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@kernel.dk>
  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         const 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 *pipe,
  55                                      struct pipe_buffer *buf)
  56 {
  57         struct page *page = buf->page;
  58         struct address_space *mapping;
  59
  60         lock_page(page);
  61
  62         mapping = page_mapping(page);
  63         if (mapping) {
  64                 WARN_ON(!PageUptodate(page));
  65
  66                 /*
  67                  * At least for ext2 with nobh option, we need to wait on
  68                  * writeback completing on this page, since we'll remove it
  69                  * from the pagecache.  Otherwise truncate wont wait on the
  70                  * page, allowing the disk blocks to be reused by someone else
  71                  * before we actually wrote our data to them. fs corruption
  72                  * ensues.
  73                  */
  74                 wait_on_page_writeback(page);
  75
  76                 if (PagePrivate(page))
  77                         try_to_release_page(page, GFP_KERNEL);
  78
  79                 /*
  80                  * If we succeeded in removing the mapping, set LRU flag
  81                  * and return good.
  82                  */
  83                 if (remove_mapping(mapping, page)) {
  84                         buf->flags |= PIPE_BUF_FLAG_LRU;
  85                         return 0;
  86                 }
  87         }
  88
  89         /*
  90          * Raced with truncate or failed to remove page from current
  91          * address space, unlock and return failure.
  92          */
  93         unlock_page(page);
  94         return 1;
  95 }
  96
  97 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  98                                         struct pipe_buffer *buf)
  99 {
 100         page_cache_release(buf->page);
 101         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 102 }
 103
 104 static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
 105                                    struct pipe_buffer *buf)
 106 {
 107         struct page *page = buf->page;
 108         int err;
 109
 110         if (!PageUptodate(page)) {
 111                 lock_page(page);
 112
 113                 /*
 114                  * Page got truncated/unhashed. This will cause a 0-byte
 115                  * splice, if this is the first page.
 116                  */
 117                 if (!page->mapping) {
 118                         err = -ENODATA;
 119                         goto error;
 120                 }
 121
 122                 /*
 123                  * Uh oh, read-error from disk.
 124                  */
 125                 if (!PageUptodate(page)) {
 126                         err = -EIO;
 127                         goto error;
 128                 }
 129
 130                 /*
 131                  * Page is ok afterall, we are done.
 132                  */
 133                 unlock_page(page);
 134         }
 135
 136         return 0;
 137 error:
 138         unlock_page(page);
 139         return err;
 140 }
 141
 142 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 143         .can_merge = 0,
 144         .map = generic_pipe_buf_map,
 145         .unmap = generic_pipe_buf_unmap,
 146         .pin = page_cache_pipe_buf_pin,
 147         .release = page_cache_pipe_buf_release,
 148         .steal = page_cache_pipe_buf_steal,
 149         .get = generic_pipe_buf_get,
 150 };
 151
 152 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 153                                     struct pipe_buffer *buf)
 154 {
 155         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 156                 return 1;
 157
 158         buf->flags |= PIPE_BUF_FLAG_LRU;
 159         return generic_pipe_buf_steal(pipe, buf);
 160 }
 161
 162 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 163         .can_merge = 0,
 164         .map = generic_pipe_buf_map,
 165         .unmap = generic_pipe_buf_unmap,
 166         .pin = generic_pipe_buf_pin,
 167         .release = page_cache_pipe_buf_release,
 168         .steal = user_page_pipe_buf_steal,
 169         .get = generic_pipe_buf_get,
 170 };
 171
 172 /*
 173  * Pipe output worker. This sets up our pipe format with the page cache
 174  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 175  */
 176 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 177                               struct splice_pipe_desc *spd)
 178 {
 179         int ret, do_wakeup, page_nr;
 180
 181         ret = 0;
 182         do_wakeup = 0;
 183         page_nr = 0;
 184
 185         if (pipe->inode)
 186                 mutex_lock(&pipe->inode->i_mutex);
 187
 188         for (;;) {
 189                 if (!pipe->readers) {
 190                         send_sig(SIGPIPE, current, 0);
 191                         if (!ret)
 192                                 ret = -EPIPE;
 193                         break;
 194                 }
 195
 196                 if (pipe->nrbufs < PIPE_BUFFERS) {
 197                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 198                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 199
 200                         buf->page = spd->pages[page_nr];
 201                         buf->offset = spd->partial[page_nr].offset;
 202                         buf->len = spd->partial[page_nr].len;
 203                         buf->ops = spd->ops;
 204                         if (spd->flags & SPLICE_F_GIFT)
 205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 206
 207                         pipe->nrbufs++;
 208                         page_nr++;
 209                         ret += buf->len;
 210
 211                         if (pipe->inode)
 212                                 do_wakeup = 1;
 213
 214                         if (!--spd->nr_pages)
 215                                 break;
 216                         if (pipe->nrbufs < PIPE_BUFFERS)
 217                                 continue;
 218
 219                         break;
 220                 }
 221
 222                 if (spd->flags & SPLICE_F_NONBLOCK) {
 223                         if (!ret)
 224                                 ret = -EAGAIN;
 225                         break;
 226                 }
 227
 228                 if (signal_pending(current)) {
 229                         if (!ret)
 230                                 ret = -ERESTARTSYS;
 231                         break;
 232                 }
 233
 234                 if (do_wakeup) {
 235                         smp_mb();
 236                         if (waitqueue_active(&pipe->wait))
 237                                 wake_up_interruptible_sync(&pipe->wait);
 238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 239                         do_wakeup = 0;
 240                 }
 241
 242                 pipe->waiting_writers++;
