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 /*
53  * Replace the pages in the rq_argpages array with the pages from the SGE in
54  * the RDMA_RECV completion. The SGL should contain full pages up until the
55  * last one.
56  */
57 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
58 			       struct svc_rdma_op_ctxt *ctxt,
59 			       u32 byte_count)
60 {
61 	struct page *page;
62 	u32 bc;
63 	int sge_no;
64 
65 	/* Swap the page in the SGE with the page in argpages */
66 	page = ctxt->pages[0];
67 	put_page(rqstp->rq_pages[0]);
68 	rqstp->rq_pages[0] = page;
69 
70 	/* Set up the XDR head */
71 	rqstp->rq_arg.head[0].iov_base = page_address(page);
72 	rqstp->rq_arg.head[0].iov_len = min(byte_count, ctxt->sge[0].length);
73 	rqstp->rq_arg.len = byte_count;
74 	rqstp->rq_arg.buflen = byte_count;
75 
76 	/* Compute bytes past head in the SGL */
77 	bc = byte_count - rqstp->rq_arg.head[0].iov_len;
78 
79 	/* If data remains, store it in the pagelist */
80 	rqstp->rq_arg.page_len = bc;
81 	rqstp->rq_arg.page_base = 0;
82 	rqstp->rq_arg.pages = &rqstp->rq_pages[1];
83 	sge_no = 1;
84 	while (bc && sge_no < ctxt->count) {
85 		page = ctxt->pages[sge_no];
86 		put_page(rqstp->rq_pages[sge_no]);
87 		rqstp->rq_pages[sge_no] = page;
88 		bc -= min(bc, ctxt->sge[sge_no].length);
89 		rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
90 		sge_no++;
91 	}
92 	rqstp->rq_respages = &rqstp->rq_pages[sge_no];
93 
94 	/* We should never run out of SGE because the limit is defined to
95 	 * support the max allowed RPC data length
96 	 */
97 	BUG_ON(bc && (sge_no == ctxt->count));
98 	BUG_ON((rqstp->rq_arg.head[0].iov_len + rqstp->rq_arg.page_len)
99 	       != byte_count);
100 	BUG_ON(rqstp->rq_arg.len != byte_count);
101 
102 	/* If not all pages were used from the SGL, free the remaining ones */
103 	bc = sge_no;
104 	while (sge_no < ctxt->count) {
105 		page = ctxt->pages[sge_no++];
106 		put_page(page);
107 	}
108 	ctxt->count = bc;
109 
110 	/* Set up tail */
111 	rqstp->rq_arg.tail[0].iov_base = NULL;
112 	rqstp->rq_arg.tail[0].iov_len = 0;
113 }
114 
115 /* Encode a read-chunk-list as an array of IB SGE
116  *
117  * Assumptions:
118  * - chunk[0]->position points to pages[0] at an offset of 0
119  * - pages[] is not physically or virtually contiguous and consists of
120  *   PAGE_SIZE elements.
121  *
122  * Output:
123  * - sge array pointing into pages[] array.
124  * - chunk_sge array specifying sge index and count for each
125  *   chunk in the read list
126  *
127  */
128 static int map_read_chunks(struct svcxprt_rdma *xprt,
129 			   struct svc_rqst *rqstp,
130 			   struct svc_rdma_op_ctxt *head,
131 			   struct rpcrdma_msg *rmsgp,
132 			   struct svc_rdma_req_map *rpl_map,
133 			   struct svc_rdma_req_map *chl_map,
134 			   int ch_count,
135 			   int byte_count)
136 {
137 	int sge_no;
138 	int sge_bytes;
139 	int page_off;
140 	int page_no;
141 	int ch_bytes;
142 	int ch_no;
143 	struct rpcrdma_read_chunk *ch;
144 
145 	sge_no = 0;
146 	page_no = 0;
147 	page_off = 0;
148 	ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
149 	ch_no = 0;
150 	ch_bytes = ch->rc_target.rs_length;
151 	head->arg.head[0] = rqstp->rq_arg.head[0];
152 	head->arg.tail[0] = rqstp->rq_arg.tail[0];
153 	head->arg.pages = &head->pages[head->count];
154 	head->hdr_count = head->count; /* save count of hdr pages */
155 	head->arg.page_base = 0;
156 	head->arg.page_len = ch_bytes;
157 	head->arg.len = rqstp->rq_arg.len + ch_bytes;
158 	head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
159 	head->count++;
160 	chl_map->ch[0].start = 0;
161 	while (byte_count) {
162 		rpl_map->sge[sge_no].iov_base =
163 			page_address(rqstp->rq_arg.pages[page_no]) + page_off;
164 		sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
165 		rpl_map->sge[sge_no].iov_len = sge_bytes;
166 		/*
167 		 * Don't bump head->count here because the same page
168 		 * may be used by multiple SGE.
