Loading...
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 */
66static 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 */
89static 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 */
112void 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 */
130static 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 */
162static 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 */
172static bool
173rpcrdma_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/* 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 */
189static struct rpcrdma_mr_seg *
190rpcrdma_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
219static int
220rpcrdma_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_ep->re_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_ep->re_implicit_roundup)
267 goto out;
268
269 if (xdrbuf->tail[0].iov_len)
270 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
271
272out:
273 if (unlikely(n > RPCRDMA_MAX_SEGS))
274 return -EIO;
275 return n;
276}
277
278static int
279encode_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
291static int
292encode_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
307static 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 trace_xprtrdma_mr_get(req);
319 (*mr)->mr_req = req;
320 }
321
322 rpcrdma_mr_push(*mr, &req->rl_registered);
323 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
324
325out_getmr_err:
326 trace_xprtrdma_nomrs(req);
327 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
328 rpcrdma_mrs_refresh(r_xprt);
329 return ERR_PTR(-EAGAIN);
330}
331
332/* Register and XDR encode the Read list. Supports encoding a list of read
333 * segments that belong to a single read chunk.
334 *
335 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
336 *
337 * Read chunklist (a linked list):
338 * N elements, position P (same P for all chunks of same arg!):
339 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
340 *
341 * Returns zero on success, or a negative errno if a failure occurred.
342 * @xdr is advanced to the next position in the stream.
343 *
344 * Only a single @pos value is currently supported.
345 */
346static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
347 struct rpcrdma_req *req,
348 struct rpc_rqst *rqst,
349 enum rpcrdma_chunktype rtype)
350{
351 struct xdr_stream *xdr = &req->rl_stream;
352 struct rpcrdma_mr_seg *seg;
353 struct rpcrdma_mr *mr;
354 unsigned int pos;
355 int nsegs;
356
357 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
358 goto done;
359
360 pos = rqst->rq_snd_buf.head[0].iov_len;
361 if (rtype == rpcrdma_areadch)
362 pos = 0;
363 seg = req->rl_segments;
364 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
365 rtype, seg);
366 if (nsegs < 0)
367 return nsegs;
368
369 do {
370 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
371 if (IS_ERR(seg))
372 return PTR_ERR(seg);
373
374 if (encode_read_segment(xdr, mr, pos) < 0)
375 return -EMSGSIZE;
376
377 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
378 r_xprt->rx_stats.read_chunk_count++;
379 nsegs -= mr->mr_nents;
380 } while (nsegs);
381
382done:
383 if (xdr_stream_encode_item_absent(xdr) < 0)
384 return -EMSGSIZE;
385 return 0;
386}
387
388/* Register and XDR encode the Write list. Supports encoding a list
389 * containing one array of plain segments that belong to a single
390 * write chunk.
391 *
392 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
393 *
394 * Write chunklist (a list of (one) counted array):
395 * N elements:
396 * 1 - N - HLOO - HLOO - ... - HLOO - 0
397 *
398 * Returns zero on success, or a negative errno if a failure occurred.
399 * @xdr is advanced to the next position in the stream.
400 *
401 * Only a single Write chunk is currently supported.
402 */
403static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
404 struct rpcrdma_req *req,
405 struct rpc_rqst *rqst,
406 enum rpcrdma_chunktype wtype)
407{
408 struct xdr_stream *xdr = &req->rl_stream;
409 struct rpcrdma_mr_seg *seg;
410 struct rpcrdma_mr *mr;
411 int nsegs, nchunks;
412 __be32 *segcount;
413
414 if (wtype != rpcrdma_writech)
415 goto done;
416
417 seg = req->rl_segments;
418 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
419 rqst->rq_rcv_buf.head[0].iov_len,
420 wtype, seg);
421 if (nsegs < 0)
422 return nsegs;
423
424 if (xdr_stream_encode_item_present(xdr) < 0)
425 return -EMSGSIZE;
426 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
427 if (unlikely(!segcount))
428 return -EMSGSIZE;
429 /* Actual value encoded below */
430
431 nchunks = 0;
432 do {
433 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
434 if (IS_ERR(seg))
435 return PTR_ERR(seg);
436
437 if (encode_rdma_segment(xdr, mr) < 0)
438 return -EMSGSIZE;
439
440 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
441 r_xprt->rx_stats.write_chunk_count++;
442 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
443 nchunks++;
444 nsegs -= mr->mr_nents;
445 } while (nsegs);
446
447 /* Update count of segments in this Write chunk */
448 *segcount = cpu_to_be32(nchunks);
449
450done:
451 if (xdr_stream_encode_item_absent(xdr) < 0)
452 return -EMSGSIZE;
453 return 0;
454}
455
456/* Register and XDR encode the Reply chunk. Supports encoding an array
457 * of plain segments that belong to a single write (reply) chunk.
458 *
459 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
460 *
461 * Reply chunk (a counted array):
462 * N elements:
463 * 1 - N - HLOO - HLOO - ... - HLOO
464 *
465 * Returns zero on success, or a negative errno if a failure occurred.
466 * @xdr is advanced to the next position in the stream.
467 */
468static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
469 struct rpcrdma_req *req,
470 struct rpc_rqst *rqst,
471 enum rpcrdma_chunktype wtype)
472{
473 struct xdr_stream *xdr = &req->rl_stream;
474 struct rpcrdma_mr_seg *seg;
475 struct rpcrdma_mr *mr;
476 int nsegs, nchunks;
477 __be32 *segcount;
478
479 if (wtype != rpcrdma_replych) {
480 if (xdr_stream_encode_item_absent(xdr) < 0)
481 return -EMSGSIZE;
482 return 0;
483 }
484
485 seg = req->rl_segments;
486 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
487 if (nsegs < 0)
488 return nsegs;
489
490 if (xdr_stream_encode_item_present(xdr) < 0)
491 return -EMSGSIZE;
492 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
493 if (unlikely(!segcount))
494 return -EMSGSIZE;
495 /* Actual value encoded below */
496
497 nchunks = 0;
498 do {
499 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
500 if (IS_ERR(seg))
501 return PTR_ERR(seg);
502
503 if (encode_rdma_segment(xdr, mr) < 0)
504 return -EMSGSIZE;
505
506 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
507 r_xprt->rx_stats.reply_chunk_count++;
508 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
509 nchunks++;
510 nsegs -= mr->mr_nents;
511 } while (nsegs);
512
513 /* Update count of segments in the Reply chunk */
514 *segcount = cpu_to_be32(nchunks);
515
516 return 0;
517}
518
519static void rpcrdma_sendctx_done(struct kref *kref)
520{
521 struct rpcrdma_req *req =
522 container_of(kref, struct rpcrdma_req, rl_kref);
523 struct rpcrdma_rep *rep = req->rl_reply;
524
525 rpcrdma_complete_rqst(rep);
526 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
527}
528
529/**
530 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
531 * @sc: sendctx containing SGEs to unmap
532 *
533 */
534void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
535{
536 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
537 struct ib_sge *sge;
538
539 if (!sc->sc_unmap_count)
540 return;
541
542 /* The first two SGEs contain the transport header and
543 * the inline buffer. These are always left mapped so
544 * they can be cheaply re-used.
545 */
546 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
547 ++sge, --sc->sc_unmap_count)
548 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
549 DMA_TO_DEVICE);
550
551 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
552}
553
554/* Prepare an SGE for the RPC-over-RDMA transport header.
555 */
556static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
557 struct rpcrdma_req *req, u32 len)
558{
559 struct rpcrdma_sendctx *sc = req->rl_sendctx;
560 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
561 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
562
563 sge->addr = rdmab_addr(rb);
564 sge->length = len;
565 sge->lkey = rdmab_lkey(rb);
566
567 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
568 DMA_TO_DEVICE);
569}
570
571/* The head iovec is straightforward, as it is usually already
572 * DMA-mapped. Sync the content that has changed.
