2 * Copyright (c) 2014-2017 Oracle. All rights reserved.
3 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 * This file contains the guts of the RPC RDMA protocol, and
45 * does marshaling/unmarshaling, etc. It is also where interfacing
46 * to the Linux RPC framework lives.
49 #include "xprt_rdma.h"
51 #include <linux/highmem.h>
53 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
54 # define RPCDBG_FACILITY RPCDBG_TRANS
57 static const char transfertypes[][12] = {
58 "inline", /* no chunks */
59 "read list", /* some argument via rdma read */
60 "*read list", /* entire request via rdma read */
61 "write list", /* some result via rdma write */
62 "reply chunk" /* entire reply via rdma write */
65 /* Returns size of largest RPC-over-RDMA header in a Call message
67 * The largest Call header contains a full-size Read list and a
68 * minimal Reply chunk.
70 static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
74 /* Fixed header fields and list discriminators */
75 size = RPCRDMA_HDRLEN_MIN;
77 /* Maximum Read list size */
78 maxsegs += 2; /* segment for head and tail buffers */
79 size = maxsegs * rpcrdma_readchunk_maxsz * sizeof(__be32);
81 /* Minimal Read chunk size */
82 size += sizeof(__be32); /* segment count */
83 size += rpcrdma_segment_maxsz * sizeof(__be32);
84 size += sizeof(__be32); /* list discriminator */
86 dprintk("RPC: %s: max call header size = %u\n",
91 /* Returns size of largest RPC-over-RDMA header in a Reply message
93 * There is only one Write list or one Reply chunk per Reply
94 * message. The larger list is the Write list.
96 static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
100 /* Fixed header fields and list discriminators */
101 size = RPCRDMA_HDRLEN_MIN;
103 /* Maximum Write list size */
104 maxsegs += 2; /* segment for head and tail buffers */
105 size = sizeof(__be32); /* segment count */
106 size += maxsegs * rpcrdma_segment_maxsz * sizeof(__be32);
107 size += sizeof(__be32); /* list discriminator */
109 dprintk("RPC: %s: max reply header size = %u\n",
114 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *r_xprt)
116 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
117 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
118 unsigned int maxsegs = ia->ri_max_segs;
120 ia->ri_max_inline_write = cdata->inline_wsize -
121 rpcrdma_max_call_header_size(maxsegs);
122 ia->ri_max_inline_read = cdata->inline_rsize -
123 rpcrdma_max_reply_header_size(maxsegs);
126 /* The client can send a request inline as long as the RPCRDMA header
127 * plus the RPC call fit under the transport's inline limit. If the
128 * combined call message size exceeds that limit, the client must use
129 * a Read chunk for this operation.
131 * A Read chunk is also required if sending the RPC call inline would
132 * exceed this device's max_sge limit.
134 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
135 struct rpc_rqst *rqst)
137 struct xdr_buf *xdr = &rqst->rq_snd_buf;
138 unsigned int count, remaining, offset;
140 if (xdr->len > r_xprt->rx_ia.ri_max_inline_write)
144 remaining = xdr->page_len;
145 offset = offset_in_page(xdr->page_base);
146 count = RPCRDMA_MIN_SEND_SGES;
148 remaining -= min_t(unsigned int,
149 PAGE_SIZE - offset, remaining);
151 if (++count > r_xprt->rx_ia.ri_max_send_sges)
159 /* The client can't know how large the actual reply will be. Thus it
160 * plans for the largest possible reply for that particular ULP
161 * operation. If the maximum combined reply message size exceeds that
162 * limit, the client must provide a write list or a reply chunk for
165 static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
166 struct rpc_rqst *rqst)
168 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
170 return rqst->rq_rcv_buf.buflen <= ia->ri_max_inline_read;
173 /* Split @vec on page boundaries into SGEs. FMR registers pages, not
174 * a byte range. Other modes coalesce these SGEs into a single MR
177 * Returns pointer to next available SGE, and bumps the total number
180 static struct rpcrdma_mr_seg *
181 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
184 u32 remaining, page_offset;
187 base = vec->iov_base;
188 page_offset = offset_in_page(base);
189 remaining = vec->iov_len;
192 seg->mr_offset = base;
193 seg->mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
194 remaining -= seg->mr_len;
203 /* Convert @xdrbuf into SGEs no larger than a page each. As they
204 * are registered, these SGEs are then coalesced into RDMA segments
205 * when the selected memreg mode supports it.
