fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@suse.de>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/pipe_fs_i.h>
  24 #include <linux/mm_inline.h>
  25 #include <linux/swap.h>
  26 #include <linux/writeback.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/module.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31
  32 struct partial_page {
  33         unsigned int offset;
  34         unsigned int len;
  35 };
  36
  37 /*
  38  * Passed to splice_to_pipe
  39  */
  40 struct splice_pipe_desc {
  41         struct page **pages;            /* page map */
  42         struct partial_page *partial;   /* pages[] may not be contig */
  43         int nr_pages;                   /* number of pages in map */
  44         unsigned int flags;             /* splice flags */
  45         struct pipe_buf_operations *ops;/* ops associated with output pipe */
  46 };
  47
  48 /*
  49  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  50  * a vm helper function, it's already simplified quite a bit by the
  51  * addition of remove_mapping(). If success is returned, the caller may
  52  * attempt to reuse this page for another destination.
  53  */
  54 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  55                                      struct pipe_buffer *buf)
  56 {
  57         struct page *page = buf->page;
  58         struct address_space *mapping = page_mapping(page);
  59
  60         lock_page(page);
  61
  62         WARN_ON(!PageUptodate(page));
  63
  64         /*
  65          * At least for ext2 with nobh option, we need to wait on writeback
  66          * completing on this page, since we'll remove it from the pagecache.
  67          * Otherwise truncate wont wait on the page, allowing the disk
  68          * blocks to be reused by someone else before we actually wrote our
  69          * data to them. fs corruption ensues.
  70          */
  71         wait_on_page_writeback(page);
  72
  73         if (PagePrivate(page))
  74                 try_to_release_page(page, mapping_gfp_mask(mapping));
  75
  76         if (!remove_mapping(mapping, page)) {
  77                 unlock_page(page);
  78                 return 1;
  79         }
  80
  81         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
  82         return 0;
  83 }
  84
  85 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  86                                         struct pipe_buffer *buf)
  87 {
  88         page_cache_release(buf->page);
  89         buf->page = NULL;
  90         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
  91 }
  92
  93 static void *page_cache_pipe_buf_map(struct file *file,
  94                                      struct pipe_inode_info *info,
  95                                      struct pipe_buffer *buf)
  96 {
  97         struct page *page = buf->page;
  98         int err;
  99
 100         if (!PageUptodate(page)) {
 101                 lock_page(page);
 102
 103                 /*
 104                  * Page got truncated/unhashed. This will cause a 0-byte
 105                  * splice, if this is the first page.
 106                  */
 107                 if (!page->mapping) {
 108                         err = -ENODATA;
 109                         goto error;
 110                 }
 111
 112                 /*
 113                  * Uh oh, read-error from disk.
 114                  */
 115                 if (!PageUptodate(page)) {
 116                         err = -EIO;
 117                         goto error;
 118                 }
 119
 120                 /*
 121                  * Page is ok afterall, fall through to mapping.
 122                  */
 123                 unlock_page(page);
 124         }
 125
 126         return kmap(page);
 127 error:
 128         unlock_page(page);
 129         return ERR_PTR(err);
 130 }
 131
 132 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
 133                                       struct pipe_buffer *buf)
 134 {
 135         kunmap(buf->page);
 136 }
 137
 138 static void *user_page_pipe_buf_map(struct file *file,
 139                                     struct pipe_inode_info *pipe,
 140                                     struct pipe_buffer *buf)
 141 {
 142         return kmap(buf->page);
 143 }
 144
 145 static void user_page_pipe_buf_unmap(struct pipe_inode_info *pipe,
 146                                      struct pipe_buffer *buf)
 147 {
 148         kunmap(buf->page);
 149 }
 150
 151 static void page_cache_pipe_buf_get(struct pipe_inode_info *info,
 152                                     struct pipe_buffer *buf)
 153 {
 154         page_cache_get(buf->page);
 155 }
 156
 157 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 158         .can_merge = 0,
 159         .map = page_cache_pipe_buf_map,
 160         .unmap = page_cache_pipe_buf_unmap,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = page_cache_pipe_buf_steal,
 163         .get = page_cache_pipe_buf_get,
 164 };
 165
 166 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 167                                     struct pipe_buffer *buf)
 168 {
 169         return 1;
 170 }
 171
 172 static struct pipe_buf_operations user_page_pipe_buf_ops = {
 173         .can_merge = 0,
 174         .map = user_page_pipe_buf_map,
 175         .unmap = user_page_pipe_buf_unmap,
 176         .release = page_cache_pipe_buf_release,
 177         .steal = user_page_pipe_buf_steal,
 178         .get = page_cache_pipe_buf_get,
 179 };
 180
 181 /*
 182  * Pipe output worker. This sets up our pipe format with the page cache
 183  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 184  */
 185 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                               struct splice_pipe_desc *spd)
 187 {
 188         int ret, do_wakeup, page_nr;
 189
 190         ret = 0;
 191         do_wakeup = 0;
 192         page_nr = 0;
 193
 194         if (pipe->inode)
 195                 mutex_lock(&pipe->inode->i_mutex);
 196
 197         for (;;) {
 198                 if (!pipe->readers) {
 199                         send_sig(SIGPIPE, current, 0);
 200                         if (!ret)
 201                                 ret = -EPIPE;
 202                         break;
 203                 }
 204
 205                 if (pipe->nrbufs < PIPE_BUFFERS) {
 206                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 207                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 208
 209                         buf->page = spd->pages[page_nr];
 210                         buf->offset = spd->partial[page_nr].offset;
 211                         buf->len = spd->partial[page_nr].len;
 212                         buf->ops = spd->ops;
 213                         pipe->nrbufs++;
 214                         page_nr++;
 215                         ret += buf->len;
 216
 217                         if (pipe->inode)
 218                                 do_wakeup = 1;
 219
 220                         if (!--spd->nr_pages)
 221                                 break;
 222                         if (pipe->nrbufs < PIPE_BUFFERS)
 223                                 continue;
 224
 225                         break;
 226                 }
 227
 228                 if (spd->flags & SPLICE_F_NONBLOCK) {
