1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2020, Oracle and/or its affiliates.
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 * @ep: endpoint 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_ep *ep)
114 unsigned int maxsegs = ep->re_max_rdma_segs;
116 ep->re_max_inline_send =
117 ep->re_inline_send - rpcrdma_max_call_header_size(maxsegs);
118 ep->re_max_inline_recv =
119 ep->re_inline_recv - rpcrdma_max_reply_header_size(maxsegs);
122 /* The client can send a request inline as long as the RPCRDMA header
123 * plus the RPC call fit under the transport's inline limit. If the
124 * combined call message size exceeds that limit, the client must use
125 * a Read chunk for this operation.
127 * A Read chunk is also required if sending the RPC call inline would
128 * exceed this device's max_sge limit.
130 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
131 struct rpc_rqst *rqst)
133 struct xdr_buf *xdr = &rqst->rq_snd_buf;
134 struct rpcrdma_ep *ep = r_xprt->rx_ep;
135 unsigned int count, remaining, offset;
137 if (xdr->len > ep->re_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 > ep->re_attr.cap.max_send_sge)
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->re_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->re_max_inline_recv;
182 /* ACL likes to be lazy in allocating pages. For TCP, these
183 * pages can be allocated during receive processing. Not true
184 * for RDMA, which must always provision receive buffers
188 rpcrdma_alloc_sparse_pages(struct xdr_buf *buf)
190 struct page **ppages;
194 ppages = buf->pages + (buf->page_base >> PAGE_SHIFT);
197 *ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
207 /* Convert @vec to a single SGL element.
209 * Returns pointer to next available SGE, and bumps the total number
212 static struct rpcrdma_mr_seg *
213 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
216 seg->mr_page = virt_to_page(vec->iov_base);
217 seg->mr_offset = vec->iov_base;
218 seg->mr_len = vec->iov_len;
224 /* Convert @xdrbuf into SGEs no larger than a page each. As they
225 * are registered, these SGEs are then coalesced into RDMA segments
226 * when the selected memreg mode supports it.
228 * Returns positive number of SGEs consumed, or a negative errno.
232 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
233 unsigned int pos, enum rpcrdma_chunktype type,
234 struct rpcrdma_mr_seg *seg)
236 unsigned long page_base;
238 struct page **ppages;
242 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
244 len = xdrbuf->page_len;
245 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
246 page_base = offset_in_page(xdrbuf->page_base);
248 seg->mr_page = *ppages;
249 seg->mr_offset = (char *)page_base;
250 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
258 /* When encoding a Read chunk, the tail iovec contains an
259 * XDR pad and may be omitted.
261 if (type == rpcrdma_readch && r_xprt->rx_ep->re_implicit_roundup)
264 /* When encoding a Write chunk, some servers need to see an
265 * extra segment for non-XDR-aligned Write chunks. The upper
266 * layer provides space in the tail iovec that may be used
269 if (type == rpcrdma_writech && r_xprt->rx_ep->re_implicit_roundup)
272 if (xdrbuf->tail[0].iov_len)
273 rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
276 if (unlikely(n > RPCRDMA_MAX_SEGS))
282 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
286 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
290 xdr_encode_rdma_segment(p, mr->mr_handle, mr->mr_length, mr->mr_offset);
295 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
300 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
304 *p++ = xdr_one; /* Item present */
305 xdr_encode_read_segment(p, position, mr->mr_handle, mr->mr_length,
310 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
311 struct rpcrdma_req *req,
312 struct rpcrdma_mr_seg *seg,
313 int nsegs, bool writing,
314 struct rpcrdma_mr **mr)
316 *mr = rpcrdma_mr_pop(&req->rl_free_mrs);
318 *mr = rpcrdma_mr_get(r_xprt);
324 rpcrdma_mr_push(*mr, &req->rl_registered);
325 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
328 trace_xprtrdma_nomrs_err(r_xprt, req);
329 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
330 rpcrdma_mrs_refresh(r_xprt);
331 return ERR_PTR(-EAGAIN);
334 /* Register and XDR encode the Read list. Supports encoding a list of read
335 * segments that belong to a single read chunk.
