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 contigous 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_page(xprt
->sc_cm_id
->device
, page
, 0,
132 PAGE_SIZE
, DMA_TO_DEVICE
);
133 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
134 frmr
->page_list
->page_list
[page_no
]))
137 atomic_inc(&xprt
->sc_dma_used
);
138 page_off
= 0; /* reset for next time through loop */
139 frmr
->map_len
+= PAGE_SIZE
;
140 frmr
->page_list_len
++;
146 if (0 == xdr
->tail
[0].iov_len
)
150 vec
->sge
[sge_no
].iov_len
= xdr
->tail
[0].iov_len
;
152 if (((unsigned long)xdr
->tail
[0].iov_base
& PAGE_MASK
) ==
153 ((unsigned long)xdr
->head
[0].iov_base
& PAGE_MASK
)) {
155 * If head and tail use the same page, we don't need
158 vec
->sge
[sge_no
].iov_base
= xdr
->tail
[0].iov_base
;
162 /* Map another page for the tail */
163 page_off
= (unsigned long)xdr
->tail
[0].iov_base
& ~PAGE_MASK
;
164 va
= (void *)((unsigned long)xdr
->tail
[0].iov_base
& PAGE_MASK
);
165 vec
->sge
[sge_no
].iov_base
= frva
+ frmr
->map_len
+ page_off
;
167 frmr
->page_list
->page_list
[page_no
] =
168 ib_dma_map_single(xprt
->sc_cm_id
->device
, va
, PAGE_SIZE
,
170 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
171 frmr
->page_list
->page_list
[page_no
]))
173 atomic_inc(&xprt
->sc_dma_used
);
174 frmr
->map_len
+= PAGE_SIZE
;
175 frmr
->page_list_len
++;
179 if (svc_rdma_fastreg(xprt
, frmr
))
185 printk("svcrdma: Error fast registering memory for xprt %p\n", xprt
);
186 svc_rdma_put_frmr(xprt
, frmr
);
190 static int map_xdr(struct svcxprt_rdma
*xprt
,
192 struct svc_rdma_req_map
*vec
)
194 int sge_max
= (xdr
->len
+PAGE_SIZE
-1) / PAGE_SIZE
+ 3;
202 (xdr
->head
[0].iov_len
+ xdr
->page_len
+ xdr
->tail
[0].iov_len
));
204 if (xprt
->sc_frmr_pg_list_len
)
205 return fast_reg_xdr(xprt
, xdr
, vec
);
207 /* Skip the first sge, this is for the RPCRDMA header */
211 vec
->sge
[sge_no
].iov_base
= xdr
->head
[0].iov_base
;
212 vec
->sge
[sge_no
].iov_len
= xdr
->head
[0].iov_len
;
217 page_bytes
= xdr
->page_len
;
218 page_off
= xdr
->page_base
;
220 vec
->sge
[sge_no
].iov_base
=
221 page_address(xdr
->pages
[page_no
]) + page_off
;
222 sge_bytes
= min_t(u32
, page_bytes
, (PAGE_SIZE
- page_off
));
223 page_bytes
-= sge_bytes
;
224 vec
->sge
[sge_no
].iov_len
= sge_bytes
;
228 page_off
= 0; /* reset for next time through loop */
232 if (xdr
->tail
[0].iov_len
) {
233 vec
->sge
[sge_no
].iov_base
= xdr
->tail
[0].iov_base
;
234 vec
->sge
[sge_no
].iov_len
= xdr
->tail
[0].iov_len
;
238 BUG_ON(sge_no
> sge_max
);
244 * - We are using FRMR
246 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
248 static int send_write(struct svcxprt_rdma
*xprt
, struct svc_rqst
*rqstp
,
250 u32 xdr_off
, int write_len
,
251 struct svc_rdma_req_map
*vec
)
253 struct ib_send_wr write_wr
;
260 struct svc_rdma_op_ctxt
*ctxt
;
262 BUG_ON(vec
->count
> RPCSVC_MAXPAGES
);
263 dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
264 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
265 rmr
, (unsigned long long)to
, xdr_off
,
266 write_len
, vec
->sge
, vec
->count
);
268 ctxt
= svc_rdma_get_context(xprt
);
269 ctxt
->direction
= DMA_TO_DEVICE
;
272 /* Find the SGE associated with xdr_off */
273 for (bc
= xdr_off
, xdr_sge_no
= 1; bc
&& xdr_sge_no
< vec
->count
;
275 if (vec
->sge
[xdr_sge_no
].iov_len
> bc
)
277 bc
-= vec
->sge
[xdr_sge_no
].iov_len
;
