Loading...
1/*
2 * Copyright (c) 2014-2017 Oracle. All rights reserved.
3 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
9 * license below:
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 *
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 *
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
22 *
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
26 * permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 */
40
41/*
42 * rpc_rdma.c
43 *
44 * This file contains the guts of the RPC RDMA protocol, and
45 * does marshaling/unmarshaling, etc. It is also where interfacing
46 * to the Linux RPC framework lives.
47 */
48
49#include "xprt_rdma.h"
50
51#include <linux/highmem.h>
52
53#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
54# define RPCDBG_FACILITY RPCDBG_TRANS
55#endif
56
57static const char transfertypes[][12] = {
58 "inline", /* no chunks */
59 "read list", /* some argument via rdma read */
60 "*read list", /* entire request via rdma read */
61 "write list", /* some result via rdma write */
62 "reply chunk" /* entire reply via rdma write */
63};
64
65/* Returns size of largest RPC-over-RDMA header in a Call message
66 *
67 * The largest Call header contains a full-size Read list and a
68 * minimal Reply chunk.
69 */
70static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
71{
72 unsigned int size;
73
74 /* Fixed header fields and list discriminators */
75 size = RPCRDMA_HDRLEN_MIN;
76
77 /* Maximum Read list size */
78 maxsegs += 2; /* segment for head and tail buffers */
79 size = maxsegs * rpcrdma_readchunk_maxsz * sizeof(__be32);
80
81 /* Minimal Read chunk size */
82 size += sizeof(__be32); /* segment count */
83 size += rpcrdma_segment_maxsz * sizeof(__be32);
84 size += sizeof(__be32); /* list discriminator */
85
86 dprintk("RPC: %s: max call header size = %u\n",
87 __func__, size);
88 return size;
89}
90
91/* Returns size of largest RPC-over-RDMA header in a Reply message
92 *
93 * There is only one Write list or one Reply chunk per Reply
94 * message. The larger list is the Write list.
95 */
96static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
97{
98 unsigned int size;
99
100 /* Fixed header fields and list discriminators */
101 size = RPCRDMA_HDRLEN_MIN;
102
103 /* Maximum Write list size */
104 maxsegs += 2; /* segment for head and tail buffers */
105 size = sizeof(__be32); /* segment count */
106 size += maxsegs * rpcrdma_segment_maxsz * sizeof(__be32);
107 size += sizeof(__be32); /* list discriminator */
108
109 dprintk("RPC: %s: max reply header size = %u\n",
110 __func__, size);
111 return size;
112}
113
114void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *r_xprt)
115{
116 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
117 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
118 unsigned int maxsegs = ia->ri_max_segs;
119
120 ia->ri_max_inline_write = cdata->inline_wsize -
121 rpcrdma_max_call_header_size(maxsegs);
122 ia->ri_max_inline_read = cdata->inline_rsize -
123 rpcrdma_max_reply_header_size(maxsegs);
124}
125
126/* The client can send a request inline as long as the RPCRDMA header
127 * plus the RPC call fit under the transport's inline limit. If the
128 * combined call message size exceeds that limit, the client must use
129 * a Read chunk for this operation.
130 *
131 * A Read chunk is also required if sending the RPC call inline would
132 * exceed this device's max_sge limit.
133 */
134static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
135 struct rpc_rqst *rqst)
136{
137 struct xdr_buf *xdr = &rqst->rq_snd_buf;
138 unsigned int count, remaining, offset;
139
140 if (xdr->len > r_xprt->rx_ia.ri_max_inline_write)
141 return false;
142
143 if (xdr->page_len) {
144 remaining = xdr->page_len;
145 offset = offset_in_page(xdr->page_base);
146 count = RPCRDMA_MIN_SEND_SGES;
147 while (remaining) {
148 remaining -= min_t(unsigned int,
149 PAGE_SIZE - offset, remaining);
150 offset = 0;
151 if (++count > r_xprt->rx_ia.ri_max_send_sges)
152 return false;
153 }
154 }
155
156 return true;
157}
158
159/* The client can't know how large the actual reply will be. Thus it
160 * plans for the largest possible reply for that particular ULP
161 * operation. If the maximum combined reply message size exceeds that
162 * limit, the client must provide a write list or a reply chunk for
163 * this request.
164 */
165static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
166 struct rpc_rqst *rqst)
167{
168 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
169
170 return rqst->rq_rcv_buf.buflen <= ia->ri_max_inline_read;
171}
172
173/* Split @vec on page boundaries into SGEs. FMR registers pages, not
174 * a byte range. Other modes coalesce these SGEs into a single MR
175 * when they can.
176 *
177 * Returns pointer to next available SGE, and bumps the total number
178 * of SGEs consumed.
179 */
180static struct rpcrdma_mr_seg *
181rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
182 unsigned int *n)
183{
184 u32 remaining, page_offset;
185 char *base;
186
187 base = vec->iov_base;
188 page_offset = offset_in_page(base);
189 remaining = vec->iov_len;
190 while (remaining) {
191 seg->mr_page = NULL;
192 seg->mr_offset = base;
193 seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
194 remaining -= seg->mr_len;
195 base += seg->mr_len;
196 ++seg;
197 ++(*n);
198 page_offset = 0;
199 }
200 return seg;
201}
202
203/* Convert @xdrbuf into SGEs no larger than a page each. As they
204 * are registered, these SGEs are then coalesced into RDMA segments
205 * when the selected memreg mode supports it.
206 *
207 * Returns positive number of SGEs consumed, or a negative errno.
208 */
209
210static int
211rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
212 unsigned int pos, enum rpcrdma_chunktype type,
213 struct rpcrdma_mr_seg *seg)
214{
215 unsigned long page_base;
216 unsigned int len, n;
217 struct page **ppages;
218
219 n = 0;
220 if (pos == 0)
221 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
222
223 len = xdrbuf->page_len;
224 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
225 page_base = offset_in_page(xdrbuf->page_base);
226 while (len) {
227 if (unlikely(!*ppages)) {
228 /* XXX: Certain upper layer operations do
229 * not provide receive buffer pages.
230 */
231 *ppages = alloc_page(GFP_ATOMIC);
232 if (!*ppages)
233 return -EAGAIN;
234 }
235 seg->mr_page = *ppages;
236 seg->mr_offset = (char *)page_base;
237 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
238 len -= seg->mr_len;
239 ++ppages;
240 ++seg;
241 ++n;
242 page_base = 0;
243 }
244
245 /* When encoding a Read chunk, the tail iovec contains an
246 * XDR pad and may be omitted.
247 */
248 if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
249 goto out;
250
251 /* When encoding a Write chunk, some servers need to see an
252 * extra segment for non-XDR-aligned Write chunks. The upper
253 * layer provides space in the tail iovec that may be used
254 * for this purpose.
