SAFE public projects git trees. - safe/jmp/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                 sge[sge_no].addr =
 167                         ib_dma_map_single(xprt->sc_cm_id->device,
 168                                           (void *)
 169                                           vec->sge[xdr_sge_no].iov_base + sge_off,
 170                                           sge_bytes, DMA_TO_DEVICE);
 171                 if (dma_mapping_error(sge[sge_no].addr))
 172                         goto err;
 173                 sge_off = 0;
 174                 sge_no++;
 175                 ctxt->count++;
 176                 xdr_sge_no++;
 177                 bc -= sge_bytes;
 178         }
 179
 180         BUG_ON(bc != 0);
 181         BUG_ON(xdr_sge_no > vec->count);
 182
 183         /* Prepare WRITE WR */
 184         memset(&write_wr, 0, sizeof write_wr);
 185         ctxt->wr_op = IB_WR_RDMA_WRITE;
 186         write_wr.wr_id = (unsigned long)ctxt;
 187         write_wr.sg_list = &sge[0];
 188         write_wr.num_sge = sge_no;
 189         write_wr.opcode = IB_WR_RDMA_WRITE;
 190         write_wr.send_flags = IB_SEND_SIGNALED;
 191         write_wr.wr.rdma.rkey = rmr;
 192         write_wr.wr.rdma.remote_addr = to;
 193
 194         /* Post It */
 195         atomic_inc(&rdma_stat_write);
 196         if (svc_rdma_send(xprt, &write_wr))
 197                 goto err;
 198         return 0;
 199  err:
 200         svc_rdma_put_context(ctxt, 0);
 201         /* Fatal error, close transport */
 202         return -EIO;
 203 }
 204
 205 static int send_write_chunks(struct svcxprt_rdma *xprt,
 206                              struct rpcrdma_msg *rdma_argp,
 207                              struct rpcrdma_msg *rdma_resp,
 208                              struct svc_rqst *rqstp,
 209                              struct svc_rdma_req_map *vec)
 210 {
 211         u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 212         int write_len;
 213         int max_write;
 214         u32 xdr_off;
 215         int chunk_off;
 216         int chunk_no;
 217         struct rpcrdma_write_array *arg_ary;
 218         struct rpcrdma_write_array *res_ary;
 219         int ret;
 220
 221         arg_ary = svc_rdma_get_write_array(rdma_argp);
 222         if (!arg_ary)
 223                 return 0;
 224         res_ary = (struct rpcrdma_write_array *)
 225                 &rdma_resp->rm_body.rm_chunks[1];
 226
 227         max_write = xprt->sc_max_sge * PAGE_SIZE;
 228
 229         /* Write chunks start at the pagelist */
 230         for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
 231              xfer_len && chunk_no < arg_ary->wc_nchunks;
 232              chunk_no++) {
 233                 struct rpcrdma_segment *arg_ch;
 234                 u64 rs_offset;
 235
 236                 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
 237                 write_len = min(xfer_len, arg_ch->rs_length);
 238
 239                 /* Prepare the response chunk given the length actually
 240                  * written */
 241                 rs_offset = get_unaligned(&(arg_ch->rs_offset));
 242                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 243                                             arg_ch->rs_handle,
 244                                             rs_offset,
 245                                             write_len);
 246                 chunk_off = 0;
 247                 while (write_len) {
 248                         int this_write;
 249                         this_write = min(write_len, max_write);
 250                         ret = send_write(xprt, rqstp,
 251                                          arg_ch->rs_handle,
 252                                          rs_offset + chunk_off,
 253                                          xdr_off,
 254                                          this_write,
 255                                          vec);
 256                         if (ret) {
 257                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 258                                         ret);
 259                                 return -EIO;
 260                         }
 261                         chunk_off += this_write;
 262                         xdr_off += this_write;
 263                         xfer_len -= this_write;
 264                         write_len -= this_write;
 265                 }
 266         }
 267         /* Update the req with the number of chunks actually used */
 268         svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
 269
 270         return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
 271 }
 272
 273 static int send_reply_chunks(struct svcxprt_rdma *xprt,
 274                              struct rpcrdma_msg *rdma_argp,
 275                              struct rpcrdma_msg *rdma_resp,
 276                              struct svc_rqst *rqstp,
 277                              struct svc_rdma_req_map *vec)
 278 {
 279         u32 xfer_len = rqstp->rq_res.len;
 280         int write_len;
 281         int max_write;
 282         u32 xdr_off;
 283         int chunk_no;
 284         int chunk_off;
 285         struct rpcrdma_segment *ch;
 286         struct rpcrdma_write_array *arg_ary;
 287         struct rpcrdma_write_array *res_ary;
 288         int ret;
 289
 290         arg_ary = svc_rdma_get_reply_array(rdma_argp);
 291         if (!arg_ary)
 292                 return 0;
 293         /* XXX: need to fix when reply lists occur with read-list and or
