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 #ifdef RPC_DEBUG
  57 static const char transfertypes[][12] = {
  58         "pure inline",  /* no chunks */
  59         " read chunk",  /* some argument via rdma read */
  60         "*read chunk",  /* entire request via rdma read */
  61         "write chunk",  /* some result via rdma write */
  62         "reply chunk"   /* entire reply via rdma write */
  63 };
  64 #endif
  65
  66 /*
  67  * Chunk assembly from upper layer xdr_buf.
  68  *
  69  * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
  70  * elements. Segments are then coalesced when registered, if possible
  71  * within the selected memreg mode.
  72  *
  73  * Returns positive number of segments converted, or a negative errno.
  74  */
  75
  76 static int
  77 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
  78         enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
  79 {
  80         int len, n = 0, p;
  81         int page_base;
  82         struct page **ppages;
  83
  84         if (pos == 0 && xdrbuf->head[0].iov_len) {
  85                 seg[n].mr_page = NULL;
  86                 seg[n].mr_offset = xdrbuf->head[0].iov_base;
  87                 seg[n].mr_len = xdrbuf->head[0].iov_len;
  88                 ++n;
  89         }
  90
  91         len = xdrbuf->page_len;
  92         ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
  93         page_base = xdrbuf->page_base & ~PAGE_MASK;
  94         p = 0;
  95         while (len && n < nsegs) {
  96                 if (!ppages[p]) {
  97                         /* alloc the pagelist for receiving buffer */
  98                         ppages[p] = alloc_page(GFP_ATOMIC);
  99                         if (!ppages[p])
 100                                 return -ENOMEM;
 101                 }
 102                 seg[n].mr_page = ppages[p];
 103                 seg[n].mr_offset = (void *)(unsigned long) page_base;
 104                 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
 105                 if (seg[n].mr_len > PAGE_SIZE)
 106                         return -EIO;
 107                 len -= seg[n].mr_len;
 108                 ++n;
 109                 ++p;
 110                 page_base = 0;  /* page offset only applies to first page */
 111         }
 112
 113         /* Message overflows the seg array */
 114         if (len && n == nsegs)
 115                 return -EIO;
 116
 117         if (xdrbuf->tail[0].iov_len) {
 118                 /* the rpcrdma protocol allows us to omit any trailing
 119                  * xdr pad bytes, saving the server an RDMA operation. */
 120                 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
 121                         return n;
 122                 if (n == nsegs)
 123                         /* Tail remains, but we're out of segments */
 124                         return -EIO;
 125                 seg[n].mr_page = NULL;
 126                 seg[n].mr_offset = xdrbuf->tail[0].iov_base;
 127                 seg[n].mr_len = xdrbuf->tail[0].iov_len;
 128                 ++n;
 129         }
 130
 131         return n;
 132 }
 133
 134 /*
 135  * Create read/write chunk lists, and reply chunks, for RDMA
 136  *
 137  *   Assume check against THRESHOLD has been done, and chunks are required.
 138  *   Assume only encoding one list entry for read|write chunks. The NFSv3
 139  *     protocol is simple enough to allow this as it only has a single "bulk
 140  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
 141  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
 142  *
 143  * When used for a single reply chunk (which is a special write
 144  * chunk used for the entire reply, rather than just the data), it
 145  * is used primarily for READDIR and READLINK which would otherwise
 146  * be severely size-limited by a small rdma inline read max. The server
 147  * response will come back as an RDMA Write, followed by a message
 148  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
 149  * chunks do not provide data alignment, however they do not require
 150  * "fixup" (moving the response to the upper layer buffer) either.
 151  *
 152  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 153  *
 154  *  Read chunklist (a linked list):
 155  *   N elements, position P (same P for all chunks of same arg!):
 156  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
 157  *
 158  *  Write chunklist (a list of (one) counted array):
 159  *   N elements:
 160  *    1 - N - HLOO - HLOO - ... - HLOO - 0
 161  *
 162  *  Reply chunk (a counted array):
 163  *   N elements:
 164  *    1 - N - HLOO - HLOO - ... - HLOO
 165  *
 166  * Returns positive RPC/RDMA header size, or negative errno.