 243                 pipe_wait(pipe);
 244                 pipe->waiting_writers--;
 245         }
 246
 247         if (pipe->inode)
 248                 mutex_unlock(&pipe->inode->i_mutex);
 249
 250         if (do_wakeup) {
 251                 smp_mb();
 252                 if (waitqueue_active(&pipe->wait))
 253                         wake_up_interruptible(&pipe->wait);
 254                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 255         }
 256
 257         while (page_nr < spd->nr_pages)
 258                 page_cache_release(spd->pages[page_nr++]);
 259
 260         return ret;
 261 }
 262
 263 static int
 264 __generic_file_splice_read(struct file *in, loff_t *ppos,
 265                            struct pipe_inode_info *pipe, size_t len,
 266                            unsigned int flags)
 267 {
 268         struct address_space *mapping = in->f_mapping;
 269         unsigned int loff, nr_pages;
 270         struct page *pages[PIPE_BUFFERS];
 271         struct partial_page partial[PIPE_BUFFERS];
 272         struct page *page;
 273         pgoff_t index, end_index;
 274         loff_t isize;
 275         size_t total_len;
 276         int error, page_nr;
 277         struct splice_pipe_desc spd = {
 278                 .pages = pages,
 279                 .partial = partial,
 280                 .flags = flags,
 281                 .ops = &page_cache_pipe_buf_ops,
 282         };
 283
 284         index = *ppos >> PAGE_CACHE_SHIFT;
 285         loff = *ppos & ~PAGE_CACHE_MASK;
 286         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 287
 288         if (nr_pages > PIPE_BUFFERS)
 289                 nr_pages = PIPE_BUFFERS;
 290
 291         /*
 292          * Initiate read-ahead on this page range. however, don't call into
 293          * read-ahead if this is a non-zero offset (we are likely doing small
 294          * chunk splice and the page is already there) for a single page.
 295          */
 296         if (!loff || nr_pages > 1)
 297                 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 298
 299         /*
 300          * Now fill in the holes:
 301          */
 302         error = 0;
 303         total_len = 0;
 304
 305         /*
 306          * Lookup the (hopefully) full range of pages we need.
 307          */
 308         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 309
 310         /*
 311          * If find_get_pages_contig() returned fewer pages than we needed,
 312          * allocate the rest.
 313          */
 314         index += spd.nr_pages;
 315         while (spd.nr_pages < nr_pages) {
 316                 /*
 317                  * Page could be there, find_get_pages_contig() breaks on
 318                  * the first hole.
 319                  */
 320                 page = find_get_page(mapping, index);
 321                 if (!page) {
 322                         /*
 323                          * Make sure the read-ahead engine is notified
 324                          * about this failure.
 325                          */
 326                         handle_ra_miss(mapping, &in->f_ra, index);
 327
 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                                               GFP_KERNEL);
 337                         if (unlikely(error)) {
 338                                 page_cache_release(page);
 339                                 if (error == -EEXIST)
 340                                         continue;
 341                                 break;
 342                         }
 343                         /*
 344                          * add_to_page_cache() locks the page, unlock it
 345                          * to avoid convoluting the logic below even more.
 346                          */
 347                         unlock_page(page);
 348                 }
 349
 350                 pages[spd.nr_pages++] = page;
 351                 index++;
 352         }
 353
 354         /*
 355          * Now loop over the map and see if we need to start IO on any
 356          * pages, fill in the partial map, etc.
 357          */
 358         index = *ppos >> PAGE_CACHE_SHIFT;
 359         nr_pages = spd.nr_pages;
 360         spd.nr_pages = 0;
 361         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 362                 unsigned int this_len;
 363
 364                 if (!len)
 365                         break;
 366
 367                 /*
 368                  * this_len is the max we'll use from this page
 369                  */
 370                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 371                 page = pages[page_nr];
 372
 373                 /*
 374                  * If the page isn't uptodate, we may need to start io on it
 375                  */
 376                 if (!PageUptodate(page)) {
 377                         /*
 378                          * If in nonblock mode then dont block on waiting
 379                          * for an in-flight io page
 380                          */
 381                         if (flags & SPLICE_F_NONBLOCK)
 382                                 break;
 383
 384                         lock_page(page);
 385
 386                         /*
 387                          * page was truncated, stop here. if this isn't the
 388                          * first page, we'll just complete what we already
 389                          * added
 390                          */
 391                         if (!page->mapping) {
 392                                 unlock_page(page);
 393                                 break;
 394                         }
 395                         /*
 396                          * page was already under io and is now done, great
 397                          */
 398                         if (PageUptodate(page)) {
 399                                 unlock_page(page);
 400                                 goto fill_it;
 401                         }
 402
 403                         /*
 404                          * need to read in the page
 405                          */
 406                         error = mapping->a_ops->readpage(in, page);
 407                         if (unlikely(error)) {
 408                                 /*
 409                                  * We really should re-lookup the page here,
 410                                  * but it complicates things a lot. Instead
 411                                  * lets just do what we already stored, and
 412                                  * we'll get it the next time we are called.
 413                                  */
 414                                 if (error == AOP_TRUNCATED_PAGE)
 415                                         error = 0;
 416
 417                                 break;
 418                         }
 419
 420                         /*
 421                          * i_size must be checked after ->readpage().
