Loading...
1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2/*
3 * Copyright (c) 2014-2020, Oracle and/or its affiliates.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
10 * license below:
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 *
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 *
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
23 *
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
27 * permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 */
41
42/*
43 * rpc_rdma.c
44 *
45 * This file contains the guts of the RPC RDMA protocol, and
46 * does marshaling/unmarshaling, etc. It is also where interfacing
47 * to the Linux RPC framework lives.
48 */
49
50#include <linux/highmem.h>
51
52#include <linux/sunrpc/svc_rdma.h>
53
54#include "xprt_rdma.h"
55#include <trace/events/rpcrdma.h>
56
57#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58# define RPCDBG_FACILITY RPCDBG_TRANS
59#endif
60
61/* Returns size of largest RPC-over-RDMA header in a Call message
62 *
63 * The largest Call header contains a full-size Read list and a
64 * minimal Reply chunk.
65 */
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/* ACL likes to be lazy in allocating pages. For TCP, these
183 * pages can be allocated during receive processing. Not true
184 * for RDMA, which must always provision receive buffers
185 * up front.
186 */
187static noinline int
188rpcrdma_alloc_sparse_pages(struct xdr_buf *buf)
189{
190 struct page **ppages;
191 int len;
192
193 len = buf->page_len;
194 ppages = buf->pages + (buf->page_base >> PAGE_SHIFT);
195 while (len > 0) {
196 if (!*ppages)
197 *ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
198 if (!*ppages)
199 return -ENOBUFS;
200 ppages++;
201 len -= PAGE_SIZE;
202 }
203
204 return 0;
205}
206
207/* Convert @vec to a single SGL element.
208 *
209 * Returns pointer to next available SGE, and bumps the total number
210 * of SGEs consumed.
211 */
212static struct rpcrdma_mr_seg *
213rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
214 unsigned int *n)
215{
216 seg->mr_page = virt_to_page(vec->iov_base);
217 seg->mr_offset = offset_in_page(vec->iov_base);
218 seg->mr_len = vec->iov_len;
219 ++seg;
220 ++(*n);
221 return seg;
222}
223
224/* Convert @xdrbuf into SGEs no larger than a page each. As they
225 * are registered, these SGEs are then coalesced into RDMA segments
226 * when the selected memreg mode supports it.
227 *
228 * Returns positive number of SGEs consumed, or a negative errno.
229 */
230
231static int
232rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
233 unsigned int pos, enum rpcrdma_chunktype type,
234 struct rpcrdma_mr_seg *seg)
235{
236 unsigned long page_base;
237 unsigned int len, n;
238 struct page **ppages;
239
240 n = 0;
241 if (pos == 0)
242 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
243
244 len = xdrbuf->page_len;
245 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
246 page_base = offset_in_page(xdrbuf->page_base);
247 while (len) {
248 seg->mr_page = *ppages;
249 seg->mr_offset = page_base;
250 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
251 len -= seg->mr_len;
252 ++ppages;
253 ++seg;
254 ++n;
255 page_base = 0;
256 }
257
258 if (type == rpcrdma_readch)
259 goto out;
260
261 /* When encoding a Write chunk, some servers need to see an
262 * extra segment for non-XDR-aligned Write chunks. The upper
263 * layer provides space in the tail iovec that may be used
264 * for this purpose.
265 */
266 if (type == rpcrdma_writech && r_xprt->rx_ep->re_implicit_roundup)
267 goto out;
268
269 if (xdrbuf->tail[0].iov_len)
270 rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
271
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 (*mr)->mr_req = req;
319 }
320
321 rpcrdma_mr_push(*mr, &req->rl_registered);
322 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
323
324out_getmr_err:
325 trace_xprtrdma_nomrs_err(r_xprt, req);
326 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
327 rpcrdma_mrs_refresh(r_xprt);
328 return ERR_PTR(-EAGAIN);
329}
330
331/* Register and XDR encode the Read list. Supports encoding a list of read
332 * segments that belong to a single read chunk.
333 *
334 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
335 *
336 * Read chunklist (a linked list):
337 * N elements, position P (same P for all chunks of same arg!):
338 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
339 *
340 * Returns zero on success, or a negative errno if a failure occurred.
341 * @xdr is advanced to the next position in the stream.
342 *
343 * Only a single @pos value is currently supported.
344 */
345static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
346 struct rpcrdma_req *req,
347 struct rpc_rqst *rqst,
348 enum rpcrdma_chunktype rtype)
349{
350 struct xdr_stream *xdr = &req->rl_stream;
351 struct rpcrdma_mr_seg *seg;
352 struct rpcrdma_mr *mr;
353 unsigned int pos;
354 int nsegs;
355
356 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
357 goto done;
358
359 pos = rqst->rq_snd_buf.head[0].iov_len;
360 if (rtype == rpcrdma_areadch)
361 pos = 0;
362 seg = req->rl_segments;
363 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
364 rtype, seg);
365 if (nsegs < 0)
366 return nsegs;
367
368 do {
369 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
370 if (IS_ERR(seg))
371 return PTR_ERR(seg);
372
373 if (encode_read_segment(xdr, mr, pos) < 0)
374 return -EMSGSIZE;
375
376 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
377 r_xprt->rx_stats.read_chunk_count++;
378 nsegs -= mr->mr_nents;
379 } while (nsegs);
380
381done:
382 if (xdr_stream_encode_item_absent(xdr) < 0)
383 return -EMSGSIZE;
384 return 0;
385}
386
387/* Register and XDR encode the Write list. Supports encoding a list
388 * containing one array of plain segments that belong to a single
389 * write chunk.
