net/sunrpc/xprtrdma/svc_rdma_sendto.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/debug.h>
  43 #include <linux/sunrpc/rpc_rdma.h>
  44 #include <linux/spinlock.h>
  45 #include <asm/unaligned.h>
  46 #include <rdma/ib_verbs.h>
  47 #include <rdma/rdma_cm.h>
  48 #include <linux/sunrpc/svc_rdma.h>
  49
  50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
  51
  52 /* Encode an XDR as an array of IB SGE
  53  *
  54  * Assumptions:
  55  * - head[0] is physically contiguous.
  56  * - tail[0] is physically contiguous.
  57  * - pages[] is not physically or virtually contiguous and consists of
  58  *   PAGE_SIZE elements.
  59  *
  60  * Output:
  61  * SGE[0]              reserved for RCPRDMA header
  62  * SGE[1]              data from xdr->head[]
  63  * SGE[2..sge_count-2] data from xdr->pages[]
  64  * SGE[sge_count-1]    data from xdr->tail.
  65  *
  66  * The max SGE we need is the length of the XDR / pagesize + one for
  67  * head + one for tail + one for RPCRDMA header. Since RPCSVC_MAXPAGES
  68  * reserves a page for both the request and the reply header, and this
  69  * array is only concerned with the reply we are assured that we have
  70  * on extra page for the RPCRMDA header.
  71  */
  72 static int fast_reg_xdr(struct svcxprt_rdma *xprt,
  73                         struct xdr_buf *xdr,
  74                         struct svc_rdma_req_map *vec)
  75 {
  76         int sge_no;
  77         u32 sge_bytes;
  78         u32 page_bytes;
  79         u32 page_off;
  80         int page_no = 0;
  81         u8 *frva;
  82         struct svc_rdma_fastreg_mr *frmr;
  83
  84         frmr = svc_rdma_get_frmr(xprt);
  85         if (IS_ERR(frmr))
  86                 return -ENOMEM;
  87         vec->frmr = frmr;
  88
  89         /* Skip the RPCRDMA header */
  90         sge_no = 1;
  91
  92         /* Map the head. */
  93         frva = (void *)((unsigned long)(xdr->head[0].iov_base) & PAGE_MASK);
  94         vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
  95         vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
  96         vec->count = 2;
  97         sge_no++;
  98
  99         /* Map the XDR head */
 100         frmr->kva = frva;
 101         frmr->direction = DMA_TO_DEVICE;
 102         frmr->access_flags = 0;
 103         frmr->map_len = PAGE_SIZE;
 104         frmr->page_list_len = 1;
 105         page_off = (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
 106         frmr->page_list->page_list[page_no] =
 107                 ib_dma_map_page(xprt->sc_cm_id->device,
 108                                 virt_to_page(xdr->head[0].iov_base),
 109                                 page_off,
 110                                 PAGE_SIZE - page_off,
 111                                 DMA_TO_DEVICE);
 112         if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 113                                  frmr->page_list->page_list[page_no]))
 114                 goto fatal_err;
 115         atomic_inc(&xprt->sc_dma_used);
 116
 117         /* Map the XDR page list */
 118         page_off = xdr->page_base;
 119         page_bytes = xdr->page_len + page_off;
 120         if (!page_bytes)
 121                 goto encode_tail;
 122
 123         /* Map the pages */
 124         vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
 125         vec->sge[sge_no].iov_len = page_bytes;
 126         sge_no++;
 127         while (page_bytes) {
 128                 struct page *page;
 129
 130                 page = xdr->pages[page_no++];
 131                 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
 132                 page_bytes -= sge_bytes;
 133
 134                 frmr->page_list->page_list[page_no] =
 135                         ib_dma_map_page(xprt->sc_cm_id->device,
 136                                         page, page_off,
 137                                         sge_bytes, DMA_TO_DEVICE);
 138                 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 139                                          frmr->page_list->page_list[page_no]))
 140                         goto fatal_err;
 141
 142                 atomic_inc(&xprt->sc_dma_used);
 143                 page_off = 0; /* reset for next time through loop */
 144                 frmr->map_len += PAGE_SIZE;
 145                 frmr->page_list_len++;
 146         }
 147         vec->count++;
 148
 149  encode_tail:
 150         /* Map tail */
 151         if (0 == xdr->tail[0].iov_len)
 152                 goto done;
 153
 154         vec->count++;
 155         vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
 156
 157         if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) ==
 158             ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) {
 159                 /*
 160                  * If head and tail use the same page, we don't need
 161                  * to map it again.
