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 = offset_in_page(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 = page_base;
250 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
258 if (type == rpcrdma_readch)
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_ep->re_implicit_roundup)
269 if (xdrbuf->tail[0].iov_len)
270 rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
273 if (unlikely(n > RPCRDMA_MAX_SEGS))
279 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
283 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
287 xdr_encode_rdma_segment(p, mr->mr_handle, mr->mr_length, mr->mr_offset);
292 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
297 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
301 *p++ = xdr_one; /* Item present */
302 xdr_encode_read_segment(p, position, mr->mr_handle, mr->mr_length,
307 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
308 struct rpcrdma_req *req,
309 struct rpcrdma_mr_seg *seg,
310 int nsegs, bool writing,
311 struct rpcrdma_mr **mr)
313 *mr = rpcrdma_mr_pop(&req->rl_free_mrs);
315 *mr = rpcrdma_mr_get(r_xprt);
321 rpcrdma_mr_push(*mr, &req->rl_registered);
322 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
325 trace_xprtrdma_nomrs_err(r_xprt, req);
326 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
327 rpcrdma_mrs_refresh(r_xprt);
328 return ERR_PTR(-EAGAIN);
331 /* Register and XDR encode the Read list. Supports encoding a list of read
332 * segments that belong to a single read chunk.
334 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
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
340 * Returns zero on success, or a negative errno if a failure occurred.
341 * @xdr is advanced to the next position in the stream.
343 * Only a single @pos value is currently supported.
345 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
346 struct rpcrdma_req *req,
347 struct rpc_rqst *rqst,
348 enum rpcrdma_chunktype rtype)
350 struct xdr_stream *xdr = &req->rl_stream;
351 struct rpcrdma_mr_seg *seg;
352 struct rpcrdma_mr *mr;
356 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
359 pos = rqst->rq_snd_buf.head[0].iov_len;
360 if (rtype == rpcrdma_areadch)
362 seg = req->rl_segments;
363 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
369 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
373 if (encode_read_segment(xdr, mr, pos) < 0)
376 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
377 r_xprt->rx_stats.read_chunk_count++;
378 nsegs -= mr->mr_nents;
382 if (xdr_stream_encode_item_absent(xdr) < 0)
387 /* Register and XDR encode the Write list. Supports encoding a list
388 * containing one array of plain segments that belong to a single
391 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
393 * Write chunklist (a list of (one) counted array):
395 * 1 - N - HLOO - HLOO - ... - HLOO - 0
397 * Returns zero on success, or a negative errno if a failure occurred.
398 * @xdr is advanced to the next position in the stream.
400 * Only a single Write chunk is currently supported.
402 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
403 struct rpcrdma_req *req,
404 struct rpc_rqst *rqst,
405 enum rpcrdma_chunktype wtype)
407 struct xdr_stream *xdr = &req->rl_stream;
408 struct rpcrdma_mr_seg *seg;
409 struct rpcrdma_mr *mr;
413 if (wtype != rpcrdma_writech)
416 seg = req->rl_segments;
417 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
418 rqst->rq_rcv_buf.head[0].iov_len,
423 if (xdr_stream_encode_item_present(xdr) < 0)
425 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
426 if (unlikely(!segcount))
428 /* Actual value encoded below */
432 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
436 if (encode_rdma_segment(xdr, mr) < 0)
439 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
440 r_xprt->rx_stats.write_chunk_count++;
441 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
443 nsegs -= mr->mr_nents;
446 /* Update count of segments in this Write chunk */
447 *segcount = cpu_to_be32(nchunks);
450 if (xdr_stream_encode_item_absent(xdr) < 0)
455 /* Register and XDR encode the Reply chunk. Supports encoding an array
456 * of plain segments that belong to a single write (reply) chunk.
458 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
460 * Reply chunk (a counted array):
462 * 1 - N - HLOO - HLOO - ... - HLOO
464 * Returns zero on success, or a negative errno if a failure occurred.
465 * @xdr is advanced to the next position in the stream.