337 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
339 * Read chunklist (a linked list):
340 * N elements, position P (same P for all chunks of same arg!):
341 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
343 * Returns zero on success, or a negative errno if a failure occurred.
344 * @xdr is advanced to the next position in the stream.
346 * Only a single @pos value is currently supported.
348 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
349 struct rpcrdma_req *req,
350 struct rpc_rqst *rqst,
351 enum rpcrdma_chunktype rtype)
353 struct xdr_stream *xdr = &req->rl_stream;
354 struct rpcrdma_mr_seg *seg;
355 struct rpcrdma_mr *mr;
359 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
362 pos = rqst->rq_snd_buf.head[0].iov_len;
363 if (rtype == rpcrdma_areadch)
365 seg = req->rl_segments;
366 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
372 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
376 if (encode_read_segment(xdr, mr, pos) < 0)
379 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
380 r_xprt->rx_stats.read_chunk_count++;
381 nsegs -= mr->mr_nents;
385 if (xdr_stream_encode_item_absent(xdr) < 0)
390 /* Register and XDR encode the Write list. Supports encoding a list
391 * containing one array of plain segments that belong to a single
394 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
396 * Write chunklist (a list of (one) counted array):
398 * 1 - N - HLOO - HLOO - ... - HLOO - 0
400 * Returns zero on success, or a negative errno if a failure occurred.
401 * @xdr is advanced to the next position in the stream.
403 * Only a single Write chunk is currently supported.
405 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
406 struct rpcrdma_req *req,
407 struct rpc_rqst *rqst,
408 enum rpcrdma_chunktype wtype)
410 struct xdr_stream *xdr = &req->rl_stream;
411 struct rpcrdma_mr_seg *seg;
412 struct rpcrdma_mr *mr;
416 if (wtype != rpcrdma_writech)
419 seg = req->rl_segments;
420 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
421 rqst->rq_rcv_buf.head[0].iov_len,
426 if (xdr_stream_encode_item_present(xdr) < 0)
428 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
429 if (unlikely(!segcount))
431 /* Actual value encoded below */
435 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
439 if (encode_rdma_segment(xdr, mr) < 0)
442 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
443 r_xprt->rx_stats.write_chunk_count++;
444 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
446 nsegs -= mr->mr_nents;
449 /* Update count of segments in this Write chunk */
450 *segcount = cpu_to_be32(nchunks);
453 if (xdr_stream_encode_item_absent(xdr) < 0)
458 /* Register and XDR encode the Reply chunk. Supports encoding an array
459 * of plain segments that belong to a single write (reply) chunk.
461 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
463 * Reply chunk (a counted array):
465 * 1 - N - HLOO - HLOO - ... - HLOO
467 * Returns zero on success, or a negative errno if a failure occurred.
468 * @xdr is advanced to the next position in the stream.
470 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
471 struct rpcrdma_req *req,
472 struct rpc_rqst *rqst,
473 enum rpcrdma_chunktype wtype)
475 struct xdr_stream *xdr = &req->rl_stream;
476 struct rpcrdma_mr_seg *seg;
477 struct rpcrdma_mr *mr;
481 if (wtype != rpcrdma_replych) {
482 if (xdr_stream_encode_item_absent(xdr) < 0)
487 seg = req->rl_segments;
488 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
492 if (xdr_stream_encode_item_present(xdr) < 0)
494 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
495 if (unlikely(!segcount))
497 /* Actual value encoded below */
501 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
505 if (encode_rdma_segment(xdr, mr) < 0)
508 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
509 r_xprt->rx_stats.reply_chunk_count++;
510 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
512 nsegs -= mr->mr_nents;
515 /* Update count of segments in the Reply chunk */
516 *segcount = cpu_to_be32(nchunks);
521 static void rpcrdma_sendctx_done(struct kref *kref)
523 struct rpcrdma_req *req =
524 container_of(kref, struct rpcrdma_req, rl_kref);
525 struct rpcrdma_rep *rep = req->rl_reply;
527 rpcrdma_complete_rqst(rep);
528 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
532 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
533 * @sc: sendctx containing SGEs to unmap
536 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
538 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
541 if (!sc->sc_unmap_count)
544 /* The first two SGEs contain the transport header and
545 * the inline buffer. These are always left mapped so
546 * they can be cheaply re-used.
