xref: /openbmc/linux/net/sunrpc/xprtrdma/rpc_rdma.c (revision 9659281c)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2020, Oracle and/or its affiliates.
4  * Copyright (c) 2003-2007 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 
42 /*
43  * rpc_rdma.c
44  *
45  * This file contains the guts of the RPC RDMA protocol, and
46  * does marshaling/unmarshaling, etc. It is also where interfacing
47  * to the Linux RPC framework lives.
48  */
49 
50 #include <linux/highmem.h>
51 
52 #include <linux/sunrpc/svc_rdma.h>
53 
54 #include "xprt_rdma.h"
55 #include <trace/events/rpcrdma.h>
56 
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY	RPCDBG_TRANS
59 #endif
60 
61 /* Returns size of largest RPC-over-RDMA header in a Call message
62  *
63  * The largest Call header contains a full-size Read list and a
64  * minimal Reply chunk.
65  */
66 static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
67 {
68 	unsigned int size;
69 
70 	/* Fixed header fields and list discriminators */
71 	size = RPCRDMA_HDRLEN_MIN;
72 
73 	/* Maximum Read list size */
74 	size += maxsegs * rpcrdma_readchunk_maxsz * sizeof(__be32);
75 
76 	/* Minimal Read chunk size */
77 	size += sizeof(__be32);	/* segment count */
78 	size += rpcrdma_segment_maxsz * sizeof(__be32);
79 	size += sizeof(__be32);	/* list discriminator */
80 
81 	return size;
82 }
83 
84 /* Returns size of largest RPC-over-RDMA header in a Reply message
85  *
86  * There is only one Write list or one Reply chunk per Reply
87  * message.  The larger list is the Write list.
88  */
89 static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
90 {
91 	unsigned int size;
92 
93 	/* Fixed header fields and list discriminators */
94 	size = RPCRDMA_HDRLEN_MIN;
95 
96 	/* Maximum Write list size */
97 	size += sizeof(__be32);		/* segment count */
98 	size += maxsegs * rpcrdma_segment_maxsz * sizeof(__be32);
99 	size += sizeof(__be32);	/* list discriminator */
100 
101 	return size;
102 }
103 
104 /**
105  * rpcrdma_set_max_header_sizes - Initialize inline payload sizes
106  * @ep: endpoint to initialize
107  *
108  * The max_inline fields contain the maximum size of an RPC message
109  * so the marshaling code doesn't have to repeat this calculation
110  * for every RPC.
111  */
112 void rpcrdma_set_max_header_sizes(struct rpcrdma_ep *ep)
113 {
114 	unsigned int maxsegs = ep->re_max_rdma_segs;
115 
116 	ep->re_max_inline_send =
117 		ep->re_inline_send - rpcrdma_max_call_header_size(maxsegs);
118 	ep->re_max_inline_recv =
119 		ep->re_inline_recv - rpcrdma_max_reply_header_size(maxsegs);
120 }
121 
122 /* The client can send a request inline as long as the RPCRDMA header
123  * plus the RPC call fit under the transport's inline limit. If the
124  * combined call message size exceeds that limit, the client must use
125  * a Read chunk for this operation.
126  *
127  * A Read chunk is also required if sending the RPC call inline would
128  * exceed this device's max_sge limit.
129  */
130 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
131 				struct rpc_rqst *rqst)
132 {
133 	struct xdr_buf *xdr = &rqst->rq_snd_buf;
134 	struct rpcrdma_ep *ep = r_xprt->rx_ep;
135 	unsigned int count, remaining, offset;
136 
137 	if (xdr->len > ep->re_max_inline_send)
138 		return false;
139 
140 	if (xdr->page_len) {
141 		remaining = xdr->page_len;
142 		offset = offset_in_page(xdr->page_base);
143 		count = RPCRDMA_MIN_SEND_SGES;
144 		while (remaining) {
145 			remaining -= min_t(unsigned int,
146 					   PAGE_SIZE - offset, remaining);
147 			offset = 0;
148 			if (++count > ep->re_attr.cap.max_send_sge)
149 				return false;
150 		}
151 	}
152 
153 	return true;
154 }
155 
156 /* The client can't know how large the actual reply will be. Thus it
157  * plans for the largest possible reply for that particular ULP
158  * operation. If the maximum combined reply message size exceeds that
159  * limit, the client must provide a write list or a reply chunk for
160  * this request.
161  */
162 static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
163 				   struct rpc_rqst *rqst)
164 {
165 	return rqst->rq_rcv_buf.buflen <= r_xprt->rx_ep->re_max_inline_recv;
166 }
167 
168 /* The client is required to provide a Reply chunk if the maximum
169  * size of the non-payload part of the RPC Reply is larger than
170  * the inline threshold.
171  */
172 static bool
173 rpcrdma_nonpayload_inline(const struct rpcrdma_xprt *r_xprt,
174 			  const struct rpc_rqst *rqst)
175 {
176 	const struct xdr_buf *buf = &rqst->rq_rcv_buf;
177 
178 	return (buf->head[0].iov_len + buf->tail[0].iov_len) <
179 		r_xprt->rx_ep->re_max_inline_recv;
180 }
181 
182 /* ACL likes to be lazy in allocating pages. For TCP, these
183  * pages can be allocated during receive processing. Not true
184  * for RDMA, which must always provision receive buffers
185  * up front.
186  */
187 static noinline int
188 rpcrdma_alloc_sparse_pages(struct xdr_buf *buf)
189 {
190 	struct page **ppages;
191 	int len;
192 
193 	len = buf->page_len;
194 	ppages = buf->pages + (buf->page_base >> PAGE_SHIFT);
195 	while (len > 0) {
196 		if (!*ppages)
197 			*ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
198 		if (!*ppages)
199 			return -ENOBUFS;
200 		ppages++;
201 		len -= PAGE_SIZE;
202 	}
203 
204 	return 0;
205 }
206 
207 /* Convert @vec to a single SGL element.
208  *
209  * Returns pointer to next available SGE, and bumps the total number
210  * of SGEs consumed.
