git.oblomov.eu Git - linux-2.6/blob - 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  * 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 void xdr_to_sge(struct svcxprt_rdma *xprt,
  73                        struct xdr_buf *xdr,
  74                        struct svc_rdma_req_map *vec)
  75 {
  76         int sge_max = (xdr->len+PAGE_SIZE-1) / PAGE_SIZE + 3;
  77         int sge_no;
  78         u32 sge_bytes;
  79         u32 page_bytes;
  80         u32 page_off;
  81         int page_no;
  82
  83         BUG_ON(xdr->len !=
  84                (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
  85
  86         /* Skip the first sge, this is for the RPCRDMA header */
  87         sge_no = 1;
  88
  89         /* Head SGE */
  90         vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
  91         vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
  92         sge_no++;
  93
  94         /* pages SGE */
  95         page_no = 0;
  96         page_bytes = xdr->page_len;
  97         page_off = xdr->page_base;
  98         while (page_bytes) {
  99                 vec->sge[sge_no].iov_base =
 100                         page_address(xdr->pages[page_no]) + page_off;
 101                 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
 102                 page_bytes -= sge_bytes;
 103                 vec->sge[sge_no].iov_len = sge_bytes;
 104
 105                 sge_no++;
 106                 page_no++;
 107                 page_off = 0; /* reset for next time through loop */
 108         }
 109
 110         /* Tail SGE */
 111         if (xdr->tail[0].iov_len) {
 112                 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
 113                 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
 114                 sge_no++;
 115         }
 116
 117         BUG_ON(sge_no > sge_max);
 118         vec->count = sge_no;
 119 }
 120
 121 /* Assumptions:
 122  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
 123  */
 124 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
 125                       u32 rmr, u64 to,
 126                       u32 xdr_off, int write_len,
 127                       struct svc_rdma_req_map *vec)
 128 {
 129         struct ib_send_wr write_wr;
 130         struct ib_sge *sge;
 131         int xdr_sge_no;
 132         int sge_no;
 133         int sge_bytes;
 134         int sge_off;
 135         int bc;
 136         struct svc_rdma_op_ctxt *ctxt;
 137
 138         BUG_ON(vec->count > RPCSVC_MAXPAGES);
 139         dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
 140                 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
 141                 rmr, (unsigned long long)to, xdr_off,
 142                 write_len, vec->sge, vec->count);
 143
 144         ctxt = svc_rdma_get_context(xprt);
 145         ctxt->direction = DMA_TO_DEVICE;
 146         sge = ctxt->sge;
 147
 148         /* Find the SGE associated with xdr_off */
 149         for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
 150              xdr_sge_no++) {
 151                 if (vec->sge[xdr_sge_no].iov_len > bc)
 152                         break;
 153                 bc -= vec->sge[xdr_sge_no].iov_len;
 154         }
 155
 156         sge_off = bc;
 157         bc = write_len;
 158         sge_no = 0;
 159
 160         /* Copy the remaining SGE */
 161         while (bc != 0 && xdr_sge_no < vec->count) {
 162                 sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 163                 sge_bytes = min((size_t)bc,
 164                                 (size_t)(vec->sge[xdr_sge_no].iov_len-sge_off));
 165                 sge[sge_no].length = sge_bytes;
 166                 atomic_inc(&xprt->sc_dma_used);
 167                 sge[sge_no].addr =
 168                         ib_dma_map_single(xprt->sc_cm_id->device,
 169                                           (void *)
 170                                           vec->sge[xdr_sge_no].iov_base + sge_off,
 171                                           sge_bytes, DMA_TO_DEVICE);
 172                 if (dma_mapping_error(sge[sge_no].addr))
 173                         goto err;
 174                 sge_off = 0;
 175                 sge_no++;
 176                 ctxt->count++;
 177                 xdr_sge_no++;
 178                 bc -= sge_bytes;
 179         }