 422                          */
 423                         isize = i_size_read(mapping->host);
 424                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 425                         if (unlikely(!isize || index > end_index))
 426                                 break;
 427
 428                         /*
 429                          * if this is the last page, see if we need to shrink
 430                          * the length and stop
 431                          */
 432                         if (end_index == index) {
 433                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 434                                 if (total_len + loff > isize)
 435                                         break;
 436                                 /*
 437                                  * force quit after adding this page
 438                                  */
 439                                 len = this_len;
 440                                 this_len = min(this_len, loff);
 441                                 loff = 0;
 442                         }
 443                 }
 444 fill_it:
 445                 partial[page_nr].offset = loff;
 446                 partial[page_nr].len = this_len;
 447                 len -= this_len;
 448                 total_len += this_len;
 449                 loff = 0;
 450                 spd.nr_pages++;
 451                 index++;
 452         }
 453
 454         /*
 455          * Release any pages at the end, if we quit early. 'i' is how far
 456          * we got, 'nr_pages' is how many pages are in the map.
 457          */
 458         while (page_nr < nr_pages)
 459                 page_cache_release(pages[page_nr++]);
 460
 461         if (spd.nr_pages)
 462                 return splice_to_pipe(pipe, &spd);
 463
 464         return error;
 465 }
 466
 467 /**
 468  * generic_file_splice_read - splice data from file to a pipe
 469  * @in:         file to splice from
 470  * @pipe:       pipe to splice to
 471  * @len:        number of bytes to splice
 472  * @flags:      splice modifier flags
 473  *
 474  * Will read pages from given file and fill them into a pipe.
 475  */
 476 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 477                                  struct pipe_inode_info *pipe, size_t len,
 478                                  unsigned int flags)
 479 {
 480         ssize_t spliced;
 481         int ret;
 482
 483         ret = 0;
 484         spliced = 0;
 485
 486         while (len) {
 487                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 488
 489                 if (ret < 0)
 490                         break;
 491                 else if (!ret) {
 492                         if (spliced)
 493                                 break;
 494                         if (flags & SPLICE_F_NONBLOCK) {
 495                                 ret = -EAGAIN;
 496                                 break;
 497                         }
 498                 }
 499
 500                 *ppos += ret;
 501                 len -= ret;
 502                 spliced += ret;
 503         }
 504
 505         if (spliced)
 506                 return spliced;
 507
 508         return ret;
 509 }
 510
 511 EXPORT_SYMBOL(generic_file_splice_read);
 512
 513 /*
 514  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 515  * using sendpage(). Return the number of bytes sent.
 516  */
 517 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 518                             struct pipe_buffer *buf, struct splice_desc *sd)
 519 {
 520         struct file *file = sd->file;
 521         loff_t pos = sd->pos;
 522         int ret, more;
 523
 524         ret = buf->ops->pin(pipe, buf);
 525         if (!ret) {
 526                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 527
 528                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 529                                            sd->len, &pos, more);
 530         }
 531
 532         return ret;
 533 }
 534
 535 /*
 536  * This is a little more tricky than the file -> pipe splicing. There are
 537  * basically three cases:
 538  *
 539  *      - Destination page already exists in the address space and there
 540  *        are users of it. For that case we have no other option that
 541  *        copying the data. Tough luck.
 542  *      - Destination page already exists in the address space, but there
 543  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 544  *        through to last case.
 545  *      - Destination page does not exist, we can add the pipe page to
 546  *        the page cache and avoid the copy.
 547  *
 548  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 549  * sd->flags), we attempt to migrate pages from the pipe to the output
 550  * file address space page cache. This is possible if no one else has
 551  * the pipe page referenced outside of the pipe and page cache. If
 552  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 553  * a new page in the output file page cache and fill/dirty that.
 554  */
 555 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 556                         struct splice_desc *sd)
 557 {
 558         struct file *file = sd->file;
 559         struct address_space *mapping = file->f_mapping;
 560         unsigned int offset, this_len;
 561         struct page *page;
 562         pgoff_t index;
 563         int ret;
 564
 565         /*
 566          * make sure the data in this buffer is uptodate
 567          */
 568         ret = buf->ops->pin(pipe, buf);
 569         if (unlikely(ret))
 570                 return ret;
 571
 572         index = sd->pos >> PAGE_CACHE_SHIFT;
 573         offset = sd->pos & ~PAGE_CACHE_MASK;
 574
 575         this_len = sd->len;
 576         if (this_len + offset > PAGE_CACHE_SIZE)
 577                 this_len = PAGE_CACHE_SIZE - offset;
 578
 579 find_page:
 580         page = find_lock_page(mapping, index);
 581         if (!page) {
 582                 ret = -ENOMEM;
 583                 page = page_cache_alloc_cold(mapping);
 584                 if (unlikely(!page))
 585                         goto out_ret;
 586
 587                 /*
 588                  * This will also lock the page
 589                  */
 590                 ret = add_to_page_cache_lru(page, mapping, index,
 591                                             GFP_KERNEL);
 592                 if (unlikely(ret))
 593                         goto out;
 594         }
 595
 596         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 597         if (unlikely(ret)) {
 598                 loff_t isize = i_size_read(mapping->host);
 599
 600                 if (ret != AOP_TRUNCATED_PAGE)
 601                         unlock_page(page);
 602                 page_cache_release(page);
 603                 if (ret == AOP_TRUNCATED_PAGE)
 604                         goto find_page;
 605
 606                 /*
 607                  * prepare_write() may have instantiated a few blocks
 608                  * outside i_size.  Trim these off again.
 609                  */
 610                 if (sd->pos + this_len > isize)
 611                         vmtruncate(mapping->host, isize);
 612
 613                 goto out_ret;
 614         }
 615
 616         if (buf->page != page) {
 617                 /*
 618                  * Careful, ->map() uses KM_USER0!
