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         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, mapping_gfp_mask(mapping));
  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 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 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                                               mapping_gfp_mask(mapping));
 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         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 561         unsigned int offset, this_len;
 562         struct page *page;
 563         pgoff_t index;
 564         int ret;
 565
 566         /*
 567          * make sure the data in this buffer is uptodate
 568          */
 569         ret = buf->ops->pin(pipe, buf);
 570         if (unlikely(ret))
 571                 return ret;
 572
 573         index = sd->pos >> PAGE_CACHE_SHIFT;
 574         offset = sd->pos & ~PAGE_CACHE_MASK;
 575
 576         this_len = sd->len;
 577         if (this_len + offset > PAGE_CACHE_SIZE)
 578                 this_len = PAGE_CACHE_SIZE - offset;
 579
 580         /*
 581          * Reuse buf page, if SPLICE_F_MOVE is set and we are doing a full
 582          * page.
 583          */
 584         if ((sd->flags & SPLICE_F_MOVE) && this_len == PAGE_CACHE_SIZE) {
 585                 /*
 586                  * If steal succeeds, buf->page is now pruned from the
 587                  * pagecache and we can reuse it. The page will also be
 588                  * locked on successful return.
 589                  */
 590                 if (buf->ops->steal(pipe, buf))
 591                         goto find_page;
 592
 593                 page = buf->page;
 594                 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
 595                         unlock_page(page);
 596                         goto find_page;
 597                 }
 598
 599                 page_cache_get(page);
 600
 601                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 602                         lru_cache_add(page);
 603         } else {
 604 find_page:
 605                 page = find_lock_page(mapping, index);
 606                 if (!page) {
 607                         ret = -ENOMEM;
 608                         page = page_cache_alloc_cold(mapping);
 609                         if (unlikely(!page))
 610                                 goto out_ret;
 611
 612                         /*
 613                          * This will also lock the page
 614                          */
 615                         ret = add_to_page_cache_lru(page, mapping, index,
 616                                                     gfp_mask);
 617                         if (unlikely(ret))
 618                                 goto out;
 619                 }
 620
 621                 /*
 622                  * We get here with the page locked. If the page is also
 623                  * uptodate, we don't need to do more. If it isn't, we
 624                  * may need to bring it in if we are not going to overwrite
 625                  * the full page.
 626                  */
 627                 if (!PageUptodate(page)) {
 628                         if (this_len < PAGE_CACHE_SIZE) {
 629                                 ret = mapping->a_ops->readpage(file, page);
 630                                 if (unlikely(ret))
 631                                         goto out;
 632
 633                                 lock_page(page);
 634
 635                                 if (!PageUptodate(page)) {
 636                                         /*
 637                                          * Page got invalidated, repeat.
 638                                          */
 639                                         if (!page->mapping) {
 640                                                 unlock_page(page);
 641                                                 page_cache_release(page);
 642                                                 goto find_page;
 643                                         }
 644                                         ret = -EIO;
 645                                         goto out;
 646                                 }
 647                         } else
 648                                 SetPageUptodate(page);
 649                 }
 650         }
 651
 652         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 653         if (unlikely(ret)) {
 654                 loff_t isize = i_size_read(mapping->host);
 655
 656                 if (ret != AOP_TRUNCATED_PAGE)
 657                         unlock_page(page);
 658                 page_cache_release(page);
 659                 if (ret == AOP_TRUNCATED_PAGE)
 660                         goto find_page;
 661
 662                 /*
 663                  * prepare_write() may have instantiated a few blocks
 664                  * outside i_size.  Trim these off again.
 665                  */
 666                 if (sd->pos + this_len > isize)
 667                         vmtruncate(mapping->host, isize);
 668
 669                 goto out_ret;
 670         }
 671
 672         if (buf->page != page) {
 673                 /*
 674                  * Careful, ->map() uses KM_USER0!
