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 contigous 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  */
  67 static struct ib_sge *xdr_to_sge(struct svcxprt_rdma *xprt,
  68                                  struct xdr_buf *xdr,
  69                                  struct ib_sge *sge,
  70                                  int *sge_count)
  71 {
  72         /* Max we need is the length of the XDR / pagesize + one for
  73          * head + one for tail + one for RPCRDMA header
  74          */
  75         int sge_max = (xdr->len+PAGE_SIZE-1) / PAGE_SIZE + 3;
  76         int sge_no;
  77         u32 byte_count = xdr->len;
  78         u32 sge_bytes;
  79         u32 page_bytes;
  80         int page_off;
  81         int page_no;
  82
  83         /* Skip the first sge, this is for the RPCRDMA header */
  84         sge_no = 1;
  85
  86         /* Head SGE */
  87         sge[sge_no].addr = ib_dma_map_single(xprt->sc_cm_id->device,
  88                                              xdr->head[0].iov_base,
  89                                              xdr->head[0].iov_len,
  90                                              DMA_TO_DEVICE);
  91         sge_bytes = min_t(u32, byte_count, xdr->head[0].iov_len);
  92         byte_count -= sge_bytes;
  93         sge[sge_no].length = sge_bytes;
  94         sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
  95         sge_no++;
  96
  97         /* pages SGE */
  98         page_no = 0;
  99         page_bytes = xdr->page_len;
 100         page_off = xdr->page_base;
 101         while (byte_count && page_bytes) {
 102                 sge_bytes = min_t(u32, byte_count, (PAGE_SIZE-page_off));
 103                 sge[sge_no].addr =
 104                         ib_dma_map_page(xprt->sc_cm_id->device,
 105                                         xdr->pages[page_no], page_off,
 106                                         sge_bytes, DMA_TO_DEVICE);
 107                 sge_bytes = min(sge_bytes, page_bytes);
 108                 byte_count -= sge_bytes;
 109                 page_bytes -= sge_bytes;
 110                 sge[sge_no].length = sge_bytes;
 111                 sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 112
 113                 sge_no++;
 114                 page_no++;
 115                 page_off = 0; /* reset for next time through loop */
 116         }
 117
 118         /* Tail SGE */
 119         if (byte_count && xdr->tail[0].iov_len) {
 120                 sge[sge_no].addr =
 121                         ib_dma_map_single(xprt->sc_cm_id->device,
 122                                           xdr->tail[0].iov_base,
 123                                           xdr->tail[0].iov_len,
 124                                           DMA_TO_DEVICE);
 125                 sge_bytes = min_t(u32, byte_count, xdr->tail[0].iov_len);
 126                 byte_count -= sge_bytes;
 127                 sge[sge_no].length = sge_bytes;
 128                 sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 129                 sge_no++;
 130         }
 131
 132         BUG_ON(sge_no > sge_max);
 133         BUG_ON(byte_count != 0);
 134
 135         *sge_count = sge_no;
 136         return sge;
 137 }
 138
 139
 140 /* Assumptions:
 141  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
 142  */
 143 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
 144                       u32 rmr, u64 to,
 145                       u32 xdr_off, int write_len,
 146                       struct ib_sge *xdr_sge, int sge_count)
 147 {
 148         struct svc_rdma_op_ctxt *tmp_sge_ctxt;
 149         struct ib_send_wr write_wr;
 150         struct ib_sge *sge;
 151         int xdr_sge_no;
 152         int sge_no;
 153         int sge_bytes;
 154         int sge_off;
 155         int bc;
 156         struct svc_rdma_op_ctxt *ctxt;
 157         int ret = 0;
 158
 159         BUG_ON(sge_count >= 32);
 160         dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
 161                 "write_len=%d, xdr_sge=%p, sge_count=%d\n",
 162                 rmr, to, xdr_off, write_len, xdr_sge, sge_count);
 163
 164         ctxt = svc_rdma_get_context(xprt);
 165         ctxt->count = 0;
 166         tmp_sge_ctxt = svc_rdma_get_context(xprt);
 167         sge = tmp_sge_ctxt->sge;
 168
 169         /* Find the SGE associated with xdr_off */
 170         for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < sge_count;
 171              xdr_sge_no++) {
 172                 if (xdr_sge[xdr_sge_no].length > bc)
 173                         break;
 174                 bc -= xdr_sge[xdr_sge_no].length;
 175         }
 176
 177         sge_off = bc;
 178         bc = write_len;
 179         sge_no = 0;
 180
 181         /* Copy the remaining SGE */
 182         while (bc != 0 && xdr_sge_no < sge_count) {
 183                 sge[sge_no].addr = xdr_sge[xdr_sge_no].addr + sge_off;
 184                 sge[sge_no].lkey = xdr_sge[xdr_sge_no].lkey;
 185                 sge_bytes = min((size_t)bc,
