1 /*
2  * Copyright (c) 2016 Oracle. All rights reserved.
3  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
4  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  *
41  * Author: Tom Tucker <tom@opengridcomputing.com>
42  */
43 
44 /* Operation
45  *
46  * The main entry point is svc_rdma_sendto. This is called by the
47  * RPC server when an RPC Reply is ready to be transmitted to a client.
48  *
49  * The passed-in svc_rqst contains a struct xdr_buf which holds an
50  * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
51  * transport header, post all Write WRs needed for this Reply, then post
52  * a Send WR conveying the transport header and the RPC message itself to
53  * the client.
54  *
55  * svc_rdma_sendto must fully transmit the Reply before returning, as
56  * the svc_rqst will be recycled as soon as sendto returns. Remaining
57  * resources referred to by the svc_rqst are also recycled at that time.
58  * Therefore any resources that must remain longer must be detached
59  * from the svc_rqst and released later.
60  *
61  * Page Management
62  *
63  * The I/O that performs Reply transmission is asynchronous, and may
64  * complete well after sendto returns. Thus pages under I/O must be
65  * removed from the svc_rqst before sendto returns.
66  *
67  * The logic here depends on Send Queue and completion ordering. Since
68  * the Send WR is always posted last, it will always complete last. Thus
69  * when it completes, it is guaranteed that all previous Write WRs have
70  * also completed.
71  *
72  * Write WRs are constructed and posted. Each Write segment gets its own
73  * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
74  * DMA-unmap the pages under I/O for that Write segment. The Write
75  * completion handler does not release any pages.
76  *
77  * When the Send WR is constructed, it also gets its own svc_rdma_op_ctxt.
78  * The ownership of all of the Reply's pages are transferred into that
79  * ctxt, the Send WR is posted, and sendto returns.
80  *
81  * The svc_rdma_op_ctxt is presented when the Send WR completes. The
82  * Send completion handler finally releases the Reply's pages.
83  *
84  * This mechanism also assumes that completions on the transport's Send
85  * Completion Queue do not run in parallel. Otherwise a Write completion
86  * and Send completion running at the same time could release pages that
87  * are still DMA-mapped.
88  *
89  * Error Handling
90  *
91  * - If the Send WR is posted successfully, it will either complete
92  *   successfully, or get flushed. Either way, the Send completion
93  *   handler releases the Reply's pages.
94  * - If the Send WR cannot be not posted, the forward path releases
95  *   the Reply's pages.
96  *
97  * This handles the case, without the use of page reference counting,
98  * where two different Write segments send portions of the same page.
99  */
100 
101 #include <linux/sunrpc/debug.h>
102 #include <linux/sunrpc/rpc_rdma.h>
103 #include <linux/spinlock.h>
104 #include <asm/unaligned.h>
105 #include <rdma/ib_verbs.h>
106 #include <rdma/rdma_cm.h>
107 #include <linux/sunrpc/svc_rdma.h>
108 
109 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
110 
111 static u32 xdr_padsize(u32 len)
112 {
113 	return (len & 3) ? (4 - (len & 3)) : 0;
114 }
115 
116 /* Returns length of transport header, in bytes.
117  */
118 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
119 {
120 	unsigned int nsegs;
121 	__be32 *p;
122 
123 	p = rdma_resp;
124 
125 	/* RPC-over-RDMA V1 replies never have a Read list. */
126 	p += rpcrdma_fixed_maxsz + 1;
127 
128 	/* Skip Write list. */
129 	while (*p++ != xdr_zero) {
130 		nsegs = be32_to_cpup(p++);
131 		p += nsegs * rpcrdma_segment_maxsz;
132 	}
133 
134 	/* Skip Reply chunk. */
135 	if (*p++ != xdr_zero) {
136 		nsegs = be32_to_cpup(p++);
137 		p += nsegs * rpcrdma_segment_maxsz;
138 	}
139 
140 	return (unsigned long)p - (unsigned long)rdma_resp;
141 }
142 
143 /* One Write chunk is copied from Call transport header to Reply
144  * transport header. Each segment's length field is updated to
145  * reflect number of bytes consumed in the segment.