 675                  */
 676                 char *src = buf->ops->map(pipe, buf, 1);
 677                 char *dst = kmap_atomic(page, KM_USER1);
 678
 679                 memcpy(dst + offset, src + buf->offset, this_len);
 680                 flush_dcache_page(page);
 681                 kunmap_atomic(dst, KM_USER1);
 682                 buf->ops->unmap(pipe, buf, src);
 683         }
 684
 685         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 686         if (!ret) {
 687                 /*
 688                  * Return the number of bytes written and mark page as
 689                  * accessed, we are now done!
 690                  */
 691                 ret = this_len;
 692                 mark_page_accessed(page);
 693                 balance_dirty_pages_ratelimited(mapping);
 694         } else if (ret == AOP_TRUNCATED_PAGE) {
 695                 page_cache_release(page);
 696                 goto find_page;
 697         }
 698 out:
 699         page_cache_release(page);
 700         unlock_page(page);
 701 out_ret:
 702         return ret;
 703 }
 704
 705 /*
 706  * Pipe input worker. Most of this logic works like a regular pipe, the
 707  * key here is the 'actor' worker passed in that actually moves the data
 708  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 709  */
 710 static ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 711                                   struct file *out, loff_t *ppos, size_t len,
 712                                   unsigned int flags, splice_actor *actor)
 713 {
 714         int ret, do_wakeup, err;
 715         struct splice_desc sd;
 716
 717         ret = 0;
 718         do_wakeup = 0;
 719
 720         sd.total_len = len;
 721         sd.flags = flags;
 722         sd.file = out;
 723         sd.pos = *ppos;
 724
 725         for (;;) {
 726                 if (pipe->nrbufs) {
 727                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 728                         struct pipe_buf_operations *ops = buf->ops;
 729
 730                         sd.len = buf->len;
 731                         if (sd.len > sd.total_len)
 732                                 sd.len = sd.total_len;
 733
 734                         err = actor(pipe, buf, &sd);
 735                         if (err <= 0) {
 736                                 if (!ret && err != -ENODATA)
 737                                         ret = err;
 738
 739                                 break;
 740                         }
 741
 742                         ret += err;
 743                         buf->offset += err;
 744                         buf->len -= err;
 745
 746                         sd.len -= err;
 747                         sd.pos += err;
 748                         sd.total_len -= err;
 749                         if (sd.len)
 750                                 continue;
 751
 752                         if (!buf->len) {
 753                                 buf->ops = NULL;
 754                                 ops->release(pipe, buf);
 755                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 756                                 pipe->nrbufs--;
 757                                 if (pipe->inode)
 758                                         do_wakeup = 1;
 759                         }
 760
 761                         if (!sd.total_len)
 762                                 break;
 763                 }
 764
 765                 if (pipe->nrbufs)
 766                         continue;
 767                 if (!pipe->writers)
 768                         break;
 769                 if (!pipe->waiting_writers) {
 770                         if (ret)
 771                                 break;
 772                 }
 773
 774                 if (flags & SPLICE_F_NONBLOCK) {
 775                         if (!ret)
 776                                 ret = -EAGAIN;
 777                         break;
 778                 }
 779
 780                 if (signal_pending(current)) {
 781                         if (!ret)
 782                                 ret = -ERESTARTSYS;
 783                         break;
 784                 }
 785
 786                 if (do_wakeup) {
 787                         smp_mb();
 788                         if (waitqueue_active(&pipe->wait))
 789                                 wake_up_interruptible_sync(&pipe->wait);
 790                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 791                         do_wakeup = 0;
 792                 }
 793
 794                 pipe_wait(pipe);
 795         }
 796
 797         if (do_wakeup) {
 798                 smp_mb();
 799                 if (waitqueue_active(&pipe->wait))
 800                         wake_up_interruptible(&pipe->wait);
 801                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 802         }
 803
 804         return ret;
 805 }
 806
 807 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 808                          loff_t *ppos, size_t len, unsigned int flags,
 809                          splice_actor *actor)
 810 {
 811         ssize_t ret;
 812         struct inode *inode = out->f_mapping->host;
 813
 814         /*
 815          * The actor worker might be calling ->prepare_write and
 816          * ->commit_write. Most of the time, these expect i_mutex to
 817          * be held. Since this may result in an ABBA deadlock with
 818          * pipe->inode, we have to order lock acquiry here.