390 *
391 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
392 *
393 * Write chunklist (a list of (one) counted array):
394 * N elements:
395 * 1 - N - HLOO - HLOO - ... - HLOO - 0
396 *
397 * Returns zero on success, or a negative errno if a failure occurred.
398 * @xdr is advanced to the next position in the stream.
399 *
400 * Only a single Write chunk is currently supported.
401 */
402static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
403 struct rpcrdma_req *req,
404 struct rpc_rqst *rqst,
405 enum rpcrdma_chunktype wtype)
406{
407 struct xdr_stream *xdr = &req->rl_stream;
408 struct rpcrdma_mr_seg *seg;
409 struct rpcrdma_mr *mr;
410 int nsegs, nchunks;
411 __be32 *segcount;
412
413 if (wtype != rpcrdma_writech)
414 goto done;
415
416 seg = req->rl_segments;
417 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
418 rqst->rq_rcv_buf.head[0].iov_len,
419 wtype, seg);
420 if (nsegs < 0)
421 return nsegs;
422
423 if (xdr_stream_encode_item_present(xdr) < 0)
424 return -EMSGSIZE;
425 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
426 if (unlikely(!segcount))
427 return -EMSGSIZE;
428 /* Actual value encoded below */
429
430 nchunks = 0;
431 do {
432 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
433 if (IS_ERR(seg))
434 return PTR_ERR(seg);
435
436 if (encode_rdma_segment(xdr, mr) < 0)
437 return -EMSGSIZE;
438
439 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
440 r_xprt->rx_stats.write_chunk_count++;
441 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
442 nchunks++;
443 nsegs -= mr->mr_nents;
444 } while (nsegs);
445
446 /* Update count of segments in this Write chunk */
447 *segcount = cpu_to_be32(nchunks);
448
449done:
450 if (xdr_stream_encode_item_absent(xdr) < 0)
451 return -EMSGSIZE;
452 return 0;
453}
454
455/* Register and XDR encode the Reply chunk. Supports encoding an array
456 * of plain segments that belong to a single write (reply) chunk.
457 *
458 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
459 *
460 * Reply chunk (a counted array):
461 * N elements:
462 * 1 - N - HLOO - HLOO - ... - HLOO
463 *
464 * Returns zero on success, or a negative errno if a failure occurred.
465 * @xdr is advanced to the next position in the stream.
466 */
467static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
468 struct rpcrdma_req *req,
469 struct rpc_rqst *rqst,
470 enum rpcrdma_chunktype wtype)
471{
472 struct xdr_stream *xdr = &req->rl_stream;
473 struct rpcrdma_mr_seg *seg;
474 struct rpcrdma_mr *mr;
475 int nsegs, nchunks;
476 __be32 *segcount;
477
478 if (wtype != rpcrdma_replych) {
479 if (xdr_stream_encode_item_absent(xdr) < 0)
480 return -EMSGSIZE;
481 return 0;
482 }
483
484 seg = req->rl_segments;
485 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
486 if (nsegs < 0)
487 return nsegs;
488
489 if (xdr_stream_encode_item_present(xdr) < 0)
490 return -EMSGSIZE;
491 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
492 if (unlikely(!segcount))
493 return -EMSGSIZE;
494 /* Actual value encoded below */
495
496 nchunks = 0;
497 do {
498 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
499 if (IS_ERR(seg))
500 return PTR_ERR(seg);
501
502 if (encode_rdma_segment(xdr, mr) < 0)
503 return -EMSGSIZE;
504
505 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
506 r_xprt->rx_stats.reply_chunk_count++;
507 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
508 nchunks++;
509 nsegs -= mr->mr_nents;
510 } while (nsegs);
511
512 /* Update count of segments in the Reply chunk */
513 *segcount = cpu_to_be32(nchunks);
514
515 return 0;
516}
517
518static void rpcrdma_sendctx_done(struct kref *kref)
519{
520 struct rpcrdma_req *req =
521 container_of(kref, struct rpcrdma_req, rl_kref);
522 struct rpcrdma_rep *rep = req->rl_reply;
523
524 rpcrdma_complete_rqst(rep);
525 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
526}
527
528/**
529 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
530 * @sc: sendctx containing SGEs to unmap
531 *
532 */
533void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
534{
535 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
536 struct ib_sge *sge;
537
538 if (!sc->sc_unmap_count)
539 return;
540
541 /* The first two SGEs contain the transport header and
542 * the inline buffer. These are always left mapped so
543 * they can be cheaply re-used.
544 */
545 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
546 ++sge, --sc->sc_unmap_count)
547 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
548 DMA_TO_DEVICE);
549
550 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
551}
552
553/* Prepare an SGE for the RPC-over-RDMA transport header.
554 */
555static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
556 struct rpcrdma_req *req, u32 len)
557{
558 struct rpcrdma_sendctx *sc = req->rl_sendctx;
559 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
560 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
561
562 sge->addr = rdmab_addr(rb);
563 sge->length = len;
564 sge->lkey = rdmab_lkey(rb);
565
566 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
567 DMA_TO_DEVICE);
568}
569
570/* The head iovec is straightforward, as it is usually already
571 * DMA-mapped. Sync the content that has changed.