255 */
256 if (type == rpcrdma_writech && r_xprt->rx_ia.ri_implicit_roundup)
257 goto out;
258
259 if (xdrbuf->tail[0].iov_len)
260 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
261
262out:
263 if (unlikely(n > RPCRDMA_MAX_SEGS))
264 return -EIO;
265 return n;
266}
267
268static inline int
269encode_item_present(struct xdr_stream *xdr)
270{
271 __be32 *p;
272
273 p = xdr_reserve_space(xdr, sizeof(*p));
274 if (unlikely(!p))
275 return -EMSGSIZE;
276
277 *p = xdr_one;
278 return 0;
279}
280
281static inline int
282encode_item_not_present(struct xdr_stream *xdr)
283{
284 __be32 *p;
285
286 p = xdr_reserve_space(xdr, sizeof(*p));
287 if (unlikely(!p))
288 return -EMSGSIZE;
289
290 *p = xdr_zero;
291 return 0;
292}
293
294static void
295xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr *mr)
296{
297 *iptr++ = cpu_to_be32(mr->mr_handle);
298 *iptr++ = cpu_to_be32(mr->mr_length);
299 xdr_encode_hyper(iptr, mr->mr_offset);
300}
301
302static int
303encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
304{
305 __be32 *p;
306
307 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
308 if (unlikely(!p))
309 return -EMSGSIZE;
310
311 xdr_encode_rdma_segment(p, mr);
312 return 0;
313}
314
315static int
316encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
317 u32 position)
318{
319 __be32 *p;
320
321 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
322 if (unlikely(!p))
323 return -EMSGSIZE;
324
325 *p++ = xdr_one; /* Item present */
326 *p++ = cpu_to_be32(position);
327 xdr_encode_rdma_segment(p, mr);
328 return 0;
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 noinline int
346rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
347 struct rpc_rqst *rqst, enum rpcrdma_chunktype rtype)
348{
349 struct xdr_stream *xdr = &req->rl_stream;
350 struct rpcrdma_mr_seg *seg;
351 struct rpcrdma_mr *mr;
352 unsigned int pos;
353 int nsegs;
354
355 pos = rqst->rq_snd_buf.head[0].iov_len;
356 if (rtype == rpcrdma_areadch)
357 pos = 0;
358 seg = req->rl_segments;
359 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
360 rtype, seg);
361 if (nsegs < 0)
362 return nsegs;
363
364 do {
365 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
366 false, &mr);
367 if (IS_ERR(seg))
368 goto out_maperr;
369 rpcrdma_mr_push(mr, &req->rl_registered);
370
371 if (encode_read_segment(xdr, mr, pos) < 0)
372 return -EMSGSIZE;
373
374 trace_xprtrdma_read_chunk(rqst->rq_task, pos, mr, nsegs);
375 r_xprt->rx_stats.read_chunk_count++;
376 nsegs -= mr->mr_nents;
377 } while (nsegs);
378
379 return 0;
380
381out_maperr:
382 if (PTR_ERR(seg) == -EAGAIN)
383 xprt_wait_for_buffer_space(rqst->rq_task, NULL);
384 return PTR_ERR(seg);
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 noinline int
403rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
404 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
405{
406 struct xdr_stream *xdr = &req->rl_stream;
407 struct rpcrdma_mr_seg *seg;
408 struct rpcrdma_mr *mr;
409 int nsegs, nchunks;
410 __be32 *segcount;
411
412 seg = req->rl_segments;
413 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
414 rqst->rq_rcv_buf.head[0].iov_len,
415 wtype, seg);
416 if (nsegs < 0)
417 return nsegs;
418
419 if (encode_item_present(xdr) < 0)
420 return -EMSGSIZE;
421 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
422 if (unlikely(!segcount))
423 return -EMSGSIZE;
424 /* Actual value encoded below */
425
426 nchunks = 0;
427 do {
428 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
429 true, &mr);
430 if (IS_ERR(seg))
431 goto out_maperr;
432 rpcrdma_mr_push(mr, &req->rl_registered);
433
434 if (encode_rdma_segment(xdr, mr) < 0)
435 return -EMSGSIZE;
436
437 trace_xprtrdma_write_chunk(rqst->rq_task, mr, nsegs);
438 r_xprt->rx_stats.write_chunk_count++;
439 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
440 nchunks++;
441 nsegs -= mr->mr_nents;
442 } while (nsegs);
443
444 /* Update count of segments in this Write chunk */
445 *segcount = cpu_to_be32(nchunks);
446
447 return 0;
448
449out_maperr:
450 if (PTR_ERR(seg) == -EAGAIN)
451 xprt_wait_for_buffer_space(rqst->rq_task, NULL);
452 return PTR_ERR(seg);
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 noinline int
468rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
469 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
470{
471 struct xdr_stream *xdr = &req->rl_stream;
472 struct rpcrdma_mr_seg *seg;
473 struct rpcrdma_mr *mr;
474 int nsegs, nchunks;
475 __be32 *segcount;
476
477 seg = req->rl_segments;
478 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
479 if (nsegs < 0)
480 return nsegs;
481
482 if (encode_item_present(xdr) < 0)
483 return -EMSGSIZE;
484 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
485 if (unlikely(!segcount))
486 return -EMSGSIZE;
487 /* Actual value encoded below */
488
489 nchunks = 0;
490 do {
491 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
492 true, &mr);
493 if (IS_ERR(seg))
494 goto out_maperr;
495 rpcrdma_mr_push(mr, &req->rl_registered);
496
497 if (encode_rdma_segment(xdr, mr) < 0)
498 return -EMSGSIZE;
499
500 trace_xprtrdma_reply_chunk(rqst->rq_task, mr, nsegs);
501 r_xprt->rx_stats.reply_chunk_count++;
502 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
503 nchunks++;
504 nsegs -= mr->mr_nents;
505 } while (nsegs);
506
507 /* Update count of segments in the Reply chunk */
508 *segcount = cpu_to_be32(nchunks);
509
510 return 0;
511
512out_maperr:
513 if (PTR_ERR(seg) == -EAGAIN)
514 xprt_wait_for_buffer_space(rqst->rq_task, NULL);
515 return PTR_ERR(seg);
516}
517
518/**
519 * rpcrdma_unmap_sendctx - DMA-unmap Send buffers
520 * @sc: sendctx containing SGEs to unmap
521 *
522 */
523void
524rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc)
525{
526 struct rpcrdma_ia *ia = &sc->sc_xprt->rx_ia;
527 struct ib_sge *sge;
528 unsigned int count;
529
530 /* The first two SGEs contain the transport header and
531 * the inline buffer. These are always left mapped so
532 * they can be cheaply re-used.
533 */
534 sge = &sc->sc_sges[2];
535 for (count = sc->sc_unmap_count; count; ++sge, --count)
536 ib_dma_unmap_page(ia->ri_device,
537 sge->addr, sge->length, DMA_TO_DEVICE);
538
539 if (test_and_clear_bit(RPCRDMA_REQ_F_TX_RESOURCES, &sc->sc_req->rl_flags)) {
540 smp_mb__after_atomic();
541 wake_up_bit(&sc->sc_req->rl_flags, RPCRDMA_REQ_F_TX_RESOURCES);
542 }
543}
544
545/* Prepare an SGE for the RPC-over-RDMA transport header.
546 */
547static bool
548rpcrdma_prepare_hdr_sge(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
549 u32 len)
550{
551 struct rpcrdma_sendctx *sc = req->rl_sendctx;
552 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
553 struct ib_sge *sge = sc->sc_sges;
554
555 if (!rpcrdma_dma_map_regbuf(ia, rb))
556 goto out_regbuf;
557 sge->addr = rdmab_addr(rb);
558 sge->length = len;
559 sge->lkey = rdmab_lkey(rb);
560
561 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr,
562 sge->length, DMA_TO_DEVICE);
563 sc->sc_wr.num_sge++;
564 return true;
565
566out_regbuf:
567 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
568 return false;
569}
570
571/* Prepare the Send SGEs. The head and tail iovec, and each entry
572 * in the page list, gets its own SGE.
573 */
574static bool
575rpcrdma_prepare_msg_sges(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
576 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
577{
578 struct rpcrdma_sendctx *sc = req->rl_sendctx;
579 unsigned int sge_no, page_base, len, remaining;
580 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
581 struct ib_device *device = ia->ri_device;
582 struct ib_sge *sge = sc->sc_sges;
583 u32 lkey = ia->ri_pd->local_dma_lkey;
584 struct page *page, **ppages;
585
586 /* The head iovec is straightforward, as it is already
587 * DMA-mapped. Sync the content that has changed.
588 */
589 if (!rpcrdma_dma_map_regbuf(ia, rb))
590 goto out_regbuf;
591 sge_no = 1;
592 sge[sge_no].addr = rdmab_addr(rb);
593 sge[sge_no].length = xdr->head[0].iov_len;
594 sge[sge_no].lkey = rdmab_lkey(rb);
595 ib_dma_sync_single_for_device(rdmab_device(rb), sge[sge_no].addr,
596 sge[sge_no].length, DMA_TO_DEVICE);
597
598 /* If there is a Read chunk, the page list is being handled
599 * via explicit RDMA, and thus is skipped here. However, the
600 * tail iovec may include an XDR pad for the page list, as
601 * well as additional content, and may not reside in the
602 * same page as the head iovec.
603 */
604 if (rtype == rpcrdma_readch) {
605 len = xdr->tail[0].iov_len;
606
607 /* Do not include the tail if it is only an XDR pad */
608 if (len < 4)
609 goto out;
610
611 page = virt_to_page(xdr->tail[0].iov_base);
612 page_base = offset_in_page(xdr->tail[0].iov_base);
613
614 /* If the content in the page list is an odd length,
615 * xdr_write_pages() has added a pad at the beginning
616 * of the tail iovec. Force the tail's non-pad content
617 * to land at the next XDR position in the Send message.
618 */
619 page_base += len & 3;
620 len -= len & 3;
621 goto map_tail;
622 }
623
624 /* If there is a page list present, temporarily DMA map
625 * and prepare an SGE for each page to be sent.
626 */
627 if (xdr->page_len) {
628 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
629 page_base = offset_in_page(xdr->page_base);
630 remaining = xdr->page_len;
631 while (remaining) {
632 sge_no++;
633 if (sge_no > RPCRDMA_MAX_SEND_SGES - 2)
634 goto out_mapping_overflow;
635
636 len = min_t(u32, PAGE_SIZE - page_base, remaining);
637 sge[sge_no].addr = ib_dma_map_page(device, *ppages,
638 page_base, len,
639 DMA_TO_DEVICE);
640 if (ib_dma_mapping_error(device, sge[sge_no].addr))
641 goto out_mapping_err;
642 sge[sge_no].length = len;
643 sge[sge_no].lkey = lkey;
644
645 sc->sc_unmap_count++;
646 ppages++;
647 remaining -= len;
648 page_base = 0;
649 }
650 }
651
652 /* The tail iovec is not always constructed in the same
653 * page where the head iovec resides (see, for example,
654 * gss_wrap_req_priv). To neatly accommodate that case,
655 * DMA map it separately.
