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 enum rpcrdma_chunktype {
  57         rpcrdma_noch = 0,
  58         rpcrdma_readch,
  59         rpcrdma_areadch,
  60         rpcrdma_writech,
  61         rpcrdma_replych
  62 };
  63
  64 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  65 static const char transfertypes[][12] = {
  66         "pure inline",  /* no chunks */
  67         " read chunk",  /* some argument via rdma read */
  68         "*read chunk",  /* entire request via rdma read */
  69         "write chunk",  /* some result via rdma write */
  70         "reply chunk"   /* entire reply via rdma write */
  71 };
  72 #endif
  73
  74 /*
  75  * Chunk assembly from upper layer xdr_buf.
  76  *
  77  * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
  78  * elements. Segments are then coalesced when registered, if possible
  79  * within the selected memreg mode.
  80  *
  81  * Returns positive number of segments converted, or a negative errno.
  82  */
  83
  84 static int
  85 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
  86         enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
  87 {
  88         int len, n = 0, p;
  89         int page_base;
  90         struct page **ppages;
  91
  92         if (pos == 0 && xdrbuf->head[0].iov_len) {
  93                 seg[n].mr_page = NULL;
  94                 seg[n].mr_offset = xdrbuf->head[0].iov_base;
  95                 seg[n].mr_len = xdrbuf->head[0].iov_len;
  96                 ++n;
  97         }
  98
  99         len = xdrbuf->page_len;
 100         ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
 101         page_base = xdrbuf->page_base & ~PAGE_MASK;
 102         p = 0;
 103         while (len && n < nsegs) {
 104                 if (!ppages[p]) {
 105                         /* alloc the pagelist for receiving buffer */
 106                         ppages[p] = alloc_page(GFP_ATOMIC);
 107                         if (!ppages[p])
 108                                 return -ENOMEM;
 109                 }
 110                 seg[n].mr_page = ppages[p];
 111                 seg[n].mr_offset = (void *)(unsigned long) page_base;
 112                 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
 113                 if (seg[n].mr_len > PAGE_SIZE)
 114                         return -EIO;
 115                 len -= seg[n].mr_len;
 116                 ++n;
 117                 ++p;
 118                 page_base = 0;  /* page offset only applies to first page */
 119         }
 120
 121         /* Message overflows the seg array */
 122         if (len && n == nsegs)
 123                 return -EIO;
 124
 125         if (xdrbuf->tail[0].iov_len) {
 126                 /* the rpcrdma protocol allows us to omit any trailing
 127                  * xdr pad bytes, saving the server an RDMA operation. */
 128                 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
 129                         return n;
 130                 if (n == nsegs)
 131                         /* Tail remains, but we're out of segments */
 132                         return -EIO;
 133                 seg[n].mr_page = NULL;
 134                 seg[n].mr_offset = xdrbuf->tail[0].iov_base;
 135                 seg[n].mr_len = xdrbuf->tail[0].iov_len;
 136                 ++n;
 137         }
 138
 139         return n;
 140 }
 141
 142 /*
 143  * Create read/write chunk lists, and reply chunks, for RDMA
 144  *
 145  *   Assume check against THRESHOLD has been done, and chunks are required.
 146  *   Assume only encoding one list entry for read|write chunks. The NFSv3
 147  *     protocol is simple enough to allow this as it only has a single "bulk
 148  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
 149  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
 150  *
 151  * When used for a single reply chunk (which is a special write
 152  * chunk used for the entire reply, rather than just the data), it
 153  * is used primarily for READDIR and READLINK which would otherwise
 154  * be severely size-limited by a small rdma inline read max. The server
 155  * response will come back as an RDMA Write, followed by a message
 156  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
 157  * chunks do not provide data alignment, however they do not require
 158  * "fixup" (moving the response to the upper layer buffer) either.
 159  *
 160  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
 161  *
 162  *  Read chunklist (a linked list):
 163  *   N elements, position P (same P for all chunks of same arg!):
 164  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
 165  *
 166  *  Write chunklist (a list of (one) counted array):
 167  *   N elements:
 168  *    1 - N - HLOO - HLOO - ... - HLOO - 0
 169  *
 170  *  Reply chunk (a counted array):
 171  *   N elements:
 172  *    1 - N - HLOO - HLOO - ... - HLOO
 173  *
 174  * Returns positive RPC/RDMA header size, or negative errno.
