xref: /openbmc/linux/net/sunrpc/xprtrdma/rpc_rdma.c (revision dc6a81c3)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
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  * @r_xprt: transport instance 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_xprt *r_xprt)
113 {
114 	unsigned int maxsegs = r_xprt->rx_ia.ri_max_rdma_segs;
115 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
116 
117 	ep->rep_max_inline_send =
118 		ep->rep_inline_send - rpcrdma_max_call_header_size(maxsegs);
119 	ep->rep_max_inline_recv =
120 		ep->rep_inline_recv - rpcrdma_max_reply_header_size(maxsegs);
121 }
122 
123 /* The client can send a request inline as long as the RPCRDMA header
124  * plus the RPC call fit under the transport's inline limit. If the
125  * combined call message size exceeds that limit, the client must use
126  * a Read chunk for this operation.
127  *
128  * A Read chunk is also required if sending the RPC call inline would
129  * exceed this device's max_sge limit.
130  */
131 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
132 				struct rpc_rqst *rqst)
133 {
134 	struct xdr_buf *xdr = &rqst->rq_snd_buf;
135 	unsigned int count, remaining, offset;
136 
137 	if (xdr->len > r_xprt->rx_ep.rep_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 > r_xprt->rx_ep.rep_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.rep_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.rep_max_inline_recv;
180 }
181 
182 /* Split @vec on page boundaries into SGEs. FMR registers pages, not
183  * a byte range. Other modes coalesce these SGEs into a single MR
184  * when they can.
185  *
186  * Returns pointer to next available SGE, and bumps the total number
187  * of SGEs consumed.
188  */
189 static struct rpcrdma_mr_seg *
190 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
191 		     unsigned int *n)
192 {
193 	u32 remaining, page_offset;
194 	char *base;
195 
196 	base = vec->iov_base;
197 	page_offset = offset_in_page(base);
198 	remaining = vec->iov_len;
199 	while (remaining) {
200 		seg->mr_page = NULL;
201 		seg->mr_offset = base;
202 		seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
203 		remaining -= seg->mr_len;
204 		base += seg->mr_len;
205 		++seg;
206 		++(*n);
207 		page_offset = 0;
208 	}
209 	return seg;
210 }
211 
212 /* Convert @xdrbuf into SGEs no larger than a page each. As they
213  * are registered, these SGEs are then coalesced into RDMA segments
214  * when the selected memreg mode supports it.
215  *
216  * Returns positive number of SGEs consumed, or a negative errno.
217  */
218 
219 static int
220 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
221 		     unsigned int pos, enum rpcrdma_chunktype type,
222 		     struct rpcrdma_mr_seg *seg)
223 {
224 	unsigned long page_base;
225 	unsigned int len, n;
226 	struct page **ppages;
227 
228 	n = 0;
229 	if (pos == 0)
230 		seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
231 
232 	len = xdrbuf->page_len;
233 	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
234 	page_base = offset_in_page(xdrbuf->page_base);
235 	while (len) {
236 		/* ACL likes to be lazy in allocating pages - ACLs
237 		 * are small by default but can get huge.
238 		 */
239 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
240 			if (!*ppages)
241 				*ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
242 			if (!*ppages)
243 				return -ENOBUFS;
244 		}
245 		seg->mr_page = *ppages;
246 		seg->mr_offset = (char *)page_base;
247 		seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
248 		len -= seg->mr_len;
249 		++ppages;
250 		++seg;
251 		++n;
252 		page_base = 0;
253 	}
254 
255 	/* When encoding a Read chunk, the tail iovec contains an
256 	 * XDR pad and may be omitted.
257 	 */
258 	if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
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_ia.ri_implicit_roundup)
267 		goto out;
268 
269 	if (xdrbuf->tail[0].iov_len)
270 		seg = 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 inline int
279 encode_item_present(struct xdr_stream *xdr)
280 {
281 	__be32 *p;
282 
283 	p = xdr_reserve_space(xdr, sizeof(*p));
284 	if (unlikely(!p))
285 		return -EMSGSIZE;
286 
287 	*p = xdr_one;
288 	return 0;
289 }
290 
291 static inline int
292 encode_item_not_present(struct xdr_stream *xdr)
293 {
294 	__be32 *p;
295 
296 	p = xdr_reserve_space(xdr, sizeof(*p));
297 	if (unlikely(!p))
298 		return -EMSGSIZE;
299 
300 	*p = xdr_zero;
301 	return 0;
302 }
303 
304 static void
305 xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr *mr)
306 {
307 	*iptr++ = cpu_to_be32(mr->mr_handle);
308 	*iptr++ = cpu_to_be32(mr->mr_length);
309 	xdr_encode_hyper(iptr, mr->mr_offset);
310 }
311 
312 static int
313 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
314 {
315 	__be32 *p;
316 
317 	p = xdr_reserve_space(xdr, 4 * sizeof(*p));
318 	if (unlikely(!p))
319 		return -EMSGSIZE;
320 
321 	xdr_encode_rdma_segment(p, mr);
322 	return 0;
323 }
324 
325 static int
326 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
327 		    u32 position)
328 {
329 	__be32 *p;
330 
331 	p = xdr_reserve_space(xdr, 6 * sizeof(*p));
332 	if (unlikely(!p))
333 		return -EMSGSIZE;
334 
335 	*p++ = xdr_one;			/* Item present */
336 	*p++ = cpu_to_be32(position);
337 	xdr_encode_rdma_segment(p, mr);
338 	return 0;
339 }
340 
341 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
342 						 struct rpcrdma_req *req,
343 						 struct rpcrdma_mr_seg *seg,
344 						 int nsegs, bool writing,
345 						 struct rpcrdma_mr **mr)
346 {
347 	*mr = rpcrdma_mr_pop(&req->rl_free_mrs);
348 	if (!*mr) {
349 		*mr = rpcrdma_mr_get(r_xprt);
350 		if (!*mr)
351 			goto out_getmr_err;
352 		trace_xprtrdma_mr_get(req);
353 		(*mr)->mr_req = req;
354 	}
355 
356 	rpcrdma_mr_push(*mr, &req->rl_registered);
357 	return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
358 
359 out_getmr_err:
360 	trace_xprtrdma_nomrs(req);
361 	xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
362 	rpcrdma_mrs_refresh(r_xprt);
363 	return ERR_PTR(-EAGAIN);
364 }
365 
366 /* Register and XDR encode the Read list. Supports encoding a list of read
367  * segments that belong to a single read chunk.