146  *
147  * Returns number of segments in this chunk.
148  */
149 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
150 					   unsigned int remaining)
151 {
152 	unsigned int i, nsegs;
153 	u32 seg_len;
154 
155 	/* Write list discriminator */
156 	*dst++ = *src++;
157 
158 	/* number of segments in this chunk */
159 	nsegs = be32_to_cpup(src);
160 	*dst++ = *src++;
161 
162 	for (i = nsegs; i; i--) {
163 		/* segment's RDMA handle */
164 		*dst++ = *src++;
165 
166 		/* bytes returned in this segment */
167 		seg_len = be32_to_cpu(*src);
168 		if (remaining >= seg_len) {
169 			/* entire segment was consumed */
170 			*dst = *src;
171 			remaining -= seg_len;
172 		} else {
173 			/* segment only partly filled */
174 			*dst = cpu_to_be32(remaining);
175 			remaining = 0;
176 		}
177 		dst++; src++;
178 
179 		/* segment's RDMA offset */
180 		*dst++ = *src++;
181 		*dst++ = *src++;
182 	}
183 
184 	return nsegs;
185 }
186 
187 /* The client provided a Write list in the Call message. Fill in
188  * the segments in the first Write chunk in the Reply's transport
189  * header with the number of bytes consumed in each segment.
190  * Remaining chunks are returned unused.
191  *
192  * Assumptions:
193  *  - Client has provided only one Write chunk
194  */
195 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
196 					   unsigned int consumed)
197 {
198 	unsigned int nsegs;
199 	__be32 *p, *q;
200 
201 	/* RPC-over-RDMA V1 replies never have a Read list. */
202 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
203 
204 	q = wr_ch;
205 	while (*q != xdr_zero) {
206 		nsegs = xdr_encode_write_chunk(p, q, consumed);
207 		q += 2 + nsegs * rpcrdma_segment_maxsz;
208 		p += 2 + nsegs * rpcrdma_segment_maxsz;
209 		consumed = 0;
210 	}
211 
212 	/* Terminate Write list */
213 	*p++ = xdr_zero;
214 
215 	/* Reply chunk discriminator; may be replaced later */
216 	*p = xdr_zero;
217 }
218 
219 /* The client provided a Reply chunk in the Call message. Fill in
220  * the segments in the Reply chunk in the Reply message with the
221  * number of bytes consumed in each segment.
222  *
223  * Assumptions:
224  * - Reply can always fit in the provided Reply chunk
225  */
226 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
227 					    unsigned int consumed)
228 {
229 	__be32 *p;
230 
231 	/* Find the Reply chunk in the Reply's xprt header.
232 	 * RPC-over-RDMA V1 replies never have a Read list.
233 	 */
234 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
235 
236 	/* Skip past Write list */
237 	while (*p++ != xdr_zero)
238 		p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
239 
240 	xdr_encode_write_chunk(p, rp_ch, consumed);
241 }
242 
243 /* Parse the RPC Call's transport header.
244  */
245 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
246 				      __be32 **write, __be32 **reply)
247 {
248 	__be32 *p;
249 
250 	p = rdma_argp + rpcrdma_fixed_maxsz;
251 
252 	/* Read list */
253 	while (*p++ != xdr_zero)
254 		p += 5;
255 
256 	/* Write list */
257 	if (*p != xdr_zero) {
258 		*write = p;
259 		while (*p++ != xdr_zero)
260 			p += 1 + be32_to_cpu(*p) * 4;
261 	} else {
262 		*write = NULL;
263 		p++;
264 	}
265 
266 	/* Reply chunk */
267 	if (*p != xdr_zero)
268 		*reply = p;
269 	else
270 		*reply = NULL;
271 }
272 
273 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
274  * Responder's choice: requester signals it can handle Send With
275  * Invalidate, and responder chooses one rkey to invalidate.
