fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/pipe_fs_i.h>
  24 #include <linux/mm_inline.h>
  25 #include <linux/swap.h>
  26 #include <linux/writeback.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/module.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31
  32 struct partial_page {
  33         unsigned int offset;
  34         unsigned int len;
  35 };
  36
  37 /*
  38  * Passed to splice_to_pipe
  39  */
  40 struct splice_pipe_desc {
  41         struct page **pages;            /* page map */
  42         struct partial_page *partial;   /* pages[] may not be contig */
  43         int nr_pages;                   /* number of pages in map */
  44         unsigned int flags;             /* splice flags */
  45         const struct pipe_buf_operations *ops;/* ops associated with output pipe */
  46 };
  47
  48 /*
  49  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  50  * a vm helper function, it's already simplified quite a bit by the
  51  * addition of remove_mapping(). If success is returned, the caller may
  52  * attempt to reuse this page for another destination.
  53  */
  54 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  55                                      struct pipe_buffer *buf)
  56 {
  57         struct page *page = buf->page;
  58         struct address_space *mapping;
  59
  60         lock_page(page);
  61
  62         mapping = page_mapping(page);
  63         if (mapping) {
  64                 WARN_ON(!PageUptodate(page));
  65
  66                 /*
  67                  * At least for ext2 with nobh option, we need to wait on
  68                  * writeback completing on this page, since we'll remove it
  69                  * from the pagecache.  Otherwise truncate wont wait on the
  70                  * page, allowing the disk blocks to be reused by someone else
  71                  * before we actually wrote our data to them. fs corruption
  72                  * ensues.
  73                  */
  74                 wait_on_page_writeback(page);
  75
  76                 if (PagePrivate(page))
  77                         try_to_release_page(page, GFP_KERNEL);
  78
  79                 /*
  80                  * If we succeeded in removing the mapping, set LRU flag
  81                  * and return good.
  82                  */
  83                 if (remove_mapping(mapping, page)) {
  84                         buf->flags |= PIPE_BUF_FLAG_LRU;
  85                         return 0;
  86                 }
  87         }
  88
  89         /*
  90          * Raced with truncate or failed to remove page from current
  91          * address space, unlock and return failure.
  92          */
  93         unlock_page(page);
  94         return 1;
  95 }
  96
  97 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  98                                         struct pipe_buffer *buf)
  99 {
 100         page_cache_release(buf->page);
 101         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 102 }
 103
 104 static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
 105                                    struct pipe_buffer *buf)
 106 {
 107         struct page *page = buf->page;
 108         int err;
 109
 110         if (!PageUptodate(page)) {
 111                 lock_page(page);
 112
 113                 /*
 114                  * Page got truncated/unhashed. This will cause a 0-byte
 115                  * splice, if this is the first page.
 116                  */
 117                 if (!page->mapping) {
 118                         err = -ENODATA;
 119                         goto error;
 120                 }
 121
 122                 /*
 123                  * Uh oh, read-error from disk.
 124                  */
 125                 if (!PageUptodate(page)) {
 126                         err = -EIO;
 127                         goto error;
 128                 }
 129
 130                 /*
 131                  * Page is ok afterall, we are done.
 132                  */
 133                 unlock_page(page);
 134         }
 135
 136         return 0;
 137 error:
 138         unlock_page(page);
 139         return err;
 140 }
 141
 142 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 143         .can_merge = 0,
 144         .map = generic_pipe_buf_map,
 145         .unmap = generic_pipe_buf_unmap,
 146         .pin = page_cache_pipe_buf_pin,
 147         .release = page_cache_pipe_buf_release,
 148         .steal = page_cache_pipe_buf_steal,
 149         .get = generic_pipe_buf_get,
 150 };
 151
 152 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 153                                     struct pipe_buffer *buf)
 154 {
 155         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 156                 return 1;
 157
 158         buf->flags |= PIPE_BUF_FLAG_LRU;
 159         return generic_pipe_buf_steal(pipe, buf);
 160 }
 161
 162 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 163         .can_merge = 0,
 164         .map = generic_pipe_buf_map,
 165         .unmap = generic_pipe_buf_unmap,
 166         .pin = generic_pipe_buf_pin,
 167         .release = page_cache_pipe_buf_release,
 168         .steal = user_page_pipe_buf_steal,
 169         .get = generic_pipe_buf_get,
 170 };
 171
 172 /*
 173  * Pipe output worker. This sets up our pipe format with the page cache
 174  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 175  */
 176 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 177                               struct splice_pipe_desc *spd)
 178 {
 179         int ret, do_wakeup, page_nr;
 180
 181         ret = 0;
 182         do_wakeup = 0;
 183         page_nr = 0;
 184
 185         if (pipe->inode)
 186                 mutex_lock(&pipe->inode->i_mutex);
 187
 188         for (;;) {
 189                 if (!pipe->readers) {
 190                         send_sig(SIGPIPE, current, 0);
 191                         if (!ret)
 192                                 ret = -EPIPE;
 193                         break;
 194                 }
 195
 196                 if (pipe->nrbufs < PIPE_BUFFERS) {
 197                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 198                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 199
 200                         buf->page = spd->pages[page_nr];
 201                         buf->offset = spd->partial[page_nr].offset;
 202                         buf->len = spd->partial[page_nr].len;
 203                         buf->ops = spd->ops;
 204                         if (spd->flags & SPLICE_F_GIFT)
 205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 206
 207                         pipe->nrbufs++;
 208                         page_nr++;
 209                         ret += buf->len;
 210
 211                         if (pipe->inode)
 212                                 do_wakeup = 1;
 213
 214                         if (!--spd->nr_pages)
 215                                 break;
 216                         if (pipe->nrbufs < PIPE_BUFFERS)
 217                                 continue;
 218
 219                         break;
 220                 }
 221
 222                 if (spd->flags & SPLICE_F_NONBLOCK) {
 223                         if (!ret)
 224                                 ret = -EAGAIN;
 225                         break;
 226                 }
 227
 228                 if (signal_pending(current)) {
 229                         if (!ret)