169 		 */
170 		head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
171 		rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
172 
173 		byte_count -= sge_bytes;
174 		ch_bytes -= sge_bytes;
175 		sge_no++;
176 		/*
177 		 * If all bytes for this chunk have been mapped to an
178 		 * SGE, move to the next SGE
179 		 */
180 		if (ch_bytes == 0) {
181 			chl_map->ch[ch_no].count =
182 				sge_no - chl_map->ch[ch_no].start;
183 			ch_no++;
184 			ch++;
185 			chl_map->ch[ch_no].start = sge_no;
186 			ch_bytes = ch->rc_target.rs_length;
187 			/* If bytes remaining account for next chunk */
188 			if (byte_count) {
189 				head->arg.page_len += ch_bytes;
190 				head->arg.len += ch_bytes;
191 				head->arg.buflen += ch_bytes;
192 			}
193 		}
194 		/*
195 		 * If this SGE consumed all of the page, move to the
196 		 * next page
197 		 */
198 		if ((sge_bytes + page_off) == PAGE_SIZE) {
199 			page_no++;
200 			page_off = 0;
201 			/*
202 			 * If there are still bytes left to map, bump
203 			 * the page count
204 			 */
205 			if (byte_count)
206 				head->count++;
207 		} else
208 			page_off += sge_bytes;
209 	}
210 	BUG_ON(byte_count != 0);
211 	return sge_no;
212 }
213 
214 /* Map a read-chunk-list to an XDR and fast register the page-list.
215  *
216  * Assumptions:
217  * - chunk[0]	position points to pages[0] at an offset of 0
218  * - pages[]	will be made physically contiguous by creating a one-off memory
219  *		region using the fastreg verb.
220  * - byte_count is # of bytes in read-chunk-list
221  * - ch_count	is # of chunks in read-chunk-list
222  *
223  * Output:
224  * - sge array pointing into pages[] array.
225  * - chunk_sge array specifying sge index and count for each
226  *   chunk in the read list
227  */
228 static int fast_reg_read_chunks(struct svcxprt_rdma *xprt,
229 				struct svc_rqst *rqstp,
230 				struct svc_rdma_op_ctxt *head,
231 				struct rpcrdma_msg *rmsgp,
232 				struct svc_rdma_req_map *rpl_map,
233 				struct svc_rdma_req_map *chl_map,
234 				int ch_count,
235 				int byte_count)
236 {
237 	int page_no;
238 	int ch_no;
239 	u32 offset;
240 	struct rpcrdma_read_chunk *ch;
241 	struct svc_rdma_fastreg_mr *frmr;
242 	int ret = 0;
243 
244 	frmr = svc_rdma_get_frmr(xprt);
245 	if (IS_ERR(frmr))
246 		return -ENOMEM;
247 
248 	head->frmr = frmr;
249 	head->arg.head[0] = rqstp->rq_arg.head[0];
250 	head->arg.tail[0] = rqstp->rq_arg.tail[0];
251 	head->arg.pages = &head->pages[head->count];
252 	head->hdr_count = head->count; /* save count of hdr pages */
253 	head->arg.page_base = 0;
254 	head->arg.page_len = byte_count;
255 	head->arg.len = rqstp->rq_arg.len + byte_count;
256 	head->arg.buflen = rqstp->rq_arg.buflen + byte_count;
257 
258 	/* Fast register the page list */
259 	frmr->kva = page_address(rqstp->rq_arg.pages[0]);
260 	frmr->direction = DMA_FROM_DEVICE;
261 	frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