276  *
277  * Find a candidate rkey to invalidate when sending a reply.  Picks the
278  * first R_key it finds in the chunk lists.
279  *
280  * Returns zero if RPC's chunk lists are empty.
281  */
282 static u32 svc_rdma_get_inv_rkey(__be32 *rdma_argp,
283 				 __be32 *wr_lst, __be32 *rp_ch)
284 {
285 	__be32 *p;
286 
287 	p = rdma_argp + rpcrdma_fixed_maxsz;
288 	if (*p != xdr_zero)
289 		p += 2;
290 	else if (wr_lst && be32_to_cpup(wr_lst + 1))
291 		p = wr_lst + 2;
292 	else if (rp_ch && be32_to_cpup(rp_ch + 1))
293 		p = rp_ch + 2;
294 	else
295 		return 0;
296 	return be32_to_cpup(p);
297 }
298 
299 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
300  * is used during completion to DMA-unmap this memory, and
301  * it uses ib_dma_unmap_page() exclusively.
302  */
303 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
304 				struct svc_rdma_op_ctxt *ctxt,
305 				unsigned int sge_no,
306 				unsigned char *base,
307 				unsigned int len)
308 {
309 	unsigned long offset = (unsigned long)base & ~PAGE_MASK;
310 	struct ib_device *dev = rdma->sc_cm_id->device;
311 	dma_addr_t dma_addr;
312 
313 	dma_addr = ib_dma_map_page(dev, virt_to_page(base),
314 				   offset, len, DMA_TO_DEVICE);
315 	if (ib_dma_mapping_error(dev, dma_addr))
316 		return -EIO;
317 
318 	ctxt->sge[sge_no].addr = dma_addr;
319 	ctxt->sge[sge_no].length = len;
320 	ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
321 	svc_rdma_count_mappings(rdma, ctxt);
322 	return 0;
323 }
324 
325 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
326 				 struct svc_rdma_op_ctxt *ctxt,
327 				 unsigned int sge_no,
328 				 struct page *page,
329 				 unsigned int offset,
330 				 unsigned int len)
331 {
332 	struct ib_device *dev = rdma->sc_cm_id->device;
333 	dma_addr_t dma_addr;
334 
335 	dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
336 	if (ib_dma_mapping_error(dev, dma_addr))
337 		return -EIO;
338 
339 	ctxt->sge[sge_no].addr = dma_addr;
340 	ctxt->sge[sge_no].length = len;
341 	ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
342 	svc_rdma_count_mappings(rdma, ctxt);
343 	return 0;
344 }
345 
346 /**
347  * svc_rdma_map_reply_hdr - DMA map the transport header buffer
348  * @rdma: controlling transport
349  * @ctxt: op_ctxt for the Send WR
350  * @rdma_resp: buffer containing transport header
351  * @len: length of transport header
352  *
353  * Returns:
354  *	%0 if the header is DMA mapped,
355  *	%-EIO if DMA mapping failed.
356  */
357 int svc_rdma_map_reply_hdr(struct svcxprt_rdma *rdma,
358 			   struct svc_rdma_op_ctxt *ctxt,
359 			   __be32 *rdma_resp,
360 			   unsigned int len)
361 {
362 	ctxt->direction = DMA_TO_DEVICE;
363 	ctxt->pages[0] = virt_to_page(rdma_resp);
364 	ctxt->count = 1;
365 	return svc_rdma_dma_map_page(rdma, ctxt, 0, ctxt->pages[0], 0, len);
366 }
367 
368 /* Load the xdr_buf into the ctxt's sge array, and DMA map each
369  * element as it is added.
370  *
371  * Returns the number of sge elements loaded on success, or
372  * a negative errno on failure.