 230                                 ret = -ERESTARTSYS;
 231                         break;
 232                 }
 233
 234                 if (do_wakeup) {
 235                         smp_mb();
 236                         if (waitqueue_active(&pipe->wait))
 237                                 wake_up_interruptible_sync(&pipe->wait);
 238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 239                         do_wakeup = 0;
 240                 }
 241
 242                 pipe->waiting_writers++;
 243                 pipe_wait(pipe);
 244                 pipe->waiting_writers--;
 245         }
 246
 247         if (pipe->inode)
 248                 mutex_unlock(&pipe->inode->i_mutex);
 249
 250         if (do_wakeup) {
 251                 smp_mb();
 252                 if (waitqueue_active(&pipe->wait))
 253                         wake_up_interruptible(&pipe->wait);
 254                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 255         }
 256
 257         while (page_nr < spd->nr_pages)
 258                 page_cache_release(spd->pages[page_nr++]);
 259
 260         return ret;
 261 }
 262
 263 static int
 264 __generic_file_splice_read(struct file *in, loff_t *ppos,
 265                            struct pipe_inode_info *pipe, size_t len,
 266                            unsigned int flags)
 267 {
 268         struct address_space *mapping = in->f_mapping;
 269         unsigned int loff, nr_pages;
 270         struct page *pages[PIPE_BUFFERS];
 271         struct partial_page partial[PIPE_BUFFERS];
 272         struct page *page;
 273         pgoff_t index, end_index;
 274         loff_t isize;
 275         size_t total_len;
 276         int error, page_nr;
 277         struct splice_pipe_desc spd = {
 278                 .pages = pages,
 279                 .partial = partial,
 280                 .flags = flags,
 281                 .ops = &page_cache_pipe_buf_ops,
 282         };
 283
 284         index = *ppos >> PAGE_CACHE_SHIFT;
 285         loff = *ppos & ~PAGE_CACHE_MASK;
 286         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 287
 288         if (nr_pages > PIPE_BUFFERS)
 289                 nr_pages = PIPE_BUFFERS;
 290
 291         /*
 292          * Initiate read-ahead on this page range. however, don't call into
 293          * read-ahead if this is a non-zero offset (we are likely doing small
 294          * chunk splice and the page is already there) for a single page.
 295          */
 296         if (!loff || nr_pages > 1)
 297                 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 298
 299         /*
 300          * Now fill in the holes:
 301          */
 302         error = 0;
 303         total_len = 0;
 304
 305         /*
 306          * Lookup the (hopefully) full range of pages we need.
 307          */
 308         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 309
 310         /*
 311          * If find_get_pages_contig() returned fewer pages than we needed,
 312          * allocate the rest.
 313          */
 314         index += spd.nr_pages;
 315         while (spd.nr_pages < nr_pages) {
 316                 /*
 317                  * Page could be there, find_get_pages_contig() breaks on
 318                  * the first hole.
 319                  */
 320                 page = find_get_page(mapping, index);
 321                 if (!page) {
 322                         /*
 323                          * Make sure the read-ahead engine is notified
 324                          * about this failure.
 325                          */
 326                         handle_ra_miss(mapping, &in->f_ra, index);
 327
 328                         /*
 329                          * page didn't exist, allocate one.
 330                          */
 331                         page = page_cache_alloc_cold(mapping);
 332                         if (!page)
 333                                 break;
 334
 335                         error = add_to_page_cache_lru(page, mapping, index,
 336                                               GFP_KERNEL);
 337                         if (unlikely(error)) {
 338                                 page_cache_release(page);
 339                                 if (error == -EEXIST)
 340                                         continue;
 341                                 break;
 342                         }
 343                         /*
 344                          * add_to_page_cache() locks the page, unlock it
 345                          * to avoid convoluting the logic below even more.
 346                          */
 347                         unlock_page(page);
 348                 }
 349
 350                 pages[spd.nr_pages++] = page;
 351                 index++;
 352         }
 353
 354         /*
 355          * Now loop over the map and see if we need to start IO on any
 356          * pages, fill in the partial map, etc.
 357          */
 358         index = *ppos >> PAGE_CACHE_SHIFT;
 359         nr_pages = spd.nr_pages;
 360         spd.nr_pages = 0;
 361         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 362                 unsigned int this_len;
 363
 364                 if (!len)
 365                         break;
 366
 367                 /*
 368                  * this_len is the max we'll use from this page
 369                  */
 370                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 371                 page = pages[page_nr];
 372
 373                 /*
 374                  * If the page isn't uptodate, we may need to start io on it
 375                  */
 376                 if (!PageUptodate(page)) {
 377                         /*
 378                          * If in nonblock mode then dont block on waiting
 379                          * for an in-flight io page
 380                          */
 381                         if (flags & SPLICE_F_NONBLOCK)
 382                                 break;
 383
 384                         lock_page(page);
 385
 386                         /*
 387                          * page was truncated, stop here. if this isn't the
 388                          * first page, we'll just complete what we already
 389                          * added
 390                          */
 391                         if (!page->mapping) {
 392                                 unlock_page(page);
 393                                 break;
 394                         }
 395                         /*
 396                          * page was already under io and is now done, great
 397                          */
 398                         if (PageUptodate(page)) {
 399                                 unlock_page(page);
 400                                 goto fill_it;
 401                         }
 402
 403                         /*
 404                          * need to read in the page
 405                          */
 406                         error = mapping->a_ops->readpage(in, page);
 407                         if (unlikely(error)) {
 408                                 /*
 409                                  * We really should re-lookup the page here,
 410                                  * but it complicates things a lot. Instead
 411                                  * lets just do what we already stored, and
 412                                  * we'll get it the next time we are called.
