1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
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
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
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.
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
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.
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.
50 #include <linux/highmem.h>
52 #include <linux/sunrpc/svc_rdma.h>
54 #include "xprt_rdma.h"
55 #include <trace/events/rpcrdma.h>
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY RPCDBG_TRANS
61 /* Returns size of largest RPC-over-RDMA header in a Call message
63 * The largest Call header contains a full-size Read list and a
64 * minimal Reply chunk.
66 static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
70 /* Fixed header fields and list discriminators */
71 size = RPCRDMA_HDRLEN_MIN;
73 /* Maximum Read list size */
74 size = maxsegs * rpcrdma_readchunk_maxsz * sizeof(__be32);
76 /* Minimal Read chunk size */
77 size += sizeof(__be32); /* segment count */
78 size += rpcrdma_segment_maxsz * sizeof(__be32);
79 size += sizeof(__be32); /* list discriminator */
84 /* Returns size of largest RPC-over-RDMA header in a Reply message
86 * There is only one Write list or one Reply chunk per Reply
87 * message. The larger list is the Write list.
89 static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
93 /* Fixed header fields and list discriminators */
94 size = RPCRDMA_HDRLEN_MIN;
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 */
105 * rpcrdma_set_max_header_sizes - Initialize inline payload sizes
106 * @r_xprt: transport instance to initialize
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
112 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *r_xprt)
114 unsigned int maxsegs = r_xprt->rx_ia.ri_max_segs;
115 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
117 ep->rep_max_inline_send =
118 ep->rep_inline_send - rpcrdma_max_call_header_size(maxsegs);
119 ep->rep_max_inline_recv =
120 ep->rep_inline_recv - rpcrdma_max_reply_header_size(maxsegs);
123 /* The client can send a request inline as long as the RPCRDMA header
124 * plus the RPC call fit under the transport's inline limit. If the
125 * combined call message size exceeds that limit, the client must use
126 * a Read chunk for this operation.
128 * A Read chunk is also required if sending the RPC call inline would
129 * exceed this device's max_sge limit.
131 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
132 struct rpc_rqst *rqst)
134 struct xdr_buf *xdr = &rqst->rq_snd_buf;
135 unsigned int count, remaining, offset;
137 if (xdr->len > r_xprt->rx_ep.rep_max_inline_send)
141 remaining = xdr->page_len;
142 offset = offset_in_page(xdr->page_base);
143 count = RPCRDMA_MIN_SEND_SGES;
145 remaining -= min_t(unsigned int,
146 PAGE_SIZE - offset, remaining);
148 if (++count > r_xprt->rx_ia.ri_max_send_sges)
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
162 static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
163 struct rpc_rqst *rqst)
165 return rqst->rq_rcv_buf.buflen <= r_xprt->rx_ep.rep_max_inline_recv;
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.
173 rpcrdma_nonpayload_inline(const struct rpcrdma_xprt *r_xprt,
174 const struct rpc_rqst *rqst)
176 const struct xdr_buf *buf = &rqst->rq_rcv_buf;
178 return (buf->head[0].iov_len + buf->tail[0].iov_len) <
179 r_xprt->rx_ep.rep_max_inline_recv;
182 /* Split @vec on page boundaries into SGEs. FMR registers pages, not
183 * a byte range. Other modes coalesce these SGEs into a single MR
186 * Returns pointer to next available SGE, and bumps the total number
189 static struct rpcrdma_mr_seg *
190 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
193 u32 remaining, page_offset;
196 base = vec->iov_base;
197 page_offset = offset_in_page(base);
198 remaining = vec->iov_len;
201 seg->mr_offset = base;
202 seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
203 remaining -= seg->mr_len;
212 /* Convert @xdrbuf into SGEs no larger than a page each. As they
213 * are registered, these SGEs are then coalesced into RDMA segments
214 * when the selected memreg mode supports it.
216 * Returns positive number of SGEs consumed, or a negative errno.
220 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
221 unsigned int pos, enum rpcrdma_chunktype type,
222 struct rpcrdma_mr_seg *seg)
224 unsigned long page_base;
226 struct page **ppages;
230 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
232 len = xdrbuf->page_len;
233 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
234 page_base = offset_in_page(xdrbuf->page_base);
236 /* ACL likes to be lazy in allocating pages - ACLs
237 * are small by default but can get huge.
