net/sunrpc/xprtrdma/svc_rdma_marshal.c

   1 /*
   2  * Copyright (c) 2005-2006 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  * Author: Tom Tucker <tom@opengridcomputing.com>
  40  */
  41
  42 #include <linux/sunrpc/xdr.h>
  43 #include <linux/sunrpc/debug.h>
  44 #include <asm/unaligned.h>
  45 #include <linux/sunrpc/rpc_rdma.h>
  46 #include <linux/sunrpc/svc_rdma.h>
  47
  48 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
  49
  50 /*
  51  * Decodes a read chunk list. The expected format is as follows:
  52  *    descrim  : xdr_one
  53  *    position : u32 offset into XDR stream
  54  *    handle   : u32 RKEY
  55  *    . . .
  56  *  end-of-list: xdr_zero
  57  */
  58 static u32 *decode_read_list(u32 *va, u32 *vaend)
  59 {
  60         struct rpcrdma_read_chunk *ch = (struct rpcrdma_read_chunk *)va;
  61
  62         while (ch->rc_discrim != xdr_zero) {
  63                 if (((unsigned long)ch + sizeof(struct rpcrdma_read_chunk)) >
  64                     (unsigned long)vaend) {
  65                         dprintk("svcrdma: vaend=%p, ch=%p\n", vaend, ch);
  66                         return NULL;
  67                 }
  68                 ch++;
  69         }
  70         return (u32 *)&ch->rc_position;
  71 }
  72
  73 /*
  74  * Determine number of chunks and total bytes in chunk list. The chunk
  75  * list has already been verified to fit within the RPCRDMA header.
  76  */
  77 void svc_rdma_rcl_chunk_counts(struct rpcrdma_read_chunk *ch,
  78                                int *ch_count, int *byte_count)
  79 {
  80         /* compute the number of bytes represented by read chunks */
  81         *byte_count = 0;
  82         *ch_count = 0;
  83         for (; ch->rc_discrim != 0; ch++) {
  84                 *byte_count = *byte_count + ntohl(ch->rc_target.rs_length);
  85                 *ch_count = *ch_count + 1;
  86         }
  87 }
  88
  89 /*
  90  * Decodes a write chunk list. The expected format is as follows:
  91  *    descrim  : xdr_one
  92  *    nchunks  : <count>
  93  *       handle   : u32 RKEY              ---+
  94  *       length   : u32 <len of segment>     |
  95  *       offset   : remove va                + <count>
  96  *       . . .                               |
  97  *                                        ---+
  98  */
  99 static u32 *decode_write_list(u32 *va, u32 *vaend)
 100 {
 101         unsigned long start, end;
 102         int nchunks;
 103
 104         struct rpcrdma_write_array *ary =
 105                 (struct rpcrdma_write_array *)va;
 106
 107         /* Check for not write-array */
 108         if (ary->wc_discrim == xdr_zero)
 109                 return (u32 *)&ary->wc_nchunks;
 110
 111         if ((unsigned long)ary + sizeof(struct rpcrdma_write_array) >
 112             (unsigned long)vaend) {
 113                 dprintk("svcrdma: ary=%p, vaend=%p\n", ary, vaend);
 114                 return NULL;
 115         }
 116         nchunks = ntohl(ary->wc_nchunks);
 117
 118         start = (unsigned long)&ary->wc_array[0];
 119         end = (unsigned long)vaend;
 120         if (nchunks < 0 ||
 121             nchunks > (SIZE_MAX - start) / sizeof(struct rpcrdma_write_chunk) ||
 122             (start + (sizeof(struct rpcrdma_write_chunk) * nchunks)) > end) {
 123                 dprintk("svcrdma: ary=%p, wc_nchunks=%d, vaend=%p\n",
 124                         ary, nchunks, vaend);