 167  */
 168
 169 static ssize_t
 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_xprt);
 175         int n, 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 nsegs;
 203
 204         do {
 205                 n = rpcrdma_register_external(seg, nsegs,
 206                                                 cur_wchunk != NULL, r_xprt);
 207                 if (n <= 0)
 208                         goto out;
 209                 if (cur_rchunk) {       /* read */
 210                         cur_rchunk->rc_discrim = xdr_one;
 211                         /* all read chunks have the same "position" */
 212                         cur_rchunk->rc_position = htonl(pos);
 213                         cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey);
 214                         cur_rchunk->rc_target.rs_length = htonl(seg->mr_len);
 215                         xdr_encode_hyper(
 216                                         (__be32 *)&cur_rchunk->rc_target.rs_offset,
 217                                         seg->mr_base);
 218                         dprintk("RPC:       %s: read chunk "
 219                                 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
 220                                 seg->mr_len, (unsigned long long)seg->mr_base,
 221                                 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
 222                         cur_rchunk++;
 223                         r_xprt->rx_stats.read_chunk_count++;
 224                 } else {                /* write/reply */
 225                         cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey);
 226                         cur_wchunk->wc_target.rs_length = htonl(seg->mr_len);
 227                         xdr_encode_hyper(
 228                                         (__be32 *)&cur_wchunk->wc_target.rs_offset,
 229                                         seg->mr_base);
 230                         dprintk("RPC:       %s: %s chunk "
 231                                 "elem %d@0x%llx:0x%x (%s)\n", __func__,
 232                                 (type == rpcrdma_replych) ? "reply" : "write",
 233                                 seg->mr_len, (unsigned long long)seg->mr_base,
 234                                 seg->mr_rkey, n < nsegs ? "more" : "last");
 235                         cur_wchunk++;
 236                         if (type == rpcrdma_replych)
 237                                 r_xprt->rx_stats.reply_chunk_count++;
 238                         else
 239                                 r_xprt->rx_stats.write_chunk_count++;
 240                         r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 241                 }
 242                 nchunks++;
 243                 seg   += n;
 244                 nsegs -= n;
 245         } while (nsegs);
 246
 247         /* success. all failures return above */
 248         req->rl_nchunks = nchunks;
 249
 250         /*
 251          * finish off header. If write, marshal discrim and nchunks.
 252          */
 253         if (cur_rchunk) {
 254                 iptr = (__be32 *) cur_rchunk;
 255                 *iptr++ = xdr_zero;     /* finish the read chunk list */
 256                 *iptr++ = xdr_zero;     /* encode a NULL write chunk list */
 257                 *iptr++ = xdr_zero;     /* encode a NULL reply chunk */
 258         } else {
 259                 warray->wc_discrim = xdr_one;
 260                 warray->wc_nchunks = htonl(nchunks);
 261                 iptr = (__be32 *) cur_wchunk;
 262                 if (type == rpcrdma_writech) {
 263                         *iptr++ = xdr_zero; /* finish the write chunk list */
 264                         *iptr++ = xdr_zero; /* encode a NULL reply chunk */
 265                 }
 266         }
 267
 268         /*
 269          * Return header size.
 270          */
 271         return (unsigned char *)iptr - (unsigned char *)headerp;
 272
 273 out:
 274         for (pos = 0; nchunks--;)
 275                 pos += rpcrdma_deregister_external(
 276                                 &req->rl_segments[pos], r_xprt);
 277         return n;
 278 }
 279
 280 /*
 281  * Marshal chunks. This routine returns the header length
 282  * consumed by marshaling.
 283  *
 284  * Returns positive RPC/RDMA header size, or negative errno.
