net/sunrpc/xprtrdma/svc_rdma_recvfrom.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 /*
  53  * Replace the pages in the rq_argpages array with the pages from the SGE in
  54  * the RDMA_RECV completion. The SGL should contain full pages up until the
  55  * last one.
  56  */
  57 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
  58                                struct svc_rdma_op_ctxt *ctxt,
  59                                u32 byte_count)
  60 {
  61         struct page *page;
  62         u32 bc;
  63         int sge_no;
  64
  65         /* Swap the page in the SGE with the page in argpages */
  66         page = ctxt->pages[0];
  67         put_page(rqstp->rq_pages[0]);
  68         rqstp->rq_pages[0] = page;
  69
  70         /* Set up the XDR head */
  71         rqstp->rq_arg.head[0].iov_base = page_address(page);
  72         rqstp->rq_arg.head[0].iov_len = min(byte_count, ctxt->sge[0].length);
  73         rqstp->rq_arg.len = byte_count;
  74         rqstp->rq_arg.buflen = byte_count;
  75
  76         /* Compute bytes past head in the SGL */
  77         bc = byte_count - rqstp->rq_arg.head[0].iov_len;
  78
  79         /* If data remains, store it in the pagelist */
  80         rqstp->rq_arg.page_len = bc;
  81         rqstp->rq_arg.page_base = 0;
  82         rqstp->rq_arg.pages = &rqstp->rq_pages[1];
  83         sge_no = 1;
  84         while (bc && sge_no < ctxt->count) {
  85                 page = ctxt->pages[sge_no];
  86                 put_page(rqstp->rq_pages[sge_no]);
  87                 rqstp->rq_pages[sge_no] = page;
  88                 bc -= min(bc, ctxt->sge[sge_no].length);
  89                 rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
  90                 sge_no++;
  91         }
  92         rqstp->rq_respages = &rqstp->rq_pages[sge_no];
  93
  94         /* We should never run out of SGE because the limit is defined to
  95          * support the max allowed RPC data length
  96          */
  97         BUG_ON(bc && (sge_no == ctxt->count));
  98         BUG_ON((rqstp->rq_arg.head[0].iov_len + rqstp->rq_arg.page_len)
  99                != byte_count);
 100         BUG_ON(rqstp->rq_arg.len != byte_count);
 101
 102         /* If not all pages were used from the SGL, free the remaining ones */
 103         bc = sge_no;
 104         while (sge_no < ctxt->count) {
 105                 page = ctxt->pages[sge_no++];
 106                 put_page(page);
 107         }
 108         ctxt->count = bc;
 109
 110         /* Set up tail */
 111         rqstp->rq_arg.tail[0].iov_base = NULL;
 112         rqstp->rq_arg.tail[0].iov_len = 0;
 113 }
 114
 115 /* Encode a read-chunk-list as an array of IB SGE
 116  *
 117  * Assumptions:
 118  * - chunk[0]->position points to pages[0] at an offset of 0
 119  * - pages[] is not physically or virtually contiguous and consists of
 120  *   PAGE_SIZE elements.
 121  *
 122  * Output:
 123  * - sge array pointing into pages[] array.
 124  * - chunk_sge array specifying sge index and count for each
 125  *   chunk in the read list
 126  *
 127  */
 128 static int map_read_chunks(struct svcxprt_rdma *xprt,
 129                            struct svc_rqst *rqstp,
 130                            struct svc_rdma_op_ctxt *head,
 131                            struct rpcrdma_msg *rmsgp,
 132                            struct svc_rdma_req_map *rpl_map,
 133                            struct svc_rdma_req_map *chl_map,
 134                            int ch_count,
 135                            int byte_count)
 136 {
 137         int sge_no;
 138         int sge_bytes;
 139         int page_off;
 140         int page_no;
 141         int ch_bytes;
 142         int ch_no;
 143         struct rpcrdma_read_chunk *ch;
 144
 145         sge_no = 0;
 146         page_no = 0;
 147         page_off = 0;
 148         ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 149         ch_no = 0;
 150         ch_bytes = ch->rc_target.rs_length;
 151         head->arg.head[0] = rqstp->rq_arg.head[0];
 152         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 153         head->arg.pages = &head->pages[head->count];
 154         head->hdr_count = head->count; /* save count of hdr pages */
 155         head->arg.page_base = 0;
 156         head->arg.page_len = ch_bytes;
 157         head->arg.len = rqstp->rq_arg.len + ch_bytes;
 158         head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
 159         head->count++;
 160         chl_map->ch[0].start = 0;
 161         while (byte_count) {
 162                 rpl_map->sge[sge_no].iov_base =
 163                         page_address(rqstp->rq_arg.pages[page_no]) + page_off;
 164                 sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
 165                 rpl_map->sge[sge_no].iov_len = sge_bytes;
 166                 /*
 167                  * Don't bump head->count here because the same page
 168                  * may be used by multiple SGE.