656 */
657 if (xdr->tail[0].iov_len) {
658 page = virt_to_page(xdr->tail[0].iov_base);
659 page_base = offset_in_page(xdr->tail[0].iov_base);
660 len = xdr->tail[0].iov_len;
661
662map_tail:
663 sge_no++;
664 sge[sge_no].addr = ib_dma_map_page(device, page,
665 page_base, len,
666 DMA_TO_DEVICE);
667 if (ib_dma_mapping_error(device, sge[sge_no].addr))
668 goto out_mapping_err;
669 sge[sge_no].length = len;
670 sge[sge_no].lkey = lkey;
671 sc->sc_unmap_count++;
672 }
673
674out:
675 sc->sc_wr.num_sge += sge_no;
676 if (sc->sc_unmap_count)
677 __set_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags);
678 return true;
679
680out_regbuf:
681 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
682 return false;
683
684out_mapping_overflow:
685 rpcrdma_unmap_sendctx(sc);
686 pr_err("rpcrdma: too many Send SGEs (%u)\n", sge_no);
687 return false;
688
689out_mapping_err:
690 rpcrdma_unmap_sendctx(sc);
691 pr_err("rpcrdma: Send mapping error\n");
692 return false;
693}
694
695/**
696 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
697 * @r_xprt: controlling transport
698 * @req: context of RPC Call being marshalled
699 * @hdrlen: size of transport header, in bytes
700 * @xdr: xdr_buf containing RPC Call
701 * @rtype: chunk type being encoded
702 *
703 * Returns 0 on success; otherwise a negative errno is returned.
704 */
705int
706rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
707 struct rpcrdma_req *req, u32 hdrlen,
708 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
709{
710 req->rl_sendctx = rpcrdma_sendctx_get_locked(&r_xprt->rx_buf);
711 if (!req->rl_sendctx)
712 return -ENOBUFS;
713 req->rl_sendctx->sc_wr.num_sge = 0;
714 req->rl_sendctx->sc_unmap_count = 0;
715 req->rl_sendctx->sc_req = req;
716 __clear_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags);
717
718 if (!rpcrdma_prepare_hdr_sge(&r_xprt->rx_ia, req, hdrlen))
719 return -EIO;
720
721 if (rtype != rpcrdma_areadch)
722 if (!rpcrdma_prepare_msg_sges(&r_xprt->rx_ia, req, xdr, rtype))
723 return -EIO;
724
725 return 0;
726}
727
728/**
729 * rpcrdma_marshal_req - Marshal and send one RPC request
730 * @r_xprt: controlling transport
731 * @rqst: RPC request to be marshaled
732 *
733 * For the RPC in "rqst", this function:
734 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
735 * - Registers Read, Write, and Reply chunks
736 * - Constructs the transport header
737 * - Posts a Send WR to send the transport header and request
738 *
739 * Returns:
740 * %0 if the RPC was sent successfully,
741 * %-ENOTCONN if the connection was lost,
742 * %-EAGAIN if the caller should call again with the same arguments,
743 * %-ENOBUFS if the caller should call again after a delay,
744 * %-EMSGSIZE if the transport header is too small,
745 * %-EIO if a permanent problem occurred while marshaling.
746 */
747int
748rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
749{
750 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
751 struct xdr_stream *xdr = &req->rl_stream;
752 enum rpcrdma_chunktype rtype, wtype;
753 bool ddp_allowed;
754 __be32 *p;
755 int ret;
756
757 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
758 xdr_init_encode(xdr, &req->rl_hdrbuf,
759 req->rl_rdmabuf->rg_base);
760
761 /* Fixed header fields */
762 ret = -EMSGSIZE;
763 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
764 if (!p)
765 goto out_err;
766 *p++ = rqst->rq_xid;
767 *p++ = rpcrdma_version;
768 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
769
770 /* When the ULP employs a GSS flavor that guarantees integrity
771 * or privacy, direct data placement of individual data items
772 * is not allowed.
773 */
774 ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
775 RPCAUTH_AUTH_DATATOUCH);
776
777 /*
778 * Chunks needed for results?
779 *
780 * o If the expected result is under the inline threshold, all ops
781 * return as inline.
782 * o Large read ops return data as write chunk(s), header as
783 * inline.
784 * o Large non-read ops return as a single reply chunk.
785 */
786 if (rpcrdma_results_inline(r_xprt, rqst))
787 wtype = rpcrdma_noch;
788 else if (ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ)
789 wtype = rpcrdma_writech;
790 else
791 wtype = rpcrdma_replych;
792
793 /*
794 * Chunks needed for arguments?
795 *
796 * o If the total request is under the inline threshold, all ops
797 * are sent as inline.
798 * o Large write ops transmit data as read chunk(s), header as
799 * inline.
800 * o Large non-write ops are sent with the entire message as a
801 * single read chunk (protocol 0-position special case).
802 *
803 * This assumes that the upper layer does not present a request
804 * that both has a data payload, and whose non-data arguments
805 * by themselves are larger than the inline threshold.
806 */
807 if (rpcrdma_args_inline(r_xprt, rqst)) {
808 *p++ = rdma_msg;
809 rtype = rpcrdma_noch;
810 } else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
811 *p++ = rdma_msg;
812 rtype = rpcrdma_readch;
813 } else {
814 r_xprt->rx_stats.nomsg_call_count++;
815 *p++ = rdma_nomsg;
816 rtype = rpcrdma_areadch;
817 }
818
819 /* If this is a retransmit, discard previously registered
820 * chunks. Very likely the connection has been replaced,
821 * so these registrations are invalid and unusable.
822 */
823 while (unlikely(!list_empty(&req->rl_registered))) {
824 struct rpcrdma_mr *mr;
825
826 mr = rpcrdma_mr_pop(&req->rl_registered);
827 rpcrdma_mr_defer_recovery(mr);
828 }
829
830 /* This implementation supports the following combinations
831 * of chunk lists in one RPC-over-RDMA Call message:
832 *
833 * - Read list
834 * - Write list
835 * - Reply chunk
836 * - Read list + Reply chunk
837 *
838 * It might not yet support the following combinations:
839 *
840 * - Read list + Write list
841 *
842 * It does not support the following combinations:
843 *
844 * - Write list + Reply chunk
845 * - Read list + Write list + Reply chunk
846 *
847 * This implementation supports only a single chunk in each
848 * Read or Write list. Thus for example the client cannot
849 * send a Call message with a Position Zero Read chunk and a
850 * regular Read chunk at the same time.
851 */
852 if (rtype != rpcrdma_noch) {
853 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
854 if (ret)
855 goto out_err;
856 }
857 ret = encode_item_not_present(xdr);
858 if (ret)
859 goto out_err;
860
861 if (wtype == rpcrdma_writech) {
862 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
863 if (ret)
864 goto out_err;
865 }
866 ret = encode_item_not_present(xdr);
867 if (ret)
868 goto out_err;
869
870 if (wtype != rpcrdma_replych)
871 ret = encode_item_not_present(xdr);
872 else
873 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
874 if (ret)
875 goto out_err;
876
877 trace_xprtrdma_marshal(rqst, xdr_stream_pos(xdr), rtype, wtype);
878
879 ret = rpcrdma_prepare_send_sges(r_xprt, req, xdr_stream_pos(xdr),
880 &rqst->rq_snd_buf, rtype);
881 if (ret)
882 goto out_err;
883 return 0;
884
885out_err:
886 r_xprt->rx_stats.failed_marshal_count++;
887 return ret;
888}
889
890/**
891 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
892 * @rqst: controlling RPC request
893 * @srcp: points to RPC message payload in receive buffer
894 * @copy_len: remaining length of receive buffer content
895 * @pad: Write chunk pad bytes needed (zero for pure inline)
896 *
897 * The upper layer has set the maximum number of bytes it can
898 * receive in each component of rq_rcv_buf. These values are set in
899 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
900 *
901 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
902 * many cases this function simply updates iov_base pointers in
903 * rq_rcv_buf to point directly to the received reply data, to
904 * avoid copying reply data.
905 *
906 * Returns the count of bytes which had to be memcopied.
907 */
908static unsigned long
909rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
910{
911 unsigned long fixup_copy_count;
912 int i, npages, curlen;
913 char *destp;
914 struct page **ppages;
915 int page_base;
916
917 /* The head iovec is redirected to the RPC reply message
918 * in the receive buffer, to avoid a memcopy.
919 */
920 rqst->rq_rcv_buf.head[0].iov_base = srcp;
921 rqst->rq_private_buf.head[0].iov_base = srcp;
922
923 /* The contents of the receive buffer that follow
924 * head.iov_len bytes are copied into the page list.
925 */
926 curlen = rqst->rq_rcv_buf.head[0].iov_len;
927 if (curlen > copy_len)
928 curlen = copy_len;
929 trace_xprtrdma_fixup(rqst, copy_len, curlen);
930 srcp += curlen;
931 copy_len -= curlen;
932
933 ppages = rqst->rq_rcv_buf.pages +
934 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
935 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
936 fixup_copy_count = 0;
937 if (copy_len && rqst->rq_rcv_buf.page_len) {
938 int pagelist_len;
939
940 pagelist_len = rqst->rq_rcv_buf.page_len;
941 if (pagelist_len > copy_len)
942 pagelist_len = copy_len;
943 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
944 for (i = 0; i < npages; i++) {
945 curlen = PAGE_SIZE - page_base;
946 if (curlen > pagelist_len)
947 curlen = pagelist_len;
948
949 trace_xprtrdma_fixup_pg(rqst, i, srcp,
950 copy_len, curlen);
951 destp = kmap_atomic(ppages[i]);
952 memcpy(destp + page_base, srcp, curlen);
953 flush_dcache_page(ppages[i]);
954 kunmap_atomic(destp);
955 srcp += curlen;
956 copy_len -= curlen;
957 fixup_copy_count += curlen;
958 pagelist_len -= curlen;
959 if (!pagelist_len)
960 break;
961 page_base = 0;
962 }
963
964 /* Implicit padding for the last segment in a Write
965 * chunk is inserted inline at the front of the tail
966 * iovec. The upper layer ignores the content of
967 * the pad. Simply ensure inline content in the tail
968 * that follows the Write chunk is properly aligned.
