1 /*
2  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
3  *
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
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
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.
21  *
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
25  *      permission.
26  *
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.
38  *
39  * Author: Tom Tucker <tom@opengridcomputing.com>
40  */
41 
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>
49 
50 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
51 
52 /* Encode an XDR as an array of IB SGE
53  *
54  * Assumptions:
55  * - head[0] is physically contiguous.
56  * - tail[0] is physically contiguous.
57  * - pages[] is not physically or virtually contiguous and consists of
58  *   PAGE_SIZE elements.
59  *
60  * Output:
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.
65  *
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.
71  */
72 static int fast_reg_xdr(struct svcxprt_rdma *xprt,
73 			struct xdr_buf *xdr,
74 			struct svc_rdma_req_map *vec)
75 {
76 	int sge_no;
77 	u32 sge_bytes;
78 	u32 page_bytes;
79 	u32 page_off;
80 	int page_no = 0;
81 	u8 *frva;
82 	struct svc_rdma_fastreg_mr *frmr;
83 
84 	frmr = svc_rdma_get_frmr(xprt);
85 	if (IS_ERR(frmr))
86 		return -ENOMEM;
87 	vec->frmr = frmr;
88 
89 	/* Skip the RPCRDMA header */
90 	sge_no = 1;
91 
92 	/* Map the head. */
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;
96 	vec->count = 2;
97 	sge_no++;
98 
99 	/* Map the XDR head */
100 	frmr->kva = frva;
101 	frmr->direction = DMA_TO_DEVICE;
102 	frmr->access_flags = 0;
103 	frmr->map_len = PAGE_SIZE;
104 	frmr->page_list_len = 1;
105 	page_off = (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
106 	frmr->page_list->page_list[page_no] =
107 		ib_dma_map_page(xprt->sc_cm_id->device,
108 				virt_to_page(xdr->head[0].iov_base),
109 				page_off,
110 				PAGE_SIZE - page_off,
111 				DMA_TO_DEVICE);
112 	if (ib_dma_mapping_error(xprt->sc_cm_id->device,
113 				 frmr->page_list->page_list[page_no]))
114 		goto fatal_err;
115 	atomic_inc(&xprt->sc_dma_used);
116 
117 	/* Map the XDR page list */
118 	page_off = xdr->page_base;
119 	page_bytes = xdr->page_len + page_off;
120 	if (!page_bytes)
121 		goto encode_tail;
122 
123 	/* Map the pages */
124 	vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
125 	vec->sge[sge_no].iov_len = page_bytes;
126 	sge_no++;
127 	while (page_bytes) {
128 		struct page *page;
129 
130 		page = xdr->pages[page_no++];
131 		sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
132 		page_bytes -= sge_bytes;
133 
134 		frmr->page_list->page_list[page_no] =
135 			ib_dma_map_page(xprt->sc_cm_id->device,
136 					page, page_off,
137 					sge_bytes, DMA_TO_DEVICE);
138 		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
139 					 frmr->page_list->page_list[page_no]))
140 			goto fatal_err;
141 
142 		atomic_inc(&xprt->sc_dma_used);
143 		page_off = 0; /* reset for next time through loop */
144 		frmr->map_len += PAGE_SIZE;
145 		frmr->page_list_len++;
146 	}
147 	vec->count++;
148 
149  encode_tail:
150 	/* Map tail */
151 	if (0 == xdr->tail[0].iov_len)
152 		goto done;
153 
154 	vec->count++;
155 	vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
156 
157 	if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) ==
158 	    ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) {
159 		/*
160 		 * If head and tail use the same page, we don't need
161 		 * to map it again.