211  */
212 static struct rpcrdma_mr_seg *
213 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
214 		     unsigned int *n)
215 {
216 	seg->mr_page = virt_to_page(vec->iov_base);
217 	seg->mr_offset = offset_in_page(vec->iov_base);
218 	seg->mr_len = vec->iov_len;
219 	++seg;
220 	++(*n);
221 	return seg;
222 }
223 
224 /* Convert @xdrbuf into SGEs no larger than a page each. As they
225  * are registered, these SGEs are then coalesced into RDMA segments
226  * when the selected memreg mode supports it.
227  *
228  * Returns positive number of SGEs consumed, or a negative errno.
229  */
230 
231 static int
232 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
233 		     unsigned int pos, enum rpcrdma_chunktype type,
234 		     struct rpcrdma_mr_seg *seg)
235 {
236 	unsigned long page_base;
237 	unsigned int len, n;
238 	struct page **ppages;
239 
240 	n = 0;
241 	if (pos == 0)
242 		seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
243 
244 	len = xdrbuf->page_len;
245 	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
246 	page_base = offset_in_page(xdrbuf->page_base);
247 	while (len) {
248 		seg->mr_page = *ppages;
249 		seg->mr_offset = page_base;
250 		seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
251 		len -= seg->mr_len;
252 		++ppages;
253 		++seg;
254 		++n;
255 		page_base = 0;
256 	}
257 
258 	if (type == rpcrdma_readch)
259 		goto out;
260 
261 	/* When encoding a Write chunk, some servers need to see an
262 	 * extra segment for non-XDR-aligned Write chunks. The upper
263 	 * layer provides space in the tail iovec that may be used
264 	 * for this purpose.
265 	 */
266 	if (type == rpcrdma_writech && r_xprt->rx_ep->re_implicit_roundup)
267 		goto out;
268 
269 	if (xdrbuf->tail[0].iov_len)
270 		rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
271 
272 out:
273 	if (unlikely(n > RPCRDMA_MAX_SEGS))
274 		return -EIO;
275 	return n;
276 }
277 
278 static int
279 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
280 {
281 	__be32 *p;
282 
283 	p = xdr_reserve_space(xdr, 4 * sizeof(*p));
284 	if (unlikely(!p))
285 		return -EMSGSIZE;
286 
287 	xdr_encode_rdma_segment(p, mr->mr_handle, mr->mr_length, mr->mr_offset);
288 	return 0;
289 }
290 
291 static int
292 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
293 		    u32 position)
294 {
295 	__be32 *p;
296 
297 	p = xdr_reserve_space(xdr, 6 * sizeof(*p));
298 	if (unlikely(!p))
299 		return -EMSGSIZE;
300 
301 	*p++ = xdr_one;			/* Item present */
302 	xdr_encode_read_segment(p, position, mr->mr_handle, mr->mr_length,
303 				mr->mr_offset);
304 	return 0;
305 }
306 
307 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
308 						 struct rpcrdma_req *req,
309 						 struct rpcrdma_mr_seg *seg,
310 						 int nsegs, bool writing,
311 						 struct rpcrdma_mr **mr)
312 {
313 	*mr = rpcrdma_mr_pop(&req->rl_free_mrs);
314 	if (!*mr) {
315 		*mr = rpcrdma_mr_get(r_xprt);
316 		if (!*mr)
317 			goto out_getmr_err;
318 		(*mr)->mr_req = req;
319 	}
320 
321 	rpcrdma_mr_push(*mr, &req->rl_registered);
322 	return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
323 
324 out_getmr_err:
325 	trace_xprtrdma_nomrs_err(r_xprt, req);
326 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
327 	rpcrdma_mrs_refresh(r_xprt);
328 	return ERR_PTR(-EAGAIN);
329 }
330 
331 /* Register and XDR encode the Read list. Supports encoding a list of read
332  * segments that belong to a single read chunk.
333  *
334  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
335  *
336  *  Read chunklist (a linked list):
337  *   N elements, position P (same P for all chunks of same arg!):
338  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
339  *
340  * Returns zero on success, or a negative errno if a failure occurred.
341  * @xdr is advanced to the next position in the stream.
342  *
343  * Only a single @pos value is currently supported.
344  */
345 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
346 				    struct rpcrdma_req *req,
347 				    struct rpc_rqst *rqst,
348 				    enum rpcrdma_chunktype rtype)
349 {
350 	struct xdr_stream *xdr = &req->rl_stream;
351 	struct rpcrdma_mr_seg *seg;
352 	struct rpcrdma_mr *mr;
353 	unsigned int pos;
354 	int nsegs;
355 
356 	if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
357 		goto done;
358 
359 	pos = rqst->rq_snd_buf.head[0].iov_len;
360 	if (rtype == rpcrdma_areadch)
361 		pos = 0;
362 	seg = req->rl_segments;
363 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
364 				     rtype, seg);
365 	if (nsegs < 0)
366 		return nsegs;
367 
368 	do {
369 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
370 		if (IS_ERR(seg))
371 			return PTR_ERR(seg);
372 
373 		if (encode_read_segment(xdr, mr, pos) < 0)
374 			return -EMSGSIZE;
375 
376 		trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
377 		r_xprt->rx_stats.read_chunk_count++;
378 		nsegs -= mr->mr_nents;
379 	} while (nsegs);
380 
381 done:
382 	if (xdr_stream_encode_item_absent(xdr) < 0)
383 		return -EMSGSIZE;
384 	return 0;
385 }
386 
387 /* Register and XDR encode the Write list. Supports encoding a list
388  * containing one array of plain segments that belong to a single
389  * write chunk.
390  *
391  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
392  *
393  *  Write chunklist (a list of (one) counted array):
394  *   N elements:
395  *    1 - N - HLOO - HLOO - ... - HLOO - 0
396  *
397  * Returns zero on success, or a negative errno if a failure occurred.
398  * @xdr is advanced to the next position in the stream.
399  *
400  * Only a single Write chunk is currently supported.
