root/net/sunrpc/xprtrdma/svc_rdma_sendto.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. svc_rdma_next_send_ctxt
  2. svc_rdma_send_ctxt_alloc
  3. svc_rdma_send_ctxts_destroy
  4. svc_rdma_send_ctxt_get
  5. svc_rdma_send_ctxt_put
  6. svc_rdma_wc_send
  7. svc_rdma_send
  8. xdr_padsize
  9. svc_rdma_reply_hdr_len
  10. xdr_encode_write_chunk
  11. svc_rdma_xdr_encode_write_list
  12. svc_rdma_xdr_encode_reply_chunk
  13. svc_rdma_get_write_arrays
  14. svc_rdma_dma_map_page
  15. svc_rdma_dma_map_buf
  16. svc_rdma_sync_reply_hdr
  17. svc_rdma_pull_up_needed
  18. svc_rdma_pull_up_reply_msg
  19. svc_rdma_map_reply_msg
  20. svc_rdma_save_io_pages
  21. svc_rdma_send_reply_msg
  22. svc_rdma_send_error_msg
  23. svc_rdma_sendto

   1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2 /*
   3  * Copyright (c) 2016-2018 Oracle. All rights reserved.
   4  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the BSD-type
  11  * license below:
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  *
  17  *      Redistributions of source code must retain the above copyright
  18  *      notice, this list of conditions and the following disclaimer.
  19  *
  20  *      Redistributions in binary form must reproduce the above
  21  *      copyright notice, this list of conditions and the following
  22  *      disclaimer in the documentation and/or other materials provided
  23  *      with the distribution.
  24  *
  25  *      Neither the name of the Network Appliance, Inc. nor the names of
  26  *      its contributors may be used to endorse or promote products
  27  *      derived from this software without specific prior written
  28  *      permission.
  29  *
  30  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  31  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  32  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  33  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  34  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  35  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  36  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  37  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  38  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  39  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  40  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  41  *
  42  * Author: Tom Tucker <tom@opengridcomputing.com>
  43  */
  44 
  45 /* Operation
  46  *
  47  * The main entry point is svc_rdma_sendto. This is called by the
  48  * RPC server when an RPC Reply is ready to be transmitted to a client.
  49  *
  50  * The passed-in svc_rqst contains a struct xdr_buf which holds an
  51  * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
  52  * transport header, post all Write WRs needed for this Reply, then post
  53  * a Send WR conveying the transport header and the RPC message itself to
  54  * the client.
  55  *
  56  * svc_rdma_sendto must fully transmit the Reply before returning, as
  57  * the svc_rqst will be recycled as soon as sendto returns. Remaining
  58  * resources referred to by the svc_rqst are also recycled at that time.
  59  * Therefore any resources that must remain longer must be detached
  60  * from the svc_rqst and released later.
  61  *
  62  * Page Management
  63  *
  64  * The I/O that performs Reply transmission is asynchronous, and may
  65  * complete well after sendto returns. Thus pages under I/O must be
  66  * removed from the svc_rqst before sendto returns.
  67  *
  68  * The logic here depends on Send Queue and completion ordering. Since
  69  * the Send WR is always posted last, it will always complete last. Thus
  70  * when it completes, it is guaranteed that all previous Write WRs have
  71  * also completed.
  72  *
  73  * Write WRs are constructed and posted. Each Write segment gets its own
  74  * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
  75  * DMA-unmap the pages under I/O for that Write segment. The Write
  76  * completion handler does not release any pages.
  77  *
  78  * When the Send WR is constructed, it also gets its own svc_rdma_send_ctxt.
  79  * The ownership of all of the Reply's pages are transferred into that
  80  * ctxt, the Send WR is posted, and sendto returns.
  81  *
  82  * The svc_rdma_send_ctxt is presented when the Send WR completes. The
  83  * Send completion handler finally releases the Reply's pages.
  84  *
  85  * This mechanism also assumes that completions on the transport's Send
  86  * Completion Queue do not run in parallel. Otherwise a Write completion
  87  * and Send completion running at the same time could release pages that
  88  * are still DMA-mapped.