 619                  */
 620                 char *src = buf->ops->map(pipe, buf, 1);
 621                 char *dst = kmap_atomic(page, KM_USER1);
 622
 623                 memcpy(dst + offset, src + buf->offset, this_len);
 624                 flush_dcache_page(page);
 625                 kunmap_atomic(dst, KM_USER1);
 626                 buf->ops->unmap(pipe, buf, src);
 627         }
 628
 629         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 630         if (!ret) {
 631                 /*
 632                  * Return the number of bytes written and mark page as
 633                  * accessed, we are now done!
 634                  */
 635                 ret = this_len;
 636                 mark_page_accessed(page);
 637                 balance_dirty_pages_ratelimited(mapping);
 638         } else if (ret == AOP_TRUNCATED_PAGE) {
 639                 page_cache_release(page);
 640                 goto find_page;
 641         }
 642 out:
 643         page_cache_release(page);
 644         unlock_page(page);
 645 out_ret:
 646         return ret;
 647 }
 648
 649 /*
 650  * Pipe input worker. Most of this logic works like a regular pipe, the
 651  * key here is the 'actor' worker passed in that actually moves the data
 652  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 653  */
 654 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 655                            struct file *out, loff_t *ppos, size_t len,
 656                            unsigned int flags, splice_actor *actor)
 657 {
 658         int ret, do_wakeup, err;
 659         struct splice_desc sd;
 660
 661         ret = 0;
 662         do_wakeup = 0;
 663
 664         sd.total_len = len;
 665         sd.flags = flags;
 666         sd.file = out;
 667         sd.pos = *ppos;
 668
 669         for (;;) {
 670                 if (pipe->nrbufs) {
 671                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 672                         const struct pipe_buf_operations *ops = buf->ops;
 673
 674                         sd.len = buf->len;
 675                         if (sd.len > sd.total_len)
 676                                 sd.len = sd.total_len;
 677
 678                         err = actor(pipe, buf, &sd);
 679                         if (err <= 0) {
 680                                 if (!ret && err != -ENODATA)
 681                                         ret = err;
 682
 683                                 break;
 684                         }
 685
 686                         ret += err;
 687                         buf->offset += err;
 688                         buf->len -= err;
 689
 690                         sd.len -= err;
 691                         sd.pos += err;
 692                         sd.total_len -= err;
 693                         if (sd.len)
 694                                 continue;
 695
 696                         if (!buf->len) {
 697                                 buf->ops = NULL;
 698                                 ops->release(pipe, buf);
 699                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 700                                 pipe->nrbufs--;
 701                                 if (pipe->inode)
 702                                         do_wakeup = 1;
 703                         }
 704
 705                         if (!sd.total_len)
 706                                 break;
 707                 }
 708
 709                 if (pipe->nrbufs)
 710                         continue;
 711                 if (!pipe->writers)
 712                         break;
 713                 if (!pipe->waiting_writers) {
 714                         if (ret)
 715                                 break;
 716                 }
 717
 718                 if (flags & SPLICE_F_NONBLOCK) {
 719                         if (!ret)
 720                                 ret = -EAGAIN;
 721                         break;
 722                 }
 723
 724                 if (signal_pending(current)) {
 725                         if (!ret)
 726                                 ret = -ERESTARTSYS;
 727                         break;
 728                 }
 729
 730                 if (do_wakeup) {
 731                         smp_mb();
 732                         if (waitqueue_active(&pipe->wait))
 733                                 wake_up_interruptible_sync(&pipe->wait);
 734                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 735                         do_wakeup = 0;
 736                 }
 737
 738                 pipe_wait(pipe);
 739         }
 740
 741         if (do_wakeup) {
 742                 smp_mb();
 743                 if (waitqueue_active(&pipe->wait))
 744                         wake_up_interruptible(&pipe->wait);
 745                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 746         }
 747
 748         return ret;
 749 }
 750 EXPORT_SYMBOL(__splice_from_pipe);
 751
 752 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 753                          loff_t *ppos, size_t len, unsigned int flags,
 754                          splice_actor *actor)
 755 {
 756         ssize_t ret;
 757         struct inode *inode = out->f_mapping->host;
 758
 759         /*
 760          * The actor worker might be calling ->prepare_write and
 761          * ->commit_write. Most of the time, these expect i_mutex to
 762          * be held. Since this may result in an ABBA deadlock with
 763          * pipe->inode, we have to order lock acquiry here.
 764          */
 765         inode_double_lock(inode, pipe->inode);
 766         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 767         inode_double_unlock(inode, pipe->inode);
 768
 769         return ret;
 770 }
 771
 772 /**
 773  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 774  * @pipe:       pipe info
 775  * @out:        file to write to
 776  * @len:        number of bytes to splice
 777  * @flags:      splice modifier flags
 778  *
 779  * Will either move or copy pages (determined by @flags options) from
 780  * the given pipe inode to the given file. The caller is responsible
 781  * for acquiring i_mutex on both inodes.
 782  *
 783  */
 784 ssize_t
 785 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 786                                  loff_t *ppos, size_t len, unsigned int flags)
 787 {
 788         struct address_space *mapping = out->f_mapping;
 789         struct inode *inode = mapping->host;
 790         ssize_t ret;
 791         int err;
 792
 793         err = remove_suid(out->f_path.dentry);
 794         if (unlikely(err))
 795                 return err;
 796
 797         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 798         if (ret > 0) {
 799                 *ppos += ret;
 800
 801                 /*
 802                  * If file or inode is SYNC and we actually wrote some data,
 803                  * sync it.
