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         unsigned int spd_pages = spd->nr_pages;
 180         int ret, do_wakeup, page_nr;
 181
 182         ret = 0;
 183         do_wakeup = 0;
 184         page_nr = 0;
 185
 186         if (pipe->inode)
 187                 mutex_lock(&pipe->inode->i_mutex);
 188
 189         for (;;) {
 190                 if (!pipe->readers) {
 191                         send_sig(SIGPIPE, current, 0);
 192                         if (!ret)
 193                                 ret = -EPIPE;
 194                         break;
 195                 }
 196
 197                 if (pipe->nrbufs < PIPE_BUFFERS) {
 198                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 199                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 200
 201                         buf->page = spd->pages[page_nr];
 202                         buf->offset = spd->partial[page_nr].offset;
 203                         buf->len = spd->partial[page_nr].len;
 204                         buf->ops = spd->ops;
 205                         if (spd->flags & SPLICE_F_GIFT)
 206                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 207
 208                         pipe->nrbufs++;
 209                         page_nr++;
 210                         ret += buf->len;
 211
 212                         if (pipe->inode)
 213                                 do_wakeup = 1;
 214
 215                         if (!--spd->nr_pages)
 216                                 break;
 217                         if (pipe->nrbufs < PIPE_BUFFERS)
 218                                 continue;
 219
 220                         break;
 221                 }
 222
 223                 if (spd->flags & SPLICE_F_NONBLOCK) {
 224                         if (!ret)
 225                                 ret = -EAGAIN;
 226                         break;
 227                 }
 228
 229                 if (signal_pending(current)) {
 230                         if (!ret)
 231                                 ret = -ERESTARTSYS;
 232                         break;
 233                 }
 234
 235                 if (do_wakeup) {
 236                         smp_mb();
 237                         if (waitqueue_active(&pipe->wait))
 238                                 wake_up_interruptible_sync(&pipe->wait);
 239                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 240                         do_wakeup = 0;
 241                 }
 242
 243                 pipe->waiting_writers++;
 244                 pipe_wait(pipe);
 245                 pipe->waiting_writers--;
 246         }
 247
 248         if (pipe->inode) {
 249                 mutex_unlock(&pipe->inode->i_mutex);
 250
 251                 if (do_wakeup) {
 252                         smp_mb();
 253                         if (waitqueue_active(&pipe->wait))
 254                                 wake_up_interruptible(&pipe->wait);
 255                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 256                 }
 257         }
 258
 259         while (page_nr < spd_pages)
 260                 page_cache_release(spd->pages[page_nr++]);
 261
 262         return ret;
 263 }
 264
 265 static int
 266 __generic_file_splice_read(struct file *in, loff_t *ppos,
 267                            struct pipe_inode_info *pipe, size_t len,
 268                            unsigned int flags)
 269 {
 270         struct address_space *mapping = in->f_mapping;
 271         unsigned int loff, nr_pages;
 272         struct page *pages[PIPE_BUFFERS];
 273         struct partial_page partial[PIPE_BUFFERS];
 274         struct page *page;
 275         pgoff_t index, end_index;
 276         loff_t isize;
 277         int error, page_nr;
 278         struct splice_pipe_desc spd = {
 279                 .pages = pages,
 280                 .partial = partial,
 281                 .flags = flags,
 282                 .ops = &page_cache_pipe_buf_ops,
 283         };
 284
 285         index = *ppos >> PAGE_CACHE_SHIFT;
 286         loff = *ppos & ~PAGE_CACHE_MASK;
 287         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 288
 289         if (nr_pages > PIPE_BUFFERS)
 290                 nr_pages = PIPE_BUFFERS;
 291
 292         /*
 293          * Don't try to 2nd guess the read-ahead logic, call into
 294          * page_cache_readahead() like the page cache reads would do.
 295          */
 296         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 297
 298         /*
 299          * Now fill in the holes:
 300          */
 301         error = 0;
 302
 303         /*
 304          * Lookup the (hopefully) full range of pages we need.
 305          */
 306         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 307
 308         /*
 309          * If find_get_pages_contig() returned fewer pages than we needed,
 310          * allocate the rest.
