git.oblomov.eu Git - linux-2.6/blob - net/sunrpc/xprtrdma/rpc_rdma.c

   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
  56 enum rpcrdma_chunktype {
  57         rpcrdma_noch = 0,
  58         rpcrdma_readch,
  59         rpcrdma_areadch,
  60         rpcrdma_writech,
  61         rpcrdma_replych
  62 };
  63
  64 #ifdef RPC_DEBUG
  65 static 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
  85 static int
  86 rpcrdma_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
  91         if (pos == 0 && xdrbuf->head[0].iov_len) {
  92                 seg[n].mr_page = NULL;
  93                 seg[n].mr_offset = xdrbuf->head[0].iov_base;
  94                 seg[n].mr_len = xdrbuf->head[0].iov_len;
  95                 ++n;
  96         }
  97
  98         if (xdrbuf->page_len && (xdrbuf->pages[0] != NULL)) {
  99                 if (n == nsegs)
 100                         return 0;
 101                 seg[n].mr_page = xdrbuf->pages[0];
 102                 seg[n].mr_offset = (void *)(unsigned long) xdrbuf->page_base;
 103                 seg[n].mr_len = min_t(u32,
 104                         PAGE_SIZE - xdrbuf->page_base, xdrbuf->page_len);
 105                 len = xdrbuf->page_len - seg[n].mr_len;
 106                 ++n;
 107                 p = 1;
 108                 while (len > 0) {
 109                         if (n == nsegs)
 110                                 return 0;
 111                         seg[n].mr_page = xdrbuf->pages[p];
 112                         seg[n].mr_offset = NULL;
 113                         seg[n].mr_len = min_t(u32, PAGE_SIZE, len);
 114                         len -= seg[n].mr_len;
 115                         ++n;
 116                         ++p;
 117                 }
 118         }
 119
 120         if (xdrbuf->tail[0].iov_len) {
 121                 if (n == nsegs)
 122                         return 0;
 123                 seg[n].mr_page = NULL;
 124                 seg[n].mr_offset = xdrbuf->tail[0].iov_base;
 125                 seg[n].mr_len = xdrbuf->tail[0].iov_len;
 126                 ++n;
 127         }
 128
 129         return n;
 130 }
 131
 132 /*
 133  * Create read/write chunk lists, and reply chunks, for RDMA
 134  *
 135  *   Assume check against THRESHOLD has been done, and chunks are required.
 136  *   Assume only encoding one list entry for read|write chunks. The NFSv3
 137  *     protocol is simple enough to allow this as it only has a single "bulk
 138  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
 139  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
 140  *
 141  * When used for a single reply chunk (which is a special write
 142  * chunk used for the entire reply, rather than just the data), it
 143  * is used primarily for READDIR and READLINK which would otherwise
 144  * be severely size-limited by a small rdma inline read max. The server
 145  * response will come back as an RDMA Write, followed by a message
 146  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
 147  * chunks do not provide data alignment, however they do not require
 148  * "fixup" (moving the response to the upper layer buffer) either.