573 */
574static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
575 struct rpcrdma_req *req, unsigned int len)
576{
577 struct rpcrdma_sendctx *sc = req->rl_sendctx;
578 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
579 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
580
581 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
582 return false;
583
584 sge->addr = rdmab_addr(rb);
585 sge->length = len;
586 sge->lkey = rdmab_lkey(rb);
587
588 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
589 DMA_TO_DEVICE);
590 return true;
591}
592
593/* If there is a page list present, DMA map and prepare an
594 * SGE for each page to be sent.
595 */
596static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
597 struct xdr_buf *xdr)
598{
599 struct rpcrdma_sendctx *sc = req->rl_sendctx;
600 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
601 unsigned int page_base, len, remaining;
602 struct page **ppages;
603 struct ib_sge *sge;
604
605 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
606 page_base = offset_in_page(xdr->page_base);
607 remaining = xdr->page_len;
608 while (remaining) {
609 sge = &sc->sc_sges[req->rl_wr.num_sge++];
610 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
611 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
612 page_base, len, DMA_TO_DEVICE);
613 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
614 goto out_mapping_err;
615
616 sge->length = len;
617 sge->lkey = rdmab_lkey(rb);
618
619 sc->sc_unmap_count++;
620 ppages++;
621 remaining -= len;
622 page_base = 0;
623 }
624
625 return true;
626
627out_mapping_err:
628 trace_xprtrdma_dma_maperr(sge->addr);
629 return false;
630}
631
632/* The tail iovec may include an XDR pad for the page list,
633 * as well as additional content, and may not reside in the
634 * same page as the head iovec.
635 */
636static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
637 struct xdr_buf *xdr,
638 unsigned int page_base, unsigned int len)
639{
640 struct rpcrdma_sendctx *sc = req->rl_sendctx;
641 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
642 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
643 struct page *page = virt_to_page(xdr->tail[0].iov_base);
644
645 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
646 DMA_TO_DEVICE);
647 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
648 goto out_mapping_err;
649
650 sge->length = len;
651 sge->lkey = rdmab_lkey(rb);
652 ++sc->sc_unmap_count;
653 return true;
654
655out_mapping_err:
656 trace_xprtrdma_dma_maperr(sge->addr);
657 return false;
658}
659
660/* Copy the tail to the end of the head buffer.
661 */
662static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
663 struct rpcrdma_req *req,
664 struct xdr_buf *xdr)
665{
666 unsigned char *dst;
667
668 dst = (unsigned char *)xdr->head[0].iov_base;
669 dst += xdr->head[0].iov_len + xdr->page_len;
670 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
671 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
672}
673
674/* Copy pagelist content into the head buffer.
675 */
676static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
677 struct rpcrdma_req *req,
678 struct xdr_buf *xdr)
679{
680 unsigned int len, page_base, remaining;
681 struct page **ppages;
682 unsigned char *src, *dst;
683
684 dst = (unsigned char *)xdr->head[0].iov_base;
685 dst += xdr->head[0].iov_len;
686 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
687 page_base = offset_in_page(xdr->page_base);
688 remaining = xdr->page_len;
689 while (remaining) {
690 src = page_address(*ppages);
691 src += page_base;
692 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
693 memcpy(dst, src, len);
694 r_xprt->rx_stats.pullup_copy_count += len;
695
696 ppages++;
697 dst += len;
698 remaining -= len;
699 page_base = 0;
700 }
701}
702
703/* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
704 * When the head, pagelist, and tail are small, a pull-up copy
705 * is considerably less costly than DMA mapping the components
706 * of @xdr.
707 *
708 * Assumptions:
709 * - the caller has already verified that the total length
710 * of the RPC Call body will fit into @rl_sendbuf.
711 */
712static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
713 struct rpcrdma_req *req,
714 struct xdr_buf *xdr)
715{
716 if (unlikely(xdr->tail[0].iov_len))
717 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
718
719 if (unlikely(xdr->page_len))
720 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
721
722 /* The whole RPC message resides in the head iovec now */
723 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
724}
725
726static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
727 struct rpcrdma_req *req,
728 struct xdr_buf *xdr)
729{
730 struct kvec *tail = &xdr->tail[0];
731
732 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
733 return false;
734 if (xdr->page_len)
735 if (!rpcrdma_prepare_pagelist(req, xdr))
736 return false;
737 if (tail->iov_len)
738 if (!rpcrdma_prepare_tail_iov(req, xdr,
739 offset_in_page(tail->iov_base),
740 tail->iov_len))
741 return false;
742
743 if (req->rl_sendctx->sc_unmap_count)
744 kref_get(&req->rl_kref);
745 return true;
746}
747
748static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
749 struct rpcrdma_req *req,
750 struct xdr_buf *xdr)
751{
752 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
753 return false;
754
755 /* If there is a Read chunk, the page list is being handled
756 * via explicit RDMA, and thus is skipped here.
757 */
758
759 /* Do not include the tail if it is only an XDR pad */
760 if (xdr->tail[0].iov_len > 3) {
761 unsigned int page_base, len;
762
763 /* If the content in the page list is an odd length,
764 * xdr_write_pages() adds a pad at the beginning of
765 * the tail iovec. Force the tail's non-pad content to
766 * land at the next XDR position in the Send message.
767 */
768 page_base = offset_in_page(xdr->tail[0].iov_base);
769 len = xdr->tail[0].iov_len;
770 page_base += len & 3;
771 len -= len & 3;
772 if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
773 return false;
774 kref_get(&req->rl_kref);
775 }
776
777 return true;
778}
779
780/**
781 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
782 * @r_xprt: controlling transport
783 * @req: context of RPC Call being marshalled
784 * @hdrlen: size of transport header, in bytes
785 * @xdr: xdr_buf containing RPC Call
786 * @rtype: chunk type being encoded
787 *
788 * Returns 0 on success; otherwise a negative errno is returned.
789 */
790inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
791 struct rpcrdma_req *req, u32 hdrlen,
792 struct xdr_buf *xdr,
793 enum rpcrdma_chunktype rtype)
794{
795 int ret;
796
797 ret = -EAGAIN;
798 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
799 if (!req->rl_sendctx)
800 goto out_nosc;
801 req->rl_sendctx->sc_unmap_count = 0;
802 req->rl_sendctx->sc_req = req;
803 kref_init(&req->rl_kref);
804 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
805 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
806 req->rl_wr.num_sge = 0;
807 req->rl_wr.opcode = IB_WR_SEND;
808
809 rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
810
811 ret = -EIO;
812 switch (rtype) {
813 case rpcrdma_noch_pullup:
814 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
815 goto out_unmap;
816 break;
817 case rpcrdma_noch_mapped:
818 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
819 goto out_unmap;
820 break;
821 case rpcrdma_readch:
822 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
823 goto out_unmap;
824 break;
825 case rpcrdma_areadch:
826 break;
827 default:
828 goto out_unmap;
829 }
830
831 return 0;
832
833out_unmap:
834 rpcrdma_sendctx_unmap(req->rl_sendctx);
835out_nosc:
836 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
837 return ret;
838}
839
840/**
841 * rpcrdma_marshal_req - Marshal and send one RPC request
842 * @r_xprt: controlling transport
843 * @rqst: RPC request to be marshaled
844 *
845 * For the RPC in "rqst", this function:
846 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
847 * - Registers Read, Write, and Reply chunks
848 * - Constructs the transport header
849 * - Posts a Send WR to send the transport header and request
850 *
851 * Returns:
852 * %0 if the RPC was sent successfully,
853 * %-ENOTCONN if the connection was lost,
854 * %-EAGAIN if the caller should call again with the same arguments,
855 * %-ENOBUFS if the caller should call again after a delay,
856 * %-EMSGSIZE if the transport header is too small,
857 * %-EIO if a permanent problem occurred while marshaling.
