net/sunrpc/xprtrdma/svc_rdma_recvfrom.c

   1 /*
   2  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
   3  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the BSD-type
   9  * license below:
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  *
  15  *      Redistributions of source code must retain the above copyright
  16  *      notice, this list of conditions and the following disclaimer.
  17  *
  18  *      Redistributions in binary form must reproduce the above
  19  *      copyright notice, this list of conditions and the following
  20  *      disclaimer in the documentation and/or other materials provided
  21  *      with the distribution.
  22  *
  23  *      Neither the name of the Network Appliance, Inc. nor the names of
  24  *      its contributors may be used to endorse or promote products
  25  *      derived from this software without specific prior written
  26  *      permission.
  27  *
  28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  39  *
  40  * Author: Tom Tucker <tom@opengridcomputing.com>
  41  */
  42
  43 #include <linux/sunrpc/debug.h>
  44 #include <linux/sunrpc/rpc_rdma.h>
  45 #include <linux/spinlock.h>
  46 #include <asm/unaligned.h>
  47 #include <rdma/ib_verbs.h>
  48 #include <rdma/rdma_cm.h>
  49 #include <linux/sunrpc/svc_rdma.h>
  50
  51 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
  52
  53 /*
  54  * Replace the pages in the rq_argpages array with the pages from the SGE in
  55  * the RDMA_RECV completion. The SGL should contain full pages up until the
  56  * last one.
  57  */
  58 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
  59                                struct svc_rdma_op_ctxt *ctxt,
  60                                u32 byte_count)
  61 {
  62         struct rpcrdma_msg *rmsgp;
  63         struct page *page;
  64         u32 bc;
  65         int sge_no;
  66
  67         /* Swap the page in the SGE with the page in argpages */
  68         page = ctxt->pages[0];
  69         put_page(rqstp->rq_pages[0]);
  70         rqstp->rq_pages[0] = page;
  71
  72         /* Set up the XDR head */
  73         rqstp->rq_arg.head[0].iov_base = page_address(page);
  74         rqstp->rq_arg.head[0].iov_len =
  75                 min_t(size_t, byte_count, ctxt->sge[0].length);
  76         rqstp->rq_arg.len = byte_count;
  77         rqstp->rq_arg.buflen = byte_count;
  78
  79         /* Compute bytes past head in the SGL */
  80         bc = byte_count - rqstp->rq_arg.head[0].iov_len;
  81
  82         /* If data remains, store it in the pagelist */
  83         rqstp->rq_arg.page_len = bc;
  84         rqstp->rq_arg.page_base = 0;
  85
  86         /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
  87         rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
  88         if (rmsgp->rm_type == rdma_nomsg)
  89                 rqstp->rq_arg.pages = &rqstp->rq_pages[0];
  90         else
  91                 rqstp->rq_arg.pages = &rqstp->rq_pages[1];
  92
  93         sge_no = 1;
  94         while (bc && sge_no < ctxt->count) {
  95                 page = ctxt->pages[sge_no];
  96                 put_page(rqstp->rq_pages[sge_no]);
  97                 rqstp->rq_pages[sge_no] = page;
  98                 bc -= min_t(u32, bc, ctxt->sge[sge_no].length);
  99                 rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
 100                 sge_no++;
 101         }
 102         rqstp->rq_respages = &rqstp->rq_pages[sge_no];
 103         rqstp->rq_next_page = rqstp->rq_respages + 1;
 104
 105         /* If not all pages were used from the SGL, free the remaining ones */
 106         bc = sge_no;
 107         while (sge_no < ctxt->count) {
 108                 page = ctxt->pages[sge_no++];
 109                 put_page(page);
 110         }
 111         ctxt->count = bc;
 112
 113         /* Set up tail */
 114         rqstp->rq_arg.tail[0].iov_base = NULL;
 115         rqstp->rq_arg.tail[0].iov_len = 0;
 116 }
 117
 118 /* Issue an RDMA_READ using the local lkey to map the data sink */
 119 int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
 120                         struct svc_rqst *rqstp,
 121                         struct svc_rdma_op_ctxt *head,
 122                         int *page_no,
 123                         u32 *page_offset,
 124                         u32 rs_handle,
 125                         u32 rs_length,
 126                         u64 rs_offset,
 127                         bool last)
 128 {
 129         struct ib_rdma_wr read_wr;
 130         int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
 131         struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
 132         int ret, read, pno;
 133         u32 pg_off = *page_offset;
 134         u32 pg_no = *page_no;