 149  *
 150  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 151  *
 152  *  Read chunklist (a linked list):
 153  *   N elements, position P (same P for all chunks of same arg!):
 154  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
 155  *
 156  *  Write chunklist (a list of (one) counted array):
 157  *   N elements:
 158  *    1 - N - HLOO - HLOO - ... - HLOO - 0
 159  *
 160  *  Reply chunk (a counted array):
 161  *   N elements:
 162  *    1 - N - HLOO - HLOO - ... - HLOO
 163  */
 164
 165 static unsigned int
 166 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
 167                 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
 168 {
 169         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 170         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_task->tk_xprt);
 171         int nsegs, nchunks = 0;
 172         unsigned int pos;
 173         struct rpcrdma_mr_seg *seg = req->rl_segments;
 174         struct rpcrdma_read_chunk *cur_rchunk = NULL;
 175         struct rpcrdma_write_array *warray = NULL;
 176         struct rpcrdma_write_chunk *cur_wchunk = NULL;
 177         __be32 *iptr = headerp->rm_body.rm_chunks;
 178
 179         if (type == rpcrdma_readch || type == rpcrdma_areadch) {
 180                 /* a read chunk - server will RDMA Read our memory */
 181                 cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
 182         } else {
 183                 /* a write or reply chunk - server will RDMA Write our memory */
 184                 *iptr++ = xdr_zero;     /* encode a NULL read chunk list */
 185                 if (type == rpcrdma_replych)
 186                         *iptr++ = xdr_zero;     /* a NULL write chunk list */
 187                 warray = (struct rpcrdma_write_array *) iptr;
 188                 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
 189         }
 190
 191         if (type == rpcrdma_replych || type == rpcrdma_areadch)
 192                 pos = 0;
 193         else
 194                 pos = target->head[0].iov_len;
 195
 196         nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
 197         if (nsegs == 0)
 198                 return 0;
 199
 200         do {
 201                 /* bind/register the memory, then build chunk from result. */
 202                 int n = rpcrdma_register_external(seg, nsegs,
 203                                                 cur_wchunk != NULL, r_xprt);
 204                 if (n <= 0)
 205                         goto out;
 206                 if (cur_rchunk) {       /* read */
 207                         cur_rchunk->rc_discrim = xdr_one;
 208                         /* all read chunks have the same "position" */
 209                         cur_rchunk->rc_position = htonl(pos);
 210                         cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey);
 211                         cur_rchunk->rc_target.rs_length = htonl(seg->mr_len);
 212                         xdr_encode_hyper(
 213                                         (__be32 *)&cur_rchunk->rc_target.rs_offset,
 214                                         seg->mr_base);
 215                         dprintk("RPC:       %s: read chunk "
 216                                 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
 217                                 seg->mr_len, (unsigned long long)seg->mr_base,
 218                                 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
 219                         cur_rchunk++;
 220                         r_xprt->rx_stats.read_chunk_count++;
 221                 } else {                /* write/reply */
 222                         cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey);
 223                         cur_wchunk->wc_target.rs_length = htonl(seg->mr_len);
 224                         xdr_encode_hyper(
 225                                         (__be32 *)&cur_wchunk->wc_target.rs_offset,
 226                                         seg->mr_base);
 227                         dprintk("RPC:       %s: %s chunk "
 228                                 "elem %d@0x%llx:0x%x (%s)\n", __func__,
 229                                 (type == rpcrdma_replych) ? "reply" : "write",
 230                                 seg->mr_len, (unsigned long long)seg->mr_base,
 231                                 seg->mr_rkey, n < nsegs ? "more" : "last");
 232                         cur_wchunk++;
 233                         if (type == rpcrdma_replych)
 234                                 r_xprt->rx_stats.reply_chunk_count++;
 235                         else
 236                                 r_xprt->rx_stats.write_chunk_count++;
 237                         r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 238                 }
 239                 nchunks++;
 240                 seg   += n;
 241                 nsegs -= n;
 242         } while (nsegs);
 243
 244         /* success. all failures return above */
 245         req->rl_nchunks = nchunks;
 246
 247         BUG_ON(nchunks == 0);
 248
 249         /*
 250          * finish off header. If write, marshal discrim and nchunks.
 251          */
 252         if (cur_rchunk) {
 253                 iptr = (__be32 *) cur_rchunk;
 254                 *iptr++ = xdr_zero;     /* finish the read chunk list */
 255                 *iptr++ = xdr_zero;     /* encode a NULL write chunk list */
 256                 *iptr++ = xdr_zero;     /* encode a NULL reply chunk */
 257         } else {
 258                 warray->wc_discrim = xdr_one;
 259                 warray->wc_nchunks = htonl(nchunks);
 260                 iptr = (__be32 *) cur_wchunk;
 261                 if (type == rpcrdma_writech) {
 262                         *iptr++ = xdr_zero; /* finish the write chunk list */
 263                         *iptr++ = xdr_zero; /* encode a NULL reply chunk */
 264                 }
 265         }
 266
 267         /*
 268          * Return header size.
 269          */
 270         return (unsigned char *)iptr - (unsigned char *)headerp;
 271
 272 out:
 273         for (pos = 0; nchunks--;)
 274                 pos += rpcrdma_deregister_external(
 275                                 &req->rl_segments[pos], r_xprt, NULL);
 276         return 0;
 277 }
 278
 279 /*
 280  * Copy write data inline.
 281  * This function is used for "small" requests. Data which is passed
 282  * to RPC via iovecs (or page list) is copied directly into the
 283  * pre-registered memory buffer for this request. For small amounts
 284  * of data, this is efficient. The cutoff value is tunable.
