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 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  53 # define RPCDBG_FACILITY        RPCDBG_TRANS
  54 #endif
  55
  56 #if IS_ENABLED(CONFIG_SUNRPC_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 = cpu_to_be32(pos);
 213                         cur_rchunk->rc_target.rs_handle =
 214                                                 cpu_to_be32(seg->mr_rkey);
 215                         cur_rchunk->rc_target.rs_length =
 216                                                 cpu_to_be32(seg->mr_len);
 217                         xdr_encode_hyper(
 218                                         (__be32 *)&cur_rchunk->rc_target.rs_offset,
 219                                         seg->mr_base);
 220                         dprintk("RPC:       %s: read chunk "
 221                                 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
 222                                 seg->mr_len, (unsigned long long)seg->mr_base,
 223                                 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
 224                         cur_rchunk++;
 225                         r_xprt->rx_stats.read_chunk_count++;
 226                 } else {                /* write/reply */
 227                         cur_wchunk->wc_target.rs_handle =
 228                                                 cpu_to_be32(seg->mr_rkey);
 229                         cur_wchunk->wc_target.rs_length =
 230                                                 cpu_to_be32(seg->mr_len);
 231                         xdr_encode_hyper(
 232                                         (__be32 *)&cur_wchunk->wc_target.rs_offset,
 233                                         seg->mr_base);
 234                         dprintk("RPC:       %s: %s chunk "
 235                                 "elem %d@0x%llx:0x%x (%s)\n", __func__,
 236                                 (type == rpcrdma_replych) ? "reply" : "write",
 237                                 seg->mr_len, (unsigned long long)seg->mr_base,
 238                                 seg->mr_rkey, n < nsegs ? "more" : "last");
 239                         cur_wchunk++;
 240                         if (type == rpcrdma_replych)
 241                                 r_xprt->rx_stats.reply_chunk_count++;
 242                         else
 243                                 r_xprt->rx_stats.write_chunk_count++;
 244                         r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 245                 }
 246                 nchunks++;
 247                 seg   += n;
 248                 nsegs -= n;
 249         } while (nsegs);
 250
 251         /* success. all failures return above */
 252         req->rl_nchunks = nchunks;
 253
 254         /*
 255          * finish off header. If write, marshal discrim and nchunks.
 256          */
 257         if (cur_rchunk) {
 258                 iptr = (__be32 *) cur_rchunk;
 259                 *iptr++ = xdr_zero;     /* finish the read chunk list */
 260                 *iptr++ = xdr_zero;     /* encode a NULL write chunk list */
 261                 *iptr++ = xdr_zero;     /* encode a NULL reply chunk */
 262         } else {
 263                 warray->wc_discrim = xdr_one;
 264                 warray->wc_nchunks = cpu_to_be32(nchunks);
 265                 iptr = (__be32 *) cur_wchunk;
 266                 if (type == rpcrdma_writech) {
 267                         *iptr++ = xdr_zero; /* finish the write chunk list */
 268                         *iptr++ = xdr_zero; /* encode a NULL reply chunk */
 269                 }
 270         }
 271
 272         /*
 273          * Return header size.
 274          */
 275         return (unsigned char *)iptr - (unsigned char *)headerp;
 276
 277 out:
 278         if (r_xprt->rx_ia.ri_memreg_strategy != RPCRDMA_FRMR) {
 279                 for (pos = 0; nchunks--;)
 280                         pos += rpcrdma_deregister_external(
 281                                         &req->rl_segments[pos], r_xprt);
 282         }
 283         return n;
 284 }
 285
 286 /*
 287  * Marshal chunks. This routine returns the header length
 288  * consumed by marshaling.
 289  *
 290  * Returns positive RPC/RDMA header size, or negative errno.
 291  */
 292
 293 ssize_t
 294 rpcrdma_marshal_chunks(struct rpc_rqst *rqst, ssize_t result)
 295 {
 296         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 297         struct rpcrdma_msg *headerp = rdmab_to_msg(req->rl_rdmabuf);
 298
 299         if (req->rl_rtype != rpcrdma_noch)
 300                 result = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
 301                                                headerp, req->rl_rtype);
 302         else if (req->rl_wtype != rpcrdma_noch)
 303                 result = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
 304                                                headerp, req->rl_wtype);
 305         return result;
 306 }
 307
 308 /*
 309  * Copy write data inline.
