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                 /* the rpcrdma protocol allows us to omit any trailing
 122                  * xdr pad bytes, saving the server an RDMA operation. */
 123                 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
 124                         return n;
 125                 if (n == nsegs)
 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
 169 static unsigned int
 170 rpcrdma_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
 278 out:
 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  */
 292 static int
 293 rpcrdma_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
 300         destp = rqst->rq_svec[0].iov_base;
 301         curlen = rqst->rq_svec[0].iov_len;
 302         destp += curlen;
 303         /*
 304          * Do optional padding where it makes sense. Alignment of write
 305          * payload can help the server, if our setting is accurate.
 306          */
 307         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 308         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 309                 pad = 0;        /* don't pad this request */
 310
 311         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 312                 __func__, pad, destp, rqst->rq_slen, curlen);
 313
 314         copy_len = rqst->rq_snd_buf.page_len;
 315
 316         if (rqst->rq_snd_buf.tail[0].iov_len) {
 317                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 318                 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
 319                         memmove(destp + copy_len,
 320                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 321                         r_xprt->rx_stats.pullup_copy_count += curlen;
 322                 }
 323                 dprintk("RPC:       %s: tail destp 0x%p len %d\n",
 324                         __func__, destp + copy_len, curlen);
 325                 rqst->rq_svec[0].iov_len += curlen;
 326         }
 327
 328         r_xprt->rx_stats.pullup_copy_count += copy_len;
 329         npages = PAGE_ALIGN(rqst->rq_snd_buf.page_base+copy_len) >> PAGE_SHIFT;
 330         for (i = 0; copy_len && i < npages; i++) {
 331                 if (i == 0)
 332                         curlen = PAGE_SIZE - rqst->rq_snd_buf.page_base;
 333                 else
 334                         curlen = PAGE_SIZE;
 335                 if (curlen > copy_len)
 336                         curlen = copy_len;
 337                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 338                         __func__, i, destp, copy_len, curlen);
 339                 srcp = kmap_atomic(rqst->rq_snd_buf.pages[i],
 340                                         KM_SKB_SUNRPC_DATA);
 341                 if (i == 0)
 342                         memcpy(destp, srcp+rqst->rq_snd_buf.page_base, curlen);
 343                 else
 344                         memcpy(destp, srcp, curlen);
 345                 kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
 346                 rqst->rq_svec[0].iov_len += curlen;
 347                 destp += curlen;
 348                 copy_len -= curlen;
 349         }
 350         /* header now contains entire send message */
 351         return pad;
 352 }
 353
 354 /*
 355  * Marshal a request: the primary job of this routine is to choose
 356  * the transfer modes. See comments below.
 357  *
 358  * Uses multiple RDMA IOVs for a request:
 359  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 360  *         preregistered buffer that already holds the RPC data in
 361  *         its middle.
 362  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 363  *  [2] -- optional padding.
 364  *  [3] -- if padded, header only in [1] and data here.
 365  */
 366
 367 int
 368 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 369 {
 370         struct rpc_xprt *xprt = rqst->rq_task->tk_xprt;
 371         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 372         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 373         char *base;
 374         size_t hdrlen, rpclen, padlen;
 375         enum rpcrdma_chunktype rtype, wtype;
 376         struct rpcrdma_msg *headerp;
 377
 378         /*
 379          * rpclen gets amount of data in first buffer, which is the
 380          * pre-registered buffer.
 381          */
 382         base = rqst->rq_svec[0].iov_base;
 383         rpclen = rqst->rq_svec[0].iov_len;
 384
 385         /* build RDMA header in private area at front */
 386         headerp = (struct rpcrdma_msg *) req->rl_base;
 387         /* don't htonl XID, it's already done in request */
 388         headerp->rm_xid = rqst->rq_xid;
 389         headerp->rm_vers = xdr_one;
 390         headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
 391         headerp->rm_type = htonl(RDMA_MSG);
 392
 393         /*
 394          * Chunks needed for results?
 395          *
 396          * o If the expected result is under the inline threshold, all ops
 397          *   return as inline (but see later).
 398          * o Large non-read ops return as a single reply chunk.
 399          * o Large read ops return data as write chunk(s), header as inline.