 804                  */
 805                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 806                         err = generic_osync_inode(inode, mapping,
 807                                                   OSYNC_METADATA|OSYNC_DATA);
 808
 809                         if (err)
 810                                 ret = err;
 811                 }
 812         }
 813
 814         return ret;
 815 }
 816
 817 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 818
 819 /**
 820  * generic_file_splice_write - splice data from a pipe to a file
 821  * @pipe:       pipe info
 822  * @out:        file to write to
 823  * @len:        number of bytes to splice
 824  * @flags:      splice modifier flags
 825  *
 826  * Will either move or copy pages (determined by @flags options) from
 827  * the given pipe inode to the given file.
 828  *
 829  */
 830 ssize_t
 831 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 832                           loff_t *ppos, size_t len, unsigned int flags)
 833 {
 834         struct address_space *mapping = out->f_mapping;
 835         struct inode *inode = mapping->host;
 836         ssize_t ret;
 837         int err;
 838
 839         err = should_remove_suid(out->f_path.dentry);
 840         if (unlikely(err)) {
 841                 mutex_lock(&inode->i_mutex);
 842                 err = __remove_suid(out->f_path.dentry, err);
 843                 mutex_unlock(&inode->i_mutex);
 844                 if (err)
 845                         return err;
 846         }
 847
 848         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 849         if (ret > 0) {
 850                 *ppos += ret;
 851
 852                 /*
 853                  * If file or inode is SYNC and we actually wrote some data,
 854                  * sync it.
 855                  */
 856                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 857                         mutex_lock(&inode->i_mutex);
 858                         err = generic_osync_inode(inode, mapping,
 859                                                   OSYNC_METADATA|OSYNC_DATA);
 860                         mutex_unlock(&inode->i_mutex);
 861
 862                         if (err)
 863                                 ret = err;
 864                 }
 865         }
 866
 867         return ret;
 868 }
 869
 870 EXPORT_SYMBOL(generic_file_splice_write);
 871
 872 /**
 873  * generic_splice_sendpage - splice data from a pipe to a socket
 874  * @inode:      pipe inode
 875  * @out:        socket to write to
 876  * @len:        number of bytes to splice
 877  * @flags:      splice modifier flags
 878  *
 879  * Will send @len bytes from the pipe to a network socket. No data copying
 880  * is involved.
 881  *
 882  */
 883 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 884                                 loff_t *ppos, size_t len, unsigned int flags)
 885 {
 886         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 887 }
 888
 889 EXPORT_SYMBOL(generic_splice_sendpage);
 890
 891 /*
 892  * Attempt to initiate a splice from pipe to file.
 893  */
 894 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 895                            loff_t *ppos, size_t len, unsigned int flags)
 896 {
 897         int ret;
 898
 899         if (unlikely(!out->f_op || !out->f_op->splice_write))
 900                 return -EINVAL;
 901
 902         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 903                 return -EBADF;
 904
 905         ret = rw_verify_area(WRITE, out, ppos, len);
 906         if (unlikely(ret < 0))
 907                 return ret;
 908
 909         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 910 }
 911
 912 /*
 913  * Attempt to initiate a splice from a file to a pipe.
 914  */
 915 static long do_splice_to(struct file *in, loff_t *ppos,
 916                          struct pipe_inode_info *pipe, size_t len,
 917                          unsigned int flags)
 918 {
 919         loff_t isize, left;
 920         int ret;
 921
 922         if (unlikely(!in->f_op || !in->f_op->splice_read))
 923                 return -EINVAL;
 924
 925         if (unlikely(!(in->f_mode & FMODE_READ)))
 926                 return -EBADF;
 927
 928         ret = rw_verify_area(READ, in, ppos, len);
 929         if (unlikely(ret < 0))
 930                 return ret;
 931
 932         isize = i_size_read(in->f_mapping->host);
 933         if (unlikely(*ppos >= isize))
 934                 return 0;
 935
 936         left = isize - *ppos;
 937         if (unlikely(left < len))
 938                 len = left;
 939
 940         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 941 }
 942
 943 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 944                       size_t len, unsigned int flags)
 945 {
 946         struct pipe_inode_info *pipe;
 947         long ret, bytes;
 948         loff_t out_off;
 949         umode_t i_mode;
 950         int i;
 951
 952         /*
 953          * We require the input being a regular file, as we don't want to
 954          * randomly drop data for eg socket -> socket splicing. Use the
 955          * piped splicing for that!
 956          */
 957         i_mode = in->f_path.dentry->d_inode->i_mode;
 958         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 959                 return -EINVAL;
 960
 961         /*
 962          * neither in nor out is a pipe, setup an internal pipe attached to
 963          * 'out' and transfer the wanted data from 'in' to 'out' through that
 964          */
 965         pipe = current->splice_pipe;
 966         if (unlikely(!pipe)) {
 967                 pipe = alloc_pipe_info(NULL);
 968                 if (!pipe)
 969                         return -ENOMEM;
 970
 971                 /*
 972                  * We don't have an immediate reader, but we'll read the stuff
 973                  * out of the pipe right after the splice_to_pipe(). So set
 974                  * PIPE_READERS appropriately.
 975                  */
 976                 pipe->readers = 1;
 977
 978                 current->splice_pipe = pipe;
 979         }
 980
 981         /*
 982          * Do the splice.
