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 /*
54  * Replace the pages in the rq_argpages array with the pages from the SGE in
55  * the RDMA_RECV completion. The SGL should contain full pages up until the
56  * last one.
57  */
58 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
59 			       struct svc_rdma_op_ctxt *ctxt,
60 			       u32 byte_count)
61 {
62 	struct rpcrdma_msg *rmsgp;
63 	struct page *page;
64 	u32 bc;
65 	int sge_no;
66 
67 	/* Swap the page in the SGE with the page in argpages */
68 	page = ctxt->pages[0];
69 	put_page(rqstp->rq_pages[0]);
70 	rqstp->rq_pages[0] = page;
71 
72 	/* Set up the XDR head */
73 	rqstp->rq_arg.head[0].iov_base = page_address(page);
74 	rqstp->rq_arg.head[0].iov_len =
75 		min_t(size_t, byte_count, ctxt->sge[0].length);
76 	rqstp->rq_arg.len = byte_count;
77 	rqstp->rq_arg.buflen = byte_count;
78 
79 	/* Compute bytes past head in the SGL */
80 	bc = byte_count - rqstp->rq_arg.head[0].iov_len;
81 
82 	/* If data remains, store it in the pagelist */
83 	rqstp->rq_arg.page_len = bc;
84 	rqstp->rq_arg.page_base = 0;
85 
86 	/* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
87 	rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
88 	if (rmsgp->rm_type == rdma_nomsg)
89 		rqstp->rq_arg.pages = &rqstp->rq_pages[0];
90 	else
91 		rqstp->rq_arg.pages = &rqstp->rq_pages[1];
92 
93 	sge_no = 1;
94 	while (bc && sge_no < ctxt->count) {
95 		page = ctxt->pages[sge_no];
96 		put_page(rqstp->rq_pages[sge_no]);
97 		rqstp->rq_pages[sge_no] = page;
98 		bc -= min_t(u32, bc, ctxt->sge[sge_no].length);
99 		rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
100 		sge_no++;
101 	}
102 	rqstp->rq_respages = &rqstp->rq_pages[sge_no];
103 	rqstp->rq_next_page = rqstp->rq_respages + 1;
104 
105 	/* If not all pages were used from the SGL, free the remaining ones */
106 	bc = sge_no;
107 	while (sge_no < ctxt->count) {
108 		page = ctxt->pages[sge_no++];
109 		put_page(page);
110 	}
111 	ctxt->count = bc;
112 
113 	/* Set up tail */
114 	rqstp->rq_arg.tail[0].iov_base = NULL;
115 	rqstp->rq_arg.tail[0].iov_len = 0;
116 }
117 
118 /* Issue an RDMA_READ using the local lkey to map the data sink */
119 int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
120 			struct svc_rqst *rqstp,
121 			struct svc_rdma_op_ctxt *head,
122 			int *page_no,
123 			u32 *page_offset,
124 			u32 rs_handle,
125 			u32 rs_length,
126 			u64 rs_offset,
127 			bool last)
128 {
129 	struct ib_rdma_wr read_wr;
130 	int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
131 	struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
132 	int ret, read, pno;
133 	u32 pg_off = *page_offset;
134 	u32 pg_no = *page_no;
135 
136 	ctxt->direction = DMA_FROM_DEVICE;
137 	ctxt->read_hdr = head;
138 	pages_needed = min_t(int, pages_needed, xprt->sc_max_sge_rd);
139 	read = min_t(int, (pages_needed << PAGE_SHIFT) - *page_offset,
140 		     rs_length);
141 
142 	for (pno = 0; pno < pages_needed; pno++) {
143 		int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
144 
145 		head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
146 		head->arg.page_len += len;
147 		head->arg.len += len;
148 		if (!pg_off)
149 			head->count++;
150 		rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
151 		rqstp->rq_next_page = rqstp->rq_respages + 1;
152 		ctxt->sge[pno].addr =