 162                  */
 163                 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
 164         } else {
 165                 void *va;
 166
 167                 /* Map another page for the tail */
 168                 page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
 169                 va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK);
 170                 vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
 171
 172                 frmr->page_list->page_list[page_no] =
 173                     ib_dma_map_page(xprt->sc_cm_id->device, virt_to_page(va),
 174                                     page_off,
 175                                     PAGE_SIZE,
 176                                     DMA_TO_DEVICE);
 177                 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 178                                          frmr->page_list->page_list[page_no]))
 179                         goto fatal_err;
 180                 atomic_inc(&xprt->sc_dma_used);
 181                 frmr->map_len += PAGE_SIZE;
 182                 frmr->page_list_len++;
 183         }
 184
 185  done:
 186         if (svc_rdma_fastreg(xprt, frmr))
 187                 goto fatal_err;
 188
 189         return 0;
 190
 191  fatal_err:
 192         printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
 193         vec->frmr = NULL;
 194         svc_rdma_put_frmr(xprt, frmr);
 195         return -EIO;
 196 }
 197
 198 static int map_xdr(struct svcxprt_rdma *xprt,
 199                    struct xdr_buf *xdr,
 200                    struct svc_rdma_req_map *vec)
 201 {
 202         int sge_no;
 203         u32 sge_bytes;
 204         u32 page_bytes;
 205         u32 page_off;
 206         int page_no;
 207
 208         BUG_ON(xdr->len !=
 209                (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
 210
 211         if (xprt->sc_frmr_pg_list_len)
 212                 return fast_reg_xdr(xprt, xdr, vec);
 213
 214         /* Skip the first sge, this is for the RPCRDMA header */
 215         sge_no = 1;
 216
 217         /* Head SGE */
 218         vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
 219         vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
 220         sge_no++;
 221
 222         /* pages SGE */
 223         page_no = 0;
 224         page_bytes = xdr->page_len;
 225         page_off = xdr->page_base;
 226         while (page_bytes) {
 227                 vec->sge[sge_no].iov_base =
 228                         page_address(xdr->pages[page_no]) + page_off;
 229                 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
 230                 page_bytes -= sge_bytes;
 231                 vec->sge[sge_no].iov_len = sge_bytes;
 232
 233                 sge_no++;
 234                 page_no++;
 235                 page_off = 0; /* reset for next time through loop */
 236         }
 237
 238         /* Tail SGE */
 239         if (xdr->tail[0].iov_len) {
 240                 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
 241                 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
 242                 sge_no++;
 243         }
 244
 245         dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
 246                 "page_base %u page_len %u head_len %zu tail_len %zu\n",
 247                 sge_no, page_no, xdr->page_base, xdr->page_len,
 248                 xdr->head[0].iov_len, xdr->tail[0].iov_len);
 249
 250         vec->count = sge_no;
 251         return 0;
 252 }
 253
 254 static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
 255                               struct xdr_buf *xdr,
 256                               u32 xdr_off, size_t len, int dir)
 257 {
 258         struct page *page;
 259         dma_addr_t dma_addr;
 260         if (xdr_off < xdr->head[0].iov_len) {
 261                 /* This offset is in the head */
 262                 xdr_off += (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
 263                 page = virt_to_page(xdr->head[0].iov_base);
 264         } else {
 265                 xdr_off -= xdr->head[0].iov_len;
 266                 if (xdr_off < xdr->page_len) {
 267                         /* This offset is in the page list */
 268                         page = xdr->pages[xdr_off >> PAGE_SHIFT];
 269                         xdr_off &= ~PAGE_MASK;
 270                 } else {
 271                         /* This offset is in the tail */
 272                         xdr_off -= xdr->page_len;
 273                         xdr_off += (unsigned long)
 274                                 xdr->tail[0].iov_base & ~PAGE_MASK;
 275                         page = virt_to_page(xdr->tail[0].iov_base);
 276                 }
 277         }
 278         dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off,
 279                                    min_t(size_t, PAGE_SIZE, len), dir);
 280         return dma_addr;
 281 }
 282
 283 /* Assumptions:
 284  * - We are using FRMR
 285  *     - or -
 286  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
 287  */
 288 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
 289                       u32 rmr, u64 to,
