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 = 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         if (r_xprt->rx_ia.ri_memreg_strategy != RPCRDMA_FRMR) {
 275                 for (pos = 0; nchunks--;)
 276                         pos += rpcrdma_deregister_external(
 277                                         &req->rl_segments[pos], r_xprt);
 278         }
 279         return n;
 280 }
 281
 282 /*
 283  * Marshal chunks. This routine returns the header length
 284  * consumed by marshaling.
 285  *
 286  * Returns positive RPC/RDMA header size, or negative errno.
 287  */
 288
 289 ssize_t
 290 rpcrdma_marshal_chunks(struct rpc_rqst *rqst, ssize_t result)
 291 {
 292         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 293         struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)req->rl_base;
 294
 295         if (req->rl_rtype != rpcrdma_noch)
 296                 result = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
 297                                                headerp, req->rl_rtype);
 298         else if (req->rl_wtype != rpcrdma_noch)
 299                 result = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
 300                                                headerp, req->rl_wtype);
 301         return result;
 302 }
 303
 304 /*
 305  * Copy write data inline.
 306  * This function is used for "small" requests. Data which is passed
 307  * to RPC via iovecs (or page list) is copied directly into the
 308  * pre-registered memory buffer for this request. For small amounts
 309  * of data, this is efficient. The cutoff value is tunable.
 310  */
 311 static int
 312 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
 313 {
 314         int i, npages, curlen;
 315         int copy_len;
 316         unsigned char *srcp, *destp;
 317         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 318         int page_base;
 319         struct page **ppages;
 320
 321         destp = rqst->rq_svec[0].iov_base;
 322         curlen = rqst->rq_svec[0].iov_len;
 323         destp += curlen;
 324         /*
 325          * Do optional padding where it makes sense. Alignment of write
 326          * payload can help the server, if our setting is accurate.
 327          */
 328         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 329         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 330                 pad = 0;        /* don't pad this request */
 331
 332         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 333                 __func__, pad, destp, rqst->rq_slen, curlen);
 334
 335         copy_len = rqst->rq_snd_buf.page_len;
 336
 337         if (rqst->rq_snd_buf.tail[0].iov_len) {
 338                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 339                 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
 340                         memmove(destp + copy_len,
 341                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 342                         r_xprt->rx_stats.pullup_copy_count += curlen;
 343                 }
 344                 dprintk("RPC:       %s: tail destp 0x%p len %d\n",
 345                         __func__, destp + copy_len, curlen);
 346                 rqst->rq_svec[0].iov_len += curlen;
 347         }
 348         r_xprt->rx_stats.pullup_copy_count += copy_len;
 349
 350         page_base = rqst->rq_snd_buf.page_base;
 351         ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
 352         page_base &= ~PAGE_MASK;
 353         npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
 354         for (i = 0; copy_len && i < npages; i++) {
 355                 curlen = PAGE_SIZE - page_base;
 356                 if (curlen > copy_len)
 357                         curlen = copy_len;
 358                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 359                         __func__, i, destp, copy_len, curlen);
 360                 srcp = kmap_atomic(ppages[i]);
 361                 memcpy(destp, srcp+page_base, curlen);
 362                 kunmap_atomic(srcp);
 363                 rqst->rq_svec[0].iov_len += curlen;
 364                 destp += curlen;
 365                 copy_len -= curlen;
 366                 page_base = 0;
 367         }
 368         /* header now contains entire send message */
 369         return pad;
 370 }
 371
 372 /*
 373  * Marshal a request: the primary job of this routine is to choose
 374  * the transfer modes. See comments below.
 375  *
 376  * Uses multiple RDMA IOVs for a request:
 377  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 378  *         preregistered buffer that already holds the RPC data in
 379  *         its middle.
 380  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 381  *  [2] -- optional padding.
 382  *  [3] -- if padded, header only in [1] and data here.
 383  *
 384  * Returns zero on success, otherwise a negative errno.
 385  */
 386
 387 int
 388 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 389 {
 390         struct rpc_xprt *xprt = rqst->rq_xprt;
 391         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 392         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 393         char *base;
 394         size_t rpclen, padlen;
 395         ssize_t hdrlen;
 396         struct rpcrdma_msg *headerp;
 397
 398         /*
 399          * rpclen gets amount of data in first buffer, which is the
 400          * pre-registered buffer.