467 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
468 struct rpcrdma_req *req,
469 struct rpc_rqst *rqst,
470 enum rpcrdma_chunktype wtype)
472 struct xdr_stream *xdr = &req->rl_stream;
473 struct rpcrdma_mr_seg *seg;
474 struct rpcrdma_mr *mr;
478 if (wtype != rpcrdma_replych) {
479 if (xdr_stream_encode_item_absent(xdr) < 0)
484 seg = req->rl_segments;
485 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
489 if (xdr_stream_encode_item_present(xdr) < 0)
491 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
492 if (unlikely(!segcount))
494 /* Actual value encoded below */
498 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
502 if (encode_rdma_segment(xdr, mr) < 0)
505 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
506 r_xprt->rx_stats.reply_chunk_count++;
507 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
509 nsegs -= mr->mr_nents;
512 /* Update count of segments in the Reply chunk */
513 *segcount = cpu_to_be32(nchunks);
518 static void rpcrdma_sendctx_done(struct kref *kref)
520 struct rpcrdma_req *req =
521 container_of(kref, struct rpcrdma_req, rl_kref);
522 struct rpcrdma_rep *rep = req->rl_reply;
524 rpcrdma_complete_rqst(rep);
525 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
529 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
530 * @sc: sendctx containing SGEs to unmap
533 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
535 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
538 if (!sc->sc_unmap_count)
541 /* The first two SGEs contain the transport header and
542 * the inline buffer. These are always left mapped so
543 * they can be cheaply re-used.
545 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
546 ++sge, --sc->sc_unmap_count)
547 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
550 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
553 /* Prepare an SGE for the RPC-over-RDMA transport header.
555 static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
556 struct rpcrdma_req *req, u32 len)
558 struct rpcrdma_sendctx *sc = req->rl_sendctx;
559 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
560 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
562 sge->addr = rdmab_addr(rb);
564 sge->lkey = rdmab_lkey(rb);
566 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
570 /* The head iovec is straightforward, as it is usually already
571 * DMA-mapped. Sync the content that has changed.
573 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
574 struct rpcrdma_req *req, unsigned int len)
576 struct rpcrdma_sendctx *sc = req->rl_sendctx;
577 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
578 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
580 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
583 sge->addr = rdmab_addr(rb);
585 sge->lkey = rdmab_lkey(rb);
587 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
592 /* If there is a page list present, DMA map and prepare an
593 * SGE for each page to be sent.
595 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
598 struct rpcrdma_sendctx *sc = req->rl_sendctx;
599 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
600 unsigned int page_base, len, remaining;
601 struct page **ppages;
604 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
605 page_base = offset_in_page(xdr->page_base);
606 remaining = xdr->page_len;
608 sge = &sc->sc_sges[req->rl_wr.num_sge++];
609 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
610 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
611 page_base, len, DMA_TO_DEVICE);
612 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
613 goto out_mapping_err;
616 sge->lkey = rdmab_lkey(rb);
618 sc->sc_unmap_count++;
627 trace_xprtrdma_dma_maperr(sge->addr);
631 /* The tail iovec might not reside in the same page as the
634 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
636 unsigned int page_base, unsigned int len)
638 struct rpcrdma_sendctx *sc = req->rl_sendctx;
639 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
640 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
641 struct page *page = virt_to_page(xdr->tail[0].iov_base);
643 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
645 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
646 goto out_mapping_err;
649 sge->lkey = rdmab_lkey(rb);
650 ++sc->sc_unmap_count;
654 trace_xprtrdma_dma_maperr(sge->addr);
658 /* Copy the tail to the end of the head buffer.
660 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
661 struct rpcrdma_req *req,
666 dst = (unsigned char *)xdr->head[0].iov_base;
667 dst += xdr->head[0].iov_len + xdr->page_len;
668 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
669 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
672 /* Copy pagelist content into the head buffer.
674 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
675 struct rpcrdma_req *req,
678 unsigned int len, page_base, remaining;
679 struct page **ppages;
680 unsigned char *src, *dst;
682 dst = (unsigned char *)xdr->head[0].iov_base;
683 dst += xdr->head[0].iov_len;
684 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
685 page_base = offset_in_page(xdr->page_base);
686 remaining = xdr->page_len;
688 src = page_address(*ppages);
690 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
691 memcpy(dst, src, len);
692 r_xprt->rx_stats.pullup_copy_count += len;
701 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
702 * When the head, pagelist, and tail are small, a pull-up copy
703 * is considerably less costly than DMA mapping the components
707 * - the caller has already verified that the total length
708 * of the RPC Call body will fit into @rl_sendbuf.