 294          * write-list */
 295         res_ary = (struct rpcrdma_write_array *)
 296                 &rdma_resp->rm_body.rm_chunks[2];
 297
 298         max_write = xprt->sc_max_sge * PAGE_SIZE;
 299
 300         /* xdr offset starts at RPC message */
 301         for (xdr_off = 0, chunk_no = 0;
 302              xfer_len && chunk_no < arg_ary->wc_nchunks;
 303              chunk_no++) {
 304                 u64 rs_offset;
 305                 ch = &arg_ary->wc_array[chunk_no].wc_target;
 306                 write_len = min(xfer_len, ch->rs_length);
 307
 308
 309                 /* Prepare the reply chunk given the length actually
 310                  * written */
 311                 rs_offset = get_unaligned(&(ch->rs_offset));
 312                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
 313                                             ch->rs_handle, rs_offset,
 314                                             write_len);
 315                 chunk_off = 0;
 316                 while (write_len) {
 317                         int this_write;
 318
 319                         this_write = min(write_len, max_write);
 320                         ret = send_write(xprt, rqstp,
 321                                          ch->rs_handle,
 322                                          rs_offset + chunk_off,
 323                                          xdr_off,
 324                                          this_write,
 325                                          vec);
 326                         if (ret) {
 327                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
 328                                         ret);
 329                                 return -EIO;
 330                         }
 331                         chunk_off += this_write;
 332                         xdr_off += this_write;
 333                         xfer_len -= this_write;
 334                         write_len -= this_write;
 335                 }
 336         }
 337         /* Update the req with the number of chunks actually used */
 338         svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
 339
 340         return rqstp->rq_res.len;
 341 }
 342
 343 /* This function prepares the portion of the RPCRDMA message to be
 344  * sent in the RDMA_SEND. This function is called after data sent via
 345  * RDMA has already been transmitted. There are three cases:
 346  * - The RPCRDMA header, RPC header, and payload are all sent in a
 347  *   single RDMA_SEND. This is the "inline" case.
 348  * - The RPCRDMA header and some portion of the RPC header and data
 349  *   are sent via this RDMA_SEND and another portion of the data is
 350  *   sent via RDMA.
 351  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
 352  *   header and data are all transmitted via RDMA.
 353  * In all three cases, this function prepares the RPCRDMA header in
 354  * sge[0], the 'type' parameter indicates the type to place in the
 355  * RPCRDMA header, and the 'byte_count' field indicates how much of
 356  * the XDR to include in this RDMA_SEND.
 357  */
 358 static int send_reply(struct svcxprt_rdma *rdma,
 359                       struct svc_rqst *rqstp,
 360                       struct page *page,
 361                       struct rpcrdma_msg *rdma_resp,
 362                       struct svc_rdma_op_ctxt *ctxt,
 363                       struct svc_rdma_req_map *vec,
 364                       int byte_count)
 365 {
 366         struct ib_send_wr send_wr;
 367         int sge_no;
 368         int sge_bytes;
 369         int page_no;
 370         int ret;
 371
 372         /* Post a recv buffer to handle another request. */
 373         ret = svc_rdma_post_recv(rdma);
 374         if (ret) {
 375                 printk(KERN_INFO
 376                        "svcrdma: could not post a receive buffer, err=%d."
 377                        "Closing transport %p.\n", ret, rdma);
 378                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
 379                 svc_rdma_put_context(ctxt, 0);
 380                 return -ENOTCONN;
 381         }
 382
 383         /* Prepare the context */
 384         ctxt->pages[0] = page;
 385         ctxt->count = 1;
 386
 387         /* Prepare the SGE for the RPCRDMA Header */
 388         ctxt->sge[0].addr =
 389                 ib_dma_map_page(rdma->sc_cm_id->device,
 390                                 page, 0, PAGE_SIZE, DMA_TO_DEVICE);
 391         ctxt->direction = DMA_TO_DEVICE;
 392         ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
 393         ctxt->sge[0].lkey = rdma->sc_phys_mr->lkey;
 394
 395         /* Determine how many of our SGE are to be transmitted */
 396         for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
 397                 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
 398                 byte_count -= sge_bytes;
 399                 ctxt->sge[sge_no].addr =
 400                         ib_dma_map_single(rdma->sc_cm_id->device,
 401                                           vec->sge[sge_no].iov_base,
 402                                           sge_bytes, DMA_TO_DEVICE);
 403                 ctxt->sge[sge_no].length = sge_bytes;
 404                 ctxt->sge[sge_no].lkey = rdma->sc_phys_mr->lkey;
 405         }
 406         BUG_ON(byte_count != 0);
 407
 408         /* Save all respages in the ctxt and remove them from the
 409          * respages array. They are our pages until the I/O
 410          * completes.
