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         rqstp->rq_next_page = rqstp->rq_respages + 1;
  94
  95         /* We should never run out of SGE because the limit is defined to
  96          * support the max allowed RPC data length
  97          */
  98         BUG_ON(bc && (sge_no == ctxt->count));
  99         BUG_ON((rqstp->rq_arg.head[0].iov_len + rqstp->rq_arg.page_len)
 100                != byte_count);
 101         BUG_ON(rqstp->rq_arg.len != byte_count);
 102
 103         /* If not all pages were used from the SGL, free the remaining ones */
 104         bc = sge_no;
 105         while (sge_no < ctxt->count) {
 106                 page = ctxt->pages[sge_no++];
 107                 put_page(page);
 108         }
 109         ctxt->count = bc;
 110
 111         /* Set up tail */
 112         rqstp->rq_arg.tail[0].iov_base = NULL;
 113         rqstp->rq_arg.tail[0].iov_len = 0;
 114 }
 115
 116 /* Encode a read-chunk-list as an array of IB SGE
 117  *
 118  * Assumptions:
 119  * - chunk[0]->position points to pages[0] at an offset of 0
 120  * - pages[] is not physically or virtually contiguous and consists of
 121  *   PAGE_SIZE elements.
 122  *
 123  * Output:
 124  * - sge array pointing into pages[] array.
 125  * - chunk_sge array specifying sge index and count for each
 126  *   chunk in the read list
 127  *
 128  */
 129 static int map_read_chunks(struct svcxprt_rdma *xprt,
 130                            struct svc_rqst *rqstp,
 131                            struct svc_rdma_op_ctxt *head,
 132                            struct rpcrdma_msg *rmsgp,
 133                            struct svc_rdma_req_map *rpl_map,
 134                            struct svc_rdma_req_map *chl_map,
 135                            int ch_count,
 136                            int byte_count)
 137 {
 138         int sge_no;
 139         int sge_bytes;
 140         int page_off;
 141         int page_no;
 142         int ch_bytes;
 143         int ch_no;
 144         struct rpcrdma_read_chunk *ch;
 145
 146         sge_no = 0;
 147         page_no = 0;
 148         page_off = 0;
 149         ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 150         ch_no = 0;
 151         ch_bytes = ntohl(ch->rc_target.rs_length);
 152         head->arg.head[0] = rqstp->rq_arg.head[0];
 153         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 154         head->arg.pages = &head->pages[head->count];
 155         head->hdr_count = head->count; /* save count of hdr pages */
 156         head->arg.page_base = 0;
 157         head->arg.page_len = ch_bytes;
 158         head->arg.len = rqstp->rq_arg.len + ch_bytes;
 159         head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
 160         head->count++;
 161         chl_map->ch[0].start = 0;
 162         while (byte_count) {
 163                 rpl_map->sge[sge_no].iov_base =
 164                         page_address(rqstp->rq_arg.pages[page_no]) + page_off;
 165                 sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
 166                 rpl_map->sge[sge_no].iov_len = sge_bytes;
 167                 /*
 168                  * Don't bump head->count here because the same page
 169                  * may be used by multiple SGE.