 310  * This function is used for "small" requests. Data which is passed
 311  * to RPC via iovecs (or page list) is copied directly into the
 312  * pre-registered memory buffer for this request. For small amounts
 313  * of data, this is efficient. The cutoff value is tunable.
 314  */
 315 static int
 316 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
 317 {
 318         int i, npages, curlen;
 319         int copy_len;
 320         unsigned char *srcp, *destp;
 321         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 322         int page_base;
 323         struct page **ppages;
 324
 325         destp = rqst->rq_svec[0].iov_base;
 326         curlen = rqst->rq_svec[0].iov_len;
 327         destp += curlen;
 328         /*
 329          * Do optional padding where it makes sense. Alignment of write
 330          * payload can help the server, if our setting is accurate.
 331          */
 332         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 333         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 334                 pad = 0;        /* don't pad this request */
 335
 336         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 337                 __func__, pad, destp, rqst->rq_slen, curlen);
 338
 339         copy_len = rqst->rq_snd_buf.page_len;
 340
 341         if (rqst->rq_snd_buf.tail[0].iov_len) {
 342                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 343                 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
 344                         memmove(destp + copy_len,
 345                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 346                         r_xprt->rx_stats.pullup_copy_count += curlen;
 347                 }
 348                 dprintk("RPC:       %s: tail destp 0x%p len %d\n",
 349                         __func__, destp + copy_len, curlen);
 350                 rqst->rq_svec[0].iov_len += curlen;
 351         }
 352         r_xprt->rx_stats.pullup_copy_count += copy_len;
 353
 354         page_base = rqst->rq_snd_buf.page_base;
 355         ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
 356         page_base &= ~PAGE_MASK;
 357         npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
 358         for (i = 0; copy_len && i < npages; i++) {
 359                 curlen = PAGE_SIZE - page_base;
 360                 if (curlen > copy_len)
 361                         curlen = copy_len;
 362                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 363                         __func__, i, destp, copy_len, curlen);
 364                 srcp = kmap_atomic(ppages[i]);
 365                 memcpy(destp, srcp+page_base, curlen);
 366                 kunmap_atomic(srcp);
 367                 rqst->rq_svec[0].iov_len += curlen;
 368                 destp += curlen;
 369                 copy_len -= curlen;
 370                 page_base = 0;
 371         }
 372         /* header now contains entire send message */
 373         return pad;
 374 }
 375
 376 /*
 377  * Marshal a request: the primary job of this routine is to choose
 378  * the transfer modes. See comments below.
 379  *
 380  * Uses multiple RDMA IOVs for a request:
 381  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 382  *         preregistered buffer that already holds the RPC data in
 383  *         its middle.
 384  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 385  *  [2] -- optional padding.
 386  *  [3] -- if padded, header only in [1] and data here.
 387  *
 388  * Returns zero on success, otherwise a negative errno.
 389  */
 390
 391 int
 392 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 393 {
 394         struct rpc_xprt *xprt = rqst->rq_xprt;
 395         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 396         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 397         char *base;
 398         size_t rpclen, padlen;
 399         ssize_t hdrlen;
 400         struct rpcrdma_msg *headerp;
 401
 402         /*
 403          * rpclen gets amount of data in first buffer, which is the
 404          * pre-registered buffer.
 405          */
 406         base = rqst->rq_svec[0].iov_base;
 407         rpclen = rqst->rq_svec[0].iov_len;
 408
 409         headerp = rdmab_to_msg(req->rl_rdmabuf);
 410         /* don't byte-swap XID, it's already done in request */
 411         headerp->rm_xid = rqst->rq_xid;
 412         headerp->rm_vers = rpcrdma_version;
 413         headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
 414         headerp->rm_type = rdma_msg;
 415
 416         /*
 417          * Chunks needed for results?
 418          *
 419          * o If the expected result is under the inline threshold, all ops
 420          *   return as inline (but see later).
 421          * o Large non-read ops return as a single reply chunk.
 422          * o Large read ops return data as write chunk(s), header as inline.