368  *
369  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
370  *
371  *  Read chunklist (a linked list):
372  *   N elements, position P (same P for all chunks of same arg!):
373  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
374  *
375  * Returns zero on success, or a negative errno if a failure occurred.
376  * @xdr is advanced to the next position in the stream.
377  *
378  * Only a single @pos value is currently supported.
379  */
380 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
381 				    struct rpcrdma_req *req,
382 				    struct rpc_rqst *rqst,
383 				    enum rpcrdma_chunktype rtype)
384 {
385 	struct xdr_stream *xdr = &req->rl_stream;
386 	struct rpcrdma_mr_seg *seg;
387 	struct rpcrdma_mr *mr;
388 	unsigned int pos;
389 	int nsegs;
390 
391 	if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
392 		goto done;
393 
394 	pos = rqst->rq_snd_buf.head[0].iov_len;
395 	if (rtype == rpcrdma_areadch)
396 		pos = 0;
397 	seg = req->rl_segments;
398 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
399 				     rtype, seg);
400 	if (nsegs < 0)
401 		return nsegs;
402 
403 	do {
404 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
405 		if (IS_ERR(seg))
406 			return PTR_ERR(seg);
407 
408 		if (encode_read_segment(xdr, mr, pos) < 0)
409 			return -EMSGSIZE;
410 
411 		trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
412 		r_xprt->rx_stats.read_chunk_count++;
413 		nsegs -= mr->mr_nents;
414 	} while (nsegs);
415 
416 done:
417 	return encode_item_not_present(xdr);
418 }
419 
420 /* Register and XDR encode the Write list. Supports encoding a list
421  * containing one array of plain segments that belong to a single
422  * write chunk.
423  *
424  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
425  *
426  *  Write chunklist (a list of (one) counted array):
427  *   N elements:
428  *    1 - N - HLOO - HLOO - ... - HLOO - 0
429  *
430  * Returns zero on success, or a negative errno if a failure occurred.
431  * @xdr is advanced to the next position in the stream.
432  *
433  * Only a single Write chunk is currently supported.
434  */
435 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
436 				     struct rpcrdma_req *req,
437 				     struct rpc_rqst *rqst,
438 				     enum rpcrdma_chunktype wtype)
439 {
440 	struct xdr_stream *xdr = &req->rl_stream;
441 	struct rpcrdma_mr_seg *seg;
442 	struct rpcrdma_mr *mr;
443 	int nsegs, nchunks;
444 	__be32 *segcount;
445 
446 	if (wtype != rpcrdma_writech)
447 		goto done;
448 
449 	seg = req->rl_segments;
450 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
451 				     rqst->rq_rcv_buf.head[0].iov_len,
452 				     wtype, seg);
453 	if (nsegs < 0)
454 		return nsegs;
455 
456 	if (encode_item_present(xdr) < 0)
457 		return -EMSGSIZE;
458 	segcount = xdr_reserve_space(xdr, sizeof(*segcount));
459 	if (unlikely(!segcount))
460 		return -EMSGSIZE;
461 	/* Actual value encoded below */
462 
463 	nchunks = 0;
464 	do {
465 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
466 		if (IS_ERR(seg))
467 			return PTR_ERR(seg);
468 
469 		if (encode_rdma_segment(xdr, mr) < 0)
470 			return -EMSGSIZE;
471 
472 		trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
473 		r_xprt->rx_stats.write_chunk_count++;
474 		r_xprt->rx_stats.total_rdma_request += mr->mr_length;
475 		nchunks++;
476 		nsegs -= mr->mr_nents;
477 	} while (nsegs);
478 
479 	/* Update count of segments in this Write chunk */
480 	*segcount = cpu_to_be32(nchunks);
481 
482 done:
483 	return encode_item_not_present(xdr);
484 }
485 
486 /* Register and XDR encode the Reply chunk. Supports encoding an array
487  * of plain segments that belong to a single write (reply) chunk.
488  *
489  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
490  *
491  *  Reply chunk (a counted array):
492  *   N elements:
493  *    1 - N - HLOO - HLOO - ... - HLOO
494  *
495  * Returns zero on success, or a negative errno if a failure occurred.
496  * @xdr is advanced to the next position in the stream.