 169                  */
 170                 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
 171                 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
 172
 173                 byte_count -= sge_bytes;
 174                 ch_bytes -= sge_bytes;
 175                 sge_no++;
 176                 /*
 177                  * If all bytes for this chunk have been mapped to an
 178                  * SGE, move to the next SGE
 179                  */
 180                 if (ch_bytes == 0) {
 181                         chl_map->ch[ch_no].count =
 182                                 sge_no - chl_map->ch[ch_no].start;
 183                         ch_no++;
 184                         ch++;
 185                         chl_map->ch[ch_no].start = sge_no;
 186                         ch_bytes = ch->rc_target.rs_length;
 187                         /* If bytes remaining account for next chunk */
 188                         if (byte_count) {
 189                                 head->arg.page_len += ch_bytes;
 190                                 head->arg.len += ch_bytes;
 191                                 head->arg.buflen += ch_bytes;
 192                         }
 193                 }
 194                 /*
 195                  * If this SGE consumed all of the page, move to the
 196                  * next page
 197                  */
 198                 if ((sge_bytes + page_off) == PAGE_SIZE) {
 199                         page_no++;
 200                         page_off = 0;
 201                         /*
 202                          * If there are still bytes left to map, bump
 203                          * the page count
 204                          */
 205                         if (byte_count)
 206                                 head->count++;
 207                 } else
 208                         page_off += sge_bytes;
 209         }
 210         BUG_ON(byte_count != 0);
 211         return sge_no;
 212 }
 213
 214 /* Map a read-chunk-list to an XDR and fast register the page-list.
 215  *
 216  * Assumptions:
 217  * - chunk[0]   position points to pages[0] at an offset of 0
 218  * - pages[]    will be made physically contiguous by creating a one-off memory
 219  *              region using the fastreg verb.
 220  * - byte_count is # of bytes in read-chunk-list
 221  * - ch_count   is # of chunks in read-chunk-list
 222  *
 223  * Output:
 224  * - sge array pointing into pages[] array.
 225  * - chunk_sge array specifying sge index and count for each
 226  *   chunk in the read list
 227  */
 228 static int fast_reg_read_chunks(struct svcxprt_rdma *xprt,
 229                                 struct svc_rqst *rqstp,
 230                                 struct svc_rdma_op_ctxt *head,
 231                                 struct rpcrdma_msg *rmsgp,
 232                                 struct svc_rdma_req_map *rpl_map,
 233                                 struct svc_rdma_req_map *chl_map,
 234                                 int ch_count,
 235                                 int byte_count)
 236 {
 237         int page_no;
 238         int ch_no;
 239         u32 offset;
 240         struct rpcrdma_read_chunk *ch;
 241         struct svc_rdma_fastreg_mr *frmr;
 242         int ret = 0;
 243
 244         frmr = svc_rdma_get_frmr(xprt);
 245         if (IS_ERR(frmr))
 246                 return -ENOMEM;
 247
 248         head->frmr = frmr;
 249         head->arg.head[0] = rqstp->rq_arg.head[0];
 250         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 251         head->arg.pages = &head->pages[head->count];
 252         head->hdr_count = head->count; /* save count of hdr pages */
 253         head->arg.page_base = 0;
 254         head->arg.page_len = byte_count;
 255         head->arg.len = rqstp->rq_arg.len + byte_count;
 256         head->arg.buflen = rqstp->rq_arg.buflen + byte_count;
 257
 258         /* Fast register the page list */
 259         frmr->kva = page_address(rqstp->rq_arg.pages[0]);
 260         frmr->direction = DMA_FROM_DEVICE;
 261         frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
 262         frmr->map_len = byte_count;
 263         frmr->page_list_len = PAGE_ALIGN(byte_count) >> PAGE_SHIFT;
 264         for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