 413                                  */
 414                                 if (error == AOP_TRUNCATED_PAGE)
 415                                         error = 0;
 416
 417                                 break;
 418                         }
 419
 420                         /*
 421                          * i_size must be checked after ->readpage().
 422                          */
 423                         isize = i_size_read(mapping->host);
 424                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 425                         if (unlikely(!isize || index > end_index))
 426                                 break;
 427
 428                         /*
 429                          * if this is the last page, see if we need to shrink
 430                          * the length and stop
 431                          */
 432                         if (end_index == index) {
 433                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 434                                 if (total_len + loff > isize)
 435                                         break;
 436                                 /*
 437                                  * force quit after adding this page
 438                                  */
 439                                 len = this_len;
 440                                 this_len = min(this_len, loff);
 441                                 loff = 0;
 442                         }
 443                 }
 444 fill_it:
 445                 partial[page_nr].offset = loff;
 446                 partial[page_nr].len = this_len;
 447                 len -= this_len;
 448                 total_len += this_len;
 449                 loff = 0;
 450                 spd.nr_pages++;
 451                 index++;
 452         }
 453
 454         /*
 455          * Release any pages at the end, if we quit early. 'i' is how far
 456          * we got, 'nr_pages' is how many pages are in the map.
 457          */
 458         while (page_nr < nr_pages)
 459                 page_cache_release(pages[page_nr++]);
 460
 461         if (spd.nr_pages)
 462                 return splice_to_pipe(pipe, &spd);
 463
 464         return error;
 465 }
 466
 467 /**
 468  * generic_file_splice_read - splice data from file to a pipe
 469  * @in:         file to splice from
 470  * @pipe:       pipe to splice to
 471  * @len:        number of bytes to splice
 472  * @flags:      splice modifier flags
 473  *
 474  * Will read pages from given file and fill them into a pipe.
 475  */
 476 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 477                                  struct pipe_inode_info *pipe, size_t len,
 478                                  unsigned int flags)
 479 {
 480         ssize_t spliced;
 481         int ret;
 482
 483         ret = 0;
 484         spliced = 0;
 485
 486         while (len) {
 487                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 488
 489                 if (ret < 0)
 490                         break;
 491                 else if (!ret) {
 492                         if (spliced)
 493                                 break;
 494                         if (flags & SPLICE_F_NONBLOCK) {
 495                                 ret = -EAGAIN;
 496                                 break;
 497                         }
 498                 }
 499
 500                 *ppos += ret;
 501                 len -= ret;
 502                 spliced += ret;
 503         }
 504
 505         if (spliced)
 506                 return spliced;
 507
 508         return ret;
 509 }
 510
 511 EXPORT_SYMBOL(generic_file_splice_read);
 512
 513 /*
 514  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 515  * using sendpage(). Return the number of bytes sent.
 516  */
 517 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 518                             struct pipe_buffer *buf, struct splice_desc *sd)
 519 {
 520         struct file *file = sd->file;
 521         loff_t pos = sd->pos;
 522         int ret, more;
 523
 524         ret = buf->ops->pin(pipe, buf);
 525         if (!ret) {
 526                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 527
 528                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 529                                            sd->len, &pos, more);
 530         }
 531
 532         return ret;
 533 }
 534
 535 /*
 536  * This is a little more tricky than the file -> pipe splicing. There are
 537  * basically three cases:
 538  *
 539  *      - Destination page already exists in the address space and there
 540  *        are users of it. For that case we have no other option that
 541  *        copying the data. Tough luck.
 542  *      - Destination page already exists in the address space, but there
 543  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 544  *        through to last case.
 545  *      - Destination page does not exist, we can add the pipe page to
 546  *        the page cache and avoid the copy.
 547  *
 548  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 549  * sd->flags), we attempt to migrate pages from the pipe to the output
 550  * file address space page cache. This is possible if no one else has
 551  * the pipe page referenced outside of the pipe and page cache. If
 552  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 553  * a new page in the output file page cache and fill/dirty that.
 554  */
 555 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 556                         struct splice_desc *sd)
 557 {
 558         struct file *file = sd->file;
 559         struct address_space *mapping = file->f_mapping;
 560         unsigned int offset, this_len;
 561         struct page *page;
 562         pgoff_t index;
 563         int ret;
 564
 565         /*
 566          * make sure the data in this buffer is uptodate
 567          */
 568         ret = buf->ops->pin(pipe, buf);
 569         if (unlikely(ret))
 570                 return ret;
 571
 572         index = sd->pos >> PAGE_CACHE_SHIFT;
 573         offset = sd->pos & ~PAGE_CACHE_MASK;
 574
 575         this_len = sd->len;
 576         if (this_len + offset > PAGE_CACHE_SIZE)
 577                 this_len = PAGE_CACHE_SIZE - offset;
 578
 579 find_page:
 580         page = find_lock_page(mapping, index);
 581         if (!page) {
 582                 ret = -ENOMEM;
 583                 page = page_cache_alloc_cold(mapping);
 584                 if (unlikely(!page))
 585                         goto out_ret;
 586
 587                 /*
 588                  * This will also lock the page
 589                  */
 590                 ret = add_to_page_cache_lru(page, mapping, index,
 591                                             GFP_KERNEL);
 592                 if (unlikely(ret))
 593                         goto out;
 594         }
 595
 596         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 597         if (unlikely(ret)) {
 598                 loff_t isize = i_size_read(mapping->host);
 599
 600                 if (ret != AOP_TRUNCATED_PAGE)
 601                         unlock_page(page);
 602                 page_cache_release(page);
 603                 if (ret == AOP_TRUNCATED_PAGE)
 604                         goto find_page;
 605
 606                 /*
 607                  * prepare_write() may have instantiated a few blocks
 608                  * outside i_size.  Trim these off again.