 125                 return NULL;
 126         }
 127         /*
 128          * rs_length is the 2nd 4B field in wc_target and taking its
 129          * address skips the list terminator
 130          */
 131         return (u32 *)&ary->wc_array[nchunks].wc_target.rs_length;
 132 }
 133
 134 static u32 *decode_reply_array(u32 *va, u32 *vaend)
 135 {
 136         unsigned long start, end;
 137         int nchunks;
 138         struct rpcrdma_write_array *ary =
 139                 (struct rpcrdma_write_array *)va;
 140
 141         /* Check for no reply-array */
 142         if (ary->wc_discrim == xdr_zero)
 143                 return (u32 *)&ary->wc_nchunks;
 144
 145         if ((unsigned long)ary + sizeof(struct rpcrdma_write_array) >
 146             (unsigned long)vaend) {
 147                 dprintk("svcrdma: ary=%p, vaend=%p\n", ary, vaend);
 148                 return NULL;
 149         }
 150         nchunks = ntohl(ary->wc_nchunks);
 151
 152         start = (unsigned long)&ary->wc_array[0];
 153         end = (unsigned long)vaend;
 154         if (nchunks < 0 ||
 155             nchunks > (SIZE_MAX - start) / sizeof(struct rpcrdma_write_chunk) ||
 156             (start + (sizeof(struct rpcrdma_write_chunk) * nchunks)) > end) {
 157                 dprintk("svcrdma: ary=%p, wc_nchunks=%d, vaend=%p\n",
 158                         ary, nchunks, vaend);
 159                 return NULL;
 160         }
 161         return (u32 *)&ary->wc_array[nchunks];
 162 }
 163
 164 int svc_rdma_xdr_decode_req(struct rpcrdma_msg **rdma_req,
 165                             struct svc_rqst *rqstp)
 166 {
 167         struct rpcrdma_msg *rmsgp = NULL;
 168         u32 *va;
 169         u32 *vaend;
 170         u32 hdr_len;
 171
 172         rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
 173
 174         /* Verify that there's enough bytes for header + something */
 175         if (rqstp->rq_arg.len <= RPCRDMA_HDRLEN_MIN) {
 176                 dprintk("svcrdma: header too short = %d\n",
 177                         rqstp->rq_arg.len);
 178                 return -EINVAL;
 179         }
 180
 181         /* Decode the header */
 182         rmsgp->rm_xid = ntohl(rmsgp->rm_xid);
 183         rmsgp->rm_vers = ntohl(rmsgp->rm_vers);
 184         rmsgp->rm_credit = ntohl(rmsgp->rm_credit);
 185         rmsgp->rm_type = ntohl(rmsgp->rm_type);
 186
 187         if (rmsgp->rm_vers != RPCRDMA_VERSION)
 188                 return -ENOSYS;
 189
 190         /* Pull in the extra for the padded case and bump our pointer */
 191         if (rmsgp->rm_type == RDMA_MSGP) {
 192                 int hdrlen;
 193                 rmsgp->rm_body.rm_padded.rm_align =
 194                         ntohl(rmsgp->rm_body.rm_padded.rm_align);
 195                 rmsgp->rm_body.rm_padded.rm_thresh =
 196                         ntohl(rmsgp->rm_body.rm_padded.rm_thresh);
 197
 198                 va = &rmsgp->rm_body.rm_padded.rm_pempty[4];
 199                 rqstp->rq_arg.head[0].iov_base = va;
 200                 hdrlen = (u32)((unsigned long)va - (unsigned long)rmsgp);
 201                 rqstp->rq_arg.head[0].iov_len -= hdrlen;
 202                 if (hdrlen > rqstp->rq_arg.len)
 203                         return -EINVAL;
 204                 return hdrlen;
 205         }
 206
 207         /* The chunk list may contain either a read chunk list or a write
 208          * chunk list and a reply chunk list.