 186                                 (size_t)(xdr_sge[xdr_sge_no].length-sge_off));
 187                 sge[sge_no].length = sge_bytes;
 188
 189                 sge_off = 0;
 190                 sge_no++;
 191                 xdr_sge_no++;
 192                 bc -= sge_bytes;
 193         }
 194
 195         BUG_ON(bc != 0);
 196         BUG_ON(xdr_sge_no > sge_count);
 197
 198         /* Prepare WRITE WR */
 199         memset(&write_wr, 0, sizeof write_wr);
 200         ctxt->wr_op = IB_WR_RDMA_WRITE;
 201         write_wr.wr_id = (unsigned long)ctxt;
 202         write_wr.sg_list = &sge[0];
 203         write_wr.num_sge = sge_no;
 204         write_wr.opcode = IB_WR_RDMA_WRITE;
 205         write_wr.send_flags = IB_SEND_SIGNALED;
 206         write_wr.wr.rdma.rkey = rmr;
 207         write_wr.wr.rdma.remote_addr = to;
 208
 209         /* Post It */
 210         atomic_inc(&rdma_stat_write);
 211         if (svc_rdma_send(xprt, &write_wr)) {
 212                 svc_rdma_put_context(ctxt, 1);
 213                 /* Fatal error, close transport */
 214                 ret = -EIO;
 215         }
 216         svc_rdma_put_context(tmp_sge_ctxt, 0);
 217         return ret;
 218 }
 219
 220 static int send_write_chunks(struct svcxprt_rdma *xprt,
 221                              struct rpcrdma_msg *rdma_argp,
 222                              struct rpcrdma_msg *rdma_resp,
 223                              struct svc_rqst *rqstp,
 224                              struct ib_sge *sge,
 225                              int sge_count)
 226 {
 227         u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 228         int write_len;
 229         int max_write;
 230         u32 xdr_off;
 231         int chunk_off;
 232         int chunk_no;
 233         struct rpcrdma_write_array *arg_ary;
 234         struct rpcrdma_write_array *res_ary;
 235         int ret;
 236
 237         arg_ary = svc_rdma_get_write_array(rdma_argp);
 238         if (!arg_ary)
 239                 return 0;
 240         res_ary = (struct rpcrdma_write_array *)
 241                 &rdma_resp->rm_body.rm_chunks[1];
 242
 243         max_write = xprt->sc_max_sge * PAGE_SIZE;
 244
 245         /* Write chunks start at the pagelist */
 246         for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
 247              xfer_len && chunk_no < arg_ary->wc_nchunks;
 248              chunk_no++) {
 249                 struct rpcrdma_segment *arg_ch;
 250                 u64 rs_offset;
 251
 252                 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
 253                 write_len = min(xfer_len, arg_ch->rs_length);
 254
 255                 /* Prepare the response chunk given the length actually
 256                  * written */
 257                 rs_offset = get_unaligned(&(arg_ch->rs_offset));
 258                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 259                                             arg_ch->rs_handle,
 260                                             rs_offset,
 261                                             write_len);
 262                 chunk_off = 0;
 263                 while (write_len) {
 264                         int this_write;
 265                         this_write = min(write_len, max_write);
 266                         ret = send_write(xprt, rqstp,
 267                                          arg_ch->rs_handle,
 268                                          rs_offset + chunk_off,
 269                                          xdr_off,
 270                                          this_write,
 271                                          sge,
 272                                          sge_count);
 273                         if (ret) {
 274                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 275                                         ret);
 276                                 return -EIO;
 277                         }
 278                         chunk_off += this_write;
 279                         xdr_off += this_write;
 280                         xfer_len -= this_write;
 281                         write_len -= this_write;
 282                 }
 283         }
 284         /* Update the req with the number of chunks actually used */
 285         svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
 286
 287         return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 288 }
 289
 290 static int send_reply_chunks(struct svcxprt_rdma *xprt,
 291                              struct rpcrdma_msg *rdma_argp,
 292                              struct rpcrdma_msg *rdma_resp,
 293                              struct svc_rqst *rqstp,
 294                              struct ib_sge *sge,
 295                              int sge_count)
 296 {
 297         u32 xfer_len = rqstp->rq_res.len;
 298         int write_len;
 299         int max_write;
 300         u32 xdr_off;
 301         int chunk_no;
 302         int chunk_off;
 303         struct rpcrdma_segment *ch;