153 			ib_dma_map_page(xprt->sc_cm_id->device,
154 					head->arg.pages[pg_no], pg_off,
155 					PAGE_SIZE - pg_off,
156 					DMA_FROM_DEVICE);
157 		ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
158 					   ctxt->sge[pno].addr);
159 		if (ret)
160 			goto err;
161 		atomic_inc(&xprt->sc_dma_used);
162 
163 		/* The lkey here is either a local dma lkey or a dma_mr lkey */
164 		ctxt->sge[pno].lkey = xprt->sc_dma_lkey;
165 		ctxt->sge[pno].length = len;
166 		ctxt->count++;
167 
168 		/* adjust offset and wrap to next page if needed */
169 		pg_off += len;
170 		if (pg_off == PAGE_SIZE) {
171 			pg_off = 0;
172 			pg_no++;
173 		}
174 		rs_length -= len;
175 	}
176 
177 	if (last && rs_length == 0)
178 		set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
179 	else
180 		clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
181 
182 	memset(&read_wr, 0, sizeof(read_wr));
183 	read_wr.wr.wr_id = (unsigned long)ctxt;
184 	read_wr.wr.opcode = IB_WR_RDMA_READ;
185 	ctxt->wr_op = read_wr.wr.opcode;
186 	read_wr.wr.send_flags = IB_SEND_SIGNALED;
187 	read_wr.rkey = rs_handle;
188 	read_wr.remote_addr = rs_offset;
189 	read_wr.wr.sg_list = ctxt->sge;
190 	read_wr.wr.num_sge = pages_needed;
191 
192 	ret = svc_rdma_send(xprt, &read_wr.wr);
193 	if (ret) {
194 		pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
195 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
196 		goto err;
197 	}
198 
199 	/* return current location in page array */
200 	*page_no = pg_no;
201 	*page_offset = pg_off;
202 	ret = read;
203 	atomic_inc(&rdma_stat_read);
204 	return ret;
205  err:
206 	svc_rdma_unmap_dma(ctxt);
207 	svc_rdma_put_context(ctxt, 0);
208 	return ret;
209 }
210 
211 /* Issue an RDMA_READ using an FRMR to map the data sink */
212 int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
213 			 struct svc_rqst *rqstp,
214 			 struct svc_rdma_op_ctxt *head,
215 			 int *page_no,
216 			 u32 *page_offset,
217 			 u32 rs_handle,
218 			 u32 rs_length,
219 			 u64 rs_offset,
220 			 bool last)
221 {
222 	struct ib_rdma_wr read_wr;
223 	struct ib_send_wr inv_wr;
224 	struct ib_reg_wr reg_wr;
225 	u8 key;
226 	int nents = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
227 	struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
228 	struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt);
229 	int ret, read, pno, dma_nents, n;
230 	u32 pg_off = *page_offset;
231 	u32 pg_no = *page_no;
232 
233 	if (IS_ERR(frmr))
234 		return -ENOMEM;
235 
236 	ctxt->direction = DMA_FROM_DEVICE;
237 	ctxt->frmr = frmr;
238 	nents = min_t(unsigned int, nents, xprt->sc_frmr_pg_list_len);
239 	read = min_t(int, (nents << PAGE_SHIFT) - *page_offset, rs_length);
240 
241 	frmr->direction = DMA_FROM_DEVICE;
242 	frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
243 	frmr->sg_nents = nents;
244 
245 	for (pno = 0; pno < nents; pno++) {
246 		int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
247 
248 		head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
249 		head->arg.page_len += len;
250 		head->arg.len += len;
251 		if (!pg_off)
252 			head->count++;
253 
254 		sg_set_page(&frmr->sg[pno], rqstp->rq_arg.pages[pg_no],
255 			    len, pg_off);
256 
257 		rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
258 		rqstp->rq_next_page = rqstp->rq_respages + 1;
259 