401  */
402 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
403 				     struct rpcrdma_req *req,
404 				     struct rpc_rqst *rqst,
405 				     enum rpcrdma_chunktype wtype)
406 {
407 	struct xdr_stream *xdr = &req->rl_stream;
408 	struct rpcrdma_mr_seg *seg;
409 	struct rpcrdma_mr *mr;
410 	int nsegs, nchunks;
411 	__be32 *segcount;
412 
413 	if (wtype != rpcrdma_writech)
414 		goto done;
415 
416 	seg = req->rl_segments;
417 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
418 				     rqst->rq_rcv_buf.head[0].iov_len,
419 				     wtype, seg);
420 	if (nsegs < 0)
421 		return nsegs;
422 
423 	if (xdr_stream_encode_item_present(xdr) < 0)
424 		return -EMSGSIZE;
425 	segcount = xdr_reserve_space(xdr, sizeof(*segcount));
426 	if (unlikely(!segcount))
427 		return -EMSGSIZE;
428 	/* Actual value encoded below */
429 
430 	nchunks = 0;
431 	do {
432 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
433 		if (IS_ERR(seg))
434 			return PTR_ERR(seg);
435 
436 		if (encode_rdma_segment(xdr, mr) < 0)
437 			return -EMSGSIZE;
438 
439 		trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
440 		r_xprt->rx_stats.write_chunk_count++;
441 		r_xprt->rx_stats.total_rdma_request += mr->mr_length;
442 		nchunks++;
443 		nsegs -= mr->mr_nents;
444 	} while (nsegs);
445 
446 	/* Update count of segments in this Write chunk */
447 	*segcount = cpu_to_be32(nchunks);
448 
449 done:
450 	if (xdr_stream_encode_item_absent(xdr) < 0)
451 		return -EMSGSIZE;
452 	return 0;
453 }
454 
455 /* Register and XDR encode the Reply chunk. Supports encoding an array
456  * of plain segments that belong to a single write (reply) chunk.
457  *
458  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
459  *
460  *  Reply chunk (a counted array):
461  *   N elements:
462  *    1 - N - HLOO - HLOO - ... - HLOO
463  *
464  * Returns zero on success, or a negative errno if a failure occurred.
465  * @xdr is advanced to the next position in the stream.
466  */
467 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
468 				      struct rpcrdma_req *req,
469 				      struct rpc_rqst *rqst,
470 				      enum rpcrdma_chunktype wtype)
471 {
472 	struct xdr_stream *xdr = &req->rl_stream;
473 	struct rpcrdma_mr_seg *seg;
474 	struct rpcrdma_mr *mr;
475 	int nsegs, nchunks;
476 	__be32 *segcount;
477 
478 	if (wtype != rpcrdma_replych) {
479 		if (xdr_stream_encode_item_absent(xdr) < 0)
480 			return -EMSGSIZE;
481 		return 0;
482 	}
483 
484 	seg = req->rl_segments;
485 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
486 	if (nsegs < 0)
487 		return nsegs;
488 
489 	if (xdr_stream_encode_item_present(xdr) < 0)
490 		return -EMSGSIZE;
491 	segcount = xdr_reserve_space(xdr, sizeof(*segcount));
492 	if (unlikely(!segcount))
493 		return -EMSGSIZE;
494 	/* Actual value encoded below */
495 
496 	nchunks = 0;
497 	do {
498 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
499 		if (IS_ERR(seg))
500 			return PTR_ERR(seg);
501 
502 		if (encode_rdma_segment(xdr, mr) < 0)
503 			return -EMSGSIZE;
504 
505 		trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
506 		r_xprt->rx_stats.reply_chunk_count++;
507 		r_xprt->rx_stats.total_rdma_request += mr->mr_length;
508 		nchunks++;
509 		nsegs -= mr->mr_nents;
510 	} while (nsegs);
511 
512 	/* Update count of segments in the Reply chunk */
513 	*segcount = cpu_to_be32(nchunks);
514 
515 	return 0;
516 }
517 
518 static void rpcrdma_sendctx_done(struct kref *kref)
519 {
520 	struct rpcrdma_req *req =
521 		container_of(kref, struct rpcrdma_req, rl_kref);
522 	struct rpcrdma_rep *rep = req->rl_reply;
523 
524 	rpcrdma_complete_rqst(rep);
525 	rep->rr_rxprt->rx_stats.reply_waits_for_send++;
526 }
527 
528 /**
529  * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
530  * @sc: sendctx containing SGEs to unmap
531  *
532  */
533 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
534 {
535 	struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
536 	struct ib_sge *sge;
537 
538 	if (!sc->sc_unmap_count)
539 		return;
540 
541 	/* The first two SGEs contain the transport header and
542 	 * the inline buffer. These are always left mapped so
543 	 * they can be cheaply re-used.
544 	 */
545 	for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
546 	     ++sge, --sc->sc_unmap_count)
547 		ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
548 				  DMA_TO_DEVICE);
549 
550 	kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
551 }
552 
553 /* Prepare an SGE for the RPC-over-RDMA transport header.
554  */
555 static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
556 				    struct rpcrdma_req *req, u32 len)
557 {
558 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
559 	struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
560 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
561 
562 	sge->addr = rdmab_addr(rb);
563 	sge->length = len;
564 	sge->lkey = rdmab_lkey(rb);
565 
566 	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
567 				      DMA_TO_DEVICE);
568 }
569 
570 /* The head iovec is straightforward, as it is usually already
571  * DMA-mapped. Sync the content that has changed.
572  */
573 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
574 				     struct rpcrdma_req *req, unsigned int len)
575 {
576 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
577 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
578 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
579 
580 	if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
581 		return false;
582 
583 	sge->addr = rdmab_addr(rb);
584 	sge->length = len;
585 	sge->lkey = rdmab_lkey(rb);
586 
587 	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
588 				      DMA_TO_DEVICE);
589 	return true;
590 }
591 
592 /* If there is a page list present, DMA map and prepare an
593  * SGE for each page to be sent.