 290                       u32 xdr_off, int write_len,
 291                       struct svc_rdma_req_map *vec)
 292 {
 293         struct ib_send_wr write_wr;
 294         struct ib_sge *sge;
 295         int xdr_sge_no;
 296         int sge_no;
 297         int sge_bytes;
 298         int sge_off;
 299         int bc;
 300         struct svc_rdma_op_ctxt *ctxt;
 301
 302         BUG_ON(vec->count > RPCSVC_MAXPAGES);
 303         dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
 304                 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
 305                 rmr, (unsigned long long)to, xdr_off,
 306                 write_len, vec->sge, vec->count);
 307
 308         ctxt = svc_rdma_get_context(xprt);
 309         ctxt->direction = DMA_TO_DEVICE;
 310         sge = ctxt->sge;
 311
 312         /* Find the SGE associated with xdr_off */
 313         for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
 314              xdr_sge_no++) {
 315                 if (vec->sge[xdr_sge_no].iov_len > bc)
 316                         break;
 317                 bc -= vec->sge[xdr_sge_no].iov_len;
 318         }
 319
 320         sge_off = bc;
 321         bc = write_len;
 322         sge_no = 0;
 323
 324         /* Copy the remaining SGE */
 325         while (bc != 0) {
 326                 sge_bytes = min_t(size_t,
 327                           bc, vec->sge[xdr_sge_no].iov_len-sge_off);
 328                 sge[sge_no].length = sge_bytes;
 329                 if (!vec->frmr) {
 330                         sge[sge_no].addr =
 331                                 dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
 332                                             sge_bytes, DMA_TO_DEVICE);
 333                         xdr_off += sge_bytes;
 334                         if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 335                                                  sge[sge_no].addr))
 336                                 goto err;
 337                         atomic_inc(&xprt->sc_dma_used);
 338                         sge[sge_no].lkey = xprt->sc_dma_lkey;
 339                 } else {
 340                         sge[sge_no].addr = (unsigned long)
 341                                 vec->sge[xdr_sge_no].iov_base + sge_off;
 342                         sge[sge_no].lkey = vec->frmr->mr->lkey;
 343                 }
 344                 ctxt->count++;
 345                 ctxt->frmr = vec->frmr;
 346                 sge_off = 0;
 347                 sge_no++;
 348                 xdr_sge_no++;
 349                 BUG_ON(xdr_sge_no > vec->count);
 350                 bc -= sge_bytes;
 351         }
 352
 353         /* Prepare WRITE WR */
 354         memset(&write_wr, 0, sizeof write_wr);
 355         ctxt->wr_op = IB_WR_RDMA_WRITE;
 356         write_wr.wr_id = (unsigned long)ctxt;
 357         write_wr.sg_list = &sge[0];
 358         write_wr.num_sge = sge_no;
 359         write_wr.opcode = IB_WR_RDMA_WRITE;
 360         write_wr.send_flags = IB_SEND_SIGNALED;
 361         write_wr.wr.rdma.rkey = rmr;
 362         write_wr.wr.rdma.remote_addr = to;
 363
 364         /* Post It */
 365         atomic_inc(&rdma_stat_write);
 366         if (svc_rdma_send(xprt, &write_wr))
 367                 goto err;
 368         return 0;
 369  err:
 370         svc_rdma_unmap_dma(ctxt);
 371         svc_rdma_put_frmr(xprt, vec->frmr);
 372         svc_rdma_put_context(ctxt, 0);
 373         /* Fatal error, close transport */
 374         return -EIO;
 375 }
 376
 377 static int send_write_chunks(struct svcxprt_rdma *xprt,
 378                              struct rpcrdma_msg *rdma_argp,
 379                              struct rpcrdma_msg *rdma_resp,
 380                              struct svc_rqst *rqstp,
 381                              struct svc_rdma_req_map *vec)
 382 {
 383         u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 384         int write_len;
 385         int max_write;
 386         u32 xdr_off;
 387         int chunk_off;
 388         int chunk_no;
 389         struct rpcrdma_write_array *arg_ary;
 390         struct rpcrdma_write_array *res_ary;
 391         int ret;
 392
 393         arg_ary = svc_rdma_get_write_array(rdma_argp);
 394         if (!arg_ary)
 395                 return 0;
 396         res_ary = (struct rpcrdma_write_array *)
 397                 &rdma_resp->rm_body.rm_chunks[1];
 398
 399         if (vec->frmr)
 400                 max_write = vec->frmr->map_len;
 401         else
 402                 max_write = xprt->sc_max_sge * PAGE_SIZE;
 403
 404         /* Write chunks start at the pagelist */
 405         for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
 406              xfer_len && chunk_no < arg_ary->wc_nchunks;
 407              chunk_no++) {
 408                 struct rpcrdma_segment *arg_ch;
 409                 u64 rs_offset;
 410
 411                 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
 412                 write_len = min(xfer_len, ntohl(arg_ch->rs_length));
 413
 414                 /* Prepare the response chunk given the length actually
 415                  * written */
 416                 xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