239 if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
241 *ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
245 seg->mr_page = *ppages;
246 seg->mr_offset = (char *)page_base;
247 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
255 /* When encoding a Read chunk, the tail iovec contains an
256 * XDR pad and may be omitted.
258 if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
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
266 if (type == rpcrdma_writech && r_xprt->rx_ia.ri_implicit_roundup)
269 if (xdrbuf->tail[0].iov_len)
270 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
273 if (unlikely(n > RPCRDMA_MAX_SEGS))
279 encode_item_present(struct xdr_stream *xdr)
283 p = xdr_reserve_space(xdr, sizeof(*p));
292 encode_item_not_present(struct xdr_stream *xdr)
296 p = xdr_reserve_space(xdr, sizeof(*p));
305 xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr *mr)
307 *iptr++ = cpu_to_be32(mr->mr_handle);
308 *iptr++ = cpu_to_be32(mr->mr_length);
309 xdr_encode_hyper(iptr, mr->mr_offset);
313 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
317 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
321 xdr_encode_rdma_segment(p, mr);
326 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
331 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
335 *p++ = xdr_one; /* Item present */
336 *p++ = cpu_to_be32(position);
337 xdr_encode_rdma_segment(p, mr);
341 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
342 struct rpcrdma_req *req,
343 struct rpcrdma_mr_seg *seg,
344 int nsegs, bool writing,
345 struct rpcrdma_mr **mr)
347 *mr = rpcrdma_mr_pop(&req->rl_free_mrs);
349 *mr = rpcrdma_mr_get(r_xprt);
352 trace_xprtrdma_mr_get(req);
356 rpcrdma_mr_push(*mr, &req->rl_registered);
357 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
360 trace_xprtrdma_nomrs(req);
361 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
362 rpcrdma_mrs_refresh(r_xprt);
363 return ERR_PTR(-EAGAIN);
366 /* Register and XDR encode the Read list. Supports encoding a list of read
367 * segments that belong to a single read chunk.
369 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
371 * Read chunklist (a linked list):
372 * N elements, position P (same P for all chunks of same arg!):
373 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
375 * Returns zero on success, or a negative errno if a failure occurred.
376 * @xdr is advanced to the next position in the stream.
378 * Only a single @pos value is currently supported.
380 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
381 struct rpcrdma_req *req,
382 struct rpc_rqst *rqst,
383 enum rpcrdma_chunktype rtype)
385 struct xdr_stream *xdr = &req->rl_stream;
386 struct rpcrdma_mr_seg *seg;
387 struct rpcrdma_mr *mr;
391 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
394 pos = rqst->rq_snd_buf.head[0].iov_len;
395 if (rtype == rpcrdma_areadch)
397 seg = req->rl_segments;
398 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
404 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
408 if (encode_read_segment(xdr, mr, pos) < 0)
411 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
412 r_xprt->rx_stats.read_chunk_count++;
413 nsegs -= mr->mr_nents;
417 return encode_item_not_present(xdr);
420 /* Register and XDR encode the Write list. Supports encoding a list
421 * containing one array of plain segments that belong to a single
424 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
426 * Write chunklist (a list of (one) counted array):
428 * 1 - N - HLOO - HLOO - ... - HLOO - 0
430 * Returns zero on success, or a negative errno if a failure occurred.
431 * @xdr is advanced to the next position in the stream.
433 * Only a single Write chunk is currently supported.
435 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
436 struct rpcrdma_req *req,
437 struct rpc_rqst *rqst,
438 enum rpcrdma_chunktype wtype)
440 struct xdr_stream *xdr = &req->rl_stream;
441 struct rpcrdma_mr_seg *seg;
442 struct rpcrdma_mr *mr;
446 if (wtype != rpcrdma_writech)
449 seg = req->rl_segments;
450 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
451 rqst->rq_rcv_buf.head[0].iov_len,
456 if (encode_item_present(xdr) < 0)
458 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
459 if (unlikely(!segcount))
461 /* Actual value encoded below */
465 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
469 if (encode_rdma_segment(xdr, mr) < 0)
472 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
473 r_xprt->rx_stats.write_chunk_count++;
474 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
476 nsegs -= mr->mr_nents;
479 /* Update count of segments in this Write chunk */
480 *segcount = cpu_to_be32(nchunks);
483 return encode_item_not_present(xdr);
486 /* Register and XDR encode the Reply chunk. Supports encoding an array
487 * of plain segments that belong to a single write (reply) chunk.
