1 /*
2  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
4  *
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
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
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.
22  *
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
26  *      permission.
27  *
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.
39  *
40  * Author: Tom Tucker <tom@opengridcomputing.com>
41  */
42 
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <asm/unaligned.h>
47 #include <rdma/ib_verbs.h>
48 #include <rdma/rdma_cm.h>
49 #include <linux/sunrpc/svc_rdma.h>
50 
51 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
52 
53 static u32 xdr_padsize(u32 len)
54 {
55 	return (len & 3) ? (4 - (len & 3)) : 0;
56 }
57 
58 int svc_rdma_map_xdr(struct svcxprt_rdma *xprt,
59 		     struct xdr_buf *xdr,
60 		     struct svc_rdma_req_map *vec,
61 		     bool write_chunk_present)
62 {
63 	int sge_no;
64 	u32 sge_bytes;
65 	u32 page_bytes;
66 	u32 page_off;
67 	int page_no;
68 
69 	if (xdr->len !=
70 	    (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)) {
71 		pr_err("svcrdma: %s: XDR buffer length error\n", __func__);
72 		return -EIO;
73 	}
74 
75 	/* Skip the first sge, this is for the RPCRDMA header */
76 	sge_no = 1;
77 
78 	/* Head SGE */
79 	vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
80 	vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
81 	sge_no++;
82 
83 	/* pages SGE */
84 	page_no = 0;
85 	page_bytes = xdr->page_len;
86 	page_off = xdr->page_base;
87 	while (page_bytes) {
88 		vec->sge[sge_no].iov_base =
89 			page_address(xdr->pages[page_no]) + page_off;
90 		sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
91 		page_bytes -= sge_bytes;
92 		vec->sge[sge_no].iov_len = sge_bytes;
93 
94 		sge_no++;
95 		page_no++;
96 		page_off = 0; /* reset for next time through loop */
97 	}
98 
99 	/* Tail SGE */
100 	if (xdr->tail[0].iov_len) {
101 		unsigned char *base = xdr->tail[0].iov_base;
102 		size_t len = xdr->tail[0].iov_len;
103 		u32 xdr_pad = xdr_padsize(xdr->page_len);
104 
105 		if (write_chunk_present && xdr_pad) {
106 			base += xdr_pad;
107 			len -= xdr_pad;
108 		}
109 
110 		if (len) {
111 			vec->sge[sge_no].iov_base = base;
112 			vec->sge[sge_no].iov_len = len;
113 			sge_no++;
114 		}
115 	}
116 
117 	dprintk("svcrdma: %s: sge_no %d page_no %d "
118 		"page_base %u page_len %u head_len %zu tail_len %zu\n",
119 		__func__, sge_no, page_no, xdr->page_base, xdr->page_len,
120 		xdr->head[0].iov_len, xdr->tail[0].iov_len);
121 
122 	vec->count = sge_no;
123 	return 0;
124 }
125 
126 static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
127 			      struct xdr_buf *xdr,
128 			      u32 xdr_off, size_t len, int dir)
129 {
130 	struct page *page;
131 	dma_addr_t dma_addr;
132 	if (xdr_off < xdr->head[0].iov_len) {
133 		/* This offset is in the head */
134 		xdr_off += (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
135 		page = virt_to_page(xdr->head[0].iov_base);
136 	} else {
137 		xdr_off -= xdr->head[0].iov_len;
138 		if (xdr_off < xdr->page_len) {