710 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
711 struct rpcrdma_req *req,
714 if (unlikely(xdr->tail[0].iov_len))
715 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
717 if (unlikely(xdr->page_len))
718 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
720 /* The whole RPC message resides in the head iovec now */
721 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
724 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
725 struct rpcrdma_req *req,
728 struct kvec *tail = &xdr->tail[0];
730 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
733 if (!rpcrdma_prepare_pagelist(req, xdr))
736 if (!rpcrdma_prepare_tail_iov(req, xdr,
737 offset_in_page(tail->iov_base),
741 if (req->rl_sendctx->sc_unmap_count)
742 kref_get(&req->rl_kref);
746 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
747 struct rpcrdma_req *req,
750 struct kvec *tail = &xdr->tail[0];
752 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
755 /* If there is a Read chunk, the page list is handled
756 * via explicit RDMA, and thus is skipped here.
760 if (!rpcrdma_prepare_tail_iov(req, xdr,
761 offset_in_page(tail->iov_base),
764 kref_get(&req->rl_kref);
771 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
772 * @r_xprt: controlling transport
773 * @req: context of RPC Call being marshalled
774 * @hdrlen: size of transport header, in bytes
775 * @xdr: xdr_buf containing RPC Call
776 * @rtype: chunk type being encoded
778 * Returns 0 on success; otherwise a negative errno is returned.
780 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
781 struct rpcrdma_req *req, u32 hdrlen,
783 enum rpcrdma_chunktype rtype)
788 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
789 if (!req->rl_sendctx)
791 req->rl_sendctx->sc_unmap_count = 0;
792 req->rl_sendctx->sc_req = req;
793 kref_init(&req->rl_kref);
794 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
795 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
796 req->rl_wr.num_sge = 0;
797 req->rl_wr.opcode = IB_WR_SEND;
799 rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
803 case rpcrdma_noch_pullup:
804 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
807 case rpcrdma_noch_mapped:
808 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
812 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
815 case rpcrdma_areadch:
824 rpcrdma_sendctx_unmap(req->rl_sendctx);
826 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
831 * rpcrdma_marshal_req - Marshal and send one RPC request
832 * @r_xprt: controlling transport
833 * @rqst: RPC request to be marshaled
835 * For the RPC in "rqst", this function:
836 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
837 * - Registers Read, Write, and Reply chunks
838 * - Constructs the transport header
839 * - Posts a Send WR to send the transport header and request
842 * %0 if the RPC was sent successfully,
843 * %-ENOTCONN if the connection was lost,
844 * %-EAGAIN if the caller should call again with the same arguments,
845 * %-ENOBUFS if the caller should call again after a delay,
846 * %-EMSGSIZE if the transport header is too small,
847 * %-EIO if a permanent problem occurred while marshaling.
850 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
852 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
853 struct xdr_stream *xdr = &req->rl_stream;
854 enum rpcrdma_chunktype rtype, wtype;
855 struct xdr_buf *buf = &rqst->rq_snd_buf;
860 if (unlikely(rqst->rq_rcv_buf.flags & XDRBUF_SPARSE_PAGES)) {
861 ret = rpcrdma_alloc_sparse_pages(&rqst->rq_rcv_buf);
866 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
867 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
870 /* Fixed header fields */
872 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
876 *p++ = rpcrdma_version;
877 *p++ = r_xprt->rx_buf.rb_max_requests;
879 /* When the ULP employs a GSS flavor that guarantees integrity
880 * or privacy, direct data placement of individual data items
883 ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
884 &rqst->rq_cred->cr_auth->au_flags);
887 * Chunks needed for results?
889 * o If the expected result is under the inline threshold, all ops
891 * o Large read ops return data as write chunk(s), header as
893 * o Large non-read ops return as a single reply chunk.
895 if (rpcrdma_results_inline(r_xprt, rqst))
896 wtype = rpcrdma_noch;
897 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
898 rpcrdma_nonpayload_inline(r_xprt, rqst))
899 wtype = rpcrdma_writech;
901 wtype = rpcrdma_replych;
904 * Chunks needed for arguments?
906 * o If the total request is under the inline threshold, all ops
907 * are sent as inline.
908 * o Large write ops transmit data as read chunk(s), header as
910 * o Large non-write ops are sent with the entire message as a
911 * single read chunk (protocol 0-position special case).