 401          */
 402         base = rqst->rq_svec[0].iov_base;
 403         rpclen = rqst->rq_svec[0].iov_len;
 404
 405         /* build RDMA header in private area at front */
 406         headerp = (struct rpcrdma_msg *) req->rl_base;
 407         /* don't htonl XID, it's already done in request */
 408         headerp->rm_xid = rqst->rq_xid;
 409         headerp->rm_vers = xdr_one;
 410         headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
 411         headerp->rm_type = htonl(RDMA_MSG);
 412
 413         /*
 414          * Chunks needed for results?
 415          *
 416          * o If the expected result is under the inline threshold, all ops
 417          *   return as inline (but see later).
 418          * o Large non-read ops return as a single reply chunk.
 419          * o Large read ops return data as write chunk(s), header as inline.
 420          *
 421          * Note: the NFS code sending down multiple result segments implies
 422          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 423          */
 424
 425         /*
 426          * This code can handle read chunks, write chunks OR reply
 427          * chunks -- only one type. If the request is too big to fit
 428          * inline, then we will choose read chunks. If the request is
 429          * a READ, then use write chunks to separate the file data
 430          * into pages; otherwise use reply chunks.
 431          */
 432         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 433                 req->rl_wtype = rpcrdma_noch;
 434         else if (rqst->rq_rcv_buf.page_len == 0)
 435                 req->rl_wtype = rpcrdma_replych;
 436         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 437                 req->rl_wtype = rpcrdma_writech;
 438         else
 439                 req->rl_wtype = rpcrdma_replych;
 440
 441         /*
 442          * Chunks needed for arguments?
 443          *
 444          * o If the total request is under the inline threshold, all ops
 445          *   are sent as inline.
 446          * o Large non-write ops are sent with the entire message as a
 447          *   single read chunk (protocol 0-position special case).
 448          * o Large write ops transmit data as read chunk(s), header as
 449          *   inline.
 450          *
 451          * Note: the NFS code sending down multiple argument segments
 452          * implies the op is a write.
 453          * TBD check NFSv4 setacl
 454          */
 455         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 456                 req->rl_rtype = rpcrdma_noch;
 457         else if (rqst->rq_snd_buf.page_len == 0)
 458                 req->rl_rtype = rpcrdma_areadch;
 459         else
 460                 req->rl_rtype = rpcrdma_readch;
 461
 462         /* The following simplification is not true forever */
 463         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype == rpcrdma_replych)
 464                 req->rl_wtype = rpcrdma_noch;
 465         if (req->rl_rtype != rpcrdma_noch && req->rl_wtype != rpcrdma_noch) {
 466                 dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
 467                         __func__);
 468                 return -EIO;
 469         }
 470
 471         hdrlen = 28; /*sizeof *headerp;*/
 472         padlen = 0;
 473
 474         /*
 475          * Pull up any extra send data into the preregistered buffer.
 476          * When padding is in use and applies to the transfer, insert
 477          * it and change the message type.
 478          */
 479         if (req->rl_rtype == rpcrdma_noch) {
 480
 481                 padlen = rpcrdma_inline_pullup(rqst,
 482                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 483
 484                 if (padlen) {
 485                         headerp->rm_type = htonl(RDMA_MSGP);
 486                         headerp->rm_body.rm_padded.rm_align =
 487                                 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
 488                         headerp->rm_body.rm_padded.rm_thresh =
 489                                 htonl(RPCRDMA_INLINE_PAD_THRESH);
 490                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 491                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 492                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 493                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 494                         if (req->rl_wtype != rpcrdma_noch) {
 495                                 dprintk("RPC:       %s: invalid chunk list\n",
 496                                         __func__);
 497                                 return -EIO;
 498                         }
 499                 } else {
 500                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 501                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 502                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 503                         /* new length after pullup */
 504                         rpclen = rqst->rq_svec[0].iov_len;
 505                         /*
 506                          * Currently we try to not actually use read inline.
