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 struct chunk_sge {
 116         int start;              /* sge no for this chunk */
 117         int count;              /* sge count for this chunk */
 118 };
 119
 120 /* Encode a read-chunk-list as an array of IB SGE
 121  *
 122  * Assumptions:
 123  * - chunk[0]->position points to pages[0] at an offset of 0
 124  * - pages[] is not physically or virtually contigous and consists of
 125  *   PAGE_SIZE elements.
 126  *
 127  * Output:
 128  * - sge array pointing into pages[] array.
 129  * - chunk_sge array specifying sge index and count for each
 130  *   chunk in the read list
 131  *
 132  */
 133 static int rdma_rcl_to_sge(struct svcxprt_rdma *xprt,
 134                            struct svc_rqst *rqstp,
 135                            struct svc_rdma_op_ctxt *head,
 136                            struct rpcrdma_msg *rmsgp,
 137                            struct ib_sge *sge,
 138                            struct chunk_sge *ch_sge_ary,
 139                            int ch_count,
 140                            int byte_count)
 141 {
 142         int sge_no;
 143         int sge_bytes;
 144         int page_off;
 145         int page_no;
 146         int ch_bytes;
 147         int ch_no;
 148         struct rpcrdma_read_chunk *ch;
 149
 150         sge_no = 0;
 151         page_no = 0;
 152         page_off = 0;
 153         ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 154         ch_no = 0;
 155         ch_bytes = ch->rc_target.rs_length;
 156         head->arg.head[0] = rqstp->rq_arg.head[0];
 157         head->arg.tail[0] = rqstp->rq_arg.tail[0];
 158         head->arg.pages = &head->pages[head->count];
 159         head->sge[0].length = head->count; /* save count of hdr pages */
 160         head->arg.page_base = 0;
 161         head->arg.page_len = ch_bytes;
 162         head->arg.len = rqstp->rq_arg.len + ch_bytes;
 163         head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
 164         head->count++;
 165         ch_sge_ary[0].start = 0;
 166         while (byte_count) {
 167                 sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
 168                 sge[sge_no].addr =
 169                         ib_dma_map_page(xprt->sc_cm_id->device,
 170                                         rqstp->rq_arg.pages[page_no],
 171                                         page_off, sge_bytes,
 172                                         DMA_FROM_DEVICE);
 173                 sge[sge_no].length = sge_bytes;
 174                 sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 175                 /*
 176                  * Don't bump head->count here because the same page
 177                  * may be used by multiple SGE.
 178                  */
 179                 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
 180                 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
 181
 182                 byte_count -= sge_bytes;
 183                 ch_bytes -= sge_bytes;
 184                 sge_no++;
 185                 /*
 186                  * If all bytes for this chunk have been mapped to an
 187                  * SGE, move to the next SGE
 188                  */
 189                 if (ch_bytes == 0) {
 190                         ch_sge_ary[ch_no].count =
 191                                 sge_no - ch_sge_ary[ch_no].start;
 192                         ch_no++;
 193                         ch++;
 194                         ch_sge_ary[ch_no].start = sge_no;
 195                         ch_bytes = ch->rc_target.rs_length;
 196                         /* If bytes remaining account for next chunk */
 197                         if (byte_count) {
 198                                 head->arg.page_len += ch_bytes;
 199                                 head->arg.len += ch_bytes;
 200                                 head->arg.buflen += ch_bytes;
 201                         }
 202                 }
 203                 /*
 204                  * If this SGE consumed all of the page, move to the
 205                  * next page
 206                  */
 207                 if ((sge_bytes + page_off) == PAGE_SIZE) {
 208                         page_no++;
 209                         page_off = 0;
 210                         /*
 211                          * If there are still bytes left to map, bump
 212                          * the page count
 213                          */
 214                         if (byte_count)
 215                                 head->count++;
 216                 } else
 217                         page_off += sge_bytes;
 218         }
 219         BUG_ON(byte_count != 0);
 220         return sge_no;
 221 }
 222
 223 static void rdma_set_ctxt_sge(struct svc_rdma_op_ctxt *ctxt,
 224                               struct ib_sge *sge,
 225                               u64 *sgl_offset,
 226                               int count)
 227 {
 228         int i;
 229
 230         ctxt->count = count;
 231         for (i = 0; i < count; i++) {
 232                 ctxt->sge[i].addr = sge[i].addr;
 233                 ctxt->sge[i].length = sge[i].length;
 234                 *sgl_offset = *sgl_offset + sge[i].length;
 235         }
 236 }
 237
 238 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
 239 {
 240         if ((RDMA_TRANSPORT_IWARP ==
 241              rdma_node_get_transport(xprt->sc_cm_id->
 242                                      device->node_type))
 243             && sge_count > 1)
 244                 return 1;
 245         else
 246                 return min_t(int, sge_count, xprt->sc_max_sge);
 247 }
 248
 249 /*
 250  * Use RDMA_READ to read data from the advertised client buffer into the
 251  * XDR stream starting at rq_arg.head[0].iov_base.