548 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
549 ++sge, --sc->sc_unmap_count)
550 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
553 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
556 /* Prepare an SGE for the RPC-over-RDMA transport header.
558 static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
559 struct rpcrdma_req *req, u32 len)
561 struct rpcrdma_sendctx *sc = req->rl_sendctx;
562 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
563 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
565 sge->addr = rdmab_addr(rb);
567 sge->lkey = rdmab_lkey(rb);
569 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
573 /* The head iovec is straightforward, as it is usually already
574 * DMA-mapped. Sync the content that has changed.
576 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
577 struct rpcrdma_req *req, unsigned int len)
579 struct rpcrdma_sendctx *sc = req->rl_sendctx;
580 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
581 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
583 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
586 sge->addr = rdmab_addr(rb);
588 sge->lkey = rdmab_lkey(rb);
590 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
595 /* If there is a page list present, DMA map and prepare an
596 * SGE for each page to be sent.
598 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
601 struct rpcrdma_sendctx *sc = req->rl_sendctx;
602 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
603 unsigned int page_base, len, remaining;
604 struct page **ppages;
607 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
608 page_base = offset_in_page(xdr->page_base);
609 remaining = xdr->page_len;
611 sge = &sc->sc_sges[req->rl_wr.num_sge++];
612 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
613 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
614 page_base, len, DMA_TO_DEVICE);
615 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
616 goto out_mapping_err;
619 sge->lkey = rdmab_lkey(rb);
621 sc->sc_unmap_count++;
630 trace_xprtrdma_dma_maperr(sge->addr);
634 /* The tail iovec may include an XDR pad for the page list,
635 * as well as additional content, and may not reside in the
636 * same page as the head iovec.
638 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
640 unsigned int page_base, unsigned int len)
642 struct rpcrdma_sendctx *sc = req->rl_sendctx;
643 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
644 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
645 struct page *page = virt_to_page(xdr->tail[0].iov_base);
647 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
649 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
650 goto out_mapping_err;
653 sge->lkey = rdmab_lkey(rb);
654 ++sc->sc_unmap_count;
658 trace_xprtrdma_dma_maperr(sge->addr);
662 /* Copy the tail to the end of the head buffer.
664 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
665 struct rpcrdma_req *req,
670 dst = (unsigned char *)xdr->head[0].iov_base;
671 dst += xdr->head[0].iov_len + xdr->page_len;
672 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
673 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
676 /* Copy pagelist content into the head buffer.
678 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
679 struct rpcrdma_req *req,
682 unsigned int len, page_base, remaining;
683 struct page **ppages;
684 unsigned char *src, *dst;
686 dst = (unsigned char *)xdr->head[0].iov_base;
687 dst += xdr->head[0].iov_len;
688 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
689 page_base = offset_in_page(xdr->page_base);
690 remaining = xdr->page_len;
692 src = page_address(*ppages);
694 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
695 memcpy(dst, src, len);
696 r_xprt->rx_stats.pullup_copy_count += len;
705 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
706 * When the head, pagelist, and tail are small, a pull-up copy
707 * is considerably less costly than DMA mapping the components
711 * - the caller has already verified that the total length
712 * of the RPC Call body will fit into @rl_sendbuf.
714 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
715 struct rpcrdma_req *req,
718 if (unlikely(xdr->tail[0].iov_len))
719 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
721 if (unlikely(xdr->page_len))
722 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
724 /* The whole RPC message resides in the head iovec now */
725 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
728 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
729 struct rpcrdma_req *req,
732 struct kvec *tail = &xdr->tail[0];
734 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
737 if (!rpcrdma_prepare_pagelist(req, xdr))
740 if (!rpcrdma_prepare_tail_iov(req, xdr,
741 offset_in_page(tail->iov_base),
745 if (req->rl_sendctx->sc_unmap_count)
746 kref_get(&req->rl_kref);
750 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
751 struct rpcrdma_req *req,
754 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
757 /* If there is a Read chunk, the page list is being handled
758 * via explicit RDMA, and thus is skipped here.