572 */
573static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
574 struct rpcrdma_req *req, unsigned int len)
575{
576 struct rpcrdma_sendctx *sc = req->rl_sendctx;
577 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
578 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
579
580 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
581 return false;
582
583 sge->addr = rdmab_addr(rb);
584 sge->length = len;
585 sge->lkey = rdmab_lkey(rb);
586
587 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
588 DMA_TO_DEVICE);
589 return true;
590}
591
592/* If there is a page list present, DMA map and prepare an
593 * SGE for each page to be sent.
594 */
595static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
596 struct xdr_buf *xdr)
597{
598 struct rpcrdma_sendctx *sc = req->rl_sendctx;
599 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
600 unsigned int page_base, len, remaining;
601 struct page **ppages;
602 struct ib_sge *sge;
603
604 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
605 page_base = offset_in_page(xdr->page_base);
606 remaining = xdr->page_len;
607 while (remaining) {
608 sge = &sc->sc_sges[req->rl_wr.num_sge++];
609 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
610 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
611 page_base, len, DMA_TO_DEVICE);
612 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
613 goto out_mapping_err;
614
615 sge->length = len;
616 sge->lkey = rdmab_lkey(rb);
617
618 sc->sc_unmap_count++;
619 ppages++;
620 remaining -= len;
621 page_base = 0;
622 }
623
624 return true;
625
626out_mapping_err:
627 trace_xprtrdma_dma_maperr(sge->addr);
628 return false;
629}
630
631/* The tail iovec may include an XDR pad for the page list,
632 * as well as additional content, and may not reside in the
633 * same page as the head iovec.
634 */
635static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
636 struct xdr_buf *xdr,
637 unsigned int page_base, unsigned int len)
638{
639 struct rpcrdma_sendctx *sc = req->rl_sendctx;
640 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
641 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
642 struct page *page = virt_to_page(xdr->tail[0].iov_base);
643
644 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
645 DMA_TO_DEVICE);
646 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
647 goto out_mapping_err;
648
649 sge->length = len;
650 sge->lkey = rdmab_lkey(rb);
651 ++sc->sc_unmap_count;
652 return true;
653
654out_mapping_err:
655 trace_xprtrdma_dma_maperr(sge->addr);
656 return false;
657}
658
659/* Copy the tail to the end of the head buffer.
660 */
661static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
662 struct rpcrdma_req *req,
663 struct xdr_buf *xdr)
664{
665 unsigned char *dst;
666
667 dst = (unsigned char *)xdr->head[0].iov_base;
668 dst += xdr->head[0].iov_len + xdr->page_len;
669 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
670 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
671}
672
673/* Copy pagelist content into the head buffer.
674 */
675static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
676 struct rpcrdma_req *req,
677 struct xdr_buf *xdr)
678{
679 unsigned int len, page_base, remaining;
680 struct page **ppages;
681 unsigned char *src, *dst;
682
683 dst = (unsigned char *)xdr->head[0].iov_base;
684 dst += xdr->head[0].iov_len;
685 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
686 page_base = offset_in_page(xdr->page_base);
687 remaining = xdr->page_len;
688 while (remaining) {
689 src = page_address(*ppages);
690 src += page_base;
691 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
692 memcpy(dst, src, len);
693 r_xprt->rx_stats.pullup_copy_count += len;
694
695 ppages++;
696 dst += len;
697 remaining -= len;
698 page_base = 0;
699 }
700}
701
702/* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
703 * When the head, pagelist, and tail are small, a pull-up copy
704 * is considerably less costly than DMA mapping the components
705 * of @xdr.
706 *
707 * Assumptions:
708 * - the caller has already verified that the total length
709 * of the RPC Call body will fit into @rl_sendbuf.
710 */
711static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
712 struct rpcrdma_req *req,
713 struct xdr_buf *xdr)
714{
715 if (unlikely(xdr->tail[0].iov_len))
716 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
717
718 if (unlikely(xdr->page_len))
719 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
720
721 /* The whole RPC message resides in the head iovec now */
722 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
723}
724
725static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
726 struct rpcrdma_req *req,
727 struct xdr_buf *xdr)
728{
729 struct kvec *tail = &xdr->tail[0];
730
731 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
732 return false;
733 if (xdr->page_len)
734 if (!rpcrdma_prepare_pagelist(req, xdr))
735 return false;
736 if (tail->iov_len)
737 if (!rpcrdma_prepare_tail_iov(req, xdr,
738 offset_in_page(tail->iov_base),
739 tail->iov_len))
740 return false;
741
742 if (req->rl_sendctx->sc_unmap_count)
743 kref_get(&req->rl_kref);
744 return true;
745}
746
747static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
748 struct rpcrdma_req *req,
749 struct xdr_buf *xdr)
750{
751 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
752 return false;
753
754 /* If there is a Read chunk, the page list is being handled
755 * via explicit RDMA, and thus is skipped here.
756 */
757
758 /* Do not include the tail if it is only an XDR pad */
759 if (xdr->tail[0].iov_len > 3) {
760 unsigned int page_base, len;
761
762 /* If the content in the page list is an odd length,
763 * xdr_write_pages() adds a pad at the beginning of
764 * the tail iovec. Force the tail's non-pad content to
765 * land at the next XDR position in the Send message.