 170                  */
 171                 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
 172                 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
 173                 rqstp->rq_next_page = rqstp->rq_respages + 1;
 174
 175                 byte_count -= sge_bytes;
 176                 ch_bytes -= sge_bytes;
 177                 sge_no++;
 178                 /*
 179                  * If all bytes for this chunk have been mapped to an
 180                  * SGE, move to the next SGE
 181                  */
 182                 if (ch_bytes == 0) {
 183                         chl_map->ch[ch_no].count =
 184                                 sge_no - chl_map->ch[ch_no].start;
 185                         ch_no++;
 186                         ch++;
 187                         chl_map->ch[ch_no].start = sge_no;
 188                         ch_bytes = ntohl(ch->rc_target.rs_length);
 189                         /* If bytes remaining account for next chunk */
 190                         if (byte_count) {
 191                                 head->arg.page_len += ch_bytes;
 192                                 head->arg.len += ch_bytes;
 193                                 head->arg.buflen += ch_bytes;
 194                         }
 195                 }
 196                 /*
 197                  * If this SGE consumed all of the page, move to the
 198                  * next page
 199                  */
 200                 if ((sge_bytes + page_off) == PAGE_SIZE) {
 201                         page_no++;
 202                         page_off = 0;
 203                         /*
 204                          * If there are still bytes left to map, bump
 205                          * the page count
 206                          */
 207                         if (byte_count)
 208                                 head->count++;
 209                 } else
 210                         page_off += sge_bytes;
 211         }
 212         BUG_ON(byte_count != 0);
 213         return sge_no;
 214 }
 215
 216 /* Map a read-chunk-list to an XDR and fast register the page-list.
 217  *
 218  * Assumptions:
 219  * - chunk[0]   position points to pages[0] at an offset of 0
 220  * - pages[]    will be made physically contiguous by creating a one-off memory
 221  *              region using the fastreg verb.
 222  * - byte_count is # of bytes in read-chunk-list
 223  * - ch_count   is # of chunks in read-chunk-list
 224  *
 225  * Output:
 226  * - sge array pointing into pages[] array.
 227  * - chunk_sge array specifying sge index and count for each
 228  *   chunk in the read list
 229  */
 230 static int fast_reg_read_chunks(struct svcxprt_rdma *xprt,
 231                                 struct svc_rqst *rqstp,
 232                                 struct svc_rdma_op_ctxt *head,
 233                                 struct rpcrdma_msg *rmsgp,
 234                                 struct svc_rdma_req_map *rpl_map,
 235                                 struct svc_rdma_req_map *chl_map,
 236                                 int ch_count,
 237                                 int byte_count)
 238 {
 239         int page_no;
 240         int ch_no;
 241         u32 offset;
 242         struct rpcrdma_read_chunk *ch;
 243         struct svc_rdma_fastreg_mr *frmr;
 244         int ret = 0;
 245
 246         frmr = svc_rdma_get_frmr(xprt);
 247         if (IS_ERR(frmr))
 248                 return -ENOMEM;
 249
 250         head->frmr = frmr;
 251         head->arg.head[0] = rqstp->rq_arg.head[0];
 252         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 253         head->arg.pages = &head->pages[head->count];
 254         head->hdr_count = head->count; /* save count of hdr pages */
 255         head->arg.page_base = 0;
 256         head->arg.page_len = byte_count;
 257         head->arg.len = rqstp->rq_arg.len + byte_count;
 258         head->arg.buflen = rqstp->rq_arg.buflen + byte_count;
 259
 260         /* Fast register the page list */
 261         frmr->kva = page_address(rqstp->rq_arg.pages[0]);
 262         frmr->direction = DMA_FROM_DEVICE;
 263         frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
 264         frmr->map_len = byte_count;
 265         frmr->page_list_len = PAGE_ALIGN(byte_count) >> PAGE_SHIFT;
 266         for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
 267                 frmr->page_list->page_list[page_no] =
 268                         ib_dma_map_page(xprt->sc_cm_id->device,
 269                                         rqstp->rq_arg.pages[page_no], 0,
 270                                         PAGE_SIZE, DMA_FROM_DEVICE);