 285  */
 286 static int
 287 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
 288 {
 289         int i, npages, curlen;
 290         int copy_len;
 291         unsigned char *srcp, *destp;
 292         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 293
 294         destp = rqst->rq_svec[0].iov_base;
 295         curlen = rqst->rq_svec[0].iov_len;
 296         destp += curlen;
 297         /*
 298          * Do optional padding where it makes sense. Alignment of write
 299          * payload can help the server, if our setting is accurate.
 300          */
 301         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 302         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 303                 pad = 0;        /* don't pad this request */
 304
 305         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 306                 __func__, pad, destp, rqst->rq_slen, curlen);
 307
 308         copy_len = rqst->rq_snd_buf.page_len;
 309         r_xprt->rx_stats.pullup_copy_count += copy_len;
 310         npages = PAGE_ALIGN(rqst->rq_snd_buf.page_base+copy_len) >> PAGE_SHIFT;
 311         for (i = 0; copy_len && i < npages; i++) {
 312                 if (i == 0)
 313                         curlen = PAGE_SIZE - rqst->rq_snd_buf.page_base;
 314                 else
 315                         curlen = PAGE_SIZE;
 316                 if (curlen > copy_len)
 317                         curlen = copy_len;
 318                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 319                         __func__, i, destp, copy_len, curlen);
 320                 srcp = kmap_atomic(rqst->rq_snd_buf.pages[i],
 321                                         KM_SKB_SUNRPC_DATA);
 322                 if (i == 0)
 323                         memcpy(destp, srcp+rqst->rq_snd_buf.page_base, curlen);
 324                 else
 325                         memcpy(destp, srcp, curlen);
 326                 kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
 327                 rqst->rq_svec[0].iov_len += curlen;
 328                 destp += curlen;
 329                 copy_len -= curlen;
 330         }
 331         if (rqst->rq_snd_buf.tail[0].iov_len) {
 332                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 333                 if (destp != rqst->rq_snd_buf.tail[0].iov_base) {
 334                         memcpy(destp,
 335                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 336                         r_xprt->rx_stats.pullup_copy_count += curlen;
 337                 }
 338                 dprintk("RPC:       %s: tail destp 0x%p len %d curlen %d\n",
 339                         __func__, destp, copy_len, curlen);
 340                 rqst->rq_svec[0].iov_len += curlen;
 341         }
 342         /* header now contains entire send message */
 343         return pad;
 344 }
 345
 346 /*
 347  * Marshal a request: the primary job of this routine is to choose
 348  * the transfer modes. See comments below.
 349  *
 350  * Uses multiple RDMA IOVs for a request:
 351  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 352  *         preregistered buffer that already holds the RPC data in
 353  *         its middle.
 354  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 355  *  [2] -- optional padding.
 356  *  [3] -- if padded, header only in [1] and data here.
 357  */
 358
 359 int
 360 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 361 {
 362         struct rpc_xprt *xprt = rqst->rq_task->tk_xprt;
 363         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 364         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 365         char *base;
 366         size_t hdrlen, rpclen, padlen;
 367         enum rpcrdma_chunktype rtype, wtype;
 368         struct rpcrdma_msg *headerp;
 369
 370         /*
 371          * rpclen gets amount of data in first buffer, which is the
 372          * pre-registered buffer.
 373          */
 374         base = rqst->rq_svec[0].iov_base;
 375         rpclen = rqst->rq_svec[0].iov_len;
 376
 377         /* build RDMA header in private area at front */
 378         headerp = (struct rpcrdma_msg *) req->rl_base;
 379         /* don't htonl XID, it's already done in request */
 380         headerp->rm_xid = rqst->rq_xid;
 381         headerp->rm_vers = xdr_one;
 382         headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
 383         headerp->rm_type = htonl(RDMA_MSG);
 384
 385         /*
 386          * Chunks needed for results?
 387          *
 388          * o If the expected result is under the inline threshold, all ops
 389          *   return as inline (but see later).
 390          * o Large non-read ops return as a single reply chunk.
 391          * o Large read ops return data as write chunk(s), header as inline.