766 */
767 page_base = offset_in_page(xdr->tail[0].iov_base);
768 len = xdr->tail[0].iov_len;
769 page_base += len & 3;
770 len -= len & 3;
771 if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
772 return false;
773 kref_get(&req->rl_kref);
774 }
775
776 return true;
777}
778
779/**
780 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
781 * @r_xprt: controlling transport
782 * @req: context of RPC Call being marshalled
783 * @hdrlen: size of transport header, in bytes
784 * @xdr: xdr_buf containing RPC Call
785 * @rtype: chunk type being encoded
786 *
787 * Returns 0 on success; otherwise a negative errno is returned.
788 */
789inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
790 struct rpcrdma_req *req, u32 hdrlen,
791 struct xdr_buf *xdr,
792 enum rpcrdma_chunktype rtype)
793{
794 int ret;
795
796 ret = -EAGAIN;
797 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
798 if (!req->rl_sendctx)
799 goto out_nosc;
800 req->rl_sendctx->sc_unmap_count = 0;
801 req->rl_sendctx->sc_req = req;
802 kref_init(&req->rl_kref);
803 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
804 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
805 req->rl_wr.num_sge = 0;
806 req->rl_wr.opcode = IB_WR_SEND;
807
808 rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
809
810 ret = -EIO;
811 switch (rtype) {
812 case rpcrdma_noch_pullup:
813 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
814 goto out_unmap;
815 break;
816 case rpcrdma_noch_mapped:
817 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
818 goto out_unmap;
819 break;
820 case rpcrdma_readch:
821 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
822 goto out_unmap;
823 break;
824 case rpcrdma_areadch:
825 break;
826 default:
827 goto out_unmap;
828 }
829
830 return 0;
831
832out_unmap:
833 rpcrdma_sendctx_unmap(req->rl_sendctx);
834out_nosc:
835 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
836 return ret;
837}
838
839/**
840 * rpcrdma_marshal_req - Marshal and send one RPC request
841 * @r_xprt: controlling transport
842 * @rqst: RPC request to be marshaled
843 *
844 * For the RPC in "rqst", this function:
845 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
846 * - Registers Read, Write, and Reply chunks
847 * - Constructs the transport header
848 * - Posts a Send WR to send the transport header and request
849 *
850 * Returns:
851 * %0 if the RPC was sent successfully,
852 * %-ENOTCONN if the connection was lost,
853 * %-EAGAIN if the caller should call again with the same arguments,
854 * %-ENOBUFS if the caller should call again after a delay,
855 * %-EMSGSIZE if the transport header is too small,
856 * %-EIO if a permanent problem occurred while marshaling.
857 */
858int
859rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
860{
861 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
862 struct xdr_stream *xdr = &req->rl_stream;
863 enum rpcrdma_chunktype rtype, wtype;
864 struct xdr_buf *buf = &rqst->rq_snd_buf;
865 bool ddp_allowed;
866 __be32 *p;
867 int ret;
868
869 if (unlikely(rqst->rq_rcv_buf.flags & XDRBUF_SPARSE_PAGES)) {
870 ret = rpcrdma_alloc_sparse_pages(&rqst->rq_rcv_buf);
871 if (ret)
872 return ret;
873 }
874
875 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
876 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
877 rqst);
878
879 /* Fixed header fields */
880 ret = -EMSGSIZE;
881 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
882 if (!p)
883 goto out_err;
884 *p++ = rqst->rq_xid;
885 *p++ = rpcrdma_version;
886 *p++ = r_xprt->rx_buf.rb_max_requests;
887
888 /* When the ULP employs a GSS flavor that guarantees integrity
889 * or privacy, direct data placement of individual data items
890 * is not allowed.
891 */
892 ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
893 &rqst->rq_cred->cr_auth->au_flags);
894
895 /*
896 * Chunks needed for results?
897 *
898 * o If the expected result is under the inline threshold, all ops
899 * return as inline.
900 * o Large read ops return data as write chunk(s), header as
901 * inline.
902 * o Large non-read ops return as a single reply chunk.
903 */
904 if (rpcrdma_results_inline(r_xprt, rqst))
905 wtype = rpcrdma_noch;
906 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
907 rpcrdma_nonpayload_inline(r_xprt, rqst))
908 wtype = rpcrdma_writech;
909 else
910 wtype = rpcrdma_replych;
911
912 /*
913 * Chunks needed for arguments?
914 *
915 * o If the total request is under the inline threshold, all ops
916 * are sent as inline.
917 * o Large write ops transmit data as read chunk(s), header as
918 * inline.
919 * o Large non-write ops are sent with the entire message as a
920 * single read chunk (protocol 0-position special case).
921 *
922 * This assumes that the upper layer does not present a request
923 * that both has a data payload, and whose non-data arguments
924 * by themselves are larger than the inline threshold.
925 */
926 if (rpcrdma_args_inline(r_xprt, rqst)) {
927 *p++ = rdma_msg;
928 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
929 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
930 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
931 *p++ = rdma_msg;
932 rtype = rpcrdma_readch;
933 } else {
934 r_xprt->rx_stats.nomsg_call_count++;
935 *p++ = rdma_nomsg;
936 rtype = rpcrdma_areadch;
937 }
938
939 /* This implementation supports the following combinations
940 * of chunk lists in one RPC-over-RDMA Call message:
941 *
942 * - Read list
943 * - Write list
944 * - Reply chunk
945 * - Read list + Reply chunk
946 *
947 * It might not yet support the following combinations:
948 *
949 * - Read list + Write list
950 *
951 * It does not support the following combinations:
952 *
953 * - Write list + Reply chunk
954 * - Read list + Write list + Reply chunk
955 *
956 * This implementation supports only a single chunk in each
957 * Read or Write list. Thus for example the client cannot
958 * send a Call message with a Position Zero Read chunk and a
959 * regular Read chunk at the same time.