 175  */
 176
 177 static ssize_t
 178 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
 179                 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
 180 {
 181         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 182         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 183         int n, nsegs, nchunks = 0;
 184         unsigned int pos;
 185         struct rpcrdma_mr_seg *seg = req->rl_segments;
 186         struct rpcrdma_read_chunk *cur_rchunk = NULL;
 187         struct rpcrdma_write_array *warray = NULL;
 188         struct rpcrdma_write_chunk *cur_wchunk = NULL;
 189         __be32 *iptr = headerp->rm_body.rm_chunks;
 190         int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool);
 191
 192         if (type == rpcrdma_readch || type == rpcrdma_areadch) {
 193                 /* a read chunk - server will RDMA Read our memory */
 194                 cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
 195         } else {
 196                 /* a write or reply chunk - server will RDMA Write our memory */
 197                 *iptr++ = xdr_zero;     /* encode a NULL read chunk list */
 198                 if (type == rpcrdma_replych)
 199                         *iptr++ = xdr_zero;     /* a NULL write chunk list */
 200                 warray = (struct rpcrdma_write_array *) iptr;
 201                 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
 202         }
 203
 204         if (type == rpcrdma_replych || type == rpcrdma_areadch)
 205                 pos = 0;
 206         else
 207                 pos = target->head[0].iov_len;
 208
 209         nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
 210         if (nsegs < 0)
 211                 return nsegs;
 212
 213         map = r_xprt->rx_ia.ri_ops->ro_map;
 214         do {
 215                 n = map(r_xprt, seg, nsegs, cur_wchunk != NULL);
 216                 if (n <= 0)
 217                         goto out;
 218                 if (cur_rchunk) {       /* read */
 219                         cur_rchunk->rc_discrim = xdr_one;
 220                         /* all read chunks have the same "position" */
 221                         cur_rchunk->rc_position = cpu_to_be32(pos);
 222                         cur_rchunk->rc_target.rs_handle =
 223                                                 cpu_to_be32(seg->mr_rkey);
 224                         cur_rchunk->rc_target.rs_length =
 225                                                 cpu_to_be32(seg->mr_len);
 226                         xdr_encode_hyper(
 227                                         (__be32 *)&cur_rchunk->rc_target.rs_offset,
 228                                         seg->mr_base);
 229                         dprintk("RPC:       %s: read chunk "
 230                                 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
 231                                 seg->mr_len, (unsigned long long)seg->mr_base,
 232                                 seg->mr_rkey, pos, n < nsegs ? "more" : "last");
 233                         cur_rchunk++;
 234                         r_xprt->rx_stats.read_chunk_count++;
 235                 } else {                /* write/reply */
 236                         cur_wchunk->wc_target.rs_handle =
 237                                                 cpu_to_be32(seg->mr_rkey);
 238                         cur_wchunk->wc_target.rs_length =
 239                                                 cpu_to_be32(seg->mr_len);
 240                         xdr_encode_hyper(
 241                                         (__be32 *)&cur_wchunk->wc_target.rs_offset,
 242                                         seg->mr_base);
 243                         dprintk("RPC:       %s: %s chunk "
 244                                 "elem %d@0x%llx:0x%x (%s)\n", __func__,
 245                                 (type == rpcrdma_replych) ? "reply" : "write",
 246                                 seg->mr_len, (unsigned long long)seg->mr_base,
 247                                 seg->mr_rkey, n < nsegs ? "more" : "last");
 248                         cur_wchunk++;
 249                         if (type == rpcrdma_replych)
 250                                 r_xprt->rx_stats.reply_chunk_count++;
 251                         else
 252                                 r_xprt->rx_stats.write_chunk_count++;
 253                         r_xprt->rx_stats.total_rdma_request += seg->mr_len;
 254                 }
 255                 nchunks++;
 256                 seg   += n;
 257                 nsegs -= n;
 258         } while (nsegs);
 259
 260         /* success. all failures return above */
 261         req->rl_nchunks = nchunks;
 262
 263         /*
 264          * finish off header. If write, marshal discrim and nchunks.
 265          */
 266         if (cur_rchunk) {
 267                 iptr = (__be32 *) cur_rchunk;
 268                 *iptr++ = xdr_zero;     /* finish the read chunk list */
 269                 *iptr++ = xdr_zero;     /* encode a NULL write chunk list */
 270                 *iptr++ = xdr_zero;     /* encode a NULL reply chunk */
 271         } else {
 272                 warray->wc_discrim = xdr_one;
 273                 warray->wc_nchunks = cpu_to_be32(nchunks);
 274                 iptr = (__be32 *) cur_wchunk;
 275                 if (type == rpcrdma_writech) {
 276                         *iptr++ = xdr_zero; /* finish the write chunk list */
 277                         *iptr++ = xdr_zero; /* encode a NULL reply chunk */
 278                 }
 279         }
 280
 281         /*
 282          * Return header size.
