1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * This file contains the guts of the RPC RDMA protocol, and
46 * does marshaling/unmarshaling, etc. It is also where interfacing
47 * to the Linux RPC framework lives.
50 #include <linux/highmem.h>
52 #include <linux/sunrpc/svc_rdma.h>
54 #include "xprt_rdma.h"
55 #include <trace/events/rpcrdma.h>
57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
58 # define RPCDBG_FACILITY RPCDBG_TRANS
61 /* Returns size of largest RPC-over-RDMA header in a Call message
63 * The largest Call header contains a full-size Read list and a
64 * minimal Reply chunk.
66 static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
70 /* Fixed header fields and list discriminators */
71 size = RPCRDMA_HDRLEN_MIN;
73 /* Maximum Read list size */
74 size = maxsegs * rpcrdma_readchunk_maxsz * sizeof(__be32);
76 /* Minimal Read chunk size */
77 size += sizeof(__be32); /* segment count */
78 size += rpcrdma_segment_maxsz * sizeof(__be32);
79 size += sizeof(__be32); /* list discriminator */
81 dprintk("RPC: %s: max call header size = %u\n",
86 /* Returns size of largest RPC-over-RDMA header in a Reply message
88 * There is only one Write list or one Reply chunk per Reply
89 * message. The larger list is the Write list.
91 static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
95 /* Fixed header fields and list discriminators */
96 size = RPCRDMA_HDRLEN_MIN;
98 /* Maximum Write list size */
99 size = sizeof(__be32); /* segment count */
100 size += maxsegs * rpcrdma_segment_maxsz * sizeof(__be32);
101 size += sizeof(__be32); /* list discriminator */
103 dprintk("RPC: %s: max reply header size = %u\n",
109 * rpcrdma_set_max_header_sizes - Initialize inline payload sizes
110 * @r_xprt: transport instance to initialize
112 * The max_inline fields contain the maximum size of an RPC message
113 * so the marshaling code doesn't have to repeat this calculation
116 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *r_xprt)
118 unsigned int maxsegs = r_xprt->rx_ia.ri_max_segs;
119 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
121 ep->rep_max_inline_send =
122 ep->rep_inline_send - rpcrdma_max_call_header_size(maxsegs);
123 ep->rep_max_inline_recv =
124 ep->rep_inline_recv - rpcrdma_max_reply_header_size(maxsegs);
127 /* The client can send a request inline as long as the RPCRDMA header
128 * plus the RPC call fit under the transport's inline limit. If the
129 * combined call message size exceeds that limit, the client must use
130 * a Read chunk for this operation.
132 * A Read chunk is also required if sending the RPC call inline would
133 * exceed this device's max_sge limit.
135 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
136 struct rpc_rqst *rqst)
138 struct xdr_buf *xdr = &rqst->rq_snd_buf;
139 unsigned int count, remaining, offset;
141 if (xdr->len > r_xprt->rx_ep.rep_max_inline_send)
145 remaining = xdr->page_len;
146 offset = offset_in_page(xdr->page_base);
147 count = RPCRDMA_MIN_SEND_SGES;
149 remaining -= min_t(unsigned int,
150 PAGE_SIZE - offset, remaining);
152 if (++count > r_xprt->rx_ia.ri_max_send_sges)
160 /* The client can't know how large the actual reply will be. Thus it
161 * plans for the largest possible reply for that particular ULP
162 * operation. If the maximum combined reply message size exceeds that
163 * limit, the client must provide a write list or a reply chunk for
166 static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
167 struct rpc_rqst *rqst)
169 return rqst->rq_rcv_buf.buflen <= r_xprt->rx_ep.rep_max_inline_recv;
172 /* The client is required to provide a Reply chunk if the maximum
173 * size of the non-payload part of the RPC Reply is larger than
174 * the inline threshold.