 271                 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 272                                          frmr->page_list->page_list[page_no]))
 273                         goto fatal_err;
 274                 atomic_inc(&xprt->sc_dma_used);
 275                 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
 276         }
 277         head->count += page_no;
 278
 279         /* rq_respages points one past arg pages */
 280         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
 281         rqstp->rq_next_page = rqstp->rq_respages + 1;
 282
 283         /* Create the reply and chunk maps */
 284         offset = 0;
 285         ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 286         for (ch_no = 0; ch_no < ch_count; ch_no++) {
 287                 int len = ntohl(ch->rc_target.rs_length);
 288                 rpl_map->sge[ch_no].iov_base = frmr->kva + offset;
 289                 rpl_map->sge[ch_no].iov_len = len;
 290                 chl_map->ch[ch_no].count = 1;
 291                 chl_map->ch[ch_no].start = ch_no;
 292                 offset += len;
 293                 ch++;
 294         }
 295
 296         ret = svc_rdma_fastreg(xprt, frmr);
 297         if (ret)
 298                 goto fatal_err;
 299
 300         return ch_no;
 301
 302  fatal_err:
 303         printk("svcrdma: error fast registering xdr for xprt %p", xprt);
 304         svc_rdma_put_frmr(xprt, frmr);
 305         return -EIO;
 306 }
 307
 308 static int rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
 309                              struct svc_rdma_op_ctxt *ctxt,
 310                              struct svc_rdma_fastreg_mr *frmr,
 311                              struct kvec *vec,
 312                              u64 *sgl_offset,
 313                              int count)
 314 {
 315         int i;
 316         unsigned long off;
 317
 318         ctxt->count = count;
 319         ctxt->direction = DMA_FROM_DEVICE;
 320         for (i = 0; i < count; i++) {
 321                 ctxt->sge[i].length = 0; /* in case map fails */
 322                 if (!frmr) {
 323                         BUG_ON(!virt_to_page(vec[i].iov_base));
 324                         off = (unsigned long)vec[i].iov_base & ~PAGE_MASK;
 325                         ctxt->sge[i].addr =
 326                                 ib_dma_map_page(xprt->sc_cm_id->device,
 327                                                 virt_to_page(vec[i].iov_base),
 328                                                 off,
 329                                                 vec[i].iov_len,
 330                                                 DMA_FROM_DEVICE);
 331                         if (ib_dma_mapping_error(xprt->sc_cm_id->device,
 332                                                  ctxt->sge[i].addr))
 333                                 return -EINVAL;
 334                         ctxt->sge[i].lkey = xprt->sc_dma_lkey;
 335                         atomic_inc(&xprt->sc_dma_used);
 336                 } else {
 337                         ctxt->sge[i].addr = (unsigned long)vec[i].iov_base;
 338                         ctxt->sge[i].lkey = frmr->mr->lkey;
 339                 }
 340                 ctxt->sge[i].length = vec[i].iov_len;
 341                 *sgl_offset = *sgl_offset + vec[i].iov_len;
 342         }
 343         return 0;
 344 }
 345
 346 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
 347 {
 348         if ((rdma_node_get_transport(xprt->sc_cm_id->device->node_type) ==
 349              RDMA_TRANSPORT_IWARP) &&
 350             sge_count > 1)
 351                 return 1;
 352         else
 353                 return min_t(int, sge_count, xprt->sc_max_sge);
 354 }
 355
 356 /*
 357  * Use RDMA_READ to read data from the advertised client buffer into the
 358  * XDR stream starting at rq_arg.head[0].iov_base.
 359  * Each chunk in the array
 360  * contains the following fields:
 361  * discrim      - '1', This isn't used for data placement
 362  * position     - The xdr stream offset (the same for every chunk)
 363  * handle       - RMR for client memory region
 364  * length       - data transfer length
 365  * offset       - 64 bit tagged offset in remote memory region
 366  *
 367  * On our side, we need to read into a pagelist. The first page immediately
 368  * follows the RPC header.
 369  *
 370  * This function returns:
 371  * 0 - No error and no read-list found.
 372  *
 373  * 1 - Successful read-list processing. The data is not yet in
 374  * the pagelist and therefore the RPC request must be deferred. The
 375  * I/O completion will enqueue the transport again and
 376  * svc_rdma_recvfrom will complete the request.
 377  *
 378  * <0 - Error processing/posting read-list.
