Linux Audio

Check our new training course

Loading...
v4.10.11
   1/*
   2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the BSD-type
   8 * license below:
   9 *
  10 * Redistribution and use in source and binary forms, with or without
  11 * modification, are permitted provided that the following conditions
  12 * are met:
  13 *
  14 *      Redistributions of source code must retain the above copyright
  15 *      notice, this list of conditions and the following disclaimer.
  16 *
  17 *      Redistributions in binary form must reproduce the above
  18 *      copyright notice, this list of conditions and the following
  19 *      disclaimer in the documentation and/or other materials provided
  20 *      with the distribution.
  21 *
  22 *      Neither the name of the Network Appliance, Inc. nor the names of
  23 *      its contributors may be used to endorse or promote products
  24 *      derived from this software without specific prior written
  25 *      permission.
  26 *
  27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  38 */
  39
  40/*
  41 * rpc_rdma.c
  42 *
  43 * This file contains the guts of the RPC RDMA protocol, and
  44 * does marshaling/unmarshaling, etc. It is also where interfacing
  45 * to the Linux RPC framework lives.
  46 */
  47
  48#include "xprt_rdma.h"
  49
  50#include <linux/highmem.h>
  51
  52#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  53# define RPCDBG_FACILITY	RPCDBG_TRANS
  54#endif
  55
 
 
 
 
 
 
 
 
 
  56static const char transfertypes[][12] = {
  57	"inline",	/* no chunks */
  58	"read list",	/* some argument via rdma read */
  59	"*read list",	/* entire request via rdma read */
  60	"write list",	/* some result via rdma write */
  61	"reply chunk"	/* entire reply via rdma write */
  62};
  63
  64/* Returns size of largest RPC-over-RDMA header in a Call message
  65 *
  66 * The largest Call header contains a full-size Read list and a
  67 * minimal Reply chunk.
  68 */
  69static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs)
  70{
  71	unsigned int size;
  72
  73	/* Fixed header fields and list discriminators */
  74	size = RPCRDMA_HDRLEN_MIN;
  75
  76	/* Maximum Read list size */
  77	maxsegs += 2;	/* segment for head and tail buffers */
  78	size = maxsegs * sizeof(struct rpcrdma_read_chunk);
  79
  80	/* Minimal Read chunk size */
  81	size += sizeof(__be32);	/* segment count */
  82	size += sizeof(struct rpcrdma_segment);
  83	size += sizeof(__be32);	/* list discriminator */
  84
  85	dprintk("RPC:       %s: max call header size = %u\n",
  86		__func__, size);
  87	return size;
  88}
  89
  90/* Returns size of largest RPC-over-RDMA header in a Reply message
  91 *
  92 * There is only one Write list or one Reply chunk per Reply
  93 * message.  The larger list is the Write list.
  94 */
  95static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs)
  96{
  97	unsigned int size;
  98
  99	/* Fixed header fields and list discriminators */
 100	size = RPCRDMA_HDRLEN_MIN;
 101
 102	/* Maximum Write list size */
 103	maxsegs += 2;	/* segment for head and tail buffers */
 104	size = sizeof(__be32);		/* segment count */
 105	size += maxsegs * sizeof(struct rpcrdma_segment);
 106	size += sizeof(__be32);	/* list discriminator */
 107
 108	dprintk("RPC:       %s: max reply header size = %u\n",
 109		__func__, size);
 110	return size;
 111}
 112
 113void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *r_xprt)
 114{
 115	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 116	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 117	unsigned int maxsegs = ia->ri_max_segs;
 118
 119	ia->ri_max_inline_write = cdata->inline_wsize -
 120				  rpcrdma_max_call_header_size(maxsegs);
 121	ia->ri_max_inline_read = cdata->inline_rsize -
 122				 rpcrdma_max_reply_header_size(maxsegs);
 123}
 124
 125/* The client can send a request inline as long as the RPCRDMA header
 126 * plus the RPC call fit under the transport's inline limit. If the
 127 * combined call message size exceeds that limit, the client must use
 128 * a Read chunk for this operation.
 129 *
 130 * A Read chunk is also required if sending the RPC call inline would
 131 * exceed this device's max_sge limit.
 132 */
 133static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt,
 134				struct rpc_rqst *rqst)
 135{
 136	struct xdr_buf *xdr = &rqst->rq_snd_buf;
 137	unsigned int count, remaining, offset;
 138
 139	if (xdr->len > r_xprt->rx_ia.ri_max_inline_write)
 140		return false;
 141
 142	if (xdr->page_len) {
 143		remaining = xdr->page_len;
 144		offset = xdr->page_base & ~PAGE_MASK;
 145		count = 0;
 146		while (remaining) {
 147			remaining -= min_t(unsigned int,
 148					   PAGE_SIZE - offset, remaining);
 149			offset = 0;
 150			if (++count > r_xprt->rx_ia.ri_max_send_sges)
 151				return false;
 152		}
 153	}
 154
 155	return true;
 156}
 157
 158/* The client can't know how large the actual reply will be. Thus it
 159 * plans for the largest possible reply for that particular ULP
 160 * operation. If the maximum combined reply message size exceeds that
 161 * limit, the client must provide a write list or a reply chunk for
 162 * this request.
 163 */
 164static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
 165				   struct rpc_rqst *rqst)
 166{
 167	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 168
 169	return rqst->rq_rcv_buf.buflen <= ia->ri_max_inline_read;
 170}
 171
 172/* Split "vec" on page boundaries into segments. FMR registers pages,
 173 * not a byte range. Other modes coalesce these segments into a single
 174 * MR when they can.
 175 */
 176static int
 177rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg, int n)
 178{
 179	size_t page_offset;
 180	u32 remaining;
 181	char *base;
 182
 183	base = vec->iov_base;
 184	page_offset = offset_in_page(base);
 185	remaining = vec->iov_len;
 186	while (remaining && n < RPCRDMA_MAX_SEGS) {
 187		seg[n].mr_page = NULL;
 188		seg[n].mr_offset = base;
 189		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
 190		remaining -= seg[n].mr_len;
 191		base += seg[n].mr_len;
 192		++n;
 193		page_offset = 0;
 194	}
 195	return n;
 196}
 197
 198/*
 199 * Chunk assembly from upper layer xdr_buf.
 200 *
 201 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
 202 * elements. Segments are then coalesced when registered, if possible
 203 * within the selected memreg mode.
 204 *
 205 * Returns positive number of segments converted, or a negative errno.
 
 206 */
 207
 208static int
 209rpcrdma_convert_iovs(struct rpcrdma_xprt *r_xprt, struct xdr_buf *xdrbuf,
 210		     unsigned int pos, enum rpcrdma_chunktype type,
 211		     struct rpcrdma_mr_seg *seg)
 212{
 213	int len, n, p, page_base;
 
 214	struct page **ppages;
 215
 216	n = 0;
 217	if (pos == 0) {
 218		n = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, n);
 219		if (n == RPCRDMA_MAX_SEGS)
 220			goto out_overflow;
 221	}
 222
 223	len = xdrbuf->page_len;
 224	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
 225	page_base = xdrbuf->page_base & ~PAGE_MASK;
 226	p = 0;
 227	while (len && n < RPCRDMA_MAX_SEGS) {
 228		if (!ppages[p]) {
 229			/* alloc the pagelist for receiving buffer */
 230			ppages[p] = alloc_page(GFP_ATOMIC);
 231			if (!ppages[p])
 232				return -EAGAIN;
 233		}
 234		seg[n].mr_page = ppages[p];
 235		seg[n].mr_offset = (void *)(unsigned long) page_base;
 236		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
 237		if (seg[n].mr_len > PAGE_SIZE)
 238			goto out_overflow;
 239		len -= seg[n].mr_len;
 240		++n;
 241		++p;
 242		page_base = 0;	/* page offset only applies to first page */
 243	}
 244
 245	/* Message overflows the seg array */
 246	if (len && n == RPCRDMA_MAX_SEGS)
 247		goto out_overflow;
 248
 249	/* When encoding a Read chunk, the tail iovec contains an
 250	 * XDR pad and may be omitted.
 251	 */
 252	if (type == rpcrdma_readch && r_xprt->rx_ia.ri_implicit_roundup)
 253		return n;
 254
 255	/* When encoding a Write chunk, some servers need to see an
 256	 * extra segment for non-XDR-aligned Write chunks. The upper
 257	 * layer provides space in the tail iovec that may be used
 258	 * for this purpose.
 259	 */
 260	if (type == rpcrdma_writech && r_xprt->rx_ia.ri_implicit_roundup)
 261		return n;
 262
 263	if (xdrbuf->tail[0].iov_len) {
 264		n = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, n);
 265		if (n == RPCRDMA_MAX_SEGS)
 266			goto out_overflow;
 
 
 
 
 
 
 
 
 267	}
 268
 269	return n;
 270
 271out_overflow:
 272	pr_err("rpcrdma: segment array overflow\n");
 273	return -EIO;
 274}
 275
 276static inline __be32 *
 277xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mw *mw)
 278{
 279	*iptr++ = cpu_to_be32(mw->mw_handle);
 280	*iptr++ = cpu_to_be32(mw->mw_length);
 281	return xdr_encode_hyper(iptr, mw->mw_offset);
 282}
 283
 284/* XDR-encode the Read list. Supports encoding a list of read
 285 * segments that belong to a single read chunk.
 