177 rpcrdma_nonpayload_inline(const struct rpcrdma_xprt *r_xprt,
178 const struct rpc_rqst *rqst)
180 const struct xdr_buf *buf = &rqst->rq_rcv_buf;
182 return (buf->head[0].iov_len + buf->tail[0].iov_len) <
183 r_xprt->rx_ep.rep_max_inline_recv;
186 /* Split @vec on page boundaries into SGEs. FMR registers pages, not
187 * a byte range. Other modes coalesce these SGEs into a single MR
190 * Returns pointer to next available SGE, and bumps the total number
193 static struct rpcrdma_mr_seg *
194 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
197 u32 remaining, page_offset;
200 base = vec->iov_base;
201 page_offset = offset_in_page(base);
202 remaining = vec->iov_len;
205 seg->mr_offset = base;
206 seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
207 remaining -= seg->mr_len;
216 /* Convert @xdrbuf into SGEs no larger than a page each. As they
217 * are registered, these SGEs are then coalesced into RDMA segments
218 * when the selected memreg mode supports it.
220 * Returns positive number of SGEs consumed, or a negative errno.
224 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
225 unsigned int pos, enum rpcrdma_chunktype type,
226 struct rpcrdma_mr_seg *seg)
228 unsigned long page_base;
230 struct page **ppages;
234 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
236 len = xdrbuf->page_len;
237 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
238 page_base = offset_in_page(xdrbuf->page_base);
240 /* ACL likes to be lazy in allocating pages - ACLs
241 * are small by default but can get huge.
243 if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
245 *ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
249 seg->mr_page = *ppages;
250 seg->mr_offset = (char *)page_base;
251 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
259 /* When encoding a Read chunk, the tail iovec contains an
260 * XDR pad and may be omitted.
262 if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
265 /* When encoding a Write chunk, some servers need to see an
266 * extra segment for non-XDR-aligned Write chunks. The upper
267 * layer provides space in the tail iovec that may be used
270 if (type == rpcrdma_writech && r_xprt->rx_ia.ri_implicit_roundup)
273 if (xdrbuf->tail[0].iov_len)
274 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
277 if (unlikely(n > RPCRDMA_MAX_SEGS))
283 encode_item_present(struct xdr_stream *xdr)
287 p = xdr_reserve_space(xdr, sizeof(*p));
296 encode_item_not_present(struct xdr_stream *xdr)
300 p = xdr_reserve_space(xdr, sizeof(*p));
309 xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr *mr)
311 *iptr++ = cpu_to_be32(mr->mr_handle);
312 *iptr++ = cpu_to_be32(mr->mr_length);
313 xdr_encode_hyper(iptr, mr->mr_offset);
317 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
321 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
325 xdr_encode_rdma_segment(p, mr);
330 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
335 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
339 *p++ = xdr_one; /* Item present */
340 *p++ = cpu_to_be32(position);
341 xdr_encode_rdma_segment(p, mr);
345 /* Register and XDR encode the Read list. Supports encoding a list of read
346 * segments that belong to a single read chunk.
348 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
350 * Read chunklist (a linked list):
351 * N elements, position P (same P for all chunks of same arg!):
352 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
354 * Returns zero on success, or a negative errno if a failure occurred.
355 * @xdr is advanced to the next position in the stream.
357 * Only a single @pos value is currently supported.
360 rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
361 struct rpc_rqst *rqst, enum rpcrdma_chunktype rtype)
363 struct xdr_stream *xdr = &req->rl_stream;
364 struct rpcrdma_mr_seg *seg;
365 struct rpcrdma_mr *mr;
369 if (rtype == rpcrdma_noch)
372 pos = rqst->rq_snd_buf.head[0].iov_len;
373 if (rtype == rpcrdma_areadch)
375 seg = req->rl_segments;
376 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
382 seg = frwr_map(r_xprt, seg, nsegs, false, rqst->rq_xid, &mr);
385 rpcrdma_mr_push(mr, &req->rl_registered);
387 if (encode_read_segment(xdr, mr, pos) < 0)
390 trace_xprtrdma_chunk_read(rqst->rq_task, pos, mr, nsegs);
391 r_xprt->rx_stats.read_chunk_count++;
392 nsegs -= mr->mr_nents;
396 return encode_item_not_present(xdr);
399 /* Register and XDR encode the Write list. Supports encoding a list
400 * containing one array of plain segments that belong to a single
403 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
405 * Write chunklist (a list of (one) counted array):
407 * 1 - N - HLOO - HLOO - ... - HLOO - 0
409 * Returns zero on success, or a negative errno if a failure occurred.