162 		 */
163 		vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
164 	} else {
165 		void *va;
166 
167 		/* Map another page for the tail */
168 		page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
169 		va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK);
170 		vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
171 
172 		frmr->page_list->page_list[page_no] =
173 		    ib_dma_map_page(xprt->sc_cm_id->device, virt_to_page(va),
174 				    page_off,
175 				    PAGE_SIZE,
176 				    DMA_TO_DEVICE);
177 		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
178 					 frmr->page_list->page_list[page_no]))
179 			goto fatal_err;
180 		atomic_inc(&xprt->sc_dma_used);
181 		frmr->map_len += PAGE_SIZE;
182 		frmr->page_list_len++;
183 	}
184 
185  done:
186 	if (svc_rdma_fastreg(xprt, frmr))
187 		goto fatal_err;
188 
189 	return 0;
190 
191  fatal_err:
192 	printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
193 	vec->frmr = NULL;
194 	svc_rdma_put_frmr(xprt, frmr);
195 	return -EIO;
196 }
197 
198 static int map_xdr(struct svcxprt_rdma *xprt,
199 		   struct xdr_buf *xdr,
200 		   struct svc_rdma_req_map *vec)
201 {
202 	int sge_no;
203 	u32 sge_bytes;
204 	u32 page_bytes;
205 	u32 page_off;
206 	int page_no;
207 
208 	BUG_ON(xdr->len !=
209 	       (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
210 
211 	if (xprt->sc_frmr_pg_list_len)
212 		return fast_reg_xdr(xprt, xdr, vec);
213 
214 	/* Skip the first sge, this is for the RPCRDMA header */
215 	sge_no = 1;
216 
217 	/* Head SGE */
218 	vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
219 	vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
220 	sge_no++;
221 
222 	/* pages SGE */
223 	page_no = 0;
224 	page_bytes = xdr->page_len;
225 	page_off = xdr->page_base;
226 	while (page_bytes) {
227 		vec->sge[sge_no].iov_base =
228 			page_address(xdr->pages[page_no]) + page_off;
229 		sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
230 		page_bytes -= sge_bytes;
231 		vec->sge[sge_no].iov_len = sge_bytes;
232 
233 		sge_no++;
234 		page_no++;
235 		page_off = 0; /* reset for next time through loop */
236 	}
237 
238 	/* Tail SGE */
239 	if (xdr->tail[0].iov_len) {
240 		vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
241 		vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
242 		sge_no++;
243 	}
244 
245 	dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
246 		"page_base %u page_len %u head_len %zu tail_len %zu\n",
247 		sge_no, page_no, xdr->page_base, xdr->page_len,
248 		xdr->head[0].iov_len, xdr->tail[0].iov_len);
249 
250 	vec->count = sge_no;
251 	return 0;
252 }
253 
254 static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
255 			      struct xdr_buf *xdr,
256 			      u32 xdr_off, size_t len, int dir)
257 {
258 	struct page *page;
259 	dma_addr_t dma_addr;
260 	if (xdr_off < xdr->head[0].iov_len) {
261 		/* This offset is in the head */
262 		xdr_off += (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
263 		page = virt_to_page(xdr->head[0].iov_base);
264 	} else {
265 		xdr_off -= xdr->head[0].iov_len;
266 		if (xdr_off < xdr->page_len) {
267 			/* This offset is in the page list */
268 			page = xdr->pages[xdr_off >> PAGE_SHIFT];
269 			xdr_off &= ~PAGE_MASK;
270 		} else {
271 			/* This offset is in the tail */
272 			xdr_off -= xdr->page_len;
273 			xdr_off += (unsigned long)
274 				xdr->tail[0].iov_base & ~PAGE_MASK;
275 			page = virt_to_page(xdr->tail[0].iov_base);
276 		}
277 	}
278 	dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off,
279 				   min_t(size_t, PAGE_SIZE, len), dir);
280 	return dma_addr;
281 }
282 
283 /* Assumptions:
284  * - We are using FRMR
285  *     - or -