207 * Returns positive number of SGEs consumed, or a negative errno.
211 rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
212 unsigned int pos, enum rpcrdma_chunktype type,
213 struct rpcrdma_mr_seg *seg)
215 unsigned long page_base;
217 struct page **ppages;
221 seg = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, &n);
223 len = xdrbuf->page_len;
224 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
225 page_base = offset_in_page(xdrbuf->page_base);
227 if (unlikely(!*ppages)) {
228 /* XXX: Certain upper layer operations do
229 * not provide receive buffer pages.
231 *ppages = alloc_page(GFP_ATOMIC);
235 seg->mr_page = *ppages;
236 seg->mr_offset = (char *)page_base;
237 seg->mr_len = min_t(u32, PAGE_SIZE - page_base, len);
245 /* When encoding a Read chunk, the tail iovec contains an
246 * XDR pad and may be omitted.
248 if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
251 /* When encoding a Write chunk, some servers need to see an
252 * extra segment for non-XDR-aligned Write chunks. The upper
253 * layer provides space in the tail iovec that may be used
256 if (type == rpcrdma_writech && r_xprt->rx_ia.ri_implicit_roundup)
259 if (xdrbuf->tail[0].iov_len)
260 seg = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, &n);
263 if (unlikely(n > RPCRDMA_MAX_SEGS))
269 encode_item_present(struct xdr_stream *xdr)
273 p = xdr_reserve_space(xdr, sizeof(*p));
282 encode_item_not_present(struct xdr_stream *xdr)
286 p = xdr_reserve_space(xdr, sizeof(*p));
295 xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr *mr)
297 *iptr++ = cpu_to_be32(mr->mr_handle);
298 *iptr++ = cpu_to_be32(mr->mr_length);
299 xdr_encode_hyper(iptr, mr->mr_offset);
303 encode_rdma_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr)
307 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
311 xdr_encode_rdma_segment(p, mr);
316 encode_read_segment(struct xdr_stream *xdr, struct rpcrdma_mr *mr,
321 p = xdr_reserve_space(xdr, 6 * sizeof(*p));
325 *p++ = xdr_one; /* Item present */
326 *p++ = cpu_to_be32(position);
327 xdr_encode_rdma_segment(p, mr);
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.
346 rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
347 struct rpc_rqst *rqst, enum rpcrdma_chunktype rtype)
349 struct xdr_stream *xdr = &req->rl_stream;
350 struct rpcrdma_mr_seg *seg;
351 struct rpcrdma_mr *mr;
355 pos = rqst->rq_snd_buf.head[0].iov_len;
356 if (rtype == rpcrdma_areadch)
358 seg = req->rl_segments;
359 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
365 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
369 rpcrdma_mr_push(mr, &req->rl_registered);
371 if (encode_read_segment(xdr, mr, pos) < 0)
374 trace_xprtrdma_read_chunk(rqst->rq_task, pos, mr, nsegs);
375 r_xprt->rx_stats.read_chunk_count++;
376 nsegs -= mr->mr_nents;
382 /* Register and XDR encode the Write list. Supports encoding a list
383 * containing one array of plain segments that belong to a single
386 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
388 * Write chunklist (a list of (one) counted array):
390 * 1 - N - HLOO - HLOO - ... - HLOO - 0
392 * Returns zero on success, or a negative errno if a failure occurred.
393 * @xdr is advanced to the next position in the stream.
395 * Only a single Write chunk is currently supported.
398 rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
399 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
401 struct xdr_stream *xdr = &req->rl_stream;
402 struct rpcrdma_mr_seg *seg;
403 struct rpcrdma_mr *mr;
407 seg = req->rl_segments;
408 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
409 rqst->rq_rcv_buf.head[0].iov_len,
414 if (encode_item_present(xdr) < 0)
416 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
417 if (unlikely(!segcount))
419 /* Actual value encoded below */
423 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
427 rpcrdma_mr_push(mr, &req->rl_registered);
429 if (encode_rdma_segment(xdr, mr) < 0)
432 trace_xprtrdma_write_chunk(rqst->rq_task, mr, nsegs);
433 r_xprt->rx_stats.write_chunk_count++;
434 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
436 nsegs -= mr->mr_nents;
439 /* Update count of segments in this Write chunk */
440 *segcount = cpu_to_be32(nchunks);
445 /* Register and XDR encode the Reply chunk. Supports encoding an array
446 * of plain segments that belong to a single write (reply) chunk.