 819          */
 820         inode_double_lock(inode, pipe->inode);
 821         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 822         inode_double_unlock(inode, pipe->inode);
 823
 824         return ret;
 825 }
 826
 827 /**
 828  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 829  * @pipe:       pipe info
 830  * @out:        file to write to
 831  * @len:        number of bytes to splice
 832  * @flags:      splice modifier flags
 833  *
 834  * Will either move or copy pages (determined by @flags options) from
 835  * the given pipe inode to the given file. The caller is responsible
 836  * for acquiring i_mutex on both inodes.
 837  *
 838  */
 839 ssize_t
 840 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 841                                  loff_t *ppos, size_t len, unsigned int flags)
 842 {
 843         struct address_space *mapping = out->f_mapping;
 844         struct inode *inode = mapping->host;
 845         ssize_t ret;
 846         int err;
 847
 848         err = remove_suid(out->f_dentry);
 849         if (unlikely(err))
 850                 return err;
 851
 852         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 853         if (ret > 0) {
 854                 *ppos += ret;
 855
 856                 /*
 857                  * If file or inode is SYNC and we actually wrote some data,
 858                  * sync it.
 859                  */
 860                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 861                         err = generic_osync_inode(inode, mapping,
 862                                                   OSYNC_METADATA|OSYNC_DATA);
 863
 864                         if (err)
 865                                 ret = err;
 866                 }
 867         }
 868
 869         return ret;
 870 }
 871
 872 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 873
 874 /**
 875  * generic_file_splice_write - splice data from a pipe to a file
 876  * @pipe:       pipe info
 877  * @out:        file to write to
 878  * @len:        number of bytes to splice
 879  * @flags:      splice modifier flags
 880  *
 881  * Will either move or copy pages (determined by @flags options) from
 882  * the given pipe inode to the given file.
 883  *
 884  */
 885 ssize_t
 886 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 887                           loff_t *ppos, size_t len, unsigned int flags)
 888 {
 889         struct address_space *mapping = out->f_mapping;
 890         struct inode *inode = mapping->host;
 891         ssize_t ret;
 892         int err;
 893
 894         err = should_remove_suid(out->f_dentry);
 895         if (unlikely(err)) {
 896                 mutex_lock(&inode->i_mutex);
 897                 err = __remove_suid(out->f_dentry, err);
 898                 mutex_unlock(&inode->i_mutex);
 899                 if (err)
 900                         return err;
 901         }
 902
 903         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 904         if (ret > 0) {
 905                 *ppos += ret;
 906
 907                 /*
 908                  * If file or inode is SYNC and we actually wrote some data,
 909                  * sync it.
 910                  */
 911                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 912                         mutex_lock(&inode->i_mutex);
 913                         err = generic_osync_inode(inode, mapping,
 914                                                   OSYNC_METADATA|OSYNC_DATA);
 915                         mutex_unlock(&inode->i_mutex);
 916
 917                         if (err)
 918                                 ret = err;
 919                 }
 920         }
 921
 922         return ret;
 923 }
 924
 925 EXPORT_SYMBOL(generic_file_splice_write);
 926
 927 /**
 928  * generic_splice_sendpage - splice data from a pipe to a socket
 929  * @inode:      pipe inode
 930  * @out:        socket to write to
 931  * @len:        number of bytes to splice
 932  * @flags:      splice modifier flags
 933  *
 934  * Will send @len bytes from the pipe to a network socket. No data copying
 935  * is involved.
 936  *
 937  */
 938 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 939                                 loff_t *ppos, size_t len, unsigned int flags)
 940 {
 941         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 942 }
 943
 944 EXPORT_SYMBOL(generic_splice_sendpage);
 945
 946 /*
 947  * Attempt to initiate a splice from pipe to file.