 417                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 418                                                 arg_ch->rs_handle,
 419                                                 arg_ch->rs_offset,
 420                                                 write_len);
 421                 chunk_off = 0;
 422                 while (write_len) {
 423                         int this_write;
 424                         this_write = min(write_len, max_write);
 425                         ret = send_write(xprt, rqstp,
 426                                          ntohl(arg_ch->rs_handle),
 427                                          rs_offset + chunk_off,
 428                                          xdr_off,
 429                                          this_write,
 430                                          vec);
 431                         if (ret) {
 432                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 433                                         ret);
 434                                 return -EIO;
 435                         }
 436                         chunk_off += this_write;
 437                         xdr_off += this_write;
 438                         xfer_len -= this_write;
 439                         write_len -= this_write;
 440                 }
 441         }
 442         /* Update the req with the number of chunks actually used */
 443         svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
 444
 445         return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 446 }
 447
 448 static int send_reply_chunks(struct svcxprt_rdma *xprt,
 449                              struct rpcrdma_msg *rdma_argp,
 450                              struct rpcrdma_msg *rdma_resp,
 451                              struct svc_rqst *rqstp,
 452                              struct svc_rdma_req_map *vec)
 453 {
 454         u32 xfer_len = rqstp->rq_res.len;
 455         int write_len;
 456         int max_write;
 457         u32 xdr_off;
 458         int chunk_no;
 459         int chunk_off;
 460         int nchunks;
 461         struct rpcrdma_segment *ch;
 462         struct rpcrdma_write_array *arg_ary;
 463         struct rpcrdma_write_array *res_ary;
 464         int ret;
 465
 466         arg_ary = svc_rdma_get_reply_array(rdma_argp);
 467         if (!arg_ary)
 468                 return 0;
 469         /* XXX: need to fix when reply lists occur with read-list and or
 470          * write-list */
 471         res_ary = (struct rpcrdma_write_array *)
 472                 &rdma_resp->rm_body.rm_chunks[2];
 473
 474         if (vec->frmr)
 475                 max_write = vec->frmr->map_len;
 476         else
 477                 max_write = xprt->sc_max_sge * PAGE_SIZE;
 478
 479         /* xdr offset starts at RPC message */
 480         nchunks = ntohl(arg_ary->wc_nchunks);
 481         for (xdr_off = 0, chunk_no = 0;
 482              xfer_len && chunk_no < nchunks;
 483              chunk_no++) {
 484                 u64 rs_offset;
 485                 ch = &arg_ary->wc_array[chunk_no].wc_target;
 486                 write_len = min(xfer_len, htonl(ch->rs_length));
 487
 488                 /* Prepare the reply chunk given the length actually
 489                  * written */
 490                 xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset);
 491                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 492                                                 ch->rs_handle, ch->rs_offset,
 493                                                 write_len);
 494                 chunk_off = 0;
 495                 while (write_len) {
 496                         int this_write;
 497
 498                         this_write = min(write_len, max_write);
 499                         ret = send_write(xprt, rqstp,
 500                                          ntohl(ch->rs_handle),
 501                                          rs_offset + chunk_off,
 502                                          xdr_off,
 503                                          this_write,
 504                                          vec);
 505                         if (ret) {
 506                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 507                                         ret);
 508                                 return -EIO;
 509                         }
 510                         chunk_off += this_write;
 511                         xdr_off += this_write;
 512                         xfer_len -= this_write;
 513                         write_len -= this_write;
 514                 }
 515         }
 516         /* Update the req with the number of chunks actually used */
 517         svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
 518
 519         return rqstp->rq_res.len;
 520 }
 521
 522 /* This function prepares the portion of the RPCRDMA message to be
 523  * sent in the RDMA_SEND. This function is called after data sent via
 524  * RDMA has already been transmitted. There are three cases:
 525  * - The RPCRDMA header, RPC header, and payload are all sent in a
 526  *   single RDMA_SEND. This is the "inline" case.