 283          */
 284         return (unsigned char *)iptr - (unsigned char *)headerp;
 285
 286 out:
 287         for (pos = 0; nchunks--;)
 288                 pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
 289                                                       &req->rl_segments[pos]);
 290         return n;
 291 }
 292
 293 /*
 294  * Copy write data inline.
 295  * This function is used for "small" requests. Data which is passed
 296  * to RPC via iovecs (or page list) is copied directly into the
 297  * pre-registered memory buffer for this request. For small amounts
 298  * of data, this is efficient. The cutoff value is tunable.
 299  */
 300 static int
 301 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
 302 {
 303         int i, npages, curlen;
 304         int copy_len;
 305         unsigned char *srcp, *destp;
 306         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 307         int page_base;
 308         struct page **ppages;
 309
 310         destp = rqst->rq_svec[0].iov_base;
 311         curlen = rqst->rq_svec[0].iov_len;
 312         destp += curlen;
 313         /*
 314          * Do optional padding where it makes sense. Alignment of write
 315          * payload can help the server, if our setting is accurate.
 316          */
 317         pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
 318         if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
 319                 pad = 0;        /* don't pad this request */
 320
 321         dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
 322                 __func__, pad, destp, rqst->rq_slen, curlen);
 323
 324         copy_len = rqst->rq_snd_buf.page_len;
 325
 326         if (rqst->rq_snd_buf.tail[0].iov_len) {
 327                 curlen = rqst->rq_snd_buf.tail[0].iov_len;
 328                 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
 329                         memmove(destp + copy_len,
 330                                 rqst->rq_snd_buf.tail[0].iov_base, curlen);
 331                         r_xprt->rx_stats.pullup_copy_count += curlen;
 332                 }
 333                 dprintk("RPC:       %s: tail destp 0x%p len %d\n",
 334                         __func__, destp + copy_len, curlen);
 335                 rqst->rq_svec[0].iov_len += curlen;
 336         }
 337         r_xprt->rx_stats.pullup_copy_count += copy_len;
 338
 339         page_base = rqst->rq_snd_buf.page_base;
 340         ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
 341         page_base &= ~PAGE_MASK;
 342         npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
 343         for (i = 0; copy_len && i < npages; i++) {
 344                 curlen = PAGE_SIZE - page_base;
 345                 if (curlen > copy_len)
 346                         curlen = copy_len;
 347                 dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
 348                         __func__, i, destp, copy_len, curlen);
 349                 srcp = kmap_atomic(ppages[i]);
 350                 memcpy(destp, srcp+page_base, curlen);
 351                 kunmap_atomic(srcp);
 352                 rqst->rq_svec[0].iov_len += curlen;
 353                 destp += curlen;
 354                 copy_len -= curlen;
 355                 page_base = 0;
 356         }
 357         /* header now contains entire send message */
 358         return pad;
 359 }
 360
 361 /*
 362  * Marshal a request: the primary job of this routine is to choose
 363  * the transfer modes. See comments below.
 364  *
 365  * Uses multiple RDMA IOVs for a request:
 366  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
 367  *         preregistered buffer that already holds the RPC data in
 368  *         its middle.
 369  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
 370  *  [2] -- optional padding.
 371  *  [3] -- if padded, header only in [1] and data here.
 372  *
 373  * Returns zero on success, otherwise a negative errno.
 374  */
 375
 376 int
 377 rpcrdma_marshal_req(struct rpc_rqst *rqst)
 378 {
 379         struct rpc_xprt *xprt = rqst->rq_xprt;
 380         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 381         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 382         char *base;
 383         size_t rpclen, padlen;
 384         ssize_t hdrlen;
 385         enum rpcrdma_chunktype rtype, wtype;
 386         struct rpcrdma_msg *headerp;
 387
 388         /*
 389          * rpclen gets amount of data in first buffer, which is the
 390          * pre-registered buffer.
 391          */
 392         base = rqst->rq_svec[0].iov_base;
 393         rpclen = rqst->rq_svec[0].iov_len;
 394
 395         headerp = rdmab_to_msg(req->rl_rdmabuf);
 396         /* don't byte-swap XID, it's already done in request */
 397         headerp->rm_xid = rqst->rq_xid;
 398         headerp->rm_vers = rpcrdma_version;
 399         headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
 400         headerp->rm_type = rdma_msg;
 401
 402         /*
 403          * Chunks needed for results?