 948  */
 949 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 950                            loff_t *ppos, size_t len, unsigned int flags)
 951 {
 952         int ret;
 953
 954         if (unlikely(!out->f_op || !out->f_op->splice_write))
 955                 return -EINVAL;
 956
 957         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 958                 return -EBADF;
 959
 960         ret = rw_verify_area(WRITE, out, ppos, len);
 961         if (unlikely(ret < 0))
 962                 return ret;
 963
 964         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 965 }
 966
 967 /*
 968  * Attempt to initiate a splice from a file to a pipe.
 969  */
 970 static long do_splice_to(struct file *in, loff_t *ppos,
 971                          struct pipe_inode_info *pipe, size_t len,
 972                          unsigned int flags)
 973 {
 974         loff_t isize, left;
 975         int ret;
 976
 977         if (unlikely(!in->f_op || !in->f_op->splice_read))
 978                 return -EINVAL;
 979
 980         if (unlikely(!(in->f_mode & FMODE_READ)))
 981                 return -EBADF;
 982
 983         ret = rw_verify_area(READ, in, ppos, len);
 984         if (unlikely(ret < 0))
 985                 return ret;
 986
 987         isize = i_size_read(in->f_mapping->host);
 988         if (unlikely(*ppos >= isize))
 989                 return 0;
 990
 991         left = isize - *ppos;
 992         if (unlikely(left < len))
 993                 len = left;
 994
 995         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 996 }
 997
 998 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 999                       size_t len, unsigned int flags)
1000 {
1001         struct pipe_inode_info *pipe;
1002         long ret, bytes;
1003         loff_t out_off;
1004         umode_t i_mode;
1005         int i;
1006
1007         /*
1008          * We require the input being a regular file, as we don't want to
1009          * randomly drop data for eg socket -> socket splicing. Use the
1010          * piped splicing for that!
1011          */
1012         i_mode = in->f_dentry->d_inode->i_mode;
1013         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1014                 return -EINVAL;
1015
1016         /*
1017          * neither in nor out is a pipe, setup an internal pipe attached to
1018          * 'out' and transfer the wanted data from 'in' to 'out' through that
1019          */
1020         pipe = current->splice_pipe;
1021         if (unlikely(!pipe)) {
1022                 pipe = alloc_pipe_info(NULL);
1023                 if (!pipe)
1024                         return -ENOMEM;
1025
1026                 /*
1027                  * We don't have an immediate reader, but we'll read the stuff
1028                  * out of the pipe right after the splice_to_pipe(). So set
1029                  * PIPE_READERS appropriately.
1030                  */
1031                 pipe->readers = 1;
1032
1033                 current->splice_pipe = pipe;
1034         }
1035
1036         /*
1037          * Do the splice.