 229                         if (!ret)
 230                                 ret = -EAGAIN;
 231                         break;
 232                 }
 233
 234                 if (signal_pending(current)) {
 235                         if (!ret)
 236                                 ret = -ERESTARTSYS;
 237                         break;
 238                 }
 239
 240                 if (do_wakeup) {
 241                         smp_mb();
 242                         if (waitqueue_active(&pipe->wait))
 243                                 wake_up_interruptible_sync(&pipe->wait);
 244                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 245                         do_wakeup = 0;
 246                 }
 247
 248                 pipe->waiting_writers++;
 249                 pipe_wait(pipe);
 250                 pipe->waiting_writers--;
 251         }
 252
 253         if (pipe->inode)
 254                 mutex_unlock(&pipe->inode->i_mutex);
 255
 256         if (do_wakeup) {
 257                 smp_mb();
 258                 if (waitqueue_active(&pipe->wait))
 259                         wake_up_interruptible(&pipe->wait);
 260                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 261         }
 262
 263         while (page_nr < spd->nr_pages)
 264                 page_cache_release(spd->pages[page_nr++]);
 265
 266         return ret;
 267 }
 268
 269 static int
 270 __generic_file_splice_read(struct file *in, loff_t *ppos,
 271                            struct pipe_inode_info *pipe, size_t len,
 272                            unsigned int flags)
 273 {
 274         struct address_space *mapping = in->f_mapping;
 275         unsigned int loff, nr_pages;
 276         struct page *pages[PIPE_BUFFERS];
 277         struct partial_page partial[PIPE_BUFFERS];
 278         struct page *page;
 279         pgoff_t index, end_index;
 280         loff_t isize;
 281         size_t total_len;
 282         int error, page_nr;
 283         struct splice_pipe_desc spd = {
 284                 .pages = pages,
 285                 .partial = partial,
 286                 .flags = flags,
 287                 .ops = &page_cache_pipe_buf_ops,
 288         };
 289
 290         index = *ppos >> PAGE_CACHE_SHIFT;
 291         loff = *ppos & ~PAGE_CACHE_MASK;
 292         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 293
 294         if (nr_pages > PIPE_BUFFERS)
 295                 nr_pages = PIPE_BUFFERS;
 296
 297         /*
 298          * Initiate read-ahead on this page range. however, don't call into
 299          * read-ahead if this is a non-zero offset (we are likely doing small
 300          * chunk splice and the page is already there) for a single page.
 301          */
 302         if (!loff || nr_pages > 1)
 303                 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 304
 305         /*
 306          * Now fill in the holes:
 307          */
 308         error = 0;
 309         total_len = 0;
 310
 311         /*
 312          * Lookup the (hopefully) full range of pages we need.
 313          */
 314         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 315
 316         /*
 317          * If find_get_pages_contig() returned fewer pages than we needed,
 318          * allocate the rest.
 319          */
 320         index += spd.nr_pages;
 321         while (spd.nr_pages < nr_pages) {
 322                 /*
 323                  * Page could be there, find_get_pages_contig() breaks on
 324                  * the first hole.
 325                  */
 326                 page = find_get_page(mapping, index);
 327                 if (!page) {
 328                         /*
 329                          * page didn't exist, allocate one.
 330                          */
 331                         page = page_cache_alloc_cold(mapping);
 332                         if (!page)
 333                                 break;
 334
 335                         error = add_to_page_cache_lru(page, mapping, index,
 336                                               mapping_gfp_mask(mapping));
 337                         if (unlikely(error)) {
 338                                 page_cache_release(page);
 339                                 break;
 340                         }
 341                         /*
 342                          * add_to_page_cache() locks the page, unlock it
 343                          * to avoid convoluting the logic below even more.
 344                          */
 345                         unlock_page(page);
 346                 }
 347
 348                 pages[spd.nr_pages++] = page;
 349                 index++;
 350         }
 351
 352         /*
 353          * Now loop over the map and see if we need to start IO on any
 354          * pages, fill in the partial map, etc.
 355          */
 356         index = *ppos >> PAGE_CACHE_SHIFT;
 357         nr_pages = spd.nr_pages;
 358         spd.nr_pages = 0;
 359         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 360                 unsigned int this_len;
 361
 362                 if (!len)
 363                         break;
 364
 365                 /*
 366                  * this_len is the max we'll use from this page
 367                  */
 368                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 369                 page = pages[page_nr];
 370
 371                 /*
 372                  * If the page isn't uptodate, we may need to start io on it
 373                  */
 374                 if (!PageUptodate(page)) {
 375                         /*
 376                          * If in nonblock mode then dont block on waiting
 377                          * for an in-flight io page
 378                          */
 379                         if (flags & SPLICE_F_NONBLOCK)
 380                                 break;
 381
 382                         lock_page(page);
 383
 384                         /*
 385                          * page was truncated, stop here. if this isn't the
 386                          * first page, we'll just complete what we already
 387                          * added
 388                          */
 389                         if (!page->mapping) {
 390                                 unlock_page(page);
 391                                 break;
 392                         }
 393                         /*
 394                          * page was already under io and is now done, great
 395                          */
 396                         if (PageUptodate(page)) {
 397                                 unlock_page(page);
 398                                 goto fill_it;
 399                         }
 400
 401                         /*
 402                          * need to read in the page
 403                          */
 404                         error = mapping->a_ops->readpage(in, page);
 405                         if (unlikely(error)) {
 406                                 /*
 407                                  * We really should re-lookup the page here,
 408                                  * but it complicates things a lot. Instead
 409                                  * lets just do what we already stored, and
 410                                  * we'll get it the next time we are called.