 180
 181         BUG_ON(bc != 0);
 182         BUG_ON(xdr_sge_no > vec->count);
 183
 184         /* Prepare WRITE WR */
 185         memset(&write_wr, 0, sizeof write_wr);
 186         ctxt->wr_op = IB_WR_RDMA_WRITE;
 187         write_wr.wr_id = (unsigned long)ctxt;
 188         write_wr.sg_list = &sge[0];
 189         write_wr.num_sge = sge_no;
 190         write_wr.opcode = IB_WR_RDMA_WRITE;
 191         write_wr.send_flags = IB_SEND_SIGNALED;
 192         write_wr.wr.rdma.rkey = rmr;
 193         write_wr.wr.rdma.remote_addr = to;
 194
 195         /* Post It */
 196         atomic_inc(&rdma_stat_write);
 197         if (svc_rdma_send(xprt, &write_wr))
 198                 goto err;
 199         return 0;
 200  err:
 201         svc_rdma_put_context(ctxt, 0);
 202         /* Fatal error, close transport */
 203         return -EIO;
 204 }
 205
 206 static int send_write_chunks(struct svcxprt_rdma *xprt,
 207                              struct rpcrdma_msg *rdma_argp,
 208                              struct rpcrdma_msg *rdma_resp,
 209                              struct svc_rqst *rqstp,
 210                              struct svc_rdma_req_map *vec)
 211 {
 212         u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 213         int write_len;
 214         int max_write;
 215         u32 xdr_off;
 216         int chunk_off;
 217         int chunk_no;
 218         struct rpcrdma_write_array *arg_ary;
 219         struct rpcrdma_write_array *res_ary;
 220         int ret;
 221
 222         arg_ary = svc_rdma_get_write_array(rdma_argp);
 223         if (!arg_ary)
 224                 return 0;
 225         res_ary = (struct rpcrdma_write_array *)
 226                 &rdma_resp->rm_body.rm_chunks[1];
 227
 228         max_write = xprt->sc_max_sge * PAGE_SIZE;
 229
 230         /* Write chunks start at the pagelist */
 231         for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
 232              xfer_len && chunk_no < arg_ary->wc_nchunks;
 233              chunk_no++) {
 234                 struct rpcrdma_segment *arg_ch;
 235                 u64 rs_offset;
 236
 237                 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
 238                 write_len = min(xfer_len, arg_ch->rs_length);
 239
 240                 /* Prepare the response chunk given the length actually
 241                  * written */
 242                 rs_offset = get_unaligned(&(arg_ch->rs_offset));
 243                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 244                                             arg_ch->rs_handle,
 245                                             rs_offset,
 246                                             write_len);
 247                 chunk_off = 0;
 248                 while (write_len) {
 249                         int this_write;
 250                         this_write = min(write_len, max_write);
 251                         ret = send_write(xprt, rqstp,
 252                                          arg_ch->rs_handle,
 253                                          rs_offset + chunk_off,
 254                                          xdr_off,
 255                                          this_write,
 256                                          vec);
 257                         if (ret) {
 258                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 259                                         ret);
 260                                 return -EIO;
 261                         }
 262                         chunk_off += this_write;
 263                         xdr_off += this_write;
 264                         xfer_len -= this_write;
 265                         write_len -= this_write;
 266                 }
 267         }
 268         /* Update the req with the number of chunks actually used */
 269         svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
 270
 271         return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 272 }
 273
 274 static int send_reply_chunks(struct svcxprt_rdma *xprt,
 275                              struct rpcrdma_msg *rdma_argp,
 276                              struct rpcrdma_msg *rdma_resp,
 277                              struct svc_rqst *rqstp,
 278                              struct svc_rdma_req_map *vec)
 279 {
 280         u32 xfer_len = rqstp->rq_res.len;