1038          */
1039         ret = 0;
1040         bytes = 0;
1041         out_off = 0;
1042
1043         while (len) {
1044                 size_t read_len, max_read_len;
1045
1046                 /*
1047                  * Do at most PIPE_BUFFERS pages worth of transfer:
1048                  */
1049                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1050
1051                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1052                 if (unlikely(ret < 0))
1053                         goto out_release;
1054
1055                 read_len = ret;
1056
1057                 /*
1058                  * NOTE: nonblocking mode only applies to the input. We
1059                  * must not do the output in nonblocking mode as then we
1060                  * could get stuck data in the internal pipe:
1061                  */
1062                 ret = do_splice_from(pipe, out, &out_off, read_len,
1063                                      flags & ~SPLICE_F_NONBLOCK);
1064                 if (unlikely(ret < 0))
1065                         goto out_release;
1066
1067                 bytes += ret;
1068                 len -= ret;
1069
1070                 /*
1071                  * In nonblocking mode, if we got back a short read then
1072                  * that was due to either an IO error or due to the
1073                  * pagecache entry not being there. In the IO error case
1074                  * the _next_ splice attempt will produce a clean IO error
1075                  * return value (not a short read), so in both cases it's
1076                  * correct to break out of the loop here:
1077                  */
1078                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1079                         break;
1080         }
1081
1082         pipe->nrbufs = pipe->curbuf = 0;
1083
1084         return bytes;
1085
1086 out_release:
1087         /*
1088          * If we did an incomplete transfer we must release
1089          * the pipe buffers in question:
1090          */
1091         for (i = 0; i < PIPE_BUFFERS; i++) {
1092                 struct pipe_buffer *buf = pipe->bufs + i;
1093
1094                 if (buf->ops) {
1095                         buf->ops->release(pipe, buf);
1096                         buf->ops = NULL;
1097                 }
1098         }
1099         pipe->nrbufs = pipe->curbuf = 0;
1100
1101         /*
1102          * If we transferred some data, return the number of bytes:
1103          */
1104         if (bytes > 0)
1105                 return bytes;
1106
1107         return ret;
1108 }
1109
1110 EXPORT_SYMBOL(do_splice_direct);
1111
1112 /*
1113  * Determine where to splice to/from.
1114  */
1115 static long do_splice(struct file *in, loff_t __user *off_in,
1116                       struct file *out, loff_t __user *off_out,
1117                       size_t len, unsigned int flags)
1118 {
1119         struct pipe_inode_info *pipe;
1120         loff_t offset, *off;
1121         long ret;
1122
1123         pipe = in->f_dentry->d_inode->i_pipe;
1124         if (pipe) {
1125                 if (off_in)
1126                         return -ESPIPE;
1127                 if (off_out) {
1128                         if (out->f_op->llseek == no_llseek)
1129                                 return -EINVAL;
1130                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1131                                 return -EFAULT;
1132                         off = &offset;
1133                 } else
1134                         off = &out->f_pos;
1135
1136                 ret = do_splice_from(pipe, out, off, len, flags);
1137
1138                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1139                         ret = -EFAULT;
1140
1141                 return ret;
1142         }
1143
1144         pipe = out->f_dentry->d_inode->i_pipe;
1145         if (pipe) {
1146                 if (off_out)
1147                         return -ESPIPE;
1148                 if (off_in) {
1149                         if (in->f_op->llseek == no_llseek)
1150                                 return -EINVAL;
1151                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1152                                 return -EFAULT;
1153                         off = &offset;
1154                 } else
1155                         off = &in->f_pos;
1156
1157                 ret = do_splice_to(in, off, pipe, len, flags);
1158
1159                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1160                         ret = -EFAULT;
1161
1162                 return ret;
1163         }
1164
1165         return -EINVAL;
1166 }
1167
1168 /*
1169  * Map an iov into an array of pages and offset/length tupples. With the
1170  * partial_page structure, we can map several non-contiguous ranges into
1171  * our ones pages[] map instead of splitting that operation into pieces.
1172  * Could easily be exported as a generic helper for other users, in which
1173  * case one would probably want to add a 'max_nr_pages' parameter as well.
1174  */
1175 static int get_iovec_page_array(const struct iovec __user *iov,
1176                                 unsigned int nr_vecs, struct page **pages,
1177                                 struct partial_page *partial, int aligned)
1178 {
1179         int buffers = 0, error = 0;
1180
1181         /*
1182          * It's ok to take the mmap_sem for reading, even
1183          * across a "get_user()".
1184          */
1185         down_read(&current->mm->mmap_sem);
1186
1187         while (nr_vecs) {
1188                 unsigned long off, npages;
1189                 void __user *base;
1190                 size_t len;
1191                 int i;
1192
1193                 /*
1194                  * Get user address base and length for this iovec.
1195                  */
1196                 error = get_user(base, &iov->iov_base);
1197                 if (unlikely(error))
1198                         break;
1199                 error = get_user(len, &iov->iov_len);
1200                 if (unlikely(error))
1201                         break;
1202
1203                 /*
1204                  * Sanity check this iovec. 0 read succeeds.