761 /* Do not include the tail if it is only an XDR pad */
762 if (xdr->tail[0].iov_len > 3) {
763 unsigned int page_base, len;
765 /* If the content in the page list is an odd length,
766 * xdr_write_pages() adds a pad at the beginning of
767 * the tail iovec. Force the tail's non-pad content to
768 * land at the next XDR position in the Send message.
770 page_base = offset_in_page(xdr->tail[0].iov_base);
771 len = xdr->tail[0].iov_len;
772 page_base += len & 3;
774 if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
776 kref_get(&req->rl_kref);
783 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
784 * @r_xprt: controlling transport
785 * @req: context of RPC Call being marshalled
786 * @hdrlen: size of transport header, in bytes
787 * @xdr: xdr_buf containing RPC Call
788 * @rtype: chunk type being encoded
790 * Returns 0 on success; otherwise a negative errno is returned.
792 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
793 struct rpcrdma_req *req, u32 hdrlen,
795 enum rpcrdma_chunktype rtype)
800 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
801 if (!req->rl_sendctx)
803 req->rl_sendctx->sc_unmap_count = 0;
804 req->rl_sendctx->sc_req = req;
805 kref_init(&req->rl_kref);
806 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
807 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
808 req->rl_wr.num_sge = 0;
809 req->rl_wr.opcode = IB_WR_SEND;
811 rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
815 case rpcrdma_noch_pullup:
816 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
819 case rpcrdma_noch_mapped:
820 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
824 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
827 case rpcrdma_areadch:
836 rpcrdma_sendctx_unmap(req->rl_sendctx);
838 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
843 * rpcrdma_marshal_req - Marshal and send one RPC request
844 * @r_xprt: controlling transport
845 * @rqst: RPC request to be marshaled
847 * For the RPC in "rqst", this function:
848 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
849 * - Registers Read, Write, and Reply chunks
850 * - Constructs the transport header
851 * - Posts a Send WR to send the transport header and request
854 * %0 if the RPC was sent successfully,
855 * %-ENOTCONN if the connection was lost,
856 * %-EAGAIN if the caller should call again with the same arguments,
857 * %-ENOBUFS if the caller should call again after a delay,
858 * %-EMSGSIZE if the transport header is too small,
859 * %-EIO if a permanent problem occurred while marshaling.
862 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
864 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
865 struct xdr_stream *xdr = &req->rl_stream;
866 enum rpcrdma_chunktype rtype, wtype;
867 struct xdr_buf *buf = &rqst->rq_snd_buf;
872 if (unlikely(rqst->rq_rcv_buf.flags & XDRBUF_SPARSE_PAGES)) {
873 ret = rpcrdma_alloc_sparse_pages(&rqst->rq_rcv_buf);
878 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
879 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
882 /* Fixed header fields */
884 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
888 *p++ = rpcrdma_version;
889 *p++ = r_xprt->rx_buf.rb_max_requests;
891 /* When the ULP employs a GSS flavor that guarantees integrity
892 * or privacy, direct data placement of individual data items
895 ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
896 &rqst->rq_cred->cr_auth->au_flags);
899 * Chunks needed for results?
901 * o If the expected result is under the inline threshold, all ops
903 * o Large read ops return data as write chunk(s), header as
905 * o Large non-read ops return as a single reply chunk.
907 if (rpcrdma_results_inline(r_xprt, rqst))
908 wtype = rpcrdma_noch;
909 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
910 rpcrdma_nonpayload_inline(r_xprt, rqst))
911 wtype = rpcrdma_writech;
913 wtype = rpcrdma_replych;
916 * Chunks needed for arguments?
918 * o If the total request is under the inline threshold, all ops
919 * are sent as inline.