 265                 frmr->page_list->page_list[page_no] =
 266                         ib_dma_map_single(xprt->sc_cm_id->device,
 267                                           page_address(rqstp->rq_arg.pages[page_no]),
 268                                           PAGE_SIZE, DMA_FROM_DEVICE);
 269                 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 270                                          frmr->page_list->page_list[page_no]))
 271                         goto fatal_err;
 272                 atomic_inc(&xprt->sc_dma_used);
 273                 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
 274         }
 275         head->count += page_no;
 276
 277         /* rq_respages points one past arg pages */
 278         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
 279
 280         /* Create the reply and chunk maps */
 281         offset = 0;
 282         ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 283         for (ch_no = 0; ch_no < ch_count; ch_no++) {
 284                 rpl_map->sge[ch_no].iov_base = frmr->kva + offset;
 285                 rpl_map->sge[ch_no].iov_len = ch->rc_target.rs_length;
 286                 chl_map->ch[ch_no].count = 1;
 287                 chl_map->ch[ch_no].start = ch_no;
 288                 offset += ch->rc_target.rs_length;
 289                 ch++;
 290         }
 291
 292         ret = svc_rdma_fastreg(xprt, frmr);
 293         if (ret)
 294                 goto fatal_err;
 295
 296         return ch_no;
 297
 298  fatal_err:
 299         printk("svcrdma: error fast registering xdr for xprt %p", xprt);
 300         svc_rdma_put_frmr(xprt, frmr);
 301         return -EIO;
 302 }
 303
 304 static int rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
 305                              struct svc_rdma_op_ctxt *ctxt,
 306                              struct svc_rdma_fastreg_mr *frmr,
 307                              struct kvec *vec,
 308                              u64 *sgl_offset,
 309                              int count)
 310 {
 311         int i;
 312
 313         ctxt->count = count;
 314         ctxt->direction = DMA_FROM_DEVICE;
 315         for (i = 0; i < count; i++) {
 316                 ctxt->sge[i].length = 0; /* in case map fails */
 317                 if (!frmr) {
 318                         ctxt->sge[i].addr =
 319                                 ib_dma_map_single(xprt->sc_cm_id->device,
 320                                                   vec[i].iov_base,
 321                                                   vec[i].iov_len,
 322                                                   DMA_FROM_DEVICE);
 323                         if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 324                                                  ctxt->sge[i].addr))
 325                                 return -EINVAL;
 326                         ctxt->sge[i].lkey = xprt->sc_dma_lkey;
 327                         atomic_inc(&xprt->sc_dma_used);
 328                 } else {
 329                         ctxt->sge[i].addr = (unsigned long)vec[i].iov_base;
 330                         ctxt->sge[i].lkey = frmr->mr->lkey;
 331                 }
 332                 ctxt->sge[i].length = vec[i].iov_len;
 333                 *sgl_offset = *sgl_offset + vec[i].iov_len;
 334         }
 335         return 0;
 336 }
 337
 338 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
 339 {
 340         if ((rdma_node_get_transport(xprt->sc_cm_id->device->node_type) ==
 341              RDMA_TRANSPORT_IWARP) &&
 342             sge_count > 1)
 343                 return 1;
 344         else
 345                 return min_t(int, sge_count, xprt->sc_max_sge);
 346 }
 347
 348 /*
 349  * Use RDMA_READ to read data from the advertised client buffer into the
 350  * XDR stream starting at rq_arg.head[0].iov_base.
 351  * Each chunk in the array
 352  * contains the following fields:
 353  * discrim      - '1', This isn't used for data placement
 354  * position     - The xdr stream offset (the same for every chunk)
 355  * handle       - RMR for client memory region
 356  * length       - data transfer length
 357  * offset       - 64 bit tagged offset in remote memory region
 358  *
 359  * On our side, we need to read into a pagelist. The first page immediately
 360  * follows the RPC header.