1205                  */
1206                 if (unlikely(!len))
1207                         break;
1208                 error = -EFAULT;
1209                 if (unlikely(!base))
1210                         break;
1211
1212                 /*
1213                  * Get this base offset and number of pages, then map
1214                  * in the user pages.
1215                  */
1216                 off = (unsigned long) base & ~PAGE_MASK;
1217
1218                 /*
1219                  * If asked for alignment, the offset must be zero and the
1220                  * length a multiple of the PAGE_SIZE.
1221                  */
1222                 error = -EINVAL;
1223                 if (aligned && (off || len & ~PAGE_MASK))
1224                         break;
1225
1226                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1227                 if (npages > PIPE_BUFFERS - buffers)
1228                         npages = PIPE_BUFFERS - buffers;
1229
1230                 error = get_user_pages(current, current->mm,
1231                                        (unsigned long) base, npages, 0, 0,
1232                                        &pages[buffers], NULL);
1233
1234                 if (unlikely(error <= 0))
1235                         break;
1236
1237                 /*
1238                  * Fill this contiguous range into the partial page map.
1239                  */
1240                 for (i = 0; i < error; i++) {
1241                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1242
1243                         partial[buffers].offset = off;
1244                         partial[buffers].len = plen;
1245
1246                         off = 0;
1247                         len -= plen;
1248                         buffers++;
1249                 }
1250
1251                 /*
1252                  * We didn't complete this iov, stop here since it probably
1253                  * means we have to move some of this into a pipe to
1254                  * be able to continue.
1255                  */
1256                 if (len)
1257                         break;
1258
1259                 /*
1260                  * Don't continue if we mapped fewer pages than we asked for,
1261                  * or if we mapped the max number of pages that we have
1262                  * room for.
1263                  */
1264                 if (error < npages || buffers == PIPE_BUFFERS)
1265                         break;
1266
1267                 nr_vecs--;
1268                 iov++;
1269         }
1270
1271         up_read(&current->mm->mmap_sem);
1272
1273         if (buffers)
1274                 return buffers;
1275
1276         return error;
1277 }
1278
1279 /*
1280  * vmsplice splices a user address range into a pipe. It can be thought of
1281  * as splice-from-memory, where the regular splice is splice-from-file (or
1282  * to file). In both cases the output is a pipe, naturally.
1283  *
1284  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1285  * not the other way around. Splicing from user memory is a simple operation
1286  * that can be supported without any funky alignment restrictions or nasty
1287  * vm tricks. We simply map in the user memory and fill them into a pipe.
1288  * The reverse isn't quite as easy, though. There are two possible solutions
1289  * for that:
1290  *
1291  *      - memcpy() the data internally, at which point we might as well just
1292  *        do a regular read() on the buffer anyway.
1293  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1294  *        has restriction limitations on both ends of the pipe).
1295  *
1296  * Alas, it isn't here.