260 		/* adjust offset and wrap to next page if needed */
261 		pg_off += len;
262 		if (pg_off == PAGE_SIZE) {
263 			pg_off = 0;
264 			pg_no++;
265 		}
266 		rs_length -= len;
267 	}
268 
269 	if (last && rs_length == 0)
270 		set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
271 	else
272 		clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
273 
274 	dma_nents = ib_dma_map_sg(xprt->sc_cm_id->device,
275 				  frmr->sg, frmr->sg_nents,
276 				  frmr->direction);
277 	if (!dma_nents) {
278 		pr_err("svcrdma: failed to dma map sg %p\n",
279 		       frmr->sg);
280 		return -ENOMEM;
281 	}
282 	atomic_inc(&xprt->sc_dma_used);
283 
284 	n = ib_map_mr_sg(frmr->mr, frmr->sg, frmr->sg_nents, PAGE_SIZE);
285 	if (unlikely(n != frmr->sg_nents)) {
286 		pr_err("svcrdma: failed to map mr %p (%d/%d elements)\n",
287 		       frmr->mr, n, frmr->sg_nents);
288 		return n < 0 ? n : -EINVAL;
289 	}
290 
291 	/* Bump the key */
292 	key = (u8)(frmr->mr->lkey & 0x000000FF);
293 	ib_update_fast_reg_key(frmr->mr, ++key);
294 
295 	ctxt->sge[0].addr = frmr->mr->iova;
296 	ctxt->sge[0].lkey = frmr->mr->lkey;
297 	ctxt->sge[0].length = frmr->mr->length;
298 	ctxt->count = 1;
299 	ctxt->read_hdr = head;
300 
301 	/* Prepare REG WR */
302 	reg_wr.wr.opcode = IB_WR_REG_MR;
303 	reg_wr.wr.wr_id = 0;
304 	reg_wr.wr.send_flags = IB_SEND_SIGNALED;
305 	reg_wr.wr.num_sge = 0;
306 	reg_wr.mr = frmr->mr;
307 	reg_wr.key = frmr->mr->lkey;
308 	reg_wr.access = frmr->access_flags;
309 	reg_wr.wr.next = &read_wr.wr;
310 
311 	/* Prepare RDMA_READ */
312 	memset(&read_wr, 0, sizeof(read_wr));
313 	read_wr.wr.send_flags = IB_SEND_SIGNALED;
314 	read_wr.rkey = rs_handle;
315 	read_wr.remote_addr = rs_offset;
316 	read_wr.wr.sg_list = ctxt->sge;
317 	read_wr.wr.num_sge = 1;
318 	if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_READ_W_INV) {
319 		read_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV;
320 		read_wr.wr.wr_id = (unsigned long)ctxt;
321 		read_wr.wr.ex.invalidate_rkey = ctxt->frmr->mr->lkey;
322 	} else {
323 		read_wr.wr.opcode = IB_WR_RDMA_READ;
324 		read_wr.wr.next = &inv_wr;
325 		/* Prepare invalidate */
326 		memset(&inv_wr, 0, sizeof(inv_wr));
327 		inv_wr.wr_id = (unsigned long)ctxt;
328 		inv_wr.opcode = IB_WR_LOCAL_INV;
329 		inv_wr.send_flags = IB_SEND_SIGNALED | IB_SEND_FENCE;
330 		inv_wr.ex.invalidate_rkey = frmr->mr->lkey;
331 	}
332 	ctxt->wr_op = read_wr.wr.opcode;
333 
334 	/* Post the chain */
335 	ret = svc_rdma_send(xprt, &reg_wr.wr);
336 	if (ret) {
337 		pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
338 		set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
339 		goto err;
340 	}
341 
342 	/* return current location in page array */
343 	*page_no = pg_no;
344 	*page_offset = pg_off;
345 	ret = read;
346 	atomic_inc(&rdma_stat_read);
347 	return ret;
348  err:
349 	ib_dma_unmap_sg(xprt->sc_cm_id->device,
350 			frmr->sg, frmr->sg_nents, frmr->direction);
351 	svc_rdma_put_context(ctxt, 0);
352 	svc_rdma_put_frmr(xprt, frmr);
353 	return ret;
354 }
355 
356 static unsigned int
357 rdma_rcl_chunk_count(struct rpcrdma_read_chunk *ch)
358 {
359 	unsigned int count;
360 
361 	for (count = 0; ch->rc_discrim != xdr_zero; ch++)
362 		count++;
363 	return count;
364 }
365 
366 /* If there was additional inline content, append it to the end of arg.pages.