 423          *
 424          * Note: the NFS code sending down multiple result segments implies
 425          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 426          */
 427
 428         /*
 429          * This code can handle read chunks, write chunks OR reply
 430          * chunks -- only one type. If the request is too big to fit
 431          * inline, then we will choose read chunks. If the request is
 432          * a READ, then use write chunks to separate the file data
 433          * into pages; otherwise use reply chunks.
 434          */
 435         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 436                 req->rl_wtype = rpcrdma_noch;
 437         else if (rqst->rq_rcv_buf.page_len == 0)
 438                 req->rl_wtype = rpcrdma_replych;
 439         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 440                 req->rl_wtype = rpcrdma_writech;
 441         else
 442                 req->rl_wtype = rpcrdma_replych;
 443
 444         /*
 445          * Chunks needed for arguments?
 446          *
 447          * o If the total request is under the inline threshold, all ops
 448          *   are sent as inline.
 449          * o Large non-write ops are sent with the entire message as a
 450          *   single read chunk (protocol 0-position special case).
 451          * o Large write ops transmit data as read chunk(s), header as
 452          *   inline.
 453          *
 454          * Note: the NFS code sending down multiple argument segments
 455          * implies the op is a write.
 456          * TBD check NFSv4 setacl
 457          */
 458         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 459                 req->rl_rtype = rpcrdma_noch;
 460         else if (rqst->rq_snd_buf.page_len == 0)
 461                 req->rl_rtype = rpcrdma_areadch;
 462         else
 463                 req->rl_rtype = rpcrdma_readch;
 464
 465         /* The following simplification is not true forever */
 466         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype == rpcrdma_replych)
 467                 req->rl_wtype = rpcrdma_noch;
 468         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype != rpcrdma_noch) {
 469                 dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
 470                         __func__);
 471                 return -EIO;
 472         }
 473
 474         hdrlen = RPCRDMA_HDRLEN_MIN;
 475         padlen = 0;
 476
 477         /*
 478          * Pull up any extra send data into the preregistered buffer.
 479          * When padding is in use and applies to the transfer, insert
 480          * it and change the message type.
 481          */
 482         if (req->rl_rtype == rpcrdma_noch) {
 483
 484                 padlen = rpcrdma_inline_pullup(rqst,
 485                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 486
 487                 if (padlen) {
 488                         headerp->rm_type = rdma_msgp;
 489                         headerp->rm_body.rm_padded.rm_align =
 490                                 cpu_to_be32(RPCRDMA_INLINE_PAD_VALUE(rqst));
 491                         headerp->rm_body.rm_padded.rm_thresh =
 492                                 cpu_to_be32(RPCRDMA_INLINE_PAD_THRESH);
 493                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 494                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 495                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 496                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 497                         if (req->rl_wtype != rpcrdma_noch) {
 498                                 dprintk("RPC:       %s: invalid chunk list\n",
 499                                         __func__);
 500                                 return -EIO;
 501                         }
 502                 } else {
 503                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 504                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 505                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 506                         /* new length after pullup */
 507                         rpclen = rqst->rq_svec[0].iov_len;
 508                         /*
 509                          * Currently we try to not actually use read inline.
 510                          * Reply chunks have the desirable property that
 511                          * they land, packed, directly in the target buffers
 512                          * without headers, so they require no fixup. The
 513                          * additional RDMA Write op sends the same amount
 514                          * of data, streams on-the-wire and adds no overhead
 515                          * on receive. Therefore, we request a reply chunk
 516                          * for non-writes wherever feasible and efficient.
 517                          */
 518                         if (req->rl_wtype == rpcrdma_noch)
 519                                 req->rl_wtype = rpcrdma_replych;
 520                 }
 521         }
 522
 523         hdrlen = rpcrdma_marshal_chunks(rqst, hdrlen);
 524         if (hdrlen < 0)
 525                 return hdrlen;
 526
 527         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 528                 " headerp 0x%p base 0x%p lkey 0x%x\n",
 529                 __func__, transfertypes[req->rl_wtype], hdrlen, rpclen, padlen,
 530                 headerp, base, rdmab_lkey(req->rl_rdmabuf));
 531
 532         /*
 533          * initialize send_iov's - normally only two: rdma chunk header and
 534          * single preregistered RPC header buffer, but if padding is present,
 535          * then use a preregistered (and zeroed) pad buffer between the RPC
 536          * header and any write data. In all non-rdma cases, any following
 537          * data has been copied into the RPC header buffer.