 400          *
 401          * Note: the NFS code sending down multiple result segments implies
 402          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 403          */
 404
 405         /*
 406          * This code can handle read chunks, write chunks OR reply
 407          * chunks -- only one type. If the request is too big to fit
 408          * inline, then we will choose read chunks. If the request is
 409          * a READ, then use write chunks to separate the file data
 410          * into pages; otherwise use reply chunks.
 411          */
 412         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 413                 wtype = rpcrdma_noch;
 414         else if (rqst->rq_rcv_buf.page_len == 0)
 415                 wtype = rpcrdma_replych;
 416         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 417                 wtype = rpcrdma_writech;
 418         else
 419                 wtype = rpcrdma_replych;
 420
 421         /*
 422          * Chunks needed for arguments?
 423          *
 424          * o If the total request is under the inline threshold, all ops
 425          *   are sent as inline.
 426          * o Large non-write ops are sent with the entire message as a
 427          *   single read chunk (protocol 0-position special case).
 428          * o Large write ops transmit data as read chunk(s), header as
 429          *   inline.
 430          *
 431          * Note: the NFS code sending down multiple argument segments
 432          * implies the op is a write.
 433          * TBD check NFSv4 setacl
 434          */
 435         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 436                 rtype = rpcrdma_noch;
 437         else if (rqst->rq_snd_buf.page_len == 0)
 438                 rtype = rpcrdma_areadch;
 439         else
 440                 rtype = rpcrdma_readch;
 441
 442         /* The following simplification is not true forever */
 443         if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
 444                 wtype = rpcrdma_noch;
 445         BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch);
 446
 447         if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS &&
 448             (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) {
 449                 /* forced to "pure inline"? */
 450                 dprintk("RPC:       %s: too much data (%d/%d) for inline\n",
 451                         __func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len);
 452                 return -1;
 453         }
 454
 455         hdrlen = 28; /*sizeof *headerp;*/
 456         padlen = 0;
 457
 458         /*
 459          * Pull up any extra send data into the preregistered buffer.
 460          * When padding is in use and applies to the transfer, insert
 461          * it and change the message type.
 462          */
 463         if (rtype == rpcrdma_noch) {
 464
 465                 padlen = rpcrdma_inline_pullup(rqst,
 466                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 467
 468                 if (padlen) {
 469                         headerp->rm_type = htonl(RDMA_MSGP);
 470                         headerp->rm_body.rm_padded.rm_align =
 471                                 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
 472                         headerp->rm_body.rm_padded.rm_thresh =
 473                                 htonl(RPCRDMA_INLINE_PAD_THRESH);
 474                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 475                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 476                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 477                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 478                         BUG_ON(wtype != rpcrdma_noch);
 479
 480                 } else {
 481                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 482                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 483                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 484                         /* new length after pullup */
 485                         rpclen = rqst->rq_svec[0].iov_len;
 486                         /*
 487                          * Currently we try to not actually use read inline.
 488                          * Reply chunks have the desirable property that
 489                          * they land, packed, directly in the target buffers
 490                          * without headers, so they require no fixup. The
 491                          * additional RDMA Write op sends the same amount
 492                          * of data, streams on-the-wire and adds no overhead
 493                          * on receive. Therefore, we request a reply chunk
 494                          * for non-writes wherever feasible and efficient.
 495                          */
 496                         if (wtype == rpcrdma_noch &&
 497                             r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
 498                                 wtype = rpcrdma_replych;
 499                 }
 500         }
 501
 502         /*
 503          * Marshal chunks. This routine will return the header length
 504          * consumed by marshaling.
 505          */
 506         if (rtype != rpcrdma_noch) {
 507                 hdrlen = rpcrdma_create_chunks(rqst,
 508                                         &rqst->rq_snd_buf, headerp, rtype);
 509                 wtype = rtype;  /* simplify dprintk */
 510
 511         } else if (wtype != rpcrdma_noch) {
 512                 hdrlen = rpcrdma_create_chunks(rqst,
 513                                         &rqst->rq_rcv_buf, headerp, wtype);
 514         }
 515
 516         if (hdrlen == 0)
 517                 return -1;
 518
 519         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 520                 " headerp 0x%p base 0x%p lkey 0x%x\n",
 521                 __func__, transfertypes[wtype], hdrlen, rpclen, padlen,
 522                 headerp, base, req->rl_iov.lkey);
 523
 524         /*
 525          * initialize send_iov's - normally only two: rdma chunk header and
 526          * single preregistered RPC header buffer, but if padding is present,
 527          * then use a preregistered (and zeroed) pad buffer between the RPC
 528          * header and any write data. In all non-rdma cases, any following
 529          * data has been copied into the RPC header buffer.