913 * This assumes that the upper layer does not present a request
914 * that both has a data payload, and whose non-data arguments
915 * by themselves are larger than the inline threshold.
917 if (rpcrdma_args_inline(r_xprt, rqst)) {
919 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
920 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
921 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
923 rtype = rpcrdma_readch;
925 r_xprt->rx_stats.nomsg_call_count++;
927 rtype = rpcrdma_areadch;
930 /* This implementation supports the following combinations
931 * of chunk lists in one RPC-over-RDMA Call message:
936 * - Read list + Reply chunk
938 * It might not yet support the following combinations:
940 * - Read list + Write list
942 * It does not support the following combinations:
944 * - Write list + Reply chunk
945 * - Read list + Write list + Reply chunk
947 * This implementation supports only a single chunk in each
948 * Read or Write list. Thus for example the client cannot
949 * send a Call message with a Position Zero Read chunk and a
950 * regular Read chunk at the same time.
952 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
955 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
958 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
962 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
967 trace_xprtrdma_marshal(req, rtype, wtype);
971 trace_xprtrdma_marshal_failed(rqst, ret);
972 r_xprt->rx_stats.failed_marshal_count++;
977 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
978 struct rpcrdma_buffer *buf,
981 buf->rb_credits = grant;
982 xprt->cwnd = grant << RPC_CWNDSHIFT;
985 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
987 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
989 spin_lock(&xprt->transport_lock);
990 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
991 spin_unlock(&xprt->transport_lock);
995 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
996 * @r_xprt: controlling transport instance
998 * Prepare @r_xprt for the next connection by reinitializing
999 * its credit grant to one (see RFC 8166, Section 3.3.3).
1001 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1003 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1005 spin_lock(&xprt->transport_lock);
1007 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1008 spin_unlock(&xprt->transport_lock);
1012 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1013 * @rqst: controlling RPC request
1014 * @srcp: points to RPC message payload in receive buffer
1015 * @copy_len: remaining length of receive buffer content
1016 * @pad: Write chunk pad bytes needed (zero for pure inline)
1018 * The upper layer has set the maximum number of bytes it can
1019 * receive in each component of rq_rcv_buf. These values are set in
1020 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1022 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1023 * many cases this function simply updates iov_base pointers in
1024 * rq_rcv_buf to point directly to the received reply data, to
1025 * avoid copying reply data.
1027 * Returns the count of bytes which had to be memcopied.
1029 static unsigned long
1030 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1032 unsigned long fixup_copy_count;
1033 int i, npages, curlen;
1035 struct page **ppages;
1038 /* The head iovec is redirected to the RPC reply message
1039 * in the receive buffer, to avoid a memcopy.
1041 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1042 rqst->rq_private_buf.head[0].iov_base = srcp;
1044 /* The contents of the receive buffer that follow
1045 * head.iov_len bytes are copied into the page list.
1047 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1048 if (curlen > copy_len)
1053 ppages = rqst->rq_rcv_buf.pages +
1054 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1055 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1056 fixup_copy_count = 0;
1057 if (copy_len && rqst->rq_rcv_buf.page_len) {
1060 pagelist_len = rqst->rq_rcv_buf.page_len;
1061 if (pagelist_len > copy_len)
1062 pagelist_len = copy_len;
1063 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1064 for (i = 0; i < npages; i++) {
1065 curlen = PAGE_SIZE - page_base;
1066 if (curlen > pagelist_len)
1067 curlen = pagelist_len;
1069 destp = kmap_atomic(ppages[i]);
1070 memcpy(destp + page_base, srcp, curlen);
1071 flush_dcache_page(ppages[i]);
1072 kunmap_atomic(destp);
1075 fixup_copy_count += curlen;
1076 pagelist_len -= curlen;
1082 /* Implicit padding for the last segment in a Write
1083 * chunk is inserted inline at the front of the tail
1084 * iovec. The upper layer ignores the content of
1085 * the pad. Simply ensure inline content in the tail
1086 * that follows the Write chunk is properly aligned.
1092 /* The tail iovec is redirected to the remaining data
1093 * in the receive buffer, to avoid a memcopy.