 361  *
 362  * This function returns:
 363  * 0 - No error and no read-list found.
 364  *
 365  * 1 - Successful read-list processing. The data is not yet in
 366  * the pagelist and therefore the RPC request must be deferred. The
 367  * I/O completion will enqueue the transport again and
 368  * svc_rdma_recvfrom will complete the request.
 369  *
 370  * <0 - Error processing/posting read-list.
 371  *
 372  * NOTE: The ctxt must not be touched after the last WR has been posted
 373  * because the I/O completion processing may occur on another
 374  * processor and free / modify the context. Ne touche pas!
 375  */
 376 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 377                          struct rpcrdma_msg *rmsgp,
 378                          struct svc_rqst *rqstp,
 379                          struct svc_rdma_op_ctxt *hdr_ctxt)
 380 {
 381         struct ib_send_wr read_wr;
 382         struct ib_send_wr inv_wr;
 383         int err = 0;
 384         int ch_no;
 385         int ch_count;
 386         int byte_count;
 387         int sge_count;
 388         u64 sgl_offset;
 389         struct rpcrdma_read_chunk *ch;
 390         struct svc_rdma_op_ctxt *ctxt = NULL;
 391         struct svc_rdma_req_map *rpl_map;
 392         struct svc_rdma_req_map *chl_map;
 393
 394         /* If no read list is present, return 0 */
 395         ch = svc_rdma_get_read_chunk(rmsgp);
 396         if (!ch)
 397                 return 0;
 398
 399         svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
 400         if (ch_count > RPCSVC_MAXPAGES)
 401                 return -EINVAL;
 402
 403         /* Allocate temporary reply and chunk maps */
 404         rpl_map = svc_rdma_get_req_map();
 405         chl_map = svc_rdma_get_req_map();
 406
 407         if (!xprt->sc_frmr_pg_list_len)
 408                 sge_count = map_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
 409                                             rpl_map, chl_map, ch_count,
 410                                             byte_count);
 411         else
 412                 sge_count = fast_reg_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
 413                                                  rpl_map, chl_map, ch_count,
 414                                                  byte_count);
 415         if (sge_count < 0) {
 416                 err = -EIO;
 417                 goto out;
 418         }
 419
 420         sgl_offset = 0;
 421         ch_no = 0;
 422
 423         for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 424              ch->rc_discrim != 0; ch++, ch_no++) {
 425 next_sge:
 426                 ctxt = svc_rdma_get_context(xprt);
 427                 ctxt->direction = DMA_FROM_DEVICE;
 428                 ctxt->frmr = hdr_ctxt->frmr;
 429                 ctxt->read_hdr = NULL;
 430                 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 431                 clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 432
 433                 /* Prepare READ WR */
 434                 memset(&read_wr, 0, sizeof read_wr);
 435                 read_wr.wr_id = (unsigned long)ctxt;
 436                 read_wr.opcode = IB_WR_RDMA_READ;
 437                 ctxt->wr_op = read_wr.opcode;
 438                 read_wr.send_flags = IB_SEND_SIGNALED;
 439                 read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
 440                 read_wr.wr.rdma.remote_addr =
 441                         get_unaligned(&(ch->rc_target.rs_offset)) +
 442                         sgl_offset;
 443                 read_wr.sg_list = ctxt->sge;
 444                 read_wr.num_sge =
 445                         rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
 446                 err = rdma_set_ctxt_sge(xprt, ctxt, hdr_ctxt->frmr,
 447                                         &rpl_map->sge[chl_map->ch[ch_no].start],
 448                                         &sgl_offset,
 449                                         read_wr.num_sge);
 450                 if (err) {
 451                         svc_rdma_unmap_dma(ctxt);
 452                         svc_rdma_put_context(ctxt, 0);
 453                         goto out;
 454                 }
 455                 if (((ch+1)->rc_discrim == 0) &&
 456                     (read_wr.num_sge == chl_map->ch[ch_no].count)) {
 457                         /*
 458                          * Mark the last RDMA_READ with a bit to
 459                          * indicate all RPC data has been fetched from
 460                          * the client and the RPC needs to be enqueued.