 507                          * Reply chunks have the desirable property that
 508                          * they land, packed, directly in the target buffers
 509                          * without headers, so they require no fixup. The
 510                          * additional RDMA Write op sends the same amount
 511                          * of data, streams on-the-wire and adds no overhead
 512                          * on receive. Therefore, we request a reply chunk
 513                          * for non-writes wherever feasible and efficient.
 514                          */
 515                         if (req->rl_wtype == rpcrdma_noch)
 516                                 req->rl_wtype = rpcrdma_replych;
 517                 }
 518         }
 519
 520         hdrlen = rpcrdma_marshal_chunks(rqst, hdrlen);
 521         if (hdrlen < 0)
 522                 return hdrlen;
 523
 524         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 525                 " headerp 0x%p base 0x%p lkey 0x%x\n",
 526                 __func__, transfertypes[req->rl_wtype], hdrlen, rpclen, padlen,
 527                 headerp, base, req->rl_iov.lkey);
 528
 529         /*
 530          * initialize send_iov's - normally only two: rdma chunk header and
 531          * single preregistered RPC header buffer, but if padding is present,
 532          * then use a preregistered (and zeroed) pad buffer between the RPC
 533          * header and any write data. In all non-rdma cases, any following
 534          * data has been copied into the RPC header buffer.
 535          */
 536         req->rl_send_iov[0].addr = req->rl_iov.addr;
 537         req->rl_send_iov[0].length = hdrlen;
 538         req->rl_send_iov[0].lkey = req->rl_iov.lkey;
 539
 540         req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
 541         req->rl_send_iov[1].length = rpclen;
 542         req->rl_send_iov[1].lkey = req->rl_iov.lkey;
 543
 544         req->rl_niovs = 2;
 545
 546         if (padlen) {
 547                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 548
 549                 req->rl_send_iov[2].addr = ep->rep_pad.addr;
 550                 req->rl_send_iov[2].length = padlen;
 551                 req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
 552
 553                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 554                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 555                 req->rl_send_iov[3].lkey = req->rl_iov.lkey;
 556
 557                 req->rl_niovs = 4;
 558         }
 559
 560         return 0;
 561 }
 562
 563 /*
 564  * Chase down a received write or reply chunklist to get length
 565  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 566  */
 567 static int
 568 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 569 {
 570         unsigned int i, total_len;
 571         struct rpcrdma_write_chunk *cur_wchunk;
 572
 573         i = ntohl(**iptrp);     /* get array count */
 574         if (i > max)
 575                 return -1;
 576         cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
 577         total_len = 0;
 578         while (i--) {
 579                 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
 580                 ifdebug(FACILITY) {
 581                         u64 off;
 582                         xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
 583                         dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
 584                                 __func__,
 585                                 ntohl(seg->rs_length),
 586                                 (unsigned long long)off,
 587                                 ntohl(seg->rs_handle));
 588                 }
 589                 total_len += ntohl(seg->rs_length);
 590                 ++cur_wchunk;
 591         }
 592         /* check and adjust for properly terminated write chunk */
 593         if (wrchunk) {
 594                 __be32 *w = (__be32 *) cur_wchunk;
 595                 if (*w++ != xdr_zero)
 596                         return -1;
 597                 cur_wchunk = (struct rpcrdma_write_chunk *) w;
 598         }
 599         if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
 600                 return -1;
 601
 602         *iptrp = (__be32 *) cur_wchunk;
 603         return total_len;
 604 }
 605
 606 /*
 607  * Scatter inline received data back into provided iov's.