 530          */
 531         req->rl_send_iov[0].addr = req->rl_iov.addr;
 532         req->rl_send_iov[0].length = hdrlen;
 533         req->rl_send_iov[0].lkey = req->rl_iov.lkey;
 534
 535         req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
 536         req->rl_send_iov[1].length = rpclen;
 537         req->rl_send_iov[1].lkey = req->rl_iov.lkey;
 538
 539         req->rl_niovs = 2;
 540
 541         if (padlen) {
 542                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 543
 544                 req->rl_send_iov[2].addr = ep->rep_pad.addr;
 545                 req->rl_send_iov[2].length = padlen;
 546                 req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
 547
 548                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 549                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 550                 req->rl_send_iov[3].lkey = req->rl_iov.lkey;
 551
 552                 req->rl_niovs = 4;
 553         }
 554
 555         return 0;
 556 }
 557
 558 /*
 559  * Chase down a received write or reply chunklist to get length
 560  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 561  */
 562 static int
 563 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 564 {
 565         unsigned int i, total_len;
 566         struct rpcrdma_write_chunk *cur_wchunk;
 567
 568         i = ntohl(**iptrp);     /* get array count */
 569         if (i > max)
 570                 return -1;
 571         cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 572         total_len = 0;
 573         while (i--) {
 574                 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 575                 ifdebug(FACILITY) {
 576                         u64 off;
 577                         xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 578                         dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
 579                                 __func__,
 580                                 ntohl(seg->rs_length),
 581                                 (unsigned long long)off,
 582                                 ntohl(seg->rs_handle));
 583                 }
 584                 total_len += ntohl(seg->rs_length);
 585                 ++cur_wchunk;
 586         }
 587         /* check and adjust for properly terminated write chunk */
 588         if (wrchunk) {
 589                 __be32 *w = (__be32 *) cur_wchunk;
 590                 if (*w++ != xdr_zero)
 591                         return -1;
 592                 cur_wchunk = (struct rpcrdma_write_chunk *) w;
 593         }
 594         if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
 595                 return -1;
 596
 597         *iptrp = (__be32 *) cur_wchunk;
 598         return total_len;
 599 }
 600
 601 /*
 602  * Scatter inline received data back into provided iov's.
 603  */
 604 static void
 605 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
 606 {
 607         int i, npages, curlen, olen;
 608         char *destp;
 609
 610         curlen = rqst->rq_rcv_buf.head[0].iov_len;
 611         if (curlen > copy_len) {        /* write chunk header fixup */
 612                 curlen = copy_len;
 613                 rqst->rq_rcv_buf.head[0].iov_len = curlen;
 614         }
 615
 616         dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 617                 __func__, srcp, copy_len, curlen);
 618
 619         /* Shift pointer for first receive segment only */
 620         rqst->rq_rcv_buf.head[0].iov_base = srcp;
 621         srcp += curlen;
 622         copy_len -= curlen;
 623
 624         olen = copy_len;
 625         i = 0;
 626         rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
 627         if (copy_len && rqst->rq_rcv_buf.page_len) {
 628                 npages = PAGE_ALIGN(rqst->rq_rcv_buf.page_base +
 629                         rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
 630                 for (; i < npages; i++) {
 631                         if (i == 0)
 632                                 curlen = PAGE_SIZE - rqst->rq_rcv_buf.page_base;
 633                         else
 634                                 curlen = PAGE_SIZE;
 635                         if (curlen > copy_len)
 636                                 curlen = copy_len;
 637                         dprintk("RPC:       %s: page %d"
 638                                 " srcp 0x%p len %d curlen %d\n",
 639                                 __func__, i, srcp, copy_len, curlen);
 640                         destp = kmap_atomic(rqst->rq_rcv_buf.pages[i],
 641                                                 KM_SKB_SUNRPC_DATA);
 642                         if (i == 0)
 643                                 memcpy(destp + rqst->rq_rcv_buf.page_base,
 644                                                 srcp, curlen);
 645                         else
 646                                 memcpy(destp, srcp, curlen);
 647                         flush_dcache_page(rqst->rq_rcv_buf.pages[i]);
 648                         kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
 649                         srcp += curlen;
 650                         copy_len -= curlen;
 651                         if (copy_len == 0)
 652                                 break;
 653                 }
 654                 rqst->rq_rcv_buf.page_len = olen - copy_len;
 655         } else