489 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
491 * Reply chunk (a counted array):
493 * 1 - N - HLOO - HLOO - ... - HLOO
495 * Returns zero on success, or a negative errno if a failure occurred.
496 * @xdr is advanced to the next position in the stream.
498 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
499 struct rpcrdma_req *req,
500 struct rpc_rqst *rqst,
501 enum rpcrdma_chunktype wtype)
503 struct xdr_stream *xdr = &req->rl_stream;
504 struct rpcrdma_mr_seg *seg;
505 struct rpcrdma_mr *mr;
509 if (wtype != rpcrdma_replych)
510 return encode_item_not_present(xdr);
512 seg = req->rl_segments;
513 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
517 if (encode_item_present(xdr) < 0)
519 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
520 if (unlikely(!segcount))
522 /* Actual value encoded below */
526 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
530 if (encode_rdma_segment(xdr, mr) < 0)
533 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
534 r_xprt->rx_stats.reply_chunk_count++;
535 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
537 nsegs -= mr->mr_nents;
540 /* Update count of segments in the Reply chunk */
541 *segcount = cpu_to_be32(nchunks);
546 static void rpcrdma_sendctx_done(struct kref *kref)
548 struct rpcrdma_req *req =
549 container_of(kref, struct rpcrdma_req, rl_kref);
550 struct rpcrdma_rep *rep = req->rl_reply;
552 rpcrdma_complete_rqst(rep);
553 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
557 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
558 * @sc: sendctx containing SGEs to unmap
561 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
563 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
566 if (!sc->sc_unmap_count)
569 /* The first two SGEs contain the transport header and
570 * the inline buffer. These are always left mapped so
571 * they can be cheaply re-used.
573 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
574 ++sge, --sc->sc_unmap_count)
575 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
578 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
581 /* Prepare an SGE for the RPC-over-RDMA transport header.
583 static bool rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
584 struct rpcrdma_req *req, u32 len)
586 struct rpcrdma_sendctx *sc = req->rl_sendctx;
587 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
588 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
590 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
592 sge->addr = rdmab_addr(rb);
594 sge->lkey = rdmab_lkey(rb);
596 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
601 /* The head iovec is straightforward, as it is usually already
602 * DMA-mapped. Sync the content that has changed.
604 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
605 struct rpcrdma_req *req, unsigned int len)
607 struct rpcrdma_sendctx *sc = req->rl_sendctx;
608 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
609 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
611 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
614 sge->addr = rdmab_addr(rb);
616 sge->lkey = rdmab_lkey(rb);
618 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
623 /* If there is a page list present, DMA map and prepare an
624 * SGE for each page to be sent.
626 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
629 struct rpcrdma_sendctx *sc = req->rl_sendctx;
630 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
631 unsigned int page_base, len, remaining;
632 struct page **ppages;
635 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
636 page_base = offset_in_page(xdr->page_base);
637 remaining = xdr->page_len;
639 sge = &sc->sc_sges[req->rl_wr.num_sge++];
640 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
641 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
642 page_base, len, DMA_TO_DEVICE);
643 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
644 goto out_mapping_err;
647 sge->lkey = rdmab_lkey(rb);
649 sc->sc_unmap_count++;
658 trace_xprtrdma_dma_maperr(sge->addr);
662 /* The tail iovec may include an XDR pad for the page list,
663 * as well as additional content, and may not reside in the
664 * same page as the head iovec.
666 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
668 unsigned int page_base, unsigned int len)
670 struct rpcrdma_sendctx *sc = req->rl_sendctx;
671 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
672 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
673 struct page *page = virt_to_page(xdr->tail[0].iov_base);
675 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
677 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
678 goto out_mapping_err;
681 sge->lkey = rdmab_lkey(rb);
682 ++sc->sc_unmap_count;
686 trace_xprtrdma_dma_maperr(sge->addr);
690 /* Copy the tail to the end of the head buffer.
692 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
693 struct rpcrdma_req *req,
698 dst = (unsigned char *)xdr->head[0].iov_base;
699 dst += xdr->head[0].iov_len + xdr->page_len;
700 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
701 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
704 /* Copy pagelist content into the head buffer.