286  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
287  */
288 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
289 		      u32 rmr, u64 to,
290 		      u32 xdr_off, int write_len,
291 		      struct svc_rdma_req_map *vec)
292 {
293 	struct ib_send_wr write_wr;
294 	struct ib_sge *sge;
295 	int xdr_sge_no;
296 	int sge_no;
297 	int sge_bytes;
298 	int sge_off;
299 	int bc;
300 	struct svc_rdma_op_ctxt *ctxt;
301 
302 	BUG_ON(vec->count > RPCSVC_MAXPAGES);
303 	dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
304 		"write_len=%d, vec->sge=%p, vec->count=%lu\n",
305 		rmr, (unsigned long long)to, xdr_off,
306 		write_len, vec->sge, vec->count);
307 
308 	ctxt = svc_rdma_get_context(xprt);
309 	ctxt->direction = DMA_TO_DEVICE;
310 	sge = ctxt->sge;
311 
312 	/* Find the SGE associated with xdr_off */
313 	for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
314 	     xdr_sge_no++) {
315 		if (vec->sge[xdr_sge_no].iov_len > bc)
316 			break;
317 		bc -= vec->sge[xdr_sge_no].iov_len;
318 	}
319 
320 	sge_off = bc;
321 	bc = write_len;
322 	sge_no = 0;
323 
324 	/* Copy the remaining SGE */
325 	while (bc != 0) {
326 		sge_bytes = min_t(size_t,
327 			  bc, vec->sge[xdr_sge_no].iov_len-sge_off);
328 		sge[sge_no].length = sge_bytes;
329 		if (!vec->frmr) {
330 			sge[sge_no].addr =
331 				dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
332 					    sge_bytes, DMA_TO_DEVICE);
333 			xdr_off += sge_bytes;
334 			if (ib_dma_mapping_error(xprt->sc_cm_id->device,
335 						 sge[sge_no].addr))
336 				goto err;
337 			atomic_inc(&xprt->sc_dma_used);
338 			sge[sge_no].lkey = xprt->sc_dma_lkey;
339 		} else {
340 			sge[sge_no].addr = (unsigned long)
341 				vec->sge[xdr_sge_no].iov_base + sge_off;
342 			sge[sge_no].lkey = vec->frmr->mr->lkey;
343 		}
344 		ctxt->count++;
345 		ctxt->frmr = vec->frmr;
346 		sge_off = 0;
347 		sge_no++;
348 		xdr_sge_no++;
349 		BUG_ON(xdr_sge_no > vec->count);
350 		bc -= sge_bytes;
351 	}
352 
353 	/* Prepare WRITE WR */
354 	memset(&write_wr, 0, sizeof write_wr);
355 	ctxt->wr_op = IB_WR_RDMA_WRITE;
356 	write_wr.wr_id = (unsigned long)ctxt;
357 	write_wr.sg_list = &sge[0];
358 	write_wr.num_sge = sge_no;
359 	write_wr.opcode = IB_WR_RDMA_WRITE;
360 	write_wr.send_flags = IB_SEND_SIGNALED;
361 	write_wr.wr.rdma.rkey = rmr;
362 	write_wr.wr.rdma.remote_addr = to;
363 
364 	/* Post It */
365 	atomic_inc(&rdma_stat_write);
366 	if (svc_rdma_send(xprt, &write_wr))
367 		goto err;
368 	return 0;
369  err:
370 	svc_rdma_unmap_dma(ctxt);
371 	svc_rdma_put_frmr(xprt, vec->frmr);
372 	svc_rdma_put_context(ctxt, 0);
373 	/* Fatal error, close transport */
374 	return -EIO;
375 }
376 
377 static int send_write_chunks(struct svcxprt_rdma *xprt,
378 			     struct rpcrdma_msg *rdma_argp,
379 			     struct rpcrdma_msg *rdma_resp,
380 			     struct svc_rqst *rqstp,
381 			     struct svc_rdma_req_map *vec)
382 {
383 	u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
384 	int write_len;
385 	int max_write;
386 	u32 xdr_off;
387 	int chunk_off;
388 	int chunk_no;
389 	struct rpcrdma_write_array *arg_ary;
390 	struct rpcrdma_write_array *res_ary;
391 	int ret;
392 
393 	arg_ary = svc_rdma_get_write_array(rdma_argp);
394 	if (!arg_ary)
395 		return 0;
396 	res_ary = (struct rpcrdma_write_array *)
397 		&rdma_resp->rm_body.rm_chunks[1];
398 
399 	if (vec->frmr)
400 		max_write = vec->frmr->map_len;
401 	else
402 		max_write = xprt->sc_max_sge * PAGE_SIZE;
403 
404 	/* Write chunks start at the pagelist */
405 	for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
406 	     xfer_len && chunk_no < arg_ary->wc_nchunks;
407 	     chunk_no++) {
408 		struct rpcrdma_segment *arg_ch;
409 		u64 rs_offset;
410 