 404          *
 405          * o If the expected result is under the inline threshold, all ops
 406          *   return as inline (but see later).
 407          * o Large non-read ops return as a single reply chunk.
 408          * o Large read ops return data as write chunk(s), header as inline.
 409          *
 410          * Note: the NFS code sending down multiple result segments implies
 411          * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
 412          */
 413
 414         /*
 415          * This code can handle read chunks, write chunks OR reply
 416          * chunks -- only one type. If the request is too big to fit
 417          * inline, then we will choose read chunks. If the request is
 418          * a READ, then use write chunks to separate the file data
 419          * into pages; otherwise use reply chunks.
 420          */
 421         if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
 422                 wtype = rpcrdma_noch;
 423         else if (rqst->rq_rcv_buf.page_len == 0)
 424                 wtype = rpcrdma_replych;
 425         else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
 426                 wtype = rpcrdma_writech;
 427         else
 428                 wtype = rpcrdma_replych;
 429
 430         /*
 431          * Chunks needed for arguments?
 432          *
 433          * o If the total request is under the inline threshold, all ops
 434          *   are sent as inline.
 435          * o Large non-write ops are sent with the entire message as a
 436          *   single read chunk (protocol 0-position special case).
 437          * o Large write ops transmit data as read chunk(s), header as
 438          *   inline.
 439          *
 440          * Note: the NFS code sending down multiple argument segments
 441          * implies the op is a write.
 442          * TBD check NFSv4 setacl
 443          */
 444         if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
 445                 rtype = rpcrdma_noch;
 446         else if (rqst->rq_snd_buf.page_len == 0)
 447                 rtype = rpcrdma_areadch;
 448         else
 449                 rtype = rpcrdma_readch;
 450
 451         /* The following simplification is not true forever */
 452         if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
 453                 wtype = rpcrdma_noch;
 454         if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
 455                 dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
 456                         __func__);
 457                 return -EIO;
 458         }
 459
 460         hdrlen = RPCRDMA_HDRLEN_MIN;
 461         padlen = 0;
 462
 463         /*
 464          * Pull up any extra send data into the preregistered buffer.
 465          * When padding is in use and applies to the transfer, insert
 466          * it and change the message type.
 467          */
 468         if (rtype == rpcrdma_noch) {
 469
 470                 padlen = rpcrdma_inline_pullup(rqst,
 471                                                 RPCRDMA_INLINE_PAD_VALUE(rqst));
 472
 473                 if (padlen) {
 474                         headerp->rm_type = rdma_msgp;
 475                         headerp->rm_body.rm_padded.rm_align =
 476                                 cpu_to_be32(RPCRDMA_INLINE_PAD_VALUE(rqst));
 477                         headerp->rm_body.rm_padded.rm_thresh =
 478                                 cpu_to_be32(RPCRDMA_INLINE_PAD_THRESH);
 479                         headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
 480                         headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
 481                         headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
 482                         hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
 483                         if (wtype != rpcrdma_noch) {
 484                                 dprintk("RPC:       %s: invalid chunk list\n",
 485                                         __func__);
 486                                 return -EIO;
 487                         }
 488                 } else {
 489                         headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
 490                         headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
 491                         headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
 492                         /* new length after pullup */
 493                         rpclen = rqst->rq_svec[0].iov_len;
 494                         /*
 495                          * Currently we try to not actually use read inline.
 496                          * Reply chunks have the desirable property that
 497                          * they land, packed, directly in the target buffers
 498                          * without headers, so they require no fixup. The
 499                          * additional RDMA Write op sends the same amount
 500                          * of data, streams on-the-wire and adds no overhead
 501                          * on receive. Therefore, we request a reply chunk
 502                          * for non-writes wherever feasible and efficient.