 608  */
 609 static void
 610 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
 611 {
 612         int i, npages, curlen, olen;
 613         char *destp;
 614         struct page **ppages;
 615         int page_base;
 616
 617         curlen = rqst->rq_rcv_buf.head[0].iov_len;
 618         if (curlen > copy_len) {        /* write chunk header fixup */
 619                 curlen = copy_len;
 620                 rqst->rq_rcv_buf.head[0].iov_len = curlen;
 621         }
 622
 623         dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
 624                 __func__, srcp, copy_len, curlen);
 625
 626         /* Shift pointer for first receive segment only */
 627         rqst->rq_rcv_buf.head[0].iov_base = srcp;
 628         srcp += curlen;
 629         copy_len -= curlen;
 630
 631         olen = copy_len;
 632         i = 0;
 633         rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
 634         page_base = rqst->rq_rcv_buf.page_base;
 635         ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
 636         page_base &= ~PAGE_MASK;
 637
 638         if (copy_len && rqst->rq_rcv_buf.page_len) {
 639                 npages = PAGE_ALIGN(page_base +
 640                         rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
 641                 for (; i < npages; i++) {
 642                         curlen = PAGE_SIZE - page_base;
 643                         if (curlen > copy_len)
 644                                 curlen = copy_len;
 645                         dprintk("RPC:       %s: page %d"
 646                                 " srcp 0x%p len %d curlen %d\n",
 647                                 __func__, i, srcp, copy_len, curlen);
 648                         destp = kmap_atomic(ppages[i]);
 649                         memcpy(destp + page_base, srcp, curlen);
 650                         flush_dcache_page(ppages[i]);
 651                         kunmap_atomic(destp);
 652                         srcp += curlen;
 653                         copy_len -= curlen;
 654                         if (copy_len == 0)
 655                                 break;
 656                         page_base = 0;
 657                 }
 658         }
 659
 660         if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
 661                 curlen = copy_len;
 662                 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
 663                         curlen = rqst->rq_rcv_buf.tail[0].iov_len;
 664                 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
 665                         memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
 666                 dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
 667                         __func__, srcp, copy_len, curlen);
 668                 rqst->rq_rcv_buf.tail[0].iov_len = curlen;
 669                 copy_len -= curlen; ++i;
 670         } else
 671                 rqst->rq_rcv_buf.tail[0].iov_len = 0;
 672
 673         if (pad) {
 674                 /* implicit padding on terminal chunk */
 675                 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
 676                 while (pad--)
 677                         p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
 678         }
 679
 680         if (copy_len)
 681                 dprintk("RPC:       %s: %d bytes in"
 682                         " %d extra segments (%d lost)\n",
 683                         __func__, olen, i, copy_len);
 684
 685         /* TBD avoid a warning from call_decode() */
 686         rqst->rq_private_buf = rqst->rq_rcv_buf;
 687 }
 688
 689 void
 690 rpcrdma_connect_worker(struct work_struct *work)
 691 {
 692         struct rpcrdma_ep *ep =
 693                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 694         struct rpc_xprt *xprt = ep->rep_xprt;
 695
 696         spin_lock_bh(&xprt->transport_lock);
 697         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 698                 ++xprt->connect_cookie;
 699         if (ep->rep_connected > 0) {
 700                 if (!xprt_test_and_set_connected(xprt))
 701                         xprt_wake_pending_tasks(xprt, 0);
 702         } else {
 703                 if (xprt_test_and_clear_connected(xprt))
 704                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 705         }
 706         spin_unlock_bh(&xprt->transport_lock);
 707 }
 708
 709 /*
 710  * This function is called when an async event is posted to
 711  * the connection which changes the connection state. All it
 712  * does at this point is mark the connection up/down, the rpc
 713  * timers do the rest.
 714  */
 715 void
 716 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 717 {
 718         schedule_delayed_work(&ep->rep_connect_worker, 0);
 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 rdmalen, status;
 736         unsigned long cwnd;
 737
 738         /* Check status. If bad, signal disconnect and return rep to pool */
 739         if (rep->rr_len == ~0U) {
 740                 rpcrdma_recv_buffer_put(rep);
 741                 if (r_xprt->rx_ep.rep_connected == 1) {
 742                         r_xprt->rx_ep.rep_connected = -EIO;
 743                         rpcrdma_conn_func(&r_xprt->rx_ep);
 744                 }
 745                 return;
 746         }
 747         if (rep->rr_len < 28) {
 748                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 749                 goto repost;
 750         }
 751         headerp = (struct rpcrdma_msg *) rep->rr_base;
 752         if (headerp->rm_vers != xdr_one) {
 753                 dprintk("RPC:       %s: invalid version %d\n",
 754                         __func__, ntohl(headerp->rm_vers));
 755                 goto repost;
 756         }
 757
 758         /* Get XID and try for a match. */
 759         spin_lock(&xprt->transport_lock);
 760         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 761         if (rqst == NULL) {
