_SEND. This function is called after data sent via
484 * RDMA has already been transmitted. There are three cases:
485 * - The RPCRDMA header, RPC header, and payload are all sent in a
486 * single RDMA_SEND. This is the "inline" case.
487 * - The RPCRDMA header and some portion of the RPC header and data
488 * are sent via this RDMA_SEND and another portion of the data is
490 * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
491 * header and data are all transmitted via RDMA.
492 * In all three cases, this function prepares the RPCRDMA header in
493 * sge[0], the 'type' parameter indicates the type to place in the
494 * RPCRDMA header, and the 'byte_count' field indicates how much of
495 * the XDR to include in this RDMA_SEND.
497 static int send_reply(struct svcxprt_rdma
*rdma
,
498 struct svc_rqst
*rqstp
,
500 struct rpcrdma_msg
*rdma_resp
,
501 struct svc_rdma_op_ctxt
*ctxt
,
502 struct svc_rdma_req_map
*vec
,
505 struct ib_send_wr send_wr
;
506 struct ib_send_wr inv_wr
;
512 /* Post a recv buffer to handle another request. */
513 ret
= svc_rdma_post_recv(rdma
);
516 "svcrdma: could not post a receive buffer, err=%d."
517 "Closing transport %p.\n", ret
, rdma
);
518 set_bit(XPT_CLOSE
, &rdma
->sc_xprt
.xpt_flags
);
519 svc_rdma_put_frmr(rdma
, vec
->frmr
);
520 svc_rdma_put_context(ctxt
, 0);
524 /* Prepare the context */
525 ctxt
->pages
[0] = page
;
527 ctxt
->frmr
= vec
->frmr
;
529 set_bit(RDMACTXT_F_FAST_UNREG
, &ctxt
->flags
);
531 clear_bit(RDMACTXT_F_FAST_UNREG
, &ctxt
->flags
);
533 /* Prepare the SGE for the RPCRDMA Header */
534 ctxt
->sge
[0].lkey
= rdma
->sc_dma_lkey
;
535 ctxt
->sge
[0].length
= svc_rdma_xdr_get_reply_hdr_len(rdma_resp
);
537 ib_dma_map_single(rdma
->sc_cm_id
->device
, page_address(page
),
538 ctxt
->sge
[0].length
, DMA_TO_DEVICE
);
539 if (ib_dma_mapping_error(rdma
->sc_cm_id
->device
, ctxt
->sge
[0].addr
))
541 atomic_inc(&rdma
->sc_dma_used
);
543 ctxt
->direction
= DMA_TO_DEVICE
;
545 /* Determine how many of our SGE are to be transmitted */
546 for (sge_no
= 1; byte_count
&& sge_no
< vec
->count
; sge_no
++) {
547 sge_bytes
= min_t(size_t, vec
->sge
[sge_no
].iov_len
, byte_count
);
548 byte_count
-= sge_bytes
;
550 ctxt
->sge
[sge_no
].addr
=
551 ib_dma_map_single(rdma
->sc_cm_id
->device
,
552 vec
->sge
[sge_no
].iov_base
,
553 sge_bytes
, DMA_TO_DEVICE
);
554 if (ib_dma_mapping_error(rdma
->sc_cm_id
->device
,
555 ctxt
->sge
[sge_no
].addr
))
557 atomic_inc(&rdma
->sc_dma_used
);
558 ctxt
->sge
[sge_no
].lkey
= rdma
->sc_dma_lkey
;
560 ctxt
->sge
[sge_no
].addr
= (unsigned long)
561 vec
->sge
[sge_no
].iov_base
;
562 ctxt
->sge
[sge_no
].lkey
= vec
->frmr
->mr
->lkey
;
564 ctxt
->sge
[sge_no
].length
= sge_bytes
;
566 BUG_ON(byte_count
!= 0);
568 /* Save all respages in the ctxt and remove them from the
569 * respages array. They are our pages until the I/O
572 for (page_no
= 0; page_no
< rqstp
->rq_resused
; page_no
++) {
573 ctxt
->pages
[page_no
