SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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         return 0;
  82 }
  83
  84 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  85                                         struct pipe_buffer *buf)
  86 {
  87         page_cache_release(buf->page);
  88 }
  89
  90 static int page_cache_pipe_buf_pin(struct pipe_inode_info *info,
  91                                    struct pipe_buffer *buf)
  92 {
  93         struct page *page = buf->page;
  94         int err;
  95
  96         if (!PageUptodate(page)) {
  97                 lock_page(page);
  98
  99                 /*
 100                  * Page got truncated/unhashed. This will cause a 0-byte
 101                  * splice, if this is the first page.
 102                  */
 103                 if (!page->mapping) {
 104                         err = -ENODATA;
 105                         goto error;
 106                 }
 107
 108                 /*
 109                  * Uh oh, read-error from disk.
 110                  */
 111                 if (!PageUptodate(page)) {
 112                         err = -EIO;
 113                         goto error;
 114                 }
 115
 116                 /*
 117                  * Page is ok afterall, we are done.
 118                  */
 119                 unlock_page(page);
 120         }
 121
 122         return 0;
 123 error:
 124         unlock_page(page);
 125         return err;
 126 }
 127
 128 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 129         .can_merge = 0,
 130         .map = generic_pipe_buf_map,
 131         .unmap = generic_pipe_buf_unmap,
 132         .pin = page_cache_pipe_buf_pin,
 133         .release = page_cache_pipe_buf_release,
 134         .steal = page_cache_pipe_buf_steal,
 135         .get = generic_pipe_buf_get,
 136 };
 137
 138 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 139                                     struct pipe_buffer *buf)
 140 {
 141         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 142                 return 1;
 143
 144         return 0;
 145 }
 146
 147 static struct pipe_buf_operations user_page_pipe_buf_ops = {
 148         .can_merge = 0,
 149         .map = generic_pipe_buf_map,
 150         .unmap = generic_pipe_buf_unmap,
 151         .pin = generic_pipe_buf_pin,
 152         .release = page_cache_pipe_buf_release,
 153         .steal = user_page_pipe_buf_steal,
 154         .get = generic_pipe_buf_get,
 155 };
 156
 157 /*
 158  * Pipe output worker. This sets up our pipe format with the page cache
 159  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 160  */
 161 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 162                               struct splice_pipe_desc *spd)
 163 {
 164         int ret, do_wakeup, page_nr;
 165
 166         ret = 0;
 167         do_wakeup = 0;
 168         page_nr = 0;
 169
 170         if (pipe->inode)
 171                 mutex_lock(&pipe->inode->i_mutex);
 172
 173         for (;;) {
 174                 if (!pipe->readers) {
 175                         send_sig(SIGPIPE, current, 0);
 176                         if (!ret)
 177                                 ret = -EPIPE;
 178                         break;
 179                 }
 180
 181                 if (pipe->nrbufs < PIPE_BUFFERS) {
 182                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 183                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 184
 185                         buf->page = spd->pages[page_nr];
 186                         buf->offset = spd->partial[page_nr].offset;
 187                         buf->len = spd->partial[page_nr].len;
 188                         buf->ops = spd->ops;
 189                         if (spd->flags & SPLICE_F_GIFT)
 190                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 191
 192                         pipe->nrbufs++;
 193                         page_nr++;
 194                         ret += buf->len;
 195
 196                         if (pipe->inode)
 197                                 do_wakeup = 1;
 198
 199                         if (!--spd->nr_pages)
 200                                 break;
 201                         if (pipe->nrbufs < PIPE_BUFFERS)
 202                                 continue;
 203
 204                         break;
 205                 }
 206
 207                 if (spd->flags & SPLICE_F_NONBLOCK) {
 208                         if (!ret)
 209                                 ret = -EAGAIN;
 210                         break;
 211                 }
 212
 213                 if (signal_pending(current)) {
 214                         if (!ret)
 215                                 ret = -ERESTARTSYS;
 216                         break;
 217                 }
 218
 219                 if (do_wakeup) {
 220                         smp_mb();
 221                         if (waitqueue_active(&pipe->wait))
 222                                 wake_up_interruptible_sync(&pipe->wait);
 223                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 224                         do_wakeup = 0;
 225                 }
 226
 227                 pipe->waiting_writers++;
 228                 pipe_wait(pipe);
 229                 pipe->waiting_writers--;
 230         }
 231
 232         if (pipe->inode)
 233                 mutex_unlock(&pipe->inode->i_mutex);
 234
 235         if (do_wakeup) {
 236                 smp_mb();
 237                 if (waitqueue_active(&pipe->wait))
 238                         wake_up_interruptible(&pipe->wait);
 239                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 240         }
 241
 242         while (page_nr < spd->nr_pages)
 243                 page_cache_release(spd->pages[page_nr++]);
 244
 245         return ret;
 246 }
 247
 248 static int
 249 __generic_file_splice_read(struct file *in, loff_t *ppos,
 250                            struct pipe_inode_info *pipe, size_t len,
 251                            unsigned int flags)
 252 {
 253         struct address_space *mapping = in->f_mapping;
 254         unsigned int loff, nr_pages;
 255         struct page *pages[PIPE_BUFFERS];
 256         struct partial_page partial[PIPE_BUFFERS];
 257         struct page *page;
 258         pgoff_t index, end_index;
 259         loff_t isize;
 260         size_t total_len;
 261         int error, page_nr;
 262         struct splice_pipe_desc spd = {
 263                 .pages = pages,
 264                 .partial = partial,
 265                 .flags = flags,
 266                 .ops = &page_cache_pipe_buf_ops,
 267         };
 268
 269         index = *ppos >> PAGE_CACHE_SHIFT;
 270         loff = *ppos & ~PAGE_CACHE_MASK;
 271         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 272
 273         if (nr_pages > PIPE_BUFFERS)
 274                 nr_pages = PIPE_BUFFERS;
 275
 276         /*
 277          * Initiate read-ahead on this page range. however, don't call into
 278          * read-ahead if this is a non-zero offset (we are likely doing small
 279          * chunk splice and the page is already there) for a single page.