262 	frmr->map_len = byte_count;
263 	frmr->page_list_len = PAGE_ALIGN(byte_count) >> PAGE_SHIFT;
264 	for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
265 		frmr->page_list->page_list[page_no] =
266 			ib_dma_map_single(xprt->sc_cm_id->device,
267 					  page_address(rqstp->rq_arg.pages[page_no]),
268 					  PAGE_SIZE, DMA_FROM_DEVICE);
269 		if (ib_dma_mapping_error(xprt->sc_cm_id->device,
270 					 frmr->page_list->page_list[page_no]))
271 			goto fatal_err;
272 		atomic_inc(&xprt->sc_dma_used);
273 		head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
274 	}
275 	head->count += page_no;
276 
277 	/* rq_respages points one past arg pages */
278 	rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
279 
280 	/* Create the reply and chunk maps */
281 	offset = 0;
282 	ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
283 	for (ch_no = 0; ch_no < ch_count; ch_no++) {
284 		rpl_map->sge[ch_no].iov_base = frmr->kva + offset;
285 		rpl_map->sge[ch_no].iov_len = ch->rc_target.rs_length;
286 		chl_map->ch[ch_no].count = 1;
287 		chl_map->ch[ch_no].start = ch_no;
288 		offset += ch->rc_target.rs_length;
289 		ch++;
290 	}
291 
292 	ret = svc_rdma_fastreg(xprt, frmr);
293 	if (ret)
294 		goto fatal_err;
295 
296 	return ch_no;
297 
298  fatal_err:
299 	printk("svcrdma: error fast registering xdr for xprt %p", xprt);
300 	svc_rdma_put_frmr(xprt, frmr);
301 	return -EIO;
302 }
303 
304 static int rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
305 			     struct svc_rdma_op_ctxt *ctxt,
306 			     struct svc_rdma_fastreg_mr *frmr,
307 			     struct kvec *vec,
308 			     u64 *sgl_offset,
309 			     int count)
310 {
311 	int i;
312 
313 	ctxt->count = count;
314 	ctxt->direction = DMA_FROM_DEVICE;
315 	for (i = 0; i < count; i++) {
316 		ctxt->sge[i].length = 0; /* in case map fails */
317 		if (!frmr) {
318 			ctxt->sge[i].addr =
319 				ib_dma_map_single(xprt->sc_cm_id->device,
320 						  vec[i].iov_base,
321 						  vec[i].iov_len,
322 						  DMA_FROM_DEVICE);
323 			if (ib_dma_mapping_error(xprt->sc_cm_id->device,
324 						 ctxt->sge[i].addr))
325 				return -EINVAL;
326 			ctxt->sge[i].lkey = xprt->sc_dma_lkey;
327 			atomic_inc(&xprt->sc_dma_used);
328 		} else {
329 			ctxt->sge[i].addr = (unsigned long)vec[i].iov_base;
330 			ctxt->sge[i].lkey = frmr->mr->lkey;
331 		}
332 		ctxt->sge[i].length = vec[i].iov_len;
333 		*sgl_offset = *sgl_offset + vec[i].iov_len;
334 	}
335 	return 0;
336 }
337 
338 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
339 {
340 	if ((RDMA_TRANSPORT_IWARP ==
341 	     rdma_node_get_transport(xprt->sc_cm_id->
342 				     device->node_type))
343 	    && sge_count > 1)
344 		return 1;
345 	else
346 		return min_t(int, sge_count, xprt->sc_max_sge);
347 }
348 
349 /*
350  * Use RDMA_READ to read data from the advertised client buffer into the
351  * XDR stream starting at rq_arg.head[0].iov_base.