 411                                  */
 412                                 if (error == AOP_TRUNCATED_PAGE)
 413                                         error = 0;
 414
 415                                 break;
 416                         }
 417
 418                         /*
 419                          * i_size must be checked after ->readpage().
 420                          */
 421                         isize = i_size_read(mapping->host);
 422                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 423                         if (unlikely(!isize || index > end_index))
 424                                 break;
 425
 426                         /*
 427                          * if this is the last page, see if we need to shrink
 428                          * the length and stop
 429                          */
 430                         if (end_index == index) {
 431                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 432                                 if (total_len + loff > isize)
 433                                         break;
 434                                 /*
 435                                  * force quit after adding this page
 436                                  */
 437                                 len = this_len;
 438                                 this_len = min(this_len, loff);
 439                                 loff = 0;
 440                         }
 441                 }
 442 fill_it:
 443                 partial[page_nr].offset = loff;
 444                 partial[page_nr].len = this_len;
 445                 len -= this_len;
 446                 total_len += this_len;
 447                 loff = 0;
 448                 spd.nr_pages++;
 449                 index++;
 450         }
 451
 452         /*
 453          * Release any pages at the end, if we quit early. 'i' is how far
 454          * we got, 'nr_pages' is how many pages are in the map.
 455          */
 456         while (page_nr < nr_pages)
 457                 page_cache_release(pages[page_nr++]);
 458
 459         if (spd.nr_pages)
 460                 return splice_to_pipe(pipe, &spd);
 461
 462         return error;
 463 }
 464
 465 /**
 466  * generic_file_splice_read - splice data from file to a pipe
 467  * @in:         file to splice from
 468  * @pipe:       pipe to splice to
 469  * @len:        number of bytes to splice
 470  * @flags:      splice modifier flags
 471  *
 472  * Will read pages from given file and fill them into a pipe.
 473  */
 474 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 475                                  struct pipe_inode_info *pipe, size_t len,
 476                                  unsigned int flags)
 477 {
 478         ssize_t spliced;
 479         int ret;
 480
 481         ret = 0;
 482         spliced = 0;
 483
 484         while (len) {
 485                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 486
 487                 if (ret < 0)
 488                         break;
 489                 else if (!ret) {
 490                         if (spliced)
 491                                 break;
 492                         if (flags & SPLICE_F_NONBLOCK) {
 493                                 ret = -EAGAIN;
 494                                 break;
 495                         }
 496                 }
 497
 498                 *ppos += ret;
 499                 len -= ret;
 500                 spliced += ret;
 501         }
 502
 503         if (spliced)
 504                 return spliced;
 505
 506         return ret;
 507 }
 508
 509 EXPORT_SYMBOL(generic_file_splice_read);
 510
 511 /*
 512  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 513  * using sendpage(). Return the number of bytes sent.
 514  */
 515 static int pipe_to_sendpage(struct pipe_inode_info *info,
 516                             struct pipe_buffer *buf, struct splice_desc *sd)
 517 {
 518         struct file *file = sd->file;
 519         loff_t pos = sd->pos;
 520         ssize_t ret;
 521         void *ptr;
 522         int more;
 523
 524         /*
 525          * Sub-optimal, but we are limited by the pipe ->map. We don't
 526          * need a kmap'ed buffer here, we just want to make sure we
 527          * have the page pinned if the pipe page originates from the
 528          * page cache.
 529          */
 530         ptr = buf->ops->map(file, info, buf);
 531         if (IS_ERR(ptr))
 532                 return PTR_ERR(ptr);
 533
 534         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 535
 536         ret = file->f_op->sendpage(file, buf->page, buf->offset, sd->len,
 537                                    &pos, more);
 538
 539         buf->ops->unmap(info, buf);
 540         return ret;
 541 }
 542
 543 /*
 544  * This is a little more tricky than the file -> pipe splicing. There are
 545  * basically three cases:
 546  *
 547  *      - Destination page already exists in the address space and there
 548  *        are users of it. For that case we have no other option that
 549  *        copying the data. Tough luck.
 550  *      - Destination page already exists in the address space, but there
 551  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 552  *        through to last case.
 553  *      - Destination page does not exist, we can add the pipe page to
 554  *        the page cache and avoid the copy.
 555  *
 556  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 557  * sd->flags), we attempt to migrate pages from the pipe to the output
 558  * file address space page cache. This is possible if no one else has
 559  * the pipe page referenced outside of the pipe and page cache. If
 560  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 561  * a new page in the output file page cache and fill/dirty that.