284 /* Copy the remaining SGE */
286 sge_bytes
= min_t(size_t,
287 bc
, vec
->sge
[xdr_sge_no
].iov_len
-sge_off
);
288 sge
[sge_no
].length
= sge_bytes
;
291 ib_dma_map_single(xprt
->sc_cm_id
->device
,
293 vec
->sge
[xdr_sge_no
].iov_base
+ sge_off
,
294 sge_bytes
, DMA_TO_DEVICE
);
295 if (ib_dma_mapping_error(xprt
->sc_cm_id
->device
,
298 atomic_inc(&xprt
->sc_dma_used
);
299 sge
[sge_no
].lkey
= xprt
->sc_dma_lkey
;
301 sge
[sge_no
].addr
= (unsigned long)
302 vec
->sge
[xdr_sge_no
].iov_base
+ sge_off
;
303 sge
[sge_no
].lkey
= vec
->frmr
->mr
->lkey
;
306 ctxt
->frmr
= vec
->frmr
;
310 BUG_ON(xdr_sge_no
> vec
->count
);
314 /* Prepare WRITE WR */
315 memset(&write_wr
, 0, sizeof write_wr
);
316 ctxt
->wr_op
= IB_WR_RDMA_WRITE
;
317 write_wr
.wr_id
= (unsigned long)ctxt
;
318 write_wr
.sg_list
= &sge
[0];
319 write_wr
.num_sge
= sge_no
;
320 write_wr
.opcode
= IB_WR_RDMA_WRITE
;
321 write_wr
.send_flags
= IB_SEND_SIGNALED
;
322 write_wr
.wr
.rdma
.rkey
= rmr
;
323 write_wr
.wr
.rdma
.remote_addr
= to
;
326 atomic_inc(&rdma_stat_write
);
327 if (svc_rdma_send(xprt
, &write_wr
))
331 svc_rdma_put_context(ctxt
, 0);
332 /* Fatal error, close transport */
336 static int send_write_chunks(struct svcxprt_rdma
*xprt
,
337 struct rpcrdma_msg
*rdma_argp
,
338 struct rpcrdma_msg
*rdma_resp
,
339 struct svc_rqst
*rqstp
,
340 struct svc_rdma_req_map
*vec
)
342 u32 xfer_len
= rqstp
->rq_res
.page_len
+ rqstp
->rq_res
.tail
[0].iov_len
;
348 struct rpcrdma_write_array
*arg_ary
;
349 struct rpcrdma_write_array
*res_ary
;
352 arg_ary
= svc_rdma_get_write_array(rdma_argp
);
355 res_ary
= (struct rpcrdma_write_array
*)
356 &rdma_resp
->rm_body
.rm_chunks
[1];
359 max_write
= vec
->frmr
->map_len
;
361 max_write
= xprt
->sc_max_sge
* PAGE_SIZE
;
363 /* Write chunks start at the pagelist */
364 for (xdr_off
= rqstp
->rq_res
.head
[0].iov_len
, chunk_no
= 0;
365 xfer_len
&& chunk_no
< arg_ary
->wc_nchunks
;
367 struct rpcrdma_segment
*arg_ch
;
370 arg_ch
= &arg_ary
->wc_array
[chunk_no
].wc_target
;
371 write_len
= min(xfer_len
, arg_ch
->rs_length
);
373 /* Prepare the response chunk given the length actually
375 rs_offset
= get_unaligned(&(arg_ch
->rs_offset
));
376 svc_rdma_xdr_encode_array_chunk(res_ary
, chunk_no
,
383 this_write
= min(write_len
, max_write
);
384 ret
= send_write(xprt
, rqstp
,
386 rs_offset
+ chunk_off
,
391 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
395 chunk_off
+= this_write
;
396 xdr_off
+= this_write
;
397 xfer_len
-= this_write
;
398 write_len
-= this_write
;
401 /* Update the req with the number of chunks actually used */
402 svc_rdma_xdr_encode_write_list(rdma_resp
, chunk_no
);
404 return rqstp
->rq_res
.page_len
+ rqstp
->rq_res
.tail
[0].iov_len
;
407 static int send_reply_chunks(struct svcxprt_rdma
*xprt
,
408 struct rpcrdma_msg
*rdma_argp
,
409 struct rpcrdma_msg
*rdma_resp
,
410 struct svc_rqst
*rqstp
,
411 struct svc_rdma_req_map
*vec
)
413 u32 xfer_len
= rqstp
->rq_res
.len
;
419 struct rpcrdma_segment
*ch
;
420 struct rpcrdma_write_array
*arg_ary
;
421 struct rpcrdma_write_array
*res_ary
;
424 arg_ary
= svc_rdma_get_reply_array(rdma_argp
);
427 /* XXX: need to fix when reply lists occur with read-list and or
429 res_ary
= (struct rpcrdma_write_array
*)
430 &rdma_resp
->rm_body
.rm_chunks
[2];
433 max_write
= vec
->frmr
->map_len
;
435 max_write
= xprt
->sc_max_sge
* PAGE_SIZE
;