497  */
498 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
499 				      struct rpcrdma_req *req,
500 				      struct rpc_rqst *rqst,
501 				      enum rpcrdma_chunktype wtype)
502 {
503 	struct xdr_stream *xdr = &req->rl_stream;
504 	struct rpcrdma_mr_seg *seg;
505 	struct rpcrdma_mr *mr;
506 	int nsegs, nchunks;
507 	__be32 *segcount;
508 
509 	if (wtype != rpcrdma_replych)
510 		return encode_item_not_present(xdr);
511 
512 	seg = req->rl_segments;
513 	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
514 	if (nsegs < 0)
515 		return nsegs;
516 
517 	if (encode_item_present(xdr) < 0)
518 		return -EMSGSIZE;
519 	segcount = xdr_reserve_space(xdr, sizeof(*segcount));
520 	if (unlikely(!segcount))
521 		return -EMSGSIZE;
522 	/* Actual value encoded below */
523 
524 	nchunks = 0;
525 	do {
526 		seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
527 		if (IS_ERR(seg))
528 			return PTR_ERR(seg);
529 
530 		if (encode_rdma_segment(xdr, mr) < 0)
531 			return -EMSGSIZE;
532 
533 		trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
534 		r_xprt->rx_stats.reply_chunk_count++;
535 		r_xprt->rx_stats.total_rdma_request += mr->mr_length;
536 		nchunks++;
537 		nsegs -= mr->mr_nents;
538 	} while (nsegs);
539 
540 	/* Update count of segments in the Reply chunk */
541 	*segcount = cpu_to_be32(nchunks);
542 
543 	return 0;
544 }
545 
546 static void rpcrdma_sendctx_done(struct kref *kref)
547 {
548 	struct rpcrdma_req *req =
549 		container_of(kref, struct rpcrdma_req, rl_kref);
550 	struct rpcrdma_rep *rep = req->rl_reply;
551 
552 	rpcrdma_complete_rqst(rep);
553 	rep->rr_rxprt->rx_stats.reply_waits_for_send++;
554 }
555 
556 /**
557  * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
558  * @sc: sendctx containing SGEs to unmap
559  *
560  */
561 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
562 {
563 	struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
564 	struct ib_sge *sge;
565 
566 	if (!sc->sc_unmap_count)
567 		return;
568 
569 	/* The first two SGEs contain the transport header and
570 	 * the inline buffer. These are always left mapped so
571 	 * they can be cheaply re-used.
572 	 */
573 	for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
574 	     ++sge, --sc->sc_unmap_count)
575 		ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
576 				  DMA_TO_DEVICE);
577 
578 	kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
579 }
580 
581 /* Prepare an SGE for the RPC-over-RDMA transport header.
582  */
583 static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
584 				    struct rpcrdma_req *req, u32 len)
585 {
586 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
587 	struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
588 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
589 
590 	sge->addr = rdmab_addr(rb);
591 	sge->length = len;
592 	sge->lkey = rdmab_lkey(rb);
593 
594 	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
595 				      DMA_TO_DEVICE);
596 }
597 
598 /* The head iovec is straightforward, as it is usually already
599  * DMA-mapped. Sync the content that has changed.
600  */
601 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
602 				     struct rpcrdma_req *req, unsigned int len)
603 {
604 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
605 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
606 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
607 
608 	if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
609 		return false;
610 
611 	sge->addr = rdmab_addr(rb);
612 	sge->length = len;
613 	sge->lkey = rdmab_lkey(rb);
614 
615 	ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
616 				      DMA_TO_DEVICE);
617 	return true;
618 }
619 
620 /* If there is a page list present, DMA map and prepare an
621  * SGE for each page to be sent.
622  */
623 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
624 				     struct xdr_buf *xdr)
625 {
626 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
627 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
628 	unsigned int page_base, len, remaining;
629 	struct page **ppages;
630 	struct ib_sge *sge;
631 
632 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
633 	page_base = offset_in_page(xdr->page_base);
634 	remaining = xdr->page_len;
635 	while (remaining) {
636 		sge = &sc->sc_sges[req->rl_wr.num_sge++];
637 		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
638 		sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
639 					    page_base, len, DMA_TO_DEVICE);
640 		if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
641 			goto out_mapping_err;
642 
643 		sge->length = len;
644 		sge->lkey = rdmab_lkey(rb);
645 
646 		sc->sc_unmap_count++;
647 		ppages++;
648 		remaining -= len;
649 		page_base = 0;
650 	}
651 
652 	return true;
653 
654 out_mapping_err:
655 	trace_xprtrdma_dma_maperr(sge->addr);
656 	return false;
657 }
658 
659 /* The tail iovec may include an XDR pad for the page list,
660  * as well as additional content, and may not reside in the
661  * same page as the head iovec.
662  */
663 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
664 				     struct xdr_buf *xdr,
665 				     unsigned int page_base, unsigned int len)
666 {
667 	struct rpcrdma_sendctx *sc = req->rl_sendctx;
668 	struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
669 	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
670 	struct page *page = virt_to_page(xdr->tail[0].iov_base);
671 
672 	sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
673 				    DMA_TO_DEVICE);
674 	if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
675 		goto out_mapping_err;
676 
677 	sge->length = len;
678 	sge->lkey = rdmab_lkey(rb);
679 	++sc->sc_unmap_count;
680 	return true;
681 
682 out_mapping_err:
683 	trace_xprtrdma_dma_maperr(sge->addr);
684 	return false;
685 }
686 
687 /* Copy the tail to the end of the head buffer.
688  */
689 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
690 				    struct rpcrdma_req *req,
691 				    struct xdr_buf *xdr)
692 {
693 	unsigned char *dst;
694 
695 	dst = (unsigned char *)xdr->head[0].iov_base;
696 	dst += xdr->head[0].iov_len + xdr->page_len;
697 	memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
698 	r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
699 }
700 
701 /* Copy pagelist content into the head buffer.