 280          */
 281         if (!loff || nr_pages > 1)
 282                 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 283
 284         /*
 285          * Now fill in the holes:
 286          */
 287         error = 0;
 288         total_len = 0;
 289
 290         /*
 291          * Lookup the (hopefully) full range of pages we need.
 292          */
 293         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 294
 295         /*
 296          * If find_get_pages_contig() returned fewer pages than we needed,
 297          * allocate the rest.
 298          */
 299         index += spd.nr_pages;
 300         while (spd.nr_pages < nr_pages) {
 301                 /*
 302                  * Page could be there, find_get_pages_contig() breaks on
 303                  * the first hole.
 304                  */
 305                 page = find_get_page(mapping, index);
 306                 if (!page) {
 307                         /*
 308                          * Make sure the read-ahead engine is notified
 309                          * about this failure.
 310                          */
 311                         handle_ra_miss(mapping, &in->f_ra, index);
 312
 313                         /*
 314                          * page didn't exist, allocate one.
 315                          */
 316                         page = page_cache_alloc_cold(mapping);
 317                         if (!page)
 318                                 break;
 319
 320                         error = add_to_page_cache_lru(page, mapping, index,
 321                                               mapping_gfp_mask(mapping));
 322                         if (unlikely(error)) {
 323                                 page_cache_release(page);
 324                                 break;
 325                         }
 326                         /*
 327                          * add_to_page_cache() locks the page, unlock it
 328                          * to avoid convoluting the logic below even more.
 329                          */
 330                         unlock_page(page);
 331                 }
 332
 333                 pages[spd.nr_pages++] = page;
 334                 index++;
 335         }
 336
 337         /*
 338          * Now loop over the map and see if we need to start IO on any
 339          * pages, fill in the partial map, etc.
 340          */
 341         index = *ppos >> PAGE_CACHE_SHIFT;
 342         nr_pages = spd.nr_pages;
 343         spd.nr_pages = 0;
 344         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 345                 unsigned int this_len;
 346
 347                 if (!len)
 348                         break;
 349
 350                 /*
 351                  * this_len is the max we'll use from this page
 352                  */
 353                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 354                 page = pages[page_nr];
 355
 356                 /*
 357                  * If the page isn't uptodate, we may need to start io on it
 358                  */
 359                 if (!PageUptodate(page)) {
 360                         /*
 361                          * If in nonblock mode then dont block on waiting
 362                          * for an in-flight io page
 363                          */
 364                         if (flags & SPLICE_F_NONBLOCK)
 365                                 break;
 366
 367                         lock_page(page);
 368
 369                         /*
 370                          * page was truncated, stop here. if this isn't the
 371                          * first page, we'll just complete what we already
 372                          * added
 373                          */
 374                         if (!page->mapping) {
 375                                 unlock_page(page);
 376                                 break;
 377                         }
 378                         /*
 379                          * page was already under io and is now done, great
 380                          */
 381                         if (PageUptodate(page)) {
 382                                 unlock_page(page);
 383                                 goto fill_it;
 384                         }
 385
 386                         /*
 387                          * need to read in the page
 388                          */
 389                         error = mapping->a_ops->readpage(in, page);
 390                         if (unlikely(error)) {
 391                                 /*
 392                                  * We really should re-lookup the page here,
 393                                  * but it complicates things a lot. Instead
 394                                  * lets just do what we already stored, and
 395                                  * we'll get it the next time we are called.
 396                                  */
 397                                 if (error == AOP_TRUNCATED_PAGE)
 398                                         error = 0;
 399
 400                                 break;
 401                         }
 402
 403                         /*
 404                          * i_size must be checked after ->readpage().
 405                          */
 406                         isize = i_size_read(mapping->host);
 407                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 408                         if (unlikely(!isize || index > end_index))
 409                                 break;
 410
 411                         /*
 412                          * if this is the last page, see if we need to shrink
 413                          * the length and stop
 414                          */
 415                         if (end_index == index) {
 416                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 417                                 if (total_len + loff > isize)
 418                                         break;
 419                                 /*
 420                                  * force quit after adding this page
 421                                  */
 422                                 len = this_len;
 423                                 this_len = min(this_len, loff);
 424                                 loff = 0;
 425                         }
 426                 }
 427 fill_it:
 428                 partial[page_nr].offset = loff;
 429                 partial[page_nr].len = this_len;
 430                 len -= this_len;
 431                 total_len += this_len;
 432                 loff = 0;
 433                 spd.nr_pages++;
 434                 index++;
 435         }
 436
 437         /*
 438          * Release any pages at the end, if we quit early. 'i' is how far
 439          * we got, 'nr_pages' is how many pages are in the map.