 304         struct rpcrdma_write_array *arg_ary;
 305         struct rpcrdma_write_array *res_ary;
 306         int ret;
 307
 308         arg_ary = svc_rdma_get_reply_array(rdma_argp);
 309         if (!arg_ary)
 310                 return 0;
 311         /* XXX: need to fix when reply lists occur with read-list and or
 312          * write-list */
 313         res_ary = (struct rpcrdma_write_array *)
 314                 &rdma_resp->rm_body.rm_chunks[2];
 315
 316         max_write = xprt->sc_max_sge * PAGE_SIZE;
 317
 318         /* xdr offset starts at RPC message */
 319         for (xdr_off = 0, chunk_no = 0;
 320              xfer_len && chunk_no < arg_ary->wc_nchunks;
 321              chunk_no++) {
 322                 u64 rs_offset;
 323                 ch = &arg_ary->wc_array[chunk_no].wc_target;
 324                 write_len = min(xfer_len, ch->rs_length);
 325
 326
 327                 /* Prepare the reply chunk given the length actually
 328                  * written */
 329                 rs_offset = get_unaligned(&(ch->rs_offset));
 330                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 331                                             ch->rs_handle, rs_offset,
 332                                             write_len);
 333                 chunk_off = 0;
 334                 while (write_len) {
 335                         int this_write;
 336
 337                         this_write = min(write_len, max_write);
 338                         ret = send_write(xprt, rqstp,
 339                                          ch->rs_handle,
 340                                          rs_offset + chunk_off,
 341                                          xdr_off,
 342                                          this_write,
 343                                          sge,
 344                                          sge_count);
 345                         if (ret) {
 346                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 347                                         ret);
 348                                 return -EIO;
 349                         }
 350                         chunk_off += this_write;
 351                         xdr_off += this_write;
 352                         xfer_len -= this_write;
 353                         write_len -= this_write;
 354                 }
 355         }
 356         /* Update the req with the number of chunks actually used */
 357         svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
 358
 359         return rqstp->rq_res.len;
 360 }
 361
 362 /* This function prepares the portion of the RPCRDMA message to be
 363  * sent in the RDMA_SEND. This function is called after data sent via
 364  * RDMA has already been transmitted. There are three cases:
 365  * - The RPCRDMA header, RPC header, and payload are all sent in a
 366  *   single RDMA_SEND. This is the "inline" case.
 367  * - The RPCRDMA header and some portion of the RPC header and data
 368  *   are sent via this RDMA_SEND and another portion of the data is
 369  *   sent via RDMA.
 370  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
 371  *   header and data are all transmitted via RDMA.
 372  * In all three cases, this function prepares the RPCRDMA header in
 373  * sge[0], the 'type' parameter indicates the type to place in the
 374  * RPCRDMA header, and the 'byte_count' field indicates how much of
 375  * the XDR to include in this RDMA_SEND.
 376  */
 377 static int send_reply(struct svcxprt_rdma *rdma,
 378                       struct svc_rqst *rqstp,
 379                       struct page *page,
 380                       struct rpcrdma_msg *rdma_resp,
 381                       struct svc_rdma_op_ctxt *ctxt,
 382                       int sge_count,
 383                       int byte_count)
 384 {
 385         struct ib_send_wr send_wr;
 386         int sge_no;
 387         int sge_bytes;
 388         int page_no;
 389         int ret;
 390
 391         /* Prepare the context */
 392         ctxt->pages[0] = page;
 393         ctxt->count = 1;
 394
 395         /* Prepare the SGE for the RPCRDMA Header */
 396         ctxt->sge[0].addr =
 397                 ib_dma_map_page(rdma->sc_cm_id->device,
 398                                 page, 0, PAGE_SIZE, DMA_TO_DEVICE);
 399         ctxt->direction = DMA_TO_DEVICE;
 400         ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
 401         ctxt->sge[0].lkey = rdma->sc_phys_mr->lkey;
 402
 403         /* Determine how many of our SGE are to be transmitted */
 404         for (sge_no = 1; byte_count && sge_no < sge_count; sge_no++) {
 405                 sge_bytes = min((size_t)ctxt->sge[sge_no].length,
 406                                 (size_t)byte_count);
 407                 byte_count -= sge_bytes;
 408         }
 409         BUG_ON(byte_count != 0);
 410
 411         /* Save all respages in the ctxt and remove them from the
 412          * respages array. They are our pages until the I/O
 413          * completes.