 461                          */
 462                         set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 463                         if (hdr_ctxt->frmr) {
 464                                 set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 465                                 /*
 466                                  * Invalidate the local MR used to map the data
 467                                  * sink.
 468                                  */
 469                                 if (xprt->sc_dev_caps &
 470                                     SVCRDMA_DEVCAP_READ_W_INV) {
 471                                         read_wr.opcode =
 472                                                 IB_WR_RDMA_READ_WITH_INV;
 473                                         ctxt->wr_op = read_wr.opcode;
 474                                         read_wr.ex.invalidate_rkey =
 475                                                 ctxt->frmr->mr->lkey;
 476                                 } else {
 477                                         /* Prepare INVALIDATE WR */
 478                                         memset(&inv_wr, 0, sizeof inv_wr);
 479                                         inv_wr.opcode = IB_WR_LOCAL_INV;
 480                                         inv_wr.send_flags = IB_SEND_SIGNALED;
 481                                         inv_wr.ex.invalidate_rkey =
 482                                                 hdr_ctxt->frmr->mr->lkey;
 483                                         read_wr.next = &inv_wr;
 484                                 }
 485                         }
 486                         ctxt->read_hdr = hdr_ctxt;
 487                 }
 488                 /* Post the read */
 489                 err = svc_rdma_send(xprt, &read_wr);
 490                 if (err) {
 491                         printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
 492                                err);
 493                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 494                         svc_rdma_put_context(ctxt, 0);
 495                         goto out;
 496                 }
 497                 atomic_inc(&rdma_stat_read);
 498
 499                 if (read_wr.num_sge < chl_map->ch[ch_no].count) {
 500                         chl_map->ch[ch_no].count -= read_wr.num_sge;
 501                         chl_map->ch[ch_no].start += read_wr.num_sge;
 502                         goto next_sge;
 503                 }
 504                 sgl_offset = 0;
 505                 err = 1;
 506         }
 507
 508  out:
 509         svc_rdma_put_req_map(rpl_map);
 510         svc_rdma_put_req_map(chl_map);
 511
 512         /* Detach arg pages. svc_recv will replenish them */
 513         for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
 514                 rqstp->rq_pages[ch_no] = NULL;
 515
 516         /*
 517          * Detach res pages. svc_release must see a resused count of
 518          * zero or it will attempt to put them.
 519          */
 520         while (rqstp->rq_resused)
 521                 rqstp->rq_respages[--rqstp->rq_resused] = NULL;
 522
 523         return err;
 524 }
 525
 526 static int rdma_read_complete(struct svc_rqst *rqstp,
 527                               struct svc_rdma_op_ctxt *head)
 528 {
 529         int page_no;
 530         int ret;
 531
 532         BUG_ON(!head);
 533
 534         /* Copy RPC pages */
 535         for (page_no = 0; page_no < head->count; page_no++) {
 536                 put_page(rqstp->rq_pages[page_no]);
 537                 rqstp->rq_pages[page_no] = head->pages[page_no];
 538         }
 539         /* Point rq_arg.pages past header */
 540         rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
 541         rqstp->rq_arg.page_len = head->arg.page_len;
 542         rqstp->rq_arg.page_base = head->arg.page_base;
 543
 544         /* rq_respages starts after the last arg page */
 545         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
 546         rqstp->rq_resused = 0;
 547
 548         /* Rebuild rq_arg head and tail. */
 549         rqstp->rq_arg.head[0] = head->arg.head[0];
 550         rqstp->rq_arg.tail[0] = head->arg.tail[0];
 551         rqstp->rq_arg.len = head->arg.len;
 552         rqstp->rq_arg.buflen = head->arg.buflen;
 553
 554         /* Free the context */
 555         svc_rdma_put_context(head, 0);
 556
 557         /* XXX: What should this be? */
 558         rqstp->rq_prot = IPPROTO_MAX;
 559         svc_xprt_copy_addrs(rqstp, rqstp->rq_xprt);
 560
 561         ret = rqstp->rq_arg.head[0].iov_len
 562                 + rqstp->rq_arg.page_len
 563                 + rqstp->rq_arg.tail[0].iov_len;
 564         dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
 565                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 566                 ret, rqstp->rq_arg.len, rqstp->rq_arg.head[0].iov_base,
 567                 rqstp->rq_arg.head[0].iov_len);
 568
 569         return ret;
 570 }
 571
 572 /*
 573  * Set up the rqstp thread context to point to the RQ buffer. If
 574  * necessary, pull additional data from the client with an RDMA_READ
 575  * request.
