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 /* Split @vec on page boundaries into SGEs. FMR registers pages, not
208 * a byte range. Other modes coalesce these SGEs into a single MR
211 * Returns pointer to next available SGE, and bumps the total number
214 static struct rpcrdma_mr_seg *
215 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
218 u32 remaining, page_offset;
221 base = vec->iov_base;
222 page_offset = offset_in_page(base);
223 remaining = vec->iov_len;
226 seg->mr_offset = base;
227 seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
228 remaining -= seg->mr_len;
237 /* Convert @xdrbuf into SGEs no larger than a page each. As they
238 * are registered, these SGEs are then coalesced into RDMA segments
239 * when the selected memreg mode supports it.
241 * Returns positive number of SGEs consumed, or a negative errno.
245 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
246 unsigned int pos, enum rpcrdma_chunktype type,
247 struct rpcrdma_mr_seg *seg)
249 unsigned long page_base;
251 struct page **ppages;
255 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
257 len = xdrbuf->page_len;
258 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
259 page_base = offset_in_page(xdrbuf->page_base);
261 seg->mr_page = *ppages;
262 seg->mr_offset = (char *)page_base;
263 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
271 /* When encoding a Read chunk, the tail iovec contains an
272 * XDR pad and may be omitted.
274 if (type == rpcrdma_readch && r_xprt->rx_ep->re_implicit_roundup)
277 /* When encoding a Write chunk, some servers need to see an
278 * extra segment for non-XDR-aligned Write chunks. The upper
279 * layer provides space in the tail iovec that may be used
282 if (type == rpcrdma_writech && r_xprt->rx_ep->re_implicit_roundup)
285 if (xdrbuf->tail[0].iov_len)
286 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
289 if (unlikely(n > RPCRDMA_MAX_SEGS))
295 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
299 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
303 xdr_encode_rdma_segment(p, mr->mr_handle, mr->mr_length, mr->mr_offset);
308 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
313 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
317 *p++ = xdr_one; /* Item present */
318 xdr_encode_read_segment(p, position, mr->mr_handle, mr->mr_length,
323 static struct rpcrdma_mr_seg *rpcrdma_mr_prepare(struct rpcrdma_xprt *r_xprt,
324 struct rpcrdma_req *req,
325 struct rpcrdma_mr_seg *seg,
326 int nsegs, bool writing,
327 struct rpcrdma_mr **mr)
329 *mr = rpcrdma_mr_pop(&req->rl_free_mrs);
331 *mr = rpcrdma_mr_get(r_xprt);
337 rpcrdma_mr_push(*mr, &req->rl_registered);
338 return frwr_map(r_xprt, seg, nsegs, writing, req->rl_slot.rq_xid, *mr);
341 trace_xprtrdma_nomrs_err(r_xprt, req);
342 xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
343 rpcrdma_mrs_refresh(r_xprt);
344 return ERR_PTR(-EAGAIN);
347 /* Register and XDR encode the Read list. Supports encoding a list of read
348 * segments that belong to a single read chunk.
350 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
352 * Read chunklist (a linked list):
353 * N elements, position P (same P for all chunks of same arg!):
354 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
356 * Returns zero on success, or a negative errno if a failure occurred.
357 * @xdr is advanced to the next position in the stream.
359 * Only a single @pos value is currently supported.
361 static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
362 struct rpcrdma_req *req,
363 struct rpc_rqst *rqst,
364 enum rpcrdma_chunktype rtype)
366 struct xdr_stream *xdr = &req->rl_stream;
367 struct rpcrdma_mr_seg *seg;
368 struct rpcrdma_mr *mr;
372 if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
375 pos = rqst->rq_snd_buf.head[0].iov_len;
376 if (rtype == rpcrdma_areadch)
378 seg = req->rl_segments;
379 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
385 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, false, &mr);
389 if (encode_read_segment(xdr, mr, pos) < 0)
392 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
393 r_xprt->rx_stats.read_chunk_count++;
394 nsegs -= mr->mr_nents;
398 if (xdr_stream_encode_item_absent(xdr) < 0)
403 /* Register and XDR encode the Write list. Supports encoding a list
404 * containing one array of plain segments that belong to a single
407 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
409 * Write chunklist (a list of (one) counted array):
411 * 1 - N - HLOO - HLOO - ... - HLOO - 0
413 * Returns zero on success, or a negative errno if a failure occurred.