+1] = rqstp
->rq_respages
[page_no
];
575 rqstp
->rq_respages
[page_no
] = NULL
;
577 * If there are more pages than SGE, terminate SGE
578 * list so that svc_rdma_unmap_dma doesn't attempt to
581 if (page_no
+1 >= sge_no
)
582 ctxt
->sge
[page_no
+1].length
= 0;
584 BUG_ON(sge_no
> rdma
->sc_max_sge
);
585 memset(&send_wr
, 0, sizeof send_wr
);
586 ctxt
->wr_op
= IB_WR_SEND
;
587 send_wr
.wr_id
= (unsigned long)ctxt
;
588 send_wr
.sg_list
= ctxt
->sge
;
589 send_wr
.num_sge
= sge_no
;
590 send_wr
.opcode
= IB_WR_SEND
;
591 send_wr
.send_flags
= IB_SEND_SIGNALED
;
593 /* Prepare INVALIDATE WR */
594 memset(&inv_wr
, 0, sizeof inv_wr
);
595 inv_wr
.opcode
= IB_WR_LOCAL_INV
;
596 inv_wr
.send_flags
= IB_SEND_SIGNALED
;
597 inv_wr
.ex
.invalidate_rkey
=
599 send_wr
.next
= &inv_wr
;
602 ret
= svc_rdma_send(rdma
, &send_wr
);
609 svc_rdma_unmap_dma(ctxt
);
610 svc_rdma_put_frmr(rdma
, vec
->frmr
);
611 svc_rdma_put_context(ctxt
, 1);
615 void svc_rdma_prep_reply_hdr(struct svc_rqst
*rqstp
)
620 * Return the start of an xdr buffer.
622 static void *xdr_start(struct xdr_buf
*xdr
)
624 return xdr
->head
[0].iov_base
-
627 xdr
->tail
[0].iov_len
-
628 xdr
->head
[0].iov_len
);
631 int svc_rdma_sendto(struct svc_rqst
*rqstp
)
633 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
634 struct svcxprt_rdma
*rdma
=
635 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
636 struct rpcrdma_msg
*rdma_argp
;
637 struct rpcrdma_msg
*rdma_resp
;
638 struct rpcrdma_write_array
*reply_ary
;
639 enum rpcrdma_proc reply_type
;
642 struct page
*res_page
;
643 struct svc_rdma_op_ctxt
*ctxt
;
644 struct svc_rdma_req_map
*vec
;
646 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp
);
648 /* Get the RDMA request header. */
649 rdma_argp
= xdr_start(&rqstp
->rq_arg
);
651 /* Build an req vec for the XDR */
652 ctxt
= svc_rdma_get_context(rdma
);
653 ctxt
->direction
= DMA_TO_DEVICE
;
654 vec
= svc_rdma_get_req_map();
655 ret
= map_xdr(rdma
, &rqstp
->rq_res
, vec
);
658 inline_bytes
= rqstp
->rq_res
.len
;
660 /* Create the RDMA response header */
661 res_page
= svc_rdma_get_page();
662 rdma_resp
= page_address(res_page
);
663 reply_ary
= svc_rdma_get_reply_array(rdma_argp
);
665 reply_type
= RDMA_NOMSG
;
667 reply_type
= RDMA_MSG
;
668 svc_rdma_xdr_encode_reply_header(rdma
, rdma_argp
,
669 rdma_resp
, reply_type
);
671 /* Send any write-chunk data and build resp write-list */
672 ret
= send_write_chunks(rdma
, rdma_argp
, rdma_resp
,
675 printk(KERN_ERR
"svcrdma: failed to send write chunks, rc=%d\n",
681 /* Send any reply-list data and update resp reply-list */
682 ret
= send_reply_chunks(rdma
, rdma_argp
, rdma_resp
,
685 printk(KERN_ERR
"svcrdma: failed to send reply chunks, rc=%d\n",
691 ret
= send_reply(rdma
, rqstp
, res_page
, rdma_resp
, ctxt
, vec
,
693 svc_rdma_put_req_map(vec
);
694 dprintk("svcrdma: send_reply returns %d\n", ret
);
700 svc_rdma_put_req_map(vec
);
701 svc_rdma_put_context(ctxt
, 0);