 762                 spin_unlock(&xprt->transport_lock);
 763                 dprintk("RPC:       %s: reply 0x%p failed "
 764                         "to match any request xid 0x%08x len %d\n",
 765                         __func__, rep, headerp->rm_xid, rep->rr_len);
 766 repost:
 767                 r_xprt->rx_stats.bad_reply_count++;
 768                 rep->rr_func = rpcrdma_reply_handler;
 769                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 770                         rpcrdma_recv_buffer_put(rep);
 771
 772                 return;
 773         }
 774
 775         /* get request object */
 776         req = rpcr_to_rdmar(rqst);
 777         if (req->rl_reply) {
 778                 spin_unlock(&xprt->transport_lock);
 779                 dprintk("RPC:       %s: duplicate reply 0x%p to RPC "
 780                         "request 0x%p: xid 0x%08x\n", __func__, rep, req,
 781                         headerp->rm_xid);
 782                 goto repost;
 783         }
 784
 785         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 786                 "                   RPC request 0x%p xid 0x%08x\n",
 787                         __func__, rep, req, rqst, headerp->rm_xid);
 788
 789         /* from here on, the reply is no longer an orphan */
 790         req->rl_reply = rep;
 791         xprt->reestablish_timeout = 0;
 792
 793         /* check for expected message types */
 794         /* The order of some of these tests is important. */
 795         switch (headerp->rm_type) {
 796         case htonl(RDMA_MSG):
 797                 /* never expect read chunks */
 798                 /* never expect reply chunks (two ways to check) */
 799                 /* never expect write chunks without having offered RDMA */
 800                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 801                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 802                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 803                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 804                      req->rl_nchunks == 0))
 805                         goto badheader;
 806                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 807                         /* count any expected write chunks in read reply */
 808                         /* start at write chunk array count */
 809                         iptr = &headerp->rm_body.rm_chunks[2];
 810                         rdmalen = rpcrdma_count_chunks(rep,
 811                                                 req->rl_nchunks, 1, &iptr);
 812                         /* check for validity, and no reply chunk after */
 813                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 814                                 goto badheader;
 815                         rep->rr_len -=
 816                             ((unsigned char *)iptr - (unsigned char *)headerp);
 817                         status = rep->rr_len + rdmalen;
 818                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 819                         /* special case - last chunk may omit padding */
 820                         if (rdmalen &= 3) {
 821                                 rdmalen = 4 - rdmalen;
 822                                 status += rdmalen;
 823                         }
 824                 } else {
 825                         /* else ordinary inline */
 826                         rdmalen = 0;
 827                         iptr = (__be32 *)((unsigned char *)headerp + 28);
 828                         rep->rr_len -= 28; /*sizeof *headerp;*/
 829                         status = rep->rr_len;
 830                 }
 831                 /* Fix up the rpc results for upper layer */
 832                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 833                 break;
 834
 835         case htonl(RDMA_NOMSG):
 836                 /* never expect read or write chunks, always reply chunks */
 837                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 838                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 839                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 840                     req->rl_nchunks == 0)
 841                         goto badheader;
 842                 iptr = (__be32 *)((unsigned char *)headerp + 28);
 843                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 844                 if (rdmalen < 0)
 845                         goto badheader;
 846                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 847                 /* Reply chunk buffer already is the reply vector - no fixup. */
 848                 status = rdmalen;
 849                 break;
 850
 851 badheader:
 852         default:
 853                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 854                                 " chunks[012] == %d %d %d"
 855                                 " expected chunks <= %d\n",
 856                                 __func__, ntohl(headerp->rm_type),
 857                                 headerp->rm_body.rm_chunks[0],
 858                                 headerp->rm_body.rm_chunks[1],
 859                                 headerp->rm_body.rm_chunks[2],
 860                                 req->rl_nchunks);
 861                 status = -EIO;
 862                 r_xprt->rx_stats.bad_reply_count++;
 863                 break;
 864         }
 865
 866         cwnd = xprt->cwnd;
 867         xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
 868         if (xprt->cwnd > cwnd)
 869                 xprt_release_rqst_cong(rqst->rq_task);
 870
 871         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 872                         __func__, xprt, rqst, status);
 873         xprt_complete_rqst(rqst->rq_task, status);
 874         spin_unlock(&xprt->transport_lock);
 875 }