 
 
 
 
 
 
 286 *
 287 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 288 *
 289 *  Read chunklist (a linked list):
 290 *   N elements, position P (same P for all chunks of same arg!):
 291 *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
 292 *
 293 * Returns a pointer to the XDR word in the RDMA header following
 294 * the end of the Read list, or an error pointer.
 295 */
 296static __be32 *
 297rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
 298			 struct rpcrdma_req *req, struct rpc_rqst *rqst,
 299			 __be32 *iptr, enum rpcrdma_chunktype rtype)
 300{
 301	struct rpcrdma_mr_seg *seg;
 302	struct rpcrdma_mw *mw;
 303	unsigned int pos;
 304	int n, nsegs;
 305
 306	if (rtype == rpcrdma_noch) {
 307		*iptr++ = xdr_zero;	/* item not present */
 308		return iptr;
 309	}
 310
 311	pos = rqst->rq_snd_buf.head[0].iov_len;
 312	if (rtype == rpcrdma_areadch)
 313		pos = 0;
 314	seg = req->rl_segments;
 315	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_snd_buf, pos,
 316				     rtype, seg);
 317	if (nsegs < 0)
 318		return ERR_PTR(nsegs);
 319
 320	do {
 321		n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
 322						 false, &mw);
 323		if (n < 0)
 324			return ERR_PTR(n);
 325		list_add(&mw->mw_list, &req->rl_registered);
 326
 327		*iptr++ = xdr_one;	/* item present */
 328
 329		/* All read segments in this chunk
 330		 * have the same "position".
 331		 */
 332		*iptr++ = cpu_to_be32(pos);
 333		iptr = xdr_encode_rdma_segment(iptr, mw);
 334
 335		dprintk("RPC: %5u %s: pos %u %u@0x%016llx:0x%08x (%s)\n",
 336			rqst->rq_task->tk_pid, __func__, pos,
 337			mw->mw_length, (unsigned long long)mw->mw_offset,
 338			mw->mw_handle, n < nsegs ? "more" : "last");
 339
 340		r_xprt->rx_stats.read_chunk_count++;
 341		seg += n;
 342		nsegs -= n;
 343	} while (nsegs);
 344
 345	/* Finish Read list */
 346	*iptr++ = xdr_zero;	/* Next item not present */
 347	return iptr;
 348}
 349
 350/* XDR-encode the Write list. Supports encoding a list containing
 351 * one array of plain segments that belong to a single write chunk.
 352 *
 353 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 354 *
 355 *  Write chunklist (a list of (one) counted array):
 356 *   N elements:
 357 *    1 - N - HLOO - HLOO - ... - HLOO - 0
 358 *
 359 * Returns a pointer to the XDR word in the RDMA header following
 360 * the end of the Write list, or an error pointer.
 361 */
 362static __be32 *
 363rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 364			  struct rpc_rqst *rqst, __be32 *iptr,
 365			  enum rpcrdma_chunktype wtype)
 366{
 367	struct rpcrdma_mr_seg *seg;
 368	struct rpcrdma_mw *mw;
 369	int n, nsegs, nchunks;
 370	__be32 *segcount;
 371
 372	if (wtype != rpcrdma_writech) {
 373		*iptr++ = xdr_zero;	/* no Write list present */
 374		return iptr;
 375	}
 376
 377	seg = req->rl_segments;
 378	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf,
 379				     rqst->rq_rcv_buf.head[0].iov_len,
 380				     wtype, seg);
 381	if (nsegs < 0)
 382		return ERR_PTR(nsegs);
 383
 384	*iptr++ = xdr_one;	/* Write list present */
 385	segcount = iptr++;	/* save location of segment count */
 386
 387	nchunks = 0;
 388	do {
 389		n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
 390						 true, &mw);
 391		if (n < 0)
 392			return ERR_PTR(n);
 393		list_add(&mw->mw_list, &req->rl_registered);
 394
 395		iptr = xdr_encode_rdma_segment(iptr, mw);
 396
 397		dprintk("RPC: %5u %s: %u@0x016%llx:0x%08x (%s)\n",
 398			rqst->rq_task->tk_pid, __func__,
 399			mw->mw_length, (unsigned long long)mw->mw_offset,
 400			mw->mw_handle, n < nsegs ? "more" : "last");
 401
 402		r_xprt->rx_stats.write_chunk_count++;
 403		r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 404		nchunks++;
 405		seg   += n;
 406		nsegs -= n;
 407	} while (nsegs);
 408
 409	/* Update count of segments in this Write chunk */
 410	*segcount = cpu_to_be32(nchunks);
 411
 412	/* Finish Write list */
 413	*iptr++ = xdr_zero;	/* Next item not present */
 414	return iptr;
 415}
 416
 417/* XDR-encode the Reply chunk. Supports encoding an array of plain
 418 * segments that belong to a single write (reply) chunk.
 419 *
 420 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 421 *
 422 *  Reply chunk (a counted array):
 423 *   N elements:
 424 *    1 - N - HLOO - HLOO - ... - HLOO
 425 *
 426 * Returns a pointer to the XDR word in the RDMA header following
 427 * the end of the Reply chunk, or an error pointer.
 428 */
 429static __be32 *
 430rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt,
 431			   struct rpcrdma_req *req, struct rpc_rqst *rqst,
 432			   __be32 *iptr, enum rpcrdma_chunktype wtype)
 433{
 434	struct rpcrdma_mr_seg *seg;
 435	struct rpcrdma_mw *mw;
 436	int n, nsegs, nchunks;
 437	__be32 *segcount;
 438
 439	if (wtype != rpcrdma_replych) {
 440		*iptr++ = xdr_zero;	/* no Reply chunk present */
 441		return iptr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 442	}
 443
 444	seg = req->rl_segments;
 445	nsegs = rpcrdma_convert_iovs(r_xprt, &rqst->rq_rcv_buf, 0, wtype, seg);
 446	if (nsegs < 0)
 447		return ERR_PTR(nsegs);
 448
 449	*iptr++ = xdr_one;	/* Reply chunk present */
 450	segcount = iptr++;	/* save location of segment count */
 
 451
 452	nchunks = 0;
 453	do {
 454		n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs,
 455						 true, &mw);
 456		if (n < 0)
 457			return ERR_PTR(n);
 458		list_add(&mw->mw_list, &req->rl_registered);
 459
 460		iptr = xdr_encode_rdma_segment(iptr, mw);
 461
 462		dprintk("RPC: %5u %s: %u@0x%016llx:0x%08x (%s)\n",
 463			rqst->rq_task->tk_pid, __func__,
 464			mw->mw_length, (unsigned long long)mw->mw_offset,
 465			mw->mw_handle, n < nsegs ? "more" : "last");
 466
 467		r_xprt->rx_stats.reply_chunk_count++;
 468		r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 469		nchunks++;
 470		seg   += n;
 471		nsegs -= n;
 472	} while (nsegs);
 473
 474	/* Update count of segments in the Reply chunk */
 475	*segcount = cpu_to_be32(nchunks);
 476
 477	return iptr;
 478}
 479
 480/* Prepare the RPC-over-RDMA header SGE.
 481 */
 482static bool
 483rpcrdma_prepare_hdr_sge(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
 484			u32 len)
 485{
 486	struct rpcrdma_regbuf *rb = req->rl_rdmabuf;
 487	struct ib_sge *sge = &req->rl_send_sge[0];
 488
 489	if (unlikely(!rpcrdma_regbuf_is_mapped(rb))) {
 490		if (!__rpcrdma_dma_map_regbuf(ia, rb))
 491			return false;
 492		sge->addr = rdmab_addr(rb);
 493		sge->lkey = rdmab_lkey(rb);
 494	}
 495	sge->length = len;
 496
 497	ib_dma_sync_single_for_device(ia->ri_device, sge->addr,
 498				      sge->length, DMA_TO_DEVICE);
 499	req->rl_send_wr.num_sge++;
 500	return true;
 501}
 502
 503/* Prepare the Send SGEs. The head and tail iovec, and each entry
 504 * in the page list, gets its own SGE.
 505 */
 506static bool
 507rpcrdma_prepare_msg_sges(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
 508			 struct xdr_buf *xdr, enum rpcrdma_chunktype rtype)
 509{
 510	unsigned int sge_no, page_base, len, remaining;
 511	struct rpcrdma_regbuf *rb = req->rl_sendbuf;
 512	struct ib_device *device = ia->ri_device;
 513	struct ib_sge *sge = req->rl_send_sge;
 514	u32 lkey = ia->ri_pd->local_dma_lkey;
 515	struct page *page, **ppages;
 516
 517	/* The head iovec is straightforward, as it is already
 518	 * DMA-mapped. Sync the content that has changed.
 519	 */
 520	if (!rpcrdma_dma_map_regbuf(ia, rb))
 521		return false;
 522	sge_no = 1;
 523	sge[sge_no].addr = rdmab_addr(rb);
 524	sge[sge_no].length = xdr->head[0].iov_len;
 525	sge[sge_no].lkey = rdmab_lkey(rb);
 526	ib_dma_sync_single_for_device(device, sge[sge_no].addr,
 527				      sge[sge_no].length, DMA_TO_DEVICE);
 528
 529	/* If there is a Read chunk, the page list is being handled
 530	 * via explicit RDMA, and thus is skipped here. However, the
 531	 * tail iovec may include an XDR pad for the page list, as
 532	 * well as additional content, and may not reside in the
 533	 * same page as the head iovec.
 534	 */
 535	if (rtype == rpcrdma_readch) {
 536		len = xdr->tail[0].iov_len;
 537
 538		/* Do not include the tail if it is only an XDR pad */
 539		if (len < 4)
 540			goto out;
 541
 542		page = virt_to_page(xdr->tail[0].iov_base);
 543		page_base = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
 