367  * Tail copy has to be done after the reader function has determined how many
368  * pages are needed for RDMA READ.
369  */
370 static int
371 rdma_copy_tail(struct svc_rqst *rqstp, struct svc_rdma_op_ctxt *head,
372 	       u32 position, u32 byte_count, u32 page_offset, int page_no)
373 {
374 	char *srcp, *destp;
375 	int ret;
376 
377 	ret = 0;
378 	srcp = head->arg.head[0].iov_base + position;
379 	byte_count = head->arg.head[0].iov_len - position;
380 	if (byte_count > PAGE_SIZE) {
381 		dprintk("svcrdma: large tail unsupported\n");
382 		return 0;
383 	}
384 
385 	/* Fit as much of the tail on the current page as possible */
386 	if (page_offset != PAGE_SIZE) {
387 		destp = page_address(rqstp->rq_arg.pages[page_no]);
388 		destp += page_offset;
389 		while (byte_count--) {
390 			*destp++ = *srcp++;
391 			page_offset++;
392 			if (page_offset == PAGE_SIZE && byte_count)
393 				goto more;
394 		}
395 		goto done;
396 	}
397 
398 more:
399 	/* Fit the rest on the next page */
400 	page_no++;
401 	destp = page_address(rqstp->rq_arg.pages[page_no]);
402 	while (byte_count--)
403 		*destp++ = *srcp++;
404 
405 	rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
406 	rqstp->rq_next_page = rqstp->rq_respages + 1;
407 
408 done:
409 	byte_count = head->arg.head[0].iov_len - position;
410 	head->arg.page_len += byte_count;
411 	head->arg.len += byte_count;
412 	head->arg.buflen += byte_count;
413 	return 1;
414 }
415 
416 static int rdma_read_chunks(struct svcxprt_rdma *xprt,
417 			    struct rpcrdma_msg *rmsgp,
418 			    struct svc_rqst *rqstp,
419 			    struct svc_rdma_op_ctxt *head)
420 {
421 	int page_no, ret;
422 	struct rpcrdma_read_chunk *ch;
423 	u32 handle, page_offset, byte_count;
424 	u32 position;
425 	u64 rs_offset;
426 	bool last;
427 
428 	/* If no read list is present, return 0 */
429 	ch = svc_rdma_get_read_chunk(rmsgp);
430 	if (!ch)
431 		return 0;
432 
433 	if (rdma_rcl_chunk_count(ch) > RPCSVC_MAXPAGES)
434 		return -EINVAL;
435 
436 	/* The request is completed when the RDMA_READs complete. The
437 	 * head context keeps all the pages that comprise the
438 	 * request.