 538          */
 539         req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
 540         req->rl_send_iov[0].length = hdrlen;
 541         req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
 542
 543         req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
 544         req->rl_send_iov[1].length = rpclen;
 545         req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
 546
 547         req->rl_niovs = 2;
 548
 549         if (padlen) {
 550                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 551
 552                 req->rl_send_iov[2].addr = rdmab_addr(ep->rep_padbuf);
 553                 req->rl_send_iov[2].length = padlen;
 554                 req->rl_send_iov[2].lkey = rdmab_lkey(ep->rep_padbuf);
 555
 556                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 557                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 558                 req->rl_send_iov[3].lkey = rdmab_lkey(req->rl_sendbuf);
 559
 560                 req->rl_niovs = 4;
 561         }
 562
 563         return 0;
 564 }
 565
 566 /*
 567  * Chase down a received write or reply chunklist to get length
 568  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 569  */
 570 static int
 571 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 572 {
 573         unsigned int i, total_len;
 574         struct rpcrdma_write_chunk *cur_wchunk;
 575         char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
 576
 577         i = be32_to_cpu(**iptrp);
 578         if (i > max)
 579                 return -1;
 580         cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 581         total_len = 0;
 582         while (i--) {
 583                 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 584                 ifdebug(FACILITY) {
 585                         u64 off;
 586                         xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 587                         dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
 588                                 __func__,
 589                                 be32_to_cpu(seg->rs_length),
 590                                 (unsigned long long)off,
 591                                 be32_to_cpu(seg->rs_handle));
 592                 }
 593                 total_len += be32_to_cpu(seg->rs_length);
 594                 ++cur_wchunk;
 595         }
 596         /* check and adjust for properly terminated write chunk */
 597         if (wrchunk) {
 598                 __be32 *w = (__be32 *) cur_wchunk;
 599                 if (*w++ != xdr_zero)
 600                         return -1;
 601                 cur_wchunk = (struct rpcrdma_write_chunk *) w;
 602         }
 603         if ((char *)cur_wchunk > base + rep->rr_len)
 604                 return -1;
 605
 606         *iptrp = (__be32 *) cur_wchunk;
 607         return total_len;
 608 }
 609
 610 /*
 611  * Scatter inline received data back into provided iov's.
 612  */
 613 static void
 614 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
 615 {
 616         int i, npages, curlen, olen;
 617         char *destp;
 618         struct page **ppages;
 619         int page_base;
 620
 621         curlen = rqst->rq_rcv_buf.head[0].iov_len;
 622         if (curlen > copy_len) {        /* write chunk header fixup */
 623                 curlen = copy_len;
 624                 rqst->rq_rcv_buf.head[0].iov_len = curlen;
 625         }
 626
 627         dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 628                 __func__, srcp, copy_len, curlen);
 629
 630         /* Shift pointer for first receive segment only */
 631         rqst->rq_rcv_buf.head[0].iov_base = srcp;
 632         srcp += curlen;
 633         copy_len -= curlen;
 634
 635         olen = copy_len;
 636         i = 0;
 637         rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
 638         page_base = rqst->rq_rcv_buf.page_base;
 639         ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
 640         page_base &= ~PAGE_MASK;
 641
 642         if (copy_len && rqst->rq_rcv_buf.page_len) {
 643                 npages = PAGE_ALIGN(page_base +
 644                         rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
 645                 for (; i < npages; i++) {
 646                         curlen = PAGE_SIZE - page_base;
 647                         if (curlen > copy_len)
 648                                 curlen = copy_len;
 649                         dprintk("RPC:       %s: page %d"
 650                                 " srcp 0x%p len %d curlen %d\n",
 651                                 __func__, i, srcp, copy_len, curlen);
 652                         destp = kmap_atomic(ppages[i]);
 653                         memcpy(destp + page_base, srcp, curlen);
 654                         flush_dcache_page(ppages[i]);
 655                         kunmap_atomic(destp);
 656                         srcp += curlen;
 657                         copy_len -= curlen;
 658                         if (copy_len == 0)
 659                                 break;
 660                         page_base = 0;
 661                 }
 662         }
 663
 664         if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
 665                 curlen = copy_len;
 666                 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