 311          */
 312         index += spd.nr_pages;
 313         while (spd.nr_pages < nr_pages) {
 314                 /*
 315                  * Page could be there, find_get_pages_contig() breaks on
 316                  * the first hole.
 317                  */
 318                 page = find_get_page(mapping, index);
 319                 if (!page) {
 320                         /*
 321                          * Make sure the read-ahead engine is notified
 322                          * about this failure.
 323                          */
 324                         handle_ra_miss(mapping, &in->f_ra, index);
 325
 326                         /*
 327                          * page didn't exist, allocate one.
 328                          */
 329                         page = page_cache_alloc_cold(mapping);
 330                         if (!page)
 331                                 break;
 332
 333                         error = add_to_page_cache_lru(page, mapping, index,
 334                                               GFP_KERNEL);
 335                         if (unlikely(error)) {
 336                                 page_cache_release(page);
 337                                 if (error == -EEXIST)
 338                                         continue;
 339                                 break;
 340                         }
 341                         /*
 342                          * add_to_page_cache() locks the page, unlock it
 343                          * to avoid convoluting the logic below even more.
 344                          */
 345                         unlock_page(page);
 346                 }
 347
 348                 pages[spd.nr_pages++] = page;
 349                 index++;
 350         }
 351
 352         /*
 353          * Now loop over the map and see if we need to start IO on any
 354          * pages, fill in the partial map, etc.
 355          */
 356         index = *ppos >> PAGE_CACHE_SHIFT;
 357         nr_pages = spd.nr_pages;
 358         spd.nr_pages = 0;
 359         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 360                 unsigned int this_len;
 361
 362                 if (!len)
 363                         break;
 364
 365                 /*
 366                  * this_len is the max we'll use from this page
 367                  */
 368                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 369                 page = pages[page_nr];
 370
 371                 /*
 372                  * If the page isn't uptodate, we may need to start io on it
 373                  */
 374                 if (!PageUptodate(page)) {
 375                         /*
 376                          * If in nonblock mode then dont block on waiting
 377                          * for an in-flight io page
 378                          */
 379                         if (flags & SPLICE_F_NONBLOCK) {
 380                                 if (TestSetPageLocked(page))
 381                                         break;
 382                         } else
 383                                 lock_page(page);
 384
 385                         /*
 386                          * page was truncated, stop here. if this isn't the
 387                          * first page, we'll just complete what we already
 388                          * added
 389                          */
 390                         if (!page->mapping) {
 391                                 unlock_page(page);
 392                                 break;
 393                         }
 394                         /*
 395                          * page was already under io and is now done, great
 396                          */
 397                         if (PageUptodate(page)) {
 398                                 unlock_page(page);
 399                                 goto fill_it;
 400                         }
 401
 402                         /*
 403                          * need to read in the page
 404                          */
 405                         error = mapping->a_ops->readpage(in, page);
 406                         if (unlikely(error)) {
 407                                 /*
 408                                  * We really should re-lookup the page here,
 409                                  * but it complicates things a lot. Instead
 410                                  * lets just do what we already stored, and
 411                                  * we'll get it the next time we are called.
 412                                  */
 413                                 if (error == AOP_TRUNCATED_PAGE)
 414                                         error = 0;
 415
 416                                 break;
 417                         }
 418                 }
 419 fill_it:
 420                 /*
 421                  * i_size must be checked after PageUptodate.
 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                         unsigned int plen;
 434
 435                         /*
 436                          * max good bytes in this page
 437                          */
 438                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 439                         if (plen <= loff)
 440                                 break;
 441
 442                         /*
 443                          * force quit after adding this page
 444                          */
 445                         this_len = min(this_len, plen - loff);
 446                         len = this_len;
 447                 }
 448
 449                 partial[page_nr].offset = loff;
 450                 partial[page_nr].len = this_len;
 451                 len -= this_len;
 452                 loff = 0;
 453                 spd.nr_pages++;
 454                 index++;
 455         }
 456
 457         /*
 458          * Release any pages at the end, if we quit early. 'page_nr' is how far
 459          * we got, 'nr_pages' is how many pages are in the map.