 392          *
 393          * Note: the NFS code sending down multiple result segments implies
 394          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 395          */
 396
 397         /*
 398          * This code can handle read chunks, write chunks OR reply
 399          * chunks -- only one type. If the request is too big to fit
 400          * inline, then we will choose read chunks. If the request is
 401          * a READ, then use write chunks to separate the file data
 402          * into pages; otherwise use reply chunks.
 403          */
 404         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 405                 wtype = rpcrdma_noch;
 406         else if (rqst->rq_rcv_buf.page_len == 0)
 407                 wtype = rpcrdma_replych;
 408         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 409                 wtype = rpcrdma_writech;
 410         else
 411                 wtype = rpcrdma_replych;
 412
 413         /*
 414          * Chunks needed for arguments?
 415          *
 416          * o If the total request is under the inline threshold, all ops
 417          *   are sent as inline.
 418          * o Large non-write ops are sent with the entire message as a
 419          *   single read chunk (protocol 0-position special case).
 420          * o Large write ops transmit data as read chunk(s), header as
 421          *   inline.
 422          *
 423          * Note: the NFS code sending down multiple argument segments
 424          * implies the op is a write.
 425          * TBD check NFSv4 setacl
 426          */
 427         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 428                 rtype = rpcrdma_noch;
 429         else if (rqst->rq_snd_buf.page_len == 0)
 430                 rtype = rpcrdma_areadch;
 431         else
 432                 rtype = rpcrdma_readch;
 433
 434         /* The following simplification is not true forever */
 435         if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
 436                 wtype = rpcrdma_noch;
 437         BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch);
 438
 439         if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS &&
 440             (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) {
 441                 /* forced to "pure inline"? */
 442                 dprintk("RPC:       %s: too much data (%d/%d) for inline\n",
 443                         __func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len);
 444                 return -1;
 445         }
 446
 447         hdrlen = 28; /*sizeof *headerp;*/
 448         padlen = 0;
 449
 450         /*
 451          * Pull up any extra send data into the preregistered buffer.
 452          * When padding is in use and applies to the transfer, insert
 453          * it and change the message type.
 454          */
 455         if (rtype == rpcrdma_noch) {
 456
 457                 padlen = rpcrdma_inline_pullup(rqst,
 458                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 459
 460                 if (padlen) {
 461                         headerp->rm_type = htonl(RDMA_MSGP);
 462                         headerp->rm_body.rm_padded.rm_align =
 463                                 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
 464                         headerp->rm_body.rm_padded.rm_thresh =
 465                                 htonl(RPCRDMA_INLINE_PAD_THRESH);
 466                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 467                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 468                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 469                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 470                         BUG_ON(wtype != rpcrdma_noch);
 471
 472                 } else {
 473                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 474                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 475                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 476                         /* new length after pullup */
 477                         rpclen = rqst->rq_svec[0].iov_len;
 478                         /*
 479                          * Currently we try to not actually use read inline.
 480                          * Reply chunks have the desirable property that
 481                          * they land, packed, directly in the target buffers
 482                          * without headers, so they require no fixup. The
 483                          * additional RDMA Write op sends the same amount
 484                          * of data, streams on-the-wire and adds no overhead
 485                          * on receive. Therefore, we request a reply chunk
 486                          * for non-writes wherever feasible and efficient.
 487                          */
 488                         if (wtype == rpcrdma_noch &&
 489                             r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
 490                                 wtype = rpcrdma_replych;
 491                 }
 492         }
 493
 494         /*
 495          * Marshal chunks. This routine will return the header length
 496          * consumed by marshaling.
 497          */
 498         if (rtype != rpcrdma_noch) {
 499                 hdrlen = rpcrdma_create_chunks(rqst,
 500                                         &rqst->rq_snd_buf, headerp, rtype);
 501                 wtype = rtype;  /* simplify dprintk */
 502
 503         } else if (wtype != rpcrdma_noch) {
 504                 hdrlen = rpcrdma_create_chunks(rqst,
 505                                         &rqst->rq_rcv_buf, headerp, wtype);
 506         }
 507
 508         if (hdrlen == 0)
 509                 return -1;
 510
 511         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd\n"
 512                 "                   headerp 0x%p base 0x%p lkey 0x%x\n",
 513                 __func__, transfertypes[wtype], hdrlen, rpclen, padlen,
 514                 headerp, base, req->rl_iov.lkey);
 515
 516         /*
 517          * initialize send_iov's - normally only two: rdma chunk header and
 518          * single preregistered RPC header buffer, but if padding is present,
 519          * then use a preregistered (and zeroed) pad buffer between the RPC
 520          * header and any write data. In all non-rdma cases, any following
 521          * data has been copied into the RPC header buffer.