 440          */
 441         while (page_nr < nr_pages)
 442                 page_cache_release(pages[page_nr++]);
 443
 444         if (spd.nr_pages)
 445                 return splice_to_pipe(pipe, &spd);
 446
 447         return error;
 448 }
 449
 450 /**
 451  * generic_file_splice_read - splice data from file to a pipe
 452  * @in:         file to splice from
 453  * @pipe:       pipe to splice to
 454  * @len:        number of bytes to splice
 455  * @flags:      splice modifier flags
 456  *
 457  * Will read pages from given file and fill them into a pipe.
 458  */
 459 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 460                                  struct pipe_inode_info *pipe, size_t len,
 461                                  unsigned int flags)
 462 {
 463         ssize_t spliced;
 464         int ret;
 465
 466         ret = 0;
 467         spliced = 0;
 468
 469         while (len) {
 470                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 471
 472                 if (ret < 0)
 473                         break;
 474                 else if (!ret) {
 475                         if (spliced)
 476                                 break;
 477                         if (flags & SPLICE_F_NONBLOCK) {
 478                                 ret = -EAGAIN;
 479                                 break;
 480                         }
 481                 }
 482
 483                 *ppos += ret;
 484                 len -= ret;
 485                 spliced += ret;
 486         }
 487
 488         if (spliced)
 489                 return spliced;
 490
 491         return ret;
 492 }
 493
 494 EXPORT_SYMBOL(generic_file_splice_read);
 495
 496 /*
 497  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 498  * using sendpage(). Return the number of bytes sent.
 499  */
 500 static int pipe_to_sendpage(struct pipe_inode_info *info,
 501                             struct pipe_buffer *buf, struct splice_desc *sd)
 502 {
 503         struct file *file = sd->file;
 504         loff_t pos = sd->pos;
 505         int ret, more;
 506
 507         ret = buf->ops->pin(info, buf);
 508         if (!ret) {
 509                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 510
 511                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 512                                            sd->len, &pos, more);
 513         }
 514
 515         return ret;
 516 }
 517
 518 /*
 519  * This is a little more tricky than the file -> pipe splicing. There are
 520  * basically three cases:
 521  *
 522  *      - Destination page already exists in the address space and there
 523  *        are users of it. For that case we have no other option that
 524  *        copying the data. Tough luck.
 525  *      - Destination page already exists in the address space, but there
 526  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 527  *        through to last case.
 528  *      - Destination page does not exist, we can add the pipe page to
 529  *        the page cache and avoid the copy.
 530  *
 531  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 532  * sd->flags), we attempt to migrate pages from the pipe to the output
 533  * file address space page cache. This is possible if no one else has
 534  * the pipe page referenced outside of the pipe and page cache. If
 535  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 536  * a new page in the output file page cache and fill/dirty that.
 537  */
 538 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 539                         struct splice_desc *sd)
 540 {
 541         struct file *file = sd->file;
 542         struct address_space *mapping = file->f_mapping;
 543         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 544         unsigned int offset, this_len;
 545         struct page *page;
 546         pgoff_t index;
 547         int ret;
 548
 549         /*
 550          * make sure the data in this buffer is uptodate
 551          */
 552         ret = buf->ops->pin(info, buf);
 553         if (unlikely(ret))
 554                 return ret;
 555
 556         index = sd->pos >> PAGE_CACHE_SHIFT;
 557         offset = sd->pos & ~PAGE_CACHE_MASK;
 558
 559         this_len = sd->len;
 560         if (this_len + offset > PAGE_CACHE_SIZE)
 561                 this_len = PAGE_CACHE_SIZE - offset;
 562
 563         /*
 564          * Reuse buf page, if SPLICE_F_MOVE is set and we are doing a full
 565          * page.
 566          */
 567         if ((sd->flags & SPLICE_F_MOVE) && this_len == PAGE_CACHE_SIZE) {
 568                 /*
 569                  * If steal succeeds, buf->page is now pruned from the vm
 570                  * side (page cache) and we can reuse it. The page will also
 571                  * be locked on successful return.
 572                  */
 573                 if (buf->ops->steal(info, buf))
 574                         goto find_page;
 575
 576                 page = buf->page;
 577                 page_cache_get(page);
 578
 579                 /*
 580                  * page must be on the LRU for adding to the pagecache.
 581                  * Check this without grabbing the zone lock, if it isn't
 582                  * the do grab the zone lock, recheck, and add if necessary.
 583                  */
 584                 if (!PageLRU(page)) {
 585                         struct zone *zone = page_zone(page);
 586
 587                         spin_lock_irq(&zone->lru_lock);
 588                         if (!PageLRU(page)) {
 589                                 SetPageLRU(page);
 590                                 add_page_to_inactive_list(zone, page);
 591                         }
 592                         spin_unlock_irq(&zone->lru_lock);
 593                 }
 594
 595                 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
 596                         page_cache_release(page);
 597                         unlock_page(page);
 598                         goto find_page;
 599                 }
 600         } else {
 601 find_page:
 602                 page = find_lock_page(mapping, index);
 603                 if (!page) {
 604                         ret = -ENOMEM;
 605                         page = page_cache_alloc_cold(mapping);
 606                         if (unlikely(!page))
 607                                 goto out_nomem;
 608
 609                         /*
 610                          * This will also lock the page
 611                          */
 612                         ret = add_to_page_cache_lru(page, mapping, index,
 613                                                     gfp_mask);
 614                         if (unlikely(ret))
 615                                 goto out;
 616                 }
 617
 618                 /*
 619                  * We get here with the page locked. If the page is also
 620                  * uptodate, we don't need to do more. If it isn't, we
 621                  * may need to bring it in if we are not going to overwrite
 622                  * the full page.