 460          */
 461         while (page_nr < nr_pages)
 462                 page_cache_release(pages[page_nr++]);
 463
 464         if (spd.nr_pages)
 465                 return splice_to_pipe(pipe, &spd);
 466
 467         return error;
 468 }
 469
 470 /**
 471  * generic_file_splice_read - splice data from file to a pipe
 472  * @in:         file to splice from
 473  * @pipe:       pipe to splice to
 474  * @len:        number of bytes to splice
 475  * @flags:      splice modifier flags
 476  *
 477  * Will read pages from given file and fill them into a pipe.
 478  */
 479 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 480                                  struct pipe_inode_info *pipe, size_t len,
 481                                  unsigned int flags)
 482 {
 483         ssize_t spliced;
 484         int ret;
 485         loff_t isize, left;
 486
 487         isize = i_size_read(in->f_mapping->host);
 488         if (unlikely(*ppos >= isize))
 489                 return 0;
 490
 491         left = isize - *ppos;
 492         if (unlikely(left < len))
 493                 len = left;
 494
 495         ret = 0;
 496         spliced = 0;
 497         while (len) {
 498                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 499
 500                 if (ret < 0)
 501                         break;
 502                 else if (!ret) {
 503                         if (spliced)
 504                                 break;
 505                         if (flags & SPLICE_F_NONBLOCK) {
 506                                 ret = -EAGAIN;
 507                                 break;
 508                         }
 509                 }
 510
 511                 *ppos += ret;
 512                 len -= ret;
 513                 spliced += ret;
 514         }
 515
 516         if (spliced)
 517                 return spliced;
 518
 519         return ret;
 520 }
 521
 522 EXPORT_SYMBOL(generic_file_splice_read);
 523
 524 /*
 525  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 526  * using sendpage(). Return the number of bytes sent.
 527  */
 528 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 529                             struct pipe_buffer *buf, struct splice_desc *sd)
 530 {
 531         struct file *file = sd->file;
 532         loff_t pos = sd->pos;
 533         int ret, more;
 534
 535         ret = buf->ops->pin(pipe, buf);
 536         if (!ret) {
 537                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 538
 539                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 540                                            sd->len, &pos, more);
 541         }
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * This is a little more tricky than the file -> pipe splicing. There are
 548  * basically three cases:
 549  *
 550  *      - Destination page already exists in the address space and there
 551  *        are users of it. For that case we have no other option that
 552  *        copying the data. Tough luck.
 553  *      - Destination page already exists in the address space, but there
 554  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 555  *        through to last case.
 556  *      - Destination page does not exist, we can add the pipe page to
 557  *        the page cache and avoid the copy.
 558  *
 559  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 560  * sd->flags), we attempt to migrate pages from the pipe to the output
 561  * file address space page cache. This is possible if no one else has
 562  * the pipe page referenced outside of the pipe and page cache. If
 563  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 564  * a new page in the output file page cache and fill/dirty that.
 565  */
 566 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 567                         struct splice_desc *sd)
 568 {
 569         struct file *file = sd->file;
 570         struct address_space *mapping = file->f_mapping;
 571         unsigned int offset, this_len;
 572         struct page *page;
 573         pgoff_t index;
 574         int ret;
 575
 576         /*
 577          * make sure the data in this buffer is uptodate
 578          */
 579         ret = buf->ops->pin(pipe, buf);
 580         if (unlikely(ret))
 581                 return ret;
 582
 583         index = sd->pos >> PAGE_CACHE_SHIFT;
 584         offset = sd->pos & ~PAGE_CACHE_MASK;
 585
 586         this_len = sd->len;
 587         if (this_len + offset > PAGE_CACHE_SIZE)
 588                 this_len = PAGE_CACHE_SIZE - offset;
 589
 590 find_page:
 591         page = find_lock_page(mapping, index);
 592         if (!page) {
 593                 ret = -ENOMEM;
 594                 page = page_cache_alloc_cold(mapping);
 595                 if (unlikely(!page))
 596                         goto out_ret;
 597
 598                 /*
 599                  * This will also lock the page
 600                  */
 601                 ret = add_to_page_cache_lru(page, mapping, index,
 602                                             GFP_KERNEL);
 603                 if (unlikely(ret))
 604                         goto out_release;
 605         }
 606
 607         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 608         if (unlikely(ret)) {
 609                 loff_t isize = i_size_read(mapping->host);
 610
 611                 if (ret != AOP_TRUNCATED_PAGE)
 612                         unlock_page(page);
 613                 page_cache_release(page);
 614                 if (ret == AOP_TRUNCATED_PAGE)
 615                         goto find_page;
 616
 617                 /*
 618                  * prepare_write() may have instantiated a few blocks
 619                  * outside i_size.  Trim these off again.