410 * @xdr is advanced to the next position in the stream.
412 * Only a single Write chunk is currently supported.
415 rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
416 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
418 struct xdr_stream *xdr = &req->rl_stream;
419 struct rpcrdma_mr_seg *seg;
420 struct rpcrdma_mr *mr;
424 if (wtype != rpcrdma_writech)
427 seg = req->rl_segments;
428 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
429 rqst->rq_rcv_buf.head[0].iov_len,
434 if (encode_item_present(xdr) < 0)
436 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
437 if (unlikely(!segcount))
439 /* Actual value encoded below */
443 seg = frwr_map(r_xprt, seg, nsegs, true, rqst->rq_xid, &mr);
446 rpcrdma_mr_push(mr, &req->rl_registered);
448 if (encode_rdma_segment(xdr, mr) < 0)
451 trace_xprtrdma_chunk_write(rqst->rq_task, mr, nsegs);
452 r_xprt->rx_stats.write_chunk_count++;
453 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
455 nsegs -= mr->mr_nents;
458 /* Update count of segments in this Write chunk */
459 *segcount = cpu_to_be32(nchunks);
462 return encode_item_not_present(xdr);
465 /* Register and XDR encode the Reply chunk. Supports encoding an array
466 * of plain segments that belong to a single write (reply) chunk.
468 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
470 * Reply chunk (a counted array):
472 * 1 - N - HLOO - HLOO - ... - HLOO
474 * Returns zero on success, or a negative errno if a failure occurred.
475 * @xdr is advanced to the next position in the stream.
478 rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
479 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
481 struct xdr_stream *xdr = &req->rl_stream;
482 struct rpcrdma_mr_seg *seg;
483 struct rpcrdma_mr *mr;
487 if (wtype != rpcrdma_replych)
488 return encode_item_not_present(xdr);
490 seg = req->rl_segments;
491 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
495 if (encode_item_present(xdr) < 0)
497 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
498 if (unlikely(!segcount))
500 /* Actual value encoded below */
504 seg = frwr_map(r_xprt, seg, nsegs, true, rqst->rq_xid, &mr);
507 rpcrdma_mr_push(mr, &req->rl_registered);
509 if (encode_rdma_segment(xdr, mr) < 0)
512 trace_xprtrdma_chunk_reply(rqst->rq_task, mr, nsegs);
513 r_xprt->rx_stats.reply_chunk_count++;
514 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
516 nsegs -= mr->mr_nents;
519 /* Update count of segments in the Reply chunk */
520 *segcount = cpu_to_be32(nchunks);
525 static void rpcrdma_sendctx_done(struct kref *kref)
527 struct rpcrdma_req *req =
528 container_of(kref, struct rpcrdma_req, rl_kref);
529 struct rpcrdma_rep *rep = req->rl_reply;
531 rpcrdma_complete_rqst(rep);
532 rep->rr_rxprt->rx_stats.reply_waits_for_send++;
536 * rpcrdma_sendctx_unmap - DMA-unmap Send buffer
537 * @sc: sendctx containing SGEs to unmap
540 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc)
544 if (!sc->sc_unmap_count)
547 /* The first two SGEs contain the transport header and
548 * the inline buffer. These are always left mapped so
549 * they can be cheaply re-used.
551 for (sge = &sc->sc_sges[2]; sc->sc_unmap_count;
552 ++sge, --sc->sc_unmap_count)
553 ib_dma_unmap_page(sc->sc_device, sge->addr, sge->length,
556 kref_put(&sc->sc_req->rl_kref, rpcrdma_sendctx_done);
559 /* Prepare an SGE for the RPC-over-RDMA transport header.