 252  * Each chunk in the array
 253  * contains the following fields:
 254  * discrim      - '1', This isn't used for data placement
 255  * position     - The xdr stream offset (the same for every chunk)
 256  * handle       - RMR for client memory region
 257  * length       - data transfer length
 258  * offset       - 64 bit tagged offset in remote memory region
 259  *
 260  * On our side, we need to read into a pagelist. The first page immediately
 261  * follows the RPC header.
 262  *
 263  * This function returns:
 264  * 0 - No error and no read-list found.
 265  *
 266  * 1 - Successful read-list processing. The data is not yet in
 267  * the pagelist and therefore the RPC request must be deferred. The
 268  * I/O completion will enqueue the transport again and
 269  * svc_rdma_recvfrom will complete the request.
 270  *
 271  * <0 - Error processing/posting read-list.
 272  *
 273  * NOTE: The ctxt must not be touched after the last WR has been posted
 274  * because the I/O completion processing may occur on another
 275  * processor and free / modify the context. Ne touche pas!
 276  */
 277 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
 278                          struct rpcrdma_msg *rmsgp,
 279                          struct svc_rqst *rqstp,
 280                          struct svc_rdma_op_ctxt *hdr_ctxt)
 281 {
 282         struct ib_send_wr read_wr;
 283         int err = 0;
 284         int ch_no;
 285         struct ib_sge *sge;
 286         int ch_count;
 287         int byte_count;
 288         int sge_count;
 289         u64 sgl_offset;
 290         struct rpcrdma_read_chunk *ch;
 291         struct svc_rdma_op_ctxt *ctxt = NULL;
 292         struct svc_rdma_op_ctxt *tmp_sge_ctxt;
 293         struct svc_rdma_op_ctxt *tmp_ch_ctxt;
 294         struct chunk_sge *ch_sge_ary;
 295
 296         /* If no read list is present, return 0 */
 297         ch = svc_rdma_get_read_chunk(rmsgp);
 298         if (!ch)
 299                 return 0;
 300
 301         /* Allocate temporary contexts to keep SGE */
 302         BUG_ON(sizeof(struct ib_sge) < sizeof(struct chunk_sge));
 303         tmp_sge_ctxt = svc_rdma_get_context(xprt);
 304         sge = tmp_sge_ctxt->sge;
 305         tmp_ch_ctxt = svc_rdma_get_context(xprt);
 306         ch_sge_ary = (struct chunk_sge *)tmp_ch_ctxt->sge;
 307
 308         svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
 309         sge_count = rdma_rcl_to_sge(xprt, rqstp, hdr_ctxt, rmsgp,
 310                                     sge, ch_sge_ary,
 311                                     ch_count, byte_count);
 312         sgl_offset = 0;
 313         ch_no = 0;
 314
 315         for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
 316              ch->rc_discrim != 0; ch++, ch_no++) {
 317 next_sge:
 318                 ctxt = svc_rdma_get_context(xprt);
 319                 ctxt->direction = DMA_FROM_DEVICE;
 320                 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 321
 322                 /* Prepare READ WR */
 323                 memset(&read_wr, 0, sizeof read_wr);
 324                 ctxt->wr_op = IB_WR_RDMA_READ;
 325                 read_wr.wr_id = (unsigned long)ctxt;
 326                 read_wr.opcode = IB_WR_RDMA_READ;
 327                 read_wr.send_flags = IB_SEND_SIGNALED;
 328                 read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
 329                 read_wr.wr.rdma.remote_addr =
 330                         get_unaligned(&(ch->rc_target.rs_offset)) +
 331                         sgl_offset;
 332                 read_wr.sg_list = &sge[ch_sge_ary[ch_no].start];
 333                 read_wr.num_sge =
 334                         rdma_read_max_sge(xprt, ch_sge_ary[ch_no].count);
 335                 rdma_set_ctxt_sge(ctxt, &sge[ch_sge_ary[ch_no].start],
 336                                   &sgl_offset,
 337                                   read_wr.num_sge);
 338                 if (((ch+1)->rc_discrim == 0) &&
 339                     (read_wr.num_sge == ch_sge_ary[ch_no].count)) {
 340                         /*
 341                          * Mark the last RDMA_READ with a bit to
 342                          * indicate all RPC data has been fetched from
 343                          * the client and the RPC needs to be enqueued.