 609                  */
 610                 if (sd->pos + this_len > isize)
 611                         vmtruncate(mapping->host, isize);
 612
 613                 goto out_ret;
 614         }
 615
 616         if (buf->page != page) {
 617                 /*
 618                  * Careful, ->map() uses KM_USER0!
 619                  */
 620                 char *src = buf->ops->map(pipe, buf, 1);
 621                 char *dst = kmap_atomic(page, KM_USER1);
 622
 623                 memcpy(dst + offset, src + buf->offset, this_len);
 624                 flush_dcache_page(page);
 625                 kunmap_atomic(dst, KM_USER1);
 626                 buf->ops->unmap(pipe, buf, src);
 627         }
 628
 629         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 630         if (ret) {
 631                 if (ret == AOP_TRUNCATED_PAGE) {
 632                         page_cache_release(page);
 633                         goto find_page;
 634                 }
 635                 if (ret < 0)
 636                         goto out;
 637                 /*
 638                  * Partial write has happened, so 'ret' already initialized by
 639                  * number of bytes written, Where is nothing we have to do here.
 640                  */
 641         } else
 642                 ret = this_len;
 643         /*
 644          * Return the number of bytes written and mark page as
 645          * accessed, we are now done!
 646          */
 647         mark_page_accessed(page);
 648         balance_dirty_pages_ratelimited(mapping);
 649 out:
 650         page_cache_release(page);
 651         unlock_page(page);
 652 out_ret:
 653         return ret;
 654 }
 655
 656 /*
 657  * Pipe input worker. Most of this logic works like a regular pipe, the
 658  * key here is the 'actor' worker passed in that actually moves the data
 659  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 660  */
 661 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 662                            struct file *out, loff_t *ppos, size_t len,
 663                            unsigned int flags, splice_actor *actor)
 664 {
 665         int ret, do_wakeup, err;
 666         struct splice_desc sd;
 667
 668         ret = 0;
 669         do_wakeup = 0;
 670
 671         sd.total_len = len;
 672         sd.flags = flags;
 673         sd.file = out;
 674         sd.pos = *ppos;
 675
 676         for (;;) {
 677                 if (pipe->nrbufs) {
 678                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 679                         const struct pipe_buf_operations *ops = buf->ops;
 680
 681                         sd.len = buf->len;
 682                         if (sd.len > sd.total_len)
 683                                 sd.len = sd.total_len;
 684
 685                         err = actor(pipe, buf, &sd);
 686                         if (err <= 0) {
 687                                 if (!ret && err != -ENODATA)
 688                                         ret = err;
 689
 690                                 break;
 691                         }
 692
 693                         ret += err;
 694                         buf->offset += err;
 695                         buf->len -= err;
 696
 697                         sd.len -= err;
 698                         sd.pos += err;
 699                         sd.total_len -= err;
 700                         if (sd.len)
 701                                 continue;
 702
 703                         if (!buf->len) {
 704                                 buf->ops = NULL;
 705                                 ops->release(pipe, buf);
 706                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 707                                 pipe->nrbufs--;
 708                                 if (pipe->inode)
 709                                         do_wakeup = 1;
 710                         }
 711
 712                         if (!sd.total_len)
 713                                 break;
 714                 }
 715
 716                 if (pipe->nrbufs)
 717                         continue;
 718                 if (!pipe->writers)
 719                         break;
 720                 if (!pipe->waiting_writers) {
 721                         if (ret)
 722                                 break;
 723                 }
 724
 725                 if (flags & SPLICE_F_NONBLOCK) {
 726                         if (!ret)
 727                                 ret = -EAGAIN;
 728                         break;
 729                 }
 730
 731                 if (signal_pending(current)) {
 732                         if (!ret)
 733                                 ret = -ERESTARTSYS;
 734                         break;
 735                 }
 736
 737                 if (do_wakeup) {
 738                         smp_mb();
 739                         if (waitqueue_active(&pipe->wait))
 740                                 wake_up_interruptible_sync(&pipe->wait);
 741                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 742                         do_wakeup = 0;
 743                 }
 744
 745                 pipe_wait(pipe);
 746         }
 747
 748         if (do_wakeup) {
 749                 smp_mb();
 750                 if (waitqueue_active(&pipe->wait))
 751                         wake_up_interruptible(&pipe->wait);
 752                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 753         }
 754
 755         return ret;
 756 }
 757 EXPORT_SYMBOL(__splice_from_pipe);
 758
 759 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 760                          loff_t *ppos, size_t len, unsigned int flags,
 761                          splice_actor *actor)
 762 {
 763         ssize_t ret;
 764         struct inode *inode = out->f_mapping->host;
 765
 766         /*
 767          * The actor worker might be calling ->prepare_write and
 768          * ->commit_write. Most of the time, these expect i_mutex to
 769          * be held. Since this may result in an ABBA deadlock with
 770          * pipe->inode, we have to order lock acquiry here.
 771          */
 772         inode_double_lock(inode, pipe->inode);
 773         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 774         inode_double_unlock(inode, pipe->inode);
 775
 776         return ret;
 777 }
 778
 779 /**
 780  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 781  * @pipe:       pipe info
 782  * @out:        file to write to
 783  * @len:        number of bytes to splice
 784  * @flags:      splice modifier flags
 785  *
 786  * Will either move or copy pages (determined by @flags options) from
 787  * the given pipe inode to the given file. The caller is responsible
 788  * for acquiring i_mutex on both inodes.