 544
 545		/* If the content in the page list is an odd length,
 546		 * xdr_write_pages() has added a pad at the beginning
 547		 * of the tail iovec. Force the tail's non-pad content
 548		 * to land at the next XDR position in the Send message.
 549		 */
 550		page_base += len & 3;
 551		len -= len & 3;
 552		goto map_tail;
 553	}
 554
 555	/* If there is a page list present, temporarily DMA map
 556	 * and prepare an SGE for each page to be sent.
 557	 */
 558	if (xdr->page_len) {
 559		ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
 560		page_base = xdr->page_base & ~PAGE_MASK;
 561		remaining = xdr->page_len;
 562		while (remaining) {
 563			sge_no++;
 564			if (sge_no > RPCRDMA_MAX_SEND_SGES - 2)
 565				goto out_mapping_overflow;
 566
 567			len = min_t(u32, PAGE_SIZE - page_base, remaining);
 568			sge[sge_no].addr = ib_dma_map_page(device, *ppages,
 569							   page_base, len,
 570							   DMA_TO_DEVICE);
 571			if (ib_dma_mapping_error(device, sge[sge_no].addr))
 572				goto out_mapping_err;
 573			sge[sge_no].length = len;
 574			sge[sge_no].lkey = lkey;
 575
 576			req->rl_mapped_sges++;
 577			ppages++;
 578			remaining -= len;
 579			page_base = 0;
 580		}
 581	}
 582
 583	/* The tail iovec is not always constructed in the same
 584	 * page where the head iovec resides (see, for example,
 585	 * gss_wrap_req_priv). To neatly accommodate that case,
 586	 * DMA map it separately.
 587	 */
 588	if (xdr->tail[0].iov_len) {
 589		page = virt_to_page(xdr->tail[0].iov_base);
 590		page_base = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
 591		len = xdr->tail[0].iov_len;
 592
 593map_tail:
 594		sge_no++;
 595		sge[sge_no].addr = ib_dma_map_page(device, page,
 596						   page_base, len,
 597						   DMA_TO_DEVICE);
 598		if (ib_dma_mapping_error(device, sge[sge_no].addr))
 599			goto out_mapping_err;
 600		sge[sge_no].length = len;
 601		sge[sge_no].lkey = lkey;
 602		req->rl_mapped_sges++;
 603	}
 604
 605out:
 606	req->rl_send_wr.num_sge = sge_no + 1;
 607	return true;
 608
 609out_mapping_overflow:
 610	pr_err("rpcrdma: too many Send SGEs (%u)\n", sge_no);
 611	return false;
 612
 613out_mapping_err:
 614	pr_err("rpcrdma: Send mapping error\n");
 615	return false;
 616}
 617
 618bool
 619rpcrdma_prepare_send_sges(struct rpcrdma_ia *ia, struct rpcrdma_req *req,
 620			  u32 hdrlen, struct xdr_buf *xdr,
 621			  enum rpcrdma_chunktype rtype)
 
 
 
 
 
 622{
 623	req->rl_send_wr.num_sge = 0;
 624	req->rl_mapped_sges = 0;
 625
 626	if (!rpcrdma_prepare_hdr_sge(ia, req, hdrlen))
 627		goto out_map;
 628
 629	if (rtype != rpcrdma_areadch)
 630		if (!rpcrdma_prepare_msg_sges(ia, req, xdr, rtype))
 631			goto out_map;
 632
 633	return true;
 634
 635out_map:
 636	pr_err("rpcrdma: failed to DMA map a Send buffer\n");
 637	return false;
 638}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 639
 640void
 641rpcrdma_unmap_sges(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
 642{
 643	struct ib_device *device = ia->ri_device;
 644	struct ib_sge *sge;
 645	int count;
 646
 647	sge = &req->rl_send_sge[2];
 648	for (count = req->rl_mapped_sges; count--; sge++)
 649		ib_dma_unmap_page(device, sge->addr, sge->length,
 650				  DMA_TO_DEVICE);
 651	req->rl_mapped_sges = 0;
 
 
 
 
 
 
 
 
 652}
 653
 654/*
 655 * Marshal a request: the primary job of this routine is to choose
 656 * the transfer modes. See comments below.
 657 *
 658 * Returns zero on success, otherwise a negative errno.
 
 
 
 
 
 
 659 */
 660
 661int
 662rpcrdma_marshal_req(struct rpc_rqst *rqst)
 663{
 664	struct rpc_xprt *xprt = rqst->rq_xprt;
 665	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 666	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 
 
 667	enum rpcrdma_chunktype rtype, wtype;
 668	struct rpcrdma_msg *headerp;
 669	bool ddp_allowed;
 670	ssize_t hdrlen;
 671	size_t rpclen;
 672	__be32 *iptr;
 673
 674#if defined(CONFIG_SUNRPC_BACKCHANNEL)
 675	if (test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state))
 676		return rpcrdma_bc_marshal_reply(rqst);
 677#endif
 
 
 678
 679	headerp = rdmab_to_msg(req->rl_rdmabuf);
 680	/* don't byte-swap XID, it's already done in request */
 
 681	headerp->rm_xid = rqst->rq_xid;
 682	headerp->rm_vers = rpcrdma_version;
 683	headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
 684	headerp->rm_type = rdma_msg;
 685
 686	/* When the ULP employs a GSS flavor that guarantees integrity
 687	 * or privacy, direct data placement of individual data items
 688	 * is not allowed.
 689	 */
 690	ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags &
 691						RPCAUTH_AUTH_DATATOUCH);
 692
 693	/*
 694	 * Chunks needed for results?
 695	 *
 696	 * o If the expected result is under the inline threshold, all ops
 697	 *   return as inline.
 698	 * o Large read ops return data as write chunk(s), header as
 699	 *   inline.
 700	 * o Large non-read ops return as a single reply chunk.
 
 
 
 
 701	 */
 702	if (rpcrdma_results_inline(r_xprt, rqst))
 
 
 
 
 
 
 
 
 703		wtype = rpcrdma_noch;
 704	else if (ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ)
 
 
 705		wtype = rpcrdma_writech;
 706	else
 707		wtype = rpcrdma_replych;
 708
 709	/*
 710	 * Chunks needed for arguments?
 711	 *
 712	 * o If the total request is under the inline threshold, all ops
 713	 *   are sent as inline.
 714	 * o Large write ops transmit data as read chunk(s), header as
 715	 *   inline.
 716	 * o Large non-write ops are sent with the entire message as a
 717	 *   single read chunk (protocol 0-position special case).
 
 
 718	 *
 719	 * This assumes that the upper layer does not present a request
 720	 * that both has a data payload, and whose non-data arguments
 721	 * by themselves are larger than the inline threshold.
 722	 */
 723	if (rpcrdma_args_inline(r_xprt, rqst)) {
 724		rtype = rpcrdma_noch;
 725		rpclen = rqst->rq_snd_buf.len;
 726	} else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
 727		rtype = rpcrdma_readch;
 728		rpclen = rqst->rq_snd_buf.head[0].iov_len +
 729			 rqst->rq_snd_buf.tail[0].iov_len;
 730	} else {
 731		r_xprt->rx_stats.nomsg_call_count++;
 732		headerp->rm_type = htonl(RDMA_NOMSG);
 733		rtype = rpcrdma_areadch;
 734		rpclen = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 735	}
 736
 737	/* This implementation supports the following combinations
 738	 * of chunk lists in one RPC-over-RDMA Call message:
 739	 *
 740	 *   - Read list
 741	 *   - Write list
 742	 *   - Reply chunk
 743	 *   - Read list + Reply chunk
 744	 *
 745	 * It might not yet support the following combinations:
 746	 *
 747	 *   - Read list + Write list
 748	 *
 749	 * It does not support the following combinations:
 750	 *
 751	 *   - Write list + Reply chunk
 752	 *   - Read list + Write list + Reply chunk
 753	 *
 754	 * This implementation supports only a single chunk in each
 755	 * Read or Write list. Thus for example the client cannot
 756	 * send a Call message with a Position Zero Read chunk and a
 757	 * regular Read chunk at the same time.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 758	 */
 759	iptr = headerp->rm_body.rm_chunks;
 760	iptr = rpcrdma_encode_read_list(r_xprt, req, rqst, iptr, rtype);
 761	if (IS_ERR(iptr))
 762		goto out_unmap;
 763	iptr = rpcrdma_encode_write_list(r_xprt, req, rqst, iptr, wtype);
 764	if (IS_ERR(iptr))
 765		goto out_unmap;
 766	iptr = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, iptr, wtype);
 767	if (IS_ERR(iptr))
 768		goto out_unmap;
 769	hdrlen = (unsigned char *)iptr - (unsigned char *)headerp;
 770
 771	dprintk("RPC: %5u %s: %s/%s: hdrlen %zd rpclen %zd\n",
 772		rqst->rq_task->tk_pid, __func__,
 773		transfertypes[rtype], transfertypes[wtype],
 774		hdrlen, rpclen);
 775
 776	if (!rpcrdma_prepare_send_sges(&r_xprt->rx_ia, req, hdrlen,
 777				       &rqst->rq_snd_buf, rtype)) {
 778		iptr = ERR_PTR(-EIO);
 779		goto out_unmap;
 780	}
 781	return 0;
 782
 783out_unmap:
 784	r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
 785	return PTR_ERR(iptr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 786}
 787
 788/*
 789 * Chase down a received write or reply chunklist to get length
 790 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 791 */
 792static int
 793rpcrdma_count_chunks(struct rpcrdma_rep *rep, int wrchunk, __be32 **iptrp)
 794{
 795	unsigned int i, total_len;
 796	struct rpcrdma_write_chunk *cur_wchunk;
 797	char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
 798
 799	i = be32_to_cpu(**iptrp);
 