352  * Each chunk in the array
353  * contains the following fields:
354  * discrim      - '1', This isn't used for data placement
355  * position     - The xdr stream offset (the same for every chunk)
356  * handle       - RMR for client memory region
357  * length       - data transfer length
358  * offset       - 64 bit tagged offset in remote memory region
359  *
360  * On our side, we need to read into a pagelist. The first page immediately
361  * follows the RPC header.
362  *
363  * This function returns:
364  * 0 - No error and no read-list found.
365  *
366  * 1 - Successful read-list processing. The data is not yet in
367  * the pagelist and therefore the RPC request must be deferred. The
368  * I/O completion will enqueue the transport again and
369  * svc_rdma_recvfrom will complete the request.
370  *
371  * <0 - Error processing/posting read-list.
372  *
373  * NOTE: The ctxt must not be touched after the last WR has been posted
374  * because the I/O completion processing may occur on another
375  * processor and free / modify the context. Ne touche pas!
376  */
377 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
378 			 struct rpcrdma_msg *rmsgp,
379 			 struct svc_rqst *rqstp,
380 			 struct svc_rdma_op_ctxt *hdr_ctxt)
381 {
382 	struct ib_send_wr read_wr;
383 	struct ib_send_wr inv_wr;
384 	int err = 0;
385 	int ch_no;
386 	int ch_count;
387 	int byte_count;
388 	int sge_count;
389 	u64 sgl_offset;
390 	struct rpcrdma_read_chunk *ch;
391 	struct svc_rdma_op_ctxt *ctxt = NULL;
392 	struct svc_rdma_req_map *rpl_map;
393 	struct svc_rdma_req_map *chl_map;
394 
395 	/* If no read list is present, return 0 */
396 	ch = svc_rdma_get_read_chunk(rmsgp);
397 	if (!ch)
398 		return 0;
399 
400 	/* Allocate temporary reply and chunk maps */
401 	rpl_map = svc_rdma_get_req_map();
402 	chl_map = svc_rdma_get_req_map();
403 
404 	svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
405 	if (ch_count > RPCSVC_MAXPAGES)
406 		return -EINVAL;
407 
408 	if (!xprt->sc_frmr_pg_list_len)
409 		sge_count = map_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
410 					    rpl_map, chl_map, ch_count,
411 					    byte_count);
412 	else
413 		sge_count = fast_reg_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
414 						 rpl_map, chl_map, ch_count,
415 						 byte_count);
416 	if (sge_count < 0) {
417 		err = -EIO;
418 		goto out;
419 	}
420 
421 	sgl_offset = 0;
422 	ch_no = 0;
423 
424 	for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
425 	     ch->rc_discrim != 0; ch++, ch_no++) {
426 next_sge:
427 		ctxt = svc_rdma_get_context(xprt);
428 		ctxt->direction = DMA_FROM_DEVICE;
429 		ctxt->frmr = hdr_ctxt->frmr;
430 		ctxt->read_hdr = NULL;
431 		clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
432 		clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
433 
434 		/* Prepare READ WR */
435 		memset(&read_wr, 0, sizeof read_wr);
436 		read_wr.wr_id = (unsigned long)ctxt;
437 		read_wr.opcode = IB_WR_RDMA_READ;
438 		ctxt->wr_op = read_wr.opcode;
439 		read_wr.send_flags = IB_SEND_SIGNALED;
440 		read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
441 		read_wr.wr.rdma.remote_addr =
442 			get_unaligned(&(ch->rc_target.rs_offset)) +
443 			sgl_offset;
444 		read_wr.sg_list = ctxt->sge;
445 		read_wr.num_sge =
446 			rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
447 		err = rdma_set_ctxt_sge(xprt, ctxt, hdr_ctxt->frmr,
448 					&rpl_map->sge[chl_map->ch[ch_no].start],
449 					&sgl_offset,
450 					read_wr.num_sge);
451 		if (err) {
452 			svc_rdma_unmap_dma(ctxt);
453 			svc_rdma_put_context(ctxt, 0);
454 			goto out;
455 		}
456 		if (((ch+1)->rc_discrim == 0) &&
457 		    (read_wr.num_sge == chl_map->ch[ch_no].count)) {
458 			/*
459 			 * Mark the last RDMA_READ with a bit to
460 			 * indicate all RPC data has been fetched from
461 			 * the client and the RPC needs to be enqueued.