561 static bool rpcrdma_prepare_hdr_sge(struct rpcrdma_xprt *r_xprt,
562 struct rpcrdma_req *req, u32 len)
564 struct rpcrdma_sendctx *sc = req->rl_sendctx;
565 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
566 struct ib_sge *sge = sc->sc_sges;
568 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
570 sge->addr = rdmab_addr(rb);
572 sge->lkey = rdmab_lkey(rb);
574 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr, sge->length,
580 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
584 /* Prepare the Send SGEs. The head and tail iovec, and each entry
585 * in the page list, gets its own SGE.
587 static bool rpcrdma_prepare_msg_sges(struct rpcrdma_xprt *r_xprt,
588 struct rpcrdma_req *req,
590 enum rpcrdma_chunktype rtype)
592 struct rpcrdma_sendctx *sc = req->rl_sendctx;
593 unsigned int sge_no, page_base, len, remaining;
594 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
595 struct ib_sge *sge = sc->sc_sges;
596 struct page *page, **ppages;
598 /* The head iovec is straightforward, as it is already
599 * DMA-mapped. Sync the content that has changed.
601 if (!rpcrdma_regbuf_dma_map(r_xprt, rb))
603 sc->sc_device = rdmab_device(rb);
605 sge[sge_no].addr = rdmab_addr(rb);
606 sge[sge_no].length = xdr->head[0].iov_len;
607 sge[sge_no].lkey = rdmab_lkey(rb);
608 ib_dma_sync_single_for_device(rdmab_device(rb), sge[sge_no].addr,
609 sge[sge_no].length, DMA_TO_DEVICE);
611 /* If there is a Read chunk, the page list is being handled
612 * via explicit RDMA, and thus is skipped here. However, the
613 * tail iovec may include an XDR pad for the page list, as
614 * well as additional content, and may not reside in the
615 * same page as the head iovec.
617 if (rtype == rpcrdma_readch) {
618 len = xdr->tail[0].iov_len;
620 /* Do not include the tail if it is only an XDR pad */
624 page = virt_to_page(xdr->tail[0].iov_base);
625 page_base = offset_in_page(xdr->tail[0].iov_base);
627 /* If the content in the page list is an odd length,
628 * xdr_write_pages() has added a pad at the beginning
629 * of the tail iovec. Force the tail's non-pad content
630 * to land at the next XDR position in the Send message.
632 page_base += len & 3;
637 /* If there is a page list present, temporarily DMA map
638 * and prepare an SGE for each page to be sent.
641 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
642 page_base = offset_in_page(xdr->page_base);
643 remaining = xdr->page_len;
646 if (sge_no > RPCRDMA_MAX_SEND_SGES - 2)
647 goto out_mapping_overflow;
649 len = min_t(u32, PAGE_SIZE - page_base, remaining);
651 ib_dma_map_page(rdmab_device(rb), *ppages,
652 page_base, len, DMA_TO_DEVICE);
653 if (ib_dma_mapping_error(rdmab_device(rb),
655 goto out_mapping_err;
656 sge[sge_no].length = len;
657 sge[sge_no].lkey = rdmab_lkey(rb);
659 sc->sc_unmap_count++;
666 /* The tail iovec is not always constructed in the same
667 * page where the head iovec resides (see, for example,
668 * gss_wrap_req_priv). To neatly accommodate that case,
669 * DMA map it separately.