969 */
970 if (pad)
971 srcp -= pad;
972 }
973
974 /* The tail iovec is redirected to the remaining data
975 * in the receive buffer, to avoid a memcopy.
976 */
977 if (copy_len || pad) {
978 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
979 rqst->rq_private_buf.tail[0].iov_base = srcp;
980 }
981
982 return fixup_copy_count;
983}
984
985/* By convention, backchannel calls arrive via rdma_msg type
986 * messages, and never populate the chunk lists. This makes
987 * the RPC/RDMA header small and fixed in size, so it is
988 * straightforward to check the RPC header's direction field.
989 */
990static bool
991rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
992#if defined(CONFIG_SUNRPC_BACKCHANNEL)
993{
994 struct xdr_stream *xdr = &rep->rr_stream;
995 __be32 *p;
996
997 if (rep->rr_proc != rdma_msg)
998 return false;
999
1000 /* Peek at stream contents without advancing. */
1001 p = xdr_inline_decode(xdr, 0);
1002
1003 /* Chunk lists */
1004 if (*p++ != xdr_zero)
1005 return false;
1006 if (*p++ != xdr_zero)
1007 return false;
1008 if (*p++ != xdr_zero)
1009 return false;
1010
1011 /* RPC header */
1012 if (*p++ != rep->rr_xid)
1013 return false;
1014 if (*p != cpu_to_be32(RPC_CALL))
1015 return false;
1016
1017 /* Now that we are sure this is a backchannel call,
1018 * advance to the RPC header.
1019 */
1020 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1021 if (unlikely(!p))
1022 goto out_short;
1023
1024 rpcrdma_bc_receive_call(r_xprt, rep);
1025 return true;
1026
1027out_short:
1028 pr_warn("RPC/RDMA short backward direction call\n");
1029 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep))
1030 xprt_disconnect_done(&r_xprt->rx_xprt);
1031 return true;
1032}
1033#else /* CONFIG_SUNRPC_BACKCHANNEL */
1034{
1035 return false;
1036}
1037#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1038
1039static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1040{
1041 u32 handle;
1042 u64 offset;
1043 __be32 *p;
1044
1045 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1046 if (unlikely(!p))
1047 return -EIO;
1048
1049 handle = be32_to_cpup(p++);
1050 *length = be32_to_cpup(p++);
1051 xdr_decode_hyper(p, &offset);
1052
1053 trace_xprtrdma_decode_seg(handle, *length, offset);
1054 return 0;
1055}
1056
1057static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1058{
1059 u32 segcount, seglength;
1060 __be32 *p;
1061
1062 p = xdr_inline_decode(xdr, sizeof(*p));
1063 if (unlikely(!p))
1064 return -EIO;
1065
1066 *length = 0;
1067 segcount = be32_to_cpup(p);
1068 while (segcount--) {
1069 if (decode_rdma_segment(xdr, &seglength))
1070 return -EIO;
1071 *length += seglength;
1072 }
1073
1074 return 0;
1075}
1076
1077/* In RPC-over-RDMA Version One replies, a Read list is never
1078 * expected. This decoder is a stub that returns an error if
1079 * a Read list is present.
1080 */
1081static int decode_read_list(struct xdr_stream *xdr)
1082{
1083 __be32 *p;
1084
1085 p = xdr_inline_decode(xdr, sizeof(*p));
1086 if (unlikely(!p))
1087 return -EIO;
1088 if (unlikely(*p != xdr_zero))
1089 return -EIO;
1090 return 0;
1091}
1092
1093/* Supports only one Write chunk in the Write list
1094 */
1095static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1096{
1097 u32 chunklen;
1098 bool first;
1099 __be32 *p;
1100
1101 *length = 0;
1102 first = true;
1103 do {
1104 p = xdr_inline_decode(xdr, sizeof(*p));
1105 if (unlikely(!p))
1106 return -EIO;
1107 if (*p == xdr_zero)
1108 break;
1109 if (!first)
1110 return -EIO;
1111
1112 if (decode_write_chunk(xdr, &chunklen))
1113 return -EIO;
1114 *length += chunklen;
1115 first = false;
1116 } while (true);
1117 return 0;
1118}
1119
1120static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1121{
1122 __be32 *p;
1123
1124 p = xdr_inline_decode(xdr, sizeof(*p));
1125 if (unlikely(!p))
1126 return -EIO;
1127
1128 *length = 0;
1129 if (*p != xdr_zero)
1130 if (decode_write_chunk(xdr, length))
1131 return -EIO;
1132 return 0;
1133}
1134
1135static int
1136rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1137 struct rpc_rqst *rqst)
1138{
1139 struct xdr_stream *xdr = &rep->rr_stream;
1140 u32 writelist, replychunk, rpclen;
1141 char *base;
1142
1143 /* Decode the chunk lists */
1144 if (decode_read_list(xdr))
1145 return -EIO;
1146 if (decode_write_list(xdr, &writelist))
1147 return -EIO;
1148 if (decode_reply_chunk(xdr, &replychunk))
1149 return -EIO;
1150
1151 /* RDMA_MSG sanity checks */
1152 if (unlikely(replychunk))
1153 return -EIO;
1154
1155 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1156 base = (char *)xdr_inline_decode(xdr, 0);
1157 rpclen = xdr_stream_remaining(xdr);
1158 r_xprt->rx_stats.fixup_copy_count +=
1159 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1160
1161 r_xprt->rx_stats.total_rdma_reply += writelist;
1162 return rpclen + xdr_align_size(writelist);
1163}
1164
1165static noinline int
1166rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1167{
1168 struct xdr_stream *xdr = &rep->rr_stream;
1169 u32 writelist, replychunk;
1170
1171 /* Decode the chunk lists */
1172 if (decode_read_list(xdr))
1173 return -EIO;
1174 if (decode_write_list(xdr, &writelist))
1175 return -EIO;
1176 if (decode_reply_chunk(xdr, &replychunk))
1177 return -EIO;
1178
1179 /* RDMA_NOMSG sanity checks */
1180 if (unlikely(writelist))
1181 return -EIO;
1182 if (unlikely(!replychunk))
1183 return -EIO;
1184
1185 /* Reply chunk buffer already is the reply vector */
1186 r_xprt->rx_stats.total_rdma_reply += replychunk;
1187 return replychunk;
1188}
1189
1190static noinline int
1191rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1192 struct rpc_rqst *rqst)
1193{
1194 struct xdr_stream *xdr = &rep->rr_stream;
1195 __be32 *p;
1196
1197 p = xdr_inline_decode(xdr, sizeof(*p));
1198 if (unlikely(!p))
1199 return -EIO;
1200
1201 switch (*p) {
1202 case err_vers:
1203 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1204 if (!p)
1205 break;
1206 dprintk("RPC: %5u: %s: server reports version error (%u-%u)\n",
1207 rqst->rq_task->tk_pid, __func__,
1208 be32_to_cpup(p), be32_to_cpu(*(p + 1)));
1209 break;
1210 case err_chunk:
1211 dprintk("RPC: %5u: %s: server reports header decoding error\n",
1212 rqst->rq_task->tk_pid, __func__);
1213 break;
1214 default:
1215 dprintk("RPC: %5u: %s: server reports unrecognized error %d\n",
1216 rqst->rq_task->tk_pid, __func__, be32_to_cpup(p));
1217 }
1218
1219 r_xprt->rx_stats.bad_reply_count++;
1220 return -EREMOTEIO;
1221}
1222
1223/* Perform XID lookup, reconstruction of the RPC reply, and
1224 * RPC completion while holding the transport lock to ensure
1225 * the rep, rqst, and rq_task pointers remain stable.
1226 */
1227void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1228{
1229 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1230 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1231 struct rpc_rqst *rqst = rep->rr_rqst;
1232 unsigned long cwnd;
1233 int status;
1234
1235 xprt->reestablish_timeout = 0;
1236
1237 switch (rep->rr_proc) {
1238 case rdma_msg:
1239 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1240 break;
1241 case rdma_nomsg:
1242 status = rpcrdma_decode_nomsg(r_xprt, rep);
1243 break;
1244 case rdma_error:
1245 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1246 break;
1247 default:
1248 status = -EIO;
1249 }
1250 if (status < 0)
1251 goto out_badheader;
1252
1253out:
1254 spin_lock(&xprt->recv_lock);
1255 cwnd = xprt->cwnd;
1256 xprt->cwnd = r_xprt->rx_buf.rb_credits << RPC_CWNDSHIFT;
1257 if (xprt->cwnd > cwnd)
1258 xprt_release_rqst_cong(rqst->rq_task);
1259
1260 xprt_complete_rqst(rqst->rq_task, status);
1261 xprt_unpin_rqst(rqst);
1262 spin_unlock(&xprt->recv_lock);
1263 return;
1264
1265/* If the incoming reply terminated a pending RPC, the next
1266 * RPC call will post a replacement receive buffer as it is
1267 * being marshaled.