594  */
595 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
596 				     struct xdr_buf *xdr)
597 {
598 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
599 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
600 	unsigned int page_base, len, remaining;
601 	struct page **ppages;
602 	struct ib_sge *sge;
603 
604 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
605 	page_base = offset_in_page(xdr->page_base);
606 	remaining = xdr->page_len;
607 	while (remaining) {
608 		sge = &sc->sc_sges[req->rl_wr.num_sge++];
609 		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
610 		sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
611 					    page_base, len, DMA_TO_DEVICE);
612 		if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
613 			goto out_mapping_err;
614 
615 		sge->length = len;
616 		sge->lkey = rdmab_lkey(rb);
617 
618 		sc->sc_unmap_count++;
619 		ppages++;
620 		remaining -= len;
621 		page_base = 0;
622 	}
623 
624 	return true;
625 
626 out_mapping_err:
627 	trace_xprtrdma_dma_maperr(sge->addr);
628 	return false;
629 }
630 
631 /* The tail iovec may include an XDR pad for the page list,
632  * as well as additional content, and may not reside in the
633  * same page as the head iovec.
634  */
635 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
636 				     struct xdr_buf *xdr,
637 				     unsigned int page_base, unsigned int len)
638 {
639 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
640 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
641 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
642 	struct page *page = virt_to_page(xdr->tail[0].iov_base);
643 
644 	sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
645 				    DMA_TO_DEVICE);
646 	if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
647 		goto out_mapping_err;
648 
649 	sge->length = len;
650 	sge->lkey = rdmab_lkey(rb);
651 	++sc->sc_unmap_count;
652 	return true;
653 
654 out_mapping_err:
655 	trace_xprtrdma_dma_maperr(sge->addr);
656 	return false;
657 }
658 
659 /* Copy the tail to the end of the head buffer.
660  */
661 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
662 				    struct rpcrdma_req *req,
663 				    struct xdr_buf *xdr)
664 {
665 	unsigned char *dst;
666 
667 	dst = (unsigned char *)xdr->head[0].iov_base;
668 	dst += xdr->head[0].iov_len + xdr->page_len;
669 	memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
670 	r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
671 }
672 
673 /* Copy pagelist content into the head buffer.
674  */
675 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
676 				    struct rpcrdma_req *req,
677 				    struct xdr_buf *xdr)
678 {
679 	unsigned int len, page_base, remaining;
680 	struct page **ppages;
681 	unsigned char *src, *dst;
682 
683 	dst = (unsigned char *)xdr->head[0].iov_base;
684 	dst += xdr->head[0].iov_len;
685 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
686 	page_base = offset_in_page(xdr->page_base);
687 	remaining = xdr->page_len;
688 	while (remaining) {
689 		src = page_address(*ppages);
690 		src += page_base;
691 		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
692 		memcpy(dst, src, len);
693 		r_xprt->rx_stats.pullup_copy_count += len;
694 
695 		ppages++;
696 		dst += len;
697 		remaining -= len;
698 		page_base = 0;
699 	}
700 }
701 
702 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
703  * When the head, pagelist, and tail are small, a pull-up copy
704  * is considerably less costly than DMA mapping the components
705  * of @xdr.
706  *
707  * Assumptions:
708  *  - the caller has already verified that the total length
709  *    of the RPC Call body will fit into @rl_sendbuf.
710  */
711 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
712 					struct rpcrdma_req *req,
713 					struct xdr_buf *xdr)
714 {
715 	if (unlikely(xdr->tail[0].iov_len))
716 		rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
717 
718 	if (unlikely(xdr->page_len))
719 		rpcrdma_pullup_pagelist(r_xprt, req, xdr);
720 
721 	/* The whole RPC message resides in the head iovec now */
722 	return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
723 }
724 
725 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
726 					struct rpcrdma_req *req,
727 					struct xdr_buf *xdr)
728 {
729 	struct kvec *tail = &xdr->tail[0];
730 
731 	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
732 		return false;
733 	if (xdr->page_len)
734 		if (!rpcrdma_prepare_pagelist(req, xdr))
735 			return false;
736 	if (tail->iov_len)
737 		if (!rpcrdma_prepare_tail_iov(req, xdr,
738 					      offset_in_page(tail->iov_base),
739 					      tail->iov_len))
740 			return false;
741 
742 	if (req->rl_sendctx->sc_unmap_count)
743 		kref_get(&req->rl_kref);
744 	return true;
745 }
746 
747 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
748 				   struct rpcrdma_req *req,
749 				   struct xdr_buf *xdr)
750 {
751 	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
752 		return false;
753 
754 	/* If there is a Read chunk, the page list is being handled
755 	 * via explicit RDMA, and thus is skipped here.
756 	 */
757 
758 	/* Do not include the tail if it is only an XDR pad */
759 	if (xdr->tail[0].iov_len > 3) {
760 		unsigned int page_base, len;
761 
762 		/* If the content in the page list is an odd length,
763 		 * xdr_write_pages() adds a pad at the beginning of
764 		 * the tail iovec. Force the tail's non-pad content to
765 		 * land at the next XDR position in the Send message.
766 		 */
767 		page_base = offset_in_page(xdr->tail[0].iov_base);
768 		len = xdr->tail[0].iov_len;
769 		page_base += len & 3;
770 		len -= len & 3;
771 		if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
772 			return false;
773 		kref_get(&req->rl_kref);
774 	}
775 
776 	return true;
777 }
778 
779 /**
780  * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
781  * @r_xprt: controlling transport
782  * @req: context of RPC Call being marshalled
783  * @hdrlen: size of transport header, in bytes
784  * @xdr: xdr_buf containing RPC Call
785  * @rtype: chunk type being encoded
786  *
787  * Returns 0 on success; otherwise a negative errno is returned.