 209          */
 210         va = &rmsgp->rm_body.rm_chunks[0];
 211         vaend = (u32 *)((unsigned long)rmsgp + rqstp->rq_arg.len);
 212         va = decode_read_list(va, vaend);
 213         if (!va)
 214                 return -EINVAL;
 215         va = decode_write_list(va, vaend);
 216         if (!va)
 217                 return -EINVAL;
 218         va = decode_reply_array(va, vaend);
 219         if (!va)
 220                 return -EINVAL;
 221
 222         rqstp->rq_arg.head[0].iov_base = va;
 223         hdr_len = (unsigned long)va - (unsigned long)rmsgp;
 224         rqstp->rq_arg.head[0].iov_len -= hdr_len;
 225
 226         *rdma_req = rmsgp;
 227         return hdr_len;
 228 }
 229
 230 int svc_rdma_xdr_decode_deferred_req(struct svc_rqst *rqstp)
 231 {
 232         struct rpcrdma_msg *rmsgp = NULL;
 233         struct rpcrdma_read_chunk *ch;
 234         struct rpcrdma_write_array *ary;
 235         u32 *va;
 236         u32 hdrlen;
 237
 238         dprintk("svcrdma: processing deferred RDMA header on rqstp=%p\n",
 239                 rqstp);
 240         rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
 241
 242         /* Pull in the extra for the padded case and bump our pointer */
 243         if (rmsgp->rm_type == RDMA_MSGP) {
 244                 va = &rmsgp->rm_body.rm_padded.rm_pempty[4];
 245                 rqstp->rq_arg.head[0].iov_base = va;
 246                 hdrlen = (u32)((unsigned long)va - (unsigned long)rmsgp);
 247                 rqstp->rq_arg.head[0].iov_len -= hdrlen;
 248                 return hdrlen;
 249         }
 250
 251         /*
 252          * Skip all chunks to find RPC msg. These were previously processed
 253          */
 254         va = &rmsgp->rm_body.rm_chunks[0];
 255
 256         /* Skip read-list */
 257         for (ch = (struct rpcrdma_read_chunk *)va;
 258              ch->rc_discrim != xdr_zero; ch++);
 259         va = (u32 *)&ch->rc_position;
 260
 261         /* Skip write-list */
 262         ary = (struct rpcrdma_write_array *)va;
 263         if (ary->wc_discrim == xdr_zero)
 264                 va = (u32 *)&ary->wc_nchunks;
 265         else
 266                 /*
 267                  * rs_length is the 2nd 4B field in wc_target and taking its
 268                  * address skips the list terminator
 269                  */
 270                 va = (u32 *)&ary->wc_array[ary->wc_nchunks].wc_target.rs_length;
 271
 272         /* Skip reply-array */
 273         ary = (struct rpcrdma_write_array *)va;
 274         if (ary->wc_discrim == xdr_zero)
 275                 va = (u32 *)&ary->wc_nchunks;
 276         else
 277                 va = (u32 *)&ary->wc_array[ary->wc_nchunks];
 278
 279         rqstp->rq_arg.head[0].iov_base = va;
 280         hdrlen = (unsigned long)va - (unsigned long)rmsgp;
 281         rqstp->rq_arg.head[0].iov_len -= hdrlen;
 282
 283         return hdrlen;
 284 }
 285
 286 int svc_rdma_xdr_encode_error(struct svcxprt_rdma *xprt,
 287                               struct rpcrdma_msg *rmsgp,
 288                               enum rpcrdma_errcode err, u32 *va)
 289 {
 290         u32 *startp = va;
 291
 292         *va++ = htonl(rmsgp->rm_xid);
 293         *va++ = htonl(rmsgp->rm_vers);
 294         *va++ = htonl(xprt->sc_max_requests);
 295         *va++ = htonl(RDMA_ERROR);
 296         *va++ = htonl(err);
 297         if (err == ERR_VERS) {
 298                 *va++ = htonl(RPCRDMA_VERSION);
 299                 *va++ = htonl(RPCRDMA_VERSION);
 300         }
 301
 302         return (int)((unsigned long)va - (unsigned long)startp);
 303 }
 304
 305 int svc_rdma_xdr_get_reply_hdr_len(struct rpcrdma_msg *rmsgp)
 306 {
 307         struct rpcrdma_write_array *wr_ary;
 308
 309         /* There is no read-list in a reply */
 310
 311         /* skip write list */
 312         wr_ary = (struct rpcrdma_write_array *)
 313                 &rmsgp->rm_body.rm_chunks[1];
 314         if (wr_ary->wc_discrim)
 315                 wr_ary = (struct rpcrdma_write_array *)
 316                         &wr_ary->wc_array[ntohl(wr_ary->wc_nchunks)].