 789  *
 790  */
 791 ssize_t
 792 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 793                                  loff_t *ppos, size_t len, unsigned int flags)
 794 {
 795         struct address_space *mapping = out->f_mapping;
 796         struct inode *inode = mapping->host;
 797         ssize_t ret;
 798         int err;
 799
 800         err = remove_suid(out->f_path.dentry);
 801         if (unlikely(err))
 802                 return err;
 803
 804         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 805         if (ret > 0) {
 806                 *ppos += ret;
 807
 808                 /*
 809                  * If file or inode is SYNC and we actually wrote some data,
 810                  * sync it.
 811                  */
 812                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 813                         err = generic_osync_inode(inode, mapping,
 814                                                   OSYNC_METADATA|OSYNC_DATA);
 815
 816                         if (err)
 817                                 ret = err;
 818                 }
 819         }
 820
 821         return ret;
 822 }
 823
 824 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 825
 826 /**
 827  * generic_file_splice_write - splice data from a pipe to a file
 828  * @pipe:       pipe info
 829  * @out:        file to write to
 830  * @len:        number of bytes to splice
 831  * @flags:      splice modifier flags
 832  *
 833  * Will either move or copy pages (determined by @flags options) from
 834  * the given pipe inode to the given file.
 835  *
 836  */
 837 ssize_t
 838 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 839                           loff_t *ppos, size_t len, unsigned int flags)
 840 {
 841         struct address_space *mapping = out->f_mapping;
 842         struct inode *inode = mapping->host;
 843         ssize_t ret;
 844         int err;
 845
 846         err = should_remove_suid(out->f_path.dentry);
 847         if (unlikely(err)) {
 848                 mutex_lock(&inode->i_mutex);
 849                 err = __remove_suid(out->f_path.dentry, err);
 850                 mutex_unlock(&inode->i_mutex);
 851                 if (err)
 852                         return err;
 853         }
 854
 855         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 856         if (ret > 0) {
 857                 *ppos += ret;
 858
 859                 /*
 860                  * If file or inode is SYNC and we actually wrote some data,
 861                  * sync it.
 862                  */
 863                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 864                         mutex_lock(&inode->i_mutex);
 865                         err = generic_osync_inode(inode, mapping,
 866                                                   OSYNC_METADATA|OSYNC_DATA);
 867                         mutex_unlock(&inode->i_mutex);
 868
 869                         if (err)
 870                                 ret = err;
 871                 }
 872         }
 873
 874         return ret;
 875 }
 876
 877 EXPORT_SYMBOL(generic_file_splice_write);
 878
 879 /**
 880  * generic_splice_sendpage - splice data from a pipe to a socket
 881  * @inode:      pipe inode
 882  * @out:        socket to write to
 883  * @len:        number of bytes to splice
 884  * @flags:      splice modifier flags
 885  *
 886  * Will send @len bytes from the pipe to a network socket. No data copying
 887  * is involved.
 888  *
 889  */
 890 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 891                                 loff_t *ppos, size_t len, unsigned int flags)
 892 {
 893         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 894 }
 895
 896 EXPORT_SYMBOL(generic_splice_sendpage);
 897
 898 /*
 899  * Attempt to initiate a splice from pipe to file.
 900  */
 901 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 902                            loff_t *ppos, size_t len, unsigned int flags)
 903 {
 904         int ret;
 905
 906         if (unlikely(!out->f_op || !out->f_op->splice_write))
 907                 return -EINVAL;
 908
 909         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 910                 return -EBADF;
 911
 912         ret = rw_verify_area(WRITE, out, ppos, len);
 913         if (unlikely(ret < 0))
 914                 return ret;
 915
 916         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 917 }
 918
 919 /*
 920  * Attempt to initiate a splice from a file to a pipe.
 921  */
 922 static long do_splice_to(struct file *in, loff_t *ppos,
 923                          struct pipe_inode_info *pipe, size_t len,
 924                          unsigned int flags)
 925 {
 926         loff_t isize, left;
 927         int ret;
 928
 929         if (unlikely(!in->f_op || !in->f_op->splice_read))
 930                 return -EINVAL;
 931
 932         if (unlikely(!(in->f_mode & FMODE_READ)))
 933                 return -EBADF;
 934
 935         ret = rw_verify_area(READ, in, ppos, len);
 936         if (unlikely(ret < 0))
 937                 return ret;
 938
 939         isize = i_size_read(in->f_mapping->host);
 940         if (unlikely(*ppos >= isize))
 941                 return 0;
 942
 943         left = isize - *ppos;
 944         if (unlikely(left < len))
 945                 len = left;
 946
 947         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 948 }
 949
 950 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 951                       size_t len, unsigned int flags)
 952 {
 953         struct pipe_inode_info *pipe;
 954         long ret, bytes;
 955         loff_t out_off;
 956         umode_t i_mode;
 957         int i;
 958
 959         /*
 960          * We require the input being a regular file, as we don't want to
 961          * randomly drop data for eg socket -> socket splicing. Use the
 962          * piped splicing for that!
 963          */
 964         i_mode = in->f_path.dentry->d_inode->i_mode;
 965         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 966                 return -EINVAL;
 967
 968         /*
 969          * neither in nor out is a pipe, setup an internal pipe attached to
 970          * 'out' and transfer the wanted data from 'in' to 'out' through that
 971          */
 972         pipe = current->splice_pipe;
 973         if (unlikely(!pipe)) {
 974                 pipe = alloc_pipe_info(NULL);
 975                 if (!pipe)
 976                         return -ENOMEM;
 977
 978                 /*
 979                  * We don't have an immediate reader, but we'll read the stuff
 980                  * out of the pipe right after the splice_to_pipe(). So set
 981                  * PIPE_READERS appropriately.