 620                  */
 621                 if (sd->pos + this_len > isize)
 622                         vmtruncate(mapping->host, isize);
 623
 624                 goto out_ret;
 625         }
 626
 627         if (buf->page != page) {
 628                 /*
 629                  * Careful, ->map() uses KM_USER0!
 630                  */
 631                 char *src = buf->ops->map(pipe, buf, 1);
 632                 char *dst = kmap_atomic(page, KM_USER1);
 633
 634                 memcpy(dst + offset, src + buf->offset, this_len);
 635                 flush_dcache_page(page);
 636                 kunmap_atomic(dst, KM_USER1);
 637                 buf->ops->unmap(pipe, buf, src);
 638         }
 639
 640         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 641         if (ret) {
 642                 if (ret == AOP_TRUNCATED_PAGE) {
 643                         page_cache_release(page);
 644                         goto find_page;
 645                 }
 646                 if (ret < 0)
 647                         goto out;
 648                 /*
 649                  * Partial write has happened, so 'ret' already initialized by
 650                  * number of bytes written, Where is nothing we have to do here.
 651                  */
 652         } else
 653                 ret = this_len;
 654         /*
 655          * Return the number of bytes written and mark page as
 656          * accessed, we are now done!
 657          */
 658         mark_page_accessed(page);
 659 out:
 660         unlock_page(page);
 661 out_release:
 662         page_cache_release(page);
 663 out_ret:
 664         return ret;
 665 }
 666
 667 /*
 668  * Pipe input worker. Most of this logic works like a regular pipe, the
 669  * key here is the 'actor' worker passed in that actually moves the data
 670  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 671  */
 672 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 673                            struct file *out, loff_t *ppos, size_t len,
 674                            unsigned int flags, splice_actor *actor)
 675 {
 676         int ret, do_wakeup, err;
 677         struct splice_desc sd;
 678
 679         ret = 0;
 680         do_wakeup = 0;
 681
 682         sd.total_len = len;
 683         sd.flags = flags;
 684         sd.file = out;
 685         sd.pos = *ppos;
 686
 687         for (;;) {
 688                 if (pipe->nrbufs) {
 689                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 690                         const struct pipe_buf_operations *ops = buf->ops;
 691
 692                         sd.len = buf->len;
 693                         if (sd.len > sd.total_len)
 694                                 sd.len = sd.total_len;
 695
 696                         err = actor(pipe, buf, &sd);
 697                         if (err <= 0) {
 698                                 if (!ret && err != -ENODATA)
 699                                         ret = err;
 700
 701                                 break;
 702                         }
 703
 704                         ret += err;
 705                         buf->offset += err;
 706                         buf->len -= err;
 707
 708                         sd.len -= err;
 709                         sd.pos += err;
 710                         sd.total_len -= err;
 711                         if (sd.len)
 712                                 continue;
 713
 714                         if (!buf->len) {
 715                                 buf->ops = NULL;
 716                                 ops->release(pipe, buf);
 717                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 718                                 pipe->nrbufs--;
 719                                 if (pipe->inode)
 720                                         do_wakeup = 1;
 721                         }
 722
 723                         if (!sd.total_len)
 724                                 break;
 725                 }
 726
 727                 if (pipe->nrbufs)
 728                         continue;
 729                 if (!pipe->writers)
 730                         break;
 731                 if (!pipe->waiting_writers) {
 732                         if (ret)
 733                                 break;
 734                 }
 735
 736                 if (flags & SPLICE_F_NONBLOCK) {
 737                         if (!ret)
 738                                 ret = -EAGAIN;
 739                         break;
 740                 }
 741
 742                 if (signal_pending(current)) {
 743                         if (!ret)
 744                                 ret = -ERESTARTSYS;
 745                         break;
 746                 }
 747
 748                 if (do_wakeup) {
 749                         smp_mb();
 750                         if (waitqueue_active(&pipe->wait))
 751                                 wake_up_interruptible_sync(&pipe->wait);
 752                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 753                         do_wakeup = 0;
 754                 }
 755
 756                 pipe_wait(pipe);
 757         }
 758
 759         if (do_wakeup) {
 760                 smp_mb();
 761                 if (waitqueue_active(&pipe->wait))
 762                         wake_up_interruptible(&pipe->wait);
 763                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 764         }
 765
 766         return ret;
 767 }
 768 EXPORT_SYMBOL(__splice_from_pipe);
 769
 770 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 771                          loff_t *ppos, size_t len, unsigned int flags,
 772                          splice_actor *actor)
 773 {
 774         ssize_t ret;
 775         struct inode *inode = out->f_mapping->host;
 776
 777         /*
 778          * The actor worker might be calling ->prepare_write and
 779          * ->commit_write. Most of the time, these expect i_mutex to
 780          * be held. Since this may result in an ABBA deadlock with
 781          * pipe->inode, we have to order lock acquiry here.