 281         int write_len;
 282         int max_write;
 283         u32 xdr_off;
 284         int chunk_no;
 285         int chunk_off;
 286         struct rpcrdma_segment *ch;
 287         struct rpcrdma_write_array *arg_ary;
 288         struct rpcrdma_write_array *res_ary;
 289         int ret;
 290
 291         arg_ary = svc_rdma_get_reply_array(rdma_argp);
 292         if (!arg_ary)
 293                 return 0;
 294         /* XXX: need to fix when reply lists occur with read-list and or
 295          * write-list */
 296         res_ary = (struct rpcrdma_write_array *)
 297                 &rdma_resp->rm_body.rm_chunks[2];
 298
 299         max_write = xprt->sc_max_sge * PAGE_SIZE;
 300
 301         /* xdr offset starts at RPC message */
 302         for (xdr_off = 0, chunk_no = 0;
 303              xfer_len && chunk_no < arg_ary->wc_nchunks;
 304              chunk_no++) {
 305                 u64 rs_offset;
 306                 ch = &arg_ary->wc_array[chunk_no].wc_target;
 307                 write_len = min(xfer_len, ch->rs_length);
 308
 309
 310                 /* Prepare the reply chunk given the length actually
 311                  * written */
 312                 rs_offset = get_unaligned(&(ch->rs_offset));
 313                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 314                                             ch->rs_handle, rs_offset,
 315                                             write_len);
 316                 chunk_off = 0;
 317                 while (write_len) {
 318                         int this_write;
 319
 320                         this_write = min(write_len, max_write);
 321                         ret = send_write(xprt, rqstp,
 322                                          ch->rs_handle,
 323                                          rs_offset + chunk_off,
 324                                          xdr_off,
 325                                          this_write,
 326                                          vec);
 327                         if (ret) {
 328                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 329                                         ret);
 330                                 return -EIO;
 331                         }
 332                         chunk_off += this_write;
 333                         xdr_off += this_write;
 334                         xfer_len -= this_write;
 335                         write_len -= this_write;
 336                 }
 337         }
 338         /* Update the req with the number of chunks actually used */
 339         svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
 340
 341         return rqstp->rq_res.len;
 342 }
 343
 344 /* This function prepares the portion of the RPCRDMA message to be
 345  * sent in the RDMA_SEND. This function is called after data sent via
 346  * RDMA has already been transmitted. There are three cases:
 347  * - The RPCRDMA header, RPC header, and payload are all sent in a
 348  *   single RDMA_SEND. This is the "inline" case.
 349  * - The RPCRDMA header and some portion of the RPC header and data
 350  *   are sent via this RDMA_SEND and another portion of the data is
 351  *   sent via RDMA.
 352  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
 353  *   header and data are all transmitted via RDMA.
 354  * In all three cases, this function prepares the RPCRDMA header in
 355  * sge[0], the 'type' parameter indicates the type to place in the
 356  * RPCRDMA header, and the 'byte_count' field indicates how much of
 357  * the XDR to include in this RDMA_SEND.
 358  */
 359 static int send_reply(struct svcxprt_rdma *rdma,
 360                       struct svc_rqst *rqstp,
 361                       struct page *page,
 362                       struct rpcrdma_msg *rdma_resp,
 363                       struct svc_rdma_op_ctxt *ctxt,
 364                       struct svc_rdma_req_map *vec,
 365                       int byte_count)
 366 {
 367         struct ib_send_wr send_wr;
 368         int sge_no;
 369         int sge_bytes;
 370         int page_no;
 371         int ret;
 372
 373         /* Post a recv buffer to handle another request. */
 374         ret = svc_rdma_post_recv(rdma);
 375         if (ret) {
 376                 printk(KERN_INFO
 377                        "svcrdma: could not post a receive buffer, err=%d."