414 * @xdr is advanced to the next position in the stream.
416 * Only a single Write chunk is currently supported.
418 static int rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt,
419 struct rpcrdma_req *req,
420 struct rpc_rqst *rqst,
421 enum rpcrdma_chunktype wtype)
423 struct xdr_stream *xdr = &req->rl_stream;
424 struct rpcrdma_mr_seg *seg;
425 struct rpcrdma_mr *mr;
429 if (wtype != rpcrdma_writech)
432 seg = req->rl_segments;
433 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
434 rqst->rq_rcv_buf.head[0].iov_len,
439 if (xdr_stream_encode_item_present(xdr) < 0)
441 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
442 if (unlikely(!segcount))
444 /* Actual value encoded below */
448 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
452 if (encode_rdma_segment(xdr, mr) < 0)
455 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
456 r_xprt->rx_stats.write_chunk_count++;
457 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
459 nsegs -= mr->mr_nents;
462 /* Update count of segments in this Write chunk */
463 *segcount = cpu_to_be32(nchunks);
466 if (xdr_stream_encode_item_absent(xdr) < 0)
471 /* Register and XDR encode the Reply chunk. Supports encoding an array
472 * of plain segments that belong to a single write (reply) chunk.
474 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
476 * Reply chunk (a counted array):
478 * 1 - N - HLOO - HLOO - ... - HLOO
480 * Returns zero on success, or a negative errno if a failure occurred.
481 * @xdr is advanced to the next position in the stream.
483 static int rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
484 struct rpcrdma_req *req,
485 struct rpc_rqst *rqst,
486 enum rpcrdma_chunktype wtype)
488 struct xdr_stream *xdr = &req->rl_stream;
489 struct rpcrdma_mr_seg *seg;
490 struct rpcrdma_mr *mr;
494 if (wtype != rpcrdma_replych) {
495 if (xdr_stream_encode_item_absent(xdr) < 0)
500 seg = req->rl_segments;
501 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
505 if (xdr_stream_encode_item_present(xdr) < 0)
507 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
508 if (unlikely(!segcount))
510 /* Actual value encoded below */
514 seg = rpcrdma_mr_prepare(r_xprt, req, seg, nsegs, true, &mr);
518 if (encode_rdma_segment(xdr, mr) < 0)
521 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
522 r_xprt->rx_stats.reply_chunk_count++;
523 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
525 nsegs -= mr->mr_nents;
528 /* Update count of segments in the Reply chunk */
529 *segcount = cpu_to_be32(nchunks);
534 static void rpcrdma_sendctx_done(struct kref *kref)
536 struct rpcrdma_req *req =
537 container_of(kref, struct rpcrdma_req, rl_kref);
538 struct rpcrdma_rep *rep = req->rl_reply;
540 rpcrdma_complete_rqst(rep);
541 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
545 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
546 * @sc: sendctx containing SGEs to unmap
549 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
551 struct rpcrdma_regbuf *rb = sc->sc_req->rl_sendbuf;
554 if (!sc->sc_unmap_count)
557 /* The first two SGEs contain the transport header and
558 * the inline buffer. These are always left mapped so
559 * they can be cheaply re-used.
561 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
562 ++sge, --sc->sc_unmap_count)
563 ib_dma_unmap_page(rdmab_device(rb), sge->addr, sge->length,
566 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
569 /* Prepare an SGE for the RPC-over-RDMA transport header.
571 static void rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
572 struct rpcrdma_req *req, u32 len)
574 struct rpcrdma_sendctx *sc = req->rl_sendctx;
575 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
576 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
578 sge->addr = rdmab_addr(rb);
580 sge->lkey = rdmab_lkey(rb);
582 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
586 /* The head iovec is straightforward, as it is usually already
587 * DMA-mapped. Sync the content that has changed.
589 static bool rpcrdma_prepare_head_iov(struct rpcrdma_xprt *r_xprt,
590 struct rpcrdma_req *req, unsigned int len)
592 struct rpcrdma_sendctx *sc = req->rl_sendctx;
593 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
594 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
596 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
599 sge->addr = rdmab_addr(rb);
601 sge->lkey = rdmab_lkey(rb);
603 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
608 /* If there is a page list present, DMA map and prepare an
609 * SGE for each page to be sent.