858 */
859int
860rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
861{
862 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
863 struct xdr_stream *xdr = &req->rl_stream;
864 enum rpcrdma_chunktype rtype, wtype;
865 struct xdr_buf *buf = &rqst->rq_snd_buf;
866 bool ddp_allowed;
867 __be32 *p;
868 int ret;
869
870 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
871 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
872 rqst);
873
874 /* Fixed header fields */
875 ret = -EMSGSIZE;
876 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
877 if (!p)
878 goto out_err;
879 *p++ = rqst->rq_xid;
880 *p++ = rpcrdma_version;
881 *p++ = r_xprt->rx_buf.rb_max_requests;
882
883 /* When the ULP employs a GSS flavor that guarantees integrity
884 * or privacy, direct data placement of individual data items
885 * is not allowed.
886 */
887 ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
888 &rqst->rq_cred->cr_auth->au_flags);
889
890 /*
891 * Chunks needed for results?
892 *
893 * o If the expected result is under the inline threshold, all ops
894 * return as inline.
895 * o Large read ops return data as write chunk(s), header as
896 * inline.
897 * o Large non-read ops return as a single reply chunk.
898 */
899 if (rpcrdma_results_inline(r_xprt, rqst))
900 wtype = rpcrdma_noch;
901 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
902 rpcrdma_nonpayload_inline(r_xprt, rqst))
903 wtype = rpcrdma_writech;
904 else
905 wtype = rpcrdma_replych;
906
907 /*
908 * Chunks needed for arguments?
909 *
910 * o If the total request is under the inline threshold, all ops
911 * are sent as inline.
912 * o Large write ops transmit data as read chunk(s), header as
913 * inline.
914 * o Large non-write ops are sent with the entire message as a
915 * single read chunk (protocol 0-position special case).
916 *
917 * This assumes that the upper layer does not present a request
918 * that both has a data payload, and whose non-data arguments
919 * by themselves are larger than the inline threshold.
920 */
921 if (rpcrdma_args_inline(r_xprt, rqst)) {
922 *p++ = rdma_msg;
923 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
924 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
925 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
926 *p++ = rdma_msg;
927 rtype = rpcrdma_readch;
928 } else {
929 r_xprt->rx_stats.nomsg_call_count++;
930 *p++ = rdma_nomsg;
931 rtype = rpcrdma_areadch;
932 }
933
934 /* This implementation supports the following combinations
935 * of chunk lists in one RPC-over-RDMA Call message:
936 *
937 * - Read list
938 * - Write list
939 * - Reply chunk
940 * - Read list + Reply chunk
941 *
942 * It might not yet support the following combinations:
943 *
944 * - Read list + Write list
945 *
946 * It does not support the following combinations:
947 *
948 * - Write list + Reply chunk
949 * - Read list + Write list + Reply chunk
950 *
951 * This implementation supports only a single chunk in each
952 * Read or Write list. Thus for example the client cannot
953 * send a Call message with a Position Zero Read chunk and a
954 * regular Read chunk at the same time.
955 */
956 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
957 if (ret)
958 goto out_err;
959 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
960 if (ret)
961 goto out_err;
962 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
963 if (ret)
964 goto out_err;
965
966 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
967 buf, rtype);
968 if (ret)
969 goto out_err;
970
971 trace_xprtrdma_marshal(req, rtype, wtype);
972 return 0;
973
974out_err:
975 trace_xprtrdma_marshal_failed(rqst, ret);
976 r_xprt->rx_stats.failed_marshal_count++;
977 frwr_reset(req);
978 return ret;
979}
980
981static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
982 struct rpcrdma_buffer *buf,
983 u32 grant)
984{
985 buf->rb_credits = grant;
986 xprt->cwnd = grant << RPC_CWNDSHIFT;
987}
988
989static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
990{
991 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
992
993 spin_lock(&xprt->transport_lock);
994 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
995 spin_unlock(&xprt->transport_lock);
996}
997
998/**
999 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1000 * @r_xprt: controlling transport instance
1001 *
1002 * Prepare @r_xprt for the next connection by reinitializing
1003 * its credit grant to one (see RFC 8166, Section 3.3.3).
1004 */
1005void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1006{
1007 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1008
1009 spin_lock(&xprt->transport_lock);
1010 xprt->cong = 0;
1011 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1012 spin_unlock(&xprt->transport_lock);
1013}
1014
1015/**
1016 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1017 * @rqst: controlling RPC request
1018 * @srcp: points to RPC message payload in receive buffer
1019 * @copy_len: remaining length of receive buffer content
1020 * @pad: Write chunk pad bytes needed (zero for pure inline)
1021 *
1022 * The upper layer has set the maximum number of bytes it can
1023 * receive in each component of rq_rcv_buf. These values are set in
1024 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1025 *
1026 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1027 * many cases this function simply updates iov_base pointers in
1028 * rq_rcv_buf to point directly to the received reply data, to
1029 * avoid copying reply data.
1030 *
1031 * Returns the count of bytes which had to be memcopied.
1032 */
1033static unsigned long
1034rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1035{
1036 unsigned long fixup_copy_count;
1037 int i, npages, curlen;
1038 char *destp;
1039 struct page **ppages;
1040 int page_base;
1041
1042 /* The head iovec is redirected to the RPC reply message
1043 * in the receive buffer, to avoid a memcopy.
1044 */
1045 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1046 rqst->rq_private_buf.head[0].iov_base = srcp;
1047
1048 /* The contents of the receive buffer that follow
1049 * head.iov_len bytes are copied into the page list.
1050 */
1051 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1052 if (curlen > copy_len)
1053 curlen = copy_len;
1054 srcp += curlen;
1055 copy_len -= curlen;
1056
1057 ppages = rqst->rq_rcv_buf.pages +
1058 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1059 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1060 fixup_copy_count = 0;
1061 if (copy_len && rqst->rq_rcv_buf.page_len) {
1062 int pagelist_len;
1063
1064 pagelist_len = rqst->rq_rcv_buf.page_len;
1065 if (pagelist_len > copy_len)
1066 pagelist_len = copy_len;
1067 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1068 for (i = 0; i < npages; i++) {
1069 curlen = PAGE_SIZE - page_base;
1070 if (curlen > pagelist_len)
1071 curlen = pagelist_len;
1072
1073 destp = kmap_atomic(ppages[i]);
1074 memcpy(destp + page_base, srcp, curlen);
1075 flush_dcache_page(ppages[i]);
1076 kunmap_atomic(destp);
1077 srcp += curlen;
1078 copy_len -= curlen;
1079 fixup_copy_count += curlen;
1080 pagelist_len -= curlen;
1081 if (!pagelist_len)
1082 break;
1083 page_base = 0;
1084 }
1085
1086 /* Implicit padding for the last segment in a Write
1087 * chunk is inserted inline at the front of the tail
1088 * iovec. The upper layer ignores the content of
1089 * the pad. Simply ensure inline content in the tail
1090 * that follows the Write chunk is properly aligned.
1091 */
1092 if (pad)
1093 srcp -= pad;
1094 }
1095
1096 /* The tail iovec is redirected to the remaining data
1097 * in the receive buffer, to avoid a memcopy.
1098 */
1099 if (copy_len || pad) {
1100 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1101 rqst->rq_private_buf.tail[0].iov_base = srcp;
1102 }
1103
1104 if (fixup_copy_count)
1105 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1106 return fixup_copy_count;
1107}
1108
1109/* By convention, backchannel calls arrive via rdma_msg type
1110 * messages, and never populate the chunk lists. This makes
1111 * the RPC/RDMA header small and fixed in size, so it is
1112 * straightforward to check the RPC header's direction field.
1113 */
1114static bool
1115rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1116#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1117{
1118 struct xdr_stream *xdr = &rep->rr_stream;
1119 __be32 *p;
1120
1121 if (rep->rr_proc != rdma_msg)
1122 return false;
1123
1124 /* Peek at stream contents without advancing. */
1125 p = xdr_inline_decode(xdr, 0);
1126
1127 /* Chunk lists */
1128 if (xdr_item_is_present(p++))
1129 return false;
1130 if (xdr_item_is_present(p++))
1131 return false;
1132 if (xdr_item_is_present(p++))
1133 return false;
1134
1135 /* RPC header */
1136 if (*p++ != rep->rr_xid)
1137 return false;
1138 if (*p != cpu_to_be32(RPC_CALL))
1139 return false;
1140
1141 /* Now that we are sure this is a backchannel call,
1142 * advance to the RPC header.