 378                        "Closing transport %p.\n", ret, rdma);
 379                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
 380                 svc_rdma_put_context(ctxt, 0);
 381                 return -ENOTCONN;
 382         }
 383
 384         /* Prepare the context */
 385         ctxt->pages[0] = page;
 386         ctxt->count = 1;
 387
 388         /* Prepare the SGE for the RPCRDMA Header */
 389         atomic_inc(&rdma->sc_dma_used);
 390         ctxt->sge[0].addr =
 391                 ib_dma_map_page(rdma->sc_cm_id->device,
 392                                 page, 0, PAGE_SIZE, DMA_TO_DEVICE);
 393         ctxt->direction = DMA_TO_DEVICE;
 394         ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
 395         ctxt->sge[0].lkey = rdma->sc_phys_mr->lkey;
 396
 397         /* Determine how many of our SGE are to be transmitted */
 398         for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
 399                 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
 400                 byte_count -= sge_bytes;
 401                 atomic_inc(&rdma->sc_dma_used);
 402                 ctxt->sge[sge_no].addr =
 403                         ib_dma_map_single(rdma->sc_cm_id->device,
 404                                           vec->sge[sge_no].iov_base,
 405                                           sge_bytes, DMA_TO_DEVICE);
 406                 ctxt->sge[sge_no].length = sge_bytes;
 407                 ctxt->sge[sge_no].lkey = rdma->sc_phys_mr->lkey;
 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                 /* If there are more pages than SGE, terminate SGE list */
 420                 if (page_no+1 >= sge_no)
 421                         ctxt->sge[page_no+1].length = 0;
 422         }
 423         BUG_ON(sge_no > rdma->sc_max_sge);
 424         memset(&send_wr, 0, sizeof send_wr);
 425         ctxt->wr_op = IB_WR_SEND;
 426         send_wr.wr_id = (unsigned long)ctxt;
 427         send_wr.sg_list = ctxt->sge;
 428         send_wr.num_sge = sge_no;
 429         send_wr.opcode = IB_WR_SEND;
 430         send_wr.send_flags =  IB_SEND_SIGNALED;
 431
 432         ret = svc_rdma_send(rdma, &send_wr);
 433         if (ret)
 434                 svc_rdma_put_context(ctxt, 1);
 435
 436         return ret;
 437 }
 438
 439 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
 440 {
 441 }
 442
 443 /*
 444  * Return the start of an xdr buffer.
 445  */
 446 static void *xdr_start(struct xdr_buf *xdr)
 447 {
 448         return xdr->head[0].iov_base -
 449                 (xdr->len -
 450                  xdr->page_len -
 451                  xdr->tail[0].iov_len -
 452                  xdr->head[0].iov_len);
 453 }
 454
 455 int svc_rdma_sendto(struct svc_rqst *rqstp)
 456 {
 457         struct svc_xprt *xprt = rqstp->rq_xprt;
 458         struct svcxprt_rdma *rdma =
 459                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 460         struct rpcrdma_msg *rdma_argp;
 461         struct rpcrdma_msg *rdma_resp;
 462         struct rpcrdma_write_array *reply_ary;
 463         enum rpcrdma_proc reply_type;
 464         int ret;
 465         int inline_bytes;
 466         struct page *res_page;
 467         struct svc_rdma_op_ctxt *ctxt;
 468         struct svc_rdma_req_map *vec;
 469
 470         dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
 471
 472         /* Get the RDMA request header. */
 473         rdma_argp = xdr_start(&rqstp->rq_arg);
 474
 475         /* Build an req vec for the XDR */
 476         ctxt = svc_rdma_get_context(rdma);
 477         ctxt->direction = DMA_TO_DEVICE;
 478         vec = svc_rdma_get_req_map();
 479         xdr_to_sge(rdma, &rqstp->rq_res, vec);
 480
 481         inline_bytes = rqstp->rq_res.len;
 482
 483         /* Create the RDMA response header */
 484         res_page = svc_rdma_get_page();
 485         rdma_resp = page_address(res_page);
 486         reply_ary = svc_rdma_get_reply_array(rdma_argp);
 487         if (reply_ary)
 488                 reply_type = RDMA_NOMSG;
 489         else
 490                 reply_type = RDMA_MSG;
 491         svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
 492                                          rdma_resp, reply_type);
 493
 494         /* Send any write-chunk data and build resp write-list */
 495         ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
 496                                 rqstp, vec);
 497         if (ret < 0) {
 498                 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
 499                        ret);
 500                 goto error;
 501         }
 502         inline_bytes -= ret;
 503
 504         /* Send any reply-list data and update resp reply-list */
 505         ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
 506                                 rqstp, vec);
 507         if (ret < 0) {
 508                 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
 509                        ret);
 510                 goto error;
 511         }
 512         inline_bytes -= ret;
 513
 514         ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
 515                          inline_bytes);
 516         svc_rdma_put_req_map(vec);
 517         dprintk("svcrdma: send_reply returns %d\n", ret);
 518         return ret;
 519  error:
 520         svc_rdma_put_req_map(vec);
 521         svc_rdma_put_context(ctxt, 0);
 522         put_page(res_page);
 523         return ret;
 524 }