611 static bool rpcrdma_prepare_pagelist(struct rpcrdma_req *req,
614 struct rpcrdma_sendctx *sc = req->rl_sendctx;
615 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
616 unsigned int page_base, len, remaining;
617 struct page **ppages;
620 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
621 page_base = offset_in_page(xdr->page_base);
622 remaining = xdr->page_len;
624 sge = &sc->sc_sges[req->rl_wr.num_sge++];
625 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
626 sge->addr = ib_dma_map_page(rdmab_device(rb), *ppages,
627 page_base, len, DMA_TO_DEVICE);
628 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
629 goto out_mapping_err;
632 sge->lkey = rdmab_lkey(rb);
634 sc->sc_unmap_count++;
643 trace_xprtrdma_dma_maperr(sge->addr);
647 /* The tail iovec may include an XDR pad for the page list,
648 * as well as additional content, and may not reside in the
649 * same page as the head iovec.
651 static bool rpcrdma_prepare_tail_iov(struct rpcrdma_req *req,
653 unsigned int page_base, unsigned int len)
655 struct rpcrdma_sendctx *sc = req->rl_sendctx;
656 struct ib_sge *sge = &sc->sc_sges[req->rl_wr.num_sge++];
657 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
658 struct page *page = virt_to_page(xdr->tail[0].iov_base);
660 sge->addr = ib_dma_map_page(rdmab_device(rb), page, page_base, len,
662 if (ib_dma_mapping_error(rdmab_device(rb), sge->addr))
663 goto out_mapping_err;
666 sge->lkey = rdmab_lkey(rb);
667 ++sc->sc_unmap_count;
671 trace_xprtrdma_dma_maperr(sge->addr);
675 /* Copy the tail to the end of the head buffer.
677 static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
678 struct rpcrdma_req *req,
683 dst = (unsigned char *)xdr->head[0].iov_base;
684 dst += xdr->head[0].iov_len + xdr->page_len;
685 memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
686 r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
689 /* Copy pagelist content into the head buffer.
691 static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
692 struct rpcrdma_req *req,
695 unsigned int len, page_base, remaining;
696 struct page **ppages;
697 unsigned char *src, *dst;
699 dst = (unsigned char *)xdr->head[0].iov_base;
700 dst += xdr->head[0].iov_len;
701 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
702 page_base = offset_in_page(xdr->page_base);
703 remaining = xdr->page_len;
705 src = page_address(*ppages);
707 len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
708 memcpy(dst, src, len);
709 r_xprt->rx_stats.pullup_copy_count += len;
718 /* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
719 * When the head, pagelist, and tail are small, a pull-up copy
720 * is considerably less costly than DMA mapping the components
724 * - the caller has already verified that the total length
725 * of the RPC Call body will fit into @rl_sendbuf.
727 static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
728 struct rpcrdma_req *req,
731 if (unlikely(xdr->tail[0].iov_len))
732 rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
734 if (unlikely(xdr->page_len))
735 rpcrdma_pullup_pagelist(r_xprt, req, xdr);
737 /* The whole RPC message resides in the head iovec now */
738 return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
741 static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
742 struct rpcrdma_req *req,
745 struct kvec *tail = &xdr->tail[0];
747 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
750 if (!rpcrdma_prepare_pagelist(req, xdr))
753 if (!rpcrdma_prepare_tail_iov(req, xdr,
754 offset_in_page(tail->iov_base),
758 if (req->rl_sendctx->sc_unmap_count)
759 kref_get(&req->rl_kref);
763 static bool rpcrdma_prepare_readch(struct rpcrdma_xprt *r_xprt,
764 struct rpcrdma_req *req,
767 if (!rpcrdma_prepare_head_iov(r_xprt, req, xdr->head[0].iov_len))
770 /* If there is a Read chunk, the page list is being handled
771 * via explicit RDMA, and thus is skipped here.