702  */
703 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
704 				    struct rpcrdma_req *req,
705 				    struct xdr_buf *xdr)
706 {
707 	unsigned int len, page_base, remaining;
708 	struct page **ppages;
709 	unsigned char *src, *dst;
710 
711 	dst = (unsigned char *)xdr->head[0].iov_base;
712 	dst += xdr->head[0].iov_len;
713 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
714 	page_base = offset_in_page(xdr->page_base);
715 	remaining = xdr->page_len;
716 	while (remaining) {
717 		src = page_address(*ppages);
718 		src += page_base;
719 		len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
720 		memcpy(dst, src, len);
721 		r_xprt->rx_stats.pullup_copy_count += len;
722 
723 		ppages++;
724 		dst += len;
725 		remaining -= len;
726 		page_base = 0;
727 	}
728 }
729 
730 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
731  * When the head, pagelist, and tail are small, a pull-up copy
732  * is considerably less costly than DMA mapping the components
733  * of @xdr.
734  *
735  * Assumptions:
736  *  - the caller has already verified that the total length
737  *    of the RPC Call body will fit into @rl_sendbuf.
738  */
739 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
740 					struct rpcrdma_req *req,
741 					struct xdr_buf *xdr)
742 {
743 	if (unlikely(xdr->tail[0].iov_len))
744 		rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
745 
746 	if (unlikely(xdr->page_len))
747 		rpcrdma_pullup_pagelist(r_xprt, req, xdr);
748 
749 	/* The whole RPC message resides in the head iovec now */
750 	return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
751 }
752 
753 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
754 					struct rpcrdma_req *req,
755 					struct xdr_buf *xdr)
756 {
757 	struct kvec *tail = &xdr->tail[0];
758 
759 	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
760 		return false;
761 	if (xdr->page_len)
762 		if (!rpcrdma_prepare_pagelist(req, xdr))
763 			return false;
764 	if (tail->iov_len)
765 		if (!rpcrdma_prepare_tail_iov(req, xdr,
766 					      offset_in_page(tail->iov_base),
767 					      tail->iov_len))
768 			return false;
769 
770 	if (req->rl_sendctx->sc_unmap_count)
771 		kref_get(&req->rl_kref);
772 	return true;
773 }
774 
775 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
776 				   struct rpcrdma_req *req,
777 				   struct xdr_buf *xdr)
778 {
779 	if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
780 		return false;
781 
782 	/* If there is a Read chunk, the page list is being handled
783 	 * via explicit RDMA, and thus is skipped here.
784 	 */
785 
786 	/* Do not include the tail if it is only an XDR pad */
787 	if (xdr->tail[0].iov_len > 3) {
788 		unsigned int page_base, len;
789 
790 		/* If the content in the page list is an odd length,
791 		 * xdr_write_pages() adds a pad at the beginning of
792 		 * the tail iovec. Force the tail's non-pad content to
793 		 * land at the next XDR position in the Send message.
794 		 */
795 		page_base = offset_in_page(xdr->tail[0].iov_base);
796 		len = xdr->tail[0].iov_len;
797 		page_base += len & 3;
798 		len -= len & 3;
799 		if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
800 			return false;
801 		kref_get(&req->rl_kref);
802 	}
803 
804 	return true;
805 }
806 
807 /**
808  * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
809  * @r_xprt: controlling transport
810  * @req: context of RPC Call being marshalled
811  * @hdrlen: size of transport header, in bytes
812  * @xdr: xdr_buf containing RPC Call
813  * @rtype: chunk type being encoded
814  *
815  * Returns 0 on success; otherwise a negative errno is returned.
816  */
817 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
818 				     struct rpcrdma_req *req, u32 hdrlen,
819 				     struct xdr_buf *xdr,
820 				     enum rpcrdma_chunktype rtype)
821 {
822 	int ret;
823 
824 	ret = -EAGAIN;
825 	req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
826 	if (!req->rl_sendctx)
827 		goto out_nosc;
828 	req->rl_sendctx->sc_unmap_count = 0;
829 	req->rl_sendctx->sc_req = req;
830 	kref_init(&req->rl_kref);
831 	req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
832 	req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
833 	req->rl_wr.num_sge = 0;
834 	req->rl_wr.opcode = IB_WR_SEND;
835 
836 	rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
837 
838 	ret = -EIO;
839 	switch (rtype) {
840 	case rpcrdma_noch_pullup:
841 		if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
842 			goto out_unmap;
843 		break;
844 	case rpcrdma_noch_mapped:
845 		if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
846 			goto out_unmap;
847 		break;
848 	case rpcrdma_readch:
849 		if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
850 			goto out_unmap;
851 		break;
852 	case rpcrdma_areadch:
853 		break;
854 	default:
855 		goto out_unmap;
856 	}
857 
858 	return 0;
859 
860 out_unmap:
861 	rpcrdma_sendctx_unmap(req->rl_sendctx);
862 out_nosc:
863 	trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
864 	return ret;
865 }
866 
867 /**
868  * rpcrdma_marshal_req - Marshal and send one RPC request
869  * @r_xprt: controlling transport
870  * @rqst: RPC request to be marshaled
871  *
872  * For the RPC in "rqst", this function:
873  *  - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
874  *  - Registers Read, Write, and Reply chunks
875  *  - Constructs the transport header
876  *  - Posts a Send WR to send the transport header and request
877  *
878  * Returns:
879  *	%0 if the RPC was sent successfully,
880  *	%-ENOTCONN if the connection was lost,
881  *	%-EAGAIN if the caller should call again with the same arguments,
882  *	%-ENOBUFS if the caller should call again after a delay,
883  *	%-EMSGSIZE if the transport header is too small,
884  *	%-EIO if a permanent problem occurred while marshaling.