671 if (xdr->tail[0].iov_len) {
672 page = virt_to_page(xdr->tail[0].iov_base);
673 page_base = offset_in_page(xdr->tail[0].iov_base);
674 len = xdr->tail[0].iov_len;
679 ib_dma_map_page(rdmab_device(rb), page, page_base, len,
681 if (ib_dma_mapping_error(rdmab_device(rb), sge[sge_no].addr))
682 goto out_mapping_err;
683 sge[sge_no].length = len;
684 sge[sge_no].lkey = rdmab_lkey(rb);
685 sc->sc_unmap_count++;
689 sc->sc_wr.num_sge += sge_no;
690 if (sc->sc_unmap_count)
691 kref_get(&req->rl_kref);
695 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
698 out_mapping_overflow:
699 rpcrdma_sendctx_unmap(sc);
700 pr_err("rpcrdma: too many Send SGEs (%u)\n", sge_no);
704 rpcrdma_sendctx_unmap(sc);
705 trace_xprtrdma_dma_maperr(sge[sge_no].addr);
710 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
711 * @r_xprt: controlling transport
712 * @req: context of RPC Call being marshalled
713 * @hdrlen: size of transport header, in bytes
714 * @xdr: xdr_buf containing RPC Call
715 * @rtype: chunk type being encoded
717 * Returns 0 on success; otherwise a negative errno is returned.
720 rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
721 struct rpcrdma_req *req, u32 hdrlen,
722 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
727 req->rl_sendctx = rpcrdma_sendctx_get_locked(r_xprt);
728 if (!req->rl_sendctx)
730 req->rl_sendctx->sc_wr.num_sge = 0;
731 req->rl_sendctx->sc_unmap_count = 0;
732 req->rl_sendctx->sc_req = req;
733 kref_init(&req->rl_kref);
736 if (!rpcrdma_prepare_hdr_sge(r_xprt, req, hdrlen))
738 if (rtype != rpcrdma_areadch)
739 if (!rpcrdma_prepare_msg_sges(r_xprt, req, xdr, rtype))
744 trace_xprtrdma_prepsend_failed(&req->rl_slot, ret);
749 * rpcrdma_marshal_req - Marshal and send one RPC request
750 * @r_xprt: controlling transport
751 * @rqst: RPC request to be marshaled
753 * For the RPC in "rqst", this function:
754 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
755 * - Registers Read, Write, and Reply chunks
756 * - Constructs the transport header
757 * - Posts a Send WR to send the transport header and request
760 * %0 if the RPC was sent successfully,
761 * %-ENOTCONN if the connection was lost,
762 * %-EAGAIN if the caller should call again with the same arguments,
763 * %-ENOBUFS if the caller should call again after a delay,
764 * %-EMSGSIZE if the transport header is too small,
765 * %-EIO if a permanent problem occurred while marshaling.
768 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
770 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
771 struct xdr_stream *xdr = &req->rl_stream;
772 enum rpcrdma_chunktype rtype, wtype;
777 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
778 xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
781 /* Fixed header fields */
783 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
787 *p++ = rpcrdma_version;
788 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
790 /* When the ULP employs a GSS flavor that guarantees integrity
791 * or privacy, direct data placement of individual data items
794 ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
795 RPCAUTH_AUTH_DATATOUCH);
798 * Chunks needed for results?
800 * o If the expected result is under the inline threshold, all ops
802 * o Large read ops return data as write chunk(s), header as
804 * o Large non-read ops return as a single reply chunk.
806 if (rpcrdma_results_inline(r_xprt, rqst))
807 wtype = rpcrdma_noch;
808 else if ((ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) &&
809 rpcrdma_nonpayload_inline(r_xprt, rqst))
810 wtype = rpcrdma_writech;
812 wtype = rpcrdma_replych;
815 * Chunks needed for arguments?
817 * o If the total request is under the inline threshold, all ops
818 * are sent as inline.
819 * o Large write ops transmit data as read chunk(s), header as
821 * o Large non-write ops are sent with the entire message as a
822 * single read chunk (protocol 0-position special case).
824 * This assumes that the upper layer does not present a request
825 * that both has a data payload, and whose non-data arguments
826 * by themselves are larger than the inline threshold.
828 if (rpcrdma_args_inline(r_xprt, rqst)) {
830 rtype = rpcrdma_noch;
831 } else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
833 rtype = rpcrdma_readch;
835 r_xprt->rx_stats.nomsg_call_count++;
837 rtype = rpcrdma_areadch;
840 /* If this is a retransmit, discard previously registered
841 * chunks. Very likely the connection has been replaced,
842 * so these registrations are invalid and unusable.