139 			/* This offset is in the page list */
140 			xdr_off += xdr->page_base;
141 			page = xdr->pages[xdr_off >> PAGE_SHIFT];
142 			xdr_off &= ~PAGE_MASK;
143 		} else {
144 			/* This offset is in the tail */
145 			xdr_off -= xdr->page_len;
146 			xdr_off += (unsigned long)
147 				xdr->tail[0].iov_base & ~PAGE_MASK;
148 			page = virt_to_page(xdr->tail[0].iov_base);
149 		}
150 	}
151 	dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off,
152 				   min_t(size_t, PAGE_SIZE, len), dir);
153 	return dma_addr;
154 }
155 
156 /* Returns the address of the first read chunk or <nul> if no read chunk
157  * is present
158  */
159 struct rpcrdma_read_chunk *
160 svc_rdma_get_read_chunk(struct rpcrdma_msg *rmsgp)
161 {
162 	struct rpcrdma_read_chunk *ch =
163 		(struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
164 
165 	if (ch->rc_discrim == xdr_zero)
166 		return NULL;
167 	return ch;
168 }
169 
170 /* Returns the address of the first read write array element or <nul>
171  * if no write array list is present
172  */
173 static struct rpcrdma_write_array *
174 svc_rdma_get_write_array(struct rpcrdma_msg *rmsgp)
175 {
176 	if (rmsgp->rm_body.rm_chunks[0] != xdr_zero ||
177 	    rmsgp->rm_body.rm_chunks[1] == xdr_zero)
178 		return NULL;
179 	return (struct rpcrdma_write_array *)&rmsgp->rm_body.rm_chunks[1];
180 }
181 
182 /* Returns the address of the first reply array element or <nul> if no
183  * reply array is present
184  */
185 static struct rpcrdma_write_array *
186 svc_rdma_get_reply_array(struct rpcrdma_msg *rmsgp,
187 			 struct rpcrdma_write_array *wr_ary)
188 {
189 	struct rpcrdma_read_chunk *rch;
190 	struct rpcrdma_write_array *rp_ary;
191 
192 	/* XXX: Need to fix when reply chunk may occur with read list
193 	 *	and/or write list.
194 	 */
195 	if (rmsgp->rm_body.rm_chunks[0] != xdr_zero ||
196 	    rmsgp->rm_body.rm_chunks[1] != xdr_zero)
197 		return NULL;
198 
199 	rch = svc_rdma_get_read_chunk(rmsgp);
200 	if (rch) {
201 		while (rch->rc_discrim != xdr_zero)
202 			rch++;
203 
204 		/* The reply chunk follows an empty write array located
205 		 * at 'rc_position' here. The reply array is at rc_target.
206 		 */
207 		rp_ary = (struct rpcrdma_write_array *)&rch->rc_target;
208 		goto found_it;
209 	}
210 
211 	if (wr_ary) {
212 		int chunk = be32_to_cpu(wr_ary->wc_nchunks);
213 
214 		rp_ary = (struct rpcrdma_write_array *)
215 			 &wr_ary->wc_array[chunk].wc_target.rs_length;
216 		goto found_it;
217 	}
218 
219 	/* No read list, no write list */
220 	rp_ary = (struct rpcrdma_write_array *)&rmsgp->rm_body.rm_chunks[2];
221 
222  found_it:
223 	if (rp_ary->wc_discrim == xdr_zero)
224 		return NULL;
225 	return rp_ary;
226 }
227 
228 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
229  * Responder's choice: requester signals it can handle Send With
230  * Invalidate, and responder chooses one rkey to invalidate.
231  *
232  * Find a candidate rkey to invalidate when sending a reply.  Picks the
233  * first rkey it finds in the chunks lists.
234  *
235  * Returns zero if RPC's chunk lists are empty.