920 * o Large write ops transmit data as read chunk(s), header as
922 * o Large non-write ops are sent with the entire message as a
923 * single read chunk (protocol 0-position special case).
925 * This assumes that the upper layer does not present a request
926 * that both has a data payload, and whose non-data arguments
927 * by themselves are larger than the inline threshold.
929 if (rpcrdma_args_inline(r_xprt, rqst)) {
931 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
932 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
933 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
935 rtype = rpcrdma_readch;
937 r_xprt->rx_stats.nomsg_call_count++;
939 rtype = rpcrdma_areadch;
942 /* This implementation supports the following combinations
943 * of chunk lists in one RPC-over-RDMA Call message:
948 * - Read list + Reply chunk
950 * It might not yet support the following combinations:
952 * - Read list + Write list
954 * It does not support the following combinations:
956 * - Write list + Reply chunk
957 * - Read list + Write list + Reply chunk
959 * This implementation supports only a single chunk in each
960 * Read or Write list. Thus for example the client cannot
961 * send a Call message with a Position Zero Read chunk and a
962 * regular Read chunk at the same time.
964 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
967 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
970 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
974 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
979 trace_xprtrdma_marshal(req, rtype, wtype);
983 trace_xprtrdma_marshal_failed(rqst, ret);
984 r_xprt->rx_stats.failed_marshal_count++;
989 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
990 struct rpcrdma_buffer *buf,
993 buf->rb_credits = grant;
994 xprt->cwnd = grant << RPC_CWNDSHIFT;
997 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
999 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1001 spin_lock(&xprt->transport_lock);
1002 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1003 spin_unlock(&xprt->transport_lock);
1007 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1008 * @r_xprt: controlling transport instance
1010 * Prepare @r_xprt for the next connection by reinitializing
1011 * its credit grant to one (see RFC 8166, Section 3.3.3).
1013 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1015 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1017 spin_lock(&xprt->transport_lock);
1019 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1020 spin_unlock(&xprt->transport_lock);
1024 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1025 * @rqst: controlling RPC request
1026 * @srcp: points to RPC message payload in receive buffer
1027 * @copy_len: remaining length of receive buffer content
1028 * @pad: Write chunk pad bytes needed (zero for pure inline)
1030 * The upper layer has set the maximum number of bytes it can
1031 * receive in each component of rq_rcv_buf. These values are set in
1032 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1034 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1035 * many cases this function simply updates iov_base pointers in
1036 * rq_rcv_buf to point directly to the received reply data, to
1037 * avoid copying reply data.
1039 * Returns the count of bytes which had to be memcopied.
1041 static unsigned long
1042 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1044 unsigned long fixup_copy_count;
1045 int i, npages, curlen;
1047 struct page **ppages;
1050 /* The head iovec is redirected to the RPC reply message
1051 * in the receive buffer, to avoid a memcopy.
1053 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1054 rqst->rq_private_buf.head[0].iov_base = srcp;
1056 /* The contents of the receive buffer that follow
1057 * head.iov_len bytes are copied into the page list.
1059 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1060 if (curlen > copy_len)
1065 ppages = rqst->rq_rcv_buf.pages +
1066 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1067 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1068 fixup_copy_count = 0;
1069 if (copy_len && rqst->rq_rcv_buf.page_len) {
1072 pagelist_len = rqst->rq_rcv_buf.page_len;
1073 if (pagelist_len > copy_len)
1074 pagelist_len = copy_len;
1075 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1076 for (i = 0; i < npages; i++) {
1077 curlen = PAGE_SIZE - page_base;
1078 if (curlen > pagelist_len)
1079 curlen = pagelist_len;
1081 destp = kmap_atomic(ppages[i]);
1082 memcpy(destp + page_base, srcp, curlen);
1083 flush_dcache_page(ppages[i]);
1084 kunmap_atomic(destp);
1087 fixup_copy_count += curlen;
1088 pagelist_len -= curlen;
1094 /* Implicit padding for the last segment in a Write
1095 * chunk is inserted inline at the front of the tail
1096 * iovec. The upper layer ignores the content of
1097 * the pad. Simply ensure inline content in the tail
1098 * that follows the Write chunk is properly aligned.