706 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
707 struct rpcrdma_req *req,
710 unsigned int len, page_base, remaining;
711 struct page **ppages;
712 unsigned char *src, *dst;
714 dst = (unsigned char *)xdr->head[0].iov_base;
715 dst += xdr->head[0].iov_len;
716 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
717 page_base = offset_in_page(xdr->page_base);
718 remaining = xdr->page_len;
720 src = page_address(*ppages);
722 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
723 memcpy(dst, src, len);
724 r_xprt->rx_stats.pullup_copy_count += len;
733 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
734 * When the head, pagelist, and tail are small, a pull-up copy
735 * is considerably less costly than DMA mapping the components
739 * - the caller has already verified that the total length
740 * of the RPC Call body will fit into @rl_sendbuf.
742 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
743 struct rpcrdma_req *req,
746 if (unlikely(xdr->tail[0].iov_len))
747 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
749 if (unlikely(xdr->page_len))
750 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
752 /* The whole RPC message resides in the head iovec now */
753 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
756 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
757 struct rpcrdma_req *req,
760 struct kvec *tail = &xdr->tail[0];
762 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
765 if (!rpcrdma_prepare_pagelist(req, xdr))
768 if (!rpcrdma_prepare_tail_iov(req, xdr,
769 offset_in_page(tail->iov_base),
773 if (req->rl_sendctx->sc_unmap_count)
774 kref_get(&req->rl_kref);
778 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
779 struct rpcrdma_req *req,
782 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
785 /* If there is a Read chunk, the page list is being handled
786 * via explicit RDMA, and thus is skipped here.
789 /* Do not include the tail if it is only an XDR pad */
790 if (xdr->tail[0].iov_len > 3) {
791 unsigned int page_base, len;
793 /* If the content in the page list is an odd length,
794 * xdr_write_pages() adds a pad at the beginning of
795 * the tail iovec. Force the tail's non-pad content to
796 * land at the next XDR position in the Send message.
798 page_base = offset_in_page(xdr->tail[0].iov_base);
799 len = xdr->tail[0].iov_len;
800 page_base += len & 3;
802 if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
804 kref_get(&req->rl_kref);
811 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
812 * @r_xprt: controlling transport
813 * @req: context of RPC Call being marshalled
814 * @hdrlen: size of transport header, in bytes
815 * @xdr: xdr_buf containing RPC Call
816 * @rtype: chunk type being encoded
818 * Returns 0 on success; otherwise a negative errno is returned.
820 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
821 struct rpcrdma_req *req, u32 hdrlen,
823 enum rpcrdma_chunktype rtype)
828 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
829 if (!req->rl_sendctx)
831 req->rl_sendctx->sc_unmap_count = 0;
832 req->rl_sendctx->sc_req = req;
833 kref_init(&req->rl_kref);
834 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
835 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
836 req->rl_wr.num_sge = 0;
837 req->rl_wr.opcode = IB_WR_SEND;
840 if (!rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen))
844 case rpcrdma_noch_pullup:
845 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
848 case rpcrdma_noch_mapped:
849 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
853 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
856 case rpcrdma_areadch:
865 rpcrdma_sendctx_unmap(req->rl_sendctx);
867 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
872 * rpcrdma_marshal_req - Marshal and send one RPC request
873 * @r_xprt: controlling transport
874 * @rqst: RPC request to be marshaled
876 * For the RPC in "rqst", this function:
877 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
878 * - Registers Read, Write, and Reply chunks
879 * - Constructs the transport header
880 * - Posts a Send WR to send the transport header and request
883 * %0 if the RPC was sent successfully,
884 * %-ENOTCONN if the connection was lost,
885 * %-EAGAIN if the caller should call again with the same arguments,
886 * %-ENOBUFS if the caller should call again after a delay,
887 * %-EMSGSIZE if the transport header is too small,
888 * %-EIO if a permanent problem occurred while marshaling.
891 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
893 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
894 struct xdr_stream *xdr = &req->rl_stream;
895 enum rpcrdma_chunktype rtype, wtype;
896 struct xdr_buf *buf = &rqst->rq_snd_buf;
901 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
902 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
905 /* Fixed header fields */
907 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
911 *p++ = rpcrdma_version;
912 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
914 /* When the ULP employs a GSS flavor that guarantees integrity
915 * or privacy, direct data placement of individual data items
918 ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
919 RPCAUTH_AUTH_DATATOUCH);
922 * Chunks needed for results?
924 * o If the expected result is under the inline threshold, all ops
926 * o Large read ops return data as write chunk(s), header as
928 * o Large non-read ops return as a single reply chunk.