236  */
237 static u32 svc_rdma_get_inv_rkey(struct rpcrdma_msg *rdma_argp,
238 				 struct rpcrdma_write_array *wr_ary,
239 				 struct rpcrdma_write_array *rp_ary)
240 {
241 	struct rpcrdma_read_chunk *rd_ary;
242 	struct rpcrdma_segment *arg_ch;
243 	u32 inv_rkey;
244 
245 	inv_rkey = 0;
246 
247 	rd_ary = svc_rdma_get_read_chunk(rdma_argp);
248 	if (rd_ary) {
249 		inv_rkey = be32_to_cpu(rd_ary->rc_target.rs_handle);
250 		goto out;
251 	}
252 
253 	if (wr_ary && be32_to_cpu(wr_ary->wc_nchunks)) {
254 		arg_ch = &wr_ary->wc_array[0].wc_target;
255 		inv_rkey = be32_to_cpu(arg_ch->rs_handle);
256 		goto out;
257 	}
258 
259 	if (rp_ary && be32_to_cpu(rp_ary->wc_nchunks)) {
260 		arg_ch = &rp_ary->wc_array[0].wc_target;
261 		inv_rkey = be32_to_cpu(arg_ch->rs_handle);
262 		goto out;
263 	}
264 
265 out:
266 	dprintk("svcrdma: Send With Invalidate rkey=%08x\n", inv_rkey);
267 	return inv_rkey;
268 }
269 
270 /* Assumptions:
271  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
272  */
273 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
274 		      u32 rmr, u64 to,
275 		      u32 xdr_off, int write_len,
276 		      struct svc_rdma_req_map *vec)
277 {
278 	struct ib_rdma_wr write_wr;
279 	struct ib_sge *sge;
280 	int xdr_sge_no;
281 	int sge_no;
282 	int sge_bytes;
283 	int sge_off;
284 	int bc;
285 	struct svc_rdma_op_ctxt *ctxt;
286 
287 	if (vec->count > RPCSVC_MAXPAGES) {
288 		pr_err("svcrdma: Too many pages (%lu)\n", vec->count);
289 		return -EIO;
290 	}
291 
292 	dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
293 		"write_len=%d, vec->sge=%p, vec->count=%lu\n",
294 		rmr, (unsigned long long)to, xdr_off,
295 		write_len, vec->sge, vec->count);
296 
297 	ctxt = svc_rdma_get_context(xprt);
298 	ctxt->direction = DMA_TO_DEVICE;
299 	sge = ctxt->sge;
300 
301 	/* Find the SGE associated with xdr_off */
302 	for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
303 	     xdr_sge_no++) {
304 		if (vec->sge[xdr_sge_no].iov_len > bc)
305 			break;
306 		bc -= vec->sge[xdr_sge_no].iov_len;
307 	}
308 
309 	sge_off = bc;
310 	bc = write_len;
311 	sge_no = 0;
312 
313 	/* Copy the remaining SGE */
314 	while (bc != 0) {
315 		sge_bytes = min_t(size_t,
316 			  bc, vec->sge[xdr_sge_no].iov_len-sge_off);
317 		sge[sge_no].length = sge_bytes;
318 		sge[sge_no].addr =
319 			dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
320 				    sge_bytes, DMA_TO_DEVICE);
321 		xdr_off += sge_bytes;
322 		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
323 					 sge[sge_no].addr))
324 			goto err;
325 		svc_rdma_count_mappings(xprt, ctxt);
326 		sge[sge_no].lkey = xprt->sc_pd->local_dma_lkey;
327 		ctxt->count++;
328 		sge_off = 0;
329 		sge_no++;
330 		xdr_sge_no++;
331 		if (xdr_sge_no > vec->count) {
332 			pr_err("svcrdma: Too many sges (%d)\n", xdr_sge_no);
333 			goto err;
334 		}
335 		bc -= sge_bytes;
336 		if (sge_no == xprt->sc_max_sge)
337 			break;
338 	}
339 
340 	/* Prepare WRITE WR */
341 	memset(&write_wr, 0, sizeof write_wr);
342 	ctxt->cqe.done = svc_rdma_wc_write;
343 	write_wr.wr.wr_cqe = &ctxt->cqe;
344 	write_wr.wr.sg_list = &sge[0];
345 	write_wr.wr.num_sge = sge_no;
346 	write_wr.wr.opcode = IB_WR_RDMA_WRITE;
347 	write_wr.wr.send_flags = IB_SEND_SIGNALED;
348 	write_wr.rkey = rmr;
349 	write_wr.remote_addr = to;
350 
351 	/* Post It */
352 	atomic_inc(&rdma_stat_write);
353 	if (svc_rdma_send(xprt, &write_wr.wr))
354 		goto err;