885  */
886 int
887 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
888 {
889 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
890 	struct xdr_stream *xdr = &req->rl_stream;
891 	enum rpcrdma_chunktype rtype, wtype;
892 	struct xdr_buf *buf = &rqst->rq_snd_buf;
893 	bool ddp_allowed;
894 	__be32 *p;
895 	int ret;
896 
897 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
898 	xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
899 			rqst);
900 
901 	/* Fixed header fields */
902 	ret = -EMSGSIZE;
903 	p = xdr_reserve_space(xdr, 4 * sizeof(*p));
904 	if (!p)
905 		goto out_err;
906 	*p++ = rqst->rq_xid;
907 	*p++ = rpcrdma_version;
908 	*p++ = r_xprt->rx_buf.rb_max_requests;
909 
910 	/* When the ULP employs a GSS flavor that guarantees integrity
911 	 * or privacy, direct data placement of individual data items
912 	 * is not allowed.
913 	 */
914 	ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
915 						RPCAUTH_AUTH_DATATOUCH);
916 
917 	/*
918 	 * Chunks needed for results?
919 	 *
920 	 * o If the expected result is under the inline threshold, all ops
921 	 *   return as inline.
922 	 * o Large read ops return data as write chunk(s), header as
923 	 *   inline.
924 	 * o Large non-read ops return as a single reply chunk.
925 	 */
926 	if (rpcrdma_results_inline(r_xprt, rqst))
927 		wtype = rpcrdma_noch;
928 	else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
929 		 rpcrdma_nonpayload_inline(r_xprt, rqst))
930 		wtype = rpcrdma_writech;
931 	else
932 		wtype = rpcrdma_replych;
933 
934 	/*
935 	 * Chunks needed for arguments?
936 	 *
937 	 * o If the total request is under the inline threshold, all ops
938 	 *   are sent as inline.
939 	 * o Large write ops transmit data as read chunk(s), header as
940 	 *   inline.
941 	 * o Large non-write ops are sent with the entire message as a
942 	 *   single read chunk (protocol 0-position special case).
943 	 *
944 	 * This assumes that the upper layer does not present a request
945 	 * that both has a data payload, and whose non-data arguments
946 	 * by themselves are larger than the inline threshold.
947 	 */
948 	if (rpcrdma_args_inline(r_xprt, rqst)) {
949 		*p++ = rdma_msg;
950 		rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
951 			rpcrdma_noch_pullup : rpcrdma_noch_mapped;
952 	} else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
953 		*p++ = rdma_msg;
954 		rtype = rpcrdma_readch;
955 	} else {
956 		r_xprt->rx_stats.nomsg_call_count++;
957 		*p++ = rdma_nomsg;
958 		rtype = rpcrdma_areadch;
959 	}
960 
961 	/* This implementation supports the following combinations
962 	 * of chunk lists in one RPC-over-RDMA Call message:
963 	 *
964 	 *   - Read list
965 	 *   - Write list
966 	 *   - Reply chunk
967 	 *   - Read list + Reply chunk
968 	 *
969 	 * It might not yet support the following combinations:
970 	 *
971 	 *   - Read list + Write list
972 	 *
973 	 * It does not support the following combinations:
974 	 *
975 	 *   - Write list + Reply chunk
976 	 *   - Read list + Write list + Reply chunk
977 	 *
978 	 * This implementation supports only a single chunk in each
979 	 * Read or Write list. Thus for example the client cannot
980 	 * send a Call message with a Position Zero Read chunk and a
981 	 * regular Read chunk at the same time.
982 	 */
983 	ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
984 	if (ret)
985 		goto out_err;
986 	ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
987 	if (ret)
988 		goto out_err;
989 	ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
990 	if (ret)
991 		goto out_err;
992 
993 	ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
994 					buf, rtype);
995 	if (ret)
996 		goto out_err;
997 
998 	trace_xprtrdma_marshal(req, rtype, wtype);
999 	return 0;
1000 
1001 out_err:
1002 	trace_xprtrdma_marshal_failed(rqst, ret);
1003 	r_xprt->rx_stats.failed_marshal_count++;
1004 	frwr_reset(req);
1005 	return ret;
1006 }
1007 
1008 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
1009 					 struct rpcrdma_buffer *buf,
1010 					 u32 grant)
1011 {
1012 	buf->rb_credits = grant;
1013 	xprt->cwnd = grant << RPC_CWNDSHIFT;
1014 }
1015 
1016 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
1017 {
1018 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1019 
1020 	spin_lock(&xprt->transport_lock);
1021 	__rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1022 	spin_unlock(&xprt->transport_lock);
1023 }
1024 
1025 /**
1026  * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1027  * @r_xprt: controlling transport instance
1028  *
1029  * Prepare @r_xprt for the next connection by reinitializing
1030  * its credit grant to one (see RFC 8166, Section 3.3.3).