 379  *
 380  * NOTE: The ctxt must not be touched after the last WR has been posted
 381  * because the I/O completion processing may occur on another
 382  * processor and free / modify the context. Ne touche pas!
 383  */
 384 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 385                          struct rpcrdma_msg *rmsgp,
 386                          struct svc_rqst *rqstp,
 387                          struct svc_rdma_op_ctxt *hdr_ctxt)
 388 {
 389         struct ib_send_wr read_wr;
 390         struct ib_send_wr inv_wr;
 391         int err = 0;
 392         int ch_no;
 393         int ch_count;
 394         int byte_count;
 395         int sge_count;
 396         u64 sgl_offset;
 397         struct rpcrdma_read_chunk *ch;
 398         struct svc_rdma_op_ctxt *ctxt = NULL;
 399         struct svc_rdma_req_map *rpl_map;
 400         struct svc_rdma_req_map *chl_map;
 401
 402         /* If no read list is present, return 0 */
 403         ch = svc_rdma_get_read_chunk(rmsgp);
 404         if (!ch)
 405                 return 0;
 406
 407         svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
 408         if (ch_count > RPCSVC_MAXPAGES)
 409                 return -EINVAL;
 410
 411         /* Allocate temporary reply and chunk maps */
 412         rpl_map = svc_rdma_get_req_map();
 413         chl_map = svc_rdma_get_req_map();
 414
 415         if (!xprt->sc_frmr_pg_list_len)
 416                 sge_count = map_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
 417                                             rpl_map, chl_map, ch_count,
 418                                             byte_count);
 419         else
 420                 sge_count = fast_reg_read_chunks(xprt, rqstp, hdr_ctxt, rmsgp,
 421                                                  rpl_map, chl_map, ch_count,
 422                                                  byte_count);
 423         if (sge_count < 0) {
 424                 err = -EIO;
 425                 goto out;
 426         }
 427
 428         sgl_offset = 0;
 429         ch_no = 0;
 430
 431         for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 432              ch->rc_discrim != 0; ch++, ch_no++) {
 433                 u64 rs_offset;
 434 next_sge:
 435                 ctxt = svc_rdma_get_context(xprt);
 436                 ctxt->direction = DMA_FROM_DEVICE;
 437                 ctxt->frmr = hdr_ctxt->frmr;
 438                 ctxt->read_hdr = NULL;
 439                 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 440                 clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 441
 442                 /* Prepare READ WR */
 443                 memset(&read_wr, 0, sizeof read_wr);
 444                 read_wr.wr_id = (unsigned long)ctxt;
 445                 read_wr.opcode = IB_WR_RDMA_READ;
 446                 ctxt->wr_op = read_wr.opcode;
 447                 read_wr.send_flags = IB_SEND_SIGNALED;
 448                 read_wr.wr.rdma.rkey = ntohl(ch->rc_target.rs_handle);
 449                 xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset,
 450                                  &rs_offset);
 451                 read_wr.wr.rdma.remote_addr = rs_offset + sgl_offset;
 452                 read_wr.sg_list = ctxt->sge;
 453                 read_wr.num_sge =
 454                         rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
 455                 err = rdma_set_ctxt_sge(xprt, ctxt, hdr_ctxt->frmr,
 456                                         &rpl_map->sge[chl_map->ch[ch_no].start],
 457                                         &sgl_offset,
 458                                         read_wr.num_sge);
 459                 if (err) {
 460                         svc_rdma_unmap_dma(ctxt);
 461                         svc_rdma_put_context(ctxt, 0);
 462                         goto out;
 463                 }
 464                 if (((ch+1)->rc_discrim == 0) &&
 465                     (read_wr.num_sge == chl_map->ch[ch_no].count)) {
 466                         /*
 467                          * Mark the last RDMA_READ with a bit to
 468                          * indicate all RPC data has been fetched from
 469                          * the client and the RPC needs to be enqueued.