1268 */
1269out_badheader:
1270 trace_xprtrdma_reply_hdr(rep);
1271 r_xprt->rx_stats.bad_reply_count++;
1272 status = -EIO;
1273 goto out;
1274}
1275
1276void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1277{
1278 /* Invalidate and unmap the data payloads before waking
1279 * the waiting application. This guarantees the memory
1280 * regions are properly fenced from the server before the
1281 * application accesses the data. It also ensures proper
1282 * send flow control: waking the next RPC waits until this
1283 * RPC has relinquished all its Send Queue entries.
1284 */
1285 if (!list_empty(&req->rl_registered))
1286 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
1287 &req->rl_registered);
1288
1289 /* Ensure that any DMA mapped pages associated with
1290 * the Send of the RPC Call have been unmapped before
1291 * allowing the RPC to complete. This protects argument
1292 * memory not controlled by the RPC client from being
1293 * re-used before we're done with it.
1294 */
1295 if (test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1296 r_xprt->rx_stats.reply_waits_for_send++;
1297 out_of_line_wait_on_bit(&req->rl_flags,
1298 RPCRDMA_REQ_F_TX_RESOURCES,
1299 bit_wait,
1300 TASK_UNINTERRUPTIBLE);
1301 }
1302}
1303
1304/* Reply handling runs in the poll worker thread. Anything that
1305 * might wait is deferred to a separate workqueue.
1306 */
1307void rpcrdma_deferred_completion(struct work_struct *work)
1308{
1309 struct rpcrdma_rep *rep =
1310 container_of(work, struct rpcrdma_rep, rr_work);
1311 struct rpcrdma_req *req = rpcr_to_rdmar(rep->rr_rqst);
1312 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1313
1314 trace_xprtrdma_defer_cmp(rep);
1315 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1316 r_xprt->rx_ia.ri_ops->ro_reminv(rep, &req->rl_registered);
1317 rpcrdma_release_rqst(r_xprt, req);
1318 rpcrdma_complete_rqst(rep);
1319}
1320
1321/* Process received RPC/RDMA messages.
1322 *
1323 * Errors must result in the RPC task either being awakened, or
1324 * allowed to timeout, to discover the errors at that time.
1325 */
1326void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1327{
1328 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1329 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1330 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1331 struct rpcrdma_req *req;
1332 struct rpc_rqst *rqst;
1333 u32 credits;
1334 __be32 *p;
1335
1336 if (rep->rr_hdrbuf.head[0].iov_len == 0)
1337 goto out_badstatus;
1338
1339 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1340 rep->rr_hdrbuf.head[0].iov_base);
1341
1342 /* Fixed transport header fields */
1343 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1344 if (unlikely(!p))
1345 goto out_shortreply;
1346 rep->rr_xid = *p++;
1347 rep->rr_vers = *p++;
1348 credits = be32_to_cpu(*p++);
1349 rep->rr_proc = *p++;
1350
1351 if (rep->rr_vers != rpcrdma_version)
1352 goto out_badversion;
1353
1354 if (rpcrdma_is_bcall(r_xprt, rep))
1355 return;
1356
1357 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1358 * get context for handling any incoming chunks.
1359 */
1360 spin_lock(&xprt->recv_lock);
1361 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1362 if (!rqst)
1363 goto out_norqst;
1364 xprt_pin_rqst(rqst);
1365
1366 if (credits == 0)
1367 credits = 1; /* don't deadlock */
1368 else if (credits > buf->rb_max_requests)
1369 credits = buf->rb_max_requests;
1370 buf->rb_credits = credits;
1371
1372 spin_unlock(&xprt->recv_lock);
1373
1374 req = rpcr_to_rdmar(rqst);
1375 req->rl_reply = rep;
1376 rep->rr_rqst = rqst;
1377 clear_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
1378
1379 trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1380
1381 queue_work(rpcrdma_receive_wq, &rep->rr_work);
1382 return;
1383
1384out_badstatus:
1385 rpcrdma_recv_buffer_put(rep);
1386 if (r_xprt->rx_ep.rep_connected == 1) {
1387 r_xprt->rx_ep.rep_connected = -EIO;
1388 rpcrdma_conn_func(&r_xprt->rx_ep);
1389 }
1390 return;
1391
1392out_badversion:
1393 trace_xprtrdma_reply_vers(rep);
1394 goto repost;
1395
1396/* The RPC transaction has already been terminated, or the header
1397 * is corrupt.
1398 */
1399out_norqst:
1400 spin_unlock(&xprt->recv_lock);
1401 trace_xprtrdma_reply_rqst(rep);
1402 goto repost;
1403
1404out_shortreply:
1405 trace_xprtrdma_reply_short(rep);
1406
1407/* If no pending RPC transaction was matched, post a replacement
1408 * receive buffer before returning.
1409 */
1410repost:
1411 r_xprt->rx_stats.bad_reply_count++;
1412 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep))
1413 rpcrdma_recv_buffer_put(rep);
1414}
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#if IS_ENABLED(CONFIG_SUNRPC_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#if IS_ENABLED(CONFIG_SUNRPC_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/* The client can send a request inline as long as the RPCRDMA header
75 * plus the RPC call fit under the transport's inline limit. If the
76 * combined call message size exceeds that limit, the client must use
77 * the read chunk list for this operation.
78 */
79static bool rpcrdma_args_inline(struct rpc_rqst *rqst)
80{
81 unsigned int callsize = RPCRDMA_HDRLEN_MIN + rqst->rq_snd_buf.len;
82
83 return callsize <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst);
84}
85
86/* The client can't know how large the actual reply will be. Thus it
87 * plans for the largest possible reply for that particular ULP
88 * operation. If the maximum combined reply message size exceeds that
89 * limit, the client must provide a write list or a reply chunk for
90 * this request.
91 */
92static bool rpcrdma_results_inline(struct rpc_rqst *rqst)
93{
94 unsigned int repsize = RPCRDMA_HDRLEN_MIN + rqst->rq_rcv_buf.buflen;
95
96 return repsize <= RPCRDMA_INLINE_READ_THRESHOLD(rqst);
97}
98
99static int
100rpcrdma_tail_pullup(struct xdr_buf *buf)
101{
102 size_t tlen = buf->tail[0].iov_len;
103 size_t skip = tlen & 3;
104
105 /* Do not include the tail if it is only an XDR pad */
106 if (tlen < 4)
107 return 0;
108
109 /* xdr_write_pages() adds a pad at the beginning of the tail
110 * if the content in "buf->pages" is unaligned. Force the
111 * tail's actual content to land at the next XDR position
112 * after the head instead.
113 */
114 if (skip) {
115 unsigned char *src, *dst;
116 unsigned int count;
117
118 src = buf->tail[0].iov_base;
119 dst = buf->head[0].iov_base;
120 dst += buf->head[0].iov_len;
121
122 src += skip;
123 tlen -= skip;
124
125 dprintk("RPC: %s: skip=%zu, memmove(%p, %p, %zu)\n",
126 __func__, skip, dst, src, tlen);
127
128 for (count = tlen; count; count--)
129 *dst++ = *src++;
130 }
131
132 return tlen;
133}
134
135/* Split "vec" on page boundaries into segments. FMR registers pages,
136 * not a byte range. Other modes coalesce these segments into a single
137 * MR when they can.
138 */
139static int
140rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
141 int n, int nsegs)
142{
143 size_t page_offset;
144 u32 remaining;
145 char *base;
146
147 base = vec->iov_base;
148 page_offset = offset_in_page(base);
149 remaining = vec->iov_len;
150 while (remaining && n < nsegs) {
151 seg[n].mr_page = NULL;
152 seg[n].mr_offset = base;
153 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
154 remaining -= seg[n].mr_len;
155 base += seg[n].mr_len;
156 ++n;
157 page_offset = 0;
158 }
159 return n;
160}
161
162/*
163 * Chunk assembly from upper layer xdr_buf.
164 *
165 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
166 * elements. Segments are then coalesced when registered, if possible
167 * within the selected memreg mode.
168 *
169 * Returns positive number of segments converted, or a negative errno.
170 */
171
172static int
173rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
174 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
175{
176 int len, n = 0, p;
177 int page_base;
178 struct page **ppages;
179
180 if (pos == 0) {
181 n = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, n, nsegs);
182 if (n == nsegs)
183 return -EIO;
184 }
185
186 len = xdrbuf->page_len;
187 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
188 page_base = xdrbuf->page_base & ~PAGE_MASK;
189 p = 0;
190 while (len && n < nsegs) {
191 if (!ppages[p]) {
192 /* alloc the pagelist for receiving buffer */
193 ppages[p] = alloc_page(GFP_ATOMIC);
194 if (!ppages[p])
195 return -ENOMEM;
196 }
197 seg[n].mr_page = ppages[p];
198 seg[n].mr_offset = (void *)(unsigned long) page_base;
199 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
200 if (seg[n].mr_len > PAGE_SIZE)
201 return -EIO;
202 len -= seg[n].mr_len;
203 ++n;
204 ++p;
205 page_base = 0; /* page offset only applies to first page */
206 }
207
208 /* Message overflows the seg array */
209 if (len && n == nsegs)
210 return -EIO;
211
212 /* When encoding the read list, the tail is always sent inline */
213 if (type == rpcrdma_readch)
214 return n;
215
216 if (xdrbuf->tail[0].iov_len) {
217 /* the rpcrdma protocol allows us to omit any trailing
218 * xdr pad bytes, saving the server an RDMA operation. */
219 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
220 return n;
221 n = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, n, nsegs);
222 if (n == nsegs)
223 return -EIO;
224 }
225
226 return n;
227}
228
229/*
230 * Create read/write chunk lists, and reply chunks, for RDMA
231 *
232 * Assume check against THRESHOLD has been done, and chunks are required.