 
 800	cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 801	total_len = 0;
 802	while (i--) {
 803		struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 804		ifdebug(FACILITY) {
 805			u64 off;
 806			xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 807			dprintk("RPC:       %s: chunk %d@0x%016llx:0x%08x\n",
 808				__func__,
 809				be32_to_cpu(seg->rs_length),
 810				(unsigned long long)off,
 811				be32_to_cpu(seg->rs_handle));
 812		}
 813		total_len += be32_to_cpu(seg->rs_length);
 814		++cur_wchunk;
 815	}
 816	/* check and adjust for properly terminated write chunk */
 817	if (wrchunk) {
 818		__be32 *w = (__be32 *) cur_wchunk;
 819		if (*w++ != xdr_zero)
 820			return -1;
 821		cur_wchunk = (struct rpcrdma_write_chunk *) w;
 822	}
 823	if ((char *)cur_wchunk > base + rep->rr_len)
 824		return -1;
 825
 826	*iptrp = (__be32 *) cur_wchunk;
 827	return total_len;
 828}
 829
 830/**
 831 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs
 832 * @rqst: controlling RPC request
 833 * @srcp: points to RPC message payload in receive buffer
 834 * @copy_len: remaining length of receive buffer content
 835 * @pad: Write chunk pad bytes needed (zero for pure inline)
 836 *
 837 * The upper layer has set the maximum number of bytes it can
 838 * receive in each component of rq_rcv_buf. These values are set in
 839 * the head.iov_len, page_len, tail.iov_len, and buflen fields.
 840 *
 841 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in
 842 * many cases this function simply updates iov_base pointers in
 843 * rq_rcv_buf to point directly to the received reply data, to
 844 * avoid copying reply data.
 845 *
 846 * Returns the count of bytes which had to be memcopied.
 847 */
 848static unsigned long
 849rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
 850{
 851	unsigned long fixup_copy_count;
 852	int i, npages, curlen;
 853	char *destp;
 854	struct page **ppages;
 855	int page_base;
 856
 857	/* The head iovec is redirected to the RPC reply message
 858	 * in the receive buffer, to avoid a memcopy.
 859	 */
 860	rqst->rq_rcv_buf.head[0].iov_base = srcp;
 861	rqst->rq_private_buf.head[0].iov_base = srcp;
 862
 863	/* The contents of the receive buffer that follow
 864	 * head.iov_len bytes are copied into the page list.
 865	 */
 866	curlen = rqst->rq_rcv_buf.head[0].iov_len;
 867	if (curlen > copy_len)
 868		curlen = copy_len;
 
 
 
 869	dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 870		__func__, srcp, copy_len, curlen);
 
 
 
 871	srcp += curlen;
 872	copy_len -= curlen;
 873
 
 
 
 874	page_base = rqst->rq_rcv_buf.page_base;
 875	ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
 876	page_base &= ~PAGE_MASK;
 877	fixup_copy_count = 0;
 878	if (copy_len && rqst->rq_rcv_buf.page_len) {
 879		int pagelist_len;
 880
 881		pagelist_len = rqst->rq_rcv_buf.page_len;
 882		if (pagelist_len > copy_len)
 883			pagelist_len = copy_len;
 884		npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT;
 885		for (i = 0; i < npages; i++) {
 886			curlen = PAGE_SIZE - page_base;
 887			if (curlen > pagelist_len)
 888				curlen = pagelist_len;
 889
 890			dprintk("RPC:       %s: page %d"
 891				" srcp 0x%p len %d curlen %d\n",
 892				__func__, i, srcp, copy_len, curlen);
 893			destp = kmap_atomic(ppages[i]);
 894			memcpy(destp + page_base, srcp, curlen);
 895			flush_dcache_page(ppages[i]);
 896			kunmap_atomic(destp);
 897			srcp += curlen;
 898			copy_len -= curlen;
 899			fixup_copy_count += curlen;
 900			pagelist_len -= curlen;
 901			if (!pagelist_len)
 902				break;
 903			page_base = 0;
 904		}
 
 
 
 905
 906		/* Implicit padding for the last segment in a Write
 907		 * chunk is inserted inline at the front of the tail
 908		 * iovec. The upper layer ignores the content of
 909		 * the pad. Simply ensure inline content in the tail
 910		 * that follows the Write chunk is properly aligned.
 911		 */
 912		if (pad)
 913			srcp -= pad;
 914	}
 915
 916	/* The tail iovec is redirected to the remaining data
 917	 * in the receive buffer, to avoid a memcopy.
 918	 */
 919	if (copy_len || pad) {
 920		rqst->rq_rcv_buf.tail[0].iov_base = srcp;
 921		rqst->rq_private_buf.tail[0].iov_base = srcp;
 922	}
 
 
 
 
 
 
 
 923
 924	return fixup_copy_count;
 
 925}
 926
 927#if defined(CONFIG_SUNRPC_BACKCHANNEL)
 928/* By convention, backchannel calls arrive via rdma_msg type
 929 * messages, and never populate the chunk lists. This makes
 930 * the RPC/RDMA header small and fixed in size, so it is
 931 * straightforward to check the RPC header's direction field.
 932 */
 933static bool
 934rpcrdma_is_bcall(struct rpcrdma_msg *headerp)
 935{
 936	__be32 *p = (__be32 *)headerp;
 937
 938	if (headerp->rm_type != rdma_msg)
 939		return false;
 940	if (headerp->rm_body.rm_chunks[0] != xdr_zero)
 941		return false;
 942	if (headerp->rm_body.rm_chunks[1] != xdr_zero)
 943		return false;
 944	if (headerp->rm_body.rm_chunks[2] != xdr_zero)
 945		return false;
 946
 947	/* sanity */
 948	if (p[7] != headerp->rm_xid)
 949		return false;
 950	/* call direction */
 951	if (p[8] != cpu_to_be32(RPC_CALL))
 952		return false;
 953
 954	return true;
 
 
 
 
 
 
 
 955}
 956#endif	/* CONFIG_SUNRPC_BACKCHANNEL */
 957
 958/* Process received RPC/RDMA messages.
 959 *
 960 * Errors must result in the RPC task either being awakened, or
 961 * allowed to timeout, to discover the errors at that time.
 962 */
 963void
 964rpcrdma_reply_handler(struct work_struct *work)
 965{
 966	struct rpcrdma_rep *rep =
 967			container_of(work, struct rpcrdma_rep, rr_work);
 968	struct rpcrdma_msg *headerp;
 969	struct rpcrdma_req *req;
 970	struct rpc_rqst *rqst;
 971	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
 972	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 973	__be32 *iptr;
 974	int rdmalen, status, rmerr;
 975	unsigned long cwnd;
 976
 977	dprintk("RPC:       %s: incoming rep %p\n", __func__, rep);
 978
 979	if (rep->rr_len == RPCRDMA_BAD_LEN)
 980		goto out_badstatus;
 981	if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
 982		goto out_shortreply;
 983
 984	headerp = rdmab_to_msg(rep->rr_rdmabuf);
 985#if defined(CONFIG_SUNRPC_BACKCHANNEL)
 986	if (rpcrdma_is_bcall(headerp))
 987		goto out_bcall;
 988#endif
 
 
 
 
 
 
 
 
 
 989
 990	/* Match incoming rpcrdma_rep to an rpcrdma_req to
 991	 * get context for handling any incoming chunks.
 992	 */
 993	spin_lock_bh(&xprt->transport_lock);
 994	rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 995	if (!rqst)
 996		goto out_nomatch;
 
 
 
 
 
 
 
 
 997
 
 
 
 
 998	req = rpcr_to_rdmar(rqst);
 999	if (req->rl_reply)
1000		goto out_duplicate;
1001
1002	/* Sanity checking has passed. We are now committed
1003	 * to complete this transaction.
1004	 */
1005	list_del_init(&rqst->rq_list);
1006	spin_unlock_bh(&xprt->transport_lock);
1007	dprintk("RPC:       %s: reply %p completes request %p (xid 0x%08x)\n",
1008		__func__, rep, req, be32_to_cpu(headerp->rm_xid));
1009
1010	/* from here on, the reply is no longer an orphan */
1011	req->rl_reply = rep;
1012	xprt->reestablish_timeout = 0;
1013
1014	if (headerp->rm_vers != rpcrdma_version)
1015		goto out_badversion;
1016
1017	/* check for expected message types */
1018	/* The order of some of these tests is important. */
1019	switch (headerp->rm_type) {
1020	case rdma_msg:
1021		/* never expect read chunks */
1022		/* never expect reply chunks (two ways to check) */
1023		/* never expect write chunks without having offered RDMA */
1024		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
1025		    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
1026		     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
1027		    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
1028		     list_empty(&req->rl_registered)))
1029			goto badheader;
1030		if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
1031			/* count any expected write chunks in read reply */
1032			/* start at write chunk array count */
1033			iptr = &headerp->rm_body.rm_chunks[2];
1034			rdmalen = rpcrdma_count_chunks(rep, 1, &iptr);
 