930 if (rpcrdma_results_inline(r_xprt, rqst))
931 wtype = rpcrdma_noch;
932 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
933 rpcrdma_nonpayload_inline(r_xprt, rqst))
934 wtype = rpcrdma_writech;
936 wtype = rpcrdma_replych;
939 * Chunks needed for arguments?
941 * o If the total request is under the inline threshold, all ops
942 * are sent as inline.
943 * o Large write ops transmit data as read chunk(s), header as
945 * o Large non-write ops are sent with the entire message as a
946 * single read chunk (protocol 0-position special case).
948 * This assumes that the upper layer does not present a request
949 * that both has a data payload, and whose non-data arguments
950 * by themselves are larger than the inline threshold.
952 if (rpcrdma_args_inline(r_xprt, rqst)) {
954 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
955 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
956 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
958 rtype = rpcrdma_readch;
960 r_xprt->rx_stats.nomsg_call_count++;
962 rtype = rpcrdma_areadch;
965 /* This implementation supports the following combinations
966 * of chunk lists in one RPC-over-RDMA Call message:
971 * - Read list + Reply chunk
973 * It might not yet support the following combinations:
975 * - Read list + Write list
977 * It does not support the following combinations:
979 * - Write list + Reply chunk
980 * - Read list + Write list + Reply chunk
982 * This implementation supports only a single chunk in each
983 * Read or Write list. Thus for example the client cannot
984 * send a Call message with a Position Zero Read chunk and a
985 * regular Read chunk at the same time.
987 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
990 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
993 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
997 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
1002 trace_xprtrdma_marshal(req, rtype, wtype);
1006 trace_xprtrdma_marshal_failed(rqst, ret);
1007 r_xprt->rx_stats.failed_marshal_count++;
1012 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
1013 struct rpcrdma_buffer *buf,
1016 buf->rb_credits = grant;
1017 xprt->cwnd = grant << RPC_CWNDSHIFT;
1020 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
1022 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1024 spin_lock(&xprt->transport_lock);
1025 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1026 spin_unlock(&xprt->transport_lock);
1030 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1031 * @r_xprt: controlling transport instance
1033 * Prepare @r_xprt for the next connection by reinitializing
1034 * its credit grant to one (see RFC 8166, Section 3.3.3).
1036 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1038 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1040 spin_lock(&xprt->transport_lock);
1042 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1043 spin_unlock(&xprt->transport_lock);
1047 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1048 * @rqst: controlling RPC request
1049 * @srcp: points to RPC message payload in receive buffer
1050 * @copy_len: remaining length of receive buffer content
1051 * @pad: Write chunk pad bytes needed (zero for pure inline)
1053 * The upper layer has set the maximum number of bytes it can
1054 * receive in each component of rq_rcv_buf. These values are set in
1055 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1057 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1058 * many cases this function simply updates iov_base pointers in
1059 * rq_rcv_buf to point directly to the received reply data, to
1060 * avoid copying reply data.
1062 * Returns the count of bytes which had to be memcopied.
1064 static unsigned long
1065 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1067 unsigned long fixup_copy_count;
1068 int i, npages, curlen;
1070 struct page **ppages;
1073 /* The head iovec is redirected to the RPC reply message
1074 * in the receive buffer, to avoid a memcopy.
1076 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1077 rqst->rq_private_buf.head[0].iov_base = srcp;
1079 /* The contents of the receive buffer that follow
1080 * head.iov_len bytes are copied into the page list.
1082 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1083 if (curlen > copy_len)
1088 ppages = rqst->rq_rcv_buf.pages +
1089 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1090 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1091 fixup_copy_count = 0;
1092 if (copy_len && rqst->rq_rcv_buf.page_len) {
1095 pagelist_len = rqst->rq_rcv_buf.page_len;
1096 if (pagelist_len > copy_len)
1097 pagelist_len = copy_len;
1098 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1099 for (i = 0; i < npages; i++) {
1100 curlen = PAGE_SIZE - page_base;
1101 if (curlen > pagelist_len)
1102 curlen = pagelist_len;
1104 destp = kmap_atomic(ppages[i]);
1105 memcpy(destp + page_base, srcp, curlen);
1106 flush_dcache_page(ppages[i]);
1107 kunmap_atomic(destp);
1110 fixup_copy_count += curlen;
1111 pagelist_len -= curlen;
1117 /* Implicit padding for the last segment in a Write
1118 * chunk is inserted inline at the front of the tail
1119 * iovec. The upper layer ignores the content of
1120 * the pad. Simply ensure inline content in the tail
1121 * that follows the Write chunk is properly aligned.