  89  *
  90  * Error Handling
  91  *
  92  * - If the Send WR is posted successfully, it will either complete
  93  *   successfully, or get flushed. Either way, the Send completion
  94  *   handler releases the Reply's pages.
  95  * - If the Send WR cannot be not posted, the forward path releases
  96  *   the Reply's pages.
  97  *
  98  * This handles the case, without the use of page reference counting,
  99  * where two different Write segments send portions of the same page.
 100  */
 101 
 102 #include <linux/spinlock.h>
 103 #include <asm/unaligned.h>
 104 
 105 #include <rdma/ib_verbs.h>
 106 #include <rdma/rdma_cm.h>
 107 
 108 #include <linux/sunrpc/debug.h>
 109 #include <linux/sunrpc/rpc_rdma.h>
 110 #include <linux/sunrpc/svc_rdma.h>
 111 
 112 #include "xprt_rdma.h"
 113 #include <trace/events/rpcrdma.h>
 114 
 115 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
 116 
 117 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc);
 118 
 119 static inline struct svc_rdma_send_ctxt *
 120 svc_rdma_next_send_ctxt(struct list_head *list)
 121 {
 122         return list_first_entry_or_null(list, struct svc_rdma_send_ctxt,
 123                                         sc_list);
 124 }
 125 
 126 static struct svc_rdma_send_ctxt *
 127 svc_rdma_send_ctxt_alloc(struct svcxprt_rdma *rdma)
 128 {
 129         struct svc_rdma_send_ctxt *ctxt;
 130         dma_addr_t addr;
 131         void *buffer;
 132         size_t size;
 133         int i;
 134 
 135         size = sizeof(*ctxt);
 136         size += rdma->sc_max_send_sges * sizeof(struct ib_sge);
 137         ctxt = kmalloc(size, GFP_KERNEL);
 138         if (!ctxt)
 139                 goto fail0;
 140         buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
 141         if (!buffer)
 142                 goto fail1;
 143         addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
 144                                  rdma->sc_max_req_size, DMA_TO_DEVICE);
 145         if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
 146                 goto fail2;
 147 
 148         ctxt->sc_send_wr.next = NULL;
 149         ctxt->sc_send_wr.wr_cqe = &ctxt->sc_cqe;
 150         ctxt->sc_send_wr.sg_list = ctxt->sc_sges;
 151         ctxt->sc_send_wr.send_flags = IB_SEND_SIGNALED;
 152         ctxt->sc_cqe.done = svc_rdma_wc_send;
 153         ctxt->sc_xprt_buf = buffer;
 154         ctxt->sc_sges[0].addr = addr;
 155 
 156         for (i = 0; i < rdma->sc_max_send_sges; i++)
 157                 ctxt->sc_sges[i].lkey = rdma->sc_pd->local_dma_lkey;
 158         return ctxt;
 159 
 160 fail2:
 161         kfree(buffer);
 162 fail1:
 163         kfree(ctxt);
 164 fail0:
 165         return NULL;
 166 }
 167 
 168 /**
 169  * svc_rdma_send_ctxts_destroy - Release all send_ctxt's for an xprt
 170  * @rdma: svcxprt_rdma being torn down
 171  *
 172  */
 173 void svc_rdma_send_ctxts_destroy(struct svcxprt_rdma *rdma)
 174 {
 175         struct svc_rdma_send_ctxt *ctxt;
 176 
 177         while ((ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts))) {
 178                 list_del(&ctxt->sc_list);
 179                 ib_dma_unmap_single(rdma->sc_pd->device,
 180                                     ctxt->sc_sges[0].addr,
 181                                     rdma->sc_max_req_size,
 182                                     DMA_TO_DEVICE);
 183                 kfree(ctxt->sc_xprt_buf);
 184                 kfree(ctxt);
 185         }
 186 }
 187 
 188 /**
 189  * svc_rdma_send_ctxt_get - Get a free send_ctxt
 190  * @rdma: controlling svcxprt_rdma
 191  *
 192  * Returns a ready-to-use send_ctxt, or NULL if none are
 193  * available and a fresh one cannot be allocated.