1035			/* check for validity, and no reply chunk after */
1036			if (rdmalen < 0 || *iptr++ != xdr_zero)
1037				goto badheader;
1038			rep->rr_len -=
1039			    ((unsigned char *)iptr - (unsigned char *)headerp);
1040			status = rep->rr_len + rdmalen;
1041			r_xprt->rx_stats.total_rdma_reply += rdmalen;
1042			/* special case - last chunk may omit padding */
1043			if (rdmalen &= 3) {
1044				rdmalen = 4 - rdmalen;
1045				status += rdmalen;
1046			}
1047		} else {
1048			/* else ordinary inline */
1049			rdmalen = 0;
1050			iptr = (__be32 *)((unsigned char *)headerp +
1051							RPCRDMA_HDRLEN_MIN);
1052			rep->rr_len -= RPCRDMA_HDRLEN_MIN;
1053			status = rep->rr_len;
1054		}
1055
1056		r_xprt->rx_stats.fixup_copy_count +=
1057			rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len,
1058					     rdmalen);
1059		break;
1060
1061	case rdma_nomsg:
1062		/* never expect read or write chunks, always reply chunks */
1063		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
1064		    headerp->rm_body.rm_chunks[1] != xdr_zero ||
1065		    headerp->rm_body.rm_chunks[2] != xdr_one ||
1066		    list_empty(&req->rl_registered))
1067			goto badheader;
1068		iptr = (__be32 *)((unsigned char *)headerp +
1069							RPCRDMA_HDRLEN_MIN);
1070		rdmalen = rpcrdma_count_chunks(rep, 0, &iptr);
1071		if (rdmalen < 0)
1072			goto badheader;
1073		r_xprt->rx_stats.total_rdma_reply += rdmalen;
1074		/* Reply chunk buffer already is the reply vector - no fixup. */
1075		status = rdmalen;
1076		break;
1077
1078	case rdma_error:
1079		goto out_rdmaerr;
1080
1081badheader:
1082	default:
1083		dprintk("RPC: %5u %s: invalid rpcrdma reply (type %u)\n",
1084			rqst->rq_task->tk_pid, __func__,
1085			be32_to_cpu(headerp->rm_type));
 
 
 
 
 
1086		status = -EIO;
1087		r_xprt->rx_stats.bad_reply_count++;
1088		break;
1089	}
1090
1091out:
1092	/* Invalidate and flush the data payloads before waking the
1093	 * waiting application. This guarantees the memory region is
1094	 * properly fenced from the server before the application
1095	 * accesses the data. It also ensures proper send flow
1096	 * control: waking the next RPC waits until this RPC has
1097	 * relinquished all its Send Queue entries.
1098	 */
1099	if (!list_empty(&req->rl_registered))
1100		r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);
1101
1102	spin_lock_bh(&xprt->transport_lock);
1103	cwnd = xprt->cwnd;
1104	xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
1105	if (xprt->cwnd > cwnd)
1106		xprt_release_rqst_cong(rqst->rq_task);
1107
1108	xprt_complete_rqst(rqst->rq_task, status);
1109	spin_unlock_bh(&xprt->transport_lock);
1110	dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
1111			__func__, xprt, rqst, status);
1112	return;
1113
1114out_badstatus:
1115	rpcrdma_recv_buffer_put(rep);
1116	if (r_xprt->rx_ep.rep_connected == 1) {
1117		r_xprt->rx_ep.rep_connected = -EIO;
1118		rpcrdma_conn_func(&r_xprt->rx_ep);
1119	}
1120	return;
1121
1122#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1123out_bcall:
1124	rpcrdma_bc_receive_call(r_xprt, rep);
1125	return;
1126#endif
1127
1128/* If the incoming reply terminated a pending RPC, the next
1129 * RPC call will post a replacement receive buffer as it is
1130 * being marshaled.
1131 */
1132out_badversion:
1133	dprintk("RPC:       %s: invalid version %d\n",
1134		__func__, be32_to_cpu(headerp->rm_vers));
1135	status = -EIO;
1136	r_xprt->rx_stats.bad_reply_count++;
1137	goto out;
1138
1139out_rdmaerr:
1140	rmerr = be32_to_cpu(headerp->rm_body.rm_error.rm_err);
1141	switch (rmerr) {
1142	case ERR_VERS:
1143		pr_err("%s: server reports header version error (%u-%u)\n",
1144		       __func__,
1145		       be32_to_cpu(headerp->rm_body.rm_error.rm_vers_low),
1146		       be32_to_cpu(headerp->rm_body.rm_error.rm_vers_high));
1147		break;
1148	case ERR_CHUNK:
1149		pr_err("%s: server reports header decoding error\n",
1150		       __func__);
 
1151		break;
1152	default:
1153		pr_err("%s: server reports unknown error %d\n",
1154		       __func__, rmerr);
1155	}
1156	status = -EREMOTEIO;
1157	r_xprt->rx_stats.bad_reply_count++;
1158	goto out;
1159
1160/* If no pending RPC transaction was matched, post a replacement
1161 * receive buffer before returning.
1162 */
1163out_shortreply:
1164	dprintk("RPC:       %s: short/invalid reply\n", __func__);
1165	goto repost;
1166
1167out_nomatch:
1168	spin_unlock_bh(&xprt->transport_lock);
1169	dprintk("RPC:       %s: no match for incoming xid 0x%08x len %d\n",
1170		__func__, be32_to_cpu(headerp->rm_xid),
1171		rep->rr_len);
1172	goto repost;
1173
1174out_duplicate:
1175	spin_unlock_bh(&xprt->transport_lock);
1176	dprintk("RPC:       %s: "
1177		"duplicate reply %p to RPC request %p: xid 0x%08x\n",
1178		__func__, rep, req, be32_to_cpu(headerp->rm_xid));
1179
1180repost:
1181	r_xprt->rx_stats.bad_reply_count++;
1182	if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, rep))
1183		rpcrdma_recv_buffer_put(rep);
1184}
v3.1
  1/*
  2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
  3 *
  4 * This software is available to you under a choice of one of two
  5 * licenses.  You may choose to be licensed under the terms of the GNU
  6 * General Public License (GPL) Version 2, available from the file
  7 * COPYING in the main directory of this source tree, or the BSD-type
  8 * license below:
  9 *
 10 * Redistribution and use in source and binary forms, with or without
 11 * modification, are permitted provided that the following conditions
 12 * are met:
 13 *
 14 *      Redistributions of source code must retain the above copyright
 15 *      notice, this list of conditions and the following disclaimer.
 16 *
 17 *      Redistributions in binary form must reproduce the above
 18 *      copyright notice, this list of conditions and the following
 19 *      disclaimer in the documentation and/or other materials provided
 20 *      with the distribution.
 21 *
 22 *      Neither the name of the Network Appliance, Inc. nor the names of
 23 *      its contributors may be used to endorse or promote products
 24 *      derived from this software without specific prior written
 25 *      permission.
 26 *
 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 38 */
 39
 40/*
 41 * rpc_rdma.c
 42 *
 43 * This file contains the guts of the RPC RDMA protocol, and
 44 * does marshaling/unmarshaling, etc. It is also where interfacing
 45 * to the Linux RPC framework lives.
 46 */
 47
 48#include "xprt_rdma.h"
 49
 50#include <linux/highmem.h>
 51
 52#ifdef RPC_DEBUG
 53# define RPCDBG_FACILITY	RPCDBG_TRANS
 54#endif
 55
 56enum rpcrdma_chunktype {
 57	rpcrdma_noch = 0,
 58	rpcrdma_readch,
 59	rpcrdma_areadch,
 60	rpcrdma_writech,
 61	rpcrdma_replych
 62};
 63
 64#ifdef RPC_DEBUG
 65static const char transfertypes[][12] = {
 66	"pure inline",	/* no chunks */
 67	" read chunk",	/* some argument via rdma read */
 68	"*read chunk",	/* entire request via rdma read */
 69	"write chunk",	/* some result via rdma write */
 70	"reply chunk"	/* entire reply via rdma write */
 71};
 72#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 73
 74/*
 75 * Chunk assembly from upper layer xdr_buf.
 76 *
 77 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
 78 * elements. Segments are then coalesced when registered, if possible
 79 * within the selected memreg mode.
 80 *
 81 * Note, this routine is never called if the connection's memory
 82 * registration strategy is 0 (bounce buffers).
 83 */
 84
 85static int
 86rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
 87	enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
 
 88{
 89	int len, n = 0, p;
 90	int page_base;
 91	struct page **ppages;
 92
 93	if (pos == 0 && xdrbuf->head[0].iov_len) {
 94		seg[n].mr_page = NULL;
 95		seg[n].mr_offset = xdrbuf->head[0].iov_base;
 96		seg[n].mr_len = xdrbuf->head[0].iov_len;
 97		++n;
 98	}
 99
100	len = xdrbuf->page_len;
101	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
102	page_base = xdrbuf->page_base & ~PAGE_MASK;
103	p = 0;
104	while (len && n < nsegs) {
 