 982                  */
 983                 pipe->readers = 1;
 984
 985                 current->splice_pipe = pipe;
 986         }
 987
 988         /*
 989          * Do the splice.
 990          */
 991         ret = 0;
 992         bytes = 0;
 993         out_off = 0;
 994
 995         while (len) {
 996                 size_t read_len, max_read_len;
 997
 998                 /*
 999                  * Do at most PIPE_BUFFERS pages worth of transfer:
1000                  */
1001                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1002
1003                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1004                 if (unlikely(ret < 0))
1005                         goto out_release;
1006
1007                 read_len = ret;
1008
1009                 /*
1010                  * NOTE: nonblocking mode only applies to the input. We
1011                  * must not do the output in nonblocking mode as then we
1012                  * could get stuck data in the internal pipe:
1013                  */
1014                 ret = do_splice_from(pipe, out, &out_off, read_len,
1015                                      flags & ~SPLICE_F_NONBLOCK);
1016                 if (unlikely(ret < 0))
1017                         goto out_release;
1018
1019                 bytes += ret;
1020                 len -= ret;
1021
1022                 /*
1023                  * In nonblocking mode, if we got back a short read then
1024                  * that was due to either an IO error or due to the
1025                  * pagecache entry not being there. In the IO error case
1026                  * the _next_ splice attempt will produce a clean IO error
1027                  * return value (not a short read), so in both cases it's
1028                  * correct to break out of the loop here:
1029                  */
1030                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1031                         break;
1032         }
1033
1034         pipe->nrbufs = pipe->curbuf = 0;
1035
1036         return bytes;
1037
1038 out_release:
1039         /*
1040          * If we did an incomplete transfer we must release
1041          * the pipe buffers in question:
1042          */
1043         for (i = 0; i < PIPE_BUFFERS; i++) {
1044                 struct pipe_buffer *buf = pipe->bufs + i;
1045
1046                 if (buf->ops) {
1047                         buf->ops->release(pipe, buf);
1048                         buf->ops = NULL;
1049                 }
1050         }
1051         pipe->nrbufs = pipe->curbuf = 0;
1052
1053         /*
1054          * If we transferred some data, return the number of bytes:
1055          */
1056         if (bytes > 0)
1057                 return bytes;
1058
1059         return ret;
1060 }
1061
1062 EXPORT_SYMBOL(do_splice_direct);
1063
1064 /*
1065  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1066  * location, so checking ->i_pipe is not enough to verify that this is a
1067  * pipe.
1068  */
1069 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1070 {
1071         if (S_ISFIFO(inode->i_mode))
1072                 return inode->i_pipe;
1073
1074         return NULL;
1075 }
1076
1077 /*
1078  * Determine where to splice to/from.
1079  */
1080 static long do_splice(struct file *in, loff_t __user *off_in,
1081                       struct file *out, loff_t __user *off_out,
1082                       size_t len, unsigned int flags)
1083 {
1084         struct pipe_inode_info *pipe;
1085         loff_t offset, *off;
1086         long ret;
1087
1088         pipe = pipe_info(in->f_path.dentry->d_inode);
1089         if (pipe) {
1090                 if (off_in)
1091                         return -ESPIPE;
1092                 if (off_out) {
1093                         if (out->f_op->llseek == no_llseek)
1094                                 return -EINVAL;
1095                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1096                                 return -EFAULT;
1097                         off = &offset;
1098                 } else
1099                         off = &out->f_pos;
1100
1101                 ret = do_splice_from(pipe, out, off, len, flags);
1102
1103                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1104                         ret = -EFAULT;
1105
1106                 return ret;
1107         }
1108
1109         pipe = pipe_info(out->f_path.dentry->d_inode);
1110         if (pipe) {
1111                 if (off_out)
1112                         return -ESPIPE;
1113                 if (off_in) {
1114                         if (in->f_op->llseek == no_llseek)
1115                                 return -EINVAL;
1116                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1117                                 return -EFAULT;
1118                         off = &offset;
1119                 } else
1120                         off = &in->f_pos;
1121
1122                 ret = do_splice_to(in, off, pipe, len, flags);
1123
1124                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1125                         ret = -EFAULT;
1126
1127                 return ret;
1128         }
1129
1130         return -EINVAL;
1131 }
1132
1133 /*
1134  * Map an iov into an array of pages and offset/length tupples. With the
1135  * partial_page structure, we can map several non-contiguous ranges into
1136  * our ones pages[] map instead of splitting that operation into pieces.
1137  * Could easily be exported as a generic helper for other users, in which
1138  * case one would probably want to add a 'max_nr_pages' parameter as well.
1139  */
1140 static int get_iovec_page_array(const struct iovec __user *iov,
1141                                 unsigned int nr_vecs, struct page **pages,
1142                                 struct partial_page *partial, int aligned)
1143 {
1144         int buffers = 0, error = 0;
1145
1146         /*
1147          * It's ok to take the mmap_sem for reading, even
1148          * across a "get_user()".
1149          */
1150         down_read(&current->mm->mmap_sem);
1151
1152         while (nr_vecs) {
1153                 unsigned long off, npages;
1154                 void __user *base;
1155                 size_t len;
1156                 int i;
1157
1158                 /*
1159                  * Get user address base and length for this iovec.
1160                  */
1161                 error = get_user(base, &iov->iov_base);
1162                 if (unlikely(error))
1163                         break;
1164                 error = get_user(len, &iov->iov_len);
1165                 if (unlikely(error))
1166                         break;
1167
1168                 /*
1169                  * Sanity check this iovec. 0 read succeeds.
1170                  */
1171                 if (unlikely(!len))
1172                         break;
1173                 error = -EFAULT;
1174                 if (unlikely(!base))
1175                         break;
1176
1177                 /*
1178                  * Get this base offset and number of pages, then map
1179                  * in the user pages.