 194  */
 195 struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma)
 196 {
 197         struct svc_rdma_send_ctxt *ctxt;
 198 
 199         spin_lock(&rdma->sc_send_lock);
 200         ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts);
 201         if (!ctxt)
 202                 goto out_empty;
 203         list_del(&ctxt->sc_list);
 204         spin_unlock(&rdma->sc_send_lock);
 205 
 206 out:
 207         ctxt->sc_send_wr.num_sge = 0;
 208         ctxt->sc_cur_sge_no = 0;
 209         ctxt->sc_page_count = 0;
 210         return ctxt;
 211 
 212 out_empty:
 213         spin_unlock(&rdma->sc_send_lock);
 214         ctxt = svc_rdma_send_ctxt_alloc(rdma);
 215         if (!ctxt)
 216                 return NULL;
 217         goto out;
 218 }
 219 
 220 /**
 221  * svc_rdma_send_ctxt_put - Return send_ctxt to free list
 222  * @rdma: controlling svcxprt_rdma
 223  * @ctxt: object to return to the free list
 224  *
 225  * Pages left in sc_pages are DMA unmapped and released.
 226  */
 227 void svc_rdma_send_ctxt_put(struct svcxprt_rdma *rdma,
 228                             struct svc_rdma_send_ctxt *ctxt)
 229 {
 230         struct ib_device *device = rdma->sc_cm_id->device;
 231         unsigned int i;
 232 
 233         /* The first SGE contains the transport header, which
 234          * remains mapped until @ctxt is destroyed.
 235          */
 236         for (i = 1; i < ctxt->sc_send_wr.num_sge; i++)
 237                 ib_dma_unmap_page(device,
 238                                   ctxt->sc_sges[i].addr,
 239                                   ctxt->sc_sges[i].length,
 240                                   DMA_TO_DEVICE);
 241 
 242         for (i = 0; i < ctxt->sc_page_count; ++i)
 243                 put_page(ctxt->sc_pages[i]);
 244 
 245         spin_lock(&rdma->sc_send_lock);
 246         list_add(&ctxt->sc_list, &rdma->sc_send_ctxts);
 247         spin_unlock(&rdma->sc_send_lock);
 248 }
 249 
 250 /**
 251  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
 252  * @cq: Completion Queue context
 253  * @wc: Work Completion object
 254  *
 255  * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
 256  * the Send completion handler could be running.
 257  */
 258 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 259 {
 260         struct svcxprt_rdma *rdma = cq->cq_context;
 261         struct ib_cqe *cqe = wc->wr_cqe;
 262         struct svc_rdma_send_ctxt *ctxt;
 263 
 264         trace_svcrdma_wc_send(wc);
 265 
 266         atomic_inc(&rdma->sc_sq_avail);
 267         wake_up(&rdma->sc_send_wait);
 268 
 269         ctxt = container_of(cqe, struct svc_rdma_send_ctxt, sc_cqe);
 270         svc_rdma_send_ctxt_put(rdma, ctxt);
 271 
 272         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 273                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
 274                 svc_xprt_enqueue(&rdma->sc_xprt);
 275         }
 276 
 277         svc_xprt_put(&rdma->sc_xprt);
 278 }
 279 
 280 /**
 281  * svc_rdma_send - Post a single Send WR
 282  * @rdma: transport on which to post the WR
 283  * @wr: prepared Send WR to post
 284  *
 285  * Returns zero the Send WR was posted successfully. Otherwise, a
 286  * negative errno is returned.
 287  */
 288 int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr)
 289 {
 290         int ret;
 291 
 292         might_sleep();
 293 
 294         /* If the SQ is full, wait until an SQ entry is available */
 295         while (1) {
 296                 if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) {
 297                         atomic_inc(&rdma_stat_sq_starve);
 298                         trace_svcrdma_sq_full(rdma);
 299                         atomic_inc(&rdma->sc_sq_avail);
 300                         wait_event(rdma->sc_send_wait,
 301                                    atomic_read(&rdma->sc_sq_avail) > 1);
 302                         if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
 303                                 return -ENOTCONN;
 304                         trace_svcrdma_sq_retry(rdma);
 305                         continue;
 306                 }
 307 
 308                 svc_xprt_get(&rdma->sc_xprt);
 309                 trace_svcrdma_post_send(wr);
 310                 ret = ib_post_send(rdma->sc_qp, wr, NULL);
 311                 if (ret)
 312                         break;
 313                 return 0;
 314         }
 315 
 316         trace_svcrdma_sq_post_err(rdma, ret);
 317         set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
 318         svc_xprt_put(&rdma->sc_xprt);
 319         wake_up(&rdma->sc_send_wait);
 320         return ret;
 321 }
 322 
 323 static u32 xdr_padsize(u32 len)
 324 {
 325         return (len & 3) ? (4 - (len & 3)) : 0;
 326 }
 327 
 328 /* Returns length of transport header, in bytes.