 527  * - The RPCRDMA header and some portion of the RPC header and data
 528  *   are sent via this RDMA_SEND and another portion of the data is
 529  *   sent via RDMA.
 530  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
 531  *   header and data are all transmitted via RDMA.
 532  * In all three cases, this function prepares the RPCRDMA header in
 533  * sge[0], the 'type' parameter indicates the type to place in the
 534  * RPCRDMA header, and the 'byte_count' field indicates how much of
 535  * the XDR to include in this RDMA_SEND. NB: The offset of the payload
 536  * to send is zero in the XDR.
 537  */
 538 static int send_reply(struct svcxprt_rdma *rdma,
 539                       struct svc_rqst *rqstp,
 540                       struct page *page,
 541                       struct rpcrdma_msg *rdma_resp,
 542                       struct svc_rdma_op_ctxt *ctxt,
 543                       struct svc_rdma_req_map *vec,
 544                       int byte_count)
 545 {
 546         struct ib_send_wr send_wr;
 547         struct ib_send_wr inv_wr;
 548         int sge_no;
 549         int sge_bytes;
 550         int page_no;
 551         int pages;
 552         int ret;
 553
 554         /* Post a recv buffer to handle another request. */
 555         ret = svc_rdma_post_recv(rdma);
 556         if (ret) {
 557                 printk(KERN_INFO
 558                        "svcrdma: could not post a receive buffer, err=%d."
 559                        "Closing transport %p.\n", ret, rdma);
 560                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
 561                 svc_rdma_put_frmr(rdma, vec->frmr);
 562                 svc_rdma_put_context(ctxt, 0);
 563                 return -ENOTCONN;
 564         }
 565
 566         /* Prepare the context */
 567         ctxt->pages[0] = page;
 568         ctxt->count = 1;
 569         ctxt->frmr = vec->frmr;
 570         if (vec->frmr)
 571                 set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 572         else
 573                 clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 574
 575         /* Prepare the SGE for the RPCRDMA Header */
 576         ctxt->sge[0].lkey = rdma->sc_dma_lkey;
 577         ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
 578         ctxt->sge[0].addr =
 579             ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
 580                             ctxt->sge[0].length, DMA_TO_DEVICE);
 581         if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
 582                 goto err;
 583         atomic_inc(&rdma->sc_dma_used);
 584
 585         ctxt->direction = DMA_TO_DEVICE;
 586
 587         /* Map the payload indicated by 'byte_count' */
 588         for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
 589                 int xdr_off = 0;
 590                 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
 591                 byte_count -= sge_bytes;
 592                 if (!vec->frmr) {
 593                         ctxt->sge[sge_no].addr =
 594                                 dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
 595                                             sge_bytes, DMA_TO_DEVICE);
 596                         xdr_off += sge_bytes;
 597                         if (ib_dma_mapping_error(rdma->sc_cm_id->device,
 598                                                  ctxt->sge[sge_no].addr))
 599                                 goto err;
 600                         atomic_inc(&rdma->sc_dma_used);
 601                         ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
 602                 } else {
 603                         ctxt->sge[sge_no].addr = (unsigned long)
 604                                 vec->sge[sge_no].iov_base;
 605                         ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey;
 606                 }
 607                 ctxt->sge[sge_no].length = sge_bytes;
 608         }
 609         BUG_ON(byte_count != 0);
 610
 611         /* Save all respages in the ctxt and remove them from the
 612          * respages array. They are our pages until the I/O
 613          * completes.
 614          */
 615         pages = rqstp->rq_next_page - rqstp->rq_respages;
 616         for (page_no = 0; page_no < pages; page_no++) {
 617                 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
 618                 ctxt->count++;
 619                 rqstp->rq_respages[page_no] = NULL;
 620                 /*
 621                  * If there are more pages than SGE, terminate SGE
 622                  * list so that svc_rdma_unmap_dma doesn't attempt to
 623                  * unmap garbage.