 317                         wc_target.rs_length;
 318         else
 319                 wr_ary = (struct rpcrdma_write_array *)
 320                         &wr_ary->wc_nchunks;
 321
 322         /* skip reply array */
 323         if (wr_ary->wc_discrim)
 324                 wr_ary = (struct rpcrdma_write_array *)
 325                         &wr_ary->wc_array[ntohl(wr_ary->wc_nchunks)];
 326         else
 327                 wr_ary = (struct rpcrdma_write_array *)
 328                         &wr_ary->wc_nchunks;
 329
 330         return (unsigned long) wr_ary - (unsigned long) rmsgp;
 331 }
 332
 333 void svc_rdma_xdr_encode_write_list(struct rpcrdma_msg *rmsgp, int chunks)
 334 {
 335         struct rpcrdma_write_array *ary;
 336
 337         /* no read-list */
 338         rmsgp->rm_body.rm_chunks[0] = xdr_zero;
 339
 340         /* write-array discrim */
 341         ary = (struct rpcrdma_write_array *)
 342                 &rmsgp->rm_body.rm_chunks[1];
 343         ary->wc_discrim = xdr_one;
 344         ary->wc_nchunks = htonl(chunks);
 345
 346         /* write-list terminator */
 347         ary->wc_array[chunks].wc_target.rs_handle = xdr_zero;
 348
 349         /* reply-array discriminator */
 350         ary->wc_array[chunks].wc_target.rs_length = xdr_zero;
 351 }
 352
 353 void svc_rdma_xdr_encode_reply_array(struct rpcrdma_write_array *ary,
 354                                  int chunks)
 355 {
 356         ary->wc_discrim = xdr_one;
 357         ary->wc_nchunks = htonl(chunks);
 358 }
 359
 360 void svc_rdma_xdr_encode_array_chunk(struct rpcrdma_write_array *ary,
 361                                      int chunk_no,
 362                                      __be32 rs_handle,
 363                                      __be64 rs_offset,
 364                                      u32 write_len)
 365 {
 366         struct rpcrdma_segment *seg = &ary->wc_array[chunk_no].wc_target;
 367         seg->rs_handle = rs_handle;
 368         seg->rs_offset = rs_offset;
 369         seg->rs_length = htonl(write_len);
 370 }
 371
 372 void svc_rdma_xdr_encode_reply_header(struct svcxprt_rdma *xprt,
 373                                   struct rpcrdma_msg *rdma_argp,
 374                                   struct rpcrdma_msg *rdma_resp,
 375                                   enum rpcrdma_proc rdma_type)
 376 {
 377         rdma_resp->rm_xid = htonl(rdma_argp->rm_xid);
 378         rdma_resp->rm_vers = htonl(rdma_argp->rm_vers);
 379         rdma_resp->rm_credit = htonl(xprt->sc_max_requests);
 380         rdma_resp->rm_type = htonl(rdma_type);
 381
 382         /* Encode <nul> chunks lists */
 383         rdma_resp->rm_body.rm_chunks[0] = xdr_zero;
 384         rdma_resp->rm_body.rm_chunks[1] = xdr_zero;
 385         rdma_resp->rm_body.rm_chunks[2] = xdr_zero;
 386 }