448 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
450 * Reply chunk (a counted array):
452 * 1 - N - HLOO - HLOO - ... - HLOO
454 * Returns zero on success, or a negative errno if a failure occurred.
455 * @xdr is advanced to the next position in the stream.
458 rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
459 struct rpc_rqst *rqst, enum rpcrdma_chunktype wtype)
461 struct xdr_stream *xdr = &req->rl_stream;
462 struct rpcrdma_mr_seg *seg;
463 struct rpcrdma_mr *mr;
467 seg = req->rl_segments;
468 nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
472 if (encode_item_present(xdr) < 0)
474 segcount = xdr_reserve_space(xdr, sizeof(*segcount));
475 if (unlikely(!segcount))
477 /* Actual value encoded below */
481 seg = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
485 rpcrdma_mr_push(mr, &req->rl_registered);
487 if (encode_rdma_segment(xdr, mr) < 0)
490 trace_xprtrdma_reply_chunk(rqst->rq_task, mr, nsegs);
491 r_xprt->rx_stats.reply_chunk_count++;
492 r_xprt->rx_stats.total_rdma_request += mr->mr_length;
494 nsegs -= mr->mr_nents;
497 /* Update count of segments in the Reply chunk */
498 *segcount = cpu_to_be32(nchunks);
504 * rpcrdma_unmap_sendctx - DMA-unmap Send buffers
505 * @sc: sendctx containing SGEs to unmap
509 rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc)
511 struct rpcrdma_ia *ia = &sc->sc_xprt->rx_ia;
515 /* The first two SGEs contain the transport header and
516 * the inline buffer. These are always left mapped so
517 * they can be cheaply re-used.
519 sge = &sc->sc_sges[2];
520 for (count = sc->sc_unmap_count; count; ++sge, --count)
521 ib_dma_unmap_page(ia->ri_device,
522 sge->addr, sge->length, DMA_TO_DEVICE);
524 if (test_and_clear_bit(RPCRDMA_REQ_F_TX_RESOURCES, &sc->sc_req->rl_flags)) {
525 smp_mb__after_atomic();
526 wake_up_bit(&sc->sc_req->rl_flags, RPCRDMA_REQ_F_TX_RESOURCES);
530 /* Prepare an SGE for the RPC-over-RDMA transport header.
533 rpcrdma_prepare_hdr_sge(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
536 struct rpcrdma_sendctx *sc = req->rl_sendctx;
537 struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
538 struct ib_sge *sge = sc->sc_sges;
540 if (!rpcrdma_dma_map_regbuf(ia, rb))
542 sge->addr = rdmab_addr(rb);
544 sge->lkey = rdmab_lkey(rb);
546 ib_dma_sync_single_for_device(rdmab_device(rb), sge->addr,
547 sge->length, DMA_TO_DEVICE);
552 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
556 /* Prepare the Send SGEs. The head and tail iovec, and each entry
557 * in the page list, gets its own SGE.
560 rpcrdma_prepare_msg_sges(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
561 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
563 struct rpcrdma_sendctx *sc = req->rl_sendctx;
564 unsigned int sge_no, page_base, len, remaining;
565 struct rpcrdma_regbuf *rb = req->rl_sendbuf;
566 struct ib_device *device = ia->ri_device;
567 struct ib_sge *sge = sc->sc_sges;
568 u32 lkey = ia->ri_pd->local_dma_lkey;
569 struct page *page, **ppages;
571 /* The head iovec is straightforward, as it is already
572 * DMA-mapped. Sync the content that has changed.
574 if (!rpcrdma_dma_map_regbuf(ia, rb))
577 sge[sge_no].addr = rdmab_addr(rb);
578 sge[sge_no].length = xdr->head[0].iov_len;
579 sge[sge_no].lkey = rdmab_lkey(rb);
580 ib_dma_sync_single_for_device(rdmab_device(rb), sge[sge_no].addr,
581 sge[sge_no].length, DMA_TO_DEVICE);
583 /* If there is a Read chunk, the page list is being handled
584 * via explicit RDMA, and thus is skipped here. However, the
585 * tail iovec may include an XDR pad for the page list, as
586 * well as additional content, and may not reside in the
587 * same page as the head iovec.