 782          */
 783         inode_double_lock(inode, pipe->inode);
 784         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 785         inode_double_unlock(inode, pipe->inode);
 786
 787         return ret;
 788 }
 789
 790 /**
 791  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 792  * @pipe:       pipe info
 793  * @out:        file to write to
 794  * @len:        number of bytes to splice
 795  * @flags:      splice modifier flags
 796  *
 797  * Will either move or copy pages (determined by @flags options) from
 798  * the given pipe inode to the given file. The caller is responsible
 799  * for acquiring i_mutex on both inodes.
 800  *
 801  */
 802 ssize_t
 803 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 804                                  loff_t *ppos, size_t len, unsigned int flags)
 805 {
 806         struct address_space *mapping = out->f_mapping;
 807         struct inode *inode = mapping->host;
 808         ssize_t ret;
 809         int err;
 810
 811         err = remove_suid(out->f_path.dentry);
 812         if (unlikely(err))
 813                 return err;
 814
 815         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 816         if (ret > 0) {
 817                 unsigned long nr_pages;
 818
 819                 *ppos += ret;
 820                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 821
 822                 /*
 823                  * If file or inode is SYNC and we actually wrote some data,
 824                  * sync it.
 825                  */
 826                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 827                         err = generic_osync_inode(inode, mapping,
 828                                                   OSYNC_METADATA|OSYNC_DATA);
 829
 830                         if (err)
 831                                 ret = err;
 832                 }
 833                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 834         }
 835
 836         return ret;
 837 }
 838
 839 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 840
 841 /**
 842  * generic_file_splice_write - splice data from a pipe to a file
 843  * @pipe:       pipe info
 844  * @out:        file to write to
 845  * @len:        number of bytes to splice
 846  * @flags:      splice modifier flags
 847  *
 848  * Will either move or copy pages (determined by @flags options) from
 849  * the given pipe inode to the given file.
 850  *
 851  */
 852 ssize_t
 853 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 854                           loff_t *ppos, size_t len, unsigned int flags)
 855 {
 856         struct address_space *mapping = out->f_mapping;
 857         struct inode *inode = mapping->host;
 858         ssize_t ret;
 859         int err;
 860
 861         err = should_remove_suid(out->f_path.dentry);
 862         if (unlikely(err)) {
 863                 mutex_lock(&inode->i_mutex);
 864                 err = __remove_suid(out->f_path.dentry, err);
 865                 mutex_unlock(&inode->i_mutex);
 866                 if (err)
 867                         return err;
 868         }
 869
 870         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 871         if (ret > 0) {
 872                 unsigned long nr_pages;
 873
 874                 *ppos += ret;
 875                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 876
 877                 /*
 878                  * If file or inode is SYNC and we actually wrote some data,
 879                  * sync it.