 562  */
 563 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 564                         struct splice_desc *sd)
 565 {
 566         struct file *file = sd->file;
 567         struct address_space *mapping = file->f_mapping;
 568         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 569         unsigned int offset, this_len;
 570         struct page *page;
 571         pgoff_t index;
 572         char *src;
 573         int ret;
 574
 575         /*
 576          * make sure the data in this buffer is uptodate
 577          */
 578         src = buf->ops->map(file, info, buf);
 579         if (IS_ERR(src))
 580                 return PTR_ERR(src);
 581
 582         index = sd->pos >> PAGE_CACHE_SHIFT;
 583         offset = sd->pos & ~PAGE_CACHE_MASK;
 584
 585         this_len = sd->len;
 586         if (this_len + offset > PAGE_CACHE_SIZE)
 587                 this_len = PAGE_CACHE_SIZE - offset;
 588
 589         /*
 590          * Reuse buf page, if SPLICE_F_MOVE is set.
 591          */
 592         if (sd->flags & SPLICE_F_MOVE) {
 593                 /*
 594                  * If steal succeeds, buf->page is now pruned from the vm
 595                  * side (LRU and page cache) and we can reuse it. The page
 596                  * will also be looked on successful return.
 597                  */
 598                 if (buf->ops->steal(info, buf))
 599                         goto find_page;
 600
 601                 page = buf->page;
 602                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 603                         goto find_page;
 604
 605                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 606                         lru_cache_add(page);
 607         } else {
 608 find_page:
 609                 page = find_lock_page(mapping, index);
 610                 if (!page) {
 611                         ret = -ENOMEM;
 612                         page = page_cache_alloc_cold(mapping);
 613                         if (unlikely(!page))
 614                                 goto out_nomem;
 615
 616                         /*
 617                          * This will also lock the page
 618                          */
 619                         ret = add_to_page_cache_lru(page, mapping, index,
 620                                                     gfp_mask);
 621                         if (unlikely(ret))
 622                                 goto out;
 623                 }
 624
 625                 /*
 626                  * We get here with the page locked. If the page is also
 627                  * uptodate, we don't need to do more. If it isn't, we
 628                  * may need to bring it in if we are not going to overwrite
 629                  * the full page.
 630                  */
 631                 if (!PageUptodate(page)) {
 632                         if (this_len < PAGE_CACHE_SIZE) {
 633                                 ret = mapping->a_ops->readpage(file, page);
 634                                 if (unlikely(ret))
 635                                         goto out;
 636
 637                                 lock_page(page);
 638
 639                                 if (!PageUptodate(page)) {
 640                                         /*
 641                                          * Page got invalidated, repeat.
 642                                          */
 643                                         if (!page->mapping) {
 644                                                 unlock_page(page);
 645                                                 page_cache_release(page);
 646                                                 goto find_page;
 647                                         }
 648                                         ret = -EIO;
 649                                         goto out;
 650                                 }
 651                         } else
 652                                 SetPageUptodate(page);
 653                 }
 654         }
 655
 656         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 657         if (ret == AOP_TRUNCATED_PAGE) {
 658                 page_cache_release(page);
 659                 goto find_page;
 660         } else if (ret)
 661                 goto out;
 662
 663         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 664                 char *dst = kmap_atomic(page, KM_USER0);
 665
 666                 memcpy(dst + offset, src + buf->offset, this_len);
 667                 flush_dcache_page(page);
 668                 kunmap_atomic(dst, KM_USER0);
 669         }
 670
 671         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 672         if (ret == AOP_TRUNCATED_PAGE) {
 673                 page_cache_release(page);
 674                 goto find_page;
 675         } else if (ret)
 676                 goto out;
 677
 678         /*
 679          * Return the number of bytes written.
 680          */
 681         ret = this_len;
 682         mark_page_accessed(page);
 683         balance_dirty_pages_ratelimited(mapping);
 684 out:
 685         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN))
 686                 page_cache_release(page);
 687
 688         unlock_page(page);
 689 out_nomem:
 690         buf->ops->unmap(info, buf);
 691         return ret;
 692 }
 693
 694 /*
 695  * Pipe input worker. Most of this logic works like a regular pipe, the
 696  * key here is the 'actor' worker passed in that actually moves the data
 697  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 698  */
 699 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 700                          loff_t *ppos, size_t len, unsigned int flags,
 701                          splice_actor *actor)
 702 {
 703         int ret, do_wakeup, err;
 704         struct splice_desc sd;
 705
 706         ret = 0;
 707         do_wakeup = 0;
 708
 709         sd.total_len = len;
 710         sd.flags = flags;
 711         sd.file = out;
 712         sd.pos = *ppos;
 713
 714         if (pipe->inode)
 715                 mutex_lock(&pipe->inode->i_mutex);
 716
 717         for (;;) {
 718                 if (pipe->nrbufs) {
 719                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 720                         struct pipe_buf_operations *ops = buf->ops;
 721
 722                         sd.len = buf->len;
 723                         if (sd.len > sd.total_len)
 724                                 sd.len = sd.total_len;
 725
 726                         err = actor(pipe, buf, &sd);
 727                         if (err <= 0) {
 728                                 if (!ret && err != -ENODATA)
 729                                         ret = err;
 730
 731                                 break;
 732                         }
 733
 734                         ret += err;