462 			 */
463 			set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
464 			if (hdr_ctxt->frmr) {
465 				set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
466 				/*
467 				 * Invalidate the local MR used to map the data
468 				 * sink.
469 				 */
470 				if (xprt->sc_dev_caps &
471 				    SVCRDMA_DEVCAP_READ_W_INV) {
472 					read_wr.opcode =
473 						IB_WR_RDMA_READ_WITH_INV;
474 					ctxt->wr_op = read_wr.opcode;
475 					read_wr.ex.invalidate_rkey =
476 						ctxt->frmr->mr->lkey;
477 				} else {
478 					/* Prepare INVALIDATE WR */
479 					memset(&inv_wr, 0, sizeof inv_wr);
480 					inv_wr.opcode = IB_WR_LOCAL_INV;
481 					inv_wr.send_flags = IB_SEND_SIGNALED;
482 					inv_wr.ex.invalidate_rkey =
483 						hdr_ctxt->frmr->mr->lkey;
484 					read_wr.next = &inv_wr;
485 				}
486 			}
487 			ctxt->read_hdr = hdr_ctxt;
488 		}
489 		/* Post the read */
490 		err = svc_rdma_send(xprt, &read_wr);
491 		if (err) {
492 			printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
493 			       err);
494 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
495 			svc_rdma_put_context(ctxt, 0);
496 			goto out;
497 		}
498 		atomic_inc(&rdma_stat_read);
499 
500 		if (read_wr.num_sge < chl_map->ch[ch_no].count) {
501 			chl_map->ch[ch_no].count -= read_wr.num_sge;
502 			chl_map->ch[ch_no].start += read_wr.num_sge;
503 			goto next_sge;
504 		}
505 		sgl_offset = 0;
506 		err = 1;
507 	}
508 
509  out:
510 	svc_rdma_put_req_map(rpl_map);
511 	svc_rdma_put_req_map(chl_map);
512 
513 	/* Detach arg pages. svc_recv will replenish them */
514 	for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
515 		rqstp->rq_pages[ch_no] = NULL;
516 
517 	/*
518 	 * Detach res pages. svc_release must see a resused count of
519 	 * zero or it will attempt to put them.
520 	 */
521 	while (rqstp->rq_resused)
522 		rqstp->rq_respages[--rqstp->rq_resused] = NULL;
523 
524 	return err;
525 }
526 
527 static int rdma_read_complete(struct svc_rqst *rqstp,
528 			      struct svc_rdma_op_ctxt *head)
529 {
530 	int page_no;
531 	int ret;
532 
533 	BUG_ON(!head);
534 
535 	/* Copy RPC pages */
536 	for (page_no = 0; page_no < head->count; page_no++) {
537 		put_page(rqstp->rq_pages[page_no]);
538 		rqstp->rq_pages[page_no] = head->pages[page_no];
539 	}
540 	/* Point rq_arg.pages past header */
541 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
542 	rqstp->rq_arg.page_len = head->arg.page_len;
543 	rqstp->rq_arg.page_base = head->arg.page_base;
544 
545 	/* rq_respages starts after the last arg page */
546 	rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
547 	rqstp->rq_resused = 0;
548 
549 	/* Rebuild rq_arg head and tail. */
550 	rqstp->rq_arg.head[0] = head->arg.head[0];
551 	rqstp->rq_arg.tail[0] = head->arg.tail[0];
552 	rqstp->rq_arg.len = head->arg.len;
553 	rqstp->rq_arg.buflen = head->arg.buflen;
554 
555 	/* Free the context */
556 	svc_rdma_put_context(head, 0);
557 
558 	/* XXX: What should this be? */
559 	rqstp->rq_prot = IPPROTO_MAX;
560 	svc_xprt_copy_addrs(rqstp, rqstp->rq_xprt);
561 
562 	ret = rqstp->rq_arg.head[0].iov_len
563 		+ rqstp->rq_arg.page_len
564 		+ rqstp->rq_arg.tail[0].iov_len;
565 	dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
566 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
567 		ret, rqstp->rq_arg.len,	rqstp->rq_arg.head[0].iov_base,
568 		rqstp->rq_arg.head[0].iov_len);
569 
570 	svc_xprt_received(rqstp->rq_xprt);
571 	return ret;
572 }
573 
574 /*
575  * Set up the rqstp thread context to point to the RQ buffer. If
576  * necessary, pull additional data from the client with an RDMA_READ
577  * request.
