net/sunrpc/xprtrdma/svc_rdma_sendto.c

   1 /*
   2  * Copyright (c) 2016 Oracle. All rights reserved.
   3  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
   4  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the BSD-type
  10  * license below:
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  *
  16  *      Redistributions of source code must retain the above copyright
  17  *      notice, this list of conditions and the following disclaimer.
  18  *
  19  *      Redistributions in binary form must reproduce the above
  20  *      copyright notice, this list of conditions and the following
  21  *      disclaimer in the documentation and/or other materials provided
  22  *      with the distribution.
  23  *
  24  *      Neither the name of the Network Appliance, Inc. nor the names of
  25  *      its contributors may be used to endorse or promote products
  26  *      derived from this software without specific prior written
  27  *      permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  40  *
  41  * Author: Tom Tucker <tom@opengridcomputing.com>
  42  */
  43
  44 /* Operation
  45  *
  46  * The main entry point is svc_rdma_sendto. This is called by the
  47  * RPC server when an RPC Reply is ready to be transmitted to a client.
  48  *
  49  * The passed-in svc_rqst contains a struct xdr_buf which holds an
  50  * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
  51  * transport header, post all Write WRs needed for this Reply, then post
  52  * a Send WR conveying the transport header and the RPC message itself to
  53  * the client.
  54  *
  55  * svc_rdma_sendto must fully transmit the Reply before returning, as
  56  * the svc_rqst will be recycled as soon as sendto returns. Remaining
  57  * resources referred to by the svc_rqst are also recycled at that time.
  58  * Therefore any resources that must remain longer must be detached
  59  * from the svc_rqst and released later.
  60  *
  61  * Page Management
  62  *
  63  * The I/O that performs Reply transmission is asynchronous, and may
  64  * complete well after sendto returns. Thus pages under I/O must be
  65  * removed from the svc_rqst before sendto returns.
  66  *
  67  * The logic here depends on Send Queue and completion ordering. Since
  68  * the Send WR is always posted last, it will always complete last. Thus
  69  * when it completes, it is guaranteed that all previous Write WRs have
  70  * also completed.
  71  *
  72  * Write WRs are constructed and posted. Each Write segment gets its own
  73  * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
  74  * DMA-unmap the pages under I/O for that Write segment. The Write
  75  * completion handler does not release any pages.
  76  *
  77  * When the Send WR is constructed, it also gets its own svc_rdma_op_ctxt.
  78  * The ownership of all of the Reply's pages are transferred into that
  79  * ctxt, the Send WR is posted, and sendto returns.
  80  *
  81  * The svc_rdma_op_ctxt is presented when the Send WR completes. The
  82  * Send completion handler finally releases the Reply's pages.
  83  *
  84  * This mechanism also assumes that completions on the transport's Send
  85  * Completion Queue do not run in parallel. Otherwise a Write completion
  86  * and Send completion running at the same time could release pages that
  87  * are still DMA-mapped.
  88  *
  89  * Error Handling
  90  *
  91  * - If the Send WR is posted successfully, it will either complete
  92  *   successfully, or get flushed. Either way, the Send completion
  93  *   handler releases the Reply's pages.
  94  * - If the Send WR cannot be not posted, the forward path releases
  95  *   the Reply's pages.
  96  *
  97  * This handles the case, without the use of page reference counting,
  98  * where two different Write segments send portions of the same page.
  99  */
 100
 101 #include <linux/sunrpc/debug.h>
 102 #include <linux/sunrpc/rpc_rdma.h>
 103 #include <linux/spinlock.h>
 104 #include <asm/unaligned.h>
 105 #include <rdma/ib_verbs.h>
 106 #include <rdma/rdma_cm.h>
 107 #include <linux/sunrpc/svc_rdma.h>
 108
 109 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
 110
 111 static u32 xdr_padsize(u32 len)
 112 {
 113         return (len & 3) ? (4 - (len & 3)) : 0;
 114 }
 115
 116 /* Returns length of transport header, in bytes.