 656                 rqst->rq_rcv_buf.page_len = 0;
 657
 658         if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
 659                 curlen = copy_len;
 660                 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
 661                         curlen = rqst->rq_rcv_buf.tail[0].iov_len;
 662                 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
 663                         memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
 664                 dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
 665                         __func__, srcp, copy_len, curlen);
 666                 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
 667                 copy_len -= curlen; ++i;
 668         } else
 669                 rqst->rq_rcv_buf.tail[0].iov_len = 0;
 670
 671         if (pad) {
 672                 /* implicit padding on terminal chunk */
 673                 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
 674                 while (pad--)
 675                         p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
 676         }
 677
 678         if (copy_len)
 679                 dprintk("RPC:       %s: %d bytes in"
 680                         " %d extra segments (%d lost)\n",
 681                         __func__, olen, i, copy_len);
 682
 683         /* TBD avoid a warning from call_decode() */
 684         rqst->rq_private_buf = rqst->rq_rcv_buf;
 685 }
 686
 687 /*
 688  * This function is called when an async event is posted to
 689  * the connection which changes the connection state. All it
 690  * does at this point is mark the connection up/down, the rpc
 691  * timers do the rest.
 692  */
 693 void
 694 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 695 {
 696         struct rpc_xprt *xprt = ep->rep_xprt;
 697
 698         spin_lock_bh(&xprt->transport_lock);
 699         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 700                 ++xprt->connect_cookie;
 701         if (ep->rep_connected > 0) {
 702                 if (!xprt_test_and_set_connected(xprt))
 703                         xprt_wake_pending_tasks(xprt, 0);
 704         } else {
 705                 if (xprt_test_and_clear_connected(xprt))
 706                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 707         }
 708         spin_unlock_bh(&xprt->transport_lock);
 709 }
 710
 711 /*
 712  * This function is called when memory window unbind which we are waiting
 713  * for completes. Just use rr_func (zeroed by upcall) to signal completion.
 714  */
 715 static void
 716 rpcrdma_unbind_func(struct rpcrdma_rep *rep)
 717 {
 718         wake_up(&rep->rr_unbind);
 719 }
 720
 721 /*
 722  * Called as a tasklet to do req/reply match and complete a request
 723  * Errors must result in the RPC task either being awakened, or
 724  * allowed to timeout, to discover the errors at that time.
 725  */
 726 void
 727 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 728 {
 729         struct rpcrdma_msg *headerp;
 730         struct rpcrdma_req *req;
 731         struct rpc_rqst *rqst;
 732         struct rpc_xprt *xprt = rep->rr_xprt;
 733         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 734         __be32 *iptr;
 735         int i, rdmalen, status;
 736
 737         /* Check status. If bad, signal disconnect and return rep to pool */
 738         if (rep->rr_len == ~0U) {
 739                 rpcrdma_recv_buffer_put(rep);
 740                 if (r_xprt->rx_ep.rep_connected == 1) {
 741                         r_xprt->rx_ep.rep_connected = -EIO;
 742                         rpcrdma_conn_func(&r_xprt->rx_ep);
 743                 }
 744                 return;
 745         }
 746         if (rep->rr_len < 28) {
 747                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 748                 goto repost;
 749         }
 750         headerp = (struct rpcrdma_msg *) rep->rr_base;
 751         if (headerp->rm_vers != xdr_one) {
 752                 dprintk("RPC:       %s: invalid version %d\n",
 753                         __func__, ntohl(headerp->rm_vers));
 754                 goto repost;
 755         }
 756
 757         /* Get XID and try for a match. */
 758         spin_lock(&xprt->transport_lock);
 759         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 760         if (rqst == NULL) {
 761                 spin_unlock(&xprt->transport_lock);
 762                 dprintk("RPC:       %s: reply 0x%p failed "
 763                         "to match any request xid 0x%08x len %d\n",
 764                         __func__, rep, headerp->rm_xid, rep->rr_len);
 765 repost:
 766                 r_xprt->rx_stats.bad_reply_count++;
 767                 rep->rr_func = rpcrdma_reply_handler;
 768                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 769                         rpcrdma_recv_buffer_put(rep);
 770
 771                 return;
 772         }
 773
 774         /* get request object */
 775         req = rpcr_to_rdmar(rqst);
 776
 777         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 778                 "                   RPC request 0x%p xid 0x%08x\n",
 779                         __func__, rep, req, rqst, headerp->rm_xid);
 780
 781         BUG_ON(!req || req->rl_reply);
 782
 783         /* from here on, the reply is no longer an orphan */
 784         req->rl_reply = rep;
 785
 786         /* check for expected message types */