 983          */
 984         ret = 0;
 985         bytes = 0;
 986         out_off = 0;
 987
 988         while (len) {
 989                 size_t read_len, max_read_len;
 990
 991                 /*
 992                  * Do at most PIPE_BUFFERS pages worth of transfer:
 993                  */
 994                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 995
 996                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 997                 if (unlikely(ret < 0))
 998                         goto out_release;
 999
1000                 read_len = ret;
1001
1002                 /*
1003                  * NOTE: nonblocking mode only applies to the input. We
1004                  * must not do the output in nonblocking mode as then we
1005                  * could get stuck data in the internal pipe:
1006                  */
1007                 ret = do_splice_from(pipe, out, &out_off, read_len,
1008                                      flags & ~SPLICE_F_NONBLOCK);
1009                 if (unlikely(ret < 0))
1010                         goto out_release;
1011
1012                 bytes += ret;
1013                 len -= ret;
1014
1015                 /*
1016                  * In nonblocking mode, if we got back a short read then
1017                  * that was due to either an IO error or due to the
1018                  * pagecache entry not being there. In the IO error case
1019                  * the _next_ splice attempt will produce a clean IO error
1020                  * return value (not a short read), so in both cases it's
1021                  * correct to break out of the loop here:
1022                  */
1023                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1024                         break;
1025         }
1026
1027         pipe->nrbufs = pipe->curbuf = 0;
1028
1029         return bytes;
1030
1031 out_release:
1032         /*
1033          * If we did an incomplete transfer we must release
1034          * the pipe buffers in question:
1035          */
1036         for (i = 0; i < PIPE_BUFFERS; i++) {
1037                 struct pipe_buffer *buf = pipe->bufs + i;
1038
1039                 if (buf->ops) {
1040                         buf->ops->release(pipe, buf);
1041                         buf->ops = NULL;
1042                 }
1043         }
1044         pipe->nrbufs = pipe->curbuf = 0;
1045
1046         /*
1047          * If we transferred some data, return the number of bytes:
1048          */
1049         if (bytes > 0)
1050                 return bytes;
1051
1052         return ret;
1053 }
1054
1055 EXPORT_SYMBOL(do_splice_direct);
1056
1057 /*
1058  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1059  * location, so checking ->i_pipe is not enough to verify that this is a
1060  * pipe.
1061  */
1062 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1063 {
1064         if (S_ISFIFO(inode->i_mode))
1065                 return inode->i_pipe;
1066
1067         return NULL;
1068 }
1069
1070 /*
1071  * Determine where to splice to/from.
1072  */
1073 static long do_splice(struct file *in, loff_t __user *off_in,
1074                       struct file *out, loff_t __user *off_out,
1075                       size_t len, unsigned int flags)
1076 {
1077         struct pipe_inode_info *pipe;
1078         loff_t offset, *off;
1079         long ret;
1080
1081         pipe = pipe_info(in->f_path.dentry->d_inode);
1082         if (pipe) {
1083                 if (off_in)
1084                         return -ESPIPE;
1085                 if (off_out) {
1086                         if (out->f_op->llseek == no_llseek)
1087                                 return -EINVAL;
1088                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1089                                 return -EFAULT;
1090                         off = &offset;
1091                 } else
1092                         off = &out->f_pos;
1093
1094                 ret = do_splice_from(pipe, out, off, len, flags);
1095
1096                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1097                         ret = -EFAULT;
1098
1099                 return ret;
1100         }
1101
1102         pipe = pipe_info(out->f_path.dentry->d_inode);
1103         if (pipe) {
1104                 if (off_out)
1105                         return -ESPIPE;
1106                 if (off_in) {
1107                         if (in->f_op->llseek == no_llseek)
1108                                 return -EINVAL;
1109                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1110                                 return -EFAULT;
1111                         off = &offset;
1112                 } else
1113                         off = &in->f_pos;
1114
1115                 ret = do_splice_to(in, off, pipe, len, flags);
1116
1117                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1118                         ret = -EFAULT;
1119
1120                 return ret;
1121         }
1122
1123         return -EINVAL;
1124 }
1125
1126 /*
1127  * Map an iov into an array of pages and offset/length tupples. With the
1128  * partial_page structure, we can map several non-contiguous ranges into
1129  * our ones pages[] map instead of splitting that operation into pieces.
1130  * Could easily be exported as a generic helper for other users, in which
1131  * case one would probably want to add a 'max_nr_pages' parameter as well.
1132  */
1133 static int get_iovec_page_array(const struct iovec __user *iov,
1134                                 unsigned int nr_vecs, struct page **pages,
1135                                 struct partial_page *partial, int aligned)
1136 {
1137         int buffers = 0, error = 0;
1138
1139         /*
1140          * It's ok to take the mmap_sem for reading, even
1141          * across a "get_user()".
1142          */
1143         down_read(&current->mm->mmap_sem);
1144
1145         while (nr_vecs) {
1146                 unsigned long off, npages;
1147                 void __user *base;
1148                 size_t len;
1149                 int i;
1150
1151                 /*
1152                  * Get user address base and length for this iovec.
1153                  */
1154                 error = get_user(base, &iov->iov_base);
1155                 if (unlikely(error))
1156                         break;
1157                 error = get_user(len, &iov->iov_len);
1158                 if (unlikely(error))
1159                         break;
1160
1161                 /*
1162                  * Sanity check this iovec. 0 read succeeds.
1163                  */
1164                 if (unlikely(!len))
1165                         break;
1166                 error = -EFAULT;
1167                 if (unlikely(!base))
1168                         break;
1169
1170                 /*
1171                  * Get this base offset and number of pages, then map
1172                  * in the user pages.
1173                  */
1174                 off = (unsigned long) base & ~PAGE_MASK;
1175
1176                 /*
1177                  * If asked for alignment, the offset must be zero and the
1178                  * length a multiple of the PAGE_SIZE.