 
 
 
 
 
105		seg[n].mr_page = ppages[p];
106		seg[n].mr_offset = (void *)(unsigned long) page_base;
107		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
108		BUG_ON(seg[n].mr_len > PAGE_SIZE);
 
109		len -= seg[n].mr_len;
110		++n;
111		++p;
112		page_base = 0;	/* page offset only applies to first page */
113	}
114
115	/* Message overflows the seg array */
116	if (len && n == nsegs)
117		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
118
119	if (xdrbuf->tail[0].iov_len) {
120		/* the rpcrdma protocol allows us to omit any trailing
121		 * xdr pad bytes, saving the server an RDMA operation. */
122		if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
123			return n;
124		if (n == nsegs)
125			/* Tail remains, but we're out of segments */
126			return 0;
127		seg[n].mr_page = NULL;
128		seg[n].mr_offset = xdrbuf->tail[0].iov_base;
129		seg[n].mr_len = xdrbuf->tail[0].iov_len;
130		++n;
131	}
132
133	return n;
 
 
 
 
134}
135
136/*
137 * Create read/write chunk lists, and reply chunks, for RDMA
138 *
139 *   Assume check against THRESHOLD has been done, and chunks are required.
140 *   Assume only encoding one list entry for read|write chunks. The NFSv3
141 *     protocol is simple enough to allow this as it only has a single "bulk
142 *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
143 *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
144 *
145 * When used for a single reply chunk (which is a special write
146 * chunk used for the entire reply, rather than just the data), it
147 * is used primarily for READDIR and READLINK which would otherwise
148 * be severely size-limited by a small rdma inline read max. The server
149 * response will come back as an RDMA Write, followed by a message
150 * of type RDMA_NOMSG carrying the xid and length. As a result, reply
151 * chunks do not provide data alignment, however they do not require
152 * "fixup" (moving the response to the upper layer buffer) either.
153 *
154 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
155 *
156 *  Read chunklist (a linked list):
157 *   N elements, position P (same P for all chunks of same arg!):
158 *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
159 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160 *  Write chunklist (a list of (one) counted array):
161 *   N elements:
162 *    1 - N - HLOO - HLOO - ... - HLOO - 0
163 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
164 *  Reply chunk (a counted array):
165 *   N elements:
166 *    1 - N - HLOO - HLOO - ... - HLOO
 
 
 
167 */
 
 
 
 
 
 
 
 
 
168
169static unsigned int
170rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
171		struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
172{
173	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
174	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_task->tk_xprt);
175	int nsegs, nchunks = 0;
176	unsigned int pos;
177	struct rpcrdma_mr_seg *seg = req->rl_segments;
178	struct rpcrdma_read_chunk *cur_rchunk = NULL;
179	struct rpcrdma_write_array *warray = NULL;
180	struct rpcrdma_write_chunk *cur_wchunk = NULL;
181	__be32 *iptr = headerp->rm_body.rm_chunks;
182
183	if (type == rpcrdma_readch || type == rpcrdma_areadch) {
184		/* a read chunk - server will RDMA Read our memory */
185		cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
186	} else {
187		/* a write or reply chunk - server will RDMA Write our memory */
188		*iptr++ = xdr_zero;	/* encode a NULL read chunk list */
189		if (type == rpcrdma_replych)
190			*iptr++ = xdr_zero;	/* a NULL write chunk list */
191		warray = (struct rpcrdma_write_array *) iptr;
192		cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
193	}
194
195	if (type == rpcrdma_replych || type == rpcrdma_areadch)
196		pos = 0;
197	else
198		pos = target->head[0].iov_len;
199
200	nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
201	if (nsegs == 0)
202		return 0;
203
 
204	do {
205		/* bind/register the memory, then build chunk from result. */
206		int n = rpcrdma_register_external(seg, nsegs,
207						cur_wchunk != NULL, r_xprt);
208		if (n <= 0)
209			goto out;
210		if (cur_rchunk) {	/* read */
211			cur_rchunk->rc_discrim = xdr_one;
212			/* all read chunks have the same "position" */
213			cur_rchunk->rc_position = htonl(pos);
214			cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey);
215			cur_rchunk->rc_target.rs_length = htonl(seg->mr_len);
216			xdr_encode_hyper(
217					(__be32 *)&cur_rchunk->rc_target.rs_offset,
218					seg->mr_base);
219			dprintk("RPC:       %s: read chunk "
220				"elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
221				seg->mr_len, (unsigned long long)seg->mr_base,
222				seg->mr_rkey, pos, n < nsegs ? "more" : "last");
223			cur_rchunk++;
224			r_xprt->rx_stats.read_chunk_count++;
225		} else {		/* write/reply */
226			cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey);
227			cur_wchunk->wc_target.rs_length = htonl(seg->mr_len);
228			xdr_encode_hyper(
229					(__be32 *)&cur_wchunk->wc_target.rs_offset,
230					seg->mr_base);
231			dprintk("RPC:       %s: %s chunk "
232				"elem %d@0x%llx:0x%x (%s)\n", __func__,
233				(type == rpcrdma_replych) ? "reply" : "write",
234				seg->mr_len, (unsigned long long)seg->mr_base,
235				seg->mr_rkey, n < nsegs ? "more" : "last");
236			cur_wchunk++;
237			if (type == rpcrdma_replych)
238				r_xprt->rx_stats.reply_chunk_count++;
239			else
240				r_xprt->rx_stats.write_chunk_count++;
241			r_xprt->rx_stats.total_rdma_request += seg->mr_len;
242		}
243		nchunks++;
244		seg   += n;
245		nsegs -= n;
246	} while (nsegs);
247
248	/* success. all failures return above */
249	req->rl_nchunks = nchunks;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
250
251	BUG_ON(nchunks == 0);
252	BUG_ON((r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR)
253	       && (nchunks > 3));
254
255	/*
256	 * finish off header. If write, marshal discrim and nchunks.
 
 
 
 
 
 
 
 
 
 
257	 */
258	if (cur_rchunk) {
259		iptr = (__be32 *) cur_rchunk;
260		*iptr++ = xdr_zero;	/* finish the read chunk list */
261		*iptr++ = xdr_zero;	/* encode a NULL write chunk list */
262		*iptr++ = xdr_zero;	/* encode a NULL reply chunk */
263	} else {
264		warray->wc_discrim = xdr_one;
265		warray->wc_nchunks = htonl(nchunks);
266		iptr = (__be32 *) cur_wchunk;
267		if (type == rpcrdma_writech) {
268			*iptr++ = xdr_zero; /* finish the write chunk list */
269			*iptr++ = xdr_zero; /* encode a NULL reply chunk */
 
 
 
 
 
 
 
 
 
 
270		}
271	}
272
273	/*
274	 * Return header size.
 
 
275	 */
276	return (unsigned char *)iptr - (unsigned char *)headerp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
277
278out:
279	for (pos = 0; nchunks--;)
280		pos += rpcrdma_deregister_external(
281				&req->rl_segments[pos], r_xprt, NULL);
282	return 0;
 
 
 
 
 
 
283}
284
285/*
286 * Copy write data inline.
287 * This function is used for "small" requests. Data which is passed
288 * to RPC via iovecs (or page list) is copied directly into the
289 * pre-registered memory buffer for this request. For small amounts
290 * of data, this is efficient. The cutoff value is tunable.
291 */
292static int
293rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
294{
295	int i, npages, curlen;
296	int copy_len;
297	unsigned char *srcp, *destp;
298	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
299	int page_base;
300	struct page **ppages;
 
 
 
 
 
301
302	destp = rqst->rq_svec[0].iov_base;
303	curlen = rqst->rq_svec[0].iov_len;
304	destp += curlen;
305	/*
306	 * Do optional padding where it makes sense. Alignment of write
307	 * payload can help the server, if our setting is accurate.
308	 */
309	pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
310	if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
311		pad = 0;	/* don't pad this request */
312
313	dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
314		__func__, pad, destp, rqst->rq_slen, curlen);
315
316	copy_len = rqst->rq_snd_buf.page_len;
317
318	if (rqst->rq_snd_buf.tail[0].iov_len) {
319		curlen = rqst->rq_snd_buf.tail[0].iov_len;
320		if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
321			memmove(destp + copy_len,
322				rqst->rq_snd_buf.tail[0].iov_base, curlen);
323			r_xprt->rx_stats.pullup_copy_count += curlen;
324		}
325		dprintk("RPC:       %s: tail destp 0x%p len %d\n",
326			__func__, destp + copy_len, curlen);
327		rqst->rq_svec[0].iov_len += curlen;
328	}
329	r_xprt->rx_stats.pullup_copy_count += copy_len;
330
331	page_base = rqst->rq_snd_buf.page_base;
332	ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
333	page_base &= ~PAGE_MASK;
334	npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
335	for (i = 0; copy_len && i < npages; i++) {
336		curlen = PAGE_SIZE - page_base;
337		if (curlen > copy_len)
338			curlen = copy_len;
339		dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
340			__func__, i, destp, copy_len, curlen);
341		srcp = kmap_atomic(ppages[i], KM_SKB_SUNRPC_DATA);
342		memcpy(destp, srcp+page_base, curlen);
343		kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
344		rqst->rq_svec[0].iov_len += curlen;
345		destp += curlen;
346		copy_len -= curlen;
347		page_base = 0;
348	}
349	/* header now contains entire send message */
350	return pad;
351}
352
353/*
354 * Marshal a request: the primary job of this routine is to choose
355 * the transfer modes. See comments below.
356 *
357 * Uses multiple RDMA IOVs for a request:
358 *  [0] -- RPC RDMA header, which uses memory from the *start* of the
359 *         preregistered buffer that already holds the RPC data in
360 *         its middle.
361 *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
362 *  [2] -- optional padding.
363 *  [3] -- if padded, header only in [1] and data here.
364 */
365
366int
367rpcrdma_marshal_req(struct rpc_rqst *rqst)
368{
369	struct rpc_xprt *xprt = rqst->rq_task->tk_xprt;
370	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
371	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
372	char *base;
373	size_t hdrlen, rpclen, padlen;
374	enum rpcrdma_chunktype rtype, wtype;
375	struct rpcrdma_msg *headerp;
 