 329  */
 330 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
 331 {
 332         unsigned int nsegs;
 333         __be32 *p;
 334 
 335         p = rdma_resp;
 336 
 337         /* RPC-over-RDMA V1 replies never have a Read list. */
 338         p += rpcrdma_fixed_maxsz + 1;
 339 
 340         /* Skip Write list. */
 341         while (*p++ != xdr_zero) {
 342                 nsegs = be32_to_cpup(p++);
 343                 p += nsegs * rpcrdma_segment_maxsz;
 344         }
 345 
 346         /* Skip Reply chunk. */
 347         if (*p++ != xdr_zero) {
 348                 nsegs = be32_to_cpup(p++);
 349                 p += nsegs * rpcrdma_segment_maxsz;
 350         }
 351 
 352         return (unsigned long)p - (unsigned long)rdma_resp;
 353 }
 354 
 355 /* One Write chunk is copied from Call transport header to Reply
 356  * transport header. Each segment's length field is updated to
 357  * reflect number of bytes consumed in the segment.
 358  *
 359  * Returns number of segments in this chunk.
 360  */
 361 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
 362                                            unsigned int remaining)
 363 {
 364         unsigned int i, nsegs;
 365         u32 seg_len;
 366 
 367         /* Write list discriminator */
 368         *dst++ = *src++;
 369 
 370         /* number of segments in this chunk */
 371         nsegs = be32_to_cpup(src);
 372         *dst++ = *src++;
 373 
 374         for (i = nsegs; i; i--) {
 375                 /* segment's RDMA handle */
 376                 *dst++ = *src++;
 377 
 378                 /* bytes returned in this segment */
 379                 seg_len = be32_to_cpu(*src);
 380                 if (remaining >= seg_len) {
 381                         /* entire segment was consumed */
 382                         *dst = *src;
 383                         remaining -= seg_len;
 384                 } else {
 385                         /* segment only partly filled */
 386                         *dst = cpu_to_be32(remaining);
 387                         remaining = 0;
 388                 }
 389                 dst++; src++;
 390 
 391                 /* segment's RDMA offset */
 392                 *dst++ = *src++;
 393                 *dst++ = *src++;
 394         }
 395 
 396         return nsegs;
 397 }
 398 
 399 /* The client provided a Write list in the Call message. Fill in
 400  * the segments in the first Write chunk in the Reply's transport
 401  * header with the number of bytes consumed in each segment.
 402  * Remaining chunks are returned unused.
 403  *
 404  * Assumptions:
 405  *  - Client has provided only one Write chunk
 406  */
 407 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
 408                                            unsigned int consumed)
 409 {
 410         unsigned int nsegs;
 411         __be32 *p, *q;
 412 
 413         /* RPC-over-RDMA V1 replies never have a Read list. */
 414         p = rdma_resp + rpcrdma_fixed_maxsz + 1;
 415 
 416         q = wr_ch;
 417         while (*q != xdr_zero) {
 418                 nsegs = xdr_encode_write_chunk(p, q, consumed);
 419                 q += 2 + nsegs * rpcrdma_segment_maxsz;
 420                 p += 2 + nsegs * rpcrdma_segment_maxsz;
 421                 consumed = 0;
 422         }
 423 
 424         /* Terminate Write list */
 425         *p++ = xdr_zero;
 426 
 427         /* Reply chunk discriminator; may be replaced later */
 428         *p = xdr_zero;
 429 }
 430 
 431 /* The client provided a Reply chunk in the Call message. Fill in
 432  * the segments in the Reply chunk in the Reply message with the
 433  * number of bytes consumed in each segment.