439 	 */
440 	head->arg.head[0] = rqstp->rq_arg.head[0];
441 	head->arg.tail[0] = rqstp->rq_arg.tail[0];
442 	head->hdr_count = head->count;
443 	head->arg.page_base = 0;
444 	head->arg.page_len = 0;
445 	head->arg.len = rqstp->rq_arg.len;
446 	head->arg.buflen = rqstp->rq_arg.buflen;
447 
448 	ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
449 	position = be32_to_cpu(ch->rc_position);
450 
451 	/* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
452 	if (position == 0) {
453 		head->arg.pages = &head->pages[0];
454 		page_offset = head->byte_len;
455 	} else {
456 		head->arg.pages = &head->pages[head->count];
457 		page_offset = 0;
458 	}
459 
460 	ret = 0;
461 	page_no = 0;
462 	for (; ch->rc_discrim != xdr_zero; ch++) {
463 		if (be32_to_cpu(ch->rc_position) != position)
464 			goto err;
465 
466 		handle = be32_to_cpu(ch->rc_target.rs_handle),
467 		byte_count = be32_to_cpu(ch->rc_target.rs_length);
468 		xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset,
469 				 &rs_offset);
470 
471 		while (byte_count > 0) {
472 			last = (ch + 1)->rc_discrim == xdr_zero;
473 			ret = xprt->sc_reader(xprt, rqstp, head,
474 					      &page_no, &page_offset,
475 					      handle, byte_count,
476 					      rs_offset, last);
477 			if (ret < 0)
478 				goto err;
479 			byte_count -= ret;
480 			rs_offset += ret;
481 			head->arg.buflen += ret;
482 		}
483 	}
484 
485 	/* Read list may need XDR round-up (see RFC 5666, s. 3.7) */
486 	if (page_offset & 3) {
487 		u32 pad = 4 - (page_offset & 3);
488 
489 		head->arg.page_len += pad;
490 		head->arg.len += pad;
491 		head->arg.buflen += pad;
492 		page_offset += pad;
493 	}
494 
495 	ret = 1;
496 	if (position && position < head->arg.head[0].iov_len)
497 		ret = rdma_copy_tail(rqstp, head, position,
498 				     byte_count, page_offset, page_no);
499 	head->arg.head[0].iov_len = position;
500 	head->position = position;
501 
502  err:
503 	/* Detach arg pages. svc_recv will replenish them */
504 	for (page_no = 0;
505 	     &rqstp->rq_pages[page_no] < rqstp->rq_respages; page_no++)
506 		rqstp->rq_pages[page_no] = NULL;
507 
508 	return ret;
509 }
510 
511 static int rdma_read_complete(struct svc_rqst *rqstp,
512 			      struct svc_rdma_op_ctxt *head)
513 {
514 	int page_no;
515 	int ret;
516 
517 	/* Copy RPC pages */
518 	for (page_no = 0; page_no < head->count; page_no++) {
519 		put_page(rqstp->rq_pages[page_no]);
520 		rqstp->rq_pages[page_no] = head->pages[page_no];
521 	}
522 
523 	/* Adjustments made for RDMA_NOMSG type requests */
524 	if (head->position == 0) {
525 		if (head->arg.len <= head->sge[0].length) {
526 			head->arg.head[0].iov_len = head->arg.len -
527 							head->byte_len;
528 			head->arg.page_len = 0;
529 		} else {
530 			head->arg.head[0].iov_len = head->sge[0].length -
531 								head->byte_len;
532 			head->arg.page_len = head->arg.len -
533 						head->sge[0].length;
534 		}
535 	}
536 
537 	/* Point rq_arg.pages past header */
538 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
539 	rqstp->rq_arg.page_len = head->arg.page_len;
540 	rqstp->rq_arg.page_base = head->arg.page_base;
541 
542 	/* rq_respages starts after the last arg page */
543 	rqstp->rq_respages = &rqstp->rq_pages[page_no];
544 	rqstp->rq_next_page = rqstp->rq_respages + 1;
545 
546 	/* Rebuild rq_arg head and tail. */
547 	rqstp->rq_arg.head[0] = head->arg.head[0];
548 	rqstp->rq_arg.tail[0] = head->arg.tail[0];
549 	rqstp->rq_arg.len = head->arg.len;
550 	rqstp->rq_arg.buflen = head->arg.buflen;
551 
552 	/* Free the context */
553 	svc_rdma_put_context(head, 0);
554 
555 	/* XXX: What should this be? */
556 	rqstp->rq_prot = IPPROTO_MAX;
557 	svc_xprt_copy_addrs(rqstp, rqstp->rq_xprt);
558 
559 	ret = rqstp->rq_arg.head[0].iov_len
560 		+ rqstp->rq_arg.page_len
561 		+ rqstp->rq_arg.tail[0].iov_len;
562 	dprintk("svcrdma: deferred read ret=%d, rq_arg.len=%u, "
563 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zu\n",
564 		ret, rqstp->rq_arg.len,	rqstp->rq_arg.head[0].iov_base,
565 		rqstp->rq_arg.head[0].iov_len);
566 
567 	return ret;
568 }
569 
570 /*
571  * Set up the rqstp thread context to point to the RQ buffer. If
572  * necessary, pull additional data from the client with an RDMA_READ
573  * request.