373  */
374 static int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
375 				  struct svc_rdma_op_ctxt *ctxt,
376 				  struct xdr_buf *xdr, __be32 *wr_lst)
377 {
378 	unsigned int len, sge_no, remaining, page_off;
379 	struct page **ppages;
380 	unsigned char *base;
381 	u32 xdr_pad;
382 	int ret;
383 
384 	sge_no = 1;
385 
386 	ret = svc_rdma_dma_map_buf(rdma, ctxt, sge_no++,
387 				   xdr->head[0].iov_base,
388 				   xdr->head[0].iov_len);
389 	if (ret < 0)
390 		return ret;
391 
392 	/* If a Write chunk is present, the xdr_buf's page list
393 	 * is not included inline. However the Upper Layer may
394 	 * have added XDR padding in the tail buffer, and that
395 	 * should not be included inline.
396 	 */
397 	if (wr_lst) {
398 		base = xdr->tail[0].iov_base;
399 		len = xdr->tail[0].iov_len;
400 		xdr_pad = xdr_padsize(xdr->page_len);
401 
402 		if (len && xdr_pad) {
403 			base += xdr_pad;
404 			len -= xdr_pad;
405 		}
406 
407 		goto tail;
408 	}
409 
410 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
411 	page_off = xdr->page_base & ~PAGE_MASK;
412 	remaining = xdr->page_len;
413 	while (remaining) {
414 		len = min_t(u32, PAGE_SIZE - page_off, remaining);
415 
416 		ret = svc_rdma_dma_map_page(rdma, ctxt, sge_no++,
417 					    *ppages++, page_off, len);
418 		if (ret < 0)
419 			return ret;
420 
421 		remaining -= len;
422 		page_off = 0;
423 	}
424 
425 	base = xdr->tail[0].iov_base;
426 	len = xdr->tail[0].iov_len;
427 tail:
428 	if (len) {
429 		ret = svc_rdma_dma_map_buf(rdma, ctxt, sge_no++, base, len);
430 		if (ret < 0)
431 			return ret;
432 	}
433 
434 	return sge_no - 1;
435 }
436 
437 /* The svc_rqst and all resources it owns are released as soon as
438  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
439  * so they are released by the Send completion handler.
440  */
441 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
442 				   struct svc_rdma_op_ctxt *ctxt)
443 {
444 	int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
445 
446 	ctxt->count += pages;
447 	for (i = 0; i < pages; i++) {
448 		ctxt->pages[i + 1] = rqstp->rq_respages[i];
449 		rqstp->rq_respages[i] = NULL;
450 	}
451 	rqstp->rq_next_page = rqstp->rq_respages + 1;
452 }
453 
454 /**
455  * svc_rdma_post_send_wr - Set up and post one Send Work Request
456  * @rdma: controlling transport
457  * @ctxt: op_ctxt for transmitting the Send WR
458  * @num_sge: number of SGEs to send
459  * @inv_rkey: R_key argument to Send With Invalidate, or zero
460  *
461  * Returns:
462  *	%0 if the Send* was posted successfully,
463  *	%-ENOTCONN if the connection was lost or dropped,
464  *	%-EINVAL if there was a problem with the Send we built,
465  *	%-ENOMEM if ib_post_send failed.
466  */
467 int svc_rdma_post_send_wr(struct svcxprt_rdma *rdma,
468 			  struct svc_rdma_op_ctxt *ctxt, int num_sge,
469 			  u32 inv_rkey)
470 {
471 	struct ib_send_wr *send_wr = &ctxt->send_wr;
472 
473 	dprintk("svcrdma: posting Send WR with %u sge(s)\n", num_sge);
474 
475 	send_wr->next = NULL;
476 	ctxt->cqe.done = svc_rdma_wc_send;
477 	send_wr->wr_cqe = &ctxt->cqe;
478 	send_wr->sg_list = ctxt->sge;
479 	send_wr->num_sge = num_sge;
480 	send_wr->send_flags = IB_SEND_SIGNALED;
481 	if (inv_rkey) {
482 		send_wr->opcode = IB_WR_SEND_WITH_INV;
483 		send_wr->ex.invalidate_rkey = inv_rkey;
484 	} else {
485 		send_wr->opcode = IB_WR_SEND;
486 	}
487 
488 	return svc_rdma_send(rdma, send_wr);
489 }
490 
491 /* Prepare the portion of the RPC Reply that will be transmitted
492  * via RDMA Send. The RPC-over-RDMA transport header is prepared
493  * in sge[0], and the RPC xdr_buf is prepared in following sges.