1180                  */
1181                 off = (unsigned long) base & ~PAGE_MASK;
1182
1183                 /*
1184                  * If asked for alignment, the offset must be zero and the
1185                  * length a multiple of the PAGE_SIZE.
1186                  */
1187                 error = -EINVAL;
1188                 if (aligned && (off || len & ~PAGE_MASK))
1189                         break;
1190
1191                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1192                 if (npages > PIPE_BUFFERS - buffers)
1193                         npages = PIPE_BUFFERS - buffers;
1194
1195                 error = get_user_pages(current, current->mm,
1196                                        (unsigned long) base, npages, 0, 0,
1197                                        &pages[buffers], NULL);
1198
1199                 if (unlikely(error <= 0))
1200                         break;
1201
1202                 /*
1203                  * Fill this contiguous range into the partial page map.
1204                  */
1205                 for (i = 0; i < error; i++) {
1206                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1207
1208                         partial[buffers].offset = off;
1209                         partial[buffers].len = plen;
1210
1211                         off = 0;
1212                         len -= plen;
1213                         buffers++;
1214                 }
1215
1216                 /*
1217                  * We didn't complete this iov, stop here since it probably
1218                  * means we have to move some of this into a pipe to
1219                  * be able to continue.
1220                  */
1221                 if (len)
1222                         break;
1223
1224                 /*
1225                  * Don't continue if we mapped fewer pages than we asked for,
1226                  * or if we mapped the max number of pages that we have
1227                  * room for.
1228                  */
1229                 if (error < npages || buffers == PIPE_BUFFERS)
1230                         break;
1231
1232                 nr_vecs--;
1233                 iov++;
1234         }
1235
1236         up_read(&current->mm->mmap_sem);
1237
1238         if (buffers)
1239                 return buffers;
1240
1241         return error;
1242 }
1243
1244 /*
1245  * vmsplice splices a user address range into a pipe. It can be thought of
1246  * as splice-from-memory, where the regular splice is splice-from-file (or
1247  * to file). In both cases the output is a pipe, naturally.
1248  *
1249  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1250  * not the other way around. Splicing from user memory is a simple operation
1251  * that can be supported without any funky alignment restrictions or nasty
1252  * vm tricks. We simply map in the user memory and fill them into a pipe.
1253  * The reverse isn't quite as easy, though. There are two possible solutions
1254  * for that:
1255  *
1256  *      - memcpy() the data internally, at which point we might as well just
1257  *        do a regular read() on the buffer anyway.
1258  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1259  *        has restriction limitations on both ends of the pipe).
1260  *
1261  * Alas, it isn't here.
1262  *
1263  */
1264 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1265                         unsigned long nr_segs, unsigned int flags)
1266 {
1267         struct pipe_inode_info *pipe;
1268         struct page *pages[PIPE_BUFFERS];
1269         struct partial_page partial[PIPE_BUFFERS];
1270         struct splice_pipe_desc spd = {
1271                 .pages = pages,
1272                 .partial = partial,
1273                 .flags = flags,
1274                 .ops = &user_page_pipe_buf_ops,
1275         };
1276
1277         pipe = pipe_info(file->f_path.dentry->d_inode);
1278         if (!pipe)
1279                 return -EBADF;
1280         if (unlikely(nr_segs > UIO_MAXIOV))
1281                 return -EINVAL;
1282         else if (unlikely(!nr_segs))
1283                 return 0;
1284
1285         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1286                                             flags & SPLICE_F_GIFT);
1287         if (spd.nr_pages <= 0)
1288                 return spd.nr_pages;
1289
1290         return splice_to_pipe(pipe, &spd);
1291 }
1292
1293 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1294                              unsigned long nr_segs, unsigned int flags)
1295 {
1296         struct file *file;
1297         long error;
1298         int fput;
1299
1300         error = -EBADF;
1301         file = fget_light(fd, &fput);
1302         if (file) {
1303                 if (file->f_mode & FMODE_WRITE)
1304                         error = do_vmsplice(file, iov, nr_segs, flags);
1305
1306                 fput_light(file, fput);
1307         }
1308
1309         return error;
1310 }
1311
1312 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1313                            int fd_out, loff_t __user *off_out,
1314                            size_t len, unsigned int flags)
1315 {
1316         long error;
1317         struct file *in, *out;
1318         int fput_in, fput_out;
1319
1320         if (unlikely(!len))
1321                 return 0;
1322
1323         error = -EBADF;
1324         in = fget_light(fd_in, &fput_in);
1325         if (in) {
1326                 if (in->f_mode & FMODE_READ) {
1327                         out = fget_light(fd_out, &fput_out);
1328                         if (out) {
1329                                 if (out->f_mode & FMODE_WRITE)
1330                                         error = do_splice(in, off_in,
1331                                                           out, off_out,
1332                                                           len, flags);
1333                                 fput_light(out, fput_out);
1334                         }
1335                 }
1336
1337                 fput_light(in, fput_in);
1338         }
1339
1340         return error;
1341 }
1342
1343 /*
1344  * Make sure there's data to read. Wait for input if we can, otherwise
1345  * return an appropriate error.
1346  */
1347 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1348 {
1349         int ret;
1350
1351         /*
1352          * Check ->nrbufs without the inode lock first. This function
1353          * is speculative anyways, so missing one is ok.