 117  */
 118 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
 119 {
 120         unsigned int nsegs;
 121         __be32 *p;
 122
 123         p = rdma_resp;
 124
 125         /* RPC-over-RDMA V1 replies never have a Read list. */
 126         p += rpcrdma_fixed_maxsz + 1;
 127
 128         /* Skip Write list. */
 129         while (*p++ != xdr_zero) {
 130                 nsegs = be32_to_cpup(p++);
 131                 p += nsegs * rpcrdma_segment_maxsz;
 132         }
 133
 134         /* Skip Reply chunk. */
 135         if (*p++ != xdr_zero) {
 136                 nsegs = be32_to_cpup(p++);
 137                 p += nsegs * rpcrdma_segment_maxsz;
 138         }
 139
 140         return (unsigned long)p - (unsigned long)rdma_resp;
 141 }
 142
 143 /* One Write chunk is copied from Call transport header to Reply
 144  * transport header. Each segment's length field is updated to
 145  * reflect number of bytes consumed in the segment.
 146  *
 147  * Returns number of segments in this chunk.
 148  */
 149 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
 150                                            unsigned int remaining)
 151 {
 152         unsigned int i, nsegs;
 153         u32 seg_len;
 154
 155         /* Write list discriminator */
 156         *dst++ = *src++;
 157
 158         /* number of segments in this chunk */
 159         nsegs = be32_to_cpup(src);
 160         *dst++ = *src++;
 161
 162         for (i = nsegs; i; i--) {
 163                 /* segment's RDMA handle */
 164                 *dst++ = *src++;
 165
 166                 /* bytes returned in this segment */
 167                 seg_len = be32_to_cpu(*src);
 168                 if (remaining >= seg_len) {
 169                         /* entire segment was consumed */
 170                         *dst = *src;
 171                         remaining -= seg_len;
 172                 } else {
 173                         /* segment only partly filled */
 174                         *dst = cpu_to_be32(remaining);
 175                         remaining = 0;
 176                 }
 177                 dst++; src++;
 178
 179                 /* segment's RDMA offset */
 180                 *dst++ = *src++;
 181                 *dst++ = *src++;
 182         }
 183
 184         return nsegs;
 185 }
 186
 187 /* The client provided a Write list in the Call message. Fill in
 188  * the segments in the first Write chunk in the Reply's transport
 189  * header with the number of bytes consumed in each segment.
 190  * Remaining chunks are returned unused.
 191  *
 192  * Assumptions:
 193  *  - Client has provided only one Write chunk
 194  */
 195 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
 196                                            unsigned int consumed)
 197 {
 198         unsigned int nsegs;
 199         __be32 *p, *q;
 200
 201         /* RPC-over-RDMA V1 replies never have a Read list. */
 202         p = rdma_resp + rpcrdma_fixed_maxsz + 1;
 203
 204         q = wr_ch;
 205         while (*q != xdr_zero) {
 206                 nsegs = xdr_encode_write_chunk(p, q, consumed);
 207                 q += 2 + nsegs * rpcrdma_segment_maxsz;
 208                 p += 2 + nsegs * rpcrdma_segment_maxsz;
 209                 consumed = 0;
 210         }
 211
 212         /* Terminate Write list */
 213         *p++ = xdr_zero;
 214
 215         /* Reply chunk discriminator; may be replaced later */
 216         *p = xdr_zero;
 217 }
 218
 219 /* The client provided a Reply chunk in the Call message. Fill in
 220  * the segments in the Reply chunk in the Reply message with the
 221  * number of bytes consumed in each segment.
 222  *
 223  * Assumptions:
 224  * - Reply can always fit in the provided Reply chunk
 225  */
 226 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
 227                                             unsigned int consumed)
 228 {
 229         __be32 *p;
 230
 231         /* Find the Reply chunk in the Reply's xprt header.
 232          * RPC-over-RDMA V1 replies never have a Read list.
 233          */
 234         p = rdma_resp + rpcrdma_fixed_maxsz + 1;
 235
 236         /* Skip past Write list */
 237         while (*p++ != xdr_zero)
 238                 p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
 239
 240         xdr_encode_write_chunk(p, rp_ch, consumed);
 241 }
 242
 243 /* Parse the RPC Call's transport header.