233 * Assume only encoding one list entry for read|write chunks. The NFSv3
234 * protocol is simple enough to allow this as it only has a single "bulk
235 * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
236 * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
237 *
238 * When used for a single reply chunk (which is a special write
239 * chunk used for the entire reply, rather than just the data), it
240 * is used primarily for READDIR and READLINK which would otherwise
241 * be severely size-limited by a small rdma inline read max. The server
242 * response will come back as an RDMA Write, followed by a message
243 * of type RDMA_NOMSG carrying the xid and length. As a result, reply
244 * chunks do not provide data alignment, however they do not require
245 * "fixup" (moving the response to the upper layer buffer) either.
246 *
247 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
248 *
249 * Read chunklist (a linked list):
250 * N elements, position P (same P for all chunks of same arg!):
251 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
252 *
253 * Write chunklist (a list of (one) counted array):
254 * N elements:
255 * 1 - N - HLOO - HLOO - ... - HLOO - 0
256 *
257 * Reply chunk (a counted array):
258 * N elements:
259 * 1 - N - HLOO - HLOO - ... - HLOO
260 *
261 * Returns positive RPC/RDMA header size, or negative errno.
262 */
263
264static ssize_t
265rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
266 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
267{
268 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
269 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
270 int n, nsegs, nchunks = 0;
271 unsigned int pos;
272 struct rpcrdma_mr_seg *seg = req->rl_segments;
273 struct rpcrdma_read_chunk *cur_rchunk = NULL;
274 struct rpcrdma_write_array *warray = NULL;
275 struct rpcrdma_write_chunk *cur_wchunk = NULL;
276 __be32 *iptr = headerp->rm_body.rm_chunks;
277 int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool);
278
279 if (type == rpcrdma_readch || type == rpcrdma_areadch) {
280 /* a read chunk - server will RDMA Read our memory */
281 cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
282 } else {
283 /* a write or reply chunk - server will RDMA Write our memory */
284 *iptr++ = xdr_zero; /* encode a NULL read chunk list */
285 if (type == rpcrdma_replych)
286 *iptr++ = xdr_zero; /* a NULL write chunk list */
287 warray = (struct rpcrdma_write_array *) iptr;
288 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
289 }
290
291 if (type == rpcrdma_replych || type == rpcrdma_areadch)
292 pos = 0;
293 else
294 pos = target->head[0].iov_len;
295
296 nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
297 if (nsegs < 0)
298 return nsegs;
299
300 map = r_xprt->rx_ia.ri_ops->ro_map;
301 do {
302 n = map(r_xprt, seg, nsegs, cur_wchunk != NULL);
303 if (n <= 0)
304 goto out;
305 if (cur_rchunk) { /* read */
306 cur_rchunk->rc_discrim = xdr_one;
307 /* all read chunks have the same "position" */
308 cur_rchunk->rc_position = cpu_to_be32(pos);
309 cur_rchunk->rc_target.rs_handle =
310 cpu_to_be32(seg->mr_rkey);
311 cur_rchunk->rc_target.rs_length =
312 cpu_to_be32(seg->mr_len);
313 xdr_encode_hyper(
314 (__be32 *)&cur_rchunk->rc_target.rs_offset,
315 seg->mr_base);
316 dprintk("RPC: %s: read chunk "
317 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
318 seg->mr_len, (unsigned long long)seg->mr_base,
319 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
320 cur_rchunk++;
321 r_xprt->rx_stats.read_chunk_count++;
322 } else { /* write/reply */
323 cur_wchunk->wc_target.rs_handle =
324 cpu_to_be32(seg->mr_rkey);
325 cur_wchunk->wc_target.rs_length =
326 cpu_to_be32(seg->mr_len);
327 xdr_encode_hyper(
328 (__be32 *)&cur_wchunk->wc_target.rs_offset,
329 seg->mr_base);
330 dprintk("RPC: %s: %s chunk "
331 "elem %d@0x%llx:0x%x (%s)\n", __func__,
332 (type == rpcrdma_replych) ? "reply" : "write",
333 seg->mr_len, (unsigned long long)seg->mr_base,
334 seg->mr_rkey, n < nsegs ? "more" : "last");
335 cur_wchunk++;
336 if (type == rpcrdma_replych)
337 r_xprt->rx_stats.reply_chunk_count++;
338 else
339 r_xprt->rx_stats.write_chunk_count++;
340 r_xprt->rx_stats.total_rdma_request += seg->mr_len;
341 }
342 nchunks++;
343 seg += n;
344 nsegs -= n;
345 } while (nsegs);
346
347 /* success. all failures return above */
348 req->rl_nchunks = nchunks;
349
350 /*
351 * finish off header. If write, marshal discrim and nchunks.
352 */
353 if (cur_rchunk) {
354 iptr = (__be32 *) cur_rchunk;
355 *iptr++ = xdr_zero; /* finish the read chunk list */
356 *iptr++ = xdr_zero; /* encode a NULL write chunk list */
357 *iptr++ = xdr_zero; /* encode a NULL reply chunk */
358 } else {
359 warray->wc_discrim = xdr_one;
360 warray->wc_nchunks = cpu_to_be32(nchunks);
361 iptr = (__be32 *) cur_wchunk;
362 if (type == rpcrdma_writech) {
363 *iptr++ = xdr_zero; /* finish the write chunk list */
364 *iptr++ = xdr_zero; /* encode a NULL reply chunk */
365 }
366 }
367
368 /*
369 * Return header size.
370 */
371 return (unsigned char *)iptr - (unsigned char *)headerp;
372
373out:
374 for (pos = 0; nchunks--;)
375 pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
376 &req->rl_segments[pos]);
377 return n;
378}
379
380/*
381 * Copy write data inline.
382 * This function is used for "small" requests. Data which is passed
383 * to RPC via iovecs (or page list) is copied directly into the
384 * pre-registered memory buffer for this request. For small amounts
385 * of data, this is efficient. The cutoff value is tunable.
386 */
387static void rpcrdma_inline_pullup(struct rpc_rqst *rqst)
388{
389 int i, npages, curlen;
390 int copy_len;
391 unsigned char *srcp, *destp;
392 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
393 int page_base;
394 struct page **ppages;
395
396 destp = rqst->rq_svec[0].iov_base;
397 curlen = rqst->rq_svec[0].iov_len;
398 destp += curlen;
399
400 dprintk("RPC: %s: destp 0x%p len %d hdrlen %d\n",
401 __func__, destp, rqst->rq_slen, curlen);
402
403 copy_len = rqst->rq_snd_buf.page_len;
404
405 if (rqst->rq_snd_buf.tail[0].iov_len) {
406 curlen = rqst->rq_snd_buf.tail[0].iov_len;
407 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
408 memmove(destp + copy_len,
409 rqst->rq_snd_buf.tail[0].iov_base, curlen);
410 r_xprt->rx_stats.pullup_copy_count += curlen;
411 }
412 dprintk("RPC: %s: tail destp 0x%p len %d\n",
413 __func__, destp + copy_len, curlen);
414 rqst->rq_svec[0].iov_len += curlen;
415 }
416 r_xprt->rx_stats.pullup_copy_count += copy_len;
417
418 page_base = rqst->rq_snd_buf.page_base;
419 ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
420 page_base &= ~PAGE_MASK;
421 npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
422 for (i = 0; copy_len && i < npages; i++) {
423 curlen = PAGE_SIZE - page_base;
424 if (curlen > copy_len)
425 curlen = copy_len;
426 dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n",
427 __func__, i, destp, copy_len, curlen);
428 srcp = kmap_atomic(ppages[i]);
429 memcpy(destp, srcp+page_base, curlen);
430 kunmap_atomic(srcp);
431 rqst->rq_svec[0].iov_len += curlen;
432 destp += curlen;
433 copy_len -= curlen;
434 page_base = 0;
435 }
436 /* header now contains entire send message */
437}
438
439/*
440 * Marshal a request: the primary job of this routine is to choose
441 * the transfer modes. See comments below.
442 *
443 * Uses multiple RDMA IOVs for a request:
444 * [0] -- RPC RDMA header, which uses memory from the *start* of the
445 * preregistered buffer that already holds the RPC data in
446 * its middle.
447 * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
448 * [2] -- optional padding.
449 * [3] -- if padded, header only in [1] and data here.
450 *
451 * Returns zero on success, otherwise a negative errno.
452 */
453
454int
455rpcrdma_marshal_req(struct rpc_rqst *rqst)
456{
457 struct rpc_xprt *xprt = rqst->rq_xprt;
458 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
459 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
460 char *base;
461 size_t rpclen;
462 ssize_t hdrlen;
463 enum rpcrdma_chunktype rtype, wtype;
464 struct rpcrdma_msg *headerp;
465
466#if defined(CONFIG_SUNRPC_BACKCHANNEL)
467 if (test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state))
468 return rpcrdma_bc_marshal_reply(rqst);
469#endif
470
471 /*
472 * rpclen gets amount of data in first buffer, which is the
473 * pre-registered buffer.