 135
 136         ctxt->direction = DMA_FROM_DEVICE;
 137         ctxt->read_hdr = head;
 138         pages_needed = min_t(int, pages_needed, xprt->sc_max_sge_rd);
 139         read = min_t(int, (pages_needed << PAGE_SHIFT) - *page_offset,
 140                      rs_length);
 141
 142         for (pno = 0; pno < pages_needed; pno++) {
 143                 int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
 144
 145                 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
 146                 head->arg.page_len += len;
 147
 148                 head->arg.len += len;
 149                 if (!pg_off)
 150                         head->count++;
 151                 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
 152                 rqstp->rq_next_page = rqstp->rq_respages + 1;
 153                 ctxt->sge[pno].addr =
 154                         ib_dma_map_page(xprt->sc_cm_id->device,
 155                                         head->arg.pages[pg_no], pg_off,
 156                                         PAGE_SIZE - pg_off,
 157                                         DMA_FROM_DEVICE);
 158                 ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
 159                                            ctxt->sge[pno].addr);
 160                 if (ret)
 161                         goto err;
 162                 svc_rdma_count_mappings(xprt, ctxt);
 163
 164                 ctxt->sge[pno].lkey = xprt->sc_pd->local_dma_lkey;
 165                 ctxt->sge[pno].length = len;
 166                 ctxt->count++;
 167
 168                 /* adjust offset and wrap to next page if needed */
 169                 pg_off += len;
 170                 if (pg_off == PAGE_SIZE) {
 171                         pg_off = 0;
 172                         pg_no++;
 173                 }
 174                 rs_length -= len;
 175         }
 176
 177         if (last && rs_length == 0)
 178                 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 179         else
 180                 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 181
 182         memset(&read_wr, 0, sizeof(read_wr));
 183         ctxt->cqe.done = svc_rdma_wc_read;
 184         read_wr.wr.wr_cqe = &ctxt->cqe;
 185         read_wr.wr.opcode = IB_WR_RDMA_READ;
 186         read_wr.wr.send_flags = IB_SEND_SIGNALED;
 187         read_wr.rkey = rs_handle;
 188         read_wr.remote_addr = rs_offset;
 189         read_wr.wr.sg_list = ctxt->sge;
 190         read_wr.wr.num_sge = pages_needed;
 191
 192         ret = svc_rdma_send(xprt, &read_wr.wr);
 193         if (ret) {
 194                 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
 195                 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 196                 goto err;
 197         }
 198
 199         /* return current location in page array */
 200         *page_no = pg_no;
 201         *page_offset = pg_off;
 202         ret = read;
 203         atomic_inc(&rdma_stat_read);
 204         return ret;
 205  err:
 206         svc_rdma_unmap_dma(ctxt);
 207         svc_rdma_put_context(ctxt, 0);
 208         return ret;
 209 }
 210
 211 /* Issue an RDMA_READ using an FRMR to map the data sink */
 212 int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
 213                          struct svc_rqst *rqstp,
 214                          struct svc_rdma_op_ctxt *head,
 215                          int *page_no,
 216                          u32 *page_offset,
 217                          u32 rs_handle,
 218                          u32 rs_length,
 219                          u64 rs_offset,
 220                          bool last)
 221 {
 222         struct ib_rdma_wr read_wr;
 223         struct ib_send_wr inv_wr;
 224         struct ib_reg_wr reg_wr;
 225         u8 key;
 226         int nents = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
 227         struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
 228         struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt);
 229         int ret, read, pno, dma_nents, n;
 230         u32 pg_off = *page_offset;
 231         u32 pg_no = *page_no;
 232
 233         if (IS_ERR(frmr))
 234                 return -ENOMEM;
 235
 236         ctxt->direction = DMA_FROM_DEVICE;
 237         ctxt->frmr = frmr;
 238         nents = min_t(unsigned int, nents, xprt->sc_frmr_pg_list_len);
 239         read = min_t(int, (nents << PAGE_SHIFT) - *page_offset, rs_length);
 240
 241         frmr->direction = DMA_FROM_DEVICE;
 242         frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
 243         frmr->sg_nents = nents;
 244
 245         for (pno = 0; pno < nents; pno++) {
 246                 int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
 247