 244  */
 245 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
 246                                       __be32 **write, __be32 **reply)
 247 {
 248         __be32 *p;
 249
 250         p = rdma_argp + rpcrdma_fixed_maxsz;
 251
 252         /* Read list */
 253         while (*p++ != xdr_zero)
 254                 p += 5;
 255
 256         /* Write list */
 257         if (*p != xdr_zero) {
 258                 *write = p;
 259                 while (*p++ != xdr_zero)
 260                         p += 1 + be32_to_cpu(*p) * 4;
 261         } else {
 262                 *write = NULL;
 263                 p++;
 264         }
 265
 266         /* Reply chunk */
 267         if (*p != xdr_zero)
 268                 *reply = p;
 269         else
 270                 *reply = NULL;
 271 }
 272
 273 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
 274  * Responder's choice: requester signals it can handle Send With
 275  * Invalidate, and responder chooses one rkey to invalidate.
 276  *
 277  * Find a candidate rkey to invalidate when sending a reply.  Picks the
 278  * first R_key it finds in the chunk lists.
 279  *
 280  * Returns zero if RPC's chunk lists are empty.
 281  */
 282 static u32 svc_rdma_get_inv_rkey(__be32 *rdma_argp,
 283                                  __be32 *wr_lst, __be32 *rp_ch)
 284 {
 285         __be32 *p;
 286
 287         p = rdma_argp + rpcrdma_fixed_maxsz;
 288         if (*p != xdr_zero)
 289                 p += 2;
 290         else if (wr_lst && be32_to_cpup(wr_lst + 1))
 291                 p = wr_lst + 2;
 292         else if (rp_ch && be32_to_cpup(rp_ch + 1))
 293                 p = rp_ch + 2;
 294         else
 295                 return 0;
 296         return be32_to_cpup(p);
 297 }
 298
 299 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
 300  * is used during completion to DMA-unmap this memory, and
 301  * it uses ib_dma_unmap_page() exclusively.
 302  */
 303 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
 304                                 struct svc_rdma_op_ctxt *ctxt,
 305                                 unsigned int sge_no,
 306                                 unsigned char *base,
 307                                 unsigned int len)
 308 {
 309         unsigned long offset = (unsigned long)base & ~PAGE_MASK;
 310         struct ib_device *dev = rdma->sc_cm_id->device;
 311         dma_addr_t dma_addr;
 312
 313         dma_addr = ib_dma_map_page(dev, virt_to_page(base),
 314                                    offset, len, DMA_TO_DEVICE);
 315         if (ib_dma_mapping_error(dev, dma_addr))
 316                 goto out_maperr;
 317
 318         ctxt->sge[sge_no].addr = dma_addr;
 319         ctxt->sge[sge_no].length = len;
 320         ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
 321         svc_rdma_count_mappings(rdma, ctxt);
 322         return 0;
 323
 324 out_maperr:
 325         pr_err("svcrdma: failed to map buffer\n");
 326         return -EIO;
 327 }
 328
 329 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
 330                                  struct svc_rdma_op_ctxt *ctxt,
 331                                  unsigned int sge_no,
 332                                  struct page *page,
 333                                  unsigned int offset,
 334                                  unsigned int len)
 335 {
 336         struct ib_device *dev = rdma->sc_cm_id->device;
 337         dma_addr_t dma_addr;
 338
 339         dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
 340         if (ib_dma_mapping_error(dev, dma_addr))
 341                 goto out_maperr;
 342
 343         ctxt->sge[sge_no].addr = dma_addr;
 344         ctxt->sge[sge_no].length = len;
 345         ctxt->sge[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
 346         svc_rdma_count_mappings(rdma, ctxt);
 347         return 0;
 348
 349 out_maperr:
 350         pr_err("svcrdma: failed to map page\n");
 351         return -EIO;
 352 }
 353
 354 /**
 355  * svc_rdma_map_reply_hdr - DMA map the transport header buffer
 356  * @rdma: controlling transport
 357  * @ctxt: op_ctxt for the Send WR
 358  * @rdma_resp: buffer containing transport header
 359  * @len: length of transport header
 360  *
 361  * Returns:
 362  *      %0 if the header is DMA mapped,
 363  *      %-EIO if DMA mapping failed.
 364  */
 365 int svc_rdma_map_reply_hdr(struct svcxprt_rdma *rdma,
 366                            struct svc_rdma_op_ctxt *ctxt,
 367                            __be32 *rdma_resp,
 368                            unsigned int len)
 369 {
 370         ctxt->direction = DMA_TO_DEVICE;
 371         ctxt->pages[0] = virt_to_page(rdma_resp);
 372         ctxt->count = 1;
 373         return svc_rdma_dma_map_page(rdma, ctxt, 0, ctxt->pages[0], 0, len);
 374 }
 375
 376 /* Load the xdr_buf into the ctxt's sge array, and DMA map each
 377  * element as it is added.