 787         /* The order of some of these tests is important. */
 788         switch (headerp->rm_type) {
 789         case htonl(RDMA_MSG):
 790                 /* never expect read chunks */
 791                 /* never expect reply chunks (two ways to check) */
 792                 /* never expect write chunks without having offered RDMA */
 793                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 794                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 795                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 796                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 797                      req->rl_nchunks == 0))
 798                         goto badheader;
 799                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 800                         /* count any expected write chunks in read reply */
 801                         /* start at write chunk array count */
 802                         iptr = &headerp->rm_body.rm_chunks[2];
 803                         rdmalen = rpcrdma_count_chunks(rep,
 804                                                 req->rl_nchunks, 1, &iptr);
 805                         /* check for validity, and no reply chunk after */
 806                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 807                                 goto badheader;
 808                         rep->rr_len -=
 809                             ((unsigned char *)iptr - (unsigned char *)headerp);
 810                         status = rep->rr_len + rdmalen;
 811                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 812                         /* special case - last chunk may omit padding */
 813                         if (rdmalen &= 3) {
 814                                 rdmalen = 4 - rdmalen;
 815                                 status += rdmalen;
 816                         }
 817                 } else {
 818                         /* else ordinary inline */
 819                         rdmalen = 0;
 820                         iptr = (__be32 *)((unsigned char *)headerp + 28);
 821                         rep->rr_len -= 28; /*sizeof *headerp;*/
 822                         status = rep->rr_len;
 823                 }
 824                 /* Fix up the rpc results for upper layer */
 825                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 826                 break;
 827
 828         case htonl(RDMA_NOMSG):
 829                 /* never expect read or write chunks, always reply chunks */
 830                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 831                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 832                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 833                     req->rl_nchunks == 0)
 834                         goto badheader;
 835                 iptr = (__be32 *)((unsigned char *)headerp + 28);
 836                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 837                 if (rdmalen < 0)
 838                         goto badheader;
 839                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 840                 /* Reply chunk buffer already is the reply vector - no fixup. */
 841                 status = rdmalen;
 842                 break;
 843
 844 badheader:
 845         default:
 846                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 847                                 " chunks[012] == %d %d %d"
 848                                 " expected chunks <= %d\n",
 849                                 __func__, ntohl(headerp->rm_type),
 850                                 headerp->rm_body.rm_chunks[0],
 851                                 headerp->rm_body.rm_chunks[1],
 852                                 headerp->rm_body.rm_chunks[2],
 853                                 req->rl_nchunks);
 854                 status = -EIO;
 855                 r_xprt->rx_stats.bad_reply_count++;
 856                 break;
 857         }
 858
 859         /* If using mw bind, start the deregister process now. */
 860         /* (Note: if mr_free(), cannot perform it here, in tasklet context) */
 861         if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
 862         case RPCRDMA_MEMWINDOWS:
 863                 for (i = 0; req->rl_nchunks-- > 1;)
 864                         i += rpcrdma_deregister_external(
 865                                 &req->rl_segments[i], r_xprt, NULL);
 866                 /* Optionally wait (not here) for unbinds to complete */
 867                 rep->rr_func = rpcrdma_unbind_func;
 868                 (void) rpcrdma_deregister_external(&req->rl_segments[i],
 869                                                    r_xprt, rep);
 870                 break;
 871         case RPCRDMA_MEMWINDOWS_ASYNC:
 872                 for (i = 0; req->rl_nchunks--;)
 873                         i += rpcrdma_deregister_external(&req->rl_segments[i],
 874                                                          r_xprt, NULL);
 875                 break;
 876         default:
 877                 break;
 878         }
 879
 880         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 881                         __func__, xprt, rqst, status);
 882         xprt_complete_rqst(rqst->rq_task, status);
 883         spin_unlock(&xprt->transport_lock);
 884 }