1143 */
1144 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1145 if (unlikely(!p))
1146 goto out_short;
1147
1148 rpcrdma_bc_receive_call(r_xprt, rep);
1149 return true;
1150
1151out_short:
1152 pr_warn("RPC/RDMA short backward direction call\n");
1153 return true;
1154}
1155#else /* CONFIG_SUNRPC_BACKCHANNEL */
1156{
1157 return false;
1158}
1159#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1160
1161static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1162{
1163 u32 handle;
1164 u64 offset;
1165 __be32 *p;
1166
1167 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1168 if (unlikely(!p))
1169 return -EIO;
1170
1171 xdr_decode_rdma_segment(p, &handle, length, &offset);
1172 trace_xprtrdma_decode_seg(handle, *length, offset);
1173 return 0;
1174}
1175
1176static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1177{
1178 u32 segcount, seglength;
1179 __be32 *p;
1180
1181 p = xdr_inline_decode(xdr, sizeof(*p));
1182 if (unlikely(!p))
1183 return -EIO;
1184
1185 *length = 0;
1186 segcount = be32_to_cpup(p);
1187 while (segcount--) {
1188 if (decode_rdma_segment(xdr, &seglength))
1189 return -EIO;
1190 *length += seglength;
1191 }
1192
1193 return 0;
1194}
1195
1196/* In RPC-over-RDMA Version One replies, a Read list is never
1197 * expected. This decoder is a stub that returns an error if
1198 * a Read list is present.
1199 */
1200static int decode_read_list(struct xdr_stream *xdr)
1201{
1202 __be32 *p;
1203
1204 p = xdr_inline_decode(xdr, sizeof(*p));
1205 if (unlikely(!p))
1206 return -EIO;
1207 if (unlikely(xdr_item_is_present(p)))
1208 return -EIO;
1209 return 0;
1210}
1211
1212/* Supports only one Write chunk in the Write list
1213 */
1214static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1215{
1216 u32 chunklen;
1217 bool first;
1218 __be32 *p;
1219
1220 *length = 0;
1221 first = true;
1222 do {
1223 p = xdr_inline_decode(xdr, sizeof(*p));
1224 if (unlikely(!p))
1225 return -EIO;
1226 if (xdr_item_is_absent(p))
1227 break;
1228 if (!first)
1229 return -EIO;
1230
1231 if (decode_write_chunk(xdr, &chunklen))
1232 return -EIO;
1233 *length += chunklen;
1234 first = false;
1235 } while (true);
1236 return 0;
1237}
1238
1239static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1240{
1241 __be32 *p;
1242
1243 p = xdr_inline_decode(xdr, sizeof(*p));
1244 if (unlikely(!p))
1245 return -EIO;
1246
1247 *length = 0;
1248 if (xdr_item_is_present(p))
1249 if (decode_write_chunk(xdr, length))
1250 return -EIO;
1251 return 0;
1252}
1253
1254static int
1255rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1256 struct rpc_rqst *rqst)
1257{
1258 struct xdr_stream *xdr = &rep->rr_stream;
1259 u32 writelist, replychunk, rpclen;
1260 char *base;
1261
1262 /* Decode the chunk lists */
1263 if (decode_read_list(xdr))
1264 return -EIO;
1265 if (decode_write_list(xdr, &writelist))
1266 return -EIO;
1267 if (decode_reply_chunk(xdr, &replychunk))
1268 return -EIO;
1269
1270 /* RDMA_MSG sanity checks */
1271 if (unlikely(replychunk))
1272 return -EIO;
1273
1274 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1275 base = (char *)xdr_inline_decode(xdr, 0);
1276 rpclen = xdr_stream_remaining(xdr);
1277 r_xprt->rx_stats.fixup_copy_count +=
1278 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1279
1280 r_xprt->rx_stats.total_rdma_reply += writelist;
1281 return rpclen + xdr_align_size(writelist);
1282}
1283
1284static noinline int
1285rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1286{
1287 struct xdr_stream *xdr = &rep->rr_stream;
1288 u32 writelist, replychunk;
1289
1290 /* Decode the chunk lists */
1291 if (decode_read_list(xdr))
1292 return -EIO;
1293 if (decode_write_list(xdr, &writelist))
1294 return -EIO;
1295 if (decode_reply_chunk(xdr, &replychunk))
1296 return -EIO;
1297
1298 /* RDMA_NOMSG sanity checks */
1299 if (unlikely(writelist))
1300 return -EIO;
1301 if (unlikely(!replychunk))
1302 return -EIO;
1303
1304 /* Reply chunk buffer already is the reply vector */
1305 r_xprt->rx_stats.total_rdma_reply += replychunk;
1306 return replychunk;
1307}
1308
1309static noinline int
1310rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1311 struct rpc_rqst *rqst)
1312{
1313 struct xdr_stream *xdr = &rep->rr_stream;
1314 __be32 *p;
1315
1316 p = xdr_inline_decode(xdr, sizeof(*p));
1317 if (unlikely(!p))
1318 return -EIO;
1319
1320 switch (*p) {
1321 case err_vers:
1322 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1323 if (!p)
1324 break;
1325 dprintk("RPC: %s: server reports "
1326 "version error (%u-%u), xid %08x\n", __func__,
1327 be32_to_cpup(p), be32_to_cpu(*(p + 1)),
1328 be32_to_cpu(rep->rr_xid));
1329 break;
1330 case err_chunk:
1331 dprintk("RPC: %s: server reports "
1332 "header decoding error, xid %08x\n", __func__,
1333 be32_to_cpu(rep->rr_xid));
1334 break;
1335 default:
1336 dprintk("RPC: %s: server reports "
1337 "unrecognized error %d, xid %08x\n", __func__,
1338 be32_to_cpup(p), be32_to_cpu(rep->rr_xid));
1339 }
1340
1341 return -EIO;
1342}
1343
1344/* Perform XID lookup, reconstruction of the RPC reply, and
1345 * RPC completion while holding the transport lock to ensure
1346 * the rep, rqst, and rq_task pointers remain stable.
1347 */
1348void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1349{
1350 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1351 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1352 struct rpc_rqst *rqst = rep->rr_rqst;
1353 int status;
1354
1355 switch (rep->rr_proc) {
1356 case rdma_msg:
1357 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1358 break;
1359 case rdma_nomsg:
1360 status = rpcrdma_decode_nomsg(r_xprt, rep);
1361 break;
1362 case rdma_error:
1363 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1364 break;
1365 default:
1366 status = -EIO;
1367 }
1368 if (status < 0)
1369 goto out_badheader;
1370
1371out:
1372 spin_lock(&xprt->queue_lock);
1373 xprt_complete_rqst(rqst->rq_task, status);
1374 xprt_unpin_rqst(rqst);
1375 spin_unlock(&xprt->queue_lock);
1376 return;
1377
1378out_badheader:
1379 trace_xprtrdma_reply_hdr(rep);
1380 r_xprt->rx_stats.bad_reply_count++;
1381 rqst->rq_task->tk_status = status;
1382 status = 0;
1383 goto out;
1384}
1385
1386static void rpcrdma_reply_done(struct kref *kref)
1387{
1388 struct rpcrdma_req *req =
1389 container_of(kref, struct rpcrdma_req, rl_kref);
1390
1391 rpcrdma_complete_rqst(req->rl_reply);
1392}
1393
1394/**
1395 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1396 * @rep: Incoming rpcrdma_rep object to process
1397 *
1398 * Errors must result in the RPC task either being awakened, or
1399 * allowed to timeout, to discover the errors at that time.
1400 */
1401void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1402{
1403 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1404 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1405 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1406 struct rpcrdma_req *req;
1407 struct rpc_rqst *rqst;
1408 u32 credits;
1409 __be32 *p;
1410
1411 /* Any data means we had a useful conversation, so
1412 * then we don't need to delay the next reconnect.