 470                          */
 471                         set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 472                         if (hdr_ctxt->frmr) {
 473                                 set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
 474                                 /*
 475                                  * Invalidate the local MR used to map the data
 476                                  * sink.
 477                                  */
 478                                 if (xprt->sc_dev_caps &
 479                                     SVCRDMA_DEVCAP_READ_W_INV) {
 480                                         read_wr.opcode =
 481                                                 IB_WR_RDMA_READ_WITH_INV;
 482                                         ctxt->wr_op = read_wr.opcode;
 483                                         read_wr.ex.invalidate_rkey =
 484                                                 ctxt->frmr->mr->lkey;
 485                                 } else {
 486                                         /* Prepare INVALIDATE WR */
 487                                         memset(&inv_wr, 0, sizeof inv_wr);
 488                                         inv_wr.opcode = IB_WR_LOCAL_INV;
 489                                         inv_wr.send_flags = IB_SEND_SIGNALED;
 490                                         inv_wr.ex.invalidate_rkey =
 491                                                 hdr_ctxt->frmr->mr->lkey;
 492                                         read_wr.next = &inv_wr;
 493                                 }
 494                         }
 495                         ctxt->read_hdr = hdr_ctxt;
 496                 }
 497                 /* Post the read */
 498                 err = svc_rdma_send(xprt, &read_wr);
 499                 if (err) {
 500                         printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
 501                                err);
 502                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 503                         svc_rdma_unmap_dma(ctxt);
 504                         svc_rdma_put_context(ctxt, 0);
 505                         goto out;
 506                 }
 507                 atomic_inc(&rdma_stat_read);
 508
 509                 if (read_wr.num_sge < chl_map->ch[ch_no].count) {
 510                         chl_map->ch[ch_no].count -= read_wr.num_sge;
 511                         chl_map->ch[ch_no].start += read_wr.num_sge;
 512                         goto next_sge;
 513                 }
 514                 sgl_offset = 0;
 515                 err = 1;
 516         }
 517
 518  out:
 519         svc_rdma_put_req_map(rpl_map);
 520         svc_rdma_put_req_map(chl_map);
 521
 522         /* Detach arg pages. svc_recv will replenish them */
 523         for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
 524                 rqstp->rq_pages[ch_no] = NULL;
 525
 526         return err;
 527 }
 528
 529 static int rdma_read_complete(struct svc_rqst *rqstp,
 530                               struct svc_rdma_op_ctxt *head)
 531 {
 532         int page_no;
 533         int ret;
 534
 535         BUG_ON(!head);
 536
 537         /* Copy RPC pages */
 538         for (page_no = 0; page_no < head->count; page_no++) {
 539                 put_page(rqstp->rq_pages[page_no]);
 540                 rqstp->rq_pages[page_no] = head->pages[page_no];
 541         }
 542         /* Point rq_arg.pages past header */
 543         rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
 544         rqstp->rq_arg.page_len = head->arg.page_len;
 545         rqstp->rq_arg.page_base = head->arg.page_base;
 546
 547         /* rq_respages starts after the last arg page */
 548         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
 549         rqstp->rq_next_page = rqstp->rq_respages + 1;
 550
 551         /* Rebuild rq_arg head and tail. */
 552         rqstp->rq_arg.head[0] = head->arg.head[0];
 553         rqstp->rq_arg.tail[0] = head->arg.tail[0];
 554         rqstp->rq_arg.len = head->arg.len;
 555         rqstp->rq_arg.buflen = head->arg.buflen;
 556
 557         /* Free the context */
 558         svc_rdma_put_context(head, 0);
 559
 560         /* XXX: What should this be? */
 561         rqstp->rq_prot = IPPROTO_MAX;
 562         svc_xprt_copy_addrs(rqstp, rqstp->rq_xprt);
 563
 564         ret = rqstp->rq_arg.head[0].iov_len
 565                 + rqstp->rq_arg.page_len
 566                 + rqstp->rq_arg.tail[0].iov_len;
 567         dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
 568                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 569                 ret, rqstp->rq_arg.len, rqstp->rq_arg.head[0].iov_base,
 570                 rqstp->rq_arg.head[0].iov_len);
 571
 572         return ret;
 573 }
 574
 575 /*
 576  * Set up the rqstp thread context to point to the RQ buffer. If
 577  * necessary, pull additional data from the client with an RDMA_READ
 578  * request.