1127 /* The tail iovec is redirected to the remaining data
1128 * in the receive buffer, to avoid a memcopy.
1130 if (copy_len || pad) {
1131 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1132 rqst->rq_private_buf.tail[0].iov_base = srcp;
1135 if (fixup_copy_count)
1136 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1137 return fixup_copy_count;
1140 /* By convention, backchannel calls arrive via rdma_msg type
1141 * messages, and never populate the chunk lists. This makes
1142 * the RPC/RDMA header small and fixed in size, so it is
1143 * straightforward to check the RPC header's direction field.
1146 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1147 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1149 struct xdr_stream *xdr = &rep->rr_stream;
1152 if (rep->rr_proc != rdma_msg)
1155 /* Peek at stream contents without advancing. */
1156 p = xdr_inline_decode(xdr, 0);
1159 if (*p++ != xdr_zero)
1161 if (*p++ != xdr_zero)
1163 if (*p++ != xdr_zero)
1167 if (*p++ != rep->rr_xid)
1169 if (*p != cpu_to_be32(RPC_CALL))
1172 /* Now that we are sure this is a backchannel call,
1173 * advance to the RPC header.
1175 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1179 rpcrdma_bc_receive_call(r_xprt, rep);
1183 pr_warn("RPC/RDMA short backward direction call\n");
1186 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1190 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1192 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1198 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1202 handle = be32_to_cpup(p++);
1203 *length = be32_to_cpup(p++);
1204 xdr_decode_hyper(p, &offset);
1206 trace_xprtrdma_decode_seg(handle, *length, offset);
1210 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1212 u32 segcount, seglength;
1215 p = xdr_inline_decode(xdr, sizeof(*p));
1220 segcount = be32_to_cpup(p);
1221 while (segcount--) {
1222 if (decode_rdma_segment(xdr, &seglength))
1224 *length += seglength;
1230 /* In RPC-over-RDMA Version One replies, a Read list is never
1231 * expected. This decoder is a stub that returns an error if
1232 * a Read list is present.
1234 static int decode_read_list(struct xdr_stream *xdr)
1238 p = xdr_inline_decode(xdr, sizeof(*p));
1241 if (unlikely(*p != xdr_zero))
1246 /* Supports only one Write chunk in the Write list
1248 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1257 p = xdr_inline_decode(xdr, sizeof(*p));
1265 if (decode_write_chunk(xdr, &chunklen))
1267 *length += chunklen;
1273 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1277 p = xdr_inline_decode(xdr, sizeof(*p));
1283 if (decode_write_chunk(xdr, length))
1289 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1290 struct rpc_rqst *rqst)
1292 struct xdr_stream *xdr = &rep->rr_stream;
1293 u32 writelist, replychunk, rpclen;
1296 /* Decode the chunk lists */
1297 if (decode_read_list(xdr))
1299 if (decode_write_list(xdr, &writelist))
1301 if (decode_reply_chunk(xdr, &replychunk))
1304 /* RDMA_MSG sanity checks */
1305 if (unlikely(replychunk))
1308 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1309 base = (char *)xdr_inline_decode(xdr, 0);
1310 rpclen = xdr_stream_remaining(xdr);
1311 r_xprt->rx_stats.fixup_copy_count +=
1312 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1314 r_xprt->rx_stats.total_rdma_reply += writelist;
1315 return rpclen + xdr_align_size(writelist);
1319 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1321 struct xdr_stream *xdr = &rep->rr_stream;
1322 u32 writelist, replychunk;
1324 /* Decode the chunk lists */
1325 if (decode_read_list(xdr))
1327 if (decode_write_list(xdr, &writelist))
1329 if (decode_reply_chunk(xdr, &replychunk))
1332 /* RDMA_NOMSG sanity checks */
1333 if (unlikely(writelist))
1335 if (unlikely(!replychunk))
1338 /* Reply chunk buffer already is the reply vector */
1339 r_xprt->rx_stats.total_rdma_reply += replychunk;
1344 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1345 struct rpc_rqst *rqst)
1347 struct xdr_stream *xdr = &rep->rr_stream;
1350 p = xdr_inline_decode(xdr, sizeof(*p));
1356 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1359 dprintk("RPC: %s: server reports "
1360 "version error (%u-%u), xid %08x\n", __func__,
1361 be32_to_cpup(p), be32_to_cpu(*(p + 1)),
1362 be32_to_cpu(rep->rr_xid));
1365 dprintk("RPC: %s: server reports "
1366 "header decoding error, xid %08x\n", __func__,
1367 be32_to_cpu(rep->rr_xid));
1370 dprintk("RPC: %s: server reports "
1371 "unrecognized error %d, xid %08x\n", __func__,
1372 be32_to_cpup(p), be32_to_cpu(rep->rr_xid));
1375 r_xprt->rx_stats.bad_reply_count++;
1379 /* Perform XID lookup, reconstruction of the RPC reply, and
1380 * RPC completion while holding the transport lock to ensure
1381 * the rep, rqst, and rq_task pointers remain stable.