437 /* xdr offset starts at RPC message */
438 for (xdr_off
= 0, chunk_no
= 0;
439 xfer_len
&& chunk_no
< arg_ary
->wc_nchunks
;
442 ch
= &arg_ary
->wc_array
[chunk_no
].wc_target
;
443 write_len
= min(xfer_len
, ch
->rs_length
);
445 /* Prepare the reply chunk given the length actually
447 rs_offset
= get_unaligned(&(ch
->rs_offset
));
448 svc_rdma_xdr_encode_array_chunk(res_ary
, chunk_no
,
449 ch
->rs_handle
, rs_offset
,
455 this_write
= min(write_len
, max_write
);
456 ret
= send_write(xprt
, rqstp
,
458 rs_offset
+ chunk_off
,
463 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
467 chunk_off
+= this_write
;
468 xdr_off
+= this_write
;
469 xfer_len
-= this_write
;
470 write_len
-= this_write
;
473 /* Update the req with the number of chunks actually used */
474 svc_rdma_xdr_encode_reply_array(res_ary
, chunk_no
);
476 return rqstp
->rq_res
.len
;
479 /* This function prepares the portion of the RPCRDMA message to be
480 * sent in the RDMA_SEND. This function is called after data sent via
481 * RDMA has already been transmitted. There are three cases:
482 * - The RPCRDMA header, RPC header, and payload are all sent in a
483 * single RDMA_SEND. This is the "inline" case.
484 * - The RPCRDMA header and some portion of the RPC header and data
485 * are sent via this RDMA_SEND and another portion of the data is
487 * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
488 * header and data are all transmitted via RDMA.
489 * In all three cases, this function prepares the RPCRDMA header in
490 * sge[0], the 'type' parameter indicates the type to place in the
491 * RPCRDMA header, and the 'byte_count' field indicates how much of
492 * the XDR to include in this RDMA_SEND.
494 static int send_reply(struct svcxprt_rdma
*rdma
,
495 struct svc_rqst
*rqstp
,
497 struct rpcrdma_msg
*rdma_resp
,
498 struct svc_rdma_op_ctxt
*ctxt
,
499 struct svc_rdma_req_map
*vec
,
502 struct ib_send_wr send_wr
;
503 struct ib_send_wr inv_wr
;
509 /* Post a recv buffer to handle another request. */
510 ret
= svc_rdma_post_recv(rdma
);
513 "svcrdma: could not post a receive buffer, err=%d."
514 "Closing transport %p.\n", ret
, rdma
);
515 set_bit(XPT_CLOSE
, &rdma
->sc_xprt
.xpt_flags
);
516 svc_rdma_put_context(ctxt
, 0);
520 /* Prepare the context */
521 ctxt
->pages
[0] = page
;
523 ctxt
->frmr
= vec
->frmr
;
525 set_bit(RDMACTXT_F_FAST_UNREG
, &ctxt
->flags
);
527 clear_bit(RDMACTXT_F_FAST_UNREG
, &ctxt
->flags
);
529 /* Prepare the SGE for the RPCRDMA Header */
531 ib_dma_map_page(rdma
->sc_cm_id
->device
,
532 page
, 0, PAGE_SIZE
, DMA_TO_DEVICE
);
533 if (ib_dma_mapping_error(rdma
->sc_cm_id
->device
, ctxt
->sge
[0].addr
))
535 atomic_inc(&rdma
->sc_dma_used
);
537 ctxt
->direction
= DMA_TO_DEVICE
;
539 ctxt
->sge
[0].length
= svc_rdma_xdr_get_reply_hdr_len(rdma_resp
);
540 ctxt
->sge
[0].lkey
= rdma
->sc_dma_lkey
;
542 /* Determine how many of our SGE are to be transmitted */
543 for (sge_no
= 1; byte_count
&& sge_no
< vec
->count
; sge_no
++) {
544 sge_bytes
= min_t(size_t, vec
->sge
[sge_no
].iov_len
, byte_count
);
545 byte_count
-= sge_bytes
;
547 ctxt
->sge
[sge_no
].addr
=
548 ib_dma_map_single(rdma
->sc_cm_id
->device
,
549 vec
->sge
[sge_no
].iov_base
,
550 sge_bytes
, DMA_TO_DEVICE
);
551 if (ib_dma_mapping_error(rdma
->sc_cm_id
->device
,
552 ctxt
->sge
[sge_no
].addr
))
554 atomic_inc(&rdma
->sc_dma_used
);
555 ctxt
->sge
[sge_no
].lkey
= rdma