1031  */
1032 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1033 {
1034 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1035 
1036 	spin_lock(&xprt->transport_lock);
1037 	xprt->cong = 0;
1038 	__rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1039 	spin_unlock(&xprt->transport_lock);
1040 }
1041 
1042 /**
1043  * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1044  * @rqst: controlling RPC request
1045  * @srcp: points to RPC message payload in receive buffer
1046  * @copy_len: remaining length of receive buffer content
1047  * @pad: Write chunk pad bytes needed (zero for pure inline)
1048  *
1049  * The upper layer has set the maximum number of bytes it can
1050  * receive in each component of rq_rcv_buf. These values are set in
1051  * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1052  *
1053  * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1054  * many cases this function simply updates iov_base pointers in
1055  * rq_rcv_buf to point directly to the received reply data, to
1056  * avoid copying reply data.
1057  *
1058  * Returns the count of bytes which had to be memcopied.
1059  */
1060 static unsigned long
1061 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1062 {
1063 	unsigned long fixup_copy_count;
1064 	int i, npages, curlen;
1065 	char *destp;
1066 	struct page **ppages;
1067 	int page_base;
1068 
1069 	/* The head iovec is redirected to the RPC reply message
1070 	 * in the receive buffer, to avoid a memcopy.
1071 	 */
1072 	rqst->rq_rcv_buf.head[0].iov_base = srcp;
1073 	rqst->rq_private_buf.head[0].iov_base = srcp;
1074 
1075 	/* The contents of the receive buffer that follow
1076 	 * head.iov_len bytes are copied into the page list.
1077 	 */
1078 	curlen = rqst->rq_rcv_buf.head[0].iov_len;
1079 	if (curlen > copy_len)
1080 		curlen = copy_len;
1081 	srcp += curlen;
1082 	copy_len -= curlen;
1083 
1084 	ppages = rqst->rq_rcv_buf.pages +
1085 		(rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1086 	page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1087 	fixup_copy_count = 0;
1088 	if (copy_len && rqst->rq_rcv_buf.page_len) {
1089 		int pagelist_len;
1090 
1091 		pagelist_len = rqst->rq_rcv_buf.page_len;
1092 		if (pagelist_len > copy_len)
1093 			pagelist_len = copy_len;
1094 		npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1095 		for (i = 0; i < npages; i++) {
1096 			curlen = PAGE_SIZE - page_base;
1097 			if (curlen > pagelist_len)
1098 				curlen = pagelist_len;
1099 
1100 			destp = kmap_atomic(ppages[i]);
1101 			memcpy(destp + page_base, srcp, curlen);
1102 			flush_dcache_page(ppages[i]);
1103 			kunmap_atomic(destp);
1104 			srcp += curlen;
1105 			copy_len -= curlen;
1106 			fixup_copy_count += curlen;
1107 			pagelist_len -= curlen;
1108 			if (!pagelist_len)
1109 				break;
1110 			page_base = 0;
1111 		}
1112 
1113 		/* Implicit padding for the last segment in a Write
1114 		 * chunk is inserted inline at the front of the tail
1115 		 * iovec. The upper layer ignores the content of
1116 		 * the pad. Simply ensure inline content in the tail
1117 		 * that follows the Write chunk is properly aligned.
1118 		 */
1119 		if (pad)
1120 			srcp -= pad;
1121 	}
1122 
1123 	/* The tail iovec is redirected to the remaining data
1124 	 * in the receive buffer, to avoid a memcopy.
1125 	 */
1126 	if (copy_len || pad) {
1127 		rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1128 		rqst->rq_private_buf.tail[0].iov_base = srcp;
1129 	}
1130 
1131 	if (fixup_copy_count)
1132 		trace_xprtrdma_fixup(rqst, fixup_copy_count);
1133 	return fixup_copy_count;
1134 }
1135 
1136 /* By convention, backchannel calls arrive via rdma_msg type
1137  * messages, and never populate the chunk lists. This makes
1138  * the RPC/RDMA header small and fixed in size, so it is
1139  * straightforward to check the RPC header's direction field.
1140  */
1141 static bool
1142 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1143 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1144 {
1145 	struct xdr_stream *xdr = &rep->rr_stream;
1146 	__be32 *p;
1147 
1148 	if (rep->rr_proc != rdma_msg)
1149 		return false;
1150 
1151 	/* Peek at stream contents without advancing. */
1152 	p = xdr_inline_decode(xdr, 0);
1153 
1154 	/* Chunk lists */
1155 	if (*p++ != xdr_zero)
1156 		return false;
1157 	if (*p++ != xdr_zero)
1158 		return false;
1159 	if (*p++ != xdr_zero)
1160 		return false;
1161 
1162 	/* RPC header */
1163 	if (*p++ != rep->rr_xid)
1164 		return false;
1165 	if (*p != cpu_to_be32(RPC_CALL))
1166 		return false;
1167 
1168 	/* Now that we are sure this is a backchannel call,
1169 	 * advance to the RPC header.