1179                  */
1180                 error = -EINVAL;
1181                 if (aligned && (off || len & ~PAGE_MASK))
1182                         break;
1183
1184                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1185                 if (npages > PIPE_BUFFERS - buffers)
1186                         npages = PIPE_BUFFERS - buffers;
1187
1188                 error = get_user_pages(current, current->mm,
1189                                        (unsigned long) base, npages, 0, 0,
1190                                        &pages[buffers], NULL);
1191
1192                 if (unlikely(error <= 0))
1193                         break;
1194
1195                 /*
1196                  * Fill this contiguous range into the partial page map.
1197                  */
1198                 for (i = 0; i < error; i++) {
1199                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1200
1201                         partial[buffers].offset = off;
1202                         partial[buffers].len = plen;
1203
1204                         off = 0;
1205                         len -= plen;
1206                         buffers++;
1207                 }
1208
1209                 /*
1210                  * We didn't complete this iov, stop here since it probably
1211                  * means we have to move some of this into a pipe to
1212                  * be able to continue.
1213                  */
1214                 if (len)
1215                         break;
1216
1217                 /*
1218                  * Don't continue if we mapped fewer pages than we asked for,
1219                  * or if we mapped the max number of pages that we have
1220                  * room for.
1221                  */
1222                 if (error < npages || buffers == PIPE_BUFFERS)
1223                         break;
1224
1225                 nr_vecs--;
1226                 iov++;
1227         }
1228
1229         up_read(&current->mm->mmap_sem);
1230
1231         if (buffers)
1232                 return buffers;
1233
1234         return error;
1235 }
1236
1237 /*
1238  * vmsplice splices a user address range into a pipe. It can be thought of
1239  * as splice-from-memory, where the regular splice is splice-from-file (or
1240  * to file). In both cases the output is a pipe, naturally.
1241  *
1242  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1243  * not the other way around. Splicing from user memory is a simple operation
1244  * that can be supported without any funky alignment restrictions or nasty
1245  * vm tricks. We simply map in the user memory and fill them into a pipe.
1246  * The reverse isn't quite as easy, though. There are two possible solutions
1247  * for that:
1248  *
1249  *      - memcpy() the data internally, at which point we might as well just
1250  *        do a regular read() on the buffer anyway.
1251  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1252  *        has restriction limitations on both ends of the pipe).
1253  *
1254  * Alas, it isn't here.
1255  *
1256  */
1257 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1258                         unsigned long nr_segs, unsigned int flags)
1259 {
1260         struct pipe_inode_info *pipe;
1261         struct page *pages[PIPE_BUFFERS];
1262         struct partial_page partial[PIPE_BUFFERS];
1263         struct splice_pipe_desc spd = {
1264                 .pages = pages,
1265                 .partial = partial,
1266                 .flags = flags,
1267                 .ops = &user_page_pipe_buf_ops,
1268         };
1269
1270         pipe = pipe_info(file->f_path.dentry->d_inode);
1271         if (!pipe)
1272                 return -EBADF;
1273         if (unlikely(nr_segs > UIO_MAXIOV))
1274                 return -EINVAL;
1275         else if (unlikely(!nr_segs))
1276                 return 0;
1277
1278         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1279                                             flags & SPLICE_F_GIFT);
1280         if (spd.nr_pages <= 0)
1281                 return spd.nr_pages;
1282
1283         return splice_to_pipe(pipe, &spd);
1284 }
1285
1286 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1287                              unsigned long nr_segs, unsigned int flags)
1288 {
1289         struct file *file;
1290         long error;
1291         int fput;
1292
1293         error = -EBADF;
1294         file = fget_light(fd, &fput);
1295         if (file) {
1296                 if (file->f_mode & FMODE_WRITE)
1297                         error = do_vmsplice(file, iov, nr_segs, flags);
1298
1299                 fput_light(file, fput);
1300         }
1301
1302         return error;
1303 }
1304
1305 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1306                            int fd_out, loff_t __user *off_out,
1307                            size_t len, unsigned int flags)
1308 {
1309         long error;
1310         struct file *in, *out;
1311         int fput_in, fput_out;
1312
1313         if (unlikely(!len))
1314                 return 0;
1315
1316         error = -EBADF;
1317         in = fget_light(fd_in, &fput_in);
1318         if (in) {
1319                 if (in->f_mode & FMODE_READ) {
1320                         out = fget_light(fd_out, &fput_out);
1321                         if (out) {
1322                                 if (out->f_mode & FMODE_WRITE)
1323                                         error = do_splice(in, off_in,
1324                                                           out, off_out,
1325                                                           len, flags);
1326                                 fput_light(out, fput_out);
1327                         }
1328                 }
1329
1330                 fput_light(in, fput_in);
1331         }
1332
1333         return error;
1334 }
1335
1336 /*
1337  * Make sure there's data to read. Wait for input if we can, otherwise
1338  * return an appropriate error.
1339  */
1340 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1341 {
1342         int ret;
1343
1344         /*
1345          * Check ->nrbufs without the inode lock first. This function
1346          * is speculative anyways, so missing one is ok.
1347          */
1348         if (pipe->nrbufs)
1349                 return 0;
1350
1351         ret = 0;
1352         mutex_lock(&pipe->inode->i_mutex);
1353
1354         while (!pipe->nrbufs) {
1355                 if (signal_pending(current)) {
1356                         ret = -ERESTARTSYS;
1357                         break;
1358                 }
1359                 if (!pipe->writers)
1360                         break;
1361                 if (!pipe->waiting_writers) {
1362                         if (flags & SPLICE_F_NONBLOCK) {
1363                                 ret = -EAGAIN;
1364                                 break;
1365                         }
1366                 }
1367                 pipe_wait(pipe);
1368         }
1369
1370         mutex_unlock(&pipe->inode->i_mutex);
1371         return ret;
1372 }
1373
1374 /*
1375  * Make sure there's writeable room. Wait for room if we can, otherwise
1376  * return an appropriate error.