 880                  */
 881                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 882                         mutex_lock(&inode->i_mutex);
 883                         err = generic_osync_inode(inode, mapping,
 884                                                   OSYNC_METADATA|OSYNC_DATA);
 885                         mutex_unlock(&inode->i_mutex);
 886
 887                         if (err)
 888                                 ret = err;
 889                 }
 890                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 891         }
 892
 893         return ret;
 894 }
 895
 896 EXPORT_SYMBOL(generic_file_splice_write);
 897
 898 /**
 899  * generic_splice_sendpage - splice data from a pipe to a socket
 900  * @inode:      pipe inode
 901  * @out:        socket to write to
 902  * @len:        number of bytes to splice
 903  * @flags:      splice modifier flags
 904  *
 905  * Will send @len bytes from the pipe to a network socket. No data copying
 906  * is involved.
 907  *
 908  */
 909 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 910                                 loff_t *ppos, size_t len, unsigned int flags)
 911 {
 912         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 913 }
 914
 915 EXPORT_SYMBOL(generic_splice_sendpage);
 916
 917 /*
 918  * Attempt to initiate a splice from pipe to file.
 919  */
 920 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 921                            loff_t *ppos, size_t len, unsigned int flags)
 922 {
 923         int ret;
 924
 925         if (unlikely(!out->f_op || !out->f_op->splice_write))
 926                 return -EINVAL;
 927
 928         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 929                 return -EBADF;
 930
 931         ret = rw_verify_area(WRITE, out, ppos, len);
 932         if (unlikely(ret < 0))
 933                 return ret;
 934
 935         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 936 }
 937
 938 /*
 939  * Attempt to initiate a splice from a file to a pipe.
 940  */
 941 static long do_splice_to(struct file *in, loff_t *ppos,
 942                          struct pipe_inode_info *pipe, size_t len,
 943                          unsigned int flags)
 944 {
 945         int ret;
 946
 947         if (unlikely(!in->f_op || !in->f_op->splice_read))
 948                 return -EINVAL;
 949
 950         if (unlikely(!(in->f_mode & FMODE_READ)))
 951                 return -EBADF;
 952
 953         ret = rw_verify_area(READ, in, ppos, len);
 954         if (unlikely(ret < 0))
 955                 return ret;
 956
 957         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 958 }
 959
 960 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 961                       size_t len, unsigned int flags)
 962 {
 963         struct pipe_inode_info *pipe;
 964         long ret, bytes;
 965         loff_t out_off;
 966         umode_t i_mode;
 967         int i;
 968
 969         /*
 970          * We require the input being a regular file, as we don't want to
 971          * randomly drop data for eg socket -> socket splicing. Use the
 972          * piped splicing for that!
 973          */
 974         i_mode = in->f_path.dentry->d_inode->i_mode;
 975         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 976                 return -EINVAL;
 977
 978         /*
 979          * neither in nor out is a pipe, setup an internal pipe attached to
 980          * 'out' and transfer the wanted data from 'in' to 'out' through that
 981          */
 982         pipe = current->splice_pipe;
 983         if (unlikely(!pipe)) {
 984                 pipe = alloc_pipe_info(NULL);
 985                 if (!pipe)
 986                         return -ENOMEM;
 987
 988                 /*
 989                  * We don't have an immediate reader, but we'll read the stuff
 990                  * out of the pipe right after the splice_to_pipe(). So set
 991                  * PIPE_READERS appropriately.
 992                  */
 993                 pipe->readers = 1;
 994
 995                 current->splice_pipe = pipe;
 996         }
 997
 998         /*
 999          * Do the splice.