 248                 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
 249                 head->arg.page_len += len;
 250                 head->arg.len += len;
 251                 if (!pg_off)
 252                         head->count++;
 253
 254                 sg_set_page(&frmr->sg[pno], rqstp->rq_arg.pages[pg_no],
 255                             len, pg_off);
 256
 257                 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
 258                 rqstp->rq_next_page = rqstp->rq_respages + 1;
 259
 260                 /* adjust offset and wrap to next page if needed */
 261                 pg_off += len;
 262                 if (pg_off == PAGE_SIZE) {
 263                         pg_off = 0;
 264                         pg_no++;
 265                 }
 266                 rs_length -= len;
 267         }
 268
 269         if (last && rs_length == 0)
 270                 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 271         else
 272                 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 273
 274         dma_nents = ib_dma_map_sg(xprt->sc_cm_id->device,
 275                                   frmr->sg, frmr->sg_nents,
 276                                   frmr->direction);
 277         if (!dma_nents) {
 278                 pr_err("svcrdma: failed to dma map sg %p\n",
 279                        frmr->sg);
 280                 return -ENOMEM;
 281         }
 282
 283         n = ib_map_mr_sg(frmr->mr, frmr->sg, frmr->sg_nents, NULL, PAGE_SIZE);
 284         if (unlikely(n != frmr->sg_nents)) {
 285                 pr_err("svcrdma: failed to map mr %p (%d/%d elements)\n",
 286                        frmr->mr, n, frmr->sg_nents);
 287                 return n < 0 ? n : -EINVAL;
 288         }
 289
 290         /* Bump the key */
 291         key = (u8)(frmr->mr->lkey & 0x000000FF);
 292         ib_update_fast_reg_key(frmr->mr, ++key);
 293
 294         ctxt->sge[0].addr = frmr->mr->iova;
 295         ctxt->sge[0].lkey = frmr->mr->lkey;
 296         ctxt->sge[0].length = frmr->mr->length;
 297         ctxt->count = 1;
 298         ctxt->read_hdr = head;
 299
 300         /* Prepare REG WR */
 301         ctxt->reg_cqe.done = svc_rdma_wc_reg;
 302         reg_wr.wr.wr_cqe = &ctxt->reg_cqe;
 303         reg_wr.wr.opcode = IB_WR_REG_MR;
 304         reg_wr.wr.send_flags = IB_SEND_SIGNALED;
 305         reg_wr.wr.num_sge = 0;
 306         reg_wr.mr = frmr->mr;
 307         reg_wr.key = frmr->mr->lkey;
 308         reg_wr.access = frmr->access_flags;
 309         reg_wr.wr.next = &read_wr.wr;
 310
 311         /* Prepare RDMA_READ */
 312         memset(&read_wr, 0, sizeof(read_wr));
 313         ctxt->cqe.done = svc_rdma_wc_read;
 314         read_wr.wr.wr_cqe = &ctxt->cqe;
 315         read_wr.wr.send_flags = IB_SEND_SIGNALED;
 316         read_wr.rkey = rs_handle;
 317         read_wr.remote_addr = rs_offset;
 318         read_wr.wr.sg_list = ctxt->sge;
 319         read_wr.wr.num_sge = 1;
 320         if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_READ_W_INV) {
 321                 read_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV;
 322                 read_wr.wr.ex.invalidate_rkey = ctxt->frmr->mr->lkey;
 323         } else {
 324                 read_wr.wr.opcode = IB_WR_RDMA_READ;
 325                 read_wr.wr.next = &inv_wr;
 326                 /* Prepare invalidate */
 327                 memset(&inv_wr, 0, sizeof(inv_wr));
 328                 ctxt->inv_cqe.done = svc_rdma_wc_inv;
 329                 inv_wr.wr_cqe = &ctxt->inv_cqe;
 330                 inv_wr.opcode = IB_WR_LOCAL_INV;
 331                 inv_wr.send_flags = IB_SEND_SIGNALED | IB_SEND_FENCE;
 332                 inv_wr.ex.invalidate_rkey = frmr->mr->lkey;
 333         }
 334
 335         /* Post the chain */
 336         ret = svc_rdma_send(xprt, &reg_wr.wr);
 337         if (ret) {
 338                 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
 339                 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 340                 goto err;
 341         }
 342
 343         /* return current location in page array */
 344         *page_no = pg_no;
 345         *page_offset = pg_off;
 346         ret = read;
 347         atomic_inc(&rdma_stat_read);
 348         return ret;
 349  err:
 350         svc_rdma_put_context(ctxt, 0);
 351         svc_rdma_put_frmr(xprt, frmr);
 352         return ret;
 353 }
 354
 355 static unsigned int
 356 rdma_rcl_chunk_count(struct rpcrdma_read_chunk *ch)
 357 {
 358         unsigned int count;
 359
 360         for (count = 0; ch->rc_discrim != xdr_zero; ch++)
 361                 count++;
 362         return count;
 363 }
 364
 365 /* If there was additional inline content, append it to the end of arg.pages.