->sc_dma_lkey
;
557 ctxt
->sge
[sge_no
].addr
= (unsigned long)
558 vec
->sge
[sge_no
].iov_base
;
559 ctxt
->sge
[sge_no
].lkey
= vec
->frmr
->mr
->lkey
;
561 ctxt
->sge
[sge_no
].length
= sge_bytes
;
563 BUG_ON(byte_count
!= 0);
565 /* Save all respages in the ctxt and remove them from the
566 * respages array. They are our pages until the I/O
569 for (page_no
= 0; page_no
< rqstp
->rq_resused
; page_no
++) {
570 ctxt
->pages
[page_no
+1] = rqstp
->rq_respages
[page_no
];
572 rqstp
->rq_respages
[page_no
] = NULL
;
574 * If there are more pages than SGE, terminate SGE
575 * list so that svc_rdma_unmap_dma doesn't attempt to
578 if (page_no
+1 >= sge_no
)
579 ctxt
->sge
[page_no
+1].length
= 0;
581 BUG_ON(sge_no
> rdma
->sc_max_sge
);
582 BUG_ON(sge_no
> ctxt
->count
);
583 memset(&send_wr
, 0, sizeof send_wr
);
584 ctxt
->wr_op
= IB_WR_SEND
;
585 send_wr
.wr_id
= (unsigned long)ctxt
;
586 send_wr
.sg_list
= ctxt
->sge
;
587 send_wr
.num_sge
= sge_no
;
588 send_wr
.opcode
= IB_WR_SEND
;
589 send_wr
.send_flags
= IB_SEND_SIGNALED
;
591 /* Prepare INVALIDATE WR */
592 memset(&inv_wr
, 0, sizeof inv_wr
);
593 inv_wr
.opcode
= IB_WR_LOCAL_INV
;
594 inv_wr
.send_flags
= IB_SEND_SIGNALED
;
595 inv_wr
.ex
.invalidate_rkey
=
597 send_wr
.next
= &inv_wr
;
600 ret
= svc_rdma_send(rdma
, &send_wr
);
607 svc_rdma_put_frmr(rdma
, vec
->frmr
);
608 svc_rdma_put_context(ctxt
, 1);
612 void svc_rdma_prep_reply_hdr(struct svc_rqst
*rqstp
)
617 * Return the start of an xdr buffer.
619 static void *xdr_start(struct xdr_buf
*xdr
)
621 return xdr
->head
[0].iov_base
-
624 xdr
->tail
[0].iov_len
-
625 xdr
->head
[0].iov_len
);
628 int svc_rdma_sendto(struct svc_rqst
*rqstp
)
630 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
631 struct svcxprt_rdma
*rdma
=
632 container_of(xprt
, struct svcxprt_rdma
, sc_xprt
);
633 struct rpcrdma_msg
*rdma_argp
;
634 struct rpcrdma_msg
*rdma_resp
;
635 struct rpcrdma_write_array
*reply_ary
;
636 enum rpcrdma_proc reply_type
;
639 struct page
*res_page
;
640 struct svc_rdma_op_ctxt
*ctxt
;
641 struct svc_rdma_req_map
*vec
;
643 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp
);
645 /* Get the RDMA request header. */
646 rdma_argp
= xdr_start(&rqstp
->rq_arg
);
648 /* Build an req vec for the XDR */
649 ctxt
= svc_rdma_get_context(rdma
);
650 ctxt
->direction
= DMA_TO_DEVICE
;
651 vec
= svc_rdma_get_req_map();
652 ret
= map_xdr(rdma
, &rqstp
->rq_res
, vec
);
655 inline_bytes
= rqstp
->rq_res
.len
;
657 /* Create the RDMA response header */
658 res_page
= svc_rdma_get_page();
659 rdma_resp
= page_address(res_page
);
660 reply_ary
= svc_rdma_get_reply_array(rdma_argp
);
662 reply_type
= RDMA_NOMSG
;
664 reply_type
= RDMA_MSG
;
665 svc_rdma_xdr_encode_reply_header(rdma
, rdma_argp
,
666 rdma_resp
, reply_type
);
668 /* Send any write-chunk data and build resp write-list */
669 ret
= send_write_chunks(rdma
, rdma_argp
, rdma_resp
,
672 printk(KERN_ERR
"svcrdma: failed to send write chunks, rc=%d\n",
678 /* Send any reply-list data and update resp reply-list */
679 ret
= send_reply_chunks(rdma
, rdma_argp
, rdma_resp
,
682 printk(KERN_ERR
"svcrdma: failed to send reply chunks, rc=%d\n",
688 ret
= send_reply(rdma
, rqstp
, res_page
, rdma_resp
, ctxt
, vec
,
690 svc_rdma_put_req_map(vec
);
691 dprintk("svcrdma: send_reply returns %d\n", ret
);
697 svc_rdma_put_req_map(vec
);
698 svc_rdma_put_context(ctxt
, 0);