 735                         buf->offset += err;
 736                         buf->len -= err;
 737
 738                         sd.len -= err;
 739                         sd.pos += err;
 740                         sd.total_len -= err;
 741                         if (sd.len)
 742                                 continue;
 743
 744                         if (!buf->len) {
 745                                 buf->ops = NULL;
 746                                 ops->release(pipe, buf);
 747                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 748                                 pipe->nrbufs--;
 749                                 if (pipe->inode)
 750                                         do_wakeup = 1;
 751                         }
 752
 753                         if (!sd.total_len)
 754                                 break;
 755                 }
 756
 757                 if (pipe->nrbufs)
 758                         continue;
 759                 if (!pipe->writers)
 760                         break;
 761                 if (!pipe->waiting_writers) {
 762                         if (ret)
 763                                 break;
 764                 }
 765
 766                 if (flags & SPLICE_F_NONBLOCK) {
 767                         if (!ret)
 768                                 ret = -EAGAIN;
 769                         break;
 770                 }
 771
 772                 if (signal_pending(current)) {
 773                         if (!ret)
 774                                 ret = -ERESTARTSYS;
 775                         break;
 776                 }
 777
 778                 if (do_wakeup) {
 779                         smp_mb();
 780                         if (waitqueue_active(&pipe->wait))
 781                                 wake_up_interruptible_sync(&pipe->wait);
 782                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 783                         do_wakeup = 0;
 784                 }
 785
 786                 pipe_wait(pipe);
 787         }
 788
 789         if (pipe->inode)
 790                 mutex_unlock(&pipe->inode->i_mutex);
 791
 792         if (do_wakeup) {
 793                 smp_mb();
 794                 if (waitqueue_active(&pipe->wait))
 795                         wake_up_interruptible(&pipe->wait);
 796                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 797         }
 798
 799         return ret;
 800 }
 801
 802 /**
 803  * generic_file_splice_write - splice data from a pipe to a file
 804  * @pipe:       pipe info
 805  * @out:        file to write to
 806  * @len:        number of bytes to splice
 807  * @flags:      splice modifier flags
 808  *
 809  * Will either move or copy pages (determined by @flags options) from
 810  * the given pipe inode to the given file.
 811  *
 812  */
 813 ssize_t
 814 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 815                           loff_t *ppos, size_t len, unsigned int flags)
 816 {
 817         struct address_space *mapping = out->f_mapping;
 818         ssize_t ret;
 819
 820         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 821         if (ret > 0) {
 822                 struct inode *inode = mapping->host;
 823
 824                 *ppos += ret;
 825
 826                 /*
 827                  * If file or inode is SYNC and we actually wrote some data,
 828                  * sync it.
 829                  */
 830                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 831                         int err;
 832
 833                         mutex_lock(&inode->i_mutex);
 834                         err = generic_osync_inode(inode, mapping,
 835                                                   OSYNC_METADATA|OSYNC_DATA);
 836                         mutex_unlock(&inode->i_mutex);
 837
 838                         if (err)
 839                                 ret = err;
 840                 }
 841         }
 842
 843         return ret;
 844 }
 845
 846 EXPORT_SYMBOL(generic_file_splice_write);
 847
 848 /**
 849  * generic_splice_sendpage - splice data from a pipe to a socket
 850  * @inode:      pipe inode
 851  * @out:        socket to write to
 852  * @len:        number of bytes to splice
 853  * @flags:      splice modifier flags
 854  *
 855  * Will send @len bytes from the pipe to a network socket. No data copying
 856  * is involved.
 857  *
 858  */
 859 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 860                                 loff_t *ppos, size_t len, unsigned int flags)
 861 {
 862         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 863 }
 864
 865 EXPORT_SYMBOL(generic_splice_sendpage);
 866
 867 /*
 868  * Attempt to initiate a splice from pipe to file.
 869  */
 870 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 871                            loff_t *ppos, size_t len, unsigned int flags)
 872 {
 873         int ret;
 874
 875         if (unlikely(!out->f_op || !out->f_op->splice_write))
 876                 return -EINVAL;
 877
 878         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 879                 return -EBADF;
 880
 881         ret = rw_verify_area(WRITE, out, ppos, len);
 882         if (unlikely(ret < 0))
 883                 return ret;
 884
 885         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 886 }
 887
 888 /*
 889  * Attempt to initiate a splice from a file to a pipe.
 890  */
 891 static long do_splice_to(struct file *in, loff_t *ppos,
 892                          struct pipe_inode_info *pipe, size_t len,
 893                          unsigned int flags)
 894 {
 895         loff_t isize, left;
 896         int ret;
 897
 898         if (unlikely(!in->f_op || !in->f_op->splice_read))
 899                 return -EINVAL;
 900
 901         if (unlikely(!(in->f_mode & FMODE_READ)))
 902                 return -EBADF;
 903
 904         ret = rw_verify_area(READ, in, ppos, len);
 905         if (unlikely(ret < 0))
 906                 return ret;
 907
 908         isize = i_size_read(in->f_mapping->host);
 909         if (unlikely(*ppos >= isize))
 910                 return 0;
 911
 912         left = isize - *ppos;
 913         if (unlikely(left < len))
 914                 len = left;
 915
 916         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 917 }
 918
 919 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 920                       size_t len, unsigned int flags)
 921 {
 922         struct pipe_inode_info *pipe;
 923         long ret, bytes;
 924         loff_t out_off;
 925         umode_t i_mode;
 926         int i;
 927
 928         /*
 929          * We require the input being a regular file, as we don't want to
 930          * randomly drop data for eg socket -> socket splicing. Use the
 931          * piped splicing for that!