 366  * Tail copy has to be done after the reader function has determined how many
 367  * pages are needed for RDMA READ.
 368  */
 369 static int
 370 rdma_copy_tail(struct svc_rqst *rqstp, struct svc_rdma_op_ctxt *head,
 371                u32 position, u32 byte_count, u32 page_offset, int page_no)
 372 {
 373         char *srcp, *destp;
 374
 375         srcp = head->arg.head[0].iov_base + position;
 376         byte_count = head->arg.head[0].iov_len - position;
 377         if (byte_count > PAGE_SIZE) {
 378                 dprintk("svcrdma: large tail unsupported\n");
 379                 return 0;
 380         }
 381
 382         /* Fit as much of the tail on the current page as possible */
 383         if (page_offset != PAGE_SIZE) {
 384                 destp = page_address(rqstp->rq_arg.pages[page_no]);
 385                 destp += page_offset;
 386                 while (byte_count--) {
 387                         *destp++ = *srcp++;
 388                         page_offset++;
 389                         if (page_offset == PAGE_SIZE && byte_count)
 390                                 goto more;
 391                 }
 392                 goto done;
 393         }
 394
 395 more:
 396         /* Fit the rest on the next page */
 397         page_no++;
 398         destp = page_address(rqstp->rq_arg.pages[page_no]);
 399         while (byte_count--)
 400                 *destp++ = *srcp++;
 401
 402         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
 403         rqstp->rq_next_page = rqstp->rq_respages + 1;
 404
 405 done:
 406         byte_count = head->arg.head[0].iov_len - position;
 407         head->arg.page_len += byte_count;
 408         head->arg.len += byte_count;
 409         head->arg.buflen += byte_count;
 410         return 1;
 411 }
 412
 413 /* Returns the address of the first read chunk or <nul> if no read chunk
 414  * is present
 415  */
 416 static struct rpcrdma_read_chunk *
 417 svc_rdma_get_read_chunk(struct rpcrdma_msg *rmsgp)
 418 {
 419         struct rpcrdma_read_chunk *ch =
 420                 (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 421
 422         if (ch->rc_discrim == xdr_zero)
 423                 return NULL;
 424         return ch;
 425 }
 426
 427 static int rdma_read_chunks(struct svcxprt_rdma *xprt,
 428                             struct rpcrdma_msg *rmsgp,
 429                             struct svc_rqst *rqstp,
 430                             struct svc_rdma_op_ctxt *head)
 431 {
 432         int page_no, ret;
 433         struct rpcrdma_read_chunk *ch;
 434         u32 handle, page_offset, byte_count;
 435         u32 position;
 436         u64 rs_offset;
 437         bool last;
 438
 439         /* If no read list is present, return 0 */
 440         ch = svc_rdma_get_read_chunk(rmsgp);
 441         if (!ch)
 442                 return 0;
 443
 444         if (rdma_rcl_chunk_count(ch) > RPCSVC_MAXPAGES)
 445                 return -EINVAL;
 446
 447         /* The request is completed when the RDMA_READs complete. The
 448          * head context keeps all the pages that comprise the
 449          * request.