774 /* Do not include the tail if it is only an XDR pad */
775 if (xdr->tail[0].iov_len > 3) {
776 unsigned int page_base, len;
778 /* If the content in the page list is an odd length,
779 * xdr_write_pages() adds a pad at the beginning of
780 * the tail iovec. Force the tail's non-pad content to
781 * land at the next XDR position in the Send message.
783 page_base = offset_in_page(xdr->tail[0].iov_base);
784 len = xdr->tail[0].iov_len;
785 page_base += len & 3;
787 if (!rpcrdma_prepare_tail_iov(req, xdr, page_base, len))
789 kref_get(&req->rl_kref);
796 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
797 * @r_xprt: controlling transport
798 * @req: context of RPC Call being marshalled
799 * @hdrlen: size of transport header, in bytes
800 * @xdr: xdr_buf containing RPC Call
801 * @rtype: chunk type being encoded
803 * Returns 0 on success; otherwise a negative errno is returned.
805 inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
806 struct rpcrdma_req *req, u32 hdrlen,
808 enum rpcrdma_chunktype rtype)
813 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
814 if (!req->rl_sendctx)
816 req->rl_sendctx->sc_unmap_count = 0;
817 req->rl_sendctx->sc_req = req;
818 kref_init(&req->rl_kref);
819 req->rl_wr.wr_cqe = &req->rl_sendctx->sc_cqe;
820 req->rl_wr.sg_list = req->rl_sendctx->sc_sges;
821 req->rl_wr.num_sge = 0;
822 req->rl_wr.opcode = IB_WR_SEND;
824 rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen);
828 case rpcrdma_noch_pullup:
829 if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
832 case rpcrdma_noch_mapped:
833 if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
837 if (!rpcrdma_prepare_readch(r_xprt, req, xdr))
840 case rpcrdma_areadch:
849 rpcrdma_sendctx_unmap(req->rl_sendctx);
851 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
856 * rpcrdma_marshal_req - Marshal and send one RPC request
857 * @r_xprt: controlling transport
858 * @rqst: RPC request to be marshaled
860 * For the RPC in "rqst", this function:
861 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
862 * - Registers Read, Write, and Reply chunks
863 * - Constructs the transport header
864 * - Posts a Send WR to send the transport header and request
867 * %0 if the RPC was sent successfully,
868 * %-ENOTCONN if the connection was lost,
869 * %-EAGAIN if the caller should call again with the same arguments,
870 * %-ENOBUFS if the caller should call again after a delay,
871 * %-EMSGSIZE if the transport header is too small,
872 * %-EIO if a permanent problem occurred while marshaling.
875 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
877 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
878 struct xdr_stream *xdr = &req->rl_stream;
879 enum rpcrdma_chunktype rtype, wtype;
880 struct xdr_buf *buf = &rqst->rq_snd_buf;
885 if (unlikely(rqst->rq_rcv_buf.flags & XDRBUF_SPARSE_PAGES)) {
886 ret = rpcrdma_alloc_sparse_pages(&rqst->rq_rcv_buf);
891 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
892 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
895 /* Fixed header fields */
897 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
901 *p++ = rpcrdma_version;
902 *p++ = r_xprt->rx_buf.rb_max_requests;
904 /* When the ULP employs a GSS flavor that guarantees integrity
905 * or privacy, direct data placement of individual data items
908 ddp_allowed = !test_bit(RPCAUTH_AUTH_DATATOUCH,
909 &rqst->rq_cred->cr_auth->au_flags);
912 * Chunks needed for results?
914 * o If the expected result is under the inline threshold, all ops
916 * o Large read ops return data as write chunk(s), header as
918 * o Large non-read ops return as a single reply chunk.
920 if (rpcrdma_results_inline(r_xprt, rqst))
921 wtype = rpcrdma_noch;
922 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
923 rpcrdma_nonpayload_inline(r_xprt, rqst))
924 wtype = rpcrdma_writech;
926 wtype = rpcrdma_replych;
929 * Chunks needed for arguments?
931 * o If the total request is under the inline threshold, all ops
932 * are sent as inline.
933 * o Large write ops transmit data as read chunk(s), header as
935 * o Large non-write ops are sent with the entire message as a
936 * single read chunk (protocol 0-position special case).
938 * This assumes that the upper layer does not present a request
939 * that both has a data payload, and whose non-data arguments
940 * by themselves are larger than the inline threshold.