 285  */
 286
 287 ssize_t
 288 rpcrdma_marshal_chunks(struct rpc_rqst *rqst, ssize_t result)
 289 {
 290         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 291         struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)req->rl_base;
 292
 293         if (req->rl_rtype != rpcrdma_noch)
 294                 result = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
 295                                                headerp, req->rl_rtype);
 296         else if (req->rl_wtype != rpcrdma_noch)
 297                 result = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
 298                                                headerp, req->rl_wtype);
 299         return result;
 300 }
 301
 302 /*
 303  * Copy write data inline.
 304  * This function is used for "small" requests. Data which is passed
 305  * to RPC via iovecs (or page list) is copied directly into the
 306  * pre-registered memory buffer for this request. For small amounts
 307  * of data, this is efficient. The cutoff value is tunable.
 308  */
 309 static int
 310 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
 311 {
 312         int i, npages, curlen;
 313         int copy_len;
 314         unsigned char *srcp, *destp;
 315         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 316         int page_base;
 317         struct page **ppages;
 318
 319         destp = rqst->rq_svec[0].iov_base;
 320         curlen = rqst->rq_svec[0].iov_len;
 321         destp += curlen;
 322         /*
 323          * Do optional padding where it makes sense. Alignment of write
 324          * payload can help the server, if our setting is accurate.
 325          */
 326         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 327         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 328                 pad = 0;        /* don't pad this request */
 329
 330         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 331                 __func__, pad, destp, rqst->rq_slen, curlen);
 332
 333         copy_len = rqst->rq_snd_buf.page_len;
 334
 335         if (rqst->rq_snd_buf.tail[0].iov_len) {
 336                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 337                 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
 338                         memmove(destp + copy_len,
 339                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 340                         r_xprt->rx_stats.pullup_copy_count += curlen;
 341                 }
 342                 dprintk("RPC:       %s: tail destp 0x%p len %d\n",
 343                         __func__, destp + copy_len, curlen);
 344                 rqst->rq_svec[0].iov_len += curlen;
 345         }
 346         r_xprt->rx_stats.pullup_copy_count += copy_len;
 347
 348         page_base = rqst->rq_snd_buf.page_base;
 349         ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
 350         page_base &= ~PAGE_MASK;
 351         npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
 352         for (i = 0; copy_len && i < npages; i++) {
 353                 curlen = PAGE_SIZE - page_base;
 354                 if (curlen > copy_len)
 355                         curlen = copy_len;
 356                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 357                         __func__, i, destp, copy_len, curlen);
 358                 srcp = kmap_atomic(ppages[i]);
 359                 memcpy(destp, srcp+page_base, curlen);
 360                 kunmap_atomic(srcp);
 361                 rqst->rq_svec[0].iov_len += curlen;
 362                 destp += curlen;
 363                 copy_len -= curlen;
 364                 page_base = 0;
 365         }
 366         /* header now contains entire send message */
 367         return pad;
 368 }
 369
 370 /*
 371  * Marshal a request: the primary job of this routine is to choose
 372  * the transfer modes. See comments below.
 373  *
 374  * Uses multiple RDMA IOVs for a request:
 375  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 376  *         preregistered buffer that already holds the RPC data in
 377  *         its middle.
 378  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 379  *  [2] -- optional padding.
 380  *  [3] -- if padded, header only in [1] and data here.
 381  *
 382  * Returns zero on success, otherwise a negative errno.
 383  */
 384
 385 int
 386 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 387 {
 388         struct rpc_xprt *xprt = rqst->rq_xprt;
 389         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 390         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 391         char *base;
 392         size_t rpclen, padlen;
 393         ssize_t hdrlen;
 394         struct rpcrdma_msg *headerp;
 395
 396         /*
 397          * rpclen gets amount of data in first buffer, which is the
 398          * pre-registered buffer.