474 */
475 base = rqst->rq_svec[0].iov_base;
476 rpclen = rqst->rq_svec[0].iov_len;
477
478 headerp = rdmab_to_msg(req->rl_rdmabuf);
479 /* don't byte-swap XID, it's already done in request */
480 headerp->rm_xid = rqst->rq_xid;
481 headerp->rm_vers = rpcrdma_version;
482 headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
483 headerp->rm_type = rdma_msg;
484
485 /*
486 * Chunks needed for results?
487 *
488 * o Read ops return data as write chunk(s), header as inline.
489 * o If the expected result is under the inline threshold, all ops
490 * return as inline.
491 * o Large non-read ops return as a single reply chunk.
492 */
493 if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
494 wtype = rpcrdma_writech;
495 else if (rpcrdma_results_inline(rqst))
496 wtype = rpcrdma_noch;
497 else
498 wtype = rpcrdma_replych;
499
500 /*
501 * Chunks needed for arguments?
502 *
503 * o If the total request is under the inline threshold, all ops
504 * are sent as inline.
505 * o Large write ops transmit data as read chunk(s), header as
506 * inline.
507 * o Large non-write ops are sent with the entire message as a
508 * single read chunk (protocol 0-position special case).
509 *
510 * This assumes that the upper layer does not present a request
511 * that both has a data payload, and whose non-data arguments
512 * by themselves are larger than the inline threshold.
513 */
514 if (rpcrdma_args_inline(rqst)) {
515 rtype = rpcrdma_noch;
516 } else if (rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
517 rtype = rpcrdma_readch;
518 } else {
519 r_xprt->rx_stats.nomsg_call_count++;
520 headerp->rm_type = htonl(RDMA_NOMSG);
521 rtype = rpcrdma_areadch;
522 rpclen = 0;
523 }
524
525 /* The following simplification is not true forever */
526 if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
527 wtype = rpcrdma_noch;
528 if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
529 dprintk("RPC: %s: cannot marshal multiple chunk lists\n",
530 __func__);
531 return -EIO;
532 }
533
534 hdrlen = RPCRDMA_HDRLEN_MIN;
535
536 /*
537 * Pull up any extra send data into the preregistered buffer.
538 * When padding is in use and applies to the transfer, insert
539 * it and change the message type.
540 */
541 if (rtype == rpcrdma_noch) {
542
543 rpcrdma_inline_pullup(rqst);
544
545 headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
546 headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
547 headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
548 /* new length after pullup */
549 rpclen = rqst->rq_svec[0].iov_len;
550 } else if (rtype == rpcrdma_readch)
551 rpclen += rpcrdma_tail_pullup(&rqst->rq_snd_buf);
552 if (rtype != rpcrdma_noch) {
553 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
554 headerp, rtype);
555 wtype = rtype; /* simplify dprintk */
556
557 } else if (wtype != rpcrdma_noch) {
558 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
559 headerp, wtype);
560 }
561 if (hdrlen < 0)
562 return hdrlen;
563
564 dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd"
565 " headerp 0x%p base 0x%p lkey 0x%x\n",
566 __func__, transfertypes[wtype], hdrlen, rpclen,
567 headerp, base, rdmab_lkey(req->rl_rdmabuf));
568
569 /*
570 * initialize send_iov's - normally only two: rdma chunk header and
571 * single preregistered RPC header buffer, but if padding is present,
572 * then use a preregistered (and zeroed) pad buffer between the RPC
573 * header and any write data. In all non-rdma cases, any following
574 * data has been copied into the RPC header buffer.
575 */
576 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
577 req->rl_send_iov[0].length = hdrlen;
578 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
579
580 req->rl_niovs = 1;
581 if (rtype == rpcrdma_areadch)
582 return 0;
583
584 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
585 req->rl_send_iov[1].length = rpclen;
586 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
587
588 req->rl_niovs = 2;
589 return 0;
590}
591
592/*
593 * Chase down a received write or reply chunklist to get length
594 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
595 */
596static int
597rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
598{
599 unsigned int i, total_len;
600 struct rpcrdma_write_chunk *cur_wchunk;
601 char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
602
603 i = be32_to_cpu(**iptrp);
604 if (i > max)
605 return -1;
606 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
607 total_len = 0;
608 while (i--) {
609 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
610 ifdebug(FACILITY) {
611 u64 off;
612 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
613 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n",
614 __func__,
615 be32_to_cpu(seg->rs_length),
616 (unsigned long long)off,
617 be32_to_cpu(seg->rs_handle));
618 }
619 total_len += be32_to_cpu(seg->rs_length);
620 ++cur_wchunk;
621 }
622 /* check and adjust for properly terminated write chunk */
623 if (wrchunk) {
624 __be32 *w = (__be32 *) cur_wchunk;
625 if (*w++ != xdr_zero)
626 return -1;
627 cur_wchunk = (struct rpcrdma_write_chunk *) w;
628 }
629 if ((char *)cur_wchunk > base + rep->rr_len)
630 return -1;
631
632 *iptrp = (__be32 *) cur_wchunk;
633 return total_len;
634}
635
636/*
637 * Scatter inline received data back into provided iov's.
638 */
639static void
640rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
641{
642 int i, npages, curlen, olen;
643 char *destp;
644 struct page **ppages;
645 int page_base;
646
647 curlen = rqst->rq_rcv_buf.head[0].iov_len;
648 if (curlen > copy_len) { /* write chunk header fixup */
649 curlen = copy_len;
650 rqst->rq_rcv_buf.head[0].iov_len = curlen;
651 }
652
653 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n",
654 __func__, srcp, copy_len, curlen);
655
656 /* Shift pointer for first receive segment only */
657 rqst->rq_rcv_buf.head[0].iov_base = srcp;
658 srcp += curlen;
659 copy_len -= curlen;
660
661 olen = copy_len;
662 i = 0;
663 rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
664 page_base = rqst->rq_rcv_buf.page_base;
665 ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
666 page_base &= ~PAGE_MASK;
667
668 if (copy_len && rqst->rq_rcv_buf.page_len) {
669 npages = PAGE_ALIGN(page_base +
670 rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
671 for (; i < npages; i++) {
672 curlen = PAGE_SIZE - page_base;
673 if (curlen > copy_len)
674 curlen = copy_len;
675 dprintk("RPC: %s: page %d"
676 " srcp 0x%p len %d curlen %d\n",
677 __func__, i, srcp, copy_len, curlen);
678 destp = kmap_atomic(ppages[i]);
679 memcpy(destp + page_base, srcp, curlen);
680 flush_dcache_page(ppages[i]);
681 kunmap_atomic(destp);
682 srcp += curlen;
683 copy_len -= curlen;
684 if (copy_len == 0)
685 break;
686 page_base = 0;
687 }
688 }
689
690 if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
691 curlen = copy_len;
692 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
693 curlen = rqst->rq_rcv_buf.tail[0].iov_len;
694 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
695 memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
696 dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n",
697 __func__, srcp, copy_len, curlen);
698 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
699 copy_len -= curlen; ++i;
700 } else
701 rqst->rq_rcv_buf.tail[0].iov_len = 0;
702
703 if (pad) {
704 /* implicit padding on terminal chunk */
705 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
706 while (pad--)
707 p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
708 }
709
710 if (copy_len)
711 dprintk("RPC: %s: %d bytes in"
712 " %d extra segments (%d lost)\n",
713 __func__, olen, i, copy_len);
714
715 /* TBD avoid a warning from call_decode() */
716 rqst->rq_private_buf = rqst->rq_rcv_buf;
717}
718
719void
720rpcrdma_connect_worker(struct work_struct *work)
721{
722 struct rpcrdma_ep *ep =
723 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
724 struct rpcrdma_xprt *r_xprt =
725 container_of(ep, struct rpcrdma_xprt, rx_ep);
726 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
727
728 spin_lock_bh(&xprt->transport_lock);
729 if (++xprt->connect_cookie == 0) /* maintain a reserved value */
730 ++xprt->connect_cookie;
731 if (ep->rep_connected > 0) {
732 if (!xprt_test_and_set_connected(xprt))
733 xprt_wake_pending_tasks(xprt, 0);
734 } else {
735 if (xprt_test_and_clear_connected(xprt))
736 xprt_wake_pending_tasks(xprt, -ENOTCONN);
737 }
738 spin_unlock_bh(&xprt->transport_lock);
739}
740
741#if defined(CONFIG_SUNRPC_BACKCHANNEL)
742/* By convention, backchannel calls arrive via rdma_msg type
743 * messages, and never populate the chunk lists. This makes
744 * the RPC/RDMA header small and fixed in size, so it is
745 * straightforward to check the RPC header's direction field.