 503                          */
 504                         if (wtype == rpcrdma_noch)
 505                                 wtype = rpcrdma_replych;
 506                 }
 507         }
 508
 509         if (rtype != rpcrdma_noch) {
 510                 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
 511                                                headerp, rtype);
 512                 wtype = rtype;  /* simplify dprintk */
 513
 514         } else if (wtype != rpcrdma_noch) {
 515                 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
 516                                                headerp, wtype);
 517         }
 518         if (hdrlen < 0)
 519                 return hdrlen;
 520
 521         dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
 522                 " headerp 0x%p base 0x%p lkey 0x%x\n",
 523                 __func__, transfertypes[wtype], hdrlen, rpclen, padlen,
 524                 headerp, base, rdmab_lkey(req->rl_rdmabuf));
 525
 526         /*
 527          * initialize send_iov's - normally only two: rdma chunk header and
 528          * single preregistered RPC header buffer, but if padding is present,
 529          * then use a preregistered (and zeroed) pad buffer between the RPC
 530          * header and any write data. In all non-rdma cases, any following
 531          * data has been copied into the RPC header buffer.
 532          */
 533         req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
 534         req->rl_send_iov[0].length = hdrlen;
 535         req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
 536
 537         req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
 538         req->rl_send_iov[1].length = rpclen;
 539         req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
 540
 541         req->rl_niovs = 2;
 542
 543         if (padlen) {
 544                 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 545
 546                 req->rl_send_iov[2].addr = rdmab_addr(ep->rep_padbuf);
 547                 req->rl_send_iov[2].length = padlen;
 548                 req->rl_send_iov[2].lkey = rdmab_lkey(ep->rep_padbuf);
 549
 550                 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
 551                 req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
 552                 req->rl_send_iov[3].lkey = rdmab_lkey(req->rl_sendbuf);
 553
 554                 req->rl_niovs = 4;
 555         }
 556
 557         return 0;
 558 }
 559
 560 /*
 561  * Chase down a received write or reply chunklist to get length
 562  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
 563  */
 564 static int
 565 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
 566 {
 567         unsigned int i, total_len;
 568         struct rpcrdma_write_chunk *cur_wchunk;
 569         char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
 570
 571         i = be32_to_cpu(**iptrp);
 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                                 be32_to_cpu(seg->rs_length),
 584                                 (unsigned long long)off,
 585                                 be32_to_cpu(seg->rs_handle));
 586                 }
 587                 total_len += be32_to_cpu(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 > 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 rpcrdma_xprt *r_xprt =
 693                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 694         struct rpc_xprt *xprt = &r_xprt->rx_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 rpcrdma_xprt *r_xprt = rep->rr_rxprt;
 733         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 734         __be32 *iptr;
 735         int rdmalen, status;
 736         unsigned long cwnd;
 737         u32 credits;
 738
 739         /* Check status. If bad, signal disconnect and return rep to pool */
 740         if (rep->rr_len == ~0U) {
 741                 rpcrdma_recv_buffer_put(rep);
 742                 if (r_xprt->rx_ep.rep_connected == 1) {
 743                         r_xprt->rx_ep.rep_connected = -EIO;
 744                         rpcrdma_conn_func(&r_xprt->rx_ep);
 745                 }
 746                 return;
 747         }
 748         if (rep->rr_len < RPCRDMA_HDRLEN_MIN) {
 749                 dprintk("RPC:       %s: short/invalid reply\n", __func__);
 750                 goto repost;
 751         }
 752         headerp = rdmab_to_msg(rep->rr_rdmabuf);
 753         if (headerp->rm_vers != rpcrdma_version) {
 754                 dprintk("RPC:       %s: invalid version %d\n",
 755                         __func__, be32_to_cpu(headerp->rm_vers));
 756                 goto repost;
 757         }
 758
 759         /* Get XID and try for a match. */
 760         spin_lock(&xprt->transport_lock);
 761         rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
 762         if (rqst == NULL) {
 763                 spin_unlock(&xprt->transport_lock);
 764                 dprintk("RPC:       %s: reply 0x%p failed "
 765                         "to match any request xid 0x%08x len %d\n",
 766                         __func__, rep, be32_to_cpu(headerp->rm_xid),
 767                         rep->rr_len);
 768 repost:
 769                 r_xprt->rx_stats.bad_reply_count++;
 770                 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
 771                         rpcrdma_recv_buffer_put(rep);
 772
 773                 return;
 774         }
 775
 776         /* get request object */
 777         req = rpcr_to_rdmar(rqst);
 778         if (req->rl_reply) {
 779                 spin_unlock(&xprt->transport_lock);
