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