 
 
 
376
377	/*
378	 * rpclen gets amount of data in first buffer, which is the
379	 * pre-registered buffer.
380	 */
381	base = rqst->rq_svec[0].iov_base;
382	rpclen = rqst->rq_svec[0].iov_len;
383
384	/* build RDMA header in private area at front */
385	headerp = (struct rpcrdma_msg *) req->rl_base;
386	/* don't htonl XID, it's already done in request */
387	headerp->rm_xid = rqst->rq_xid;
388	headerp->rm_vers = xdr_one;
389	headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
390	headerp->rm_type = htonl(RDMA_MSG);
 
 
 
 
 
 
 
391
392	/*
393	 * Chunks needed for results?
394	 *
395	 * o If the expected result is under the inline threshold, all ops
396	 *   return as inline (but see later).
 
 
397	 * o Large non-read ops return as a single reply chunk.
398	 * o Large read ops return data as write chunk(s), header as inline.
399	 *
400	 * Note: the NFS code sending down multiple result segments implies
401	 * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
402	 */
403
404	/*
405	 * This code can handle read chunks, write chunks OR reply
406	 * chunks -- only one type. If the request is too big to fit
407	 * inline, then we will choose read chunks. If the request is
408	 * a READ, then use write chunks to separate the file data
409	 * into pages; otherwise use reply chunks.
410	 */
411	if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
412		wtype = rpcrdma_noch;
413	else if (rqst->rq_rcv_buf.page_len == 0)
414		wtype = rpcrdma_replych;
415	else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
416		wtype = rpcrdma_writech;
417	else
418		wtype = rpcrdma_replych;
419
420	/*
421	 * Chunks needed for arguments?
422	 *
423	 * o If the total request is under the inline threshold, all ops
424	 *   are sent as inline.
 
 
425	 * o Large non-write ops are sent with the entire message as a
426	 *   single read chunk (protocol 0-position special case).
427	 * o Large write ops transmit data as read chunk(s), header as
428	 *   inline.
429	 *
430	 * Note: the NFS code sending down multiple argument segments
431	 * implies the op is a write.
432	 * TBD check NFSv4 setacl
433	 */
434	if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
435		rtype = rpcrdma_noch;
436	else if (rqst->rq_snd_buf.page_len == 0)
 
 
 
 
 
 
 
437		rtype = rpcrdma_areadch;
438	else
439		rtype = rpcrdma_readch;
440
441	/* The following simplification is not true forever */
442	if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
443		wtype = rpcrdma_noch;
444	BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch);
445
446	if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS &&
447	    (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) {
448		/* forced to "pure inline"? */
449		dprintk("RPC:       %s: too much data (%d/%d) for inline\n",
450			__func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len);
451		return -1;
452	}
453
454	hdrlen = 28; /*sizeof *headerp;*/
455	padlen = 0;
456
457	/*
458	 * Pull up any extra send data into the preregistered buffer.
459	 * When padding is in use and applies to the transfer, insert
460	 * it and change the message type.
461	 */
462	if (rtype == rpcrdma_noch) {
463
464		padlen = rpcrdma_inline_pullup(rqst,
465						RPCRDMA_INLINE_PAD_VALUE(rqst));
466
467		if (padlen) {
468			headerp->rm_type = htonl(RDMA_MSGP);
469			headerp->rm_body.rm_padded.rm_align =
470				htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
471			headerp->rm_body.rm_padded.rm_thresh =
472				htonl(RPCRDMA_INLINE_PAD_THRESH);
473			headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
474			headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
475			headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
476			hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
477			BUG_ON(wtype != rpcrdma_noch);
478
479		} else {
480			headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
481			headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
482			headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
483			/* new length after pullup */
484			rpclen = rqst->rq_svec[0].iov_len;
485			/*
486			 * Currently we try to not actually use read inline.
487			 * Reply chunks have the desirable property that
488			 * they land, packed, directly in the target buffers
489			 * without headers, so they require no fixup. The
490			 * additional RDMA Write op sends the same amount
491			 * of data, streams on-the-wire and adds no overhead
492			 * on receive. Therefore, we request a reply chunk
493			 * for non-writes wherever feasible and efficient.
494			 */
495			if (wtype == rpcrdma_noch &&
496			    r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
497				wtype = rpcrdma_replych;
498		}
499	}
500
501	/*
502	 * Marshal chunks. This routine will return the header length
503	 * consumed by marshaling.
504	 */
505	if (rtype != rpcrdma_noch) {
506		hdrlen = rpcrdma_create_chunks(rqst,
507					&rqst->rq_snd_buf, headerp, rtype);
508		wtype = rtype;	/* simplify dprintk */
509
510	} else if (wtype != rpcrdma_noch) {
511		hdrlen = rpcrdma_create_chunks(rqst,
512					&rqst->rq_rcv_buf, headerp, wtype);
 
 
 
 
 
 
 
 
 
 
 
 
 
513	}
 
514
515	if (hdrlen == 0)
516		return -1;
517
518	dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
519		" headerp 0x%p base 0x%p lkey 0x%x\n",
520		__func__, transfertypes[wtype], hdrlen, rpclen, padlen,
521		headerp, base, req->rl_iov.lkey);
522
523	/*
524	 * initialize send_iov's - normally only two: rdma chunk header and
525	 * single preregistered RPC header buffer, but if padding is present,
526	 * then use a preregistered (and zeroed) pad buffer between the RPC
527	 * header and any write data. In all non-rdma cases, any following
528	 * data has been copied into the RPC header buffer.
529	 */
530	req->rl_send_iov[0].addr = req->rl_iov.addr;
531	req->rl_send_iov[0].length = hdrlen;
532	req->rl_send_iov[0].lkey = req->rl_iov.lkey;
533
534	req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
535	req->rl_send_iov[1].length = rpclen;
536	req->rl_send_iov[1].lkey = req->rl_iov.lkey;
537
538	req->rl_niovs = 2;
539
540	if (padlen) {
541		struct rpcrdma_ep *ep = &r_xprt->rx_ep;
542
543		req->rl_send_iov[2].addr = ep->rep_pad.addr;
544		req->rl_send_iov[2].length = padlen;
545		req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
546
547		req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
548		req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
549		req->rl_send_iov[3].lkey = req->rl_iov.lkey;
550
551		req->rl_niovs = 4;
552	}
553
554	return 0;
555}
556
557/*
558 * Chase down a received write or reply chunklist to get length
559 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
560 */
561static int
562rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
563{
564	unsigned int i, total_len;
565	struct rpcrdma_write_chunk *cur_wchunk;
 
566
567	i = ntohl(**iptrp);	/* get array count */
568	if (i > max)
569		return -1;
570	cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
571	total_len = 0;
572	while (i--) {
573		struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
574		ifdebug(FACILITY) {
575			u64 off;
576			xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
577			dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
578				__func__,
579				ntohl(seg->rs_length),
580				(unsigned long long)off,
581				ntohl(seg->rs_handle));
582		}
583		total_len += ntohl(seg->rs_length);
584		++cur_wchunk;
585	}
586	/* check and adjust for properly terminated write chunk */
587	if (wrchunk) {
588		__be32 *w = (__be32 *) cur_wchunk;
589		if (*w++ != xdr_zero)
590			return -1;
591		cur_wchunk = (struct rpcrdma_write_chunk *) w;
592	}
593	if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
594		return -1;
595
596	*iptrp = (__be32 *) cur_wchunk;
597	return total_len;
598}
599
600/*
601 * Scatter inline received data back into provided iov's.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
602 */
603static void
604rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
605{
606	int i, npages, curlen, olen;
 
607	char *destp;
608	struct page **ppages;
609	int page_base;
610
 
 
 
 
 
 
 
 
 
611	curlen = rqst->rq_rcv_buf.head[0].iov_len;
612	if (curlen > copy_len) {	/* write chunk header fixup */
613		curlen = copy_len;
614		rqst->rq_rcv_buf.head[0].iov_len = curlen;
615	}
616
617	dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
618		__func__, srcp, copy_len, curlen);
619
620	/* Shift pointer for first receive segment only */
621	rqst->rq_rcv_buf.head[0].iov_base = srcp;
622	srcp += curlen;
623	copy_len -= curlen;
624
625	olen = copy_len;
626	i = 0;
627	rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
628	page_base = rqst->rq_rcv_buf.page_base;
629	ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
630	page_base &= ~PAGE_MASK;
 
 
 