 667                         curlen = rqst->rq_rcv_buf.tail[0].iov_len;
 668                 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
 669                         memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
 670                 dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
 671                         __func__, srcp, copy_len, curlen);
 672                 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
 673                 copy_len -= curlen; ++i;
 674         } else
 675                 rqst->rq_rcv_buf.tail[0].iov_len = 0;
 676
 677         if (pad) {
 678                 /* implicit padding on terminal chunk */
 679                 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
 680                 while (pad--)
 681                         p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
 682         }
 683
 684         if (copy_len)
 685                 dprintk("RPC:       %s: %d bytes in"
 686                         " %d extra segments (%d lost)\n",
 687                         __func__, olen, i, copy_len);
 688
 689         /* TBD avoid a warning from call_decode() */
 690         rqst->rq_private_buf = rqst->rq_rcv_buf;
 691 }
 692
 693 void
 694 rpcrdma_connect_worker(struct work_struct *work)
 695 {
 696         struct rpcrdma_ep *ep =
 697                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 698         struct rpcrdma_xprt *r_xprt =
 699                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 700         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 701
 702         spin_lock_bh(&xprt->transport_lock);
 703         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 704                 ++xprt->connect_cookie;
 705         if (ep->rep_connected > 0) {
 706                 if (!xprt_test_and_set_connected(xprt))
 707                         xprt_wake_pending_tasks(xprt, 0);
 708         } else {
 709                 if (xprt_test_and_clear_connected(xprt))
 710                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 711         }
 712         spin_unlock_bh(&xprt->transport_lock);
 713 }
 714
 715 /*
 716  * This function is called when an async event is posted to
 717  * the connection which changes the connection state. All it
 718  * does at this point is mark the connection up/down, the rpc
 719  * timers do the rest.
 720  */
 721 void
 722 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 723 {
 724         schedule_delayed_work(&ep->rep_connect_worker, 0);
 725 }
 726
 727 /*
 728  * Called as a tasklet to do req/reply match and complete a request
 729  * Errors must result in the RPC task either being awakened, or
 730  * allowed to timeout, to discover the errors at that time.
 731  */
 732 void
 733 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 734 {
 735         struct rpcrdma_msg *headerp;
 736         struct rpcrdma_req *req;
 737         struct rpc_rqst *rqst;
 738         struct rpc_xprt *xprt = rep->rr_xprt;
 739         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 740         __be32 *iptr;
 741         int rdmalen, status;
 742         unsigned long cwnd;
 743         u32 credits;
 744
 745         /* Check status. If bad, signal disconnect and return rep to pool */
 746         if (rep->rr_len == ~0U) {
 747                 rpcrdma_recv_buffer_put(rep);
 748                 if (r_xprt->rx_ep.rep_connected == 1) {
 749                         r_xprt->rx_ep.rep_connected = -EIO;
 750                         rpcrdma_conn_func(&r_xprt->rx_ep);
 751                 }
 752                 return;
 753         }
 754         if (rep->rr_len < RPCRDMA_HDRLEN_MIN) {
 755                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 756                 goto repost;
 757         }
 758         headerp = rdmab_to_msg(rep->rr_rdmabuf);
 759         if (headerp->rm_vers != rpcrdma_version) {
 760                 dprintk("RPC:       %s: invalid version %d\n",
 761                         __func__, be32_to_cpu(headerp->rm_vers));
 762                 goto repost;
 763         }
 764
 765         /* Get XID and try for a match. */
 766         spin_lock(&xprt->transport_lock);
 767         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 768         if (rqst == NULL) {
 769                 spin_unlock(&xprt->transport_lock);
 770                 dprintk("RPC:       %s: reply 0x%p failed "
 771                         "to match any request xid 0x%08x len %d\n",
 772                         __func__, rep, be32_to_cpu(headerp->rm_xid),
 773                         rep->rr_len);
 774 repost:
 775                 r_xprt->rx_stats.bad_reply_count++;
 776                 rep->rr_func = rpcrdma_reply_handler;
 777                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 778                         rpcrdma_recv_buffer_put(rep);
 779
 780                 return;
 781         }
 782
 783         /* get request object */
 784         req = rpcr_to_rdmar(rqst);
 785         if (req->rl_reply) {
 786                 spin_unlock(&xprt->transport_lock);
 787                 dprintk("RPC:       %s: duplicate reply 0x%p to RPC "
 788                         "request 0x%p: xid 0x%08x\n", __func__, rep, req,
 789                         be32_to_cpu(headerp->rm_xid));
 790                 goto repost;