1413 */
1414 if (xprt->reestablish_timeout)
1415 xprt->reestablish_timeout = 0;
1416
1417 /* Fixed transport header fields */
1418 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1419 rep->rr_hdrbuf.head[0].iov_base, NULL);
1420 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1421 if (unlikely(!p))
1422 goto out_shortreply;
1423 rep->rr_xid = *p++;
1424 rep->rr_vers = *p++;
1425 credits = be32_to_cpu(*p++);
1426 rep->rr_proc = *p++;
1427
1428 if (rep->rr_vers != rpcrdma_version)
1429 goto out_badversion;
1430
1431 if (rpcrdma_is_bcall(r_xprt, rep))
1432 return;
1433
1434 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1435 * get context for handling any incoming chunks.
1436 */
1437 spin_lock(&xprt->queue_lock);
1438 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1439 if (!rqst)
1440 goto out_norqst;
1441 xprt_pin_rqst(rqst);
1442 spin_unlock(&xprt->queue_lock);
1443
1444 if (credits == 0)
1445 credits = 1; /* don't deadlock */
1446 else if (credits > r_xprt->rx_ep->re_max_requests)
1447 credits = r_xprt->rx_ep->re_max_requests;
1448 if (buf->rb_credits != credits)
1449 rpcrdma_update_cwnd(r_xprt, credits);
1450 rpcrdma_post_recvs(r_xprt, false);
1451
1452 req = rpcr_to_rdmar(rqst);
1453 if (req->rl_reply) {
1454 trace_xprtrdma_leaked_rep(rqst, req->rl_reply);
1455 rpcrdma_recv_buffer_put(req->rl_reply);
1456 }
1457 req->rl_reply = rep;
1458 rep->rr_rqst = rqst;
1459
1460 trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1461
1462 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1463 frwr_reminv(rep, &req->rl_registered);
1464 if (!list_empty(&req->rl_registered))
1465 frwr_unmap_async(r_xprt, req);
1466 /* LocalInv completion will complete the RPC */
1467 else
1468 kref_put(&req->rl_kref, rpcrdma_reply_done);
1469 return;
1470
1471out_badversion:
1472 trace_xprtrdma_reply_vers(rep);
1473 goto out;
1474
1475out_norqst:
1476 spin_unlock(&xprt->queue_lock);
1477 trace_xprtrdma_reply_rqst(rep);
1478 goto out;
1479
1480out_shortreply:
1481 trace_xprtrdma_reply_short(rep);
1482
1483out:
1484 rpcrdma_recv_buffer_put(rep);
1485}
1/*
2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
8 * license below:
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 *
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 *
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
21 *
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
25 * permission.
26 *
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38 */
39
40/*
41 * rpc_rdma.c
42 *
43 * This file contains the guts of the RPC RDMA protocol, and
44 * does marshaling/unmarshaling, etc. It is also where interfacing
45 * to the Linux RPC framework lives.
46 */
47
48#include "xprt_rdma.h"
49
50#include <linux/highmem.h>
51
52#ifdef RPC_DEBUG
53# define RPCDBG_FACILITY RPCDBG_TRANS
54#endif
55
56enum rpcrdma_chunktype {
57 rpcrdma_noch = 0,
58 rpcrdma_readch,
59 rpcrdma_areadch,
60 rpcrdma_writech,
61 rpcrdma_replych
62};
63
64#ifdef RPC_DEBUG
65static const char transfertypes[][12] = {
66 "pure inline", /* no chunks */
67 " read chunk", /* some argument via rdma read */
68 "*read chunk", /* entire request via rdma read */
69 "write chunk", /* some result via rdma write */
70 "reply chunk" /* entire reply via rdma write */
71};
72#endif
73
74/*
75 * Chunk assembly from upper layer xdr_buf.
76 *
77 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
78 * elements. Segments are then coalesced when registered, if possible
79 * within the selected memreg mode.
80 *
81 * Note, this routine is never called if the connection's memory
82 * registration strategy is 0 (bounce buffers).
83 */
84
85static int
86rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
87 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
88{
89 int len, n = 0, p;
90 int page_base;
91 struct page **ppages;
92
93 if (pos == 0 && xdrbuf->head[0].iov_len) {
94 seg[n].mr_page = NULL;
95 seg[n].mr_offset = xdrbuf->head[0].iov_base;
96 seg[n].mr_len = xdrbuf->head[0].iov_len;
97 ++n;
98 }
99
100 len = xdrbuf->page_len;
101 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
102 page_base = xdrbuf->page_base & ~PAGE_MASK;
103 p = 0;
104 while (len && n < nsegs) {
105 seg[n].mr_page = ppages[p];
106 seg[n].mr_offset = (void *)(unsigned long) page_base;
107 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
108 BUG_ON(seg[n].mr_len > PAGE_SIZE);
109 len -= seg[n].mr_len;
110 ++n;
111 ++p;
112 page_base = 0; /* page offset only applies to first page */
113 }
114
115 /* Message overflows the seg array */
116 if (len && n == nsegs)
117 return 0;
118
119 if (xdrbuf->tail[0].iov_len) {
120 /* the rpcrdma protocol allows us to omit any trailing
121 * xdr pad bytes, saving the server an RDMA operation. */
122 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
123 return n;
124 if (n == nsegs)
125 /* Tail remains, but we're out of segments */
126 return 0;
127 seg[n].mr_page = NULL;
128 seg[n].mr_offset = xdrbuf->tail[0].iov_base;
129 seg[n].mr_len = xdrbuf->tail[0].iov_len;
130 ++n;
131 }
132
133 return n;
134}
135
136/*
137 * Create read/write chunk lists, and reply chunks, for RDMA
138 *
139 * Assume check against THRESHOLD has been done, and chunks are required.
140 * Assume only encoding one list entry for read|write chunks. The NFSv3
141 * protocol is simple enough to allow this as it only has a single "bulk
142 * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
143 * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
144 *
145 * When used for a single reply chunk (which is a special write
146 * chunk used for the entire reply, rather than just the data), it
147 * is used primarily for READDIR and READLINK which would otherwise
148 * be severely size-limited by a small rdma inline read max. The server
149 * response will come back as an RDMA Write, followed by a message
150 * of type RDMA_NOMSG carrying the xid and length. As a result, reply
151 * chunks do not provide data alignment, however they do not require
152 * "fixup" (moving the response to the upper layer buffer) either.