 623                  */
 624                 if (!PageUptodate(page)) {
 625                         if (this_len < PAGE_CACHE_SIZE) {
 626                                 ret = mapping->a_ops->readpage(file, page);
 627                                 if (unlikely(ret))
 628                                         goto out;
 629
 630                                 lock_page(page);
 631
 632                                 if (!PageUptodate(page)) {
 633                                         /*
 634                                          * Page got invalidated, repeat.
 635                                          */
 636                                         if (!page->mapping) {
 637                                                 unlock_page(page);
 638                                                 page_cache_release(page);
 639                                                 goto find_page;
 640                                         }
 641                                         ret = -EIO;
 642                                         goto out;
 643                                 }
 644                         } else
 645                                 SetPageUptodate(page);
 646                 }
 647         }
 648
 649         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 650         if (ret == AOP_TRUNCATED_PAGE) {
 651                 page_cache_release(page);
 652                 goto find_page;
 653         } else if (ret)
 654                 goto out;
 655
 656         if (buf->page != page) {
 657                 /*
 658                  * Careful, ->map() uses KM_USER0!
 659                  */
 660                 char *src = buf->ops->map(info, buf, 1);
 661                 char *dst = kmap_atomic(page, KM_USER1);
 662
 663                 memcpy(dst + offset, src + buf->offset, this_len);
 664                 flush_dcache_page(page);
 665                 kunmap_atomic(dst, KM_USER1);
 666                 buf->ops->unmap(info, buf, src);
 667         }
 668
 669         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 670         if (!ret) {
 671                 /*
 672                  * Return the number of bytes written and mark page as
 673                  * accessed, we are now done!
 674                  */
 675                 ret = this_len;
 676                 mark_page_accessed(page);
 677                 balance_dirty_pages_ratelimited(mapping);
 678         } else if (ret == AOP_TRUNCATED_PAGE) {
 679                 page_cache_release(page);
 680                 goto find_page;
 681         }
 682 out:
 683         page_cache_release(page);
 684         unlock_page(page);
 685 out_nomem:
 686         return ret;
 687 }
 688
 689 /*
 690  * Pipe input worker. Most of this logic works like a regular pipe, the
 691  * key here is the 'actor' worker passed in that actually moves the data
 692  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 693  */
 694 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 695                          loff_t *ppos, size_t len, unsigned int flags,
 696                          splice_actor *actor)
 697 {
 698         int ret, do_wakeup, err;
 699         struct splice_desc sd;
 700
 701         ret = 0;
 702         do_wakeup = 0;
 703
 704         sd.total_len = len;
 705         sd.flags = flags;
 706         sd.file = out;
 707         sd.pos = *ppos;
 708
 709         if (pipe->inode)
 710                 mutex_lock(&pipe->inode->i_mutex);
 711
 712         for (;;) {
 713                 if (pipe->nrbufs) {
 714                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 715                         struct pipe_buf_operations *ops = buf->ops;
 716
 717                         sd.len = buf->len;
 718                         if (sd.len > sd.total_len)
 719                                 sd.len = sd.total_len;
 720
 721                         err = actor(pipe, buf, &sd);
 722                         if (err <= 0) {
 723                                 if (!ret && err != -ENODATA)
 724                                         ret = err;
 725
 726                                 break;
 727                         }
 728
 729                         ret += err;
 730                         buf->offset += err;
 731                         buf->len -= err;
 732
 733                         sd.len -= err;
 734                         sd.pos += err;
 735                         sd.total_len -= err;
 736                         if (sd.len)
 737                                 continue;
 738
 739                         if (!buf->len) {
 740                                 buf->ops = NULL;
 741                                 ops->release(pipe, buf);
 742                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 743                                 pipe->nrbufs--;
 744                                 if (pipe->inode)
 745                                         do_wakeup = 1;
 746                         }
 747
 748                         if (!sd.total_len)
 749                                 break;
 750                 }
 751
 752                 if (pipe->nrbufs)
 753                         continue;
 754                 if (!pipe->writers)
 755                         break;
 756                 if (!pipe->waiting_writers) {
 757                         if (ret)
 758                                 break;
 759                 }
 760
 761                 if (flags & SPLICE_F_NONBLOCK) {
 762                         if (!ret)
 763                                 ret = -EAGAIN;
 764                         break;
 765                 }
 766
 767                 if (signal_pending(current)) {
 768                         if (!ret)
 769                                 ret = -ERESTARTSYS;
 770                         break;
 771                 }
 772
 773                 if (do_wakeup) {
 774                         smp_mb();
 775                         if (waitqueue_active(&pipe->wait))
 776                                 wake_up_interruptible_sync(&pipe->wait);
 777                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 778                         do_wakeup = 0;
 779                 }
 780
 781                 pipe_wait(pipe);
 782         }
 783
 784         if (pipe->inode)
 785                 mutex_unlock(&pipe->inode->i_mutex);
 786
 787         if (do_wakeup) {
 788                 smp_mb();
 789                 if (waitqueue_active(&pipe->wait))
 790                         wake_up_interruptible(&pipe->wait);
 791                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 792         }
 793
 794         return ret;
 795 }
 796
 797 /**
 798  * generic_file_splice_write - splice data from a pipe to a file
 799  * @pipe:       pipe info
 800  * @out:        file to write to
 801  * @len:        number of bytes to splice
 802  * @flags:      splice modifier flags
 803  *
 804  * Will either move or copy pages (determined by @flags options) from
 805  * the given pipe inode to the given file.