355 	return write_len - bc;
356  err:
357 	svc_rdma_unmap_dma(ctxt);
358 	svc_rdma_put_context(ctxt, 0);
359 	return -EIO;
360 }
361 
362 noinline
363 static int send_write_chunks(struct svcxprt_rdma *xprt,
364 			     struct rpcrdma_write_array *wr_ary,
365 			     struct rpcrdma_msg *rdma_resp,
366 			     struct svc_rqst *rqstp,
367 			     struct svc_rdma_req_map *vec)
368 {
369 	u32 xfer_len = rqstp->rq_res.page_len;
370 	int write_len;
371 	u32 xdr_off;
372 	int chunk_off;
373 	int chunk_no;
374 	int nchunks;
375 	struct rpcrdma_write_array *res_ary;
376 	int ret;
377 
378 	res_ary = (struct rpcrdma_write_array *)
379 		&rdma_resp->rm_body.rm_chunks[1];
380 
381 	/* Write chunks start at the pagelist */
382 	nchunks = be32_to_cpu(wr_ary->wc_nchunks);
383 	for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
384 	     xfer_len && chunk_no < nchunks;
385 	     chunk_no++) {
386 		struct rpcrdma_segment *arg_ch;
387 		u64 rs_offset;
388 
389 		arg_ch = &wr_ary->wc_array[chunk_no].wc_target;
390 		write_len = min(xfer_len, be32_to_cpu(arg_ch->rs_length));
391 
392 		/* Prepare the response chunk given the length actually
393 		 * written */
394 		xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
395 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
396 						arg_ch->rs_handle,
397 						arg_ch->rs_offset,
398 						write_len);
399 		chunk_off = 0;
400 		while (write_len) {
401 			ret = send_write(xprt, rqstp,
402 					 be32_to_cpu(arg_ch->rs_handle),
403 					 rs_offset + chunk_off,
404 					 xdr_off,
405 					 write_len,
406 					 vec);
407 			if (ret <= 0)
408 				goto out_err;
409 			chunk_off += ret;
410 			xdr_off += ret;
411 			xfer_len -= ret;
412 			write_len -= ret;
413 		}
414 	}
415 	/* Update the req with the number of chunks actually used */
416 	svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
417 
418 	return rqstp->rq_res.page_len;
419 
420 out_err:
421 	pr_err("svcrdma: failed to send write chunks, rc=%d\n", ret);
422 	return -EIO;
423 }
424 
425 noinline
426 static int send_reply_chunks(struct svcxprt_rdma *xprt,
427 			     struct rpcrdma_write_array *rp_ary,
428 			     struct rpcrdma_msg *rdma_resp,
429 			     struct svc_rqst *rqstp,
430 			     struct svc_rdma_req_map *vec)
431 {
432 	u32 xfer_len = rqstp->rq_res.len;
433 	int write_len;
434 	u32 xdr_off;
435 	int chunk_no;
436 	int chunk_off;
437 	int nchunks;
438 	struct rpcrdma_segment *ch;
439 	struct rpcrdma_write_array *res_ary;
440 	int ret;
441 
442 	/* XXX: need to fix when reply lists occur with read-list and or
443 	 * write-list */
444 	res_ary = (struct rpcrdma_write_array *)
445 		&rdma_resp->rm_body.rm_chunks[2];
446 
447 	/* xdr offset starts at RPC message */
448 	nchunks = be32_to_cpu(rp_ary->wc_nchunks);
449 	for (xdr_off = 0, chunk_no = 0;
450 	     xfer_len && chunk_no < nchunks;
451 	     chunk_no++) {
452 		u64 rs_offset;
453 		ch = &rp_ary->wc_array[chunk_no].wc_target;
454 		write_len = min(xfer_len, be32_to_cpu(ch->rs_length));
455 
456 		/* Prepare the reply chunk given the length actually
457 		 * written */
458 		xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset);
459 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
460 						ch->rs_handle, ch->rs_offset,
461 						write_len);
462 		chunk_off = 0;
463 		while (write_len) {
464 			ret = send_write(xprt, rqstp,
465 					 be32_to_cpu(ch->rs_handle),
466 					 rs_offset + chunk_off,