942 if (rpcrdma_args_inline(r_xprt, rqst)) {
944 rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
945 rpcrdma_noch_pullup : rpcrdma_noch_mapped;
946 } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
948 rtype = rpcrdma_readch;
950 r_xprt->rx_stats.nomsg_call_count++;
952 rtype = rpcrdma_areadch;
955 /* This implementation supports the following combinations
956 * of chunk lists in one RPC-over-RDMA Call message:
961 * - Read list + Reply chunk
963 * It might not yet support the following combinations:
965 * - Read list + Write list
967 * It does not support the following combinations:
969 * - Write list + Reply chunk
970 * - Read list + Write list + Reply chunk
972 * This implementation supports only a single chunk in each
973 * Read or Write list. Thus for example the client cannot
974 * send a Call message with a Position Zero Read chunk and a
975 * regular Read chunk at the same time.
977 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
980 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
983 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
987 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
992 trace_xprtrdma_marshal(req, rtype, wtype);
996 trace_xprtrdma_marshal_failed(rqst, ret);
997 r_xprt->rx_stats.failed_marshal_count++;
1002 static void __rpcrdma_update_cwnd_locked(struct rpc_xprt *xprt,
1003 struct rpcrdma_buffer *buf,
1006 buf->rb_credits = grant;
1007 xprt->cwnd = grant << RPC_CWNDSHIFT;
1010 static void rpcrdma_update_cwnd(struct rpcrdma_xprt *r_xprt, u32 grant)
1012 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1014 spin_lock(&xprt->transport_lock);
1015 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, grant);
1016 spin_unlock(&xprt->transport_lock);
1020 * rpcrdma_reset_cwnd - Reset the xprt's congestion window
1021 * @r_xprt: controlling transport instance
1023 * Prepare @r_xprt for the next connection by reinitializing
1024 * its credit grant to one (see RFC 8166, Section 3.3.3).
1026 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt)
1028 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1030 spin_lock(&xprt->transport_lock);
1032 __rpcrdma_update_cwnd_locked(xprt, &r_xprt->rx_buf, 1);
1033 spin_unlock(&xprt->transport_lock);
1037 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
1038 * @rqst: controlling RPC request
1039 * @srcp: points to RPC message payload in receive buffer
1040 * @copy_len: remaining length of receive buffer content
1041 * @pad: Write chunk pad bytes needed (zero for pure inline)
1043 * The upper layer has set the maximum number of bytes it can
1044 * receive in each component of rq_rcv_buf. These values are set in
1045 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
1047 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
1048 * many cases this function simply updates iov_base pointers in
1049 * rq_rcv_buf to point directly to the received reply data, to
1050 * avoid copying reply data.
1052 * Returns the count of bytes which had to be memcopied.
1054 static unsigned long
1055 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
1057 unsigned long fixup_copy_count;
1058 int i, npages, curlen;
1060 struct page **ppages;
1063 /* The head iovec is redirected to the RPC reply message
1064 * in the receive buffer, to avoid a memcopy.
1066 rqst->rq_rcv_buf.head[0].iov_base = srcp;
1067 rqst->rq_private_buf.head[0].iov_base = srcp;
1069 /* The contents of the receive buffer that follow
1070 * head.iov_len bytes are copied into the page list.
1072 curlen = rqst->rq_rcv_buf.head[0].iov_len;
1073 if (curlen > copy_len)
1078 ppages = rqst->rq_rcv_buf.pages +
1079 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
1080 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
1081 fixup_copy_count = 0;
1082 if (copy_len && rqst->rq_rcv_buf.page_len) {
1085 pagelist_len = rqst->rq_rcv_buf.page_len;
1086 if (pagelist_len > copy_len)
1087 pagelist_len = copy_len;
1088 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
1089 for (i = 0; i < npages; i++) {
1090 curlen = PAGE_SIZE - page_base;
1091 if (curlen > pagelist_len)
1092 curlen = pagelist_len;
1094 destp = kmap_atomic(ppages[i]);
1095 memcpy(destp + page_base, srcp, curlen);
1096 flush_dcache_page(ppages[i]);
1097 kunmap_atomic(destp);
1100 fixup_copy_count += curlen;
1101 pagelist_len -= curlen;
1107 /* Implicit padding for the last segment in a Write
1108 * chunk is inserted inline at the front of the tail
1109 * iovec. The upper layer ignores the content of
1110 * the pad. Simply ensure inline content in the tail
1111 * that follows the Write chunk is properly aligned.