 932          */
 933         i_mode = in->f_dentry->d_inode->i_mode;
 934         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 935                 return -EINVAL;
 936
 937         /*
 938          * neither in nor out is a pipe, setup an internal pipe attached to
 939          * 'out' and transfer the wanted data from 'in' to 'out' through that
 940          */
 941         pipe = current->splice_pipe;
 942         if (unlikely(!pipe)) {
 943                 pipe = alloc_pipe_info(NULL);
 944                 if (!pipe)
 945                         return -ENOMEM;
 946
 947                 /*
 948                  * We don't have an immediate reader, but we'll read the stuff
 949                  * out of the pipe right after the splice_to_pipe(). So set
 950                  * PIPE_READERS appropriately.
 951                  */
 952                 pipe->readers = 1;
 953
 954                 current->splice_pipe = pipe;
 955         }
 956
 957         /*
 958          * Do the splice.
 959          */
 960         ret = 0;
 961         bytes = 0;
 962         out_off = 0;
 963
 964         while (len) {
 965                 size_t read_len, max_read_len;
 966
 967                 /*
 968                  * Do at most PIPE_BUFFERS pages worth of transfer:
 969                  */
 970                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 971
 972                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 973                 if (unlikely(ret < 0))
 974                         goto out_release;
 975
 976                 read_len = ret;
 977
 978                 /*
 979                  * NOTE: nonblocking mode only applies to the input. We
 980                  * must not do the output in nonblocking mode as then we
 981                  * could get stuck data in the internal pipe:
 982                  */
 983                 ret = do_splice_from(pipe, out, &out_off, read_len,
 984                                      flags & ~SPLICE_F_NONBLOCK);
 985                 if (unlikely(ret < 0))
 986                         goto out_release;
 987
 988                 bytes += ret;
 989                 len -= ret;
 990
 991                 /*
 992                  * In nonblocking mode, if we got back a short read then
 993                  * that was due to either an IO error or due to the
 994                  * pagecache entry not being there. In the IO error case
 995                  * the _next_ splice attempt will produce a clean IO error
 996                  * return value (not a short read), so in both cases it's
 997                  * correct to break out of the loop here:
 998                  */
 999                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1000                         break;
1001         }
1002
1003         pipe->nrbufs = pipe->curbuf = 0;
1004
1005         return bytes;
1006
1007 out_release:
1008         /*
1009          * If we did an incomplete transfer we must release
1010          * the pipe buffers in question:
1011          */
1012         for (i = 0; i < PIPE_BUFFERS; i++) {
1013                 struct pipe_buffer *buf = pipe->bufs + i;
1014
1015                 if (buf->ops) {
1016                         buf->ops->release(pipe, buf);
1017                         buf->ops = NULL;
1018                 }
1019         }
1020         pipe->nrbufs = pipe->curbuf = 0;
1021
1022         /*
1023          * If we transferred some data, return the number of bytes:
1024          */
1025         if (bytes > 0)
1026                 return bytes;
1027
1028         return ret;
1029 }
1030
1031 EXPORT_SYMBOL(do_splice_direct);
1032
1033 /*
1034  * Determine where to splice to/from.
1035  */
1036 static long do_splice(struct file *in, loff_t __user *off_in,
1037                       struct file *out, loff_t __user *off_out,
1038                       size_t len, unsigned int flags)
1039 {
1040         struct pipe_inode_info *pipe;
1041         loff_t offset, *off;
1042         long ret;
1043
1044         pipe = in->f_dentry->d_inode->i_pipe;
1045         if (pipe) {
1046                 if (off_in)
1047                         return -ESPIPE;
1048                 if (off_out) {
1049                         if (out->f_op->llseek == no_llseek)
1050                                 return -EINVAL;
1051                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1052                                 return -EFAULT;
1053                         off = &offset;
1054                 } else
1055                         off = &out->f_pos;
1056
1057                 ret = do_splice_from(pipe, out, off, len, flags);
1058
1059                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1060                         ret = -EFAULT;
1061
1062                 return ret;
1063         }
1064
1065         pipe = out->f_dentry->d_inode->i_pipe;
1066         if (pipe) {
1067                 if (off_out)
1068                         return -ESPIPE;
1069                 if (off_in) {
1070                         if (in->f_op->llseek == no_llseek)
1071                                 return -EINVAL;
1072                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1073                                 return -EFAULT;
1074                         off = &offset;
1075                 } else
1076                         off = &in->f_pos;
1077
1078                 ret = do_splice_to(in, off, pipe, len, flags);
1079
1080                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1081                         ret = -EFAULT;
1082
1083                 return ret;
1084         }
1085
1086         return -EINVAL;
1087 }
1088
1089 /*
1090  * Map an iov into an array of pages and offset/length tupples. With the
1091  * partial_page structure, we can map several non-contiguous ranges into
1092  * our ones pages[] map instead of splitting that operation into pieces.
1093  * Could easily be exported as a generic helper for other users, in which
1094  * case one would probably want to add a 'max_nr_pages' parameter as well.
1095  */
1096 static int get_iovec_page_array(const struct iovec __user *iov,
1097                                 unsigned int nr_vecs, struct page **pages,
1098                                 struct partial_page *partial)
1099 {
1100         int buffers = 0, error = 0;
1101
1102         /*
1103          * It's ok to take the mmap_sem for reading, even
1104          * across a "get_user()".
1105          */
1106         down_read(&current->mm->mmap_sem);
1107
1108         while (nr_vecs) {
1109                 unsigned long off, npages;
1110                 void __user *base;
1111                 size_t len;
1112                 int i;
1113
1114                 /*
1115                  * Get user address base and length for this iovec.
1116                  */
1117                 error = get_user(base, &iov->iov_base);
1118                 if (unlikely(error))
1119                         break;
1120                 error = get_user(len, &iov->iov_len);
1121                 if (unlikely(error))
1122                         break;
1123
1124                 /*
1125                  * Sanity check this iovec. 0 read succeeds.