960 */
961 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
962 if (ret)
963 goto out_err;
964 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
965 if (ret)
966 goto out_err;
967 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
968 if (ret)
969 goto out_err;
970
971 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
972 buf, rtype);
973 if (ret)
974 goto out_err;
975
976 trace_xprtrdma_marshal(req, rtype, wtype);
977 return 0;
978
979out_err:
980 trace_xprtrdma_marshal_failed(rqst, ret);
981 r_xprt->rx_stats.failed_marshal_count++;
982 frwr_reset(req);
983 return ret;
984}
985
986static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
987 struct rpcrdma_buffer *buf,
988 u32 grant)
989{
990 buf->rb_credits = grant;
991 xprt->cwnd = grant << RPC_CWNDSHIFT;
992}
993
994static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
995{
996 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
997
998 spin_lock(&xprt->transport_lock);
999 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1000 spin_unlock(&xprt->transport_lock);
1001}
1002
1003/**
1004 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1005 * @r_xprt: controlling transport instance
1006 *
1007 * Prepare @r_xprt for the next connection by reinitializing
1008 * its credit grant to one (see RFC 8166, Section 3.3.3).
1009 */
1010void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1011{
1012 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1013
1014 spin_lock(&xprt->transport_lock);
1015 xprt->cong = 0;
1016 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1017 spin_unlock(&xprt->transport_lock);
1018}
1019
1020/**
1021 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1022 * @rqst: controlling RPC request
1023 * @srcp: points to RPC message payload in receive buffer
1024 * @copy_len: remaining length of receive buffer content
1025 * @pad: Write chunk pad bytes needed (zero for pure inline)
1026 *
1027 * The upper layer has set the maximum number of bytes it can
1028 * receive in each component of rq_rcv_buf. These values are set in
1029 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1030 *
1031 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1032 * many cases this function simply updates iov_base pointers in
1033 * rq_rcv_buf to point directly to the received reply data, to
1034 * avoid copying reply data.
1035 *
1036 * Returns the count of bytes which had to be memcopied.
1037 */
1038static unsigned long
1039rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1040{
1041 unsigned long fixup_copy_count;
1042 int i, npages, curlen;
1043 char *destp;
1044 struct page **ppages;
1045 int page_base;
1046
1047 /* The head iovec is redirected to the RPC reply message
1048 * in the receive buffer, to avoid a memcopy.
1049 */
1050 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1051 rqst->rq_private_buf.head[0].iov_base = srcp;
1052
1053 /* The contents of the receive buffer that follow
1054 * head.iov_len bytes are copied into the page list.
1055 */
1056 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1057 if (curlen > copy_len)
1058 curlen = copy_len;
1059 srcp += curlen;
1060 copy_len -= curlen;
1061
1062 ppages = rqst->rq_rcv_buf.pages +
1063 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1064 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1065 fixup_copy_count = 0;
1066 if (copy_len && rqst->rq_rcv_buf.page_len) {
1067 int pagelist_len;
1068
1069 pagelist_len = rqst->rq_rcv_buf.page_len;
1070 if (pagelist_len > copy_len)
1071 pagelist_len = copy_len;
1072 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1073 for (i = 0; i < npages; i++) {
1074 curlen = PAGE_SIZE - page_base;
1075 if (curlen > pagelist_len)
1076 curlen = pagelist_len;
1077
1078 destp = kmap_atomic(ppages[i]);
1079 memcpy(destp + page_base, srcp, curlen);
1080 flush_dcache_page(ppages[i]);
1081 kunmap_atomic(destp);
1082 srcp += curlen;
1083 copy_len -= curlen;
1084 fixup_copy_count += curlen;
1085 pagelist_len -= curlen;
1086 if (!pagelist_len)
1087 break;
1088 page_base = 0;
1089 }
1090
1091 /* Implicit padding for the last segment in a Write
1092 * chunk is inserted inline at the front of the tail
1093 * iovec. The upper layer ignores the content of
1094 * the pad. Simply ensure inline content in the tail
1095 * that follows the Write chunk is properly aligned.
1096 */
1097 if (pad)
1098 srcp -= pad;
1099 }
1100
1101 /* The tail iovec is redirected to the remaining data
1102 * in the receive buffer, to avoid a memcopy.
1103 */
1104 if (copy_len || pad) {
1105 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1106 rqst->rq_private_buf.tail[0].iov_base = srcp;
1107 }
1108
1109 if (fixup_copy_count)
1110 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1111 return fixup_copy_count;
1112}
1113
1114/* By convention, backchannel calls arrive via rdma_msg type
1115 * messages, and never populate the chunk lists. This makes
1116 * the RPC/RDMA header small and fixed in size, so it is
1117 * straightforward to check the RPC header's direction field.