 806  *
 807  */
 808 ssize_t
 809 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 810                           loff_t *ppos, size_t len, unsigned int flags)
 811 {
 812         struct address_space *mapping = out->f_mapping;
 813         ssize_t ret;
 814
 815         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 816         if (ret > 0) {
 817                 struct inode *inode = mapping->host;
 818
 819                 *ppos += ret;
 820
 821                 /*
 822                  * If file or inode is SYNC and we actually wrote some data,
 823                  * sync it.
 824                  */
 825                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 826                         int err;
 827
 828                         mutex_lock(&inode->i_mutex);
 829                         err = generic_osync_inode(inode, mapping,
 830                                                   OSYNC_METADATA|OSYNC_DATA);
 831                         mutex_unlock(&inode->i_mutex);
 832
 833                         if (err)
 834                                 ret = err;
 835                 }
 836         }
 837
 838         return ret;
 839 }
 840
 841 EXPORT_SYMBOL(generic_file_splice_write);
 842
 843 /**
 844  * generic_splice_sendpage - splice data from a pipe to a socket
 845  * @inode:      pipe inode
 846  * @out:        socket to write to
 847  * @len:        number of bytes to splice
 848  * @flags:      splice modifier flags
 849  *
 850  * Will send @len bytes from the pipe to a network socket. No data copying
 851  * is involved.
 852  *
 853  */
 854 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 855                                 loff_t *ppos, size_t len, unsigned int flags)
 856 {
 857         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 858 }
 859
 860 EXPORT_SYMBOL(generic_splice_sendpage);
 861
 862 /*
 863  * Attempt to initiate a splice from pipe to file.
 864  */
 865 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 866                            loff_t *ppos, size_t len, unsigned int flags)
 867 {
 868         int ret;
 869
 870         if (unlikely(!out->f_op || !out->f_op->splice_write))
 871                 return -EINVAL;
 872
 873         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 874                 return -EBADF;
 875
 876         ret = rw_verify_area(WRITE, out, ppos, len);
 877         if (unlikely(ret < 0))
 878                 return ret;
 879
 880         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 881 }
 882
 883 /*
 884  * Attempt to initiate a splice from a file to a pipe.
 885  */
 886 static long do_splice_to(struct file *in, loff_t *ppos,
 887                          struct pipe_inode_info *pipe, size_t len,
 888                          unsigned int flags)
 889 {
 890         loff_t isize, left;
 891         int ret;
 892
 893         if (unlikely(!in->f_op || !in->f_op->splice_read))
 894                 return -EINVAL;
 895
 896         if (unlikely(!(in->f_mode & FMODE_READ)))
 897                 return -EBADF;
 898
 899         ret = rw_verify_area(READ, in, ppos, len);
 900         if (unlikely(ret < 0))
 901                 return ret;
 902
 903         isize = i_size_read(in->f_mapping->host);
 904         if (unlikely(*ppos >= isize))
 905                 return 0;
 906
 907         left = isize - *ppos;
 908         if (unlikely(left < len))
 909                 len = left;
 910
 911         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 912 }
 913
 914 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 915                       size_t len, unsigned int flags)
 916 {
 917         struct pipe_inode_info *pipe;
 918         long ret, bytes;
 919         loff_t out_off;
 920         umode_t i_mode;
 921         int i;
 922
 923         /*
 924          * We require the input being a regular file, as we don't want to
 925          * randomly drop data for eg socket -> socket splicing. Use the
 926          * piped splicing for that!
 927          */
 928         i_mode = in->f_dentry->d_inode->i_mode;
 929         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 930                 return -EINVAL;
 931
 932         /*
 933          * neither in nor out is a pipe, setup an internal pipe attached to
 934          * 'out' and transfer the wanted data from 'in' to 'out' through that
 935          */
 936         pipe = current->splice_pipe;
 937         if (unlikely(!pipe)) {
 938                 pipe = alloc_pipe_info(NULL);
 939                 if (!pipe)
 940                         return -ENOMEM;
 941
 942                 /*
 943                  * We don't have an immediate reader, but we'll read the stuff
 944                  * out of the pipe right after the splice_to_pipe(). So set
 945                  * PIPE_READERS appropriately.
 946                  */
 947                 pipe->readers = 1;
 948
 949                 current->splice_pipe = pipe;
 950         }
 951
 952         /*
 953          * Do the splice.