153 *
154 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
155 *
156 * Read chunklist (a linked list):
157 * N elements, position P (same P for all chunks of same arg!):
158 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
159 *
160 * Write chunklist (a list of (one) counted array):
161 * N elements:
162 * 1 - N - HLOO - HLOO - ... - HLOO - 0
163 *
164 * Reply chunk (a counted array):
165 * N elements:
166 * 1 - N - HLOO - HLOO - ... - HLOO
167 */
168
169static unsigned int
170rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
171 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
172{
173 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
174 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_task->tk_xprt);
175 int nsegs, nchunks = 0;
176 unsigned int pos;
177 struct rpcrdma_mr_seg *seg = req->rl_segments;
178 struct rpcrdma_read_chunk *cur_rchunk = NULL;
179 struct rpcrdma_write_array *warray = NULL;
180 struct rpcrdma_write_chunk *cur_wchunk = NULL;
181 __be32 *iptr = headerp->rm_body.rm_chunks;
182
183 if (type == rpcrdma_readch || type == rpcrdma_areadch) {
184 /* a read chunk - server will RDMA Read our memory */
185 cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
186 } else {
187 /* a write or reply chunk - server will RDMA Write our memory */
188 *iptr++ = xdr_zero; /* encode a NULL read chunk list */
189 if (type == rpcrdma_replych)
190 *iptr++ = xdr_zero; /* a NULL write chunk list */
191 warray = (struct rpcrdma_write_array *) iptr;
192 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
193 }
194
195 if (type == rpcrdma_replych || type == rpcrdma_areadch)
196 pos = 0;
197 else
198 pos = target->head[0].iov_len;
199
200 nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
201 if (nsegs == 0)
202 return 0;
203
204 do {
205 /* bind/register the memory, then build chunk from result. */
206 int n = rpcrdma_register_external(seg, nsegs,
207 cur_wchunk != NULL, r_xprt);
208 if (n <= 0)
209 goto out;
210 if (cur_rchunk) { /* read */
211 cur_rchunk->rc_discrim = xdr_one;
212 /* all read chunks have the same "position" */
213 cur_rchunk->rc_position = htonl(pos);
214 cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey);
215 cur_rchunk->rc_target.rs_length = htonl(seg->mr_len);
216 xdr_encode_hyper(
217 (__be32 *)&cur_rchunk->rc_target.rs_offset,
218 seg->mr_base);
219 dprintk("RPC: %s: read chunk "
220 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
221 seg->mr_len, (unsigned long long)seg->mr_base,
222 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
223 cur_rchunk++;
224 r_xprt->rx_stats.read_chunk_count++;
225 } else { /* write/reply */
226 cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey);
227 cur_wchunk->wc_target.rs_length = htonl(seg->mr_len);
228 xdr_encode_hyper(
229 (__be32 *)&cur_wchunk->wc_target.rs_offset,
230 seg->mr_base);
231 dprintk("RPC: %s: %s chunk "
232 "elem %d@0x%llx:0x%x (%s)\n", __func__,
233 (type == rpcrdma_replych) ? "reply" : "write",
234 seg->mr_len, (unsigned long long)seg->mr_base,
235 seg->mr_rkey, n < nsegs ? "more" : "last");
236 cur_wchunk++;
237 if (type == rpcrdma_replych)
238 r_xprt->rx_stats.reply_chunk_count++;
239 else
240 r_xprt->rx_stats.write_chunk_count++;
241 r_xprt->rx_stats.total_rdma_request += seg->mr_len;
242 }
243 nchunks++;
244 seg += n;
245 nsegs -= n;
246 } while (nsegs);
247
248 /* success. all failures return above */
249 req->rl_nchunks = nchunks;
250
251 BUG_ON(nchunks == 0);
252 BUG_ON((r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR)
253 && (nchunks > 3));
254
255 /*
256 * finish off header. If write, marshal discrim and nchunks.
257 */
258 if (cur_rchunk) {
259 iptr = (__be32 *) cur_rchunk;
260 *iptr++ = xdr_zero; /* finish the read chunk list */
261 *iptr++ = xdr_zero; /* encode a NULL write chunk list */
262 *iptr++ = xdr_zero; /* encode a NULL reply chunk */
263 } else {
264 warray->wc_discrim = xdr_one;
265 warray->wc_nchunks = htonl(nchunks);
266 iptr = (__be32 *) cur_wchunk;
267 if (type == rpcrdma_writech) {
268 *iptr++ = xdr_zero; /* finish the write chunk list */
269 *iptr++ = xdr_zero; /* encode a NULL reply chunk */
270 }
271 }
272
273 /*
274 * Return header size.
275 */
276 return (unsigned char *)iptr - (unsigned char *)headerp;
277
278out:
279 for (pos = 0; nchunks--;)
280 pos += rpcrdma_deregister_external(
281 &req->rl_segments[pos], r_xprt, NULL);
282 return 0;
283}
284
285/*
286 * Copy write data inline.
287 * This function is used for "small" requests. Data which is passed
288 * to RPC via iovecs (or page list) is copied directly into the
289 * pre-registered memory buffer for this request. For small amounts
290 * of data, this is efficient. The cutoff value is tunable.
291 */
292static int
293rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
294{
295 int i, npages, curlen;
296 int copy_len;
297 unsigned char *srcp, *destp;
298 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
299 int page_base;
300 struct page **ppages;
301
302 destp = rqst->rq_svec[0].iov_base;
303 curlen = rqst->rq_svec[0].iov_len;
304 destp += curlen;
305 /*
306 * Do optional padding where it makes sense. Alignment of write
307 * payload can help the server, if our setting is accurate.
308 */
309 pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
310 if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
311 pad = 0; /* don't pad this request */
312
313 dprintk("RPC: %s: pad %d destp 0x%p len %d hdrlen %d\n",
314 __func__, pad, destp, rqst->rq_slen, curlen);
315
316 copy_len = rqst->rq_snd_buf.page_len;
317
318 if (rqst->rq_snd_buf.tail[0].iov_len) {
319 curlen = rqst->rq_snd_buf.tail[0].iov_len;
320 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
321 memmove(destp + copy_len,
322 rqst->rq_snd_buf.tail[0].iov_base, curlen);
323 r_xprt->rx_stats.pullup_copy_count += curlen;
324 }
325 dprintk("RPC: %s: tail destp 0x%p len %d\n",
326 __func__, destp + copy_len, curlen);
327 rqst->rq_svec[0].iov_len += curlen;
328 }
329 r_xprt->rx_stats.pullup_copy_count += copy_len;
330
331 page_base = rqst->rq_snd_buf.page_base;
332 ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
333 page_base &= ~PAGE_MASK;
334 npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
335 for (i = 0; copy_len && i < npages; i++) {
336 curlen = PAGE_SIZE - page_base;
337 if (curlen > copy_len)
338 curlen = copy_len;
339 dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n",
340 __func__, i, destp, copy_len, curlen);
341 srcp = kmap_atomic(ppages[i]);
342 memcpy(destp, srcp+page_base, curlen);
343 kunmap_atomic(srcp);
344 rqst->rq_svec[0].iov_len += curlen;
345 destp += curlen;
346 copy_len -= curlen;
347 page_base = 0;
348 }
349 /* header now contains entire send message */
350 return pad;
351}
352
353/*
354 * Marshal a request: the primary job of this routine is to choose
355 * the transfer modes. See comments below.
356 *
357 * Uses multiple RDMA IOVs for a request:
358 * [0] -- RPC RDMA header, which uses memory from the *start* of the
359 * preregistered buffer that already holds the RPC data in
360 * its middle.
361 * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
362 * [2] -- optional padding.
363 * [3] -- if padded, header only in [1] and data here.
364 */
365
366int
367rpcrdma_marshal_req(struct rpc_rqst *rqst)
368{
369 struct rpc_xprt *xprt = rqst->rq_task->tk_xprt;
370 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
371 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
372 char *base;
373 size_t hdrlen, rpclen, padlen;
374 enum rpcrdma_chunktype rtype, wtype;
375 struct rpcrdma_msg *headerp;
376
377 /*
378 * rpclen gets amount of data in first buffer, which is the
379 * pre-registered buffer.
380 */
381 base = rqst->rq_svec[0].iov_base;
382 rpclen = rqst->rq_svec[0].iov_len;
383
384 /* build RDMA header in private area at front */
385 headerp = (struct rpcrdma_msg *) req->rl_base;
386 /* don't htonl XID, it's already done in request */
387 headerp->rm_xid = rqst->rq_xid;
388 headerp->rm_vers = xdr_one;
389 headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
390 headerp->rm_type = htonl(RDMA_MSG);
391
392 /*
393 * Chunks needed for results?
394 *
395 * o If the expected result is under the inline threshold, all ops
396 * return as inline (but see later).
397 * o Large non-read ops return as a single reply chunk.
398 * o Large read ops return data as write chunk(s), header as inline.
399 *
400 * Note: the NFS code sending down multiple result segments implies
401 * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
402 */
403
404 /*
405 * This code can handle read chunks, write chunks OR reply
406 * chunks -- only one type. If the request is too big to fit
407 * inline, then we will choose read chunks. If the request is
408 * a READ, then use write chunks to separate the file data
409 * into pages; otherwise use reply chunks.
410 */
411 if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
412 wtype = rpcrdma_noch;
413 else if (rqst->rq_rcv_buf.page_len == 0)
414 wtype = rpcrdma_replych;
415 else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
416 wtype = rpcrdma_writech;
417 else
418 wtype = rpcrdma_replych;
419
420 /*
421 * Chunks needed for arguments?