 399          */
 400         base = rqst->rq_svec[0].iov_base;
 401         rpclen = rqst->rq_svec[0].iov_len;
 402
 403         /* build RDMA header in private area at front */
 404         headerp = (struct rpcrdma_msg *) req->rl_base;
 405         /* don't htonl XID, it's already done in request */
 406         headerp->rm_xid = rqst->rq_xid;
 407         headerp->rm_vers = xdr_one;
 408         headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
 409         headerp->rm_type = htonl(RDMA_MSG);
 410
 411         /*
 412          * Chunks needed for results?
 413          *
 414          * o If the expected result is under the inline threshold, all ops
 415          *   return as inline (but see later).
 416          * o Large non-read ops return as a single reply chunk.
 417          * o Large read ops return data as write chunk(s), header as inline.
 418          *
 419          * Note: the NFS code sending down multiple result segments implies
 420          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 421          */
 422
 423         /*
 424          * This code can handle read chunks, write chunks OR reply
 425          * chunks -- only one type. If the request is too big to fit
 426          * inline, then we will choose read chunks. If the request is
 427          * a READ, then use write chunks to separate the file data
 428          * into pages; otherwise use reply chunks.
 429          */
 430         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 431                 req->rl_wtype = rpcrdma_noch;
 432         else if (rqst->rq_rcv_buf.page_len == 0)
 433                 req->rl_wtype = rpcrdma_replych;
 434         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 435                 req->rl_wtype = rpcrdma_writech;
 436         else
 437                 req->rl_wtype = rpcrdma_replych;
 438
 439         /*
 440          * Chunks needed for arguments?
 441          *
 442          * o If the total request is under the inline threshold, all ops
 443          *   are sent as inline.
 444          * o Large non-write ops are sent with the entire message as a
 445          *   single read chunk (protocol 0-position special case).
 446          * o Large write ops transmit data as read chunk(s), header as
 447          *   inline.
 448          *
 449          * Note: the NFS code sending down multiple argument segments
 450          * implies the op is a write.
 451          * TBD check NFSv4 setacl
 452          */
 453         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 454                 req->rl_rtype = rpcrdma_noch;
 455         else if (rqst->rq_snd_buf.page_len == 0)
 456                 req->rl_rtype = rpcrdma_areadch;
 457         else
 458                 req->rl_rtype = rpcrdma_readch;
 459
 460         /* The following simplification is not true forever */
 461         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype == rpcrdma_replych)
 462                 req->rl_wtype = rpcrdma_noch;
 463         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype != rpcrdma_noch) {
 464                 dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
 465                         __func__);
 466                 return -EIO;
 467         }
 468
 469         hdrlen = 28; /*sizeof *headerp;*/
 470         padlen = 0;
 471
 472         /*
 473          * Pull up any extra send data into the preregistered buffer.
 474          * When padding is in use and applies to the transfer, insert
 475          * it and change the message type.
 476          */
 477         if (req->rl_rtype == rpcrdma_noch) {
 478
 479                 padlen = rpcrdma_inline_pullup(rqst,
 480                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 481
 482                 if (padlen) {
 483                         headerp->rm_type = htonl(RDMA_MSGP);
 484                         headerp->rm_body.rm_padded.rm_align =
 485                                 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
 486                         headerp->rm_body.rm_padded.rm_thresh =
 487                                 htonl(RPCRDMA_INLINE_PAD_THRESH);
 488                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 489                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 490                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 491                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 492                         if (req->rl_wtype != rpcrdma_noch) {
 493                                 dprintk("RPC:       %s: invalid chunk list\n",
 494                                         __func__);
 495                                 return -EIO;
 496                         }
 497                 } else {
 498                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 499                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 500                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 501                         /* new length after pullup */
 502                         rpclen = rqst->rq_svec[0].iov_len;
 503                         /*
 504                          * Currently we try to not actually use read inline.