 624                  */
 625                 if (page_no+1 >= sge_no)
 626                         ctxt->sge[page_no+1].length = 0;
 627         }
 628         BUG_ON(sge_no > rdma->sc_max_sge);
 629         memset(&send_wr, 0, sizeof send_wr);
 630         ctxt->wr_op = IB_WR_SEND;
 631         send_wr.wr_id = (unsigned long)ctxt;
 632         send_wr.sg_list = ctxt->sge;
 633         send_wr.num_sge = sge_no;
 634         send_wr.opcode = IB_WR_SEND;
 635         send_wr.send_flags =  IB_SEND_SIGNALED;
 636         if (vec->frmr) {
 637                 /* Prepare INVALIDATE WR */
 638                 memset(&inv_wr, 0, sizeof inv_wr);
 639                 inv_wr.opcode = IB_WR_LOCAL_INV;
 640                 inv_wr.send_flags = IB_SEND_SIGNALED;
 641                 inv_wr.ex.invalidate_rkey =
 642                         vec->frmr->mr->lkey;
 643                 send_wr.next = &inv_wr;
 644         }
 645
 646         ret = svc_rdma_send(rdma, &send_wr);
 647         if (ret)
 648                 goto err;
 649
 650         return 0;
 651
 652  err:
 653         svc_rdma_unmap_dma(ctxt);
 654         svc_rdma_put_frmr(rdma, vec->frmr);
 655         svc_rdma_put_context(ctxt, 1);
 656         return -EIO;
 657 }
 658
 659 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
 660 {
 661 }
 662
 663 /*
 664  * Return the start of an xdr buffer.
 665  */
 666 static void *xdr_start(struct xdr_buf *xdr)
 667 {
 668         return xdr->head[0].iov_base -
 669                 (xdr->len -
 670                  xdr->page_len -
 671                  xdr->tail[0].iov_len -
 672                  xdr->head[0].iov_len);
 673 }
 674
 675 int svc_rdma_sendto(struct svc_rqst *rqstp)
 676 {
 677         struct svc_xprt *xprt = rqstp->rq_xprt;
 678         struct svcxprt_rdma *rdma =
 679                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 680         struct rpcrdma_msg *rdma_argp;
 681         struct rpcrdma_msg *rdma_resp;
 682         struct rpcrdma_write_array *reply_ary;
 683         enum rpcrdma_proc reply_type;
 684         int ret;
 685         int inline_bytes;
 686         struct page *res_page;
 687         struct svc_rdma_op_ctxt *ctxt;
 688         struct svc_rdma_req_map *vec;
 689
 690         dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
 691
 692         /* Get the RDMA request header. */
 693         rdma_argp = xdr_start(&rqstp->rq_arg);
 694
 695         /* Build an req vec for the XDR */
 696         ctxt = svc_rdma_get_context(rdma);
 697         ctxt->direction = DMA_TO_DEVICE;
 698         vec = svc_rdma_get_req_map();
 699         ret = map_xdr(rdma, &rqstp->rq_res, vec);
 700         if (ret)
 701                 goto err0;
 702         inline_bytes = rqstp->rq_res.len;
 703
 704         /* Create the RDMA response header */
 705         res_page = svc_rdma_get_page();
 706         rdma_resp = page_address(res_page);
 707         reply_ary = svc_rdma_get_reply_array(rdma_argp);
 708         if (reply_ary)
 709                 reply_type = RDMA_NOMSG;
 710         else
 711                 reply_type = RDMA_MSG;
 712         svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
 713                                          rdma_resp, reply_type);
 714
 715         /* Send any write-chunk data and build resp write-list */
 716         ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
 717                                 rqstp, vec);
 718         if (ret < 0) {
 719                 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
 720                        ret);
 721                 goto err1;
 722         }
 723         inline_bytes -= ret;
 724
 725         /* Send any reply-list data and update resp reply-list */
 726         ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
 727                                 rqstp, vec);
 728         if (ret < 0) {
 729                 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
 730                        ret);
 731                 goto err1;
 732         }
 733         inline_bytes -= ret;
 734
 735         ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
 736                          inline_bytes);
 737         svc_rdma_put_req_map(vec);
 738         dprintk("svcrdma: send_reply returns %d\n", ret);
 739         return ret;
 740
 741  err1:
 742         put_page(res_page);
 743  err0:
 744         svc_rdma_put_req_map(vec);
 745         svc_rdma_put_context(ctxt, 0);
 746         return ret;
 747 }