631
632	if (copy_len && rqst->rq_rcv_buf.page_len) {
633		npages = PAGE_ALIGN(page_base +
634			rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
635		for (; i < npages; i++) {
 
636			curlen = PAGE_SIZE - page_base;
637			if (curlen > copy_len)
638				curlen = copy_len;
 
639			dprintk("RPC:       %s: page %d"
640				" srcp 0x%p len %d curlen %d\n",
641				__func__, i, srcp, copy_len, curlen);
642			destp = kmap_atomic(ppages[i], KM_SKB_SUNRPC_DATA);
643			memcpy(destp + page_base, srcp, curlen);
644			flush_dcache_page(ppages[i]);
645			kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
646			srcp += curlen;
647			copy_len -= curlen;
648			if (copy_len == 0)
 
 
649				break;
650			page_base = 0;
651		}
652		rqst->rq_rcv_buf.page_len = olen - copy_len;
653	} else
654		rqst->rq_rcv_buf.page_len = 0;
655
656	if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
657		curlen = copy_len;
658		if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
659			curlen = rqst->rq_rcv_buf.tail[0].iov_len;
660		if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
661			memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
662		dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
663			__func__, srcp, copy_len, curlen);
664		rqst->rq_rcv_buf.tail[0].iov_len = curlen;
665		copy_len -= curlen; ++i;
666	} else
667		rqst->rq_rcv_buf.tail[0].iov_len = 0;
668
669	if (pad) {
670		/* implicit padding on terminal chunk */
671		unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
672		while (pad--)
673			p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
674	}
675
676	if (copy_len)
677		dprintk("RPC:       %s: %d bytes in"
678			" %d extra segments (%d lost)\n",
679			__func__, olen, i, copy_len);
680
681	/* TBD avoid a warning from call_decode() */
682	rqst->rq_private_buf = rqst->rq_rcv_buf;
683}
684
685/*
686 * This function is called when an async event is posted to
687 * the connection which changes the connection state. All it
688 * does at this point is mark the connection up/down, the rpc
689 * timers do the rest.
690 */
691void
692rpcrdma_conn_func(struct rpcrdma_ep *ep)
693{
694	struct rpc_xprt *xprt = ep->rep_xprt;
695
696	spin_lock_bh(&xprt->transport_lock);
697	if (++xprt->connect_cookie == 0)	/* maintain a reserved value */
698		++xprt->connect_cookie;
699	if (ep->rep_connected > 0) {
700		if (!xprt_test_and_set_connected(xprt))
701			xprt_wake_pending_tasks(xprt, 0);
702	} else {
703		if (xprt_test_and_clear_connected(xprt))
704			xprt_wake_pending_tasks(xprt, -ENOTCONN);
705	}
706	spin_unlock_bh(&xprt->transport_lock);
707}
 
 
 
708
709/*
710 * This function is called when memory window unbind which we are waiting
711 * for completes. Just use rr_func (zeroed by upcall) to signal completion.
712 */
713static void
714rpcrdma_unbind_func(struct rpcrdma_rep *rep)
715{
716	wake_up(&rep->rr_unbind);
717}
 
718
719/*
720 * Called as a tasklet to do req/reply match and complete a request
721 * Errors must result in the RPC task either being awakened, or
722 * allowed to timeout, to discover the errors at that time.
723 */
724void
725rpcrdma_reply_handler(struct rpcrdma_rep *rep)
726{
 
 
727	struct rpcrdma_msg *headerp;
728	struct rpcrdma_req *req;
729	struct rpc_rqst *rqst;
730	struct rpc_xprt *xprt = rep->rr_xprt;
731	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
732	__be32 *iptr;
733	int i, rdmalen, status;
 
 
 
734
735	/* Check status. If bad, signal disconnect and return rep to pool */
736	if (rep->rr_len == ~0U) {
737		rpcrdma_recv_buffer_put(rep);
738		if (r_xprt->rx_ep.rep_connected == 1) {
739			r_xprt->rx_ep.rep_connected = -EIO;
740			rpcrdma_conn_func(&r_xprt->rx_ep);
741		}
742		return;
743	}
744	if (rep->rr_len < 28) {
745		dprintk("RPC:       %s: short/invalid reply\n", __func__);
746		goto repost;
747	}
748	headerp = (struct rpcrdma_msg *) rep->rr_base;
749	if (headerp->rm_vers != xdr_one) {
750		dprintk("RPC:       %s: invalid version %d\n",
751			__func__, ntohl(headerp->rm_vers));
752		goto repost;
753	}
754
755	/* Get XID and try for a match. */
756	spin_lock(&xprt->transport_lock);
 
 
757	rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
758	if (rqst == NULL) {
759		spin_unlock(&xprt->transport_lock);
760		dprintk("RPC:       %s: reply 0x%p failed "
761			"to match any request xid 0x%08x len %d\n",
762			__func__, rep, headerp->rm_xid, rep->rr_len);
763repost:
764		r_xprt->rx_stats.bad_reply_count++;
765		rep->rr_func = rpcrdma_reply_handler;
766		if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
767			rpcrdma_recv_buffer_put(rep);
768
769		return;
770	}
771
772	/* get request object */
773	req = rpcr_to_rdmar(rqst);
 
 
774
775	dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
776		"                   RPC request 0x%p xid 0x%08x\n",
777			__func__, rep, req, rqst, headerp->rm_xid);
778
779	BUG_ON(!req || req->rl_reply);
 
 
780
781	/* from here on, the reply is no longer an orphan */
782	req->rl_reply = rep;
 
 
 
 
783
784	/* check for expected message types */
785	/* The order of some of these tests is important. */
786	switch (headerp->rm_type) {
787	case htonl(RDMA_MSG):
788		/* never expect read chunks */
789		/* never expect reply chunks (two ways to check) */
790		/* never expect write chunks without having offered RDMA */
791		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
792		    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
793		     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
794		    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
795		     req->rl_nchunks == 0))
796			goto badheader;
797		if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
798			/* count any expected write chunks in read reply */
799			/* start at write chunk array count */
800			iptr = &headerp->rm_body.rm_chunks[2];
801			rdmalen = rpcrdma_count_chunks(rep,
802						req->rl_nchunks, 1, &iptr);
803			/* check for validity, and no reply chunk after */
804			if (rdmalen < 0 || *iptr++ != xdr_zero)
805				goto badheader;
806			rep->rr_len -=
807			    ((unsigned char *)iptr - (unsigned char *)headerp);
808			status = rep->rr_len + rdmalen;
809			r_xprt->rx_stats.total_rdma_reply += rdmalen;
810			/* special case - last chunk may omit padding */
811			if (rdmalen &= 3) {
812				rdmalen = 4 - rdmalen;
813				status += rdmalen;
814			}
815		} else {
816			/* else ordinary inline */
817			rdmalen = 0;
818			iptr = (__be32 *)((unsigned char *)headerp + 28);
819			rep->rr_len -= 28; /*sizeof *headerp;*/
 
820			status = rep->rr_len;
821		}
822		/* Fix up the rpc results for upper layer */
823		rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 
 
824		break;
825
826	case htonl(RDMA_NOMSG):
827		/* never expect read or write chunks, always reply chunks */
828		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
829		    headerp->rm_body.rm_chunks[1] != xdr_zero ||
830		    headerp->rm_body.rm_chunks[2] != xdr_one ||
831		    req->rl_nchunks == 0)
832			goto badheader;
833		iptr = (__be32 *)((unsigned char *)headerp + 28);
834		rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 
835		if (rdmalen < 0)
836			goto badheader;
837		r_xprt->rx_stats.total_rdma_reply += rdmalen;
838		/* Reply chunk buffer already is the reply vector - no fixup. */
839		status = rdmalen;
840		break;
841
 
 
 
842badheader:
843	default:
844		dprintk("%s: invalid rpcrdma reply header (type %d):"
845				" chunks[012] == %d %d %d"
846				" expected chunks <= %d\n",
847				__func__, ntohl(headerp->rm_type),
848				headerp->rm_body.rm_chunks[0],
849				headerp->rm_body.rm_chunks[1],
850				headerp->rm_body.rm_chunks[2],
851				req->rl_nchunks);
852		status = -EIO;
853		r_xprt->rx_stats.bad_reply_count++;
854		break;
855	}
856
857	/* If using mw bind, start the deregister process now. */
858	/* (Note: if mr_free(), cannot perform it here, in tasklet context) */
859	if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
860	case RPCRDMA_MEMWINDOWS:
861		for (i = 0; req->rl_nchunks-- > 1;)
862			i += rpcrdma_deregister_external(
863				&req->rl_segments[i], r_xprt, NULL);
864		/* Optionally wait (not here) for unbinds to complete */
865		rep->rr_func = rpcrdma_unbind_func;
866		(void) rpcrdma_deregister_external(&req->rl_segments[i],
867						   r_xprt, rep);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
868		break;
869	case RPCRDMA_MEMWINDOWS_ASYNC:
870		for (i = 0; req->rl_nchunks--;)
871			i += rpcrdma_deregister_external(&req->rl_segments[i],
872							 r_xprt, NULL);
873		break;
874	default:
875		break;
 
876	}
 
 
 
 
 
 
 
 
 
 
877
878	dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
879			__func__, xprt, rqst, status);
880	xprt_complete_rqst(rqst->rq_task, status);
881	spin_unlock(&xprt->transport_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
882}