411 		arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
412 		write_len = min(xfer_len, ntohl(arg_ch->rs_length));
413 
414 		/* Prepare the response chunk given the length actually
415 		 * written */
416 		xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
417 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
418 						arg_ch->rs_handle,
419 						arg_ch->rs_offset,
420 						write_len);
421 		chunk_off = 0;
422 		while (write_len) {
423 			int this_write;
424 			this_write = min(write_len, max_write);
425 			ret = send_write(xprt, rqstp,
426 					 ntohl(arg_ch->rs_handle),
427 					 rs_offset + chunk_off,
428 					 xdr_off,
429 					 this_write,
430 					 vec);
431 			if (ret) {
432 				dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
433 					ret);
434 				return -EIO;
435 			}
436 			chunk_off += this_write;
437 			xdr_off += this_write;
438 			xfer_len -= this_write;
439 			write_len -= this_write;
440 		}
441 	}
442 	/* Update the req with the number of chunks actually used */
443 	svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
444 
445 	return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
446 }
447 
448 static int send_reply_chunks(struct svcxprt_rdma *xprt,
449 			     struct rpcrdma_msg *rdma_argp,
450 			     struct rpcrdma_msg *rdma_resp,
451 			     struct svc_rqst *rqstp,
452 			     struct svc_rdma_req_map *vec)
453 {
454 	u32 xfer_len = rqstp->rq_res.len;
455 	int write_len;
456 	int max_write;
457 	u32 xdr_off;
458 	int chunk_no;
459 	int chunk_off;
460 	int nchunks;
461 	struct rpcrdma_segment *ch;
462 	struct rpcrdma_write_array *arg_ary;
463 	struct rpcrdma_write_array *res_ary;
464 	int ret;
465 
466 	arg_ary = svc_rdma_get_reply_array(rdma_argp);
467 	if (!arg_ary)
468 		return 0;
469 	/* XXX: need to fix when reply lists occur with read-list and or
470 	 * write-list */
471 	res_ary = (struct rpcrdma_write_array *)
472 		&rdma_resp->rm_body.rm_chunks[2];
473 
474 	if (vec->frmr)
475 		max_write = vec->frmr->map_len;
476 	else
477 		max_write = xprt->sc_max_sge * PAGE_SIZE;
478 
479 	/* xdr offset starts at RPC message */
480 	nchunks = ntohl(arg_ary->wc_nchunks);
481 	for (xdr_off = 0, chunk_no = 0;
482 	     xfer_len && chunk_no < nchunks;
483 	     chunk_no++) {
484 		u64 rs_offset;
485 		ch = &arg_ary->wc_array[chunk_no].wc_target;
486 		write_len = min(xfer_len, htonl(ch->rs_length));
487 
488 		/* Prepare the reply chunk given the length actually
489 		 * written */
490 		xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset);
491 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
492 						ch->rs_handle, ch->rs_offset,
493 						write_len);
494 		chunk_off = 0;
495 		while (write_len) {
496 			int this_write;
497 
498 			this_write = min(write_len, max_write);
499 			ret = send_write(xprt, rqstp,
500 					 ntohl(ch->rs_handle),
501 					 rs_offset + chunk_off,
502 					 xdr_off,
503 					 this_write,
504 					 vec);
505 			if (ret) {
506 				dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
507 					ret);
508 				return -EIO;
509 			}
510 			chunk_off += this_write;
511 			xdr_off += this_write;
512 			xfer_len -= this_write;
513 			write_len -= this_write;
514 		}
515 	}
516 	/* Update the req with the number of chunks actually used */
517 	svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
518 
519 	return rqstp->rq_res.len;
520 }
521 
522 /* This function prepares the portion of the RPCRDMA message to be
523  * sent in the RDMA_SEND. This function is called after data sent via
524  * RDMA has already been transmitted. There are three cases:
525  * - The RPCRDMA header, RPC header, and payload are all sent in a
526  *   single RDMA_SEND. This is the "inline" case.