788  */
789 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
790 				     struct rpcrdma_req *req, u32 hdrlen,
791 				     struct xdr_buf *xdr,
792 				     enum rpcrdma_chunktype rtype)
793 {
794 	int ret;
795 
796 	ret = -EAGAIN;
797 	req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
798 	if (!req->rl_sendctx)
799 		goto out_nosc;
800 	req->rl_sendctx->sc_unmap_count = 0;
801 	req->rl_sendctx->sc_req = req;
802 	kref_init(&req->rl_kref);
803 	req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
804 	req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
805 	req->rl_wr.num_sge = 0;
806 	req->rl_wr.opcode = IB_WR_SEND;
807 
808 	rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
809 
810 	ret = -EIO;
811 	switch (rtype) {
812 	case rpcrdma_noch_pullup:
813 		if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
814 			goto out_unmap;
815 		break;
816 	case rpcrdma_noch_mapped:
817 		if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
818 			goto out_unmap;
819 		break;
820 	case rpcrdma_readch:
821 		if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
822 			goto out_unmap;
823 		break;
824 	case rpcrdma_areadch:
825 		break;
826 	default:
827 		goto out_unmap;
828 	}
829 
830 	return 0;
831 
832 out_unmap:
833 	rpcrdma_sendctx_unmap(req->rl_sendctx);
834 out_nosc:
835 	trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
836 	return ret;
837 }
838 
839 /**
840  * rpcrdma_marshal_req - Marshal and send one RPC request
841  * @r_xprt: controlling transport
842  * @rqst: RPC request to be marshaled
843  *
844  * For the RPC in "rqst", this function:
845  *  - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
846  *  - Registers Read, Write, and Reply chunks
847  *  - Constructs the transport header
848  *  - Posts a Send WR to send the transport header and request
849  *
850  * Returns:
851  *	%0 if the RPC was sent successfully,
852  *	%-ENOTCONN if the connection was lost,
853  *	%-EAGAIN if the caller should call again with the same arguments,
854  *	%-ENOBUFS if the caller should call again after a delay,
855  *	%-EMSGSIZE if the transport header is too small,
856  *	%-EIO if a permanent problem occurred while marshaling.
857  */
858 int
859 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
860 {
861 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
862 	struct xdr_stream *xdr = &req->rl_stream;
863 	enum rpcrdma_chunktype rtype, wtype;
864 	struct xdr_buf *buf = &rqst->rq_snd_buf;
865 	bool ddp_allowed;
866 	__be32 *p;
867 	int ret;
868 
869 	if (unlikely(rqst->rq_rcv_buf.flags & XDRBUF_SPARSE_PAGES)) {
870 		ret = rpcrdma_alloc_sparse_pages(&rqst->rq_rcv_buf);
871 		if (ret)
872 			return ret;
873 	}
874 
875 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
876 	xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
877 			rqst);
878 
879 	/* Fixed header fields */
880 	ret = -EMSGSIZE;
881 	p = xdr_reserve_space(xdr, 4 * sizeof(*p));
882 	if (!p)
883 		goto out_err;
884 	*p++ = rqst->rq_xid;
885 	*p++ = rpcrdma_version;
886 	*p++ = r_xprt->rx_buf.rb_max_requests;
887 
888 	/* When the ULP employs a GSS flavor that guarantees integrity
889 	 * or privacy, direct data placement of individual data items
890 	 * is not allowed.
891 	 */
892 	ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
893 				&rqst->rq_cred->cr_auth->au_flags);
894 
895 	/*
896 	 * Chunks needed for results?
897 	 *
898 	 * o If the expected result is under the inline threshold, all ops
899 	 *   return as inline.
900 	 * o Large read ops return data as write chunk(s), header as
901 	 *   inline.
902 	 * o Large non-read ops return as a single reply chunk.
903 	 */
904 	if (rpcrdma_results_inline(r_xprt, rqst))
905 		wtype = rpcrdma_noch;
906 	else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
907 		 rpcrdma_nonpayload_inline(r_xprt, rqst))
908 		wtype = rpcrdma_writech;
909 	else
910 		wtype = rpcrdma_replych;
911 
912 	/*
913 	 * Chunks needed for arguments?
914 	 *
915 	 * o If the total request is under the inline threshold, all ops
916 	 *   are sent as inline.
917 	 * o Large write ops transmit data as read chunk(s), header as
918 	 *   inline.
919 	 * o Large non-write ops are sent with the entire message as a
920 	 *   single read chunk (protocol 0-position special case).
921 	 *
922 	 * This assumes that the upper layer does not present a request
923 	 * that both has a data payload, and whose non-data arguments
924 	 * by themselves are larger than the inline threshold.
925 	 */
926 	if (rpcrdma_args_inline(r_xprt, rqst)) {
927 		*p++ = rdma_msg;
928 		rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
929 			rpcrdma_noch_pullup : rpcrdma_noch_mapped;
930 	} else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
931 		*p++ = rdma_msg;
932 		rtype = rpcrdma_readch;
933 	} else {
934 		r_xprt->rx_stats.nomsg_call_count++;
935 		*p++ = rdma_nomsg;
936 		rtype = rpcrdma_areadch;
937 	}
938 
939 	/* This implementation supports the following combinations
940 	 * of chunk lists in one RPC-over-RDMA Call message:
941 	 *
942 	 *   - Read list
943 	 *   - Write list
944 	 *   - Reply chunk
945 	 *   - Read list + Reply chunk
946 	 *
947 	 * It might not yet support the following combinations:
948 	 *
949 	 *   - Read list + Write list
950 	 *
951 	 * It does not support the following combinations:
952 	 *
953 	 *   - Write list + Reply chunk
954 	 *   - Read list + Write list + Reply chunk
955 	 *
956 	 * This implementation supports only a single chunk in each
957 	 * Read or Write list. Thus for example the client cannot
958 	 * send a Call message with a Position Zero Read chunk and a
959 	 * regular Read chunk at the same time.