494  *
495  * Depending on whether a Write list or Reply chunk is present,
496  * the server may send all, a portion of, or none of the xdr_buf.
497  * In the latter case, only the transport header (sge[0]) is
498  * transmitted.
499  *
500  * RDMA Send is the last step of transmitting an RPC reply. Pages
501  * involved in the earlier RDMA Writes are here transferred out
502  * of the rqstp and into the ctxt's page array. These pages are
503  * DMA unmapped by each Write completion, but the subsequent Send
504  * completion finally releases these pages.
505  *
506  * Assumptions:
507  * - The Reply's transport header will never be larger than a page.
508  */
509 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
510 				   __be32 *rdma_argp, __be32 *rdma_resp,
511 				   struct svc_rqst *rqstp,
512 				   __be32 *wr_lst, __be32 *rp_ch)
513 {
514 	struct svc_rdma_op_ctxt *ctxt;
515 	u32 inv_rkey;
516 	int ret;
517 
518 	dprintk("svcrdma: sending %s reply: head=%zu, pagelen=%u, tail=%zu\n",
519 		(rp_ch ? "RDMA_NOMSG" : "RDMA_MSG"),
520 		rqstp->rq_res.head[0].iov_len,
521 		rqstp->rq_res.page_len,
522 		rqstp->rq_res.tail[0].iov_len);
523 
524 	ctxt = svc_rdma_get_context(rdma);
525 
526 	ret = svc_rdma_map_reply_hdr(rdma, ctxt, rdma_resp,
527 				     svc_rdma_reply_hdr_len(rdma_resp));
528 	if (ret < 0)
529 		goto err;
530 
531 	if (!rp_ch) {
532 		ret = svc_rdma_map_reply_msg(rdma, ctxt,
533 					     &rqstp->rq_res, wr_lst);
534 		if (ret < 0)
535 			goto err;
536 	}
537 
538 	svc_rdma_save_io_pages(rqstp, ctxt);
539 
540 	inv_rkey = 0;
541 	if (rdma->sc_snd_w_inv)
542 		inv_rkey = svc_rdma_get_inv_rkey(rdma_argp, wr_lst, rp_ch);
543 	ret = svc_rdma_post_send_wr(rdma, ctxt, 1 + ret, inv_rkey);
544 	if (ret)
545 		goto err;
546 
547 	return 0;
548 
549 err:
550 	pr_err("svcrdma: failed to post Send WR (%d)\n", ret);
551 	svc_rdma_unmap_dma(ctxt);
552 	svc_rdma_put_context(ctxt, 1);
553 	return ret;
554 }
555 
556 /* Given the client-provided Write and Reply chunks, the server was not
557  * able to form a complete reply. Return an RDMA_ERROR message so the
558  * client can retire this RPC transaction. As above, the Send completion
559  * routine releases payload pages that were part of a previous RDMA Write.
560  *
561  * Remote Invalidation is skipped for simplicity.
562  */
563 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
564 				   __be32 *rdma_resp, struct svc_rqst *rqstp)
565 {
566 	struct svc_rdma_op_ctxt *ctxt;
567 	__be32 *p;
568 	int ret;
569 
570 	ctxt = svc_rdma_get_context(rdma);
571 
572 	/* Replace the original transport header with an
573 	 * RDMA_ERROR response. XID etc are preserved.