844 while (unlikely(!list_empty(&req->rl_registered))) {
845 struct rpcrdma_mr *mr;
847 mr = rpcrdma_mr_pop(&req->rl_registered);
848 rpcrdma_mr_recycle(mr);
851 /* This implementation supports the following combinations
852 * of chunk lists in one RPC-over-RDMA Call message:
857 * - Read list + Reply chunk
859 * It might not yet support the following combinations:
861 * - Read list + Write list
863 * It does not support the following combinations:
865 * - Write list + Reply chunk
866 * - Read list + Write list + Reply chunk
868 * This implementation supports only a single chunk in each
869 * Read or Write list. Thus for example the client cannot
870 * send a Call message with a Position Zero Read chunk and a
871 * regular Read chunk at the same time.
873 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
876 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
879 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
883 ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
884 &rqst->rq_snd_buf, rtype);
888 trace_xprtrdma_marshal(req, rtype, wtype);
892 trace_xprtrdma_marshal_failed(rqst, ret);
893 r_xprt->rx_stats.failed_marshal_count++;
899 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
900 * @rqst: controlling RPC request
901 * @srcp: points to RPC message payload in receive buffer
902 * @copy_len: remaining length of receive buffer content
903 * @pad: Write chunk pad bytes needed (zero for pure inline)
905 * The upper layer has set the maximum number of bytes it can
906 * receive in each component of rq_rcv_buf. These values are set in
907 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
909 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
910 * many cases this function simply updates iov_base pointers in
911 * rq_rcv_buf to point directly to the received reply data, to
912 * avoid copying reply data.
914 * Returns the count of bytes which had to be memcopied.
917 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
919 unsigned long fixup_copy_count;
920 int i, npages, curlen;
922 struct page **ppages;
925 /* The head iovec is redirected to the RPC reply message
926 * in the receive buffer, to avoid a memcopy.
928 rqst->rq_rcv_buf.head[0].iov_base = srcp;
929 rqst->rq_private_buf.head[0].iov_base = srcp;
931 /* The contents of the receive buffer that follow
932 * head.iov_len bytes are copied into the page list.
934 curlen = rqst->rq_rcv_buf.head[0].iov_len;
935 if (curlen > copy_len)
937 trace_xprtrdma_fixup(rqst, copy_len, curlen);
941 ppages = rqst->rq_rcv_buf.pages +
942 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
943 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
944 fixup_copy_count = 0;
945 if (copy_len && rqst->rq_rcv_buf.page_len) {
948 pagelist_len = rqst->rq_rcv_buf.page_len;
949 if (pagelist_len > copy_len)
950 pagelist_len = copy_len;
951 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
952 for (i = 0; i < npages; i++) {
953 curlen = PAGE_SIZE - page_base;
954 if (curlen > pagelist_len)
955 curlen = pagelist_len;
957 trace_xprtrdma_fixup_pg(rqst, i, srcp,
959 destp = kmap_atomic(ppages[i]);
960 memcpy(destp + page_base, srcp, curlen);
961 flush_dcache_page(ppages[i]);
962 kunmap_atomic(destp);
965 fixup_copy_count += curlen;
966 pagelist_len -= curlen;
972 /* Implicit padding for the last segment in a Write
973 * chunk is inserted inline at the front of the tail
974 * iovec. The upper layer ignores the content of
975 * the pad. Simply ensure inline content in the tail
976 * that follows the Write chunk is properly aligned.
982 /* The tail iovec is redirected to the remaining data
983 * in the receive buffer, to avoid a memcopy.
985 if (copy_len || pad) {
986 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
987 rqst->rq_private_buf.tail[0].iov_base = srcp;
990 return fixup_copy_count;
993 /* By convention, backchannel calls arrive via rdma_msg type
994 * messages, and never populate the chunk lists. This makes
995 * the RPC/RDMA header small and fixed in size, so it is
996 * straightforward to check the RPC header's direction field.