1000          */
1001         ret = 0;
1002         bytes = 0;
1003         out_off = 0;
1004
1005         while (len) {
1006                 size_t read_len, max_read_len;
1007
1008                 /*
1009                  * Do at most PIPE_BUFFERS pages worth of transfer:
1010                  */
1011                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1012
1013                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1014                 if (unlikely(ret <= 0))
1015                         goto out_release;
1016
1017                 read_len = ret;
1018
1019                 /*
1020                  * NOTE: nonblocking mode only applies to the input. We
1021                  * must not do the output in nonblocking mode as then we
1022                  * could get stuck data in the internal pipe:
1023                  */
1024                 ret = do_splice_from(pipe, out, &out_off, read_len,
1025                                      flags & ~SPLICE_F_NONBLOCK);
1026                 if (unlikely(ret <= 0))
1027                         goto out_release;
1028
1029                 bytes += ret;
1030                 len -= ret;
1031
1032                 /*
1033                  * In nonblocking mode, if we got back a short read then
1034                  * that was due to either an IO error or due to the
1035                  * pagecache entry not being there. In the IO error case
1036                  * the _next_ splice attempt will produce a clean IO error
1037                  * return value (not a short read), so in both cases it's
1038                  * correct to break out of the loop here:
1039                  */
1040                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1041                         break;
1042         }
1043
1044         pipe->nrbufs = pipe->curbuf = 0;
1045
1046         return bytes;
1047
1048 out_release:
1049         /*
1050          * If we did an incomplete transfer we must release
1051          * the pipe buffers in question:
1052          */
1053         for (i = 0; i < PIPE_BUFFERS; i++) {
1054                 struct pipe_buffer *buf = pipe->bufs + i;
1055
1056                 if (buf->ops) {
1057                         buf->ops->release(pipe, buf);
1058                         buf->ops = NULL;
1059                 }
1060         }
1061         pipe->nrbufs = pipe->curbuf = 0;
1062
1063         /*
1064          * If we transferred some data, return the number of bytes:
1065          */
1066         if (bytes > 0)
1067                 return bytes;
1068
1069         return ret;
1070 }
1071
1072 /*
1073  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1074  * location, so checking ->i_pipe is not enough to verify that this is a
1075  * pipe.
1076  */
1077 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1078 {
1079         if (S_ISFIFO(inode->i_mode))
1080                 return inode->i_pipe;
1081
1082         return NULL;
1083 }
1084
1085 /*
1086  * Determine where to splice to/from.
1087  */
1088 static long do_splice(struct file *in, loff_t __user *off_in,
1089                       struct file *out, loff_t __user *off_out,
1090                       size_t len, unsigned int flags)
1091 {
1092         struct pipe_inode_info *pipe;
1093         loff_t offset, *off;
1094         long ret;
1095
1096         pipe = pipe_info(in->f_path.dentry->d_inode);
1097         if (pipe) {
1098                 if (off_in)
1099                         return -ESPIPE;
1100                 if (off_out) {
1101                         if (out->f_op->llseek == no_llseek)
1102                                 return -EINVAL;
1103                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1104                                 return -EFAULT;
1105                         off = &offset;
1106                 } else
1107                         off = &out->f_pos;
1108
1109                 ret = do_splice_from(pipe, out, off, len, flags);
1110
1111                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1112                         ret = -EFAULT;
1113
1114                 return ret;
1115         }
1116
1117         pipe = pipe_info(out->f_path.dentry->d_inode);
1118         if (pipe) {
1119                 if (off_out)
1120                         return -ESPIPE;
1121                 if (off_in) {
1122                         if (in->f_op->llseek == no_llseek)
1123                                 return -EINVAL;
1124                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1125                                 return -EFAULT;
1126                         off = &offset;
1127                 } else
1128                         off = &in->f_pos;
1129
1130                 ret = do_splice_to(in, off, pipe, len, flags);
1131
1132                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1133                         ret = -EFAULT;
1134
1135                 return ret;
1136         }
1137
1138         return -EINVAL;
1139 }
1140
1141 /*
1142  * Map an iov into an array of pages and offset/length tupples. With the
1143  * partial_page structure, we can map several non-contiguous ranges into
1144  * our ones pages[] map instead of splitting that operation into pieces.
1145  * Could easily be exported as a generic helper for other users, in which
1146  * case one would probably want to add a 'max_nr_pages' parameter as well.
1147  */
1148 static int get_iovec_page_array(const struct iovec __user *iov,
1149                                 unsigned int nr_vecs, struct page **pages,
1150                                 struct partial_page *partial, int aligned)
1151 {
1152         int buffers = 0, error = 0;
1153
1154         /*
1155          * It's ok to take the mmap_sem for reading, even
1156          * across a "get_user()".
1157          */
1158         down_read(&current->mm->mmap_sem);
1159
1160         while (nr_vecs) {
1161                 unsigned long off, npages;
1162                 void __user *base;
1163                 size_t len;
1164                 int i;
1165
1166                 /*
1167                  * Get user address base and length for this iovec.