589 if (rtype == rpcrdma_readch) {
590 len = xdr->tail[0].iov_len;
592 /* Do not include the tail if it is only an XDR pad */
596 page = virt_to_page(xdr->tail[0].iov_base);
597 page_base = offset_in_page(xdr->tail[0].iov_base);
599 /* If the content in the page list is an odd length,
600 * xdr_write_pages() has added a pad at the beginning
601 * of the tail iovec. Force the tail's non-pad content
602 * to land at the next XDR position in the Send message.
604 page_base += len & 3;
609 /* If there is a page list present, temporarily DMA map
610 * and prepare an SGE for each page to be sent.
613 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
614 page_base = offset_in_page(xdr->page_base);
615 remaining = xdr->page_len;
618 if (sge_no > RPCRDMA_MAX_SEND_SGES - 2)
619 goto out_mapping_overflow;
621 len = min_t(u32, PAGE_SIZE - page_base, remaining);
622 sge[sge_no].addr = ib_dma_map_page(device, *ppages,
625 if (ib_dma_mapping_error(device, sge[sge_no].addr))
626 goto out_mapping_err;
627 sge[sge_no].length = len;
628 sge[sge_no].lkey = lkey;
630 sc->sc_unmap_count++;
637 /* The tail iovec is not always constructed in the same
638 * page where the head iovec resides (see, for example,
639 * gss_wrap_req_priv). To neatly accommodate that case,
640 * DMA map it separately.
642 if (xdr->tail[0].iov_len) {
643 page = virt_to_page(xdr->tail[0].iov_base);
644 page_base = offset_in_page(xdr->tail[0].iov_base);
645 len = xdr->tail[0].iov_len;
649 sge[sge_no].addr = ib_dma_map_page(device, page,
652 if (ib_dma_mapping_error(device, sge[sge_no].addr))
653 goto out_mapping_err;
654 sge[sge_no].length = len;
655 sge[sge_no].lkey = lkey;
656 sc->sc_unmap_count++;
660 sc->sc_wr.num_sge += sge_no;
661 if (sc->sc_unmap_count)
662 __set_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags);
666 pr_err("rpcrdma: failed to DMA map a Send buffer\n");
669 out_mapping_overflow:
670 rpcrdma_unmap_sendctx(sc);
671 pr_err("rpcrdma: too many Send SGEs (%u)\n", sge_no);
675 rpcrdma_unmap_sendctx(sc);
676 pr_err("rpcrdma: Send mapping error\n");
681 * rpcrdma_prepare_send_sges - Construct SGEs for a Send WR
682 * @r_xprt: controlling transport
683 * @req: context of RPC Call being marshalled
684 * @hdrlen: size of transport header, in bytes
685 * @xdr: xdr_buf containing RPC Call
686 * @rtype: chunk type being encoded
688 * Returns 0 on success; otherwise a negative errno is returned.
691 rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
692 struct rpcrdma_req *req, u32 hdrlen,
693 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
695 req->rl_sendctx = rpcrdma_sendctx_get_locked(&r_xprt->rx_buf);
696 if (!req->rl_sendctx)
698 req->rl_sendctx->sc_wr.num_sge = 0;
699 req->rl_sendctx->sc_unmap_count = 0;
700 req->rl_sendctx->sc_req = req;
701 __clear_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags);
703 if (!rpcrdma_prepare_hdr_sge(&r_xprt->rx_ia, req, hdrlen))
706 if (rtype != rpcrdma_areadch)
707 if (!rpcrdma_prepare_msg_sges(&r_xprt->rx_ia, req, xdr, rtype))
714 * rpcrdma_marshal_req - Marshal and send one RPC request
715 * @r_xprt: controlling transport
716 * @rqst: RPC request to be marshaled
718 * For the RPC in "rqst", this function:
719 * - Chooses the transfer mode (eg., RDMA_MSG or RDMA_NOMSG)
720 * - Registers Read, Write, and Reply chunks
721 * - Constructs the transport header
722 * - Posts a Send WR to send the transport header and request
725 * %0 if the RPC was sent successfully,
726 * %-ENOTCONN if the connection was lost,
727 * %-EAGAIN if not enough pages are available for on-demand reply buffer,
728 * %-ENOBUFS if no MRs are available to register chunks,
729 * %-EMSGSIZE if the transport header is too small,
730 * %-EIO if a permanent problem occurred while marshaling.