527  * - The RPCRDMA header and some portion of the RPC header and data
528  *   are sent via this RDMA_SEND and another portion of the data is
529  *   sent via RDMA.
530  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
531  *   header and data are all transmitted via RDMA.
532  * In all three cases, this function prepares the RPCRDMA header in
533  * sge[0], the 'type' parameter indicates the type to place in the
534  * RPCRDMA header, and the 'byte_count' field indicates how much of
535  * the XDR to include in this RDMA_SEND. NB: The offset of the payload
536  * to send is zero in the XDR.
537  */
538 static int send_reply(struct svcxprt_rdma *rdma,
539 		      struct svc_rqst *rqstp,
540 		      struct page *page,
541 		      struct rpcrdma_msg *rdma_resp,
542 		      struct svc_rdma_op_ctxt *ctxt,
543 		      struct svc_rdma_req_map *vec,
544 		      int byte_count)
545 {
546 	struct ib_send_wr send_wr;
547 	struct ib_send_wr inv_wr;
548 	int sge_no;
549 	int sge_bytes;
550 	int page_no;
551 	int pages;
552 	int ret;
553 
554 	/* Post a recv buffer to handle another request. */
555 	ret = svc_rdma_post_recv(rdma);
556 	if (ret) {
557 		printk(KERN_INFO
558 		       "svcrdma: could not post a receive buffer, err=%d."
559 		       "Closing transport %p.\n", ret, rdma);
560 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
561 		svc_rdma_put_frmr(rdma, vec->frmr);
562 		svc_rdma_put_context(ctxt, 0);
563 		return -ENOTCONN;
564 	}
565 
566 	/* Prepare the context */
567 	ctxt->pages[0] = page;
568 	ctxt->count = 1;
569 	ctxt->frmr = vec->frmr;
570 	if (vec->frmr)
571 		set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
572 	else
573 		clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
574 
575 	/* Prepare the SGE for the RPCRDMA Header */
576 	ctxt->sge[0].lkey = rdma->sc_dma_lkey;
577 	ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
578 	ctxt->sge[0].addr =
579 	    ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
580 			    ctxt->sge[0].length, DMA_TO_DEVICE);
581 	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
582 		goto err;
583 	atomic_inc(&rdma->sc_dma_used);
584 
585 	ctxt->direction = DMA_TO_DEVICE;
586 
587 	/* Map the payload indicated by 'byte_count' */
588 	for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
589 		int xdr_off = 0;
590 		sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
591 		byte_count -= sge_bytes;
592 		if (!vec->frmr) {
593 			ctxt->sge[sge_no].addr =
594 				dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
595 					    sge_bytes, DMA_TO_DEVICE);
596 			xdr_off += sge_bytes;
597 			if (ib_dma_mapping_error(rdma->sc_cm_id->device,
598 						 ctxt->sge[sge_no].addr))
599 				goto err;
600 			atomic_inc(&rdma->sc_dma_used);
601 			ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
602 		} else {
603 			ctxt->sge[sge_no].addr = (unsigned long)
604 				vec->sge[sge_no].iov_base;
605 			ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey;
606 		}
607 		ctxt->sge[sge_no].length = sge_bytes;
608 	}
609 	BUG_ON(byte_count != 0);
610 
611 	/* Save all respages in the ctxt and remove them from the
612 	 * respages array. They are our pages until the I/O
613 	 * completes.
614 	 */
615 	pages = rqstp->rq_next_page - rqstp->rq_respages;
616 	for (page_no = 0; page_no < pages; page_no++) {
617 		ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
618 		ctxt->count++;
619 		rqstp->rq_respages[page_no] = NULL;
620 		/*
621 		 * If there are more pages than SGE, terminate SGE
622 		 * list so that svc_rdma_unmap_dma doesn't attempt to
623 		 * unmap garbage.