960 	 */
961 	ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
962 	if (ret)
963 		goto out_err;
964 	ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
965 	if (ret)
966 		goto out_err;
967 	ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
968 	if (ret)
969 		goto out_err;
970 
971 	ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
972 					buf, rtype);
973 	if (ret)
974 		goto out_err;
975 
976 	trace_xprtrdma_marshal(req, rtype, wtype);
977 	return 0;
978 
979 out_err:
980 	trace_xprtrdma_marshal_failed(rqst, ret);
981 	r_xprt->rx_stats.failed_marshal_count++;
982 	frwr_reset(req);
983 	return ret;
984 }
985 
986 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
987 					 struct rpcrdma_buffer *buf,
988 					 u32 grant)
989 {
990 	buf->rb_credits = grant;
991 	xprt->cwnd = grant << RPC_CWNDSHIFT;
992 }
993 
994 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
995 {
996 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
997 
998 	spin_lock(&xprt->transport_lock);
999 	__rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1000 	spin_unlock(&xprt->transport_lock);
1001 }
1002 
1003 /**
1004  * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1005  * @r_xprt: controlling transport instance
1006  *
1007  * Prepare @r_xprt for the next connection by reinitializing
1008  * its credit grant to one (see RFC 8166, Section 3.3.3).
1009  */
1010 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1011 {
1012 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1013 
1014 	spin_lock(&xprt->transport_lock);
1015 	xprt->cong = 0;
1016 	__rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1017 	spin_unlock(&xprt->transport_lock);
1018 }
1019 
1020 /**
1021  * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1022  * @rqst: controlling RPC request
1023  * @srcp: points to RPC message payload in receive buffer
1024  * @copy_len: remaining length of receive buffer content
1025  * @pad: Write chunk pad bytes needed (zero for pure inline)
1026  *
1027  * The upper layer has set the maximum number of bytes it can
1028  * receive in each component of rq_rcv_buf. These values are set in
1029  * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1030  *
1031  * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1032  * many cases this function simply updates iov_base pointers in
1033  * rq_rcv_buf to point directly to the received reply data, to
1034  * avoid copying reply data.
1035  *
1036  * Returns the count of bytes which had to be memcopied.
1037  */
1038 static unsigned long
1039 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1040 {
1041 	unsigned long fixup_copy_count;
1042 	int i, npages, curlen;
1043 	char *destp;
1044 	struct page **ppages;
1045 	int page_base;
1046 
1047 	/* The head iovec is redirected to the RPC reply message
1048 	 * in the receive buffer, to avoid a memcopy.
1049 	 */
1050 	rqst->rq_rcv_buf.head[0].iov_base = srcp;
1051 	rqst->rq_private_buf.head[0].iov_base = srcp;
1052 
1053 	/* The contents of the receive buffer that follow
1054 	 * head.iov_len bytes are copied into the page list.
1055 	 */
1056 	curlen = rqst->rq_rcv_buf.head[0].iov_len;
1057 	if (curlen > copy_len)
1058 		curlen = copy_len;
1059 	srcp += curlen;
1060 	copy_len -= curlen;
1061 
1062 	ppages = rqst->rq_rcv_buf.pages +
1063 		(rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1064 	page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1065 	fixup_copy_count = 0;
1066 	if (copy_len && rqst->rq_rcv_buf.page_len) {
1067 		int pagelist_len;
1068 
1069 		pagelist_len = rqst->rq_rcv_buf.page_len;
1070 		if (pagelist_len > copy_len)
1071 			pagelist_len = copy_len;
1072 		npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1073 		for (i = 0; i < npages; i++) {
1074 			curlen = PAGE_SIZE - page_base;
1075 			if (curlen > pagelist_len)
1076 				curlen = pagelist_len;
1077 
1078 			destp = kmap_atomic(ppages[i]);
1079 			memcpy(destp + page_base, srcp, curlen);
1080 			flush_dcache_page(ppages[i]);
1081 			kunmap_atomic(destp);
1082 			srcp += curlen;
1083 			copy_len -= curlen;
1084 			fixup_copy_count += curlen;
1085 			pagelist_len -= curlen;
1086 			if (!pagelist_len)
1087 				break;
1088 			page_base = 0;
1089 		}
1090 
1091 		/* Implicit padding for the last segment in a Write
1092 		 * chunk is inserted inline at the front of the tail
1093 		 * iovec. The upper layer ignores the content of
1094 		 * the pad. Simply ensure inline content in the tail
1095 		 * that follows the Write chunk is properly aligned.
1096 		 */
1097 		if (pad)
1098 			srcp -= pad;
1099 	}
1100 
1101 	/* The tail iovec is redirected to the remaining data
1102 	 * in the receive buffer, to avoid a memcopy.
1103 	 */
1104 	if (copy_len || pad) {
1105 		rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1106 		rqst->rq_private_buf.tail[0].iov_base = srcp;
1107 	}
1108 
1109 	if (fixup_copy_count)
1110 		trace_xprtrdma_fixup(rqst, fixup_copy_count);
1111 	return fixup_copy_count;
1112 }
1113 
1114 /* By convention, backchannel calls arrive via rdma_msg type
1115  * messages, and never populate the chunk lists. This makes
1116  * the RPC/RDMA header small and fixed in size, so it is
1117  * straightforward to check the RPC header's direction field.
1118  */
1119 static bool
1120 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1121 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1122 {
1123 	struct xdr_stream *xdr = &rep->rr_stream;
1124 	__be32 *p;
1125 
1126 	if (rep->rr_proc != rdma_msg)
1127 		return false;
1128 
1129 	/* Peek at stream contents without advancing. */
1130 	p = xdr_inline_decode(xdr, 0);
1131 
1132 	/* Chunk lists */
1133 	if (xdr_item_is_present(p++))
1134 		return false;
1135 	if (xdr_item_is_present(p++))
1136 		return false;
1137 	if (xdr_item_is_present(p++))
1138 		return false;
1139 
1140 	/* RPC header */
1141 	if (*p++ != rep->rr_xid)
1142 		return false;
1143 	if (*p != cpu_to_be32(RPC_CALL))
1144 		return false;
1145 
1146 	/* Now that we are sure this is a backchannel call,
1147 	 * advance to the RPC header.