 344                          */
 345                         set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 346                         ctxt->read_hdr = hdr_ctxt;
 347                 }
 348                 /* Post the read */
 349                 err = svc_rdma_send(xprt, &read_wr);
 350                 if (err) {
 351                         printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
 352                                err);
 353                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 354                         svc_rdma_put_context(ctxt, 0);
 355                         goto out;
 356                 }
 357                 atomic_inc(&rdma_stat_read);
 358
 359                 if (read_wr.num_sge < ch_sge_ary[ch_no].count) {
 360                         ch_sge_ary[ch_no].count -= read_wr.num_sge;
 361                         ch_sge_ary[ch_no].start += read_wr.num_sge;
 362                         goto next_sge;
 363                 }
 364                 sgl_offset = 0;
 365                 err = 1;
 366         }
 367
 368  out:
 369         svc_rdma_put_context(tmp_sge_ctxt, 0);
 370         svc_rdma_put_context(tmp_ch_ctxt, 0);
 371
 372         /* Detach arg pages. svc_recv will replenish them */
 373         for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
 374                 rqstp->rq_pages[ch_no] = NULL;
 375
 376         /*
 377          * Detach res pages. svc_release must see a resused count of
 378          * zero or it will attempt to put them.
 379          */
 380         while (rqstp->rq_resused)
 381                 rqstp->rq_respages[--rqstp->rq_resused] = NULL;
 382
 383         return err;
 384 }
 385
 386 static int rdma_read_complete(struct svc_rqst *rqstp,
 387                               struct svc_rdma_op_ctxt *head)
 388 {
 389         int page_no;
 390         int ret;
 391
 392         BUG_ON(!head);
 393
 394         /* Copy RPC pages */
 395         for (page_no = 0; page_no < head->count; page_no++) {
 396                 put_page(rqstp->rq_pages[page_no]);
 397                 rqstp->rq_pages[page_no] = head->pages[page_no];
 398         }
 399         /* Point rq_arg.pages past header */
 400         rqstp->rq_arg.pages = &rqstp->rq_pages[head->sge[0].length];
 401         rqstp->rq_arg.page_len = head->arg.page_len;
 402         rqstp->rq_arg.page_base = head->arg.page_base;
 403
 404         /* rq_respages starts after the last arg page */
 405         rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
 406         rqstp->rq_resused = 0;
 407
 408         /* Rebuild rq_arg head and tail. */
 409         rqstp->rq_arg.head[0] = head->arg.head[0];
 410         rqstp->rq_arg.tail[0] = head->arg.tail[0];
 411         rqstp->rq_arg.len = head->arg.len;
 412         rqstp->rq_arg.buflen = head->arg.buflen;
 413
 414         /* Free the context */
 415         svc_rdma_put_context(head, 0);
 416
 417         /* XXX: What should this be? */
 418         rqstp->rq_prot = IPPROTO_MAX;
 419
 420         ret = rqstp->rq_arg.head[0].iov_len
 421                 + rqstp->rq_arg.page_len
 422                 + rqstp->rq_arg.tail[0].iov_len;
 423         dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
 424                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 425                 ret, rqstp->rq_arg.len, rqstp->rq_arg.head[0].iov_base,
 426                 rqstp->rq_arg.head[0].iov_len);
 427
 428         svc_xprt_received(rqstp->rq_xprt);
 429         return ret;
 430 }
 431
 432 /*
 433  * Set up the rqstp thread context to point to the RQ buffer. If
 434  * necessary, pull additional data from the client with an RDMA_READ
 435  * request.