1377  */
1378 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1379 {
1380         int ret;
1381
1382         /*
1383          * Check ->nrbufs without the inode lock first. This function
1384          * is speculative anyways, so missing one is ok.
1385          */
1386         if (pipe->nrbufs < PIPE_BUFFERS)
1387                 return 0;
1388
1389         ret = 0;
1390         mutex_lock(&pipe->inode->i_mutex);
1391
1392         while (pipe->nrbufs >= PIPE_BUFFERS) {
1393                 if (!pipe->readers) {
1394                         send_sig(SIGPIPE, current, 0);
1395                         ret = -EPIPE;
1396                         break;
1397                 }
1398                 if (flags & SPLICE_F_NONBLOCK) {
1399                         ret = -EAGAIN;
1400                         break;
1401                 }
1402                 if (signal_pending(current)) {
1403                         ret = -ERESTARTSYS;
1404                         break;
1405                 }
1406                 pipe->waiting_writers++;
1407                 pipe_wait(pipe);
1408                 pipe->waiting_writers--;
1409         }
1410
1411         mutex_unlock(&pipe->inode->i_mutex);
1412         return ret;
1413 }
1414
1415 /*
1416  * Link contents of ipipe to opipe.
1417  */
1418 static int link_pipe(struct pipe_inode_info *ipipe,
1419                      struct pipe_inode_info *opipe,
1420                      size_t len, unsigned int flags)
1421 {
1422         struct pipe_buffer *ibuf, *obuf;
1423         int ret = 0, i = 0, nbuf;
1424
1425         /*
1426          * Potential ABBA deadlock, work around it by ordering lock
1427          * grabbing by inode address. Otherwise two different processes
1428          * could deadlock (one doing tee from A -> B, the other from B -> A).
1429          */
1430         inode_double_lock(ipipe->inode, opipe->inode);
1431
1432         do {
1433                 if (!opipe->readers) {
1434                         send_sig(SIGPIPE, current, 0);
1435                         if (!ret)
1436                                 ret = -EPIPE;
1437                         break;
1438                 }
1439
1440                 /*
1441                  * If we have iterated all input buffers or ran out of
1442                  * output room, break.
1443                  */
1444                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1445                         break;
1446
1447                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1448                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1449
1450                 /*
1451                  * Get a reference to this pipe buffer,
1452                  * so we can copy the contents over.
1453                  */
1454                 ibuf->ops->get(ipipe, ibuf);
1455
1456                 obuf = opipe->bufs + nbuf;
1457                 *obuf = *ibuf;
1458
1459                 /*
1460                  * Don't inherit the gift flag, we need to
1461                  * prevent multiple steals of this page.
1462                  */
1463                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1464
1465                 if (obuf->len > len)
1466                         obuf->len = len;
1467
1468                 opipe->nrbufs++;
1469                 ret += obuf->len;
1470                 len -= obuf->len;
1471                 i++;
1472         } while (len);
1473
1474         inode_double_unlock(ipipe->inode, opipe->inode);
1475
1476         /*
1477          * If we put data in the output pipe, wakeup any potential readers.
1478          */
1479         if (ret > 0) {
1480                 smp_mb();
1481                 if (waitqueue_active(&opipe->wait))
1482                         wake_up_interruptible(&opipe->wait);
1483                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1484         }
1485
1486         return ret;
1487 }
1488
1489 /*
1490  * This is a tee(1) implementation that works on pipes. It doesn't copy
1491  * any data, it simply references the 'in' pages on the 'out' pipe.
1492  * The 'flags' used are the SPLICE_F_* variants, currently the only
1493  * applicable one is SPLICE_F_NONBLOCK.
1494  */
1495 static long do_tee(struct file *in, struct file *out, size_t len,
1496                    unsigned int flags)
1497 {
1498         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1499         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1500         int ret = -EINVAL;
1501
1502         /*
1503          * Duplicate the contents of ipipe to opipe without actually
1504          * copying the data.
1505          */
1506         if (ipipe && opipe && ipipe != opipe) {
1507                 /*
1508                  * Keep going, unless we encounter an error. The ipipe/opipe
1509                  * ordering doesn't really matter.
1510                  */
1511                 ret = link_ipipe_prep(ipipe, flags);
1512                 if (!ret) {
1513                         ret = link_opipe_prep(opipe, flags);
1514                         if (!ret) {
1515                                 ret = link_pipe(ipipe, opipe, len, flags);
1516                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1517                                         ret = -EAGAIN;
1518                         }
1519                 }
1520         }
1521
1522         return ret;
1523 }
1524
1525 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1526 {
1527         struct file *in;
1528         int error, fput_in;
1529
1530         if (unlikely(!len))
1531                 return 0;
1532
1533         error = -EBADF;
1534         in = fget_light(fdin, &fput_in);
1535         if (in) {
1536                 if (in->f_mode & FMODE_READ) {
1537                         int fput_out;
1538                         struct file *out = fget_light(fdout, &fput_out);
1539
1540                         if (out) {
1541                                 if (out->f_mode & FMODE_WRITE)
1542                                         error = do_tee(in, out, len, flags);
1543                                 fput_light(out, fput_out);
1544                         }
1545                 }
1546                 fput_light(in, fput_in);
1547         }
1548
1549         return error;
1550 }