1126                  */
1127                 if (unlikely(!len))
1128                         break;
1129                 error = -EFAULT;
1130                 if (unlikely(!base))
1131                         break;
1132
1133                 /*
1134                  * Get this base offset and number of pages, then map
1135                  * in the user pages.
1136                  */
1137                 off = (unsigned long) base & ~PAGE_MASK;
1138                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1139                 if (npages > PIPE_BUFFERS - buffers)
1140                         npages = PIPE_BUFFERS - buffers;
1141
1142                 error = get_user_pages(current, current->mm,
1143                                        (unsigned long) base, npages, 0, 0,
1144                                        &pages[buffers], NULL);
1145
1146                 if (unlikely(error <= 0))
1147                         break;
1148
1149                 /*
1150                  * Fill this contiguous range into the partial page map.
1151                  */
1152                 for (i = 0; i < error; i++) {
1153                         const int plen = min_t(size_t, len, PAGE_SIZE) - off;
1154
1155                         partial[buffers].offset = off;
1156                         partial[buffers].len = plen;
1157
1158                         off = 0;
1159                         len -= plen;
1160                         buffers++;
1161                 }
1162
1163                 /*
1164                  * We didn't complete this iov, stop here since it probably
1165                  * means we have to move some of this into a pipe to
1166                  * be able to continue.
1167                  */
1168                 if (len)
1169                         break;
1170
1171                 /*
1172                  * Don't continue if we mapped fewer pages than we asked for,
1173                  * or if we mapped the max number of pages that we have
1174                  * room for.
1175                  */
1176                 if (error < npages || buffers == PIPE_BUFFERS)
1177                         break;
1178
1179                 nr_vecs--;
1180                 iov++;
1181         }
1182
1183         up_read(&current->mm->mmap_sem);
1184
1185         if (buffers)
1186                 return buffers;
1187
1188         return error;
1189 }
1190
1191 /*
1192  * vmsplice splices a user address range into a pipe. It can be thought of
1193  * as splice-from-memory, where the regular splice is splice-from-file (or
1194  * to file). In both cases the output is a pipe, naturally.
1195  *
1196  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1197  * not the other way around. Splicing from user memory is a simple operation
1198  * that can be supported without any funky alignment restrictions or nasty
1199  * vm tricks. We simply map in the user memory and fill them into a pipe.
1200  * The reverse isn't quite as easy, though. There are two possible solutions
1201  * for that:
1202  *
1203  *      - memcpy() the data internally, at which point we might as well just
1204  *        do a regular read() on the buffer anyway.
1205  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1206  *        has restriction limitations on both ends of the pipe).
1207  *
1208  * Alas, it isn't here.
1209  *
1210  */
1211 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1212                         unsigned long nr_segs, unsigned int flags)
1213 {
1214         struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1215         struct page *pages[PIPE_BUFFERS];
1216         struct partial_page partial[PIPE_BUFFERS];
1217         struct splice_pipe_desc spd = {
1218                 .pages = pages,
1219                 .partial = partial,
1220                 .flags = flags,
1221                 .ops = &user_page_pipe_buf_ops,
1222         };
1223
1224         if (unlikely(!pipe))
1225                 return -EBADF;
1226         if (unlikely(nr_segs > UIO_MAXIOV))
1227                 return -EINVAL;
1228         else if (unlikely(!nr_segs))
1229                 return 0;
1230
1231         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial);
1232         if (spd.nr_pages <= 0)
1233                 return spd.nr_pages;
1234
1235         return splice_to_pipe(pipe, &spd);
1236 }
1237
1238 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1239                              unsigned long nr_segs, unsigned int flags)
1240 {
1241         struct file *file;
1242         long error;
1243         int fput;
1244
1245         error = -EBADF;
1246         file = fget_light(fd, &fput);
1247         if (file) {
1248                 if (file->f_mode & FMODE_WRITE)
1249                         error = do_vmsplice(file, iov, nr_segs, flags);
1250
1251                 fput_light(file, fput);
1252         }
1253
1254         return error;
1255 }
1256
1257 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1258                            int fd_out, loff_t __user *off_out,
1259                            size_t len, unsigned int flags)
1260 {
1261         long error;
1262         struct file *in, *out;
1263         int fput_in, fput_out;
1264
1265         if (unlikely(!len))
1266                 return 0;
1267
1268         error = -EBADF;
1269         in = fget_light(fd_in, &fput_in);
1270         if (in) {
1271                 if (in->f_mode & FMODE_READ) {
1272                         out = fget_light(fd_out, &fput_out);
1273                         if (out) {
1274                                 if (out->f_mode & FMODE_WRITE)
1275                                         error = do_splice(in, off_in,
1276                                                           out, off_out,
1277                                                           len, flags);
1278                                 fput_light(out, fput_out);
1279                         }
1280                 }
1281
1282                 fput_light(in, fput_in);
1283         }
1284
1285         return error;
1286 }
1287
1288 /*
1289  * Link contents of ipipe to opipe.
1290  */
1291 static int link_pipe(struct pipe_inode_info *ipipe,
1292                      struct pipe_inode_info *opipe,
1293                      size_t len, unsigned int flags)
1294 {
1295         struct pipe_buffer *ibuf, *obuf;
1296         int ret, do_wakeup, i, ipipe_first;
1297
1298         ret = do_wakeup = ipipe_first = 0;
1299
1300         /*
1301          * Potential ABBA deadlock, work around it by ordering lock
1302          * grabbing by inode address. Otherwise two different processes
1303          * could deadlock (one doing tee from A -> B, the other from B -> A).