467 					 xdr_off,
468 					 write_len,
469 					 vec);
470 			if (ret <= 0)
471 				goto out_err;
472 			chunk_off += ret;
473 			xdr_off += ret;
474 			xfer_len -= ret;
475 			write_len -= ret;
476 		}
477 	}
478 	/* Update the req with the number of chunks actually used */
479 	svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
480 
481 	return rqstp->rq_res.len;
482 
483 out_err:
484 	pr_err("svcrdma: failed to send reply chunks, rc=%d\n", ret);
485 	return -EIO;
486 }
487 
488 /* This function prepares the portion of the RPCRDMA message to be
489  * sent in the RDMA_SEND. This function is called after data sent via
490  * RDMA has already been transmitted. There are three cases:
491  * - The RPCRDMA header, RPC header, and payload are all sent in a
492  *   single RDMA_SEND. This is the "inline" case.
493  * - The RPCRDMA header and some portion of the RPC header and data
494  *   are sent via this RDMA_SEND and another portion of the data is
495  *   sent via RDMA.
496  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
497  *   header and data are all transmitted via RDMA.
498  * In all three cases, this function prepares the RPCRDMA header in
499  * sge[0], the 'type' parameter indicates the type to place in the
500  * RPCRDMA header, and the 'byte_count' field indicates how much of
501  * the XDR to include in this RDMA_SEND. NB: The offset of the payload
502  * to send is zero in the XDR.
503  */
504 static int send_reply(struct svcxprt_rdma *rdma,
505 		      struct svc_rqst *rqstp,
506 		      struct page *page,
507 		      struct rpcrdma_msg *rdma_resp,
508 		      struct svc_rdma_req_map *vec,
509 		      int byte_count,
510 		      u32 inv_rkey)
511 {
512 	struct svc_rdma_op_ctxt *ctxt;
513 	struct ib_send_wr send_wr;
514 	u32 xdr_off;
515 	int sge_no;
516 	int sge_bytes;
517 	int page_no;
518 	int pages;
519 	int ret = -EIO;
520 
521 	/* Prepare the context */
522 	ctxt = svc_rdma_get_context(rdma);
523 	ctxt->direction = DMA_TO_DEVICE;
524 	ctxt->pages[0] = page;
525 	ctxt->count = 1;
526 
527 	/* Prepare the SGE for the RPCRDMA Header */
528 	ctxt->sge[0].lkey = rdma->sc_pd->local_dma_lkey;
529 	ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
530 	ctxt->sge[0].addr =
531 	    ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
532 			    ctxt->sge[0].length, DMA_TO_DEVICE);
533 	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
534 		goto err;
535 	svc_rdma_count_mappings(rdma, ctxt);
536 
537 	ctxt->direction = DMA_TO_DEVICE;
538 
539 	/* Map the payload indicated by 'byte_count' */
540 	xdr_off = 0;
541 	for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
542 		sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
543 		byte_count -= sge_bytes;
544 		ctxt->sge[sge_no].addr =
545 			dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
546 				    sge_bytes, DMA_TO_DEVICE);
547 		xdr_off += sge_bytes;
548 		if (ib_dma_mapping_error(rdma->sc_cm_id->device,
549 					 ctxt->sge[sge_no].addr))
550 			goto err;
551 		svc_rdma_count_mappings(rdma, ctxt);
552 		ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
553 		ctxt->sge[sge_no].length = sge_bytes;
554 	}
555 	if (byte_count != 0) {
556 		pr_err("svcrdma: Could not map %d bytes\n", byte_count);
557 		goto err;
558 	}
559 
560 	/* Save all respages in the ctxt and remove them from the
561 	 * respages array. They are our pages until the I/O
562 	 * completes.