 414          */
 415         for (page_no = 0; page_no < rqstp->rq_resused; page_no++) {
 416                 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
 417                 ctxt->count++;
 418                 rqstp->rq_respages[page_no] = NULL;
 419         }
 420
 421         BUG_ON(sge_no > rdma->sc_max_sge);
 422         memset(&send_wr, 0, sizeof send_wr);
 423         ctxt->wr_op = IB_WR_SEND;
 424         send_wr.wr_id = (unsigned long)ctxt;
 425         send_wr.sg_list = ctxt->sge;
 426         send_wr.num_sge = sge_no;
 427         send_wr.opcode = IB_WR_SEND;
 428         send_wr.send_flags =  IB_SEND_SIGNALED;
 429
 430         ret = svc_rdma_send(rdma, &send_wr);
 431         if (ret)
 432                 svc_rdma_put_context(ctxt, 1);
 433
 434         return ret;
 435 }
 436
 437 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
 438 {
 439 }
 440
 441 /*
 442  * Return the start of an xdr buffer.
 443  */
 444 static void *xdr_start(struct xdr_buf *xdr)
 445 {
 446         return xdr->head[0].iov_base -
 447                 (xdr->len -
 448                  xdr->page_len -
 449                  xdr->tail[0].iov_len -
 450                  xdr->head[0].iov_len);
 451 }
 452
 453 int svc_rdma_sendto(struct svc_rqst *rqstp)
 454 {
 455         struct svc_xprt *xprt = rqstp->rq_xprt;
 456         struct svcxprt_rdma *rdma =
 457                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 458         struct rpcrdma_msg *rdma_argp;
 459         struct rpcrdma_msg *rdma_resp;
 460         struct rpcrdma_write_array *reply_ary;
 461         enum rpcrdma_proc reply_type;
 462         int ret;
 463         int inline_bytes;
 464         struct ib_sge *sge;
 465         int sge_count = 0;
 466         struct page *res_page;
 467         struct svc_rdma_op_ctxt *ctxt;
 468
 469         dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
 470
 471         /* Get the RDMA request header. */
 472         rdma_argp = xdr_start(&rqstp->rq_arg);
 473
 474         /* Build an SGE for the XDR */
 475         ctxt = svc_rdma_get_context(rdma);
 476         ctxt->direction = DMA_TO_DEVICE;
 477         sge = xdr_to_sge(rdma, &rqstp->rq_res, ctxt->sge, &sge_count);
 478
 479         inline_bytes = rqstp->rq_res.len;
 480
 481         /* Create the RDMA response header */
 482         res_page = svc_rdma_get_page();
 483         rdma_resp = page_address(res_page);
 484         reply_ary = svc_rdma_get_reply_array(rdma_argp);
 485         if (reply_ary)
 486                 reply_type = RDMA_NOMSG;
 487         else
 488                 reply_type = RDMA_MSG;
 489         svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
 490                                          rdma_resp, reply_type);
 491
 492         /* Send any write-chunk data and build resp write-list */
 493         ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
 494                                 rqstp, sge, sge_count);
 495         if (ret < 0) {
 496                 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
 497                        ret);
 498                 goto error;
 499         }
 500         inline_bytes -= ret;
 501
 502         /* Send any reply-list data and update resp reply-list */
 503         ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
 504                                 rqstp, sge, sge_count);
 505         if (ret < 0) {
 506                 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
 507                        ret);
 508                 goto error;
 509         }
 510         inline_bytes -= ret;
 511
 512         ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, sge_count,
 513                          inline_bytes);
 514         dprintk("svcrdma: send_reply returns %d\n", ret);
 515         return ret;
 516  error:
 517         svc_rdma_put_context(ctxt, 0);
 518         put_page(res_page);
 519         return ret;
 520 }