574  */
575 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
576 {
577 	struct svc_xprt *xprt = rqstp->rq_xprt;
578 	struct svcxprt_rdma *rdma_xprt =
579 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
580 	struct svc_rdma_op_ctxt *ctxt = NULL;
581 	struct rpcrdma_msg *rmsgp;
582 	int ret = 0;
583 	int len;
584 
585 	dprintk("svcrdma: rqstp=%p\n", rqstp);
586 
587 	spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
588 	if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
589 		ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
590 				  struct svc_rdma_op_ctxt,
591 				  dto_q);
592 		list_del_init(&ctxt->dto_q);
593 		spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
594 		return rdma_read_complete(rqstp, ctxt);
595 	} else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
596 		ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
597 				  struct svc_rdma_op_ctxt,
598 				  dto_q);
599 		list_del_init(&ctxt->dto_q);
600 	} else {
601 		atomic_inc(&rdma_stat_rq_starve);
602 		clear_bit(XPT_DATA, &xprt->xpt_flags);
603 		ctxt = NULL;
604 	}
605 	spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
606 	if (!ctxt) {
607 		/* This is the EAGAIN path. The svc_recv routine will
608 		 * return -EAGAIN, the nfsd thread will go to call into
609 		 * svc_recv again and we shouldn't be on the active
610 		 * transport list
611 		 */
612 		if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
613 			goto close_out;
614 
615 		goto out;
616 	}
617 	dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
618 		ctxt, rdma_xprt, rqstp, ctxt->wc_status);
619 	atomic_inc(&rdma_stat_recv);
620 
621 	/* Build up the XDR from the receive buffers. */
622 	rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
623 
624 	/* Decode the RDMA header. */
625 	len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
626 	rqstp->rq_xprt_hlen = len;
627 
628 	/* If the request is invalid, reply with an error */
629 	if (len < 0) {
630 		if (len == -ENOSYS)
631 			svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
632 		goto close_out;
633 	}
634 
635 	/* Read read-list data. */
636 	ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
637 	if (ret > 0) {
638 		/* read-list posted, defer until data received from client. */
639 		goto defer;
640 	} else if (ret < 0) {
641 		/* Post of read-list failed, free context. */
642 		svc_rdma_put_context(ctxt, 1);
643 		return 0;
644 	}
645 
646 	ret = rqstp->rq_arg.head[0].iov_len
647 		+ rqstp->rq_arg.page_len
648 		+ rqstp->rq_arg.tail[0].iov_len;
649 	svc_rdma_put_context(ctxt, 0);
650  out:
651 	dprintk("svcrdma: ret=%d, rq_arg.len=%u, "
652 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zd\n",
653 		ret, rqstp->rq_arg.len,
654 		rqstp->rq_arg.head[0].iov_base,
655 		rqstp->rq_arg.head[0].iov_len);
656 	rqstp->rq_prot = IPPROTO_MAX;
657 	svc_xprt_copy_addrs(rqstp, xprt);
658 	return ret;
659 
660  close_out:
661 	if (ctxt)
662 		svc_rdma_put_context(ctxt, 1);
663 	dprintk("svcrdma: transport %p is closing\n", xprt);
664 	/*
665 	 * Set the close bit and enqueue it. svc_recv will see the
666 	 * close bit and call svc_xprt_delete
667 	 */
668 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
669 defer:
670 	return 0;
671 }
672