1354          */
1355         if (pipe->nrbufs)
1356                 return 0;
1357
1358         ret = 0;
1359         mutex_lock(&pipe->inode->i_mutex);
1360
1361         while (!pipe->nrbufs) {
1362                 if (signal_pending(current)) {
1363                         ret = -ERESTARTSYS;
1364                         break;
1365                 }
1366                 if (!pipe->writers)
1367                         break;
1368                 if (!pipe->waiting_writers) {
1369                         if (flags & SPLICE_F_NONBLOCK) {
1370                                 ret = -EAGAIN;
1371                                 break;
1372                         }
1373                 }
1374                 pipe_wait(pipe);
1375         }
1376
1377         mutex_unlock(&pipe->inode->i_mutex);
1378         return ret;
1379 }
1380
1381 /*
1382  * Make sure there's writeable room. Wait for room if we can, otherwise
1383  * return an appropriate error.
1384  */
1385 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1386 {
1387         int ret;
1388
1389         /*
1390          * Check ->nrbufs without the inode lock first. This function
1391          * is speculative anyways, so missing one is ok.
1392          */
1393         if (pipe->nrbufs < PIPE_BUFFERS)
1394                 return 0;
1395
1396         ret = 0;
1397         mutex_lock(&pipe->inode->i_mutex);
1398
1399         while (pipe->nrbufs >= PIPE_BUFFERS) {
1400                 if (!pipe->readers) {
1401                         send_sig(SIGPIPE, current, 0);
1402                         ret = -EPIPE;
1403                         break;
1404                 }
1405                 if (flags & SPLICE_F_NONBLOCK) {
1406                         ret = -EAGAIN;
1407                         break;
1408                 }
1409                 if (signal_pending(current)) {
1410                         ret = -ERESTARTSYS;
1411                         break;
1412                 }
1413                 pipe->waiting_writers++;
1414                 pipe_wait(pipe);
1415                 pipe->waiting_writers--;
1416         }
1417
1418         mutex_unlock(&pipe->inode->i_mutex);
1419         return ret;
1420 }
1421
1422 /*
1423  * Link contents of ipipe to opipe.
1424  */
1425 static int link_pipe(struct pipe_inode_info *ipipe,
1426                      struct pipe_inode_info *opipe,
1427                      size_t len, unsigned int flags)
1428 {
1429         struct pipe_buffer *ibuf, *obuf;
1430         int ret = 0, i = 0, nbuf;
1431
1432         /*
1433          * Potential ABBA deadlock, work around it by ordering lock
1434          * grabbing by inode address. Otherwise two different processes
1435          * could deadlock (one doing tee from A -> B, the other from B -> A).
1436          */
1437         inode_double_lock(ipipe->inode, opipe->inode);
1438
1439         do {
1440                 if (!opipe->readers) {
1441                         send_sig(SIGPIPE, current, 0);
1442                         if (!ret)
1443                                 ret = -EPIPE;
1444                         break;
1445                 }
1446
1447                 /*
1448                  * If we have iterated all input buffers or ran out of
1449                  * output room, break.
1450                  */
1451                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1452                         break;
1453
1454                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1455                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1456
1457                 /*
1458                  * Get a reference to this pipe buffer,
1459                  * so we can copy the contents over.
1460                  */
1461                 ibuf->ops->get(ipipe, ibuf);
1462
1463                 obuf = opipe->bufs + nbuf;
1464                 *obuf = *ibuf;
1465
1466                 /*
1467                  * Don't inherit the gift flag, we need to
1468                  * prevent multiple steals of this page.
1469                  */
1470                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1471
1472                 if (obuf->len > len)
1473                         obuf->len = len;
1474
1475                 opipe->nrbufs++;
1476                 ret += obuf->len;
1477                 len -= obuf->len;
1478                 i++;
1479         } while (len);
1480
1481         inode_double_unlock(ipipe->inode, opipe->inode);
1482
1483         /*
1484          * If we put data in the output pipe, wakeup any potential readers.
1485          */
1486         if (ret > 0) {
1487                 smp_mb();
1488                 if (waitqueue_active(&opipe->wait))
1489                         wake_up_interruptible(&opipe->wait);
1490                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1491         }
1492
1493         return ret;
1494 }
1495
1496 /*
1497  * This is a tee(1) implementation that works on pipes. It doesn't copy
1498  * any data, it simply references the 'in' pages on the 'out' pipe.
1499  * The 'flags' used are the SPLICE_F_* variants, currently the only
1500  * applicable one is SPLICE_F_NONBLOCK.
1501  */
1502 static long do_tee(struct file *in, struct file *out, size_t len,
1503                    unsigned int flags)
1504 {
1505         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1506         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1507         int ret = -EINVAL;
1508
1509         /*
1510          * Duplicate the contents of ipipe to opipe without actually
1511          * copying the data.
1512          */
1513         if (ipipe && opipe && ipipe != opipe) {
1514                 /*
1515                  * Keep going, unless we encounter an error. The ipipe/opipe
1516                  * ordering doesn't really matter.
1517                  */
1518                 ret = link_ipipe_prep(ipipe, flags);
1519                 if (!ret) {
1520                         ret = link_opipe_prep(opipe, flags);
1521                         if (!ret) {
1522                                 ret = link_pipe(ipipe, opipe, len, flags);
1523                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1524                                         ret = -EAGAIN;
1525                         }
1526                 }
1527         }
1528
1529         return ret;
1530 }
1531
1532 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1533 {
1534         struct file *in;
1535         int error, fput_in;
1536
1537         if (unlikely(!len))
1538                 return 0;
1539
1540         error = -EBADF;
1541         in = fget_light(fdin, &fput_in);
1542         if (in) {
1543                 if (in->f_mode & FMODE_READ) {
1544                         int fput_out;
1545                         struct file *out = fget_light(fdout, &fput_out);
1546
1547                         if (out) {
1548                                 if (out->f_mode & FMODE_WRITE)
1549                                         error = do_tee(in, out, len, flags);
1550                                 fput_light(out, fput_out);
1551                         }
1552                 }
1553                 fput_light(in, fput_in);
1554         }
1555
1556         return error;
1557 }