 522          */
 523         req->rl_send_iov[0].addr = req->rl_iov.addr;
 524         req->rl_send_iov[0].length = hdrlen;
 525         req->rl_send_iov[0].lkey = req->rl_iov.lkey;
 526
 527         req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
 528         req->rl_send_iov[1].length = rpclen;
 529         req->rl_send_iov[1].lkey = req->rl_iov.lkey;
 530
 531         req->rl_niovs = 2;
 532
 533         if (padlen) {
 534                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 535
 536                 req->rl_send_iov[2].addr = ep->rep_pad.addr;
 537                 req->rl_send_iov[2].length = padlen;
 538                 req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
 539
 540                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 541                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 542                 req->rl_send_iov[3].lkey = req->rl_iov.lkey;
 543
 544                 req->rl_niovs = 4;
 545         }
 546
 547         return 0;
 548 }
 549
 550 /*
 551  * Chase down a received write or reply chunklist to get length
 552  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 553  */
 554 static int
 555 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 556 {
 557         unsigned int i, total_len;
 558         struct rpcrdma_write_chunk *cur_wchunk;
 559
 560         i = ntohl(**iptrp);     /* get array count */
 561         if (i > max)
 562                 return -1;
 563         cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 564         total_len = 0;
 565         while (i--) {
 566                 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 567                 ifdebug(FACILITY) {
 568                         u64 off;
 569                         xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 570                         dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
 571                                 __func__,
 572                                 ntohl(seg->rs_length),
 573                                 (unsigned long long)off,
 574                                 ntohl(seg->rs_handle));
 575                 }
 576                 total_len += ntohl(seg->rs_length);
 577                 ++cur_wchunk;
 578         }
 579         /* check and adjust for properly terminated write chunk */
 580         if (wrchunk) {
 581                 __be32 *w = (__be32 *) cur_wchunk;
 582                 if (*w++ != xdr_zero)
 583                         return -1;
 584                 cur_wchunk = (struct rpcrdma_write_chunk *) w;
 585         }
 586         if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
 587                 return -1;
 588
 589         *iptrp = (__be32 *) cur_wchunk;
 590         return total_len;
 591 }
 592
 593 /*
 594  * Scatter inline received data back into provided iov's.
 595  */
 596 static void
 597 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len)
 598 {
 599         int i, npages, curlen, olen;
 600         char *destp;
 601
 602         curlen = rqst->rq_rcv_buf.head[0].iov_len;
 603         if (curlen > copy_len) {        /* write chunk header fixup */
 604                 curlen = copy_len;
 605                 rqst->rq_rcv_buf.head[0].iov_len = curlen;
 606         }
 607
 608         dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 609                 __func__, srcp, copy_len, curlen);
 610
 611         /* Shift pointer for first receive segment only */
 612         rqst->rq_rcv_buf.head[0].iov_base = srcp;
 613         srcp += curlen;
 614         copy_len -= curlen;
 615
 616         olen = copy_len;
 617         i = 0;
 618         rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
 619         if (copy_len && rqst->rq_rcv_buf.page_len) {
 620                 npages = PAGE_ALIGN(rqst->rq_rcv_buf.page_base +
 621                         rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
 622                 for (; i < npages; i++) {
 623                         if (i == 0)
 624                                 curlen = PAGE_SIZE - rqst->rq_rcv_buf.page_base;
 625                         else
 626                                 curlen = PAGE_SIZE;
 627                         if (curlen > copy_len)
 628                                 curlen = copy_len;
 629                         dprintk("RPC:       %s: page %d"
 630                                 " srcp 0x%p len %d curlen %d\n",
 631                                 __func__, i, srcp, copy_len, curlen);
 632                         destp = kmap_atomic(rqst->rq_rcv_buf.pages[i],
 633                                                 KM_SKB_SUNRPC_DATA);
 634                         if (i == 0)
 635                                 memcpy(destp + rqst->rq_rcv_buf.page_base,
 636                                                 srcp, curlen);
 637                         else
 638                                 memcpy(destp, srcp, curlen);
 639                         flush_dcache_page(rqst->rq_rcv_buf.pages[i]);
 640                         kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
 641                         srcp += curlen;
 642                         copy_len -= curlen;