1104 /* The tail iovec is redirected to the remaining data
1105 * in the receive buffer, to avoid a memcopy.
1107 if (copy_len || pad) {
1108 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1109 rqst->rq_private_buf.tail[0].iov_base = srcp;
1112 if (fixup_copy_count)
1113 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1114 return fixup_copy_count;
1117 /* By convention, backchannel calls arrive via rdma_msg type
1118 * messages, and never populate the chunk lists. This makes
1119 * the RPC/RDMA header small and fixed in size, so it is
1120 * straightforward to check the RPC header's direction field.
1123 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1124 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1126 struct xdr_stream *xdr = &rep->rr_stream;
1129 if (rep->rr_proc != rdma_msg)
1132 /* Peek at stream contents without advancing. */
1133 p = xdr_inline_decode(xdr, 0);
1136 if (xdr_item_is_present(p++))
1138 if (xdr_item_is_present(p++))
1140 if (xdr_item_is_present(p++))
1144 if (*p++ != rep->rr_xid)
1146 if (*p != cpu_to_be32(RPC_CALL))
1149 /* Now that we are sure this is a backchannel call,
1150 * advance to the RPC header.
1152 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1156 rpcrdma_bc_receive_call(r_xprt, rep);
1160 pr_warn("RPC/RDMA short backward direction call\n");
1163 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1167 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1169 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1175 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1179 xdr_decode_rdma_segment(p, &handle, length, &offset);
1180 trace_xprtrdma_decode_seg(handle, *length, offset);
1184 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1186 u32 segcount, seglength;
1189 p = xdr_inline_decode(xdr, sizeof(*p));
1194 segcount = be32_to_cpup(p);
1195 while (segcount--) {
1196 if (decode_rdma_segment(xdr, &seglength))
1198 *length += seglength;
1204 /* In RPC-over-RDMA Version One replies, a Read list is never
1205 * expected. This decoder is a stub that returns an error if
1206 * a Read list is present.
1208 static int decode_read_list(struct xdr_stream *xdr)
1212 p = xdr_inline_decode(xdr, sizeof(*p));
1215 if (unlikely(xdr_item_is_present(p)))
1220 /* Supports only one Write chunk in the Write list
1222 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1231 p = xdr_inline_decode(xdr, sizeof(*p));
1234 if (xdr_item_is_absent(p))
1239 if (decode_write_chunk(xdr, &chunklen))
1241 *length += chunklen;
1247 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1251 p = xdr_inline_decode(xdr, sizeof(*p));
1256 if (xdr_item_is_present(p))
1257 if (decode_write_chunk(xdr, length))
1263 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1264 struct rpc_rqst *rqst)
1266 struct xdr_stream *xdr = &rep->rr_stream;
1267 u32 writelist, replychunk, rpclen;
1270 /* Decode the chunk lists */
1271 if (decode_read_list(xdr))
1273 if (decode_write_list(xdr, &writelist))
1275 if (decode_reply_chunk(xdr, &replychunk))
1278 /* RDMA_MSG sanity checks */
1279 if (unlikely(replychunk))
1282 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1283 base = (char *)xdr_inline_decode(xdr, 0);
1284 rpclen = xdr_stream_remaining(xdr);
1285 r_xprt->rx_stats.fixup_copy_count +=
1286 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1288 r_xprt->rx_stats.total_rdma_reply += writelist;
1289 return rpclen + xdr_align_size(writelist);
1293 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1295 struct xdr_stream *xdr = &rep->rr_stream;
1296 u32 writelist, replychunk;
1298 /* Decode the chunk lists */
1299 if (decode_read_list(xdr))
1301 if (decode_write_list(xdr, &writelist))
1303 if (decode_reply_chunk(xdr, &replychunk))
1306 /* RDMA_NOMSG sanity checks */
1307 if (unlikely(writelist))
1309 if (unlikely(!replychunk))
1312 /* Reply chunk buffer already is the reply vector */
1313 r_xprt->rx_stats.total_rdma_reply += replychunk;
1318 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1319 struct rpc_rqst *rqst)
1321 struct xdr_stream *xdr = &rep->rr_stream;
1324 p = xdr_inline_decode(xdr, sizeof(*p));
1330 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1333 trace_xprtrdma_err_vers(rqst, p, p + 1);
1336 trace_xprtrdma_err_chunk(rqst);
1339 trace_xprtrdma_err_unrecognized(rqst, p);
1345 /* Perform XID lookup, reconstruction of the RPC reply, and
1346 * RPC completion while holding the transport lock to ensure
1347 * the rep, rqst, and rq_task pointers remain stable.