1118 */
1119static bool
1120rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1121#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1122{
1123 struct xdr_stream *xdr = &rep->rr_stream;
1124 __be32 *p;
1125
1126 if (rep->rr_proc != rdma_msg)
1127 return false;
1128
1129 /* Peek at stream contents without advancing. */
1130 p = xdr_inline_decode(xdr, 0);
1131
1132 /* Chunk lists */
1133 if (xdr_item_is_present(p++))
1134 return false;
1135 if (xdr_item_is_present(p++))
1136 return false;
1137 if (xdr_item_is_present(p++))
1138 return false;
1139
1140 /* RPC header */
1141 if (*p++ != rep->rr_xid)
1142 return false;
1143 if (*p != cpu_to_be32(RPC_CALL))
1144 return false;
1145
1146 /* Now that we are sure this is a backchannel call,
1147 * advance to the RPC header.
1148 */
1149 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1150 if (unlikely(!p))
1151 return true;
1152
1153 rpcrdma_bc_receive_call(r_xprt, rep);
1154 return true;
1155}
1156#else /* CONFIG_SUNRPC_BACKCHANNEL */
1157{
1158 return false;
1159}
1160#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1161
1162static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1163{
1164 u32 handle;
1165 u64 offset;
1166 __be32 *p;
1167
1168 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1169 if (unlikely(!p))
1170 return -EIO;
1171
1172 xdr_decode_rdma_segment(p, &handle, length, &offset);
1173 trace_xprtrdma_decode_seg(handle, *length, offset);
1174 return 0;
1175}
1176
1177static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1178{
1179 u32 segcount, seglength;
1180 __be32 *p;
1181
1182 p = xdr_inline_decode(xdr, sizeof(*p));
1183 if (unlikely(!p))
1184 return -EIO;
1185
1186 *length = 0;
1187 segcount = be32_to_cpup(p);
1188 while (segcount--) {
1189 if (decode_rdma_segment(xdr, &seglength))
1190 return -EIO;
1191 *length += seglength;
1192 }
1193
1194 return 0;
1195}
1196
1197/* In RPC-over-RDMA Version One replies, a Read list is never
1198 * expected. This decoder is a stub that returns an error if
1199 * a Read list is present.
1200 */
1201static int decode_read_list(struct xdr_stream *xdr)
1202{
1203 __be32 *p;
1204
1205 p = xdr_inline_decode(xdr, sizeof(*p));
1206 if (unlikely(!p))
1207 return -EIO;
1208 if (unlikely(xdr_item_is_present(p)))
1209 return -EIO;
1210 return 0;
1211}
1212
1213/* Supports only one Write chunk in the Write list
1214 */
1215static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1216{
1217 u32 chunklen;
1218 bool first;
1219 __be32 *p;
1220
1221 *length = 0;
1222 first = true;
1223 do {
1224 p = xdr_inline_decode(xdr, sizeof(*p));
1225 if (unlikely(!p))
1226 return -EIO;
1227 if (xdr_item_is_absent(p))
1228 break;
1229 if (!first)
1230 return -EIO;
1231
1232 if (decode_write_chunk(xdr, &chunklen))
1233 return -EIO;
1234 *length += chunklen;
1235 first = false;
1236 } while (true);
1237 return 0;
1238}
1239
1240static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1241{
1242 __be32 *p;
1243
1244 p = xdr_inline_decode(xdr, sizeof(*p));
1245 if (unlikely(!p))
1246 return -EIO;
1247
1248 *length = 0;
1249 if (xdr_item_is_present(p))
1250 if (decode_write_chunk(xdr, length))
1251 return -EIO;
1252 return 0;
1253}
1254
1255static int
1256rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1257 struct rpc_rqst *rqst)
1258{
1259 struct xdr_stream *xdr = &rep->rr_stream;
1260 u32 writelist, replychunk, rpclen;
1261 char *base;
1262
1263 /* Decode the chunk lists */
1264 if (decode_read_list(xdr))
1265 return -EIO;
1266 if (decode_write_list(xdr, &writelist))
1267 return -EIO;
1268 if (decode_reply_chunk(xdr, &replychunk))
1269 return -EIO;
1270
1271 /* RDMA_MSG sanity checks */
1272 if (unlikely(replychunk))
1273 return -EIO;
1274
1275 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1276 base = (char *)xdr_inline_decode(xdr, 0);
1277 rpclen = xdr_stream_remaining(xdr);
1278 r_xprt->rx_stats.fixup_copy_count +=
1279 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1280
1281 r_xprt->rx_stats.total_rdma_reply += writelist;
1282 return rpclen + xdr_align_size(writelist);
1283}
1284
1285static noinline int
1286rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1287{
1288 struct xdr_stream *xdr = &rep->rr_stream;
1289 u32 writelist, replychunk;
1290
1291 /* Decode the chunk lists */
1292 if (decode_read_list(xdr))
1293 return -EIO;
1294 if (decode_write_list(xdr, &writelist))
1295 return -EIO;
1296 if (decode_reply_chunk(xdr, &replychunk))
1297 return -EIO;
1298
1299 /* RDMA_NOMSG sanity checks */
1300 if (unlikely(writelist))
1301 return -EIO;
1302 if (unlikely(!replychunk))
1303 return -EIO;
1304
1305 /* Reply chunk buffer already is the reply vector */
1306 r_xprt->rx_stats.total_rdma_reply += replychunk;
1307 return replychunk;
1308}
1309
1310static noinline int
1311rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1312 struct rpc_rqst *rqst)
1313{
1314 struct xdr_stream *xdr = &rep->rr_stream;
1315 __be32 *p;
1316
1317 p = xdr_inline_decode(xdr, sizeof(*p));
1318 if (unlikely(!p))
1319 return -EIO;
1320
1321 switch (*p) {
1322 case err_vers:
1323 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1324 if (!p)
1325 break;
1326 trace_xprtrdma_err_vers(rqst, p, p + 1);
1327 break;
1328 case err_chunk:
1329 trace_xprtrdma_err_chunk(rqst);
1330 break;
1331 default:
1332 trace_xprtrdma_err_unrecognized(rqst, p);
1333 }
1334
1335 return -EIO;
1336}
1337
1338/**
1339 * rpcrdma_unpin_rqst - Release rqst without completing it
1340 * @rep: RPC/RDMA Receive context
1341 *
1342 * This is done when a connection is lost so that a Reply
1343 * can be dropped and its matching Call can be subsequently
1344 * retransmitted on a new connection.