624 		 */
625 		if (page_no+1 >= sge_no)
626 			ctxt->sge[page_no+1].length = 0;
627 	}
628 	BUG_ON(sge_no > rdma->sc_max_sge);
629 	memset(&send_wr, 0, sizeof send_wr);
630 	ctxt->wr_op = IB_WR_SEND;
631 	send_wr.wr_id = (unsigned long)ctxt;
632 	send_wr.sg_list = ctxt->sge;
633 	send_wr.num_sge = sge_no;
634 	send_wr.opcode = IB_WR_SEND;
635 	send_wr.send_flags =  IB_SEND_SIGNALED;
636 	if (vec->frmr) {
637 		/* Prepare INVALIDATE WR */
638 		memset(&inv_wr, 0, sizeof inv_wr);
639 		inv_wr.opcode = IB_WR_LOCAL_INV;
640 		inv_wr.send_flags = IB_SEND_SIGNALED;
641 		inv_wr.ex.invalidate_rkey =
642 			vec->frmr->mr->lkey;
643 		send_wr.next = &inv_wr;
644 	}
645 
646 	ret = svc_rdma_send(rdma, &send_wr);
647 	if (ret)
648 		goto err;
649 
650 	return 0;
651 
652  err:
653 	svc_rdma_unmap_dma(ctxt);
654 	svc_rdma_put_frmr(rdma, vec->frmr);
655 	svc_rdma_put_context(ctxt, 1);
656 	return -EIO;
657 }
658 
659 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
660 {
661 }
662 
663 /*
664  * Return the start of an xdr buffer.
665  */
666 static void *xdr_start(struct xdr_buf *xdr)
667 {
668 	return xdr->head[0].iov_base -
669 		(xdr->len -
670 		 xdr->page_len -
671 		 xdr->tail[0].iov_len -
672 		 xdr->head[0].iov_len);
673 }
674 
675 int svc_rdma_sendto(struct svc_rqst *rqstp)
676 {
677 	struct svc_xprt *xprt = rqstp->rq_xprt;
678 	struct svcxprt_rdma *rdma =
679 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
680 	struct rpcrdma_msg *rdma_argp;
681 	struct rpcrdma_msg *rdma_resp;
682 	struct rpcrdma_write_array *reply_ary;
683 	enum rpcrdma_proc reply_type;
684 	int ret;
685 	int inline_bytes;
686 	struct page *res_page;
687 	struct svc_rdma_op_ctxt *ctxt;
688 	struct svc_rdma_req_map *vec;
689 
690 	dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
691 
692 	/* Get the RDMA request header. */
693 	rdma_argp = xdr_start(&rqstp->rq_arg);
694 
695 	/* Build an req vec for the XDR */
696 	ctxt = svc_rdma_get_context(rdma);
697 	ctxt->direction = DMA_TO_DEVICE;
698 	vec = svc_rdma_get_req_map();
699 	ret = map_xdr(rdma, &rqstp->rq_res, vec);
700 	if (ret)
701 		goto err0;
702 	inline_bytes = rqstp->rq_res.len;
703 
704 	/* Create the RDMA response header */
705 	res_page = svc_rdma_get_page();
706 	rdma_resp = page_address(res_page);
707 	reply_ary = svc_rdma_get_reply_array(rdma_argp);
708 	if (reply_ary)
709 		reply_type = RDMA_NOMSG;
710 	else
711 		reply_type = RDMA_MSG;
712 	svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
713 					 rdma_resp, reply_type);
714 
715 	/* Send any write-chunk data and build resp write-list */
716 	ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
717 				rqstp, vec);
718 	if (ret < 0) {
719 		printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
720 		       ret);
721 		goto err1;
722 	}
723 	inline_bytes -= ret;
724 
725 	/* Send any reply-list data and update resp reply-list */
726 	ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
727 				rqstp, vec);
728 	if (ret < 0) {
729 		printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
730 		       ret);
731 		goto err1;
732 	}
733 	inline_bytes -= ret;
734 
735 	ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
736 			 inline_bytes);
737 	svc_rdma_put_req_map(vec);
738 	dprintk("svcrdma: send_reply returns %d\n", ret);
739 	return ret;
740 
741  err1:
742 	put_page(res_page);
743  err0:
744 	svc_rdma_put_req_map(vec);
745 	svc_rdma_put_context(ctxt, 0);
746 	return ret;
747 }
748