1170 	 */
1171 	p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1172 	if (unlikely(!p))
1173 		goto out_short;
1174 
1175 	rpcrdma_bc_receive_call(r_xprt, rep);
1176 	return true;
1177 
1178 out_short:
1179 	pr_warn("RPC/RDMA short backward direction call\n");
1180 	return true;
1181 }
1182 #else	/* CONFIG_SUNRPC_BACKCHANNEL */
1183 {
1184 	return false;
1185 }
1186 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
1187 
1188 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1189 {
1190 	u32 handle;
1191 	u64 offset;
1192 	__be32 *p;
1193 
1194 	p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1195 	if (unlikely(!p))
1196 		return -EIO;
1197 
1198 	handle = be32_to_cpup(p++);
1199 	*length = be32_to_cpup(p++);
1200 	xdr_decode_hyper(p, &offset);
1201 
1202 	trace_xprtrdma_decode_seg(handle, *length, offset);
1203 	return 0;
1204 }
1205 
1206 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1207 {
1208 	u32 segcount, seglength;
1209 	__be32 *p;
1210 
1211 	p = xdr_inline_decode(xdr, sizeof(*p));
1212 	if (unlikely(!p))
1213 		return -EIO;
1214 
1215 	*length = 0;
1216 	segcount = be32_to_cpup(p);
1217 	while (segcount--) {
1218 		if (decode_rdma_segment(xdr, &seglength))
1219 			return -EIO;
1220 		*length += seglength;
1221 	}
1222 
1223 	return 0;
1224 }
1225 
1226 /* In RPC-over-RDMA Version One replies, a Read list is never
1227  * expected. This decoder is a stub that returns an error if
1228  * a Read list is present.
1229  */
1230 static int decode_read_list(struct xdr_stream *xdr)
1231 {
1232 	__be32 *p;
1233 
1234 	p = xdr_inline_decode(xdr, sizeof(*p));
1235 	if (unlikely(!p))
1236 		return -EIO;
1237 	if (unlikely(*p != xdr_zero))
1238 		return -EIO;
1239 	return 0;
1240 }
1241 
1242 /* Supports only one Write chunk in the Write list
1243  */
1244 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1245 {
1246 	u32 chunklen;
1247 	bool first;
1248 	__be32 *p;
1249 
1250 	*length = 0;
1251 	first = true;
1252 	do {
1253 		p = xdr_inline_decode(xdr, sizeof(*p));
1254 		if (unlikely(!p))
1255 			return -EIO;
1256 		if (*p == xdr_zero)
1257 			break;
1258 		if (!first)
1259 			return -EIO;
1260 
1261 		if (decode_write_chunk(xdr, &chunklen))
1262 			return -EIO;
1263 		*length += chunklen;
1264 		first = false;
1265 	} while (true);
1266 	return 0;
1267 }
1268 
1269 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1270 {
1271 	__be32 *p;
1272 
1273 	p = xdr_inline_decode(xdr, sizeof(*p));
1274 	if (unlikely(!p))
1275 		return -EIO;
1276 
1277 	*length = 0;
1278 	if (*p != xdr_zero)
1279 		if (decode_write_chunk(xdr, length))
1280 			return -EIO;
1281 	return 0;
1282 }
1283 
1284 static int
1285 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1286 		   struct rpc_rqst *rqst)
1287 {
1288 	struct xdr_stream *xdr = &rep->rr_stream;
1289 	u32 writelist, replychunk, rpclen;
1290 	char *base;
1291 
1292 	/* Decode the chunk lists */
1293 	if (decode_read_list(xdr))
1294 		return -EIO;
1295 	if (decode_write_list(xdr, &writelist))
1296 		return -EIO;
1297 	if (decode_reply_chunk(xdr, &replychunk))
1298 		return -EIO;
1299 
1300 	/* RDMA_MSG sanity checks */
1301 	if (unlikely(replychunk))
1302 		return -EIO;
1303 
1304 	/* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1305 	base = (char *)xdr_inline_decode(xdr, 0);
1306 	rpclen = xdr_stream_remaining(xdr);
1307 	r_xprt->rx_stats.fixup_copy_count +=
1308 		rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1309 
1310 	r_xprt->rx_stats.total_rdma_reply += writelist;
1311 	return rpclen + xdr_align_size(writelist);
1312 }
1313 
1314 static noinline int
1315 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1316 {
1317 	struct xdr_stream *xdr = &rep->rr_stream;
1318 	u32 writelist, replychunk;
1319 
1320 	/* Decode the chunk lists */
1321 	if (decode_read_list(xdr))
1322 		return -EIO;
1323 	if (decode_write_list(xdr, &writelist))
1324 		return -EIO;
1325 	if (decode_reply_chunk(xdr, &replychunk))
1326 		return -EIO;
1327 
1328 	/* RDMA_NOMSG sanity checks */
1329 	if (unlikely(writelist))
1330 		return -EIO;
1331 	if (unlikely(!replychunk))
1332 		return -EIO;
1333 
1334 	/* Reply chunk buffer already is the reply vector */
1335 	r_xprt->rx_stats.total_rdma_reply += replychunk;
1336 	return replychunk;
1337 }
1338 
1339 static noinline int
1340 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1341 		     struct rpc_rqst *rqst)
1342 {
1343 	struct xdr_stream *xdr = &rep->rr_stream;
1344 	__be32 *p;
1345 
1346 	p = xdr_inline_decode(xdr, sizeof(*p));
1347 	if (unlikely(!p))
1348 		return -EIO;
1349 
1350 	switch (*p) {
1351 	case err_vers:
1352 		p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1353 		if (!p)
1354 			break;
1355 		dprintk("RPC:       %s: server reports "
1356 			"version error (%u-%u), xid %08x\n", __func__,
1357 			be32_to_cpup(p), be32_to_cpu(*(p + 1)),
1358 			be32_to_cpu(rep->rr_xid));
1359 		break;
1360 	case err_chunk:
1361 		dprintk("RPC:       %s: server reports "
1362 			"header decoding error, xid %08x\n", __func__,
1363 			be32_to_cpu(rep->rr_xid));
1364 		break;
1365 	default:
1366 		dprintk("RPC:       %s: server reports "
1367 			"unrecognized error %d, xid %08x\n", __func__,
1368 			be32_to_cpup(p), be32_to_cpu(rep->rr_xid));
1369 	}
1370 
1371 	r_xprt->rx_stats.bad_reply_count++;
1372 	return -EREMOTEIO;
1373 }
1374 
1375 /* Perform XID lookup, reconstruction of the RPC reply, and
1376  * RPC completion while holding the transport lock to ensure
1377  * the rep, rqst, and rq_task pointers remain stable.