733 rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
735 struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
736 struct xdr_stream *xdr = &req->rl_stream;
737 enum rpcrdma_chunktype rtype, wtype;
742 rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
743 xdr_init_encode(xdr, &req->rl_hdrbuf,
744 req->rl_rdmabuf->rg_base);
746 /* Fixed header fields */
748 p = xdr_reserve_space(xdr, 4 * sizeof(*p));
752 *p++ = rpcrdma_version;
753 *p++ = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
755 /* When the ULP employs a GSS flavor that guarantees integrity
756 * or privacy, direct data placement of individual data items
759 ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
760 RPCAUTH_AUTH_DATATOUCH);
763 * Chunks needed for results?
765 * o If the expected result is under the inline threshold, all ops
767 * o Large read ops return data as write chunk(s), header as
769 * o Large non-read ops return as a single reply chunk.
771 if (rpcrdma_results_inline(r_xprt, rqst))
772 wtype = rpcrdma_noch;
773 else if (ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ)
774 wtype = rpcrdma_writech;
776 wtype = rpcrdma_replych;
779 * Chunks needed for arguments?
781 * o If the total request is under the inline threshold, all ops
782 * are sent as inline.
783 * o Large write ops transmit data as read chunk(s), header as
785 * o Large non-write ops are sent with the entire message as a
786 * single read chunk (protocol 0-position special case).
788 * This assumes that the upper layer does not present a request
789 * that both has a data payload, and whose non-data arguments
790 * by themselves are larger than the inline threshold.
792 if (rpcrdma_args_inline(r_xprt, rqst)) {
794 rtype = rpcrdma_noch;
795 } else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
797 rtype = rpcrdma_readch;
799 r_xprt->rx_stats.nomsg_call_count++;
801 rtype = rpcrdma_areadch;
804 /* If this is a retransmit, discard previously registered
805 * chunks. Very likely the connection has been replaced,
806 * so these registrations are invalid and unusable.
808 while (unlikely(!list_empty(&req->rl_registered))) {
809 struct rpcrdma_mr *mr;
811 mr = rpcrdma_mr_pop(&req->rl_registered);
812 rpcrdma_mr_defer_recovery(mr);
815 /* This implementation supports the following combinations
816 * of chunk lists in one RPC-over-RDMA Call message:
821 * - Read list + Reply chunk
823 * It might not yet support the following combinations:
825 * - Read list + Write list
827 * It does not support the following combinations:
829 * - Write list + Reply chunk
830 * - Read list + Write list + Reply chunk
832 * This implementation supports only a single chunk in each
833 * Read or Write list. Thus for example the client cannot
834 * send a Call message with a Position Zero Read chunk and a
835 * regular Read chunk at the same time.
837 if (rtype != rpcrdma_noch) {
838 ret = rpcrdma_encode_read_list(r_xprt, req, rqst, rtype);
842 ret = encode_item_not_present(xdr);
846 if (wtype == rpcrdma_writech) {
847 ret = rpcrdma_encode_write_list(r_xprt, req, rqst, wtype);
851 ret = encode_item_not_present(xdr);
855 if (wtype != rpcrdma_replych)
856 ret = encode_item_not_present(xdr);
858 ret = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, wtype);
862 trace_xprtrdma_marshal(rqst, xdr_stream_pos(xdr), rtype, wtype);
864 ret = rpcrdma_prepare_send_sges(r_xprt, req, xdr_stream_pos(xdr),
865 &rqst->rq_snd_buf, rtype);
871 if (ret != -ENOBUFS) {
872 pr_err("rpcrdma: header marshaling failed (%d)\n", ret);
873 r_xprt->rx_stats.failed_marshal_count++;
879 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
880 * @rqst: controlling RPC request
881 * @srcp: points to RPC message payload in receive buffer
882 * @copy_len: remaining length of receive buffer content
883 * @pad: Write chunk pad bytes needed (zero for pure inline)
885 * The upper layer has set the maximum number of bytes it can
886 * receive in each component of rq_rcv_buf. These values are set in
887 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
889 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
890 * many cases this function simply updates iov_base pointers in
891 * rq_rcv_buf to point directly to the received reply data, to
892 * avoid copying reply data.
894 * Returns the count of bytes which had to be memcopied.
897 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
899 unsigned long fixup_copy_count;
900 int i, npages, curlen;
902 struct page **ppages;
905 /* The head iovec is redirected to the RPC reply message
906 * in the receive buffer, to avoid a memcopy.