1304          */
1305         if (ipipe->inode < opipe->inode) {
1306                 ipipe_first = 1;
1307                 mutex_lock(&ipipe->inode->i_mutex);
1308                 mutex_lock(&opipe->inode->i_mutex);
1309         } else {
1310                 mutex_lock(&opipe->inode->i_mutex);
1311                 mutex_lock(&ipipe->inode->i_mutex);
1312         }
1313
1314         for (i = 0;; i++) {
1315                 if (!opipe->readers) {
1316                         send_sig(SIGPIPE, current, 0);
1317                         if (!ret)
1318                                 ret = -EPIPE;
1319                         break;
1320                 }
1321                 if (ipipe->nrbufs - i) {
1322                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1323
1324                         /*
1325                          * If we have room, fill this buffer
1326                          */
1327                         if (opipe->nrbufs < PIPE_BUFFERS) {
1328                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1329
1330                                 /*
1331                                  * Get a reference to this pipe buffer,
1332                                  * so we can copy the contents over.
1333                                  */
1334                                 ibuf->ops->get(ipipe, ibuf);
1335
1336                                 obuf = opipe->bufs + nbuf;
1337                                 *obuf = *ibuf;
1338
1339                                 if (obuf->len > len)
1340                                         obuf->len = len;
1341
1342                                 opipe->nrbufs++;
1343                                 do_wakeup = 1;
1344                                 ret += obuf->len;
1345                                 len -= obuf->len;
1346
1347                                 if (!len)
1348                                         break;
1349                                 if (opipe->nrbufs < PIPE_BUFFERS)
1350                                         continue;
1351                         }
1352
1353                         /*
1354                          * We have input available, but no output room.
1355                          * If we already copied data, return that. If we
1356                          * need to drop the opipe lock, it must be ordered
1357                          * last to avoid deadlocks.
1358                          */
1359                         if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
1360                                 if (!ret)
1361                                         ret = -EAGAIN;
1362                                 break;
1363                         }
1364                         if (signal_pending(current)) {
1365                                 if (!ret)
1366                                         ret = -ERESTARTSYS;
1367                                 break;
1368                         }
1369                         if (do_wakeup) {
1370                                 smp_mb();
1371                                 if (waitqueue_active(&opipe->wait))
1372                                         wake_up_interruptible(&opipe->wait);
1373                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1374                                 do_wakeup = 0;
1375                         }
1376
1377                         opipe->waiting_writers++;
1378                         pipe_wait(opipe);
1379                         opipe->waiting_writers--;
1380                         continue;
1381                 }
1382
1383                 /*
1384                  * No input buffers, do the usual checks for available
1385                  * writers and blocking and wait if necessary
1386                  */
1387                 if (!ipipe->writers)
1388                         break;
1389                 if (!ipipe->waiting_writers) {
1390                         if (ret)
1391                                 break;
1392                 }
1393                 /*
1394                  * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1395                  * with another process, we can only safely do that if
1396                  * the ipipe lock is ordered last.
1397                  */
1398                 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
1399                         if (!ret)
1400                                 ret = -EAGAIN;
1401                         break;
1402                 }
1403                 if (signal_pending(current)) {
1404                         if (!ret)
1405                                 ret = -ERESTARTSYS;
1406                         break;
1407                 }
1408
1409                 if (waitqueue_active(&ipipe->wait))
1410                         wake_up_interruptible_sync(&ipipe->wait);
1411                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1412
1413                 pipe_wait(ipipe);
1414         }
1415
1416         mutex_unlock(&ipipe->inode->i_mutex);
1417         mutex_unlock(&opipe->inode->i_mutex);
1418
1419         if (do_wakeup) {
1420                 smp_mb();
1421                 if (waitqueue_active(&opipe->wait))
1422                         wake_up_interruptible(&opipe->wait);
1423                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1424         }
1425
1426         return ret;
1427 }
1428
1429 /*
1430  * This is a tee(1) implementation that works on pipes. It doesn't copy
1431  * any data, it simply references the 'in' pages on the 'out' pipe.
1432  * The 'flags' used are the SPLICE_F_* variants, currently the only
1433  * applicable one is SPLICE_F_NONBLOCK.
1434  */
1435 static long do_tee(struct file *in, struct file *out, size_t len,
1436                    unsigned int flags)
1437 {
1438         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1439         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1440
1441         /*
1442          * Link ipipe to the two output pipes, consuming as we go along.
1443          */
1444         if (ipipe && opipe)
1445                 return link_pipe(ipipe, opipe, len, flags);
1446
1447         return -EINVAL;
1448 }
1449
1450 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1451 {
1452         struct file *in;
1453         int error, fput_in;
1454
1455         if (unlikely(!len))
1456                 return 0;
1457
1458         error = -EBADF;
1459         in = fget_light(fdin, &fput_in);
1460         if (in) {
1461                 if (in->f_mode & FMODE_READ) {
1462                         int fput_out;
1463                         struct file *out = fget_light(fdout, &fput_out);
1464
1465                         if (out) {
1466                                 if (out->f_mode & FMODE_WRITE)
1467                                         error = do_tee(in, out, len, flags);
1468                                 fput_light(out, fput_out);
1469                         }
1470                 }
1471                 fput_light(in, fput_in);
1472         }
1473
1474         return error;
1475 }