1378  */
1379 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1380 {
1381 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1382 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1383 	struct rpc_rqst *rqst = rep->rr_rqst;
1384 	int status;
1385 
1386 	switch (rep->rr_proc) {
1387 	case rdma_msg:
1388 		status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1389 		break;
1390 	case rdma_nomsg:
1391 		status = rpcrdma_decode_nomsg(r_xprt, rep);
1392 		break;
1393 	case rdma_error:
1394 		status = rpcrdma_decode_error(r_xprt, rep, rqst);
1395 		break;
1396 	default:
1397 		status = -EIO;
1398 	}
1399 	if (status < 0)
1400 		goto out_badheader;
1401 
1402 out:
1403 	spin_lock(&xprt->queue_lock);
1404 	xprt_complete_rqst(rqst->rq_task, status);
1405 	xprt_unpin_rqst(rqst);
1406 	spin_unlock(&xprt->queue_lock);
1407 	return;
1408 
1409 /* If the incoming reply terminated a pending RPC, the next
1410  * RPC call will post a replacement receive buffer as it is
1411  * being marshaled.
1412  */
1413 out_badheader:
1414 	trace_xprtrdma_reply_hdr(rep);
1415 	r_xprt->rx_stats.bad_reply_count++;
1416 	goto out;
1417 }
1418 
1419 static void rpcrdma_reply_done(struct kref *kref)
1420 {
1421 	struct rpcrdma_req *req =
1422 		container_of(kref, struct rpcrdma_req, rl_kref);
1423 
1424 	rpcrdma_complete_rqst(req->rl_reply);
1425 }
1426 
1427 /**
1428  * rpcrdma_reply_handler - Process received RPC/RDMA messages
1429  * @rep: Incoming rpcrdma_rep object to process
1430  *
1431  * Errors must result in the RPC task either being awakened, or
1432  * allowed to timeout, to discover the errors at that time.
1433  */
1434 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1435 {
1436 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1437 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1438 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1439 	struct rpcrdma_req *req;
1440 	struct rpc_rqst *rqst;
1441 	u32 credits;
1442 	__be32 *p;
1443 
1444 	/* Any data means we had a useful conversation, so
1445 	 * then we don't need to delay the next reconnect.
1446 	 */
1447 	if (xprt->reestablish_timeout)
1448 		xprt->reestablish_timeout = 0;
1449 
1450 	/* Fixed transport header fields */
1451 	xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1452 			rep->rr_hdrbuf.head[0].iov_base, NULL);
1453 	p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1454 	if (unlikely(!p))
1455 		goto out_shortreply;
1456 	rep->rr_xid = *p++;
1457 	rep->rr_vers = *p++;
1458 	credits = be32_to_cpu(*p++);
1459 	rep->rr_proc = *p++;
1460 
1461 	if (rep->rr_vers != rpcrdma_version)
1462 		goto out_badversion;
1463 
1464 	if (rpcrdma_is_bcall(r_xprt, rep))
1465 		return;
1466 
1467 	/* Match incoming rpcrdma_rep to an rpcrdma_req to
1468 	 * get context for handling any incoming chunks.
1469 	 */
1470 	spin_lock(&xprt->queue_lock);
1471 	rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1472 	if (!rqst)
1473 		goto out_norqst;
1474 	xprt_pin_rqst(rqst);
1475 	spin_unlock(&xprt->queue_lock);
1476 
1477 	if (credits == 0)
1478 		credits = 1;	/* don't deadlock */
1479 	else if (credits > r_xprt->rx_ep.rep_max_requests)
1480 		credits = r_xprt->rx_ep.rep_max_requests;
1481 	if (buf->rb_credits != credits)
1482 		rpcrdma_update_cwnd(r_xprt, credits);
1483 	rpcrdma_post_recvs(r_xprt, false);
1484 
1485 	req = rpcr_to_rdmar(rqst);
1486 	if (req->rl_reply) {
1487 		trace_xprtrdma_leaked_rep(rqst, req->rl_reply);
1488 		rpcrdma_recv_buffer_put(req->rl_reply);
1489 	}
1490 	req->rl_reply = rep;
1491 	rep->rr_rqst = rqst;
1492 
1493 	trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1494 
1495 	if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1496 		frwr_reminv(rep, &req->rl_registered);
1497 	if (!list_empty(&req->rl_registered))
1498 		frwr_unmap_async(r_xprt, req);
1499 		/* LocalInv completion will complete the RPC */
1500 	else
1501 		kref_put(&req->rl_kref, rpcrdma_reply_done);
1502 	return;
1503 
1504 out_badversion:
1505 	trace_xprtrdma_reply_vers(rep);
1506 	goto out;
1507 
1508 out_norqst:
1509 	spin_unlock(&xprt->queue_lock);
1510 	trace_xprtrdma_reply_rqst(rep);
1511 	goto out;
1512 
1513 out_shortreply:
1514 	trace_xprtrdma_reply_short(rep);
1515 
1516 out:
1517 	rpcrdma_recv_buffer_put(rep);
1518 }
1519