 378  *
 379  * Returns the number of sge elements loaded on success, or
 380  * a negative errno on failure.
 381  */
 382 static int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
 383                                   struct svc_rdma_op_ctxt *ctxt,
 384                                   struct xdr_buf *xdr, __be32 *wr_lst)
 385 {
 386         unsigned int len, sge_no, remaining, page_off;
 387         struct page **ppages;
 388         unsigned char *base;
 389         u32 xdr_pad;
 390         int ret;
 391
 392         sge_no = 1;
 393
 394         ret = svc_rdma_dma_map_buf(rdma, ctxt, sge_no++,
 395                                    xdr->head[0].iov_base,
 396                                    xdr->head[0].iov_len);
 397         if (ret < 0)
 398                 return ret;
 399
 400         /* If a Write chunk is present, the xdr_buf's page list
 401          * is not included inline. However the Upper Layer may
 402          * have added XDR padding in the tail buffer, and that
 403          * should not be included inline.
 404          */
 405         if (wr_lst) {
 406                 base = xdr->tail[0].iov_base;
 407                 len = xdr->tail[0].iov_len;
 408                 xdr_pad = xdr_padsize(xdr->page_len);
 409
 410                 if (len && xdr_pad) {
 411                         base += xdr_pad;
 412                         len -= xdr_pad;
 413                 }
 414
 415                 goto tail;
 416         }
 417
 418         ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
 419         page_off = xdr->page_base & ~PAGE_MASK;
 420         remaining = xdr->page_len;
 421         while (remaining) {
 422                 len = min_t(u32, PAGE_SIZE - page_off, remaining);
 423
 424                 ret = svc_rdma_dma_map_page(rdma, ctxt, sge_no++,
 425                                             *ppages++, page_off, len);
 426                 if (ret < 0)
 427                         return ret;
 428
 429                 remaining -= len;
 430                 page_off = 0;
 431         }
 432
 433         base = xdr->tail[0].iov_base;
 434         len = xdr->tail[0].iov_len;
 435 tail:
 436         if (len) {
 437                 ret = svc_rdma_dma_map_buf(rdma, ctxt, sge_no++, base, len);
 438                 if (ret < 0)
 439                         return ret;
 440         }
 441
 442         return sge_no - 1;
 443 }
 444
 445 /* The svc_rqst and all resources it owns are released as soon as
 446  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
 447  * so they are released by the Send completion handler.
 448  */
 449 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
 450                                    struct svc_rdma_op_ctxt *ctxt)
 451 {
 452         int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
 453
 454         ctxt->count += pages;
 455         for (i = 0; i < pages; i++) {
 456                 ctxt->pages[i + 1] = rqstp->rq_respages[i];
 457                 rqstp->rq_respages[i] = NULL;
 458         }
 459         rqstp->rq_next_page = rqstp->rq_respages + 1;
 460 }
 461
 462 /**
 463  * svc_rdma_post_send_wr - Set up and post one Send Work Request
 464  * @rdma: controlling transport
 465  * @ctxt: op_ctxt for transmitting the Send WR
 466  * @num_sge: number of SGEs to send
 467  * @inv_rkey: R_key argument to Send With Invalidate, or zero
 468  *
 469  * Returns:
 470  *      %0 if the Send* was posted successfully,
 471  *      %-ENOTCONN if the connection was lost or dropped,
 472  *      %-EINVAL if there was a problem with the Send we built,
 473  *      %-ENOMEM if ib_post_send failed.