1345 */
1346void rpcrdma_unpin_rqst(struct rpcrdma_rep *rep)
1347{
1348 struct rpc_xprt *xprt = &rep->rr_rxprt->rx_xprt;
1349 struct rpc_rqst *rqst = rep->rr_rqst;
1350 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
1351
1352 req->rl_reply = NULL;
1353 rep->rr_rqst = NULL;
1354
1355 spin_lock(&xprt->queue_lock);
1356 xprt_unpin_rqst(rqst);
1357 spin_unlock(&xprt->queue_lock);
1358}
1359
1360/**
1361 * rpcrdma_complete_rqst - Pass completed rqst back to RPC
1362 * @rep: RPC/RDMA Receive context
1363 *
1364 * Reconstruct the RPC reply and complete the transaction
1365 * while @rqst is still pinned to ensure the rep, rqst, and
1366 * rq_task pointers remain stable.
1367 */
1368void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1369{
1370 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1371 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1372 struct rpc_rqst *rqst = rep->rr_rqst;
1373 int status;
1374
1375 switch (rep->rr_proc) {
1376 case rdma_msg:
1377 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1378 break;
1379 case rdma_nomsg:
1380 status = rpcrdma_decode_nomsg(r_xprt, rep);
1381 break;
1382 case rdma_error:
1383 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1384 break;
1385 default:
1386 status = -EIO;
1387 }
1388 if (status < 0)
1389 goto out_badheader;
1390
1391out:
1392 spin_lock(&xprt->queue_lock);
1393 xprt_complete_rqst(rqst->rq_task, status);
1394 xprt_unpin_rqst(rqst);
1395 spin_unlock(&xprt->queue_lock);
1396 return;
1397
1398out_badheader:
1399 trace_xprtrdma_reply_hdr_err(rep);
1400 r_xprt->rx_stats.bad_reply_count++;
1401 rqst->rq_task->tk_status = status;
1402 status = 0;
1403 goto out;
1404}
1405
1406static void rpcrdma_reply_done(struct kref *kref)
1407{
1408 struct rpcrdma_req *req =
1409 container_of(kref, struct rpcrdma_req, rl_kref);
1410
1411 rpcrdma_complete_rqst(req->rl_reply);
1412}
1413
1414/**
1415 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1416 * @rep: Incoming rpcrdma_rep object to process
1417 *
1418 * Errors must result in the RPC task either being awakened, or
1419 * allowed to timeout, to discover the errors at that time.
1420 */
1421void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1422{
1423 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1424 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1425 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1426 struct rpcrdma_req *req;
1427 struct rpc_rqst *rqst;
1428 u32 credits;
1429 __be32 *p;
1430
1431 /* Any data means we had a useful conversation, so
1432 * then we don't need to delay the next reconnect.
1433 */
1434 if (xprt->reestablish_timeout)
1435 xprt->reestablish_timeout = 0;
1436
1437 /* Fixed transport header fields */
1438 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1439 rep->rr_hdrbuf.head[0].iov_base, NULL);
1440 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1441 if (unlikely(!p))
1442 goto out_shortreply;
1443 rep->rr_xid = *p++;
1444 rep->rr_vers = *p++;
1445 credits = be32_to_cpu(*p++);
1446 rep->rr_proc = *p++;
1447
1448 if (rep->rr_vers != rpcrdma_version)
1449 goto out_badversion;
1450
1451 if (rpcrdma_is_bcall(r_xprt, rep))
1452 return;
1453
1454 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1455 * get context for handling any incoming chunks.