 434  *
 435  * Assumptions:
 436  * - Reply can always fit in the provided Reply chunk
 437  */
 438 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
 439                                             unsigned int consumed)
 440 {
 441         __be32 *p;
 442 
 443         /* Find the Reply chunk in the Reply's xprt header.
 444          * RPC-over-RDMA V1 replies never have a Read list.
 445          */
 446         p = rdma_resp + rpcrdma_fixed_maxsz + 1;
 447 
 448         /* Skip past Write list */
 449         while (*p++ != xdr_zero)
 450                 p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
 451 
 452         xdr_encode_write_chunk(p, rp_ch, consumed);
 453 }
 454 
 455 /* Parse the RPC Call's transport header.
 456  */
 457 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
 458                                       __be32 **write, __be32 **reply)
 459 {
 460         __be32 *p;
 461 
 462         p = rdma_argp + rpcrdma_fixed_maxsz;
 463 
 464         /* Read list */
 465         while (*p++ != xdr_zero)
 466                 p += 5;
 467 
 468         /* Write list */
 469         if (*p != xdr_zero) {
 470                 *write = p;
 471                 while (*p++ != xdr_zero)
 472                         p += 1 + be32_to_cpu(*p) * 4;
 473         } else {
 474                 *write = NULL;
 475                 p++;
 476         }
 477 
 478         /* Reply chunk */
 479         if (*p != xdr_zero)
 480                 *reply = p;
 481         else
 482                 *reply = NULL;
 483 }
 484 
 485 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
 486                                  struct svc_rdma_send_ctxt *ctxt,
 487                                  struct page *page,
 488                                  unsigned long offset,
 489                                  unsigned int len)
 490 {
 491         struct ib_device *dev = rdma->sc_cm_id->device;
 492         dma_addr_t dma_addr;
 493 
 494         dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
 495         if (ib_dma_mapping_error(dev, dma_addr))
 496                 goto out_maperr;
 497 
 498         ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr;
 499         ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len;
 500         ctxt->sc_send_wr.num_sge++;
 501         return 0;
 502 
 503 out_maperr:
 504         trace_svcrdma_dma_map_page(rdma, page);
 505         return -EIO;
 506 }
 507 
 508 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
 509  * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively.
 510  */
 511 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
 512                                 struct svc_rdma_send_ctxt *ctxt,
 513                                 unsigned char *base,
 514                                 unsigned int len)
 515 {
 516         return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base),
 517                                      offset_in_page(base), len);
 518 }
 519 
 520 /**
 521  * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer
 522  * @rdma: controlling transport
 523  * @ctxt: send_ctxt for the Send WR
 524  * @len: length of transport header
 525  *
 526  */
 527 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma,
 528                              struct svc_rdma_send_ctxt *ctxt,
 529                              unsigned int len)
 530 {
 531         ctxt->sc_sges[0].length = len;
 532         ctxt->sc_send_wr.num_sge++;
 533         ib_dma_sync_single_for_device(rdma->sc_pd->device,
 534                                       ctxt->sc_sges[0].addr, len,
 535                                       DMA_TO_DEVICE);
 536 }
 537 
 538 /* If the xdr_buf has more elements than the device can
 539  * transmit in a single RDMA Send, then the reply will
 540  * have to be copied into a bounce buffer.
 541  */
 542 static bool svc_rdma_pull_up_needed(struct svcxprt_rdma *rdma,
 543                                     struct xdr_buf *xdr,
 544                                     __be32 *wr_lst)
 545 {
 546         int elements;
 547 
 548         /* xdr->head */
 549         elements = 1;
 550 
 551         /* xdr->pages */
 552         if (!wr_lst) {
 553                 unsigned int remaining;
 554                 unsigned long pageoff;
 555 
 556                 pageoff = xdr->page_base & ~PAGE_MASK;
 557                 remaining = xdr->page_len;
 558                 while (remaining) {
 559                         ++elements;
 560                         remaining -= min_t(u32, PAGE_SIZE - pageoff,
 561                                            remaining);
 562                         pageoff = 0;
 563                 }
 564         }
 565 
 566         /* xdr->tail */
 567         if (xdr->tail[0].iov_len)
 568                 ++elements;
 569 
 570         /* assume 1 SGE is needed for the transport header */
 571         return elements >= rdma->sc_max_send_sges;
 572 }
 573 
 574 /* The device is not capable of sending the reply directly.