1148 	 */
1149 	p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1150 	if (unlikely(!p))
1151 		return true;
1152 
1153 	rpcrdma_bc_receive_call(r_xprt, rep);
1154 	return true;
1155 }
1156 #else	/* CONFIG_SUNRPC_BACKCHANNEL */
1157 {
1158 	return false;
1159 }
1160 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
1161 
1162 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1163 {
1164 	u32 handle;
1165 	u64 offset;
1166 	__be32 *p;
1167 
1168 	p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1169 	if (unlikely(!p))
1170 		return -EIO;
1171 
1172 	xdr_decode_rdma_segment(p, &handle, length, &offset);
1173 	trace_xprtrdma_decode_seg(handle, *length, offset);
1174 	return 0;
1175 }
1176 
1177 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1178 {
1179 	u32 segcount, seglength;
1180 	__be32 *p;
1181 
1182 	p = xdr_inline_decode(xdr, sizeof(*p));
1183 	if (unlikely(!p))
1184 		return -EIO;
1185 
1186 	*length = 0;
1187 	segcount = be32_to_cpup(p);
1188 	while (segcount--) {
1189 		if (decode_rdma_segment(xdr, &seglength))
1190 			return -EIO;
1191 		*length += seglength;
1192 	}
1193 
1194 	return 0;
1195 }
1196 
1197 /* In RPC-over-RDMA Version One replies, a Read list is never
1198  * expected. This decoder is a stub that returns an error if
1199  * a Read list is present.
1200  */
1201 static int decode_read_list(struct xdr_stream *xdr)
1202 {
1203 	__be32 *p;
1204 
1205 	p = xdr_inline_decode(xdr, sizeof(*p));
1206 	if (unlikely(!p))
1207 		return -EIO;
1208 	if (unlikely(xdr_item_is_present(p)))
1209 		return -EIO;
1210 	return 0;
1211 }
1212 
1213 /* Supports only one Write chunk in the Write list
1214  */
1215 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1216 {
1217 	u32 chunklen;
1218 	bool first;
1219 	__be32 *p;
1220 
1221 	*length = 0;
1222 	first = true;
1223 	do {
1224 		p = xdr_inline_decode(xdr, sizeof(*p));
1225 		if (unlikely(!p))
1226 			return -EIO;
1227 		if (xdr_item_is_absent(p))
1228 			break;
1229 		if (!first)
1230 			return -EIO;
1231 
1232 		if (decode_write_chunk(xdr, &chunklen))
1233 			return -EIO;
1234 		*length += chunklen;
1235 		first = false;
1236 	} while (true);
1237 	return 0;
1238 }
1239 
1240 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1241 {
1242 	__be32 *p;
1243 
1244 	p = xdr_inline_decode(xdr, sizeof(*p));
1245 	if (unlikely(!p))
1246 		return -EIO;
1247 
1248 	*length = 0;
1249 	if (xdr_item_is_present(p))
1250 		if (decode_write_chunk(xdr, length))
1251 			return -EIO;
1252 	return 0;
1253 }
1254 
1255 static int
1256 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1257 		   struct rpc_rqst *rqst)
1258 {
1259 	struct xdr_stream *xdr = &rep->rr_stream;
1260 	u32 writelist, replychunk, rpclen;
1261 	char *base;
1262 
1263 	/* Decode the chunk lists */
1264 	if (decode_read_list(xdr))
1265 		return -EIO;
1266 	if (decode_write_list(xdr, &writelist))
1267 		return -EIO;
1268 	if (decode_reply_chunk(xdr, &replychunk))
1269 		return -EIO;
1270 
1271 	/* RDMA_MSG sanity checks */
1272 	if (unlikely(replychunk))
1273 		return -EIO;
1274 
1275 	/* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1276 	base = (char *)xdr_inline_decode(xdr, 0);
1277 	rpclen = xdr_stream_remaining(xdr);
1278 	r_xprt->rx_stats.fixup_copy_count +=
1279 		rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1280 
1281 	r_xprt->rx_stats.total_rdma_reply += writelist;
1282 	return rpclen + xdr_align_size(writelist);
1283 }
1284 
1285 static noinline int
1286 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1287 {
1288 	struct xdr_stream *xdr = &rep->rr_stream;
1289 	u32 writelist, replychunk;
1290 
1291 	/* Decode the chunk lists */
1292 	if (decode_read_list(xdr))
1293 		return -EIO;
1294 	if (decode_write_list(xdr, &writelist))
1295 		return -EIO;
1296 	if (decode_reply_chunk(xdr, &replychunk))
1297 		return -EIO;
1298 
1299 	/* RDMA_NOMSG sanity checks */
1300 	if (unlikely(writelist))
1301 		return -EIO;
1302 	if (unlikely(!replychunk))
1303 		return -EIO;
1304 
1305 	/* Reply chunk buffer already is the reply vector */
1306 	r_xprt->rx_stats.total_rdma_reply += replychunk;
1307 	return replychunk;
1308 }
1309 
1310 static noinline int
1311 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1312 		     struct rpc_rqst *rqst)
1313 {
1314 	struct xdr_stream *xdr = &rep->rr_stream;
1315 	__be32 *p;
1316 
1317 	p = xdr_inline_decode(xdr, sizeof(*p));
1318 	if (unlikely(!p))
1319 		return -EIO;
1320 
1321 	switch (*p) {
1322 	case err_vers:
1323 		p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1324 		if (!p)
1325 			break;
1326 		trace_xprtrdma_err_vers(rqst, p, p + 1);
1327 		break;
1328 	case err_chunk:
1329 		trace_xprtrdma_err_chunk(rqst);
1330 		break;
1331 	default:
1332 		trace_xprtrdma_err_unrecognized(rqst, p);
1333 	}
1334 
1335 	return -EIO;
1336 }
1337 
1338 /**
1339  * rpcrdma_unpin_rqst - Release rqst without completing it
1340  * @rep: RPC/RDMA Receive context
1341  *
1342  * This is done when a connection is lost so that a Reply
1343  * can be dropped and its matching Call can be subsequently
1344  * retransmitted on a new connection.