574 	 */
575 	p = rdma_resp + 3;
576 	*p++ = rdma_error;
577 	*p   = err_chunk;
578 
579 	ret = svc_rdma_map_reply_hdr(rdma, ctxt, rdma_resp, 20);
580 	if (ret < 0)
581 		goto err;
582 
583 	svc_rdma_save_io_pages(rqstp, ctxt);
584 
585 	ret = svc_rdma_post_send_wr(rdma, ctxt, 1 + ret, 0);
586 	if (ret)
587 		goto err;
588 
589 	return 0;
590 
591 err:
592 	pr_err("svcrdma: failed to post Send WR (%d)\n", ret);
593 	svc_rdma_unmap_dma(ctxt);
594 	svc_rdma_put_context(ctxt, 1);
595 	return ret;
596 }
597 
598 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
599 {
600 }
601 
602 /**
603  * svc_rdma_sendto - Transmit an RPC reply
604  * @rqstp: processed RPC request, reply XDR already in ::rq_res
605  *
606  * Any resources still associated with @rqstp are released upon return.
607  * If no reply message was possible, the connection is closed.
608  *
609  * Returns:
610  *	%0 if an RPC reply has been successfully posted,
611  *	%-ENOMEM if a resource shortage occurred (connection is lost),
612  *	%-ENOTCONN if posting failed (connection is lost).
613  */
614 int svc_rdma_sendto(struct svc_rqst *rqstp)
615 {
616 	struct svc_xprt *xprt = rqstp->rq_xprt;
617 	struct svcxprt_rdma *rdma =
618 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
619 	__be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
620 	struct xdr_buf *xdr = &rqstp->rq_res;
621 	struct page *res_page;
622 	int ret;
623 
624 	/* Find the call's chunk lists to decide how to send the reply.
625 	 * Receive places the Call's xprt header at the start of page 0.
626 	 */
627 	rdma_argp = page_address(rqstp->rq_pages[0]);
628 	svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
629 
630 	dprintk("svcrdma: preparing response for XID 0x%08x\n",
631 		be32_to_cpup(rdma_argp));
632 
633 	/* Create the RDMA response header. xprt->xpt_mutex,
634 	 * acquired in svc_send(), serializes RPC replies. The
635 	 * code path below that inserts the credit grant value
636 	 * into each transport header runs only inside this
637 	 * critical section.
638 	 */
639 	ret = -ENOMEM;
640 	res_page = alloc_page(GFP_KERNEL);
641 	if (!res_page)
642 		goto err0;
643 	rdma_resp = page_address(res_page);
644 
645 	p = rdma_resp;
646 	*p++ = *rdma_argp;
647 	*p++ = *(rdma_argp + 1);
648 	*p++ = rdma->sc_fc_credits;
649 	*p++ = rp_ch ? rdma_nomsg : rdma_msg;
650 
651 	/* Start with empty chunks */
652 	*p++ = xdr_zero;
653 	*p++ = xdr_zero;
654 	*p   = xdr_zero;
655 
656 	if (wr_lst) {
657 		/* XXX: Presume the client sent only one Write chunk */
658 		ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
659 		if (ret < 0)
660 			goto err2;
661 		svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
662 	}
663 	if (rp_ch) {
664 		ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
665 		if (ret < 0)
666 			goto err2;
667 		svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
668 	}
669 
670 	ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
671 	if (ret)
672 		goto err1;
673 	ret = svc_rdma_send_reply_msg(rdma, rdma_argp, rdma_resp, rqstp,
674 				      wr_lst, rp_ch);
675 	if (ret < 0)
676 		goto err0;
677 	return 0;
678 
679  err2:
680 	if (ret != -E2BIG)
681 		goto err1;
682 
683 	ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
684 	if (ret)
685 		goto err1;
686 	ret = svc_rdma_send_error_msg(rdma, rdma_resp, rqstp);
687 	if (ret < 0)
688 		goto err0;
689 	return 0;
690 
691  err1:
692 	put_page(res_page);
693  err0:
694 	pr_err("svcrdma: Could not send reply, err=%d. Closing transport.\n",
695 	       ret);
696 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
697 	return -ENOTCONN;
698 }
699