422 *
423 * o If the total request is under the inline threshold, all ops
424 * are sent as inline.
425 * o Large non-write ops are sent with the entire message as a
426 * single read chunk (protocol 0-position special case).
427 * o Large write ops transmit data as read chunk(s), header as
428 * inline.
429 *
430 * Note: the NFS code sending down multiple argument segments
431 * implies the op is a write.
432 * TBD check NFSv4 setacl
433 */
434 if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
435 rtype = rpcrdma_noch;
436 else if (rqst->rq_snd_buf.page_len == 0)
437 rtype = rpcrdma_areadch;
438 else
439 rtype = rpcrdma_readch;
440
441 /* The following simplification is not true forever */
442 if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
443 wtype = rpcrdma_noch;
444 BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch);
445
446 if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS &&
447 (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) {
448 /* forced to "pure inline"? */
449 dprintk("RPC: %s: too much data (%d/%d) for inline\n",
450 __func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len);
451 return -1;
452 }
453
454 hdrlen = 28; /*sizeof *headerp;*/
455 padlen = 0;
456
457 /*
458 * Pull up any extra send data into the preregistered buffer.
459 * When padding is in use and applies to the transfer, insert
460 * it and change the message type.
461 */
462 if (rtype == rpcrdma_noch) {
463
464 padlen = rpcrdma_inline_pullup(rqst,
465 RPCRDMA_INLINE_PAD_VALUE(rqst));
466
467 if (padlen) {
468 headerp->rm_type = htonl(RDMA_MSGP);
469 headerp->rm_body.rm_padded.rm_align =
470 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
471 headerp->rm_body.rm_padded.rm_thresh =
472 htonl(RPCRDMA_INLINE_PAD_THRESH);
473 headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
474 headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
475 headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
476 hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
477 BUG_ON(wtype != rpcrdma_noch);
478
479 } else {
480 headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
481 headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
482 headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
483 /* new length after pullup */
484 rpclen = rqst->rq_svec[0].iov_len;
485 /*
486 * Currently we try to not actually use read inline.
487 * Reply chunks have the desirable property that
488 * they land, packed, directly in the target buffers
489 * without headers, so they require no fixup. The
490 * additional RDMA Write op sends the same amount
491 * of data, streams on-the-wire and adds no overhead
492 * on receive. Therefore, we request a reply chunk
493 * for non-writes wherever feasible and efficient.
494 */
495 if (wtype == rpcrdma_noch &&
496 r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
497 wtype = rpcrdma_replych;
498 }
499 }
500
501 /*
502 * Marshal chunks. This routine will return the header length
503 * consumed by marshaling.
504 */
505 if (rtype != rpcrdma_noch) {
506 hdrlen = rpcrdma_create_chunks(rqst,
507 &rqst->rq_snd_buf, headerp, rtype);
508 wtype = rtype; /* simplify dprintk */
509
510 } else if (wtype != rpcrdma_noch) {
511 hdrlen = rpcrdma_create_chunks(rqst,
512 &rqst->rq_rcv_buf, headerp, wtype);
513 }
514
515 if (hdrlen == 0)
516 return -1;
517
518 dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd padlen %zd"
519 " headerp 0x%p base 0x%p lkey 0x%x\n",
520 __func__, transfertypes[wtype], hdrlen, rpclen, padlen,
521 headerp, base, req->rl_iov.lkey);
522
523 /*
524 * initialize send_iov's - normally only two: rdma chunk header and
525 * single preregistered RPC header buffer, but if padding is present,
526 * then use a preregistered (and zeroed) pad buffer between the RPC
527 * header and any write data. In all non-rdma cases, any following
528 * data has been copied into the RPC header buffer.
529 */
530 req->rl_send_iov[0].addr = req->rl_iov.addr;
531 req->rl_send_iov[0].length = hdrlen;
532 req->rl_send_iov[0].lkey = req->rl_iov.lkey;
533
534 req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
535 req->rl_send_iov[1].length = rpclen;
536 req->rl_send_iov[1].lkey = req->rl_iov.lkey;
537
538 req->rl_niovs = 2;
539
540 if (padlen) {
541 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
542
543 req->rl_send_iov[2].addr = ep->rep_pad.addr;
544 req->rl_send_iov[2].length = padlen;
545 req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
546
547 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
548 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
549 req->rl_send_iov[3].lkey = req->rl_iov.lkey;
550
551 req->rl_niovs = 4;
552 }
553
554 return 0;
555}
556
557/*
558 * Chase down a received write or reply chunklist to get length
559 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
560 */
561static int
562rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
563{
564 unsigned int i, total_len;
565 struct rpcrdma_write_chunk *cur_wchunk;
566
567 i = ntohl(**iptrp); /* get array count */
568 if (i > max)
569 return -1;
570 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
571 total_len = 0;
572 while (i--) {
573 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
574 ifdebug(FACILITY) {
575 u64 off;
576 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
577 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n",
578 __func__,
579 ntohl(seg->rs_length),
580 (unsigned long long)off,
581 ntohl(seg->rs_handle));
582 }
583 total_len += ntohl(seg->rs_length);
584 ++cur_wchunk;
585 }
586 /* check and adjust for properly terminated write chunk */
587 if (wrchunk) {
588 __be32 *w = (__be32 *) cur_wchunk;
589 if (*w++ != xdr_zero)
590 return -1;
591 cur_wchunk = (struct rpcrdma_write_chunk *) w;
592 }
593 if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
594 return -1;
595
596 *iptrp = (__be32 *) cur_wchunk;
597 return total_len;
598}
599
600/*
601 * Scatter inline received data back into provided iov's.
602 */
603static void
604rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
605{
606 int i, npages, curlen, olen;
607 char *destp;
608 struct page **ppages;
609 int page_base;
610
611 curlen = rqst->rq_rcv_buf.head[0].iov_len;
612 if (curlen > copy_len) { /* write chunk header fixup */
613 curlen = copy_len;
614 rqst->rq_rcv_buf.head[0].iov_len = curlen;
615 }
616
617 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n",
618 __func__, srcp, copy_len, curlen);
619
620 /* Shift pointer for first receive segment only */
621 rqst->rq_rcv_buf.head[0].iov_base = srcp;
622 srcp += curlen;
623 copy_len -= curlen;
624
625 olen = copy_len;
626 i = 0;
627 rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
628 page_base = rqst->rq_rcv_buf.page_base;
629 ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
630 page_base &= ~PAGE_MASK;
631
632 if (copy_len && rqst->rq_rcv_buf.page_len) {
633 npages = PAGE_ALIGN(page_base +
634 rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
635 for (; i < npages; i++) {
636 curlen = PAGE_SIZE - page_base;
637 if (curlen > copy_len)
638 curlen = copy_len;
639 dprintk("RPC: %s: page %d"
640 " srcp 0x%p len %d curlen %d\n",
641 __func__, i, srcp, copy_len, curlen);
642 destp = kmap_atomic(ppages[i]);
643 memcpy(destp + page_base, srcp, curlen);
644 flush_dcache_page(ppages[i]);
645 kunmap_atomic(destp);
646 srcp += curlen;
647 copy_len -= curlen;
648 if (copy_len == 0)
649 break;
650 page_base = 0;
651 }
652 rqst->rq_rcv_buf.page_len = olen - copy_len;
653 } else
654 rqst->rq_rcv_buf.page_len = 0;
655
656 if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
657 curlen = copy_len;
658 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
659 curlen = rqst->rq_rcv_buf.tail[0].iov_len;
660 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
661 memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
662 dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n",
663 __func__, srcp, copy_len, curlen);
664 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
665 copy_len -= curlen; ++i;
666 } else
667 rqst->rq_rcv_buf.tail[0].iov_len = 0;
668
669 if (pad) {
670 /* implicit padding on terminal chunk */
671 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
672 while (pad--)
673 p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
674 }
675
676 if (copy_len)
677 dprintk("RPC: %s: %d bytes in"
678 " %d extra segments (%d lost)\n",
679 __func__, olen, i, copy_len);
680
681 /* TBD avoid a warning from call_decode() */
682 rqst->rq_private_buf = rqst->rq_rcv_buf;
683}
684
685/*
686 * This function is called when an async event is posted to
687 * the connection which changes the connection state. All it
688 * does at this point is mark the connection up/down, the rpc
689 * timers do the rest.