1297  *
1298  */
1299 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1300                         unsigned long nr_segs, unsigned int flags)
1301 {
1302         struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1303         struct page *pages[PIPE_BUFFERS];
1304         struct partial_page partial[PIPE_BUFFERS];
1305         struct splice_pipe_desc spd = {
1306                 .pages = pages,
1307                 .partial = partial,
1308                 .flags = flags,
1309                 .ops = &user_page_pipe_buf_ops,
1310         };
1311
1312         if (unlikely(!pipe))
1313                 return -EBADF;
1314         if (unlikely(nr_segs > UIO_MAXIOV))
1315                 return -EINVAL;
1316         else if (unlikely(!nr_segs))
1317                 return 0;
1318
1319         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1320                                             flags & SPLICE_F_GIFT);
1321         if (spd.nr_pages <= 0)
1322                 return spd.nr_pages;
1323
1324         return splice_to_pipe(pipe, &spd);
1325 }
1326
1327 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1328                              unsigned long nr_segs, unsigned int flags)
1329 {
1330         struct file *file;
1331         long error;
1332         int fput;
1333
1334         error = -EBADF;
1335         file = fget_light(fd, &fput);
1336         if (file) {
1337                 if (file->f_mode & FMODE_WRITE)
1338                         error = do_vmsplice(file, iov, nr_segs, flags);
1339
1340                 fput_light(file, fput);
1341         }
1342
1343         return error;
1344 }
1345
1346 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1347                            int fd_out, loff_t __user *off_out,
1348                            size_t len, unsigned int flags)
1349 {
1350         long error;
1351         struct file *in, *out;
1352         int fput_in, fput_out;
1353
1354         if (unlikely(!len))
1355                 return 0;
1356
1357         error = -EBADF;
1358         in = fget_light(fd_in, &fput_in);
1359         if (in) {
1360                 if (in->f_mode & FMODE_READ) {
1361                         out = fget_light(fd_out, &fput_out);
1362                         if (out) {
1363                                 if (out->f_mode & FMODE_WRITE)
1364                                         error = do_splice(in, off_in,
1365                                                           out, off_out,
1366                                                           len, flags);
1367                                 fput_light(out, fput_out);
1368                         }
1369                 }
1370
1371                 fput_light(in, fput_in);
1372         }
1373
1374         return error;
1375 }
1376
1377 /*
1378  * Make sure there's data to read. Wait for input if we can, otherwise
1379  * return an appropriate error.
1380  */
1381 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1382 {
1383         int ret;
1384
1385         /*
1386          * Check ->nrbufs without the inode lock first. This function
1387          * is speculative anyways, so missing one is ok.
1388          */
1389         if (pipe->nrbufs)
1390                 return 0;
1391
1392         ret = 0;
1393         mutex_lock(&pipe->inode->i_mutex);
1394
1395         while (!pipe->nrbufs) {
1396                 if (signal_pending(current)) {
1397                         ret = -ERESTARTSYS;
1398                         break;
1399                 }
1400                 if (!pipe->writers)
1401                         break;
1402                 if (!pipe->waiting_writers) {
1403                         if (flags & SPLICE_F_NONBLOCK) {
1404                                 ret = -EAGAIN;
1405                                 break;
1406                         }
1407                 }
1408                 pipe_wait(pipe);
1409         }
1410
1411         mutex_unlock(&pipe->inode->i_mutex);
1412         return ret;
1413 }
1414
1415 /*
1416  * Make sure there's writeable room. Wait for room if we can, otherwise
1417  * return an appropriate error.
1418  */
1419 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1420 {
1421         int ret;
1422
1423         /*
1424          * Check ->nrbufs without the inode lock first. This function
1425          * is speculative anyways, so missing one is ok.
1426          */
1427         if (pipe->nrbufs < PIPE_BUFFERS)
1428                 return 0;
1429
1430         ret = 0;
1431         mutex_lock(&pipe->inode->i_mutex);
1432
1433         while (pipe->nrbufs >= PIPE_BUFFERS) {
1434                 if (!pipe->readers) {
1435                         send_sig(SIGPIPE, current, 0);
1436                         ret = -EPIPE;
1437                         break;
1438                 }
1439                 if (flags & SPLICE_F_NONBLOCK) {
1440                         ret = -EAGAIN;
1441                         break;
1442                 }
1443                 if (signal_pending(current)) {
1444                         ret = -ERESTARTSYS;
1445                         break;
1446                 }
1447                 pipe->waiting_writers++;
1448                 pipe_wait(pipe);
1449                 pipe->waiting_writers--;
1450         }
1451
1452         mutex_unlock(&pipe->inode->i_mutex);
1453         return ret;
1454 }
1455
1456 /*
1457  * Link contents of ipipe to opipe.