1349 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1351 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1352 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1353 struct rpc_rqst *rqst = rep->rr_rqst;
1356 switch (rep->rr_proc) {
1358 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1361 status = rpcrdma_decode_nomsg(r_xprt, rep);
1364 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1373 spin_lock(&xprt->queue_lock);
1374 xprt_complete_rqst(rqst->rq_task, status);
1375 xprt_unpin_rqst(rqst);
1376 spin_unlock(&xprt->queue_lock);
1380 trace_xprtrdma_reply_hdr_err(rep);
1381 r_xprt->rx_stats.bad_reply_count++;
1382 rqst->rq_task->tk_status = status;
1387 static void rpcrdma_reply_done(struct kref *kref)
1389 struct rpcrdma_req *req =
1390 container_of(kref, struct rpcrdma_req, rl_kref);
1392 rpcrdma_complete_rqst(req->rl_reply);
1396 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1397 * @rep: Incoming rpcrdma_rep object to process
1399 * Errors must result in the RPC task either being awakened, or
1400 * allowed to timeout, to discover the errors at that time.
1402 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1404 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1405 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1406 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1407 struct rpcrdma_req *req;
1408 struct rpc_rqst *rqst;
1412 /* Any data means we had a useful conversation, so
1413 * then we don't need to delay the next reconnect.
1415 if (xprt->reestablish_timeout)
1416 xprt->reestablish_timeout = 0;
1418 /* Fixed transport header fields */
1419 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1420 rep->rr_hdrbuf.head[0].iov_base, NULL);
1421 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1423 goto out_shortreply;
1425 rep->rr_vers = *p++;
1426 credits = be32_to_cpu(*p++);
1427 rep->rr_proc = *p++;
1429 if (rep->rr_vers != rpcrdma_version)
1430 goto out_badversion;
1432 if (rpcrdma_is_bcall(r_xprt, rep))
1435 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1436 * get context for handling any incoming chunks.
1438 spin_lock(&xprt->queue_lock);
1439 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1442 xprt_pin_rqst(rqst);
1443 spin_unlock(&xprt->queue_lock);
1446 credits = 1; /* don't deadlock */
1447 else if (credits > r_xprt->rx_ep->re_max_requests)
1448 credits = r_xprt->rx_ep->re_max_requests;
1449 if (buf->rb_credits != credits)
1450 rpcrdma_update_cwnd(r_xprt, credits);
1451 rpcrdma_post_recvs(r_xprt, false);
1453 req = rpcr_to_rdmar(rqst);
1454 if (unlikely(req->rl_reply))
1455 rpcrdma_recv_buffer_put(req->rl_reply);
1456 req->rl_reply = rep;
1457 rep->rr_rqst = rqst;
1459 trace_xprtrdma_reply(rqst->rq_task, rep, credits);
1461 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1462 frwr_reminv(rep, &req->rl_registered);
1463 if (!list_empty(&req->rl_registered))
1464 frwr_unmap_async(r_xprt, req);
1465 /* LocalInv completion will complete the RPC */
1467 kref_put(&req->rl_kref, rpcrdma_reply_done);
1471 trace_xprtrdma_reply_vers_err(rep);
1475 spin_unlock(&xprt->queue_lock);
1476 trace_xprtrdma_reply_rqst_err(rep);
1480 trace_xprtrdma_reply_short_err(rep);
1483 rpcrdma_recv_buffer_put(rep);