 643                         if (copy_len == 0)
 644                                 break;
 645                 }
 646                 rqst->rq_rcv_buf.page_len = olen - copy_len;
 647         } else
 648                 rqst->rq_rcv_buf.page_len = 0;
 649
 650         if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
 651                 curlen = copy_len;
 652                 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
 653                         curlen = rqst->rq_rcv_buf.tail[0].iov_len;
 654                 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
 655                         memcpy(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
 656                 dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
 657                         __func__, srcp, copy_len, curlen);
 658                 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
 659                 copy_len -= curlen; ++i;
 660         } else
 661                 rqst->rq_rcv_buf.tail[0].iov_len = 0;
 662
 663         if (copy_len)
 664                 dprintk("RPC:       %s: %d bytes in"
 665                         " %d extra segments (%d lost)\n",
 666                         __func__, olen, i, copy_len);
 667
 668         /* TBD avoid a warning from call_decode() */
 669         rqst->rq_private_buf = rqst->rq_rcv_buf;
 670 }
 671
 672 /*
 673  * This function is called when an async event is posted to
 674  * the connection which changes the connection state. All it
 675  * does at this point is mark the connection up/down, the rpc
 676  * timers do the rest.
 677  */
 678 void
 679 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 680 {
 681         struct rpc_xprt *xprt = ep->rep_xprt;
 682
 683         spin_lock_bh(&xprt->transport_lock);
 684         if (ep->rep_connected > 0) {
 685                 if (!xprt_test_and_set_connected(xprt))
 686                         xprt_wake_pending_tasks(xprt, 0);
 687         } else {
 688                 if (xprt_test_and_clear_connected(xprt))
 689                         xprt_wake_pending_tasks(xprt, ep->rep_connected);
 690         }
 691         spin_unlock_bh(&xprt->transport_lock);
 692 }
 693
 694 /*
 695  * This function is called when memory window unbind which we are waiting
 696  * for completes. Just use rr_func (zeroed by upcall) to signal completion.
 697  */
 698 static void
 699 rpcrdma_unbind_func(struct rpcrdma_rep *rep)
 700 {
 701         wake_up(&rep->rr_unbind);
 702 }
 703
 704 /*
 705  * Called as a tasklet to do req/reply match and complete a request
 706  * Errors must result in the RPC task either being awakened, or
 707  * allowed to timeout, to discover the errors at that time.
 708  */
 709 void
 710 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 711 {
 712         struct rpcrdma_msg *headerp;
 713         struct rpcrdma_req *req;
 714         struct rpc_rqst *rqst;
 715         struct rpc_xprt *xprt = rep->rr_xprt;
 716         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 717         __be32 *iptr;
 718         int i, rdmalen, status;
 719
 720         /* Check status. If bad, signal disconnect and return rep to pool */
 721         if (rep->rr_len == ~0U) {
 722                 rpcrdma_recv_buffer_put(rep);
 723                 if (r_xprt->rx_ep.rep_connected == 1) {
 724                         r_xprt->rx_ep.rep_connected = -EIO;
 725                         rpcrdma_conn_func(&r_xprt->rx_ep);
 726                 }
 727                 return;
 728         }
 729         if (rep->rr_len < 28) {
 730                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 731                 goto repost;
 732         }
 733         headerp = (struct rpcrdma_msg *) rep->rr_base;
 734         if (headerp->rm_vers != xdr_one) {
 735                 dprintk("RPC:       %s: invalid version %d\n",
 736                         __func__, ntohl(headerp->rm_vers));
 737                 goto repost;
 738         }
 739
 740         /* Get XID and try for a match. */
 741         spin_lock(&xprt->transport_lock);
 742         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 743         if (rqst == NULL) {
 744                 spin_unlock(&xprt->transport_lock);
 745                 dprintk("RPC:       %s: reply 0x%p failed "
 746                         "to match any request xid 0x%08x len %d\n",
 747                         __func__, rep, headerp->rm_xid, rep->rr_len);
 748 repost:
 749                 r_xprt->rx_stats.bad_reply_count++;
 750                 rep->rr_func = rpcrdma_reply_handler;
 751                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 752                         rpcrdma_recv_buffer_put(rep);
 753
 754                 return;
 755         }
 756
 757         /* get request object */
 758         req = rpcr_to_rdmar(rqst);
 759
 760         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 761                 "                   RPC request 0x%p xid 0x%08x\n",
 762                         __func__, rep, req, rqst, headerp->rm_xid);
 763
 764         BUG_ON(!req || req->rl_reply);
 765
 766         /* from here on, the reply is no longer an orphan */