1458  */
1459 static int link_pipe(struct pipe_inode_info *ipipe,
1460                      struct pipe_inode_info *opipe,
1461                      size_t len, unsigned int flags)
1462 {
1463         struct pipe_buffer *ibuf, *obuf;
1464         int ret = 0, i = 0, nbuf;
1465
1466         /*
1467          * Potential ABBA deadlock, work around it by ordering lock
1468          * grabbing by inode address. Otherwise two different processes
1469          * could deadlock (one doing tee from A -> B, the other from B -> A).
1470          */
1471         inode_double_lock(ipipe->inode, opipe->inode);
1472
1473         do {
1474                 if (!opipe->readers) {
1475                         send_sig(SIGPIPE, current, 0);
1476                         if (!ret)
1477                                 ret = -EPIPE;
1478                         break;
1479                 }
1480
1481                 /*
1482                  * If we have iterated all input buffers or ran out of
1483                  * output room, break.
1484                  */
1485                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1486                         break;
1487
1488                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1489                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1490
1491                 /*
1492                  * Get a reference to this pipe buffer,
1493                  * so we can copy the contents over.
1494                  */
1495                 ibuf->ops->get(ipipe, ibuf);
1496
1497                 obuf = opipe->bufs + nbuf;
1498                 *obuf = *ibuf;
1499
1500                 /*
1501                  * Don't inherit the gift flag, we need to
1502                  * prevent multiple steals of this page.
1503                  */
1504                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1505
1506                 if (obuf->len > len)
1507                         obuf->len = len;
1508
1509                 opipe->nrbufs++;
1510                 ret += obuf->len;
1511                 len -= obuf->len;
1512                 i++;
1513         } while (len);
1514
1515         inode_double_unlock(ipipe->inode, opipe->inode);
1516
1517         /*
1518          * If we put data in the output pipe, wakeup any potential readers.
1519          */
1520         if (ret > 0) {
1521                 smp_mb();
1522                 if (waitqueue_active(&opipe->wait))
1523                         wake_up_interruptible(&opipe->wait);
1524                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1525         }
1526
1527         return ret;
1528 }
1529
1530 /*
1531  * This is a tee(1) implementation that works on pipes. It doesn't copy
1532  * any data, it simply references the 'in' pages on the 'out' pipe.
1533  * The 'flags' used are the SPLICE_F_* variants, currently the only
1534  * applicable one is SPLICE_F_NONBLOCK.
1535  */
1536 static long do_tee(struct file *in, struct file *out, size_t len,
1537                    unsigned int flags)
1538 {
1539         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1540         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1541         int ret = -EINVAL;
1542
1543         /*
1544          * Duplicate the contents of ipipe to opipe without actually
1545          * copying the data.
1546          */
1547         if (ipipe && opipe && ipipe != opipe) {
1548                 /*
1549                  * Keep going, unless we encounter an error. The ipipe/opipe
1550                  * ordering doesn't really matter.
1551                  */
1552                 ret = link_ipipe_prep(ipipe, flags);
1553                 if (!ret) {
1554                         ret = link_opipe_prep(opipe, flags);
1555                         if (!ret) {
1556                                 ret = link_pipe(ipipe, opipe, len, flags);
1557                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1558                                         ret = -EAGAIN;
1559                         }
1560                 }
1561         }
1562
1563         return ret;
1564 }
1565
1566 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1567 {
1568         struct file *in;
1569         int error, fput_in;
1570
1571         if (unlikely(!len))
1572                 return 0;
1573
1574         error = -EBADF;
1575         in = fget_light(fdin, &fput_in);
1576         if (in) {
1577                 if (in->f_mode & FMODE_READ) {
1578                         int fput_out;
1579                         struct file *out = fget_light(fdout, &fput_out);
1580
1581                         if (out) {
1582                                 if (out->f_mode & FMODE_WRITE)
1583                                         error = do_tee(in, out, len, flags);
1584                                 fput_light(out, fput_out);
1585                         }
1586                 }
1587                 fput_light(in, fput_in);
1588         }
1589
1590         return error;
1591 }