999 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1000 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1002 struct xdr_stream *xdr = &rep->rr_stream;
1005 if (rep->rr_proc != rdma_msg)
1008 /* Peek at stream contents without advancing. */
1009 p = xdr_inline_decode(xdr, 0);
1012 if (*p++ != xdr_zero)
1014 if (*p++ != xdr_zero)
1016 if (*p++ != xdr_zero)
1020 if (*p++ != rep->rr_xid)
1022 if (*p != cpu_to_be32(RPC_CALL))
1025 /* Now that we are sure this is a backchannel call,
1026 * advance to the RPC header.
1028 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1032 rpcrdma_bc_receive_call(r_xprt, rep);
1036 pr_warn("RPC/RDMA short backward direction call\n");
1039 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1043 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1045 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1051 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1055 handle = be32_to_cpup(p++);
1056 *length = be32_to_cpup(p++);
1057 xdr_decode_hyper(p, &offset);
1059 trace_xprtrdma_decode_seg(handle, *length, offset);
1063 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1065 u32 segcount, seglength;
1068 p = xdr_inline_decode(xdr, sizeof(*p));
1073 segcount = be32_to_cpup(p);
1074 while (segcount--) {
1075 if (decode_rdma_segment(xdr, &seglength))
1077 *length += seglength;
1083 /* In RPC-over-RDMA Version One replies, a Read list is never
1084 * expected. This decoder is a stub that returns an error if
1085 * a Read list is present.
1087 static int decode_read_list(struct xdr_stream *xdr)
1091 p = xdr_inline_decode(xdr, sizeof(*p));
1094 if (unlikely(*p != xdr_zero))
1099 /* Supports only one Write chunk in the Write list
1101 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1110 p = xdr_inline_decode(xdr, sizeof(*p));
1118 if (decode_write_chunk(xdr, &chunklen))
1120 *length += chunklen;
1126 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1130 p = xdr_inline_decode(xdr, sizeof(*p));
1136 if (decode_write_chunk(xdr, length))
1142 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1143 struct rpc_rqst *rqst)
1145 struct xdr_stream *xdr = &rep->rr_stream;
1146 u32 writelist, replychunk, rpclen;
1149 /* Decode the chunk lists */
1150 if (decode_read_list(xdr))
1152 if (decode_write_list(xdr, &writelist))
1154 if (decode_reply_chunk(xdr, &replychunk))
1157 /* RDMA_MSG sanity checks */
1158 if (unlikely(replychunk))
1161 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1162 base = (char *)xdr_inline_decode(xdr, 0);
1163 rpclen = xdr_stream_remaining(xdr);
1164 r_xprt->rx_stats.fixup_copy_count +=
1165 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1167 r_xprt->rx_stats.total_rdma_reply += writelist;
1168 return rpclen + xdr_align_size(writelist);
1172 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1174 struct xdr_stream *xdr = &rep->rr_stream;
1175 u32 writelist, replychunk;
1177 /* Decode the chunk lists */
1178 if (decode_read_list(xdr))
1180 if (decode_write_list(xdr, &writelist))
1182 if (decode_reply_chunk(xdr, &replychunk))
1185 /* RDMA_NOMSG sanity checks */
1186 if (unlikely(writelist))
1188 if (unlikely(!replychunk))
1191 /* Reply chunk buffer already is the reply vector */
1192 r_xprt->rx_stats.total_rdma_reply += replychunk;
1197 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1198 struct rpc_rqst *rqst)
1200 struct xdr_stream *xdr = &rep->rr_stream;
1203 p = xdr_inline_decode(xdr, sizeof(*p));
1209 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1212 dprintk("RPC: %s: server reports "
1213 "version error (%u-%u), xid %08x\n", __func__,
1214 be32_to_cpup(p), be32_to_cpu(*(p + 1)),
1215 be32_to_cpu(rep->rr_xid));
1218 dprintk("RPC: %s: server reports "
1219 "header decoding error, xid %08x\n", __func__,
1220 be32_to_cpu(rep->rr_xid));
1223 dprintk("RPC: %s: server reports "
1224 "unrecognized error %d, xid %08x\n", __func__,
1225 be32_to_cpup(p), be32_to_cpu(rep->rr_xid));
1228 r_xprt->rx_stats.bad_reply_count++;
1232 /* Perform XID lookup, reconstruction of the RPC reply, and
1233 * RPC completion while holding the transport lock to ensure
1234 * the rep, rqst, and rq_task pointers remain stable.