578  */
579 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
580 {
581 	struct svc_xprt *xprt = rqstp->rq_xprt;
582 	struct svcxprt_rdma *rdma_xprt =
583 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
584 	struct svc_rdma_op_ctxt *ctxt = NULL;
585 	struct rpcrdma_msg *rmsgp;
586 	int ret = 0;
587 	int len;
588 
589 	dprintk("svcrdma: rqstp=%p\n", rqstp);
590 
591 	spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
592 	if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
593 		ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
594 				  struct svc_rdma_op_ctxt,
595 				  dto_q);
596 		list_del_init(&ctxt->dto_q);
597 	}
598 	if (ctxt) {
599 		spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
600 		return rdma_read_complete(rqstp, ctxt);
601 	}
602 
603 	if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
604 		ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
605 				  struct svc_rdma_op_ctxt,
606 				  dto_q);
607 		list_del_init(&ctxt->dto_q);
608 	} else {
609 		atomic_inc(&rdma_stat_rq_starve);
610 		clear_bit(XPT_DATA, &xprt->xpt_flags);
611 		ctxt = NULL;
612 	}
613 	spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
614 	if (!ctxt) {
615 		/* This is the EAGAIN path. The svc_recv routine will
616 		 * return -EAGAIN, the nfsd thread will go to call into
617 		 * svc_recv again and we shouldn't be on the active
618 		 * transport list
619 		 */
620 		if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
621 			goto close_out;
622 
623 		BUG_ON(ret);
624 		goto out;
625 	}
626 	dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
627 		ctxt, rdma_xprt, rqstp, ctxt->wc_status);
628 	BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
629 	atomic_inc(&rdma_stat_recv);
630 
631 	/* Build up the XDR from the receive buffers. */
632 	rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
633 
634 	/* Decode the RDMA header. */
635 	len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
636 	rqstp->rq_xprt_hlen = len;
637 
638 	/* If the request is invalid, reply with an error */
639 	if (len < 0) {
640 		if (len == -ENOSYS)
641 			svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
642 		goto close_out;
643 	}
644 
645 	/* Read read-list data. */
646 	ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
647 	if (ret > 0) {
648 		/* read-list posted, defer until data received from client. */
649 		goto defer;
650 	}
651 	if (ret < 0) {
652 		/* Post of read-list failed, free context. */
653 		svc_rdma_put_context(ctxt, 1);
654 		return 0;
655 	}
656 
657 	ret = rqstp->rq_arg.head[0].iov_len
658 		+ rqstp->rq_arg.page_len
659 		+ rqstp->rq_arg.tail[0].iov_len;
660 	svc_rdma_put_context(ctxt, 0);
661  out:
662 	dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
663 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
664 		ret, rqstp->rq_arg.len,
665 		rqstp->rq_arg.head[0].iov_base,
666 		rqstp->rq_arg.head[0].iov_len);
667 	rqstp->rq_prot = IPPROTO_MAX;
668 	svc_xprt_copy_addrs(rqstp, xprt);
669 	svc_xprt_received(xprt);
670 	return ret;
671 
672  close_out:
673 	if (ctxt)
674 		svc_rdma_put_context(ctxt, 1);
675 	dprintk("svcrdma: transport %p is closing\n", xprt);
676 	/*
677 	 * Set the close bit and enqueue it. svc_recv will see the
678 	 * close bit and call svc_xprt_delete
679 	 */
680 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
681 defer:
682 	svc_xprt_received(xprt);
683 	return 0;
684 }
685