1095 if (copy_len || pad) {
1096 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1097 rqst->rq_private_buf.tail[0].iov_base = srcp;
1100 if (fixup_copy_count)
1101 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1102 return fixup_copy_count;
1105 /* By convention, backchannel calls arrive via rdma_msg type
1106 * messages, and never populate the chunk lists. This makes
1107 * the RPC/RDMA header small and fixed in size, so it is
1108 * straightforward to check the RPC header's direction field.
1111 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1112 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1114 struct xdr_stream *xdr = &rep->rr_stream;
1117 if (rep->rr_proc != rdma_msg)
1120 /* Peek at stream contents without advancing. */
1121 p = xdr_inline_decode(xdr, 0);
1124 if (xdr_item_is_present(p++))
1126 if (xdr_item_is_present(p++))
1128 if (xdr_item_is_present(p++))
1132 if (*p++ != rep->rr_xid)
1134 if (*p != cpu_to_be32(RPC_CALL))
1137 /* Now that we are sure this is a backchannel call,
1138 * advance to the RPC header.
1140 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1144 rpcrdma_bc_receive_call(r_xprt, rep);
1147 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1151 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1153 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1159 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1163 xdr_decode_rdma_segment(p, &handle, length, &offset);
1164 trace_xprtrdma_decode_seg(handle, *length, offset);
1168 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1170 u32 segcount, seglength;
1173 p = xdr_inline_decode(xdr, sizeof(*p));
1178 segcount = be32_to_cpup(p);
1179 while (segcount--) {
1180 if (decode_rdma_segment(xdr, &seglength))
1182 *length += seglength;
1188 /* In RPC-over-RDMA Version One replies, a Read list is never
1189 * expected. This decoder is a stub that returns an error if
1190 * a Read list is present.
1192 static int decode_read_list(struct xdr_stream *xdr)
1196 p = xdr_inline_decode(xdr, sizeof(*p));
1199 if (unlikely(xdr_item_is_present(p)))
1204 /* Supports only one Write chunk in the Write list
1206 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1215 p = xdr_inline_decode(xdr, sizeof(*p));
1218 if (xdr_item_is_absent(p))
1223 if (decode_write_chunk(xdr, &chunklen))
1225 *length += chunklen;
1231 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1235 p = xdr_inline_decode(xdr, sizeof(*p));
1240 if (xdr_item_is_present(p))
1241 if (decode_write_chunk(xdr, length))
1247 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1248 struct rpc_rqst *rqst)
1250 struct xdr_stream *xdr = &rep->rr_stream;
1251 u32 writelist, replychunk, rpclen;
1254 /* Decode the chunk lists */
1255 if (decode_read_list(xdr))
1257 if (decode_write_list(xdr, &writelist))
1259 if (decode_reply_chunk(xdr, &replychunk))
1262 /* RDMA_MSG sanity checks */
1263 if (unlikely(replychunk))
1266 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1267 base = (char *)xdr_inline_decode(xdr, 0);
1268 rpclen = xdr_stream_remaining(xdr);
1269 r_xprt->rx_stats.fixup_copy_count +=
1270 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1272 r_xprt->rx_stats.total_rdma_reply += writelist;
1273 return rpclen + xdr_align_size(writelist);
1277 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1279 struct xdr_stream *xdr = &rep->rr_stream;
1280 u32 writelist, replychunk;
1282 /* Decode the chunk lists */
1283 if (decode_read_list(xdr))
1285 if (decode_write_list(xdr, &writelist))
1287 if (decode_reply_chunk(xdr, &replychunk))
1290 /* RDMA_NOMSG sanity checks */
1291 if (unlikely(writelist))
1293 if (unlikely(!replychunk))
1296 /* Reply chunk buffer already is the reply vector */
1297 r_xprt->rx_stats.total_rdma_reply += replychunk;
1302 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1303 struct rpc_rqst *rqst)
1305 struct xdr_stream *xdr = &rep->rr_stream;
1308 p = xdr_inline_decode(xdr, sizeof(*p));
1314 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1317 trace_xprtrdma_err_vers(rqst, p, p + 1);
1320 trace_xprtrdma_err_chunk(rqst);
1323 trace_xprtrdma_err_unrecognized(rqst, p);
1330 * rpcrdma_unpin_rqst - Release rqst without completing it
1331 * @rep: RPC/RDMA Receive context
1333 * This is done when a connection is lost so that a Reply
1334 * can be dropped and its matching Call can be subsequently
1335 * retransmitted on a new connection.