 767         req->rl_reply = rep;
 768
 769         /* check for expected message types */
 770         /* The order of some of these tests is important. */
 771         switch (headerp->rm_type) {
 772         case __constant_htonl(RDMA_MSG):
 773                 /* never expect read chunks */
 774                 /* never expect reply chunks (two ways to check) */
 775                 /* never expect write chunks without having offered RDMA */
 776                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 777                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 778                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 779                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 780                      req->rl_nchunks == 0))
 781                         goto badheader;
 782                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 783                         /* count any expected write chunks in read reply */
 784                         /* start at write chunk array count */
 785                         iptr = &headerp->rm_body.rm_chunks[2];
 786                         rdmalen = rpcrdma_count_chunks(rep,
 787                                                 req->rl_nchunks, 1, &iptr);
 788                         /* check for validity, and no reply chunk after */
 789                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 790                                 goto badheader;
 791                         rep->rr_len -=
 792                             ((unsigned char *)iptr - (unsigned char *)headerp);
 793                         status = rep->rr_len + rdmalen;
 794                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 795                 } else {
 796                         /* else ordinary inline */
 797                         iptr = (__be32 *)((unsigned char *)headerp + 28);
 798                         rep->rr_len -= 28; /*sizeof *headerp;*/
 799                         status = rep->rr_len;
 800                 }
 801                 /* Fix up the rpc results for upper layer */
 802                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len);
 803                 break;
 804
 805         case __constant_htonl(RDMA_NOMSG):
 806                 /* never expect read or write chunks, always reply chunks */
 807                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 808                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 809                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 810                     req->rl_nchunks == 0)
 811                         goto badheader;
 812                 iptr = (__be32 *)((unsigned char *)headerp + 28);
 813                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 814                 if (rdmalen < 0)
 815                         goto badheader;
 816                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 817                 /* Reply chunk buffer already is the reply vector - no fixup. */
 818                 status = rdmalen;
 819                 break;
 820
 821 badheader:
 822         default:
 823                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 824                                 " chunks[012] == %d %d %d"
 825                                 " expected chunks <= %d\n",
 826                                 __func__, ntohl(headerp->rm_type),
 827                                 headerp->rm_body.rm_chunks[0],
 828                                 headerp->rm_body.rm_chunks[1],
 829                                 headerp->rm_body.rm_chunks[2],
 830                                 req->rl_nchunks);
 831                 status = -EIO;
 832                 r_xprt->rx_stats.bad_reply_count++;
 833                 break;
 834         }
 835
 836         /* If using mw bind, start the deregister process now. */
 837         /* (Note: if mr_free(), cannot perform it here, in tasklet context) */
 838         if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
 839         case RPCRDMA_MEMWINDOWS:
 840                 for (i = 0; req->rl_nchunks-- > 1;)
 841                         i += rpcrdma_deregister_external(
 842                                 &req->rl_segments[i], r_xprt, NULL);
 843                 /* Optionally wait (not here) for unbinds to complete */
 844                 rep->rr_func = rpcrdma_unbind_func;
 845                 (void) rpcrdma_deregister_external(&req->rl_segments[i],
 846                                                    r_xprt, rep);
 847                 break;
 848         case RPCRDMA_MEMWINDOWS_ASYNC:
 849                 for (i = 0; req->rl_nchunks--;)
 850                         i += rpcrdma_deregister_external(&req->rl_segments[i],
 851                                                          r_xprt, NULL);
 852                 break;
 853         default:
 854                 break;
 855         }
 856
 857         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 858                         __func__, xprt, rqst, status);
 859         xprt_complete_rqst(rqst->rq_task, status);
 860         spin_unlock(&xprt->transport_lock);
 861 }