1456 */
1457 spin_lock(&xprt->queue_lock);
1458 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1459 if (!rqst)
1460 goto out_norqst;
1461 xprt_pin_rqst(rqst);
1462 spin_unlock(&xprt->queue_lock);
1463
1464 if (credits == 0)
1465 credits = 1; /* don't deadlock */
1466 else if (credits > r_xprt->rx_ep->re_max_requests)
1467 credits = r_xprt->rx_ep->re_max_requests;
1468 rpcrdma_post_recvs(r_xprt, credits + (buf->rb_bc_srv_max_requests << 1),
1469 false);
1470 if (buf->rb_credits != credits)
1471 rpcrdma_update_cwnd(r_xprt, credits);
1472
1473 req = rpcr_to_rdmar(rqst);
1474 if (unlikely(req->rl_reply))
1475 rpcrdma_rep_put(buf, req->rl_reply);
1476 req->rl_reply = rep;
1477 rep->rr_rqst = rqst;
1478
1479 trace_xprtrdma_reply(rqst->rq_task, rep, credits);
1480
1481 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1482 frwr_reminv(rep, &req->rl_registered);
1483 if (!list_empty(&req->rl_registered))
1484 frwr_unmap_async(r_xprt, req);
1485 /* LocalInv completion will complete the RPC */
1486 else
1487 kref_put(&req->rl_kref, rpcrdma_reply_done);
1488 return;
1489
1490out_badversion:
1491 trace_xprtrdma_reply_vers_err(rep);
1492 goto out;
1493
1494out_norqst:
1495 spin_unlock(&xprt->queue_lock);
1496 trace_xprtrdma_reply_rqst_err(rep);
1497 goto out;
1498
1499out_shortreply:
1500 trace_xprtrdma_reply_short_err(rep);
1501
1502out:
1503 rpcrdma_rep_put(buf, rep);
1504}
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], KM_SKB_SUNRPC_DATA);
342 memcpy(destp, srcp+page_base, curlen);
343 kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
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], KM_SKB_SUNRPC_DATA);
643 memcpy(destp + page_base, srcp, curlen);
644 flush_dcache_page(ppages[i]);
645 kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
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
775 dprintk("RPC: %s: reply 0x%p completes request 0x%p\n"
776 " RPC request 0x%p xid 0x%08x\n",
777 __func__, rep, req, rqst, headerp->rm_xid);
778
779 BUG_ON(!req || req->rl_reply);
780
781 /* from here on, the reply is no longer an orphan */
782 req->rl_reply = rep;
783
784 /* check for expected message types */
785 /* The order of some of these tests is important. */
786 switch (headerp->rm_type) {
787 case htonl(RDMA_MSG):
788 /* never expect read chunks */
789 /* never expect reply chunks (two ways to check) */
790 /* never expect write chunks without having offered RDMA */
791 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
792 (headerp->rm_body.rm_chunks[1] == xdr_zero &&
793 headerp->rm_body.rm_chunks[2] != xdr_zero) ||
794 (headerp->rm_body.rm_chunks[1] != xdr_zero &&
795 req->rl_nchunks == 0))
796 goto badheader;
797 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
798 /* count any expected write chunks in read reply */
799 /* start at write chunk array count */
800 iptr = &headerp->rm_body.rm_chunks[2];
801 rdmalen = rpcrdma_count_chunks(rep,
802 req->rl_nchunks, 1, &iptr);
803 /* check for validity, and no reply chunk after */
804 if (rdmalen < 0 || *iptr++ != xdr_zero)
805 goto badheader;
806 rep->rr_len -=
807 ((unsigned char *)iptr - (unsigned char *)headerp);
808 status = rep->rr_len + rdmalen;
809 r_xprt->rx_stats.total_rdma_reply += rdmalen;
810 /* special case - last chunk may omit padding */
811 if (rdmalen &= 3) {
812 rdmalen = 4 - rdmalen;
813 status += rdmalen;
814 }
815 } else {
816 /* else ordinary inline */
817 rdmalen = 0;
818 iptr = (__be32 *)((unsigned char *)headerp + 28);
819 rep->rr_len -= 28; /*sizeof *headerp;*/
820 status = rep->rr_len;
821 }
822 /* Fix up the rpc results for upper layer */
823 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
824 break;
825
826 case htonl(RDMA_NOMSG):
827 /* never expect read or write chunks, always reply chunks */
828 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
829 headerp->rm_body.rm_chunks[1] != xdr_zero ||
830 headerp->rm_body.rm_chunks[2] != xdr_one ||
831 req->rl_nchunks == 0)
832 goto badheader;
833 iptr = (__be32 *)((unsigned char *)headerp + 28);
834 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
835 if (rdmalen < 0)
836 goto badheader;
837 r_xprt->rx_stats.total_rdma_reply += rdmalen;
838 /* Reply chunk buffer already is the reply vector - no fixup. */
839 status = rdmalen;
840 break;
841
842badheader:
843 default:
844 dprintk("%s: invalid rpcrdma reply header (type %d):"
845 " chunks[012] == %d %d %d"
846 " expected chunks <= %d\n",
847 __func__, ntohl(headerp->rm_type),
848 headerp->rm_body.rm_chunks[0],
849 headerp->rm_body.rm_chunks[1],
850 headerp->rm_body.rm_chunks[2],
851 req->rl_nchunks);
852 status = -EIO;
853 r_xprt->rx_stats.bad_reply_count++;
854 break;
855 }
856
857 /* If using mw bind, start the deregister process now. */
858 /* (Note: if mr_free(), cannot perform it here, in tasklet context) */
859 if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
860 case RPCRDMA_MEMWINDOWS:
861 for (i = 0; req->rl_nchunks-- > 1;)
862 i += rpcrdma_deregister_external(
863 &req->rl_segments[i], r_xprt, NULL);
864 /* Optionally wait (not here) for unbinds to complete */
865 rep->rr_func = rpcrdma_unbind_func;
866 (void) rpcrdma_deregister_external(&req->rl_segments[i],
867 r_xprt, rep);
868 break;
869 case RPCRDMA_MEMWINDOWS_ASYNC:
870 for (i = 0; req->rl_nchunks--;)
871 i += rpcrdma_deregister_external(&req->rl_segments[i],
872 r_xprt, NULL);
873 break;
874 default:
875 break;
876 }
877
878 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
879 __func__, xprt, rqst, status);
880 xprt_complete_rqst(rqst->rq_task, status);
881 spin_unlock(&xprt->transport_lock);
882}