 576  */
 577 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 578 {
 579         struct svc_xprt *xprt = rqstp->rq_xprt;
 580         struct svcxprt_rdma *rdma_xprt =
 581                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 582         struct svc_rdma_op_ctxt *ctxt = NULL;
 583         struct rpcrdma_msg *rmsgp;
 584         int ret = 0;
 585         int len;
 586
 587         dprintk("svcrdma: rqstp=%p\n", rqstp);
 588
 589         spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
 590         if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
 591                 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
 592                                   struct svc_rdma_op_ctxt,
 593                                   dto_q);
 594                 list_del_init(&ctxt->dto_q);
 595         }
 596         if (ctxt) {
 597                 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
 598                 return rdma_read_complete(rqstp, ctxt);
 599         }
 600
 601         if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
 602                 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
 603                                   struct svc_rdma_op_ctxt,
 604                                   dto_q);
 605                 list_del_init(&ctxt->dto_q);
 606         } else {
 607                 atomic_inc(&rdma_stat_rq_starve);
 608                 clear_bit(XPT_DATA, &xprt->xpt_flags);
 609                 ctxt = NULL;
 610         }
 611         spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
 612         if (!ctxt) {
 613                 /* This is the EAGAIN path. The svc_recv routine will
 614                  * return -EAGAIN, the nfsd thread will go to call into
 615                  * svc_recv again and we shouldn't be on the active
 616                  * transport list
 617                  */
 618                 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
 619                         goto close_out;
 620
 621                 BUG_ON(ret);
 622                 goto out;
 623         }
 624         dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
 625                 ctxt, rdma_xprt, rqstp, ctxt->wc_status);
 626         BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
 627         atomic_inc(&rdma_stat_recv);
 628
 629         /* Build up the XDR from the receive buffers. */
 630         rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
 631
 632         /* Decode the RDMA header. */
 633         len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
 634         rqstp->rq_xprt_hlen = len;
 635
 636         /* If the request is invalid, reply with an error */
 637         if (len < 0) {
 638                 if (len == -ENOSYS)
 639                         svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
 640                 goto close_out;
 641         }
 642
 643         /* Read read-list data. */
 644         ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
 645         if (ret > 0) {
 646                 /* read-list posted, defer until data received from client. */
 647                 goto defer;
 648         }
 649         if (ret < 0) {
 650                 /* Post of read-list failed, free context. */
 651                 svc_rdma_put_context(ctxt, 1);
 652                 return 0;
 653         }
 654
 655         ret = rqstp->rq_arg.head[0].iov_len
 656                 + rqstp->rq_arg.page_len
 657                 + rqstp->rq_arg.tail[0].iov_len;
 658         svc_rdma_put_context(ctxt, 0);
 659  out:
 660         dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
 661                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 662                 ret, rqstp->rq_arg.len,
 663                 rqstp->rq_arg.head[0].iov_base,
 664                 rqstp->rq_arg.head[0].iov_len);
 665         rqstp->rq_prot = IPPROTO_MAX;
 666         svc_xprt_copy_addrs(rqstp, xprt);
 667         return ret;
 668
 669  close_out:
 670         if (ctxt)
 671                 svc_rdma_put_context(ctxt, 1);
 672         dprintk("svcrdma: transport %p is closing\n", xprt);
 673         /*
 674          * Set the close bit and enqueue it. svc_recv will see the
 675          * close bit and call svc_xprt_delete
 676          */
 677         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 678 defer:
 679         return 0;
 680 }