 954          */
 955         ret = 0;
 956         bytes = 0;
 957         out_off = 0;
 958
 959         while (len) {
 960                 size_t read_len, max_read_len;
 961
 962                 /*
 963                  * Do at most PIPE_BUFFERS pages worth of transfer:
 964                  */
 965                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 966
 967                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 968                 if (unlikely(ret < 0))
 969                         goto out_release;
 970
 971                 read_len = ret;
 972
 973                 /*
 974                  * NOTE: nonblocking mode only applies to the input. We
 975                  * must not do the output in nonblocking mode as then we
 976                  * could get stuck data in the internal pipe:
 977                  */
 978                 ret = do_splice_from(pipe, out, &out_off, read_len,
 979                                      flags & ~SPLICE_F_NONBLOCK);
 980                 if (unlikely(ret < 0))
 981                         goto out_release;
 982
 983                 bytes += ret;
 984                 len -= ret;
 985
 986                 /*
 987                  * In nonblocking mode, if we got back a short read then
 988                  * that was due to either an IO error or due to the
 989                  * pagecache entry not being there. In the IO error case
 990                  * the _next_ splice attempt will produce a clean IO error
 991                  * return value (not a short read), so in both cases it's
 992                  * correct to break out of the loop here:
 993                  */
 994                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 995                         break;
 996         }
 997
 998         pipe->nrbufs = pipe->curbuf = 0;
 999
1000         return bytes;
1001
1002 out_release:
1003         /*
1004          * If we did an incomplete transfer we must release
1005          * the pipe buffers in question:
1006          */
1007         for (i = 0; i < PIPE_BUFFERS; i++) {
1008                 struct pipe_buffer *buf = pipe->bufs + i;
1009
1010                 if (buf->ops) {
1011                         buf->ops->release(pipe, buf);
1012                         buf->ops = NULL;
1013                 }
1014         }
1015         pipe->nrbufs = pipe->curbuf = 0;
1016
1017         /*
1018          * If we transferred some data, return the number of bytes:
1019          */
1020         if (bytes > 0)
1021                 return bytes;
1022
1023         return ret;
1024 }
1025
1026 EXPORT_SYMBOL(do_splice_direct);
1027
1028 /*
1029  * Determine where to splice to/from.
1030  */
1031 static long do_splice(struct file *in, loff_t __user *off_in,
1032                       struct file *out, loff_t __user *off_out,
1033                       size_t len, unsigned int flags)
1034 {
1035         struct pipe_inode_info *pipe;
1036         loff_t offset, *off;
1037         long ret;
1038
1039         pipe = in->f_dentry->d_inode->i_pipe;
1040         if (pipe) {
1041                 if (off_in)
1042                         return -ESPIPE;
1043                 if (off_out) {
1044                         if (out->f_op->llseek == no_llseek)
1045                                 return -EINVAL;
1046                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1047                                 return -EFAULT;
1048                         off = &offset;
1049                 } else
1050                         off = &out->f_pos;
1051
1052                 ret = do_splice_from(pipe, out, off, len, flags);
1053
1054                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1055                         ret = -EFAULT;
1056
1057                 return ret;
1058         }
1059
1060         pipe = out->f_dentry->d_inode->i_pipe;
1061         if (pipe) {
1062                 if (off_out)
1063                         return -ESPIPE;
1064                 if (off_in) {
1065                         if (in->f_op->llseek == no_llseek)
1066                                 return -EINVAL;
1067                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1068                                 return -EFAULT;
1069                         off = &offset;
1070                 } else
1071                         off = &in->f_pos;
1072
1073                 ret = do_splice_to(in, off, pipe, len, flags);
1074
1075                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1076                         ret = -EFAULT;
1077
1078                 return ret;
1079         }
1080
1081         return -EINVAL;
1082 }
1083
1084 /*
1085  * Map an iov into an array of pages and offset/length tupples. With the
1086  * partial_page structure, we can map several non-contiguous ranges into
1087  * our ones pages[] map instead of splitting that operation into pieces.
1088  * Could easily be exported as a generic helper for other users, in which
1089  * case one would probably want to add a 'max_nr_pages' parameter as well.
1090  */
1091 static int get_iovec_page_array(const struct iovec __user *iov,
1092                                 unsigned int nr_vecs, struct page **pages,
1093                                 struct partial_page *partial, int aligned)
1094 {
1095         int buffers = 0, error = 0;
1096
1097         /*
1098          * It's ok to take the mmap_sem for reading, even
1099          * across a "get_user()".
1100          */
1101         down_read(&current->mm->mmap_sem);
1102
1103         while (nr_vecs) {
1104                 unsigned long off, npages;
1105                 void __user *base;
1106                 size_t len;
1107                 int i;
1108
1109                 /*
1110                  * Get user address base and length for this iovec.
1111                  */
1112                 error = get_user(base, &iov->iov_base);
1113                 if (unlikely(error))
1114                         break;
1115                 error = get_user(len, &iov->iov_len);
1116                 if (unlikely(error))
1117                         break;
1118
1119                 /*
1120                  * Sanity check this iovec. 0 read succeeds.
1121                  */
1122                 if (unlikely(!len))
1123                         break;
1124                 error = -EFAULT;
1125                 if (unlikely(!base))
1126                         break;
1127
1128                 /*
1129                  * Get this base offset and number of pages, then map
1130                  * in the user pages.
1131                  */
1132                 off = (unsigned long) base & ~PAGE_MASK;
1133
1134                 /*
1135                  * If asked for alignment, the offset must be zero and the
1136                  * length a multiple of the PAGE_SIZE.