 579  */
 580 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 581 {
 582         struct svc_xprt *xprt = rqstp->rq_xprt;
 583         struct svcxprt_rdma *rdma_xprt =
 584                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 585         struct svc_rdma_op_ctxt *ctxt = NULL;
 586         struct rpcrdma_msg *rmsgp;
 587         int ret = 0;
 588         int len;
 589
 590         dprintk("svcrdma: rqstp=%p\n", rqstp);
 591
 592         spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
 593         if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
 594                 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
 595                                   struct svc_rdma_op_ctxt,
 596                                   dto_q);
 597                 list_del_init(&ctxt->dto_q);
 598         }
 599         if (ctxt) {
 600                 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
 601                 return rdma_read_complete(rqstp, ctxt);
 602         }
 603
 604         if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
 605                 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
 606                                   struct svc_rdma_op_ctxt,
 607                                   dto_q);
 608                 list_del_init(&ctxt->dto_q);
 609         } else {
 610                 atomic_inc(&rdma_stat_rq_starve);
 611                 clear_bit(XPT_DATA, &xprt->xpt_flags);
 612                 ctxt = NULL;
 613         }
 614         spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
 615         if (!ctxt) {
 616                 /* This is the EAGAIN path. The svc_recv routine will
 617                  * return -EAGAIN, the nfsd thread will go to call into
 618                  * svc_recv again and we shouldn't be on the active
 619                  * transport list
 620                  */
 621                 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
 622                         goto close_out;
 623
 624                 BUG_ON(ret);
 625                 goto out;
 626         }
 627         dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
 628                 ctxt, rdma_xprt, rqstp, ctxt->wc_status);
 629         BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
 630         atomic_inc(&rdma_stat_recv);
 631
 632         /* Build up the XDR from the receive buffers. */
 633         rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
 634
 635         /* Decode the RDMA header. */
 636         len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
 637         rqstp->rq_xprt_hlen = len;
 638
 639         /* If the request is invalid, reply with an error */
 640         if (len < 0) {
 641                 if (len == -ENOSYS)
 642                         svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
 643                 goto close_out;
 644         }
 645
 646         /* Read read-list data. */
 647         ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
 648         if (ret > 0) {
 649                 /* read-list posted, defer until data received from client. */
 650                 goto defer;
 651         }
 652         if (ret < 0) {
 653                 /* Post of read-list failed, free context. */
 654                 svc_rdma_put_context(ctxt, 1);
 655                 return 0;
 656         }
 657
 658         ret = rqstp->rq_arg.head[0].iov_len
 659                 + rqstp->rq_arg.page_len
 660                 + rqstp->rq_arg.tail[0].iov_len;
 661         svc_rdma_put_context(ctxt, 0);
 662  out:
 663         dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
 664                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 665                 ret, rqstp->rq_arg.len,
 666                 rqstp->rq_arg.head[0].iov_base,
 667                 rqstp->rq_arg.head[0].iov_len);
 668         rqstp->rq_prot = IPPROTO_MAX;
 669         svc_xprt_copy_addrs(rqstp, xprt);
 670         return ret;
 671
 672  close_out:
 673         if (ctxt)
 674                 svc_rdma_put_context(ctxt, 1);
 675         dprintk("svcrdma: transport %p is closing\n", xprt);
 676         /*
 677          * Set the close bit and enqueue it. svc_recv will see the
 678          * close bit and call svc_xprt_delete
 679          */
 680         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 681 defer:
 682         return 0;
 683 }