 505                          * Reply chunks have the desirable property that
 506                          * they land, packed, directly in the target buffers
 507                          * without headers, so they require no fixup. The
 508                          * additional RDMA Write op sends the same amount
 509                          * of data, streams on-the-wire and adds no overhead
 510                          * on receive. Therefore, we request a reply chunk
 511                          * for non-writes wherever feasible and efficient.
 512                          */
 513                         if (req->rl_wtype == rpcrdma_noch)
 514                                 req->rl_wtype = rpcrdma_replych;
 515                 }
 516         }
 517
 518         hdrlen = rpcrdma_marshal_chunks(rqst, hdrlen);
 519         if (hdrlen < 0)
 520                 return hdrlen;
 521
 522         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 523                 " headerp 0x%p base 0x%p lkey 0x%x\n",
 524                 __func__, transfertypes[req->rl_wtype], hdrlen, rpclen, padlen,
 525                 headerp, base, req->rl_iov.lkey);
 526
 527         /*
 528          * initialize send_iov's - normally only two: rdma chunk header and
 529          * single preregistered RPC header buffer, but if padding is present,
 530          * then use a preregistered (and zeroed) pad buffer between the RPC
 531          * header and any write data. In all non-rdma cases, any following
 532          * data has been copied into the RPC header buffer.
 533          */
 534         req->rl_send_iov[0].addr = req->rl_iov.addr;
 535         req->rl_send_iov[0].length = hdrlen;
 536         req->rl_send_iov[0].lkey = req->rl_iov.lkey;
 537
 538         req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
 539         req->rl_send_iov[1].length = rpclen;
 540         req->rl_send_iov[1].lkey = req->rl_iov.lkey;
 541
 542         req->rl_niovs = 2;
 543
 544         if (padlen) {
 545                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 546
 547                 req->rl_send_iov[2].addr = ep->rep_pad.addr;
 548                 req->rl_send_iov[2].length = padlen;
 549                 req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
 550
 551                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 552                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 553                 req->rl_send_iov[3].lkey = req->rl_iov.lkey;
 554
 555                 req->rl_niovs = 4;
 556         }
 557
 558         return 0;
 559 }
 560
 561 /*
 562  * Chase down a received write or reply chunklist to get length
 563  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 564  */
 565 static int
 566 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 567 {
 568         unsigned int i, total_len;
 569         struct rpcrdma_write_chunk *cur_wchunk;
 570
 571         i = ntohl(**iptrp);     /* get array count */
 572         if (i > max)
 573                 return -1;
 574         cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 575         total_len = 0;
 576         while (i--) {
 577                 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 578                 ifdebug(FACILITY) {
 579                         u64 off;
 580                         xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 581                         dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
 582                                 __func__,
 583                                 ntohl(seg->rs_length),
 584                                 (unsigned long long)off,
 585                                 ntohl(seg->rs_handle));
 586                 }
 587                 total_len += ntohl(seg->rs_length);
 588                 ++cur_wchunk;
 589         }
 590         /* check and adjust for properly terminated write chunk */
 591         if (wrchunk) {
 592                 __be32 *w = (__be32 *) cur_wchunk;
 593                 if (*w++ != xdr_zero)
 594                         return -1;
 595                 cur_wchunk = (struct rpcrdma_write_chunk *) w;
 596         }
 597         if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
 598                 return -1;
 599
 600         *iptrp = (__be32 *) cur_wchunk;
 601         return total_len;
 602 }
 603
 604 /*
 605  * Scatter inline received data back into provided iov's.