 791         }
 792
 793         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 794                 "                   RPC request 0x%p xid 0x%08x\n",
 795                         __func__, rep, req, rqst,
 796                         be32_to_cpu(headerp->rm_xid));
 797
 798         /* from here on, the reply is no longer an orphan */
 799         req->rl_reply = rep;
 800         xprt->reestablish_timeout = 0;
 801
 802         /* check for expected message types */
 803         /* The order of some of these tests is important. */
 804         switch (headerp->rm_type) {
 805         case rdma_msg:
 806                 /* never expect read chunks */
 807                 /* never expect reply chunks (two ways to check) */
 808                 /* never expect write chunks without having offered RDMA */
 809                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 810                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 811                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 812                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 813                      req->rl_nchunks == 0))
 814                         goto badheader;
 815                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 816                         /* count any expected write chunks in read reply */
 817                         /* start at write chunk array count */
 818                         iptr = &headerp->rm_body.rm_chunks[2];
 819                         rdmalen = rpcrdma_count_chunks(rep,
 820                                                 req->rl_nchunks, 1, &iptr);
 821                         /* check for validity, and no reply chunk after */
 822                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 823                                 goto badheader;
 824                         rep->rr_len -=
 825                             ((unsigned char *)iptr - (unsigned char *)headerp);
 826                         status = rep->rr_len + rdmalen;
 827                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 828                         /* special case - last chunk may omit padding */
 829                         if (rdmalen &= 3) {
 830                                 rdmalen = 4 - rdmalen;
 831                                 status += rdmalen;
 832                         }
 833                 } else {
 834                         /* else ordinary inline */
 835                         rdmalen = 0;
 836                         iptr = (__be32 *)((unsigned char *)headerp +
 837                                                         RPCRDMA_HDRLEN_MIN);
 838                         rep->rr_len -= RPCRDMA_HDRLEN_MIN;
 839                         status = rep->rr_len;
 840                 }
 841                 /* Fix up the rpc results for upper layer */
 842                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 843                 break;
 844
 845         case rdma_nomsg:
 846                 /* never expect read or write chunks, always reply chunks */
 847                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 848                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 849                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 850                     req->rl_nchunks == 0)
 851                         goto badheader;
 852                 iptr = (__be32 *)((unsigned char *)headerp +
 853                                                         RPCRDMA_HDRLEN_MIN);
 854                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 855                 if (rdmalen < 0)
 856                         goto badheader;
 857                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 858                 /* Reply chunk buffer already is the reply vector - no fixup. */
 859                 status = rdmalen;
 860                 break;
 861
 862 badheader:
 863         default:
 864                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 865                                 " chunks[012] == %d %d %d"
 866                                 " expected chunks <= %d\n",
 867                                 __func__, be32_to_cpu(headerp->rm_type),
 868                                 headerp->rm_body.rm_chunks[0],
 869                                 headerp->rm_body.rm_chunks[1],
 870                                 headerp->rm_body.rm_chunks[2],
 871                                 req->rl_nchunks);
 872                 status = -EIO;
 873                 r_xprt->rx_stats.bad_reply_count++;
 874                 break;
 875         }
 876
 877         credits = be32_to_cpu(headerp->rm_credit);
 878         if (credits == 0)
 879                 credits = 1;    /* don't deadlock */
 880         else if (credits > r_xprt->rx_buf.rb_max_requests)
 881                 credits = r_xprt->rx_buf.rb_max_requests;
 882
 883         cwnd = xprt->cwnd;
 884         xprt->cwnd = credits << RPC_CWNDSHIFT;
 885         if (xprt->cwnd > cwnd)
 886                 xprt_release_rqst_cong(rqst->rq_task);
 887
 888         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 889                         __func__, xprt, rqst, status);
 890         xprt_complete_rqst(rqst->rq_task, status);
 891         spin_unlock(&xprt->transport_lock);
 892 }