 780                 dprintk("RPC:       %s: duplicate reply 0x%p to RPC "
 781                         "request 0x%p: xid 0x%08x\n", __func__, rep, req,
 782                         be32_to_cpu(headerp->rm_xid));
 783                 goto repost;
 784         }
 785
 786         dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
 787                 "                   RPC request 0x%p xid 0x%08x\n",
 788                         __func__, rep, req, rqst,
 789                         be32_to_cpu(headerp->rm_xid));
 790
 791         /* from here on, the reply is no longer an orphan */
 792         req->rl_reply = rep;
 793         xprt->reestablish_timeout = 0;
 794
 795         /* check for expected message types */
 796         /* The order of some of these tests is important. */
 797         switch (headerp->rm_type) {
 798         case rdma_msg:
 799                 /* never expect read chunks */
 800                 /* never expect reply chunks (two ways to check) */
 801                 /* never expect write chunks without having offered RDMA */
 802                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 803                     (headerp->rm_body.rm_chunks[1] == xdr_zero &&
 804                      headerp->rm_body.rm_chunks[2] != xdr_zero) ||
 805                     (headerp->rm_body.rm_chunks[1] != xdr_zero &&
 806                      req->rl_nchunks == 0))
 807                         goto badheader;
 808                 if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
 809                         /* count any expected write chunks in read reply */
 810                         /* start at write chunk array count */
 811                         iptr = &headerp->rm_body.rm_chunks[2];
 812                         rdmalen = rpcrdma_count_chunks(rep,
 813                                                 req->rl_nchunks, 1, &iptr);
 814                         /* check for validity, and no reply chunk after */
 815                         if (rdmalen < 0 || *iptr++ != xdr_zero)
 816                                 goto badheader;
 817                         rep->rr_len -=
 818                             ((unsigned char *)iptr - (unsigned char *)headerp);
 819                         status = rep->rr_len + rdmalen;
 820                         r_xprt->rx_stats.total_rdma_reply += rdmalen;
 821                         /* special case - last chunk may omit padding */
 822                         if (rdmalen &= 3) {
 823                                 rdmalen = 4 - rdmalen;
 824                                 status += rdmalen;
 825                         }
 826                 } else {
 827                         /* else ordinary inline */
 828                         rdmalen = 0;
 829                         iptr = (__be32 *)((unsigned char *)headerp +
 830                                                         RPCRDMA_HDRLEN_MIN);
 831                         rep->rr_len -= RPCRDMA_HDRLEN_MIN;
 832                         status = rep->rr_len;
 833                 }
 834                 /* Fix up the rpc results for upper layer */
 835                 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
 836                 break;
 837
 838         case rdma_nomsg:
 839                 /* never expect read or write chunks, always reply chunks */
 840                 if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
 841                     headerp->rm_body.rm_chunks[1] != xdr_zero ||
 842                     headerp->rm_body.rm_chunks[2] != xdr_one ||
 843                     req->rl_nchunks == 0)
 844                         goto badheader;
 845                 iptr = (__be32 *)((unsigned char *)headerp +
 846                                                         RPCRDMA_HDRLEN_MIN);
 847                 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
 848                 if (rdmalen < 0)
 849                         goto badheader;
 850                 r_xprt->rx_stats.total_rdma_reply += rdmalen;
 851                 /* Reply chunk buffer already is the reply vector - no fixup. */
 852                 status = rdmalen;
 853                 break;
 854
 855 badheader:
 856         default:
 857                 dprintk("%s: invalid rpcrdma reply header (type %d):"
 858                                 " chunks[012] == %d %d %d"
 859                                 " expected chunks <= %d\n",
 860                                 __func__, be32_to_cpu(headerp->rm_type),
 861                                 headerp->rm_body.rm_chunks[0],
 862                                 headerp->rm_body.rm_chunks[1],
 863                                 headerp->rm_body.rm_chunks[2],
 864                                 req->rl_nchunks);
 865                 status = -EIO;
 866                 r_xprt->rx_stats.bad_reply_count++;
 867                 break;
 868         }
 869
 870         credits = be32_to_cpu(headerp->rm_credit);
 871         if (credits == 0)
 872                 credits = 1;    /* don't deadlock */
 873         else if (credits > r_xprt->rx_buf.rb_max_requests)
 874                 credits = r_xprt->rx_buf.rb_max_requests;
 875
 876         cwnd = xprt->cwnd;
 877         xprt->cwnd = credits << RPC_CWNDSHIFT;
 878         if (xprt->cwnd > cwnd)
 879                 xprt_release_rqst_cong(rqst->rq_task);
 880
 881         dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
 882                         __func__, xprt, rqst, status);
 883         xprt_complete_rqst(rqst->rq_task, status);
 884         spin_unlock(&xprt->transport_lock);
 885 }