1383 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1385 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1386 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1387 struct rpc_rqst *rqst = rep->rr_rqst;
1390 switch (rep->rr_proc) {
1392 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1395 status = rpcrdma_decode_nomsg(r_xprt, rep);
1398 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1407 spin_lock(&xprt->queue_lock);
1408 xprt_complete_rqst(rqst->rq_task, status);
1409 xprt_unpin_rqst(rqst);
1410 spin_unlock(&xprt->queue_lock);
1413 /* If the incoming reply terminated a pending RPC, the next
1414 * RPC call will post a replacement receive buffer as it is
1418 trace_xprtrdma_reply_hdr(rep);
1419 r_xprt->rx_stats.bad_reply_count++;
1423 static void rpcrdma_reply_done(struct kref *kref)
1425 struct rpcrdma_req *req =
1426 container_of(kref, struct rpcrdma_req, rl_kref);
1428 rpcrdma_complete_rqst(req->rl_reply);
1432 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1433 * @rep: Incoming rpcrdma_rep object to process
1435 * Errors must result in the RPC task either being awakened, or
1436 * allowed to timeout, to discover the errors at that time.
1438 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1440 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1441 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1442 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1443 struct rpcrdma_req *req;
1444 struct rpc_rqst *rqst;
1448 /* Any data means we had a useful conversation, so
1449 * then we don't need to delay the next reconnect.
1451 if (xprt->reestablish_timeout)
1452 xprt->reestablish_timeout = 0;
1454 /* Fixed transport header fields */
1455 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1456 rep->rr_hdrbuf.head[0].iov_base, NULL);
1457 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1459 goto out_shortreply;
1461 rep->rr_vers = *p++;
1462 credits = be32_to_cpu(*p++);
1463 rep->rr_proc = *p++;
1465 if (rep->rr_vers != rpcrdma_version)
1466 goto out_badversion;
1468 if (rpcrdma_is_bcall(r_xprt, rep))
1471 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1472 * get context for handling any incoming chunks.
1474 spin_lock(&xprt->queue_lock);
1475 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1478 xprt_pin_rqst(rqst);
1479 spin_unlock(&xprt->queue_lock);
1482 credits = 1; /* don't deadlock */
1483 else if (credits > buf->rb_max_requests)
1484 credits = buf->rb_max_requests;
1485 if (buf->rb_credits != credits)
1486 rpcrdma_update_cwnd(r_xprt, credits);
1487 rpcrdma_post_recvs(r_xprt, false);
1489 req = rpcr_to_rdmar(rqst);
1490 if (req->rl_reply) {
1491 trace_xprtrdma_leaked_rep(rqst, req->rl_reply);
1492 rpcrdma_recv_buffer_put(req->rl_reply);
1494 req->rl_reply = rep;
1495 rep->rr_rqst = rqst;
1497 trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1499 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1500 frwr_reminv(rep, &req->rl_registered);
1501 if (!list_empty(&req->rl_registered))
1502 frwr_unmap_async(r_xprt, req);
1503 /* LocalInv completion will complete the RPC */
1505 kref_put(&req->rl_kref, rpcrdma_reply_done);
1509 trace_xprtrdma_reply_vers(rep);
1513 spin_unlock(&xprt->queue_lock);
1514 trace_xprtrdma_reply_rqst(rep);
1518 trace_xprtrdma_reply_short(rep);
1521 rpcrdma_recv_buffer_put(rep);