1117 /* The tail iovec is redirected to the remaining data
1118 * in the receive buffer, to avoid a memcopy.
1120 if (copy_len || pad) {
1121 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
1122 rqst->rq_private_buf.tail[0].iov_base = srcp;
1125 if (fixup_copy_count)
1126 trace_xprtrdma_fixup(rqst, fixup_copy_count);
1127 return fixup_copy_count;
1130 /* By convention, backchannel calls arrive via rdma_msg type
1131 * messages, and never populate the chunk lists. This makes
1132 * the RPC/RDMA header small and fixed in size, so it is
1133 * straightforward to check the RPC header's direction field.
1136 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1137 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1139 struct xdr_stream *xdr = &rep->rr_stream;
1142 if (rep->rr_proc != rdma_msg)
1145 /* Peek at stream contents without advancing. */
1146 p = xdr_inline_decode(xdr, 0);
1149 if (xdr_item_is_present(p++))
1151 if (xdr_item_is_present(p++))
1153 if (xdr_item_is_present(p++))
1157 if (*p++ != rep->rr_xid)
1159 if (*p != cpu_to_be32(RPC_CALL))
1162 /* Now that we are sure this is a backchannel call,
1163 * advance to the RPC header.
1165 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1169 rpcrdma_bc_receive_call(r_xprt, rep);
1173 pr_warn("RPC/RDMA short backward direction call\n");
1176 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1180 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1182 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1188 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1192 xdr_decode_rdma_segment(p, &handle, length, &offset);
1193 trace_xprtrdma_decode_seg(handle, *length, offset);
1197 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1199 u32 segcount, seglength;
1202 p = xdr_inline_decode(xdr, sizeof(*p));
1207 segcount = be32_to_cpup(p);
1208 while (segcount--) {
1209 if (decode_rdma_segment(xdr, &seglength))
1211 *length += seglength;
1217 /* In RPC-over-RDMA Version One replies, a Read list is never
1218 * expected. This decoder is a stub that returns an error if
1219 * a Read list is present.
1221 static int decode_read_list(struct xdr_stream *xdr)
1225 p = xdr_inline_decode(xdr, sizeof(*p));
1228 if (unlikely(xdr_item_is_present(p)))
1233 /* Supports only one Write chunk in the Write list
1235 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1244 p = xdr_inline_decode(xdr, sizeof(*p));
1247 if (xdr_item_is_absent(p))
1252 if (decode_write_chunk(xdr, &chunklen))
1254 *length += chunklen;
1260 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1264 p = xdr_inline_decode(xdr, sizeof(*p));
1269 if (xdr_item_is_present(p))
1270 if (decode_write_chunk(xdr, length))
1276 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1277 struct rpc_rqst *rqst)
1279 struct xdr_stream *xdr = &rep->rr_stream;
1280 u32 writelist, replychunk, rpclen;
1283 /* Decode the chunk lists */
1284 if (decode_read_list(xdr))
1286 if (decode_write_list(xdr, &writelist))
1288 if (decode_reply_chunk(xdr, &replychunk))
1291 /* RDMA_MSG sanity checks */
1292 if (unlikely(replychunk))
1295 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1296 base = (char *)xdr_inline_decode(xdr, 0);
1297 rpclen = xdr_stream_remaining(xdr);
1298 r_xprt->rx_stats.fixup_copy_count +=
1299 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1301 r_xprt->rx_stats.total_rdma_reply += writelist;
1302 return rpclen + xdr_align_size(writelist);
1306 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1308 struct xdr_stream *xdr = &rep->rr_stream;
1309 u32 writelist, replychunk;
1311 /* Decode the chunk lists */
1312 if (decode_read_list(xdr))
1314 if (decode_write_list(xdr, &writelist))
1316 if (decode_reply_chunk(xdr, &replychunk))
1319 /* RDMA_NOMSG sanity checks */
1320 if (unlikely(writelist))
1322 if (unlikely(!replychunk))
1325 /* Reply chunk buffer already is the reply vector */
1326 r_xprt->rx_stats.total_rdma_reply += replychunk;
1331 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1332 struct rpc_rqst *rqst)
1334 struct xdr_stream *xdr = &rep->rr_stream;
1337 p = xdr_inline_decode(xdr, sizeof(*p));
1343 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1346 trace_xprtrdma_err_vers(rqst, p, p + 1);
1349 trace_xprtrdma_err_chunk(rqst);
1352 trace_xprtrdma_err_unrecognized(rqst, p);
1358 /* Perform XID lookup, reconstruction of the RPC reply, and
1359 * RPC completion while holding the transport lock to ensure
1360 * the rep, rqst, and rq_task pointers remain stable.