 606  */
 607 static void
 608 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
 609 {
 610         int i, npages, curlen, olen;
 611         char *destp;
 612         struct page **ppages;
 613         int page_base;
 614
 615         curlen = rqst->rq_rcv_buf.head[0].iov_len;
 616         if (curlen > copy_len) {        /* write chunk header fixup */
 617                 curlen = copy_len;
 618                 rqst->rq_rcv_buf.head[0].iov_len = curlen;
 619         }
 620
 621         dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 622                 __func__, srcp, copy_len, curlen);
 623
 624         /* Shift pointer for first receive segment only */
 625         rqst->rq_rcv_buf.head[0].iov_base = srcp;
 626         srcp += curlen;
 627         copy_len -= curlen;
 628
 629         olen = copy_len;
 630         i = 0;
 631         rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
 632         page_base = rqst->rq_rcv_buf.page_base;
 633         ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
 634         page_base &= ~PAGE_MASK;
 635
 636         if (copy_len && rqst->rq_rcv_buf.page_len) {
 637                 npages = PAGE_ALIGN(page_base +
 638                         rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
 639                 for (; i < npages; i++) {
 640                         curlen = PAGE_SIZE - page_base;
 641                         if (curlen > copy_len)
 642                                 curlen = copy_len;
 643                         dprintk("RPC:       %s: page %d"
 644                                 " srcp 0x%p len %d curlen %d\n",
 645                                 __func__, i, srcp, copy_len, curlen);
 646                         destp = kmap_atomic(ppages[i]);
 647                         memcpy(destp + page_base, srcp, curlen);
 648                         flush_dcache_page(ppages[i]);
 649                         kunmap_atomic(destp);
 650                         srcp += curlen;
 651                         copy_len -= curlen;
 652                         if (copy_len == 0)
 653                                 break;
 654                         page_base = 0;
 655                 }
 656         }
 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 void
 688 rpcrdma_connect_worker(struct work_struct *work)
 689 {
 690         struct rpcrdma_ep *ep =
 691                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 692         struct rpc_xprt *xprt = ep->rep_xprt;
 693
 694         spin_lock_bh(&xprt->transport_lock);
 695         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 696                 ++xprt->connect_cookie;
 697         if (ep->rep_connected > 0) {
 698                 if (!xprt_test_and_set_connected(xprt))
 699                         xprt_wake_pending_tasks(xprt, 0);
 700         } else {
 701                 if (xprt_test_and_clear_connected(xprt))
 702                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 703         }
 704         spin_unlock_bh(&xprt->transport_lock);
 705 }
 706
 707 /*
 708  * This function is called when an async event is posted to
 709  * the connection which changes the connection state. All it
 710  * does at this point is mark the connection up/down, the rpc
 711  * timers do the rest.
 712  */
 713 void
 714 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 715 {
 716         schedule_delayed_work(&ep->rep_connect_worker, 0);
 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  */
 724 void
 725 rpcrdma_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 rdmalen, status;
 734         unsigned long cwnd;
 735
 736         /* Check status. If bad, signal disconnect and return rep to pool */
 737         if (rep->rr_len == ~0U) {
 738                 rpcrdma_recv_buffer_put(rep);
 739                 if (r_xprt->rx_ep.rep_connected == 1) {
 740                         r_xprt->rx_ep.rep_connected = -EIO;
 741                         rpcrdma_conn_func(&r_xprt->rx_ep);
 742                 }
 743                 return;
 744         }
 745         if (rep->rr_len < 28) {
 746                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 747                 goto repost;
 748         }
 749         headerp = (struct rpcrdma_msg *) rep->rr_base;
 750         if (headerp->rm_vers != xdr_one) {
 751                 dprintk("RPC:       %s: invalid version %d\n",
 752                         __func__, ntohl(headerp->rm_vers));
 753                 goto repost;
 754         }
 755
 756         /* Get XID and try for a match. */
 757         spin_lock(&xprt->transport_lock);
 758         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 759         if (rqst == NULL) {
 760                 spin_unlock(&xprt->transport_lock);
 761                 dprintk("RPC:       %s: reply 0x%p failed "
 762                         "to match any request xid 0x%08x len %d\n",
 763                         __func__, rep, headerp->rm_xid, rep->rr_len);
 764 repost:
 765                 r_xprt->rx_stats.bad_reply_count++;
 766                 rep->rr_func = rpcrdma_reply_handler;
 767                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 768                         rpcrdma_recv_buffer_put(rep);
 769
 770                 return;
 771         }
 772