 575  * Assemble the elements of @xdr into the transport header
 576  * buffer.
 577  */
 578 static int svc_rdma_pull_up_reply_msg(struct svcxprt_rdma *rdma,
 579                                       struct svc_rdma_send_ctxt *ctxt,
 580                                       struct xdr_buf *xdr, __be32 *wr_lst)
 581 {
 582         unsigned char *dst, *tailbase;
 583         unsigned int taillen;
 584 
 585         dst = ctxt->sc_xprt_buf;
 586         dst += ctxt->sc_sges[0].length;
 587 
 588         memcpy(dst, xdr->head[0].iov_base, xdr->head[0].iov_len);
 589         dst += xdr->head[0].iov_len;
 590 
 591         tailbase = xdr->tail[0].iov_base;
 592         taillen = xdr->tail[0].iov_len;
 593         if (wr_lst) {
 594                 u32 xdrpad;
 595 
 596                 xdrpad = xdr_padsize(xdr->page_len);
 597                 if (taillen && xdrpad) {
 598                         tailbase += xdrpad;
 599                         taillen -= xdrpad;
 600                 }
 601         } else {
 602                 unsigned int len, remaining;
 603                 unsigned long pageoff;
 604                 struct page **ppages;
 605 
 606                 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
 607                 pageoff = xdr->page_base & ~PAGE_MASK;
 608                 remaining = xdr->page_len;
 609                 while (remaining) {
 610                         len = min_t(u32, PAGE_SIZE - pageoff, remaining);
 611 
 612                         memcpy(dst, page_address(*ppages), len);
 613                         remaining -= len;
 614                         dst += len;
 615                         pageoff = 0;
 616                 }
 617         }
 618 
 619         if (taillen)
 620                 memcpy(dst, tailbase, taillen);
 621 
 622         ctxt->sc_sges[0].length += xdr->len;
 623         ib_dma_sync_single_for_device(rdma->sc_pd->device,
 624                                       ctxt->sc_sges[0].addr,
 625                                       ctxt->sc_sges[0].length,
 626                                       DMA_TO_DEVICE);
 627 
 628         return 0;
 629 }
 630 
 631 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message
 632  * @rdma: controlling transport
 633  * @ctxt: send_ctxt for the Send WR
 634  * @xdr: prepared xdr_buf containing RPC message
 635  * @wr_lst: pointer to Call header's Write list, or NULL
 636  *
 637  * Load the xdr_buf into the ctxt's sge array, and DMA map each
 638  * element as it is added.
 639  *
 640  * Returns zero on success, or a negative errno on failure.
 641  */
 642 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
 643                            struct svc_rdma_send_ctxt *ctxt,
 644                            struct xdr_buf *xdr, __be32 *wr_lst)
 645 {
 646         unsigned int len, remaining;
 647         unsigned long page_off;
 648         struct page **ppages;
 649         unsigned char *base;
 650         u32 xdr_pad;
 651         int ret;
 652 
 653         if (svc_rdma_pull_up_needed(rdma, xdr, wr_lst))
 654                 return svc_rdma_pull_up_reply_msg(rdma, ctxt, xdr, wr_lst);
 655 
 656         ++ctxt->sc_cur_sge_no;
 657         ret = svc_rdma_dma_map_buf(rdma, ctxt,
 658                                    xdr->head[0].iov_base,
 659                                    xdr->head[0].iov_len);
 660         if (ret < 0)
 661                 return ret;
 662 
 663         /* If a Write chunk is present, the xdr_buf's page list
 664          * is not included inline. However the Upper Layer may
 665          * have added XDR padding in the tail buffer, and that
 666          * should not be included inline.