563 	 */
564 	pages = rqstp->rq_next_page - rqstp->rq_respages;
565 	for (page_no = 0; page_no < pages; page_no++) {
566 		ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
567 		ctxt->count++;
568 		rqstp->rq_respages[page_no] = NULL;
569 	}
570 	rqstp->rq_next_page = rqstp->rq_respages + 1;
571 
572 	if (sge_no > rdma->sc_max_sge) {
573 		pr_err("svcrdma: Too many sges (%d)\n", sge_no);
574 		goto err;
575 	}
576 	memset(&send_wr, 0, sizeof send_wr);
577 	ctxt->cqe.done = svc_rdma_wc_send;
578 	send_wr.wr_cqe = &ctxt->cqe;
579 	send_wr.sg_list = ctxt->sge;
580 	send_wr.num_sge = sge_no;
581 	if (inv_rkey) {
582 		send_wr.opcode = IB_WR_SEND_WITH_INV;
583 		send_wr.ex.invalidate_rkey = inv_rkey;
584 	} else
585 		send_wr.opcode = IB_WR_SEND;
586 	send_wr.send_flags =  IB_SEND_SIGNALED;
587 
588 	ret = svc_rdma_send(rdma, &send_wr);
589 	if (ret)
590 		goto err;
591 
592 	return 0;
593 
594  err:
595 	svc_rdma_unmap_dma(ctxt);
596 	svc_rdma_put_context(ctxt, 1);
597 	return ret;
598 }
599 
600 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
601 {
602 }
603 
604 int svc_rdma_sendto(struct svc_rqst *rqstp)
605 {
606 	struct svc_xprt *xprt = rqstp->rq_xprt;
607 	struct svcxprt_rdma *rdma =
608 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
609 	struct rpcrdma_msg *rdma_argp;
610 	struct rpcrdma_msg *rdma_resp;
611 	struct rpcrdma_write_array *wr_ary, *rp_ary;
612 	enum rpcrdma_proc reply_type;
613 	int ret;
614 	int inline_bytes;
615 	struct page *res_page;
616 	struct svc_rdma_req_map *vec;
617 	u32 inv_rkey;
618 
619 	dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
620 
621 	/* Get the RDMA request header. The receive logic always
622 	 * places this at the start of page 0.