908 rqst->rq_rcv_buf.head[0].iov_base = srcp;
909 rqst->rq_private_buf.head[0].iov_base = srcp;
911 /* The contents of the receive buffer that follow
912 * head.iov_len bytes are copied into the page list.
914 curlen = rqst->rq_rcv_buf.head[0].iov_len;
915 if (curlen > copy_len)
917 trace_xprtrdma_fixup(rqst, copy_len, curlen);
921 ppages = rqst->rq_rcv_buf.pages +
922 (rqst->rq_rcv_buf.page_base >> PAGE_SHIFT);
923 page_base = offset_in_page(rqst->rq_rcv_buf.page_base);
924 fixup_copy_count = 0;
925 if (copy_len && rqst->rq_rcv_buf.page_len) {
928 pagelist_len = rqst->rq_rcv_buf.page_len;
929 if (pagelist_len > copy_len)
930 pagelist_len = copy_len;
931 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
932 for (i = 0; i < npages; i++) {
933 curlen = PAGE_SIZE - page_base;
934 if (curlen > pagelist_len)
935 curlen = pagelist_len;
937 trace_xprtrdma_fixup_pg(rqst, i, srcp,
939 destp = kmap_atomic(ppages[i]);
940 memcpy(destp + page_base, srcp, curlen);
941 flush_dcache_page(ppages[i]);
942 kunmap_atomic(destp);
945 fixup_copy_count += curlen;
946 pagelist_len -= curlen;
952 /* Implicit padding for the last segment in a Write
953 * chunk is inserted inline at the front of the tail
954 * iovec. The upper layer ignores the content of
955 * the pad. Simply ensure inline content in the tail
956 * that follows the Write chunk is properly aligned.
962 /* The tail iovec is redirected to the remaining data
963 * in the receive buffer, to avoid a memcopy.
965 if (copy_len || pad) {
966 rqst->rq_rcv_buf.tail[0].iov_base = srcp;
967 rqst->rq_private_buf.tail[0].iov_base = srcp;
970 return fixup_copy_count;
973 /* By convention, backchannel calls arrive via rdma_msg type
974 * messages, and never populate the chunk lists. This makes
975 * the RPC/RDMA header small and fixed in size, so it is
976 * straightforward to check the RPC header's direction field.
979 rpcrdma_is_bcall(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
980 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
982 struct xdr_stream *xdr = &rep->rr_stream;
985 if (rep->rr_proc != rdma_msg)
988 /* Peek at stream contents without advancing. */
989 p = xdr_inline_decode(xdr, 0);
992 if (*p++ != xdr_zero)
994 if (*p++ != xdr_zero)
996 if (*p++ != xdr_zero)
1000 if (*p++ != rep->rr_xid)
1002 if (*p != cpu_to_be32(RPC_CALL))
1005 /* Now that we are sure this is a backchannel call,
1006 * advance to the RPC header.
1008 p = xdr_inline_decode(xdr, 3 * sizeof(*p));
1012 rpcrdma_bc_receive_call(r_xprt, rep);
1016 pr_warn("RPC/RDMA short backward direction call\n");
1017 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep))
1018 xprt_disconnect_done(&r_xprt->rx_xprt);
1021 #else /* CONFIG_SUNRPC_BACKCHANNEL */
1025 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1027 static int decode_rdma_segment(struct xdr_stream *xdr, u32 *length)
1033 p = xdr_inline_decode(xdr, 4 * sizeof(*p));
1037 handle = be32_to_cpup(p++);
1038 *length = be32_to_cpup(p++);
1039 xdr_decode_hyper(p, &offset);
1041 trace_xprtrdma_decode_seg(handle, *length, offset);
1045 static int decode_write_chunk(struct xdr_stream *xdr, u32 *length)
1047 u32 segcount, seglength;
1050 p = xdr_inline_decode(xdr, sizeof(*p));
1055 segcount = be32_to_cpup(p);
1056 while (segcount--) {
1057 if (decode_rdma_segment(xdr, &seglength))
1059 *length += seglength;
1065 /* In RPC-over-RDMA Version One replies, a Read list is never
1066 * expected. This decoder is a stub that returns an error if
1067 * a Read list is present.