 667          */
 668         if (wr_lst) {
 669                 base = xdr->tail[0].iov_base;
 670                 len = xdr->tail[0].iov_len;
 671                 xdr_pad = xdr_padsize(xdr->page_len);
 672 
 673                 if (len && xdr_pad) {
 674                         base += xdr_pad;
 675                         len -= xdr_pad;
 676                 }
 677 
 678                 goto tail;
 679         }
 680 
 681         ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
 682         page_off = xdr->page_base & ~PAGE_MASK;
 683         remaining = xdr->page_len;
 684         while (remaining) {
 685                 len = min_t(u32, PAGE_SIZE - page_off, remaining);
 686 
 687                 ++ctxt->sc_cur_sge_no;
 688                 ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++,
 689                                             page_off, len);
 690                 if (ret < 0)
 691                         return ret;
 692 
 693                 remaining -= len;
 694                 page_off = 0;
 695         }
 696 
 697         base = xdr->tail[0].iov_base;
 698         len = xdr->tail[0].iov_len;
 699 tail:
 700         if (len) {
 701                 ++ctxt->sc_cur_sge_no;
 702                 ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len);
 703                 if (ret < 0)
 704                         return ret;
 705         }
 706 
 707         return 0;
 708 }
 709 
 710 /* The svc_rqst and all resources it owns are released as soon as
 711  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
 712  * so they are released by the Send completion handler.
 713  */
 714 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
 715                                    struct svc_rdma_send_ctxt *ctxt)
 716 {
 717         int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
 718 
 719         ctxt->sc_page_count += pages;
 720         for (i = 0; i < pages; i++) {
 721                 ctxt->sc_pages[i] = rqstp->rq_respages[i];
 722                 rqstp->rq_respages[i] = NULL;
 723         }
 724 
 725         /* Prevent svc_xprt_release from releasing pages in rq_pages */
 726         rqstp->rq_next_page = rqstp->rq_respages;
 727 }
 728 
 729 /* Prepare the portion of the RPC Reply that will be transmitted
 730  * via RDMA Send. The RPC-over-RDMA transport header is prepared
 731  * in sc_sges[0], and the RPC xdr_buf is prepared in following sges.
 732  *
 733  * Depending on whether a Write list or Reply chunk is present,
 734  * the server may send all, a portion of, or none of the xdr_buf.
 735  * In the latter case, only the transport header (sc_sges[0]) is
 736  * transmitted.
 737  *
 738  * RDMA Send is the last step of transmitting an RPC reply. Pages
 739  * involved in the earlier RDMA Writes are here transferred out
 740  * of the rqstp and into the sctxt's page array. These pages are
 741  * DMA unmapped by each Write completion, but the subsequent Send
 742  * completion finally releases these pages.
 743  *
 744  * Assumptions:
 745  * - The Reply's transport header will never be larger than a page.
 746  */
 747 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
 748                                    struct svc_rdma_send_ctxt *sctxt,
 749                                    struct svc_rdma_recv_ctxt *rctxt,
 750                                    struct svc_rqst *rqstp,
 751                                    __be32 *wr_lst, __be32 *rp_ch)
 752 {
 753         int ret;
 754 
 755         if (!rp_ch) {
 756                 ret = svc_rdma_map_reply_msg(rdma, sctxt,
 757                                              &rqstp->rq_res, wr_lst);
 758                 if (ret < 0)
 759                         return ret;
 760         }
 761 
 762         svc_rdma_save_io_pages(rqstp, sctxt);
 763 
 764         if (rctxt->rc_inv_rkey) {
 765                 sctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV;
 766                 sctxt->sc_send_wr.ex.invalidate_rkey = rctxt->rc_inv_rkey;
 767         } else {
 768                 sctxt->sc_send_wr.opcode = IB_WR_SEND;
 769         }
 770         dprintk("svcrdma: posting Send WR with %u sge(s)\n",
 771                 sctxt->sc_send_wr.num_sge);
 772         return svc_rdma_send(rdma, &sctxt->sc_send_wr);
 773 }
 774 
 775 /* Given the client-provided Write and Reply chunks, the server was not
 776  * able to form a complete reply. Return an RDMA_ERROR message so the
 777  * client can retire this RPC transaction. As above, the Send completion
 778  * routine releases payload pages that were part of a previous RDMA Write.