 474  */
 475 int svc_rdma_post_send_wr(struct svcxprt_rdma *rdma,
 476                           struct svc_rdma_op_ctxt *ctxt, int num_sge,
 477                           u32 inv_rkey)
 478 {
 479         struct ib_send_wr *send_wr = &ctxt->send_wr;
 480
 481         dprintk("svcrdma: posting Send WR with %u sge(s)\n", num_sge);
 482
 483         send_wr->next = NULL;
 484         ctxt->cqe.done = svc_rdma_wc_send;
 485         send_wr->wr_cqe = &ctxt->cqe;
 486         send_wr->sg_list = ctxt->sge;
 487         send_wr->num_sge = num_sge;
 488         send_wr->send_flags = IB_SEND_SIGNALED;
 489         if (inv_rkey) {
 490                 send_wr->opcode = IB_WR_SEND_WITH_INV;
 491                 send_wr->ex.invalidate_rkey = inv_rkey;
 492         } else {
 493                 send_wr->opcode = IB_WR_SEND;
 494         }
 495
 496         return svc_rdma_send(rdma, send_wr);
 497 }
 498
 499 /* Prepare the portion of the RPC Reply that will be transmitted
 500  * via RDMA Send. The RPC-over-RDMA transport header is prepared
 501  * in sge[0], and the RPC xdr_buf is prepared in following sges.
 502  *
 503  * Depending on whether a Write list or Reply chunk is present,
 504  * the server may send all, a portion of, or none of the xdr_buf.
 505  * In the latter case, only the transport header (sge[0]) is
 506  * transmitted.
 507  *
 508  * RDMA Send is the last step of transmitting an RPC reply. Pages
 509  * involved in the earlier RDMA Writes are here transferred out
 510  * of the rqstp and into the ctxt's page array. These pages are
 511  * DMA unmapped by each Write completion, but the subsequent Send
 512  * completion finally releases these pages.
 513  *
 514  * Assumptions:
 515  * - The Reply's transport header will never be larger than a page.
 516  */
 517 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
 518                                    __be32 *rdma_argp, __be32 *rdma_resp,
 519                                    struct svc_rqst *rqstp,
 520                                    __be32 *wr_lst, __be32 *rp_ch)
 521 {
 522         struct svc_rdma_op_ctxt *ctxt;
 523         u32 inv_rkey;
 524         int ret;
 525
 526         dprintk("svcrdma: sending %s reply: head=%zu, pagelen=%u, tail=%zu\n",
 527                 (rp_ch ? "RDMA_NOMSG" : "RDMA_MSG"),
 528                 rqstp->rq_res.head[0].iov_len,
 529                 rqstp->rq_res.page_len,
 530                 rqstp->rq_res.tail[0].iov_len);
 531
 532         ctxt = svc_rdma_get_context(rdma);
 533
 534         ret = svc_rdma_map_reply_hdr(rdma, ctxt, rdma_resp,
 535                                      svc_rdma_reply_hdr_len(rdma_resp));
 536         if (ret < 0)
 537                 goto err;
 538
 539         if (!rp_ch) {
 540                 ret = svc_rdma_map_reply_msg(rdma, ctxt,
 541                                              &rqstp->rq_res, wr_lst);
 542                 if (ret < 0)
 543                         goto err;
 544         }
 545
 546         svc_rdma_save_io_pages(rqstp, ctxt);
 547
 548         inv_rkey = 0;
 549         if (rdma->sc_snd_w_inv)
 550                 inv_rkey = svc_rdma_get_inv_rkey(rdma_argp, wr_lst, rp_ch);
 551         ret = svc_rdma_post_send_wr(rdma, ctxt, 1 + ret, inv_rkey);
 552         if (ret)
 553                 goto err;
 554
 555         return 0;
 556
 557 err:
 558         svc_rdma_unmap_dma(ctxt);
 559         svc_rdma_put_context(ctxt, 1);
 560         return ret;
 561 }
 562
 563 /* Given the client-provided Write and Reply chunks, the server was not
 564  * able to form a complete reply. Return an RDMA_ERROR message so the
 565  * client can retire this RPC transaction. As above, the Send completion
 566  * routine releases payload pages that were part of a previous RDMA Write.
 567  *
 568  * Remote Invalidation is skipped for simplicity.
 569  */
 570 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
 571                                    __be32 *rdma_resp, struct svc_rqst *rqstp)
 572 {
 573         struct svc_rdma_op_ctxt *ctxt;
 574         __be32 *p;
 575         int ret;
 576
 577         ctxt = svc_rdma_get_context(rdma);
 578
 579         /* Replace the original transport header with an
 580          * RDMA_ERROR response. XID etc are preserved.