746 */
747static bool
748rpcrdma_is_bcall(struct rpcrdma_msg *headerp)
749{
750 __be32 *p = (__be32 *)headerp;
751
752 if (headerp->rm_type != rdma_msg)
753 return false;
754 if (headerp->rm_body.rm_chunks[0] != xdr_zero)
755 return false;
756 if (headerp->rm_body.rm_chunks[1] != xdr_zero)
757 return false;
758 if (headerp->rm_body.rm_chunks[2] != xdr_zero)
759 return false;
760
761 /* sanity */
762 if (p[7] != headerp->rm_xid)
763 return false;
764 /* call direction */
765 if (p[8] != cpu_to_be32(RPC_CALL))
766 return false;
767
768 return true;
769}
770#endif /* CONFIG_SUNRPC_BACKCHANNEL */
771
772/*
773 * This function is called when an async event is posted to
774 * the connection which changes the connection state. All it
775 * does at this point is mark the connection up/down, the rpc
776 * timers do the rest.
777 */
778void
779rpcrdma_conn_func(struct rpcrdma_ep *ep)
780{
781 schedule_delayed_work(&ep->rep_connect_worker, 0);
782}
783
784/* Process received RPC/RDMA messages.
785 *
786 * Errors must result in the RPC task either being awakened, or
787 * allowed to timeout, to discover the errors at that time.
788 */
789void
790rpcrdma_reply_handler(struct rpcrdma_rep *rep)
791{
792 struct rpcrdma_msg *headerp;
793 struct rpcrdma_req *req;
794 struct rpc_rqst *rqst;
795 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
796 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
797 __be32 *iptr;
798 int rdmalen, status, rmerr;
799 unsigned long cwnd;
800
801 dprintk("RPC: %s: incoming rep %p\n", __func__, rep);
802
803 if (rep->rr_len == RPCRDMA_BAD_LEN)
804 goto out_badstatus;
805 if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
806 goto out_shortreply;
807
808 headerp = rdmab_to_msg(rep->rr_rdmabuf);
809#if defined(CONFIG_SUNRPC_BACKCHANNEL)
810 if (rpcrdma_is_bcall(headerp))
811 goto out_bcall;
812#endif
813
814 /* Match incoming rpcrdma_rep to an rpcrdma_req to
815 * get context for handling any incoming chunks.
816 */
817 spin_lock_bh(&xprt->transport_lock);
818 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
819 if (!rqst)
820 goto out_nomatch;
821
822 req = rpcr_to_rdmar(rqst);
823 if (req->rl_reply)
824 goto out_duplicate;
825
826 /* Sanity checking has passed. We are now committed
827 * to complete this transaction.
828 */
829 list_del_init(&rqst->rq_list);
830 spin_unlock_bh(&xprt->transport_lock);
831 dprintk("RPC: %s: reply %p completes request %p (xid 0x%08x)\n",
832 __func__, rep, req, be32_to_cpu(headerp->rm_xid));
833
834 /* from here on, the reply is no longer an orphan */
835 req->rl_reply = rep;
836 xprt->reestablish_timeout = 0;
837
838 if (headerp->rm_vers != rpcrdma_version)
839 goto out_badversion;
840
841 /* check for expected message types */
842 /* The order of some of these tests is important. */
843 switch (headerp->rm_type) {
844 case rdma_msg:
845 /* never expect read chunks */
846 /* never expect reply chunks (two ways to check) */
847 /* never expect write chunks without having offered RDMA */
848 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
849 (headerp->rm_body.rm_chunks[1] == xdr_zero &&
850 headerp->rm_body.rm_chunks[2] != xdr_zero) ||
851 (headerp->rm_body.rm_chunks[1] != xdr_zero &&
852 req->rl_nchunks == 0))
853 goto badheader;
854 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
855 /* count any expected write chunks in read reply */
856 /* start at write chunk array count */
857 iptr = &headerp->rm_body.rm_chunks[2];
858 rdmalen = rpcrdma_count_chunks(rep,
859 req->rl_nchunks, 1, &iptr);
860 /* check for validity, and no reply chunk after */
861 if (rdmalen < 0 || *iptr++ != xdr_zero)
862 goto badheader;
863 rep->rr_len -=
864 ((unsigned char *)iptr - (unsigned char *)headerp);
865 status = rep->rr_len + rdmalen;
866 r_xprt->rx_stats.total_rdma_reply += rdmalen;
867 /* special case - last chunk may omit padding */
868 if (rdmalen &= 3) {
869 rdmalen = 4 - rdmalen;
870 status += rdmalen;
871 }
872 } else {
873 /* else ordinary inline */
874 rdmalen = 0;
875 iptr = (__be32 *)((unsigned char *)headerp +
876 RPCRDMA_HDRLEN_MIN);
877 rep->rr_len -= RPCRDMA_HDRLEN_MIN;
878 status = rep->rr_len;
879 }
880 /* Fix up the rpc results for upper layer */
881 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
882 break;
883
884 case rdma_nomsg:
885 /* never expect read or write chunks, always reply chunks */
886 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
887 headerp->rm_body.rm_chunks[1] != xdr_zero ||
888 headerp->rm_body.rm_chunks[2] != xdr_one ||
889 req->rl_nchunks == 0)
890 goto badheader;
891 iptr = (__be32 *)((unsigned char *)headerp +
892 RPCRDMA_HDRLEN_MIN);
893 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
894 if (rdmalen < 0)
895 goto badheader;
896 r_xprt->rx_stats.total_rdma_reply += rdmalen;
897 /* Reply chunk buffer already is the reply vector - no fixup. */
898 status = rdmalen;
899 break;
900
901 case rdma_error:
902 goto out_rdmaerr;
903
904badheader:
905 default:
906 dprintk("%s: invalid rpcrdma reply header (type %d):"
907 " chunks[012] == %d %d %d"
908 " expected chunks <= %d\n",
909 __func__, be32_to_cpu(headerp->rm_type),
910 headerp->rm_body.rm_chunks[0],
911 headerp->rm_body.rm_chunks[1],
912 headerp->rm_body.rm_chunks[2],
913 req->rl_nchunks);
914 status = -EIO;
915 r_xprt->rx_stats.bad_reply_count++;
916 break;
917 }
918
919out:
920 /* Invalidate and flush the data payloads before waking the
921 * waiting application. This guarantees the memory region is
922 * properly fenced from the server before the application
923 * accesses the data. It also ensures proper send flow
924 * control: waking the next RPC waits until this RPC has
925 * relinquished all its Send Queue entries.
926 */
927 if (req->rl_nchunks)
928 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);
929
930 spin_lock_bh(&xprt->transport_lock);
931 cwnd = xprt->cwnd;
932 xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
933 if (xprt->cwnd > cwnd)
934 xprt_release_rqst_cong(rqst->rq_task);
935
936 xprt_complete_rqst(rqst->rq_task, status);
937 spin_unlock_bh(&xprt->transport_lock);
938 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
939 __func__, xprt, rqst, status);
940 return;
941
942out_badstatus:
943 rpcrdma_recv_buffer_put(rep);
944 if (r_xprt->rx_ep.rep_connected == 1) {
945 r_xprt->rx_ep.rep_connected = -EIO;
946 rpcrdma_conn_func(&r_xprt->rx_ep);
947 }
948 return;
949
950#if defined(CONFIG_SUNRPC_BACKCHANNEL)
951out_bcall:
952 rpcrdma_bc_receive_call(r_xprt, rep);
953 return;
954#endif
955
956/* If the incoming reply terminated a pending RPC, the next
957 * RPC call will post a replacement receive buffer as it is
958 * being marshaled.
959 */
960out_badversion:
961 dprintk("RPC: %s: invalid version %d\n",
962 __func__, be32_to_cpu(headerp->rm_vers));
963 status = -EIO;
964 r_xprt->rx_stats.bad_reply_count++;
965 goto out;
966
967out_rdmaerr:
968 rmerr = be32_to_cpu(headerp->rm_body.rm_error.rm_err);
969 switch (rmerr) {
970 case ERR_VERS:
971 pr_err("%s: server reports header version error (%u-%u)\n",
972 __func__,
973 be32_to_cpu(headerp->rm_body.rm_error.rm_vers_low),
974 be32_to_cpu(headerp->rm_body.rm_error.rm_vers_high));
975 break;
976 case ERR_CHUNK:
977 pr_err("%s: server reports header decoding error\n",
978 __func__);
979 break;
980 default:
981 pr_err("%s: server reports unknown error %d\n",
982 __func__, rmerr);
983 }
984 status = -EREMOTEIO;
985 r_xprt->rx_stats.bad_reply_count++;
986 goto out;
987
988/* If no pending RPC transaction was matched, post a replacement
989 * receive buffer before returning.
990 */
991out_shortreply:
992 dprintk("RPC: %s: short/invalid reply\n", __func__);
993 goto repost;
994
995out_nomatch:
996 spin_unlock_bh(&xprt->transport_lock);
997 dprintk("RPC: %s: no match for incoming xid 0x%08x len %d\n",
998 __func__, be32_to_cpu(headerp->rm_xid),
999 rep->rr_len);
1000 goto repost;
1001
1002out_duplicate:
1003 spin_unlock_bh(&xprt->transport_lock);
1004 dprintk("RPC: %s: "
1005 "duplicate reply %p to RPC request %p: xid 0x%08x\n",
1006 __func__, rep, req, be32_to_cpu(headerp->rm_xid));
1007
1008repost:
1009 r_xprt->rx_stats.bad_reply_count++;
1010 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
1011 rpcrdma_recv_buffer_put(rep);
1012}