1236 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1238 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1239 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1240 struct rpc_rqst *rqst = rep->rr_rqst;
1243 xprt->reestablish_timeout = 0;
1245 switch (rep->rr_proc) {
1247 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1250 status = rpcrdma_decode_nomsg(r_xprt, rep);
1253 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1262 spin_lock(&xprt->queue_lock);
1263 xprt_complete_rqst(rqst->rq_task, status);
1264 xprt_unpin_rqst(rqst);
1265 spin_unlock(&xprt->queue_lock);
1268 /* If the incoming reply terminated a pending RPC, the next
1269 * RPC call will post a replacement receive buffer as it is
1273 trace_xprtrdma_reply_hdr(rep);
1274 r_xprt->rx_stats.bad_reply_count++;
1278 static void rpcrdma_reply_done(struct kref *kref)
1280 struct rpcrdma_req *req =
1281 container_of(kref, struct rpcrdma_req, rl_kref);
1283 rpcrdma_complete_rqst(req->rl_reply);
1287 * rpcrdma_reply_handler - Process received RPC/RDMA messages
1288 * @rep: Incoming rpcrdma_rep object to process
1290 * Errors must result in the RPC task either being awakened, or
1291 * allowed to timeout, to discover the errors at that time.
1293 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1295 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1296 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1297 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1298 struct rpcrdma_req *req;
1299 struct rpc_rqst *rqst;
1303 /* Fixed transport header fields */
1304 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1305 rep->rr_hdrbuf.head[0].iov_base, NULL);
1306 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1308 goto out_shortreply;
1310 rep->rr_vers = *p++;
1311 credits = be32_to_cpu(*p++);
1312 rep->rr_proc = *p++;
1314 if (rep->rr_vers != rpcrdma_version)
1315 goto out_badversion;
1317 if (rpcrdma_is_bcall(r_xprt, rep))
1320 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1321 * get context for handling any incoming chunks.
1323 spin_lock(&xprt->queue_lock);
1324 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1327 xprt_pin_rqst(rqst);
1328 spin_unlock(&xprt->queue_lock);
1331 credits = 1; /* don't deadlock */
1332 else if (credits > buf->rb_max_requests)
1333 credits = buf->rb_max_requests;
1334 if (buf->rb_credits != credits) {
1335 spin_lock(&xprt->transport_lock);
1336 buf->rb_credits = credits;
1337 xprt->cwnd = credits << RPC_CWNDSHIFT;
1338 spin_unlock(&xprt->transport_lock);
1341 req = rpcr_to_rdmar(rqst);
1342 if (req->rl_reply) {
1343 trace_xprtrdma_leaked_rep(rqst, req->rl_reply);
1344 rpcrdma_recv_buffer_put(req->rl_reply);
1346 req->rl_reply = rep;
1347 rep->rr_rqst = rqst;
1349 trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1351 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1352 frwr_reminv(rep, &req->rl_registered);
1353 if (!list_empty(&req->rl_registered))
1354 frwr_unmap_async(r_xprt, req);
1355 /* LocalInv completion will complete the RPC */
1357 kref_put(&req->rl_kref, rpcrdma_reply_done);
1361 trace_xprtrdma_reply_vers(rep);
1365 spin_unlock(&xprt->queue_lock);
1366 trace_xprtrdma_reply_rqst(rep);
1370 trace_xprtrdma_reply_short(rep);
1373 rpcrdma_recv_buffer_put(rep);