1362 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1364 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1365 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1366 struct rpc_rqst *rqst = rep->rr_rqst;
1369 switch (rep->rr_proc) {
1371 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1374 status = rpcrdma_decode_nomsg(r_xprt, rep);
1377 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1386 spin_lock(&xprt->queue_lock);
1387 xprt_complete_rqst(rqst->rq_task, status);
1388 xprt_unpin_rqst(rqst);
1389 spin_unlock(&xprt->queue_lock);
1393 trace_xprtrdma_reply_hdr_err(rep);
1394 r_xprt->rx_stats.bad_reply_count++;
1395 rqst->rq_task->tk_status = status;
1400 static void rpcrdma_reply_done(struct kref *kref)
1402 struct rpcrdma_req *req =
1403 container_of(kref, struct rpcrdma_req, rl_kref);
1405 rpcrdma_complete_rqst(req->rl_reply);
1409 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1410 * @rep: Incoming rpcrdma_rep object to process
1412 * Errors must result in the RPC task either being awakened, or
1413 * allowed to timeout, to discover the errors at that time.
1415 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1417 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1418 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1419 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1420 struct rpcrdma_req *req;
1421 struct rpc_rqst *rqst;
1425 /* Any data means we had a useful conversation, so
1426 * then we don't need to delay the next reconnect.
1428 if (xprt->reestablish_timeout)
1429 xprt->reestablish_timeout = 0;
1431 /* Fixed transport header fields */
1432 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1433 rep->rr_hdrbuf.head[0].iov_base, NULL);
1434 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1436 goto out_shortreply;
1438 rep->rr_vers = *p++;
1439 credits = be32_to_cpu(*p++);
1440 rep->rr_proc = *p++;
1442 if (rep->rr_vers != rpcrdma_version)
1443 goto out_badversion;
1445 if (rpcrdma_is_bcall(r_xprt, rep))
1448 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1449 * get context for handling any incoming chunks.
1451 spin_lock(&xprt->queue_lock);
1452 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1455 xprt_pin_rqst(rqst);
1456 spin_unlock(&xprt->queue_lock);
1459 credits = 1; /* don't deadlock */
1460 else if (credits > r_xprt->rx_ep->re_max_requests)
1461 credits = r_xprt->rx_ep->re_max_requests;
1462 if (buf->rb_credits != credits)
1463 rpcrdma_update_cwnd(r_xprt, credits);
1464 rpcrdma_post_recvs(r_xprt, false);
1466 req = rpcr_to_rdmar(rqst);
1467 if (unlikely(req->rl_reply))
1468 rpcrdma_recv_buffer_put(req->rl_reply);
1469 req->rl_reply = rep;
1470 rep->rr_rqst = rqst;
1472 trace_xprtrdma_reply(rqst->rq_task, rep, credits);
1474 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1475 frwr_reminv(rep, &req->rl_registered);
1476 if (!list_empty(&req->rl_registered))
1477 frwr_unmap_async(r_xprt, req);
1478 /* LocalInv completion will complete the RPC */
1480 kref_put(&req->rl_kref, rpcrdma_reply_done);
1484 trace_xprtrdma_reply_vers_err(rep);
1488 spin_unlock(&xprt->queue_lock);
1489 trace_xprtrdma_reply_rqst_err(rep);
1493 trace_xprtrdma_reply_short_err(rep);
1496 rpcrdma_recv_buffer_put(rep);