623 	 */
624 	rdma_argp = page_address(rqstp->rq_pages[0]);
625 	wr_ary = svc_rdma_get_write_array(rdma_argp);
626 	rp_ary = svc_rdma_get_reply_array(rdma_argp, wr_ary);
627 
628 	inv_rkey = 0;
629 	if (rdma->sc_snd_w_inv)
630 		inv_rkey = svc_rdma_get_inv_rkey(rdma_argp, wr_ary, rp_ary);
631 
632 	/* Build an req vec for the XDR */
633 	vec = svc_rdma_get_req_map(rdma);
634 	ret = svc_rdma_map_xdr(rdma, &rqstp->rq_res, vec, wr_ary != NULL);
635 	if (ret)
636 		goto err0;
637 	inline_bytes = rqstp->rq_res.len;
638 
639 	/* Create the RDMA response header */
640 	ret = -ENOMEM;
641 	res_page = alloc_page(GFP_KERNEL);
642 	if (!res_page)
643 		goto err0;
644 	rdma_resp = page_address(res_page);
645 	if (rp_ary)
646 		reply_type = RDMA_NOMSG;
647 	else
648 		reply_type = RDMA_MSG;
649 	svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
650 					 rdma_resp, reply_type);
651 
652 	/* Send any write-chunk data and build resp write-list */
653 	if (wr_ary) {
654 		ret = send_write_chunks(rdma, wr_ary, rdma_resp, rqstp, vec);
655 		if (ret < 0)
656 			goto err1;
657 		inline_bytes -= ret + xdr_padsize(ret);
658 	}
659 
660 	/* Send any reply-list data and update resp reply-list */
661 	if (rp_ary) {
662 		ret = send_reply_chunks(rdma, rp_ary, rdma_resp, rqstp, vec);
663 		if (ret < 0)
664 			goto err1;
665 		inline_bytes -= ret;
666 	}
667 
668 	/* Post a fresh Receive buffer _before_ sending the reply */
669 	ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
670 	if (ret)
671 		goto err1;
672 
673 	ret = send_reply(rdma, rqstp, res_page, rdma_resp, vec,
674 			 inline_bytes, inv_rkey);
675 	if (ret < 0)
676 		goto err0;
677 
678 	svc_rdma_put_req_map(rdma, vec);
679 	dprintk("svcrdma: send_reply returns %d\n", ret);
680 	return ret;
681 
682  err1:
683 	put_page(res_page);
684  err0:
685 	svc_rdma_put_req_map(rdma, vec);
686 	pr_err("svcrdma: Could not send reply, err=%d. Closing transport.\n",
687 	       ret);
688 	set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
689 	return -ENOTCONN;
690 }
691 
692 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
693 			 int status)
694 {
695 	struct ib_send_wr err_wr;
696 	struct page *p;
697 	struct svc_rdma_op_ctxt *ctxt;
698 	enum rpcrdma_errcode err;
699 	__be32 *va;
700 	int length;
701 	int ret;
702 
703 	ret = svc_rdma_repost_recv(xprt, GFP_KERNEL);
704 	if (ret)
705 		return;
706 
707 	p = alloc_page(GFP_KERNEL);
708 	if (!p)
709 		return;
710 	va = page_address(p);
711 
712 	/* XDR encode an error reply */
713 	err = ERR_CHUNK;
714 	if (status == -EPROTONOSUPPORT)
715 		err = ERR_VERS;
716 	length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
717 
718 	ctxt = svc_rdma_get_context(xprt);
719 	ctxt->direction = DMA_TO_DEVICE;
720 	ctxt->count = 1;
721 	ctxt->pages[0] = p;
722 
723 	/* Prepare SGE for local address */
724 	ctxt->sge[0].lkey = xprt->sc_pd->local_dma_lkey;
725 	ctxt->sge[0].length = length;
726 	ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device,
727 					    p, 0, length, DMA_TO_DEVICE);
728 	if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) {
729 		dprintk("svcrdma: Error mapping buffer for protocol error\n");
730 		svc_rdma_put_context(ctxt, 1);
731 		return;
732 	}
733 	svc_rdma_count_mappings(xprt, ctxt);
734 
735 	/* Prepare SEND WR */
736 	memset(&err_wr, 0, sizeof(err_wr));
737 	ctxt->cqe.done = svc_rdma_wc_send;
738 	err_wr.wr_cqe = &ctxt->cqe;
739 	err_wr.sg_list = ctxt->sge;
740 	err_wr.num_sge = 1;
741 	err_wr.opcode = IB_WR_SEND;
742 	err_wr.send_flags = IB_SEND_SIGNALED;
743 
744 	/* Post It */
745 	ret = svc_rdma_send(xprt, &err_wr);
746 	if (ret) {
747 		dprintk("svcrdma: Error %d posting send for protocol error\n",
748 			ret);
749 		svc_rdma_unmap_dma(ctxt);
750 		svc_rdma_put_context(ctxt, 1);
751 	}
752 }
753