 581          */
 582         p = rdma_resp + 3;
 583         *p++ = rdma_error;
 584         *p   = err_chunk;
 585
 586         ret = svc_rdma_map_reply_hdr(rdma, ctxt, rdma_resp, 20);
 587         if (ret < 0)
 588                 goto err;
 589
 590         svc_rdma_save_io_pages(rqstp, ctxt);
 591
 592         ret = svc_rdma_post_send_wr(rdma, ctxt, 1 + ret, 0);
 593         if (ret)
 594                 goto err;
 595
 596         return 0;
 597
 598 err:
 599         pr_err("svcrdma: failed to post Send WR (%d)\n", ret);
 600         svc_rdma_unmap_dma(ctxt);
 601         svc_rdma_put_context(ctxt, 1);
 602         return ret;
 603 }
 604
 605 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
 606 {
 607 }
 608
 609 /**
 610  * svc_rdma_sendto - Transmit an RPC reply
 611  * @rqstp: processed RPC request, reply XDR already in ::rq_res
 612  *
 613  * Any resources still associated with @rqstp are released upon return.
 614  * If no reply message was possible, the connection is closed.
 615  *
 616  * Returns:
 617  *      %0 if an RPC reply has been successfully posted,
 618  *      %-ENOMEM if a resource shortage occurred (connection is lost),
 619  *      %-ENOTCONN if posting failed (connection is lost).
 620  */
 621 int svc_rdma_sendto(struct svc_rqst *rqstp)
 622 {
 623         struct svc_xprt *xprt = rqstp->rq_xprt;
 624         struct svcxprt_rdma *rdma =
 625                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 626         __be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
 627         struct xdr_buf *xdr = &rqstp->rq_res;
 628         struct page *res_page;
 629         int ret;
 630
 631         /* Find the call's chunk lists to decide how to send the reply.
 632          * Receive places the Call's xprt header at the start of page 0.
 633          */
 634         rdma_argp = page_address(rqstp->rq_pages[0]);
 635         svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
 636
 637         dprintk("svcrdma: preparing response for XID 0x%08x\n",
 638                 be32_to_cpup(rdma_argp));
 639
 640         /* Create the RDMA response header. xprt->xpt_mutex,
 641          * acquired in svc_send(), serializes RPC replies. The
 642          * code path below that inserts the credit grant value
 643          * into each transport header runs only inside this
 644          * critical section.
 645          */
 646         ret = -ENOMEM;
 647         res_page = alloc_page(GFP_KERNEL);
 648         if (!res_page)
 649                 goto err0;
 650         rdma_resp = page_address(res_page);
 651
 652         p = rdma_resp;
 653         *p++ = *rdma_argp;
 654         *p++ = *(rdma_argp + 1);
 655         *p++ = rdma->sc_fc_credits;
 656         *p++ = rp_ch ? rdma_nomsg : rdma_msg;
 657
 658         /* Start with empty chunks */
 659         *p++ = xdr_zero;
 660         *p++ = xdr_zero;
 661         *p   = xdr_zero;
 662
 663         if (wr_lst) {
 664                 /* XXX: Presume the client sent only one Write chunk */
 665                 ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
 666                 if (ret < 0)
 667                         goto err2;
 668                 svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
 669         }
 670         if (rp_ch) {
 671                 ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
 672                 if (ret < 0)
 673                         goto err2;
 674                 svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
 675         }
 676
 677         ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
 678         if (ret)
 679                 goto err1;
 680         ret = svc_rdma_send_reply_msg(rdma, rdma_argp, rdma_resp, rqstp,
 681                                       wr_lst, rp_ch);
 682         if (ret < 0)
 683                 goto err0;
 684         return 0;
 685
 686  err2:
 687         if (ret != -E2BIG && ret != -EINVAL)
 688                 goto err1;
 689
 690         ret = svc_rdma_post_recv(rdma, GFP_KERNEL);
 691         if (ret)
 692                 goto err1;
 693         ret = svc_rdma_send_error_msg(rdma, rdma_resp, rqstp);
 694         if (ret < 0)
 695                 goto err0;
 696         return 0;
 697
 698  err1:
 699         put_page(res_page);
 700  err0:
 701         pr_err("svcrdma: Could not send reply, err=%d. Closing transport.\n",
 702                ret);
 703         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 704         return -ENOTCONN;
 705 }