1168                  */
1169                 error = get_user(base, &iov->iov_base);
1170                 if (unlikely(error))
1171                         break;
1172                 error = get_user(len, &iov->iov_len);
1173                 if (unlikely(error))
1174                         break;
1175
1176                 /*
1177                  * Sanity check this iovec. 0 read succeeds.
1178                  */
1179                 if (unlikely(!len))
1180                         break;
1181                 error = -EFAULT;
1182                 if (unlikely(!base))
1183                         break;
1184
1185                 if (!access_ok(VERIFY_READ, base, len))
1186                         break;
1187
1188                 /*
1189                  * Get this base offset and number of pages, then map
1190                  * in the user pages.
1191                  */
1192                 off = (unsigned long) base & ~PAGE_MASK;
1193
1194                 /*
1195                  * If asked for alignment, the offset must be zero and the
1196                  * length a multiple of the PAGE_SIZE.
1197                  */
1198                 error = -EINVAL;
1199                 if (aligned && (off || len & ~PAGE_MASK))
1200                         break;
1201
1202                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1203                 if (npages > PIPE_BUFFERS - buffers)
1204                         npages = PIPE_BUFFERS - buffers;
1205
1206                 error = get_user_pages(current, current->mm,
1207                                        (unsigned long) base, npages, 0, 0,
1208                                        &pages[buffers], NULL);
1209
1210                 if (unlikely(error <= 0))
1211                         break;
1212
1213                 /*
1214                  * Fill this contiguous range into the partial page map.
1215                  */
1216                 for (i = 0; i < error; i++) {
1217                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1218
1219                         partial[buffers].offset = off;
1220                         partial[buffers].len = plen;
1221
1222                         off = 0;
1223                         len -= plen;
1224                         buffers++;
1225                 }
1226
1227                 /*
1228                  * We didn't complete this iov, stop here since it probably
1229                  * means we have to move some of this into a pipe to
1230                  * be able to continue.
1231                  */
1232                 if (len)
1233                         break;
1234
1235                 /*
1236                  * Don't continue if we mapped fewer pages than we asked for,
1237                  * or if we mapped the max number of pages that we have
1238                  * room for.
1239                  */
1240                 if (error < npages || buffers == PIPE_BUFFERS)
1241                         break;
1242
1243                 nr_vecs--;
1244                 iov++;
1245         }
1246
1247         up_read(&current->mm->mmap_sem);
1248
1249         if (buffers)
1250                 return buffers;
1251
1252         return error;
1253 }
1254
1255 /*
1256  * vmsplice splices a user address range into a pipe. It can be thought of
1257  * as splice-from-memory, where the regular splice is splice-from-file (or
1258  * to file). In both cases the output is a pipe, naturally.
1259  *
1260  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1261  * not the other way around. Splicing from user memory is a simple operation
1262  * that can be supported without any funky alignment restrictions or nasty
1263  * vm tricks. We simply map in the user memory and fill them into a pipe.
1264  * The reverse isn't quite as easy, though. There are two possible solutions
1265  * for that:
1266  *
1267  *      - memcpy() the data internally, at which point we might as well just
1268  *        do a regular read() on the buffer anyway.
1269  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1270  *        has restriction limitations on both ends of the pipe).
1271  *
1272  * Alas, it isn't here.
1273  *
1274  */
1275 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1276                         unsigned long nr_segs, unsigned int flags)
1277 {
1278         struct pipe_inode_info *pipe;
1279         struct page *pages[PIPE_BUFFERS];
1280         struct partial_page partial[PIPE_BUFFERS];
1281         struct splice_pipe_desc spd = {
1282                 .pages = pages,
1283                 .partial = partial,
1284                 .flags = flags,
1285                 .ops = &user_page_pipe_buf_ops,
1286         };
1287
1288         pipe = pipe_info(file->f_path.dentry->d_inode);
1289         if (!pipe)
1290                 return -EBADF;
1291         if (unlikely(nr_segs > UIO_MAXIOV))
1292                 return -EINVAL;
1293         else if (unlikely(!nr_segs))
1294                 return 0;
1295
1296         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1297                                