 779  *
 780  * Remote Invalidation is skipped for simplicity.
 781  */
 782 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
 783                                    struct svc_rdma_send_ctxt *ctxt,
 784                                    struct svc_rqst *rqstp)
 785 {
 786         __be32 *p;
 787         int ret;
 788 
 789         p = ctxt->sc_xprt_buf;
 790         trace_svcrdma_err_chunk(*p);
 791         p += 3;
 792         *p++ = rdma_error;
 793         *p   = err_chunk;
 794         svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR);
 795 
 796         svc_rdma_save_io_pages(rqstp, ctxt);
 797 
 798         ctxt->sc_send_wr.opcode = IB_WR_SEND;
 799         ret = svc_rdma_send(rdma, &ctxt->sc_send_wr);
 800         if (ret) {
 801                 svc_rdma_send_ctxt_put(rdma, ctxt);
 802                 return ret;
 803         }
 804 
 805         return 0;
 806 }
 807 
 808 /**
 809  * svc_rdma_sendto - Transmit an RPC reply
 810  * @rqstp: processed RPC request, reply XDR already in ::rq_res
 811  *
 812  * Any resources still associated with @rqstp are released upon return.
 813  * If no reply message was possible, the connection is closed.
 814  *
 815  * Returns:
 816  *      %0 if an RPC reply has been successfully posted,
 817  *      %-ENOMEM if a resource shortage occurred (connection is lost),
 818  *      %-ENOTCONN if posting failed (connection is lost).
 819  */
 820 int svc_rdma_sendto(struct svc_rqst *rqstp)
 821 {
 822         struct svc_xprt *xprt = rqstp->rq_xprt;
 823         struct svcxprt_rdma *rdma =
 824                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 825         struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
 826         __be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
 827         struct xdr_buf *xdr = &rqstp->rq_res;
 828         struct svc_rdma_send_ctxt *sctxt;
 829         int ret;
 830 
 831         rdma_argp = rctxt->rc_recv_buf;
 832         svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
 833 
 834         /* Create the RDMA response header. xprt->xpt_mutex,
 835          * acquired in svc_send(), serializes RPC replies. The
 836          * code path below that inserts the credit grant value
 837          * into each transport header runs only inside this
 838          * critical section.
 839          */
 840         ret = -ENOMEM;
 841         sctxt = svc_rdma_send_ctxt_get(rdma);
 842         if (!sctxt)
 843                 goto err0;
 844         rdma_resp = sctxt->sc_xprt_buf;
 845 
 846         p = rdma_resp;
 847         *p++ = *rdma_argp;
 848         *p++ = *(rdma_argp + 1);
 849         *p++ = rdma->sc_fc_credits;
 850         *p++ = rp_ch ? rdma_nomsg : rdma_msg;
 851 
 852         /* Start with empty chunks */
 853         *p++ = xdr_zero;
 854         *p++ = xdr_zero;
 855         *p   = xdr_zero;
 856 
 857         if (wr_lst) {
 858                 /* XXX: Presume the client sent only one Write chunk */
 859                 ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
 860                 if (ret < 0)
 861                         goto err2;
 862                 svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
 863         }
 864         if (rp_ch) {
 865                 ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
 866                 if (ret < 0)
 867                         goto err2;
 868                 svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
 869         }
 870 
 871         svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp));
 872         ret = svc_rdma_send_reply_msg(rdma, sctxt, rctxt, rqstp,
 873                                       wr_lst, rp_ch);
 874         if (ret < 0)
 875                 goto err1;
 876         return 0;
 877 
 878  err2:
 879         if (ret != -E2BIG && ret != -EINVAL)
 880                 goto err1;
 881 
 882         ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp);
 883         if (ret < 0)
 884                 goto err1;
 885         return 0;
 886 
 887  err1:
 888         svc_rdma_send_ctxt_put(rdma, sctxt);
 889  err0:
 890         trace_svcrdma_send_failed(rqstp, ret);
 891         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 892         return -ENOTCONN;
 893 }

/* [<][>][^][v][top][bottom][index][help] */