 411          */
 412         for (page_no = 0; page_no < rqstp->rq_resused; page_no++) {
 413                 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
 414                 ctxt->count++;
 415                 rqstp->rq_respages[page_no] = NULL;
 416                 /* If there are more pages than SGE, terminate SGE list */
 417                 if (page_no+1 >= sge_no)
 418                         ctxt->sge[page_no+1].length = 0;
 419         }
 420         BUG_ON(sge_no > rdma->sc_max_sge);
 421         memset(&send_wr, 0, sizeof send_wr);
 422         ctxt->wr_op = IB_WR_SEND;
 423         send_wr.wr_id = (unsigned long)ctxt;
 424         send_wr.sg_list = ctxt->sge;
 425         send_wr.num_sge = sge_no;
 426         send_wr.opcode = IB_WR_SEND;
 427         send_wr.send_flags =  IB_SEND_SIGNALED;
 428
 429         ret = svc_rdma_send(rdma, &send_wr);
 430         if (ret)
 431                 svc_rdma_put_context(ctxt, 1);
 432
 433         return ret;
 434 }
 435
 436 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
 437 {
 438 }
 439
 440 /*
 441  * Return the start of an xdr buffer.
 442  */
 443 static void *xdr_start(struct xdr_buf *xdr)
 444 {
 445         return xdr->head[0].iov_base -
 446                 (xdr->len -
 447                  xdr->page_len -
 448                  xdr->tail[0].iov_len -
 449                  xdr->head[0].iov_len);
 450 }
 451
 452 int svc_rdma_sendto(struct svc_rqst *rqstp)
 453 {
 454         struct svc_xprt *xprt = rqstp->rq_xprt;
 455         struct svcxprt_rdma *rdma =
 456                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 457         struct rpcrdma_msg *rdma_argp;
 458         struct rpcrdma_msg *rdma_resp;
 459         struct rpcrdma_write_array *reply_ary;
 460         enum rpcrdma_proc reply_type;
 461         int ret;
 462         int inline_bytes;
 463         struct page *res_page;
 464         struct svc_rdma_op_ctxt *ctxt;
 465         struct svc_rdma_req_map *vec;
 466
 467         dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
 468
 469         /* Get the RDMA request header. */
 470         rdma_argp = xdr_start(&rqstp->rq_arg);
 471
 472         /* Build an req vec for the XDR */
 473         ctxt = svc_rdma_get_context(rdma);
 474         ctxt->direction = DMA_TO_DEVICE;
 475         vec = svc_rdma_get_req_map();
 476         xdr_to_sge(rdma, &rqstp->rq_res, vec);
 477
 478         inline_bytes = rqstp->rq_res.len;
 479
 480         /* Create the RDMA response header */
 481         res_page = svc_rdma_get_page();
 482         rdma_resp = page_address(res_page);
 483         reply_ary = svc_rdma_get_reply_array(rdma_argp);
 484         if (reply_ary)
 485                 reply_type = RDMA_NOMSG;
 486         else
 487                 reply_type = RDMA_MSG;
 488         svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
 489                                          rdma_resp, reply_type);
 490
 491         /* Send any write-chunk data and build resp write-list */
 492         ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
 493                                 rqstp, vec);
 494         if (ret < 0) {
 495                 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
 496                        ret);
 497                 goto error;
 498         }
 499         inline_bytes -= ret;
 500
 501         /* Send any reply-list data and update resp reply-list */
 502         ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
 503                                 rqstp, vec);
 504         if (ret < 0) {
 505                 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
 506                        ret);
 507                 goto error;
 508         }
 509         inline_bytes -= ret;
 510
 511         ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
 512                          inline_bytes);
 513         svc_rdma_put_req_map(vec);
 514         dprintk("svcrdma: send_reply returns %d\n", ret);
 515         return ret;
 516  error:
 517         svc_rdma_put_req_map(vec);
 518         svc_rdma_put_context(ctxt, 0);
 519         put_page(res_page);
 520         return ret;
 521 }