1137                  */
1138                 error = -EINVAL;
1139                 if (aligned && (off || len & ~PAGE_MASK))
1140                         break;
1141
1142                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1143                 if (npages > PIPE_BUFFERS - buffers)
1144                         npages = PIPE_BUFFERS - buffers;
1145
1146                 error = get_user_pages(current, current->mm,
1147                                        (unsigned long) base, npages, 0, 0,
1148                                        &pages[buffers], NULL);
1149
1150                 if (unlikely(error <= 0))
1151                         break;
1152
1153                 /*
1154                  * Fill this contiguous range into the partial page map.
1155                  */
1156                 for (i = 0; i < error; i++) {
1157                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1158
1159                         partial[buffers].offset = off;
1160                         partial[buffers].len = plen;
1161
1162                         off = 0;
1163                         len -= plen;
1164                         buffers++;
1165                 }
1166
1167                 /*
1168                  * We didn't complete this iov, stop here since it probably
1169                  * means we have to move some of this into a pipe to
1170                  * be able to continue.
1171                  */
1172                 if (len)
1173                         break;
1174
1175                 /*
1176                  * Don't continue if we mapped fewer pages than we asked for,
1177                  * or if we mapped the max number of pages that we have
1178                  * room for.
1179                  */
1180                 if (error < npages || buffers == PIPE_BUFFERS)
1181                         break;
1182
1183                 nr_vecs--;
1184                 iov++;
1185         }
1186
1187         up_read(&current->mm->mmap_sem);
1188
1189         if (buffers)
1190                 return buffers;
1191
1192         return error;
1193 }
1194
1195 /*
1196  * vmsplice splices a user address range into a pipe. It can be thought of
1197  * as splice-from-memory, where the regular splice is splice-from-file (or
1198  * to file). In both cases the output is a pipe, naturally.
1199  *
1200  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1201  * not the other way around. Splicing from user memory is a simple operation
1202  * that can be supported without any funky alignment restrictions or nasty
1203  * vm tricks. We simply map in the user memory and fill them into a pipe.
1204  * The reverse isn't quite as easy, though. There are two possible solutions
1205  * for that:
1206  *
1207  *      - memcpy() the data internally, at which point we might as well just
1208  *        do a regular read() on the buffer anyway.
1209  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1210  *        has restriction limitations on both ends of the pipe).
1211  *
1212  * Alas, it isn't here.
1213  *
1214  */
1215 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1216                         unsigned long nr_segs, unsigned int flags)
1217 {
1218         struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1219         struct page *pages[PIPE_BUFFERS];
1220         struct partial_page partial[PIPE_BUFFERS];
1221         struct splice_pipe_desc spd = {
1222                 .pages = pages,
1223                 .partial = partial,
1224                 .flags = flags,
1225                 .ops = &user_page_pipe_buf_ops,
1226         };
1227
1228         if (unlikely(!pipe))
1229                 return -EBADF;
1230         if (unlikely(nr_segs > UIO_MAXIOV))
1231                 return -EINVAL;
1232         else if (unlikely(!nr_segs))
1233                 return 0;
1234
1235         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1236                                             flags & SPLICE_F_GIFT);
1237         if (spd.nr_pages <= 0)
1238                 return spd.nr_pages;
1239
1240         return splice_to_pipe(pipe, &spd);
1241 }
1242
1243 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1244                              unsigned long nr_segs, unsigned int flags)
1245 {
1246         struct file *file;
1247         long error;
1248         int fput;
1249
1250         error = -EBADF;
1251         file = fget_light(fd, &fput);
1252         if (file) {
1253                 if (file->f_mode & FMODE_WRITE)
1254                         error = do_vmsplice(file, iov, nr_segs, flags);
1255
1256                 fput_light(file, fput);
1257         }
1258
1259         return error;
1260 }
1261
1262 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1263                            int fd_out, loff_t __user *off_out,
1264                            size_t len, unsigned int flags)
1265 {
1266         long error;
1267         struct file *in, *out;
1268         int fput_in, fput_out;
1269
1270         if (unlikely(!len))
1271                 return 0;
1272
1273         error = -EBADF;
1274         in = fget_light(fd_in, &fput_in);
1275         if (in) {
1276                 if (in->f_mode & FMODE_READ) {
1277                         out = fget_light(fd_out, &fput_out);
1278                         if (out) {
1279                                 if (out->f_mode & FMODE_WRITE)
1280                                         error = do_splice(in, off_in,
1281                                                           out, off_out,
1282                                                           len, flags);
1283                                 fput_light(out, fput_out);
1284                         }
1285                 }
1286
1287                 fput_light(in, fput_in);
1288         }
1289
1290         return error;
1291 }
1292
1293 /*
1294  * Link contents of ipipe to opipe.
1295  */
1296 static int link_pipe(struct pipe_inode_info *ipipe,
1297                      struct pipe_inode_info *opipe,
1298                      size_t len, unsigned int flags)
1299 {
1300         struct pipe_buffer *ibuf, *obuf;
1301         int ret, do_wakeup, i, ipipe_first;
1302
1303         ret = do_wakeup = ipipe_first = 0;
1304
1305         /*
1306          * Potential ABBA deadlock, work around it by ordering lock
1307          * grabbing by inode address. Otherwise two different processes
1308          * could deadlock (one doing tee from A -> B, the other from B -> A).