1069 static int decode_read_list(struct xdr_stream *xdr)
1073 p = xdr_inline_decode(xdr, sizeof(*p));
1076 if (unlikely(*p != xdr_zero))
1081 /* Supports only one Write chunk in the Write list
1083 static int decode_write_list(struct xdr_stream *xdr, u32 *length)
1092 p = xdr_inline_decode(xdr, sizeof(*p));
1100 if (decode_write_chunk(xdr, &chunklen))
1102 *length += chunklen;
1108 static int decode_reply_chunk(struct xdr_stream *xdr, u32 *length)
1112 p = xdr_inline_decode(xdr, sizeof(*p));
1118 if (decode_write_chunk(xdr, length))
1124 rpcrdma_decode_msg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1125 struct rpc_rqst *rqst)
1127 struct xdr_stream *xdr = &rep->rr_stream;
1128 u32 writelist, replychunk, rpclen;
1131 /* Decode the chunk lists */
1132 if (decode_read_list(xdr))
1134 if (decode_write_list(xdr, &writelist))
1136 if (decode_reply_chunk(xdr, &replychunk))
1139 /* RDMA_MSG sanity checks */
1140 if (unlikely(replychunk))
1143 /* Build the RPC reply's Payload stream in rqst->rq_rcv_buf */
1144 base = (char *)xdr_inline_decode(xdr, 0);
1145 rpclen = xdr_stream_remaining(xdr);
1146 r_xprt->rx_stats.fixup_copy_count +=
1147 rpcrdma_inline_fixup(rqst, base, rpclen, writelist & 3);
1149 r_xprt->rx_stats.total_rdma_reply += writelist;
1150 return rpclen + xdr_align_size(writelist);
1154 rpcrdma_decode_nomsg(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep)
1156 struct xdr_stream *xdr = &rep->rr_stream;
1157 u32 writelist, replychunk;
1159 /* Decode the chunk lists */
1160 if (decode_read_list(xdr))
1162 if (decode_write_list(xdr, &writelist))
1164 if (decode_reply_chunk(xdr, &replychunk))
1167 /* RDMA_NOMSG sanity checks */
1168 if (unlikely(writelist))
1170 if (unlikely(!replychunk))
1173 /* Reply chunk buffer already is the reply vector */
1174 r_xprt->rx_stats.total_rdma_reply += replychunk;
1179 rpcrdma_decode_error(struct rpcrdma_xprt *r_xprt, struct rpcrdma_rep *rep,
1180 struct rpc_rqst *rqst)
1182 struct xdr_stream *xdr = &rep->rr_stream;
1185 p = xdr_inline_decode(xdr, sizeof(*p));
1191 p = xdr_inline_decode(xdr, 2 * sizeof(*p));
1194 dprintk("RPC: %5u: %s: server reports version error (%u-%u)\n",
1195 rqst->rq_task->tk_pid, __func__,
1196 be32_to_cpup(p), be32_to_cpu(*(p + 1)));
1199 dprintk("RPC: %5u: %s: server reports header decoding error\n",
1200 rqst->rq_task->tk_pid, __func__);
1203 dprintk("RPC: %5u: %s: server reports unrecognized error %d\n",
1204 rqst->rq_task->tk_pid, __func__, be32_to_cpup(p));
1207 r_xprt->rx_stats.bad_reply_count++;
1211 /* Perform XID lookup, reconstruction of the RPC reply, and
1212 * RPC completion while holding the transport lock to ensure
1213 * the rep, rqst, and rq_task pointers remain stable.
1215 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep)
1217 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1218 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1219 struct rpc_rqst *rqst = rep->rr_rqst;
1223 xprt->reestablish_timeout = 0;
1225 switch (rep->rr_proc) {
1227 status = rpcrdma_decode_msg(r_xprt, rep, rqst);
1230 status = rpcrdma_decode_nomsg(r_xprt, rep);
1233 status = rpcrdma_decode_error(r_xprt, rep, rqst);
1242 spin_lock(&xprt->recv_lock);
1244 xprt->cwnd = r_xprt->rx_buf.rb_credits << RPC_CWNDSHIFT;
1245 if (xprt->cwnd > cwnd)
1246 xprt_release_rqst_cong(rqst->rq_task);
1248 xprt_complete_rqst(rqst->rq_task, status);
1249 xprt_unpin_rqst(rqst);
1250 spin_unlock(&xprt->recv_lock);
1253 /* If the incoming reply terminated a pending RPC, the next
1254 * RPC call will post a replacement receive buffer as it is
1258 trace_xprtrdma_reply_hdr(rep);
1259 r_xprt->rx_stats.bad_reply_count++;
1264 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
1266 /* Invalidate and unmap the data payloads before waking
1267 * the waiting application. This guarantees the memory
1268 * regions are properly fenced from the server before the
1269 * application accesses the data. It also ensures proper
1270 * send flow control: waking the next RPC waits until this
1271 * RPC has relinquished all its Send Queue entries.