690 */
691void
692rpcrdma_conn_func(struct rpcrdma_ep *ep)
693{
694 struct rpc_xprt *xprt = ep->rep_xprt;
695
696 spin_lock_bh(&xprt->transport_lock);
697 if (++xprt->connect_cookie == 0) /* maintain a reserved value */
698 ++xprt->connect_cookie;
699 if (ep->rep_connected > 0) {
700 if (!xprt_test_and_set_connected(xprt))
701 xprt_wake_pending_tasks(xprt, 0);
702 } else {
703 if (xprt_test_and_clear_connected(xprt))
704 xprt_wake_pending_tasks(xprt, -ENOTCONN);
705 }
706 spin_unlock_bh(&xprt->transport_lock);
707}
708
709/*
710 * This function is called when memory window unbind which we are waiting
711 * for completes. Just use rr_func (zeroed by upcall) to signal completion.
712 */
713static void
714rpcrdma_unbind_func(struct rpcrdma_rep *rep)
715{
716 wake_up(&rep->rr_unbind);
717}
718
719/*
720 * Called as a tasklet to do req/reply match and complete a request
721 * Errors must result in the RPC task either being awakened, or
722 * allowed to timeout, to discover the errors at that time.
723 */
724void
725rpcrdma_reply_handler(struct rpcrdma_rep *rep)
726{
727 struct rpcrdma_msg *headerp;
728 struct rpcrdma_req *req;
729 struct rpc_rqst *rqst;
730 struct rpc_xprt *xprt = rep->rr_xprt;
731 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
732 __be32 *iptr;
733 int i, rdmalen, status;
734
735 /* Check status. If bad, signal disconnect and return rep to pool */
736 if (rep->rr_len == ~0U) {
737 rpcrdma_recv_buffer_put(rep);
738 if (r_xprt->rx_ep.rep_connected == 1) {
739 r_xprt->rx_ep.rep_connected = -EIO;
740 rpcrdma_conn_func(&r_xprt->rx_ep);
741 }
742 return;
743 }
744 if (rep->rr_len < 28) {
745 dprintk("RPC: %s: short/invalid reply\n", __func__);
746 goto repost;
747 }
748 headerp = (struct rpcrdma_msg *) rep->rr_base;
749 if (headerp->rm_vers != xdr_one) {
750 dprintk("RPC: %s: invalid version %d\n",
751 __func__, ntohl(headerp->rm_vers));
752 goto repost;
753 }
754
755 /* Get XID and try for a match. */
756 spin_lock(&xprt->transport_lock);
757 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
758 if (rqst == NULL) {
759 spin_unlock(&xprt->transport_lock);
760 dprintk("RPC: %s: reply 0x%p failed "
761 "to match any request xid 0x%08x len %d\n",
762 __func__, rep, headerp->rm_xid, rep->rr_len);
763repost:
764 r_xprt->rx_stats.bad_reply_count++;
765 rep->rr_func = rpcrdma_reply_handler;
766 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
767 rpcrdma_recv_buffer_put(rep);
768
769 return;
770 }
771
772 /* get request object */
773 req = rpcr_to_rdmar(rqst);
774 if (req->rl_reply) {
775 spin_unlock(&xprt->transport_lock);
776 dprintk("RPC: %s: duplicate reply 0x%p to RPC "
777 "request 0x%p: xid 0x%08x\n", __func__, rep, req,
778 headerp->rm_xid);
779 goto repost;
780 }
781
782 dprintk("RPC: %s: reply 0x%p completes request 0x%p\n"
783 " RPC request 0x%p xid 0x%08x\n",
784 __func__, rep, req, rqst, headerp->rm_xid);
785
786 /* from here on, the reply is no longer an orphan */
787 req->rl_reply = rep;
788
789 /* check for expected message types */
790 /* The order of some of these tests is important. */
791 switch (headerp->rm_type) {
792 case htonl(RDMA_MSG):
793 /* never expect read chunks */
794 /* never expect reply chunks (two ways to check) */
795 /* never expect write chunks without having offered RDMA */
796 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
797 (headerp->rm_body.rm_chunks[1] == xdr_zero &&
798 headerp->rm_body.rm_chunks[2] != xdr_zero) ||
799 (headerp->rm_body.rm_chunks[1] != xdr_zero &&
800 req->rl_nchunks == 0))
801 goto badheader;
802 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
803 /* count any expected write chunks in read reply */
804 /* start at write chunk array count */
805 iptr = &headerp->rm_body.rm_chunks[2];
806 rdmalen = rpcrdma_count_chunks(rep,
807 req->rl_nchunks, 1, &iptr);
808 /* check for validity, and no reply chunk after */
809 if (rdmalen < 0 || *iptr++ != xdr_zero)
810 goto badheader;
811 rep->rr_len -=
812 ((unsigned char *)iptr - (unsigned char *)headerp);
813 status = rep->rr_len + rdmalen;
814 r_xprt->rx_stats.total_rdma_reply += rdmalen;
815 /* special case - last chunk may omit padding */
816 if (rdmalen &= 3) {
817 rdmalen = 4 - rdmalen;
818 status += rdmalen;
819 }
820 } else {
821 /* else ordinary inline */
822 rdmalen = 0;
823 iptr = (__be32 *)((unsigned char *)headerp + 28);
824 rep->rr_len -= 28; /*sizeof *headerp;*/
825 status = rep->rr_len;
826 }
827 /* Fix up the rpc results for upper layer */
828 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
829 break;
830
831 case htonl(RDMA_NOMSG):
832 /* never expect read or write chunks, always reply chunks */
833 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
834 headerp->rm_body.rm_chunks[1] != xdr_zero ||
835 headerp->rm_body.rm_chunks[2] != xdr_one ||
836 req->rl_nchunks == 0)
837 goto badheader;
838 iptr = (__be32 *)((unsigned char *)headerp + 28);
839 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
840 if (rdmalen < 0)
841 goto badheader;
842 r_xprt->rx_stats.total_rdma_reply += rdmalen;
843 /* Reply chunk buffer already is the reply vector - no fixup. */
844 status = rdmalen;
845 break;
846
847badheader:
848 default:
849 dprintk("%s: invalid rpcrdma reply header (type %d):"
850 " chunks[012] == %d %d %d"
851 " expected chunks <= %d\n",
852 __func__, ntohl(headerp->rm_type),
853 headerp->rm_body.rm_chunks[0],
854 headerp->rm_body.rm_chunks[1],
855 headerp->rm_body.rm_chunks[2],
856 req->rl_nchunks);
857 status = -EIO;
858 r_xprt->rx_stats.bad_reply_count++;
859 break;
860 }
861
862 /* If using mw bind, start the deregister process now. */
863 /* (Note: if mr_free(), cannot perform it here, in tasklet context) */
864 if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
865 case RPCRDMA_MEMWINDOWS:
866 for (i = 0; req->rl_nchunks-- > 1;)
867 i += rpcrdma_deregister_external(
868 &req->rl_segments[i], r_xprt, NULL);
869 /* Optionally wait (not here) for unbinds to complete */
870 rep->rr_func = rpcrdma_unbind_func;
871 (void) rpcrdma_deregister_external(&req->rl_segments[i],
872 r_xprt, rep);
873 break;
874 case RPCRDMA_MEMWINDOWS_ASYNC:
875 for (i = 0; req->rl_nchunks--;)
876 i += rpcrdma_deregister_external(&req->rl_segments[i],
877 r_xprt, NULL);
878 break;
879 default:
880 break;
881 }
882
883 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
884 __func__, xprt, rqst, status);
885 xprt_complete_rqst(rqst->rq_task, status);
886 spin_unlock(&xprt->transport_lock);
887}