1309          */
1310         if (ipipe->inode < opipe->inode) {
1311                 ipipe_first = 1;
1312                 mutex_lock(&ipipe->inode->i_mutex);
1313                 mutex_lock(&opipe->inode->i_mutex);
1314         } else {
1315                 mutex_lock(&opipe->inode->i_mutex);
1316                 mutex_lock(&ipipe->inode->i_mutex);
1317         }
1318
1319         for (i = 0;; i++) {
1320                 if (!opipe->readers) {
1321                         send_sig(SIGPIPE, current, 0);
1322                         if (!ret)
1323                                 ret = -EPIPE;
1324                         break;
1325                 }
1326                 if (ipipe->nrbufs - i) {
1327                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1328
1329                         /*
1330                          * If we have room, fill this buffer
1331                          */
1332                         if (opipe->nrbufs < PIPE_BUFFERS) {
1333                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1334
1335                                 /*
1336                                  * Get a reference to this pipe buffer,
1337                                  * so we can copy the contents over.
1338                                  */
1339                                 ibuf->ops->get(ipipe, ibuf);
1340
1341                                 obuf = opipe->bufs + nbuf;
1342                                 *obuf = *ibuf;
1343
1344                                 /*
1345                                  * Don't inherit the gift flag, we need to
1346                                  * prevent multiple steals of this page.
1347                                  */
1348                                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1349
1350                                 if (obuf->len > len)
1351                                         obuf->len = len;
1352
1353                                 opipe->nrbufs++;
1354                                 do_wakeup = 1;
1355                                 ret += obuf->len;
1356                                 len -= obuf->len;
1357
1358                                 if (!len)
1359                                         break;
1360                                 if (opipe->nrbufs < PIPE_BUFFERS)
1361                                         continue;
1362                         }
1363
1364                         /*
1365                          * We have input available, but no output room.
1366                          * If we already copied data, return that. If we
1367                          * need to drop the opipe lock, it must be ordered
1368                          * last to avoid deadlocks.
1369                          */
1370                         if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
1371                                 if (!ret)
1372                                         ret = -EAGAIN;
1373                                 break;
1374                         }
1375                         if (signal_pending(current)) {
1376                                 if (!ret)
1377                                         ret = -ERESTARTSYS;
1378                                 break;
1379                         }
1380                         if (do_wakeup) {
1381                                 smp_mb();
1382                                 if (waitqueue_active(&opipe->wait))
1383                                         wake_up_interruptible(&opipe->wait);
1384                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1385                                 do_wakeup = 0;
1386                         }
1387
1388                         opipe->waiting_writers++;
1389                         pipe_wait(opipe);
1390                         opipe->waiting_writers--;
1391                         continue;
1392                 }
1393
1394                 /*
1395                  * No input buffers, do the usual checks for available
1396                  * writers and blocking and wait if necessary
1397                  */
1398                 if (!ipipe->writers)
1399                         break;
1400                 if (!ipipe->waiting_writers) {
1401                         if (ret)
1402                                 break;
1403                 }
1404                 /*
1405                  * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1406                  * with another process, we can only safely do that if
1407                  * the ipipe lock is ordered last.
1408                  */
1409                 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
1410                         if (!ret)
1411                                 ret = -EAGAIN;
1412                         break;
1413                 }
1414                 if (signal_pending(current)) {
1415                         if (!ret)
1416                                 ret = -ERESTARTSYS;
1417                         break;
1418                 }
1419
1420                 if (waitqueue_active(&ipipe->wait))
1421                         wake_up_interruptible_sync(&ipipe->wait);
1422                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1423
1424                 pipe_wait(ipipe);
1425         }
1426
1427         mutex_unlock(&ipipe->inode->i_mutex);
1428         mutex_unlock(&opipe->inode->i_mutex);
1429
1430         if (do_wakeup) {
1431                 smp_mb();
1432                 if (waitqueue_active(&opipe->wait))
1433                         wake_up_interruptible(&opipe->wait);
1434                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1435         }
1436
1437         return ret;
1438 }
1439
1440 /*
1441  * This is a tee(1) implementation that works on pipes. It doesn't copy
1442  * any data, it simply references the 'in' pages on the 'out' pipe.
1443  * The 'flags' used are the SPLICE_F_* variants, currently the only
1444  * applicable one is SPLICE_F_NONBLOCK.
1445  */
1446 static long do_tee(struct file *in, struct file *out, size_t len,
1447                    unsigned int flags)
1448 {
1449         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1450         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1451
1452         /*
1453          * Link ipipe to the two output pipes, consuming as we go along.
1454          */
1455         if (ipipe && opipe)
1456                 return link_pipe(ipipe, opipe, len, flags);
1457
1458         return -EINVAL;
1459 }
1460
1461 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1462 {
1463         struct file *in;
1464         int error, fput_in;
1465
1466         if (unlikely(!len))
1467                 return 0;
1468
1469         error = -EBADF;
1470         in = fget_light(fdin, &fput_in);
1471         if (in) {
1472                 if (in->f_mode & FMODE_READ) {
1473                         int fput_out;
1474                         struct file *out = fget_light(fdout, &fput_out);
1475
1476                         if (out) {
1477                                 if (out->f_mode & FMODE_WRITE)
1478                                         error = do_tee(in, out, len, flags);
1479                                 fput_light(out, fput_out);
1480                         }
1481                 }
1482                 fput_light(in, fput_in);
1483         }
1484
1485         return error;
1486 }