 450          */
 451         head->arg.head[0] = rqstp->rq_arg.head[0];
 452         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 453         head->hdr_count = head->count;
 454         head->arg.page_base = 0;
 455         head->arg.page_len = 0;
 456         head->arg.len = rqstp->rq_arg.len;
 457         head->arg.buflen = rqstp->rq_arg.buflen;
 458
 459         /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
 460         position = be32_to_cpu(ch->rc_position);
 461         if (position == 0) {
 462                 head->arg.pages = &head->pages[0];
 463                 page_offset = head->byte_len;
 464         } else {
 465                 head->arg.pages = &head->pages[head->count];
 466                 page_offset = 0;
 467         }
 468
 469         ret = 0;
 470         page_no = 0;
 471         for (; ch->rc_discrim != xdr_zero; ch++) {
 472                 if (be32_to_cpu(ch->rc_position) != position)
 473                         goto err;
 474
 475                 handle = be32_to_cpu(ch->rc_target.rs_handle),
 476                 byte_count = be32_to_cpu(ch->rc_target.rs_length);
 477                 xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset,
 478                                  &rs_offset);
 479
 480                 while (byte_count > 0) {
 481                         last = (ch + 1)->rc_discrim == xdr_zero;
 482                         ret = xprt->sc_reader(xprt, rqstp, head,
 483                                               &page_no, &page_offset,
 484                                               handle, byte_count,
 485                                               rs_offset, last);
 486                         if (ret < 0)
 487                                 goto err;
 488                         byte_count -= ret;
 489                         rs_offset += ret;
 490                         head->arg.buflen += ret;
 491                 }
 492         }
 493
 494         /* Read list may need XDR round-up (see RFC 5666, s. 3.7) */
 495         if (page_offset & 3) {
 496                 u32 pad = 4 - (page_offset & 3);
 497
 498                 head->arg.tail[0].iov_len += pad;
 499                 head->arg.len += pad;
 500                 head->arg.buflen += pad;
 501                 page_offset += pad;
 502         }
 503
 504         ret = 1;
 505         if (position && position < head->arg.head[0].iov_len)
 506                 ret = rdma_copy_tail(rqstp, head, position,
 507                                      byte_count, page_offset, page_no);
 508         head->arg.head[0].iov_len = position;
 509         head->position = position;
 510
 511  err:
 512         /* Detach arg pages. svc_recv will replenish them */
 513         for (page_no = 0;
 514              &rqstp->rq_pages[page_no] < rqstp->rq_respages; page_no++)
 515                 rqstp->rq_pages[page_no] = NULL;
 516
 517         return ret;
 518 }
 519
 520 static void rdma_read_complete(struct svc_rqst *rqstp,
 521                                struct svc_rdma_op_ctxt *head)
 522 {
 523         int page_no;
 524
 525         /* Copy RPC pages */
 526         for (page_no = 0; page_no < head->count; page_no++) {
 527                 put_page(rqstp->rq_pages[page_no]);
 528                 rqstp->rq_pages[page_no] = head->pages[page_no];
 529         }
 530
 531         /* Adjustments made for RDMA_NOMSG type requests */
 532         if (head->position == 0) {
 533                 if (head->arg.len <= head->sge[0].length) {
 534                         head->arg.head[0].iov_len = head->arg.len -
 535                                                         head->byte_len;
 536                         head->arg.page_len = 0;
 537                 } else {
 538                         head->arg.head[0].iov_len = head->sge[0].length -
 539                                                                 head->byte_len;
 540                         head->arg.page_len = head->arg.len -
 541                                                 head->sge[0].length;
 542                 }
 543         }
 544
 545         /* Point rq_arg.pages past header */
 546         rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
 547         rqstp->rq_arg.page_len = head->arg.page_len;
 548         rqstp->rq_arg.page_base = head->arg.page_base;
 549
 550         /* rq_respages starts after the last arg page */
 551         rqstp->rq_respages = &rqstp->rq_pages[page_no];
 552         rqstp->rq_next_page = rqstp->rq_respages + 1;
 553
 554         /* Rebuild rq_arg head and tail. */
 555         rqstp->rq_arg.head[0] = head->arg.head[0];
 556         rqstp->rq_arg.tail[0] = head->arg.tail[0];
 557         rqstp->rq_arg.len = head->arg.len;
 558         rqstp->rq_arg.buflen = head->arg.buflen;
 559 }
 560
 561 /* By convention, backchannel calls arrive via rdma_msg type
 562  * messages, and never populate the chunk lists. This makes
 563  * the RPC/RDMA header small and fixed in size, so it is
 564  * straightforward to check the RPC header's direction field.
 565  */
 566 static bool
 567 svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp)
 568 {
 569         __be32 *p = (__be32 *)rmsgp;
 570
 571         if (!xprt->xpt_bc_xprt)
 572                 return false;
 573
 574         if (rmsgp->rm_type != rdma_msg)
 575                 return false;
 576         if (rmsgp->rm_body.rm_chunks[0] != xdr_zero)
 577                 return false;
 578         if (rmsgp->rm_body.rm_chunks[1] != xdr_zero)
 579                 return false;
 580         if (rmsgp->rm_body.rm_chunks[2] != xdr_zero)
 581                 return false;
 582
 583         /* sanity */
 584         if (p[7] != rmsgp->rm_xid)
 585                 return false;
 586         /* call direction */
 587         if (p[8] == cpu_to_be32(RPC_CALL))
 588                 return false;
 589
 590         return true;
 591 }
 592
 593 /*
 594  * Set up the rqstp thread context to point to the RQ buffer. If
 595  * necessary, pull additional data from the client with an RDMA_READ
 596  * request.