 436  */
 437 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 438 {
 439         struct svc_xprt *xprt = rqstp->rq_xprt;
 440         struct svcxprt_rdma *rdma_xprt =
 441                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 442         struct svc_rdma_op_ctxt *ctxt = NULL;
 443         struct rpcrdma_msg *rmsgp;
 444         int ret = 0;
 445         int len;
 446
 447         dprintk("svcrdma: rqstp=%p\n", rqstp);
 448
 449         spin_lock_bh(&rdma_xprt->sc_read_complete_lock);
 450         if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
 451                 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
 452                                   struct svc_rdma_op_ctxt,
 453                                   dto_q);
 454                 list_del_init(&ctxt->dto_q);
 455         }
 456         spin_unlock_bh(&rdma_xprt->sc_read_complete_lock);
 457         if (ctxt)
 458                 return rdma_read_complete(rqstp, ctxt);
 459
 460         spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
 461         if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
 462                 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
 463                                   struct svc_rdma_op_ctxt,
 464                                   dto_q);
 465                 list_del_init(&ctxt->dto_q);
 466         } else {
 467                 atomic_inc(&rdma_stat_rq_starve);
 468                 clear_bit(XPT_DATA, &xprt->xpt_flags);
 469                 ctxt = NULL;
 470         }
 471         spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
 472         if (!ctxt) {
 473                 /* This is the EAGAIN path. The svc_recv routine will
 474                  * return -EAGAIN, the nfsd thread will go to call into
 475                  * svc_recv again and we shouldn't be on the active
 476                  * transport list
 477                  */
 478                 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
 479                         goto close_out;
 480
 481                 BUG_ON(ret);
 482                 goto out;
 483         }
 484         dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
 485                 ctxt, rdma_xprt, rqstp, ctxt->wc_status);
 486         BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
 487         atomic_inc(&rdma_stat_recv);
 488
 489         /* Build up the XDR from the receive buffers. */
 490         rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
 491
 492         /* Decode the RDMA header. */
 493         len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
 494         rqstp->rq_xprt_hlen = len;
 495
 496         /* If the request is invalid, reply with an error */
 497         if (len < 0) {
 498                 if (len == -ENOSYS)
 499                         (void)svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
 500                 goto close_out;
 501         }
 502
 503         /* Read read-list data. */
 504         ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
 505         if (ret > 0) {
 506                 /* read-list posted, defer until data received from client. */
 507                 svc_xprt_received(xprt);
 508                 return 0;
 509         }
 510         if (ret < 0) {
 511                 /* Post of read-list failed, free context. */
 512                 svc_rdma_put_context(ctxt, 1);
 513                 return 0;
 514         }
 515
 516         ret = rqstp->rq_arg.head[0].iov_len
 517                 + rqstp->rq_arg.page_len
 518                 + rqstp->rq_arg.tail[0].iov_len;
 519         svc_rdma_put_context(ctxt, 0);
 520  out:
 521         dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
 522                 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
 523                 ret, rqstp->rq_arg.len,
 524                 rqstp->rq_arg.head[0].iov_base,
 525                 rqstp->rq_arg.head[0].iov_len);
 526         rqstp->rq_prot = IPPROTO_MAX;
 527         svc_xprt_copy_addrs(rqstp, xprt);
 528         svc_xprt_received(xprt);
 529         return ret;
 530
 531  close_out:
 532         if (ctxt)
 533                 svc_rdma_put_context(ctxt, 1);
 534         dprintk("svcrdma: transport %p is closing\n", xprt);
 535         /*
 536          * Set the close bit and enqueue it. svc_recv will see the
 537          * close bit and call svc_xprt_delete
 538          */
 539         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 540         svc_xprt_received(xprt);
 541         return 0;
 542 }