 773         /* get request object */
 774         req = rpcr_to_rdmar(rqst);
 775         if (req->rl_reply) {
 776                 spin_unlock(&xprt->transport_lock);
 777                 dprintk("RPC:       %s: duplicate reply 0x%p to RPC "
 778                         "request 0x%p: xid 0x%08x\n", __func__, rep, req,
 779                         headerp->rm_xid);
 780                 goto repost;
 781         }
 782
 783         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 784                 "                   RPC request 0x%p xid 0x%08x\n",
 785                         __func__, rep, req, rqst, headerp->rm_xid);
 786
 787         /* from here on, the reply is no longer an orphan */
 788         req->rl_reply = rep;
 789         xprt->reestablish_timeout = 0;
 790
 791         /* check for expected message types */
 792         /* The order of some of these tests is important. */
 793         switch (headerp->rm_type) {
 794         case htonl(RDMA_MSG):
 795                 /* never expect read chunks */
 796                 /* never expect reply chunks (two ways to check) */
 797                 /* never expect write chunks without having offered RDMA */
 798                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 799                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 800                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 801                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 802                      req->rl_nchunks == 0))
 803                         goto badheader;
 804                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 805                         /* count any expected write chunks in read reply */
 806                         /* start at write chunk array count */
 807                         iptr = &headerp->rm_body.rm_chunks[2];
 808                         rdmalen = rpcrdma_count_chunks(rep,
 809                                                 req->rl_nchunks, 1, &iptr);
 810                         /* check for validity, and no reply chunk after */
 811                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 812                                 goto badheader;
 813                         rep->rr_len -=
 814                             ((unsigned char *)iptr - (unsigned char *)headerp);
 815                         status = rep->rr_len + rdmalen;
 816                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 817                         /* special case - last chunk may omit padding */
 818                         if (rdmalen &= 3) {
 819                                 rdmalen = 4 - rdmalen;
 820                                 status += rdmalen;
 821                         }
 822                 } else {
 823                         /* else ordinary inline */
 824                         rdmalen = 0;
 825                         iptr = (__be32 *)((unsigned char *)headerp + 28);
 826                         rep->rr_len -= 28; /*sizeof *headerp;*/
 827                         status = rep->rr_len;
 828                 }
 829                 /* Fix up the rpc results for upper layer */
 830                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 831                 break;
 832
 833         case htonl(RDMA_NOMSG):
 834                 /* never expect read or write chunks, always reply chunks */
 835                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 836                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 837                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 838                     req->rl_nchunks == 0)
 839                         goto badheader;
 840                 iptr = (__be32 *)((unsigned char *)headerp + 28);
 841                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 842                 if (rdmalen < 0)
 843                         goto badheader;
 844                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 845                 /* Reply chunk buffer already is the reply vector - no fixup. */
 846                 status = rdmalen;
 847                 break;
 848
 849 badheader:
 850         default:
 851                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 852                                 " chunks[012] == %d %d %d"
 853                                 " expected chunks <= %d\n",
 854                                 __func__, ntohl(headerp->rm_type),
 855                                 headerp->rm_body.rm_chunks[0],
 856                                 headerp->rm_body.rm_chunks[1],
 857                                 headerp->rm_body.rm_chunks[2],
 858                                 req->rl_nchunks);
 859                 status = -EIO;
 860                 r_xprt->rx_stats.bad_reply_count++;
 861                 break;
 862         }
 863
 864         cwnd = xprt->cwnd;
 865         xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
 866         if (xprt->cwnd > cwnd)
 867                 xprt_release_rqst_cong(rqst->rq_task);
 868
 869         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 870                         __func__, xprt, rqst, status);
 871         xprt_complete_rqst(rqst->rq_task, status);
 872         spin_unlock(&xprt->transport_lock);
 873 }