 597  */
 598 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 599 {
 600         struct svc_xprt *xprt = rqstp->rq_xprt;
 601         struct svcxprt_rdma *rdma_xprt =
 602                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 603         struct svc_rdma_op_ctxt *ctxt = NULL;
 604         struct rpcrdma_msg *rmsgp;
 605         int ret = 0;
 606
 607         dprintk("svcrdma: rqstp=%p\n", rqstp);
 608
 609         spin_lock(&rdma_xprt->sc_rq_dto_lock);
 610         if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
 611                 ctxt = list_first_entry(&rdma_xprt->sc_read_complete_q,
 612                                         struct svc_rdma_op_ctxt, list);
 613                 list_del(&ctxt->list);
 614                 spin_unlock(&rdma_xprt->sc_rq_dto_lock);
 615                 rdma_read_complete(rqstp, ctxt);
 616                 goto complete;
 617         } else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
 618                 ctxt = list_first_entry(&rdma_xprt->sc_rq_dto_q,
 619                                         struct svc_rdma_op_ctxt, list);
 620                 list_del(&ctxt->list);
 621         } else {
 622                 atomic_inc(&rdma_stat_rq_starve);
 623                 clear_bit(XPT_DATA, &xprt->xpt_flags);
 624                 ctxt = NULL;
 625         }
 626         spin_unlock(&rdma_xprt->sc_rq_dto_lock);
 627         if (!ctxt) {
 628                 /* This is the EAGAIN path. The svc_recv routine will
 629                  * return -EAGAIN, the nfsd thread will go to call into
 630                  * svc_recv again and we shouldn't be on the active
 631                  * transport list
 632                  */
 633                 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
 634                         goto defer;
 635                 goto out;
 636         }
 637         dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p\n",
 638                 ctxt, rdma_xprt, rqstp);
 639         atomic_inc(&rdma_stat_recv);
 640
 641         /* Build up the XDR from the receive buffers. */
 642         rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
 643
 644         /* Decode the RDMA header. */
 645         rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
 646         ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg);
 647         if (ret < 0)
 648                 goto out_err;
 649         if (ret == 0)
 650                 goto out_drop;
 651         rqstp->rq_xprt_hlen = ret;
 652
 653         if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) {
 654                 ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp,
 655                                                &rqstp->rq_arg);
 656                 svc_rdma_put_context(ctxt, 0);
 657                 if (ret)
 658                         goto repost;
 659                 return ret;
 660         }
 661
 662         /* Read read-list data. */
 663         ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
 664         if (ret > 0) {
 665                 /* read-list posted, defer until data received from client. */
 666                 goto defer;
 667         } else if (ret < 0) {
 668                 /* Post of read-list failed, free context. */
 669                 svc_rdma_put_context(ctxt, 1);
 670                 return 0;
 671         }
 672
 673 complete:
 674         ret = rqstp->rq_arg.head[0].iov_len
 675                 + rqstp->rq_arg.page_len
 676                 + rqstp->rq_arg.tail[0].iov_len;
 677         svc_rdma_put_context(ctxt, 0);
 678  out:
 679         dprintk("svcrdma: ret=%d, rq_arg.len=%u, "
 680                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zd\n",
 681                 ret, rqstp->rq_arg.len,
 682                 rqstp->rq_arg.head[0].iov_base,
 683                 rqstp->rq_arg.head[0].iov_len);
 684         rqstp->rq_prot = IPPROTO_MAX;
 685         svc_xprt_copy_addrs(rqstp, xprt);
 686         return ret;
 687
 688 out_err:
 689         svc_rdma_send_error(rdma_xprt, rmsgp, ret);
 690         svc_rdma_put_context(ctxt, 0);
 691         return 0;
 692
 693 defer:
 694         return 0;
 695
 696 out_drop:
 697         svc_rdma_put_context(ctxt, 1);
 698 repost:
 699         return svc_rdma_repost_recv(rdma_xprt, GFP_KERNEL);
 700 }