1337 void rpcrdma_unpin_rqst(struct rpcrdma_rep *rep)
1339 struct rpc_xprt *xprt = &rep->rr_rxprt->rx_xprt;
1340 struct rpc_rqst *rqst = rep->rr_rqst;
1341 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
1343 req->rl_reply = NULL;
1344 rep->rr_rqst = NULL;
1346 spin_lock(&xprt->queue_lock);
1347 xprt_unpin_rqst(rqst);
1348 spin_unlock(&xprt->queue_lock);
1352 * rpcrdma_complete_rqst - Pass completed rqst back to RPC
1353 * @rep: RPC/RDMA Receive context
1355 * Reconstruct the RPC reply and complete the transaction
1356 * while @rqst is still pinned to ensure the rep, rqst, and
1357 * rq_task pointers remain stable.
1359 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1361 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1362 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1363 struct rpc_rqst *rqst = rep->rr_rqst;
1366 switch (rep->rr_proc) {
1368 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1371 status = rpcrdma_decode_nomsg(r_xprt, rep);
1374 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1383 spin_lock(&xprt->queue_lock);
1384 xprt_complete_rqst(rqst->rq_task, status);
1385 xprt_unpin_rqst(rqst);
1386 spin_unlock(&xprt->queue_lock);
1390 trace_xprtrdma_reply_hdr_err(rep);
1391 r_xprt->rx_stats.bad_reply_count++;
1392 rqst->rq_task->tk_status = status;
1397 static void rpcrdma_reply_done(struct kref *kref)
1399 struct rpcrdma_req *req =
1400 container_of(kref, struct rpcrdma_req, rl_kref);
1402 rpcrdma_complete_rqst(req->rl_reply);
1406 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1407 * @rep: Incoming rpcrdma_rep object to process
1409 * Errors must result in the RPC task either being awakened, or
1410 * allowed to timeout, to discover the errors at that time.
1412 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1414 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1415 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1416 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1417 struct rpcrdma_req *req;
1418 struct rpc_rqst *rqst;
1422 /* Any data means we had a useful conversation, so
1423 * then we don't need to delay the next reconnect.
1425 if (xprt->reestablish_timeout)
1426 xprt->reestablish_timeout = 0;
1428 /* Fixed transport header fields */
1429 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1430 rep->rr_hdrbuf.head[0].iov_base, NULL);
1431 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1433 goto out_shortreply;
1435 rep->rr_vers = *p++;
1436 credits = be32_to_cpu(*p++);
1437 rep->rr_proc = *p++;
1439 if (rep->rr_vers != rpcrdma_version)
1440 goto out_badversion;
1442 if (rpcrdma_is_bcall(r_xprt, rep))
1445 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1446 * get context for handling any incoming chunks.
1448 spin_lock(&xprt->queue_lock);
1449 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1452 xprt_pin_rqst(rqst);
1453 spin_unlock(&xprt->queue_lock);
1456 credits = 1; /* don't deadlock */
1457 else if (credits > r_xprt->rx_ep->re_max_requests)
1458 credits = r_xprt->rx_ep->re_max_requests;
1459 rpcrdma_post_recvs(r_xprt, credits + (buf->rb_bc_srv_max_requests << 1),
1461 if (buf->rb_credits != credits)
1462 rpcrdma_update_cwnd(r_xprt, credits);
1464 req = rpcr_to_rdmar(rqst);
1465 if (unlikely(req->rl_reply))
1466 rpcrdma_rep_put(buf, req->rl_reply);
1467 req->rl_reply = rep;
1468 rep->rr_rqst = rqst;
1470 trace_xprtrdma_reply(rqst->rq_task, rep, credits);
1472 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1473 frwr_reminv(rep, &req->rl_registered);
1474 if (!list_empty(&req->rl_registered))
1475 frwr_unmap_async(r_xprt, req);
1476 /* LocalInv completion will complete the RPC */
1478 kref_put(&req->rl_kref, rpcrdma_reply_done);
1482 trace_xprtrdma_reply_vers_err(rep);
1486 spin_unlock(&xprt->queue_lock);
1487 trace_xprtrdma_reply_rqst_err(rep);
1491 trace_xprtrdma_reply_short_err(rep);
1494 rpcrdma_rep_put(buf, rep);