1273 if (!list_empty(&req->rl_registered))
1274 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
1275 &req->rl_registered);
1277 /* Ensure that any DMA mapped pages associated with
1278 * the Send of the RPC Call have been unmapped before
1279 * allowing the RPC to complete. This protects argument
1280 * memory not controlled by the RPC client from being
1281 * re-used before we're done with it.
1283 if (test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1284 r_xprt->rx_stats.reply_waits_for_send++;
1285 out_of_line_wait_on_bit(&req->rl_flags,
1286 RPCRDMA_REQ_F_TX_RESOURCES,
1288 TASK_UNINTERRUPTIBLE);
1292 /* Reply handling runs in the poll worker thread. Anything that
1293 * might wait is deferred to a separate workqueue.
1295 void rpcrdma_deferred_completion(struct work_struct *work)
1297 struct rpcrdma_rep *rep =
1298 container_of(work, struct rpcrdma_rep, rr_work);
1299 struct rpcrdma_req *req = rpcr_to_rdmar(rep->rr_rqst);
1300 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1302 trace_xprtrdma_defer_cmp(rep);
1303 if (rep->rr_wc_flags & IB_WC_WITH_INVALIDATE)
1304 r_xprt->rx_ia.ri_ops->ro_reminv(rep, &req->rl_registered);
1305 rpcrdma_release_rqst(r_xprt, req);
1306 rpcrdma_complete_rqst(rep);
1309 /* Process received RPC/RDMA messages.
1311 * Errors must result in the RPC task either being awakened, or
1312 * allowed to timeout, to discover the errors at that time.
1314 void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
1316 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
1317 struct rpc_xprt *xprt = &r_xprt->rx_xprt;
1318 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1319 struct rpcrdma_req *req;
1320 struct rpc_rqst *rqst;
1324 if (rep->rr_hdrbuf.head[0].iov_len == 0)
1327 xdr_init_decode(&rep->rr_stream, &rep->rr_hdrbuf,
1328 rep->rr_hdrbuf.head[0].iov_base);
1330 /* Fixed transport header fields */
1331 p = xdr_inline_decode(&rep->rr_stream, 4 * sizeof(*p));
1333 goto out_shortreply;
1335 rep->rr_vers = *p++;
1336 credits = be32_to_cpu(*p++);
1337 rep->rr_proc = *p++;
1339 if (rep->rr_vers != rpcrdma_version)
1340 goto out_badversion;
1342 if (rpcrdma_is_bcall(r_xprt, rep))
1345 /* Match incoming rpcrdma_rep to an rpcrdma_req to
1346 * get context for handling any incoming chunks.
1348 spin_lock(&xprt->recv_lock);
1349 rqst = xprt_lookup_rqst(xprt, rep->rr_xid);
1352 xprt_pin_rqst(rqst);
1355 credits = 1; /* don't deadlock */
1356 else if (credits > buf->rb_max_requests)
1357 credits = buf->rb_max_requests;
1358 buf->rb_credits = credits;
1360 spin_unlock(&xprt->recv_lock);
1362 req = rpcr_to_rdmar(rqst);
1363 req->rl_reply = rep;
1364 rep->rr_rqst = rqst;
1365 clear_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
1367 trace_xprtrdma_reply(rqst->rq_task, rep, req, credits);
1369 queue_work_on(req->rl_cpu, rpcrdma_receive_wq, &rep->rr_work);
1373 rpcrdma_recv_buffer_put(rep);
1374 if (r_xprt->rx_ep.rep_connected == 1) {
1375 r_xprt->rx_ep.rep_connected = -EIO;
1376 rpcrdma_conn_func(&r_xprt->rx_ep);
1381 trace_xprtrdma_reply_vers(rep);
1384 /* The RPC transaction has already been terminated, or the header
1388 spin_unlock(&xprt->recv_lock);
1389 trace_xprtrdma_reply_rqst(rep);
1393 trace_xprtrdma_reply_short(rep);
1395 /* If no pending RPC transaction was matched, post a replacement
1396 * receive buffer before returning.
1399 r_xprt->rx_stats.bad_reply_count++;
1400 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep))
1401 rpcrdma_recv_buffer_put(rep);