1345  */
1346 void rpcrdma_unpin_rqst(struct rpcrdma_rep *rep)
1347 {
1348 	struct rpc_xprt *xprt = &rep->rr_rxprt->rx_xprt;
1349 	struct rpc_rqst *rqst = rep->rr_rqst;
1350 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
1351 
1352 	req->rl_reply = NULL;
1353 	rep->rr_rqst = NULL;
1354 
1355 	spin_lock(&xprt->queue_lock);
1356 	xprt_unpin_rqst(rqst);
1357 	spin_unlock(&xprt->queue_lock);
1358 }
1359 
1360 /**
1361  * rpcrdma_complete_rqst - Pass completed rqst back to RPC
1362  * @rep: RPC/RDMA Receive context
1363  *
1364  * Reconstruct the RPC reply and complete the transaction
1365  * while @rqst is still pinned to ensure the rep, rqst, and
1366  * rq_task pointers remain stable.
1367  */
1368 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1369 {
1370 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1371 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1372 	struct rpc_rqst *rqst = rep->rr_rqst;
1373 	int status;
1374 
1375 	switch (rep->rr_proc) {
1376 	case rdma_msg:
1377 		status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1378 		break;
1379 	case rdma_nomsg:
1380 		status = rpcrdma_decode_nomsg(r_xprt, rep);
1381 		break;
1382 	case rdma_error:
1383 		status = rpcrdma_decode_error(r_xprt, rep, rqst);
1384 		break;
1385 	default:
1386 		status = -EIO;
1387 	}
1388 	if (status < 0)
1389 		goto out_badheader;
1390 
1391 out:
1392 	spin_lock(&xprt->queue_lock);
1393 	xprt_complete_rqst(rqst->rq_task, status);
1394 	xprt_unpin_rqst(rqst);
1395 	spin_unlock(&xprt->queue_lock);
1396 	return;
1397 
1398 out_badheader:
1399 	trace_xprtrdma_reply_hdr_err(rep);
1400 	r_xprt->rx_stats.bad_reply_count++;
1401 	rqst->rq_task->tk_status = status;
1402 	status = 0;
1403 	goto out;
1404 }
1405 
1406 static void rpcrdma_reply_done(struct kref *kref)
1407 {
1408 	struct rpcrdma_req *req =
1409 		container_of(kref, struct rpcrdma_req, rl_kref);
1410 
1411 	rpcrdma_complete_rqst(req->rl_reply);
1412 }
1413 
1414 /**
1415  * rpcrdma_reply_handler - Process received RPC/RDMA messages
1416  * @rep: Incoming rpcrdma_rep object to process
1417  *
1418  * Errors must result in the RPC task either being awakened, or
1419  * allowed to timeout, to discover the errors at that time.
1420  */
1421 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1422 {
1423 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1424 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1425 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1426 	struct rpcrdma_req *req;
1427 	struct rpc_rqst *rqst;
1428 	u32 credits;
1429 	__be32 *p;
1430 
1431 	/* Any data means we had a useful conversation, so
1432 	 * then we don't need to delay the next reconnect.
1433 	 */
1434 	if (xprt->reestablish_timeout)
1435 		xprt->reestablish_timeout = 0;
1436 
1437 	/* Fixed transport header fields */
1438 	xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1439 			rep->rr_hdrbuf.head[0].iov_base, NULL);
1440 	p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1441 	if (unlikely(!p))
1442 		goto out_shortreply;
1443 	rep->rr_xid = *p++;
1444 	rep->rr_vers = *p++;
1445 	credits = be32_to_cpu(*p++);
1446 	rep->rr_proc = *p++;
1447 
1448 	if (rep->rr_vers != rpcrdma_version)
1449 		goto out_badversion;
1450 
1451 	if (rpcrdma_is_bcall(r_xprt, rep))
1452 		return;
1453 
1454 	/* Match incoming rpcrdma_rep to an rpcrdma_req to
1455 	 * get context for handling any incoming chunks.
1456 	 */
1457 	spin_lock(&xprt->queue_lock);
1458 	rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1459 	if (!rqst)
1460 		goto out_norqst;
1461 	xprt_pin_rqst(rqst);
1462 	spin_unlock(&xprt->queue_lock);
1463 
1464 	if (credits == 0)
1465 		credits = 1;	/* don't deadlock */
1466 	else if (credits > r_xprt->rx_ep->re_max_requests)
1467 		credits = r_xprt->rx_ep->re_max_requests;
1468 	rpcrdma_post_recvs(r_xprt, credits + (buf->rb_bc_srv_max_requests << 1),
1469 			   false);
1470 	if (buf->rb_credits != credits)
1471 		rpcrdma_update_cwnd(r_xprt, credits);
1472 
1473 	req = rpcr_to_rdmar(rqst);
1474 	if (unlikely(req->rl_reply))
1475 		rpcrdma_rep_put(buf, req->rl_reply);
1476 	req->rl_reply = rep;
1477 	rep->rr_rqst = rqst;
1478 
1479 	trace_xprtrdma_reply(rqst->rq_task, rep, credits);
1480 
1481 	if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1482 		frwr_reminv(rep, &req->rl_registered);
1483 	if (!list_empty(&req->rl_registered))
1484 		frwr_unmap_async(r_xprt, req);
1485 		/* LocalInv completion will complete the RPC */
1486 	else
1487 		kref_put(&req->rl_kref, rpcrdma_reply_done);
1488 	return;
1489 
1490 out_badversion:
1491 	trace_xprtrdma_reply_vers_err(rep);
1492 	goto out;
1493 
1494 out_norqst:
1495 	spin_unlock(&xprt->queue_lock);
1496 	trace_xprtrdma_reply_rqst_err(rep);
1497 	goto out;
1498 
1499 out_shortreply:
1500 	trace_xprtrdma_reply_short_err(rep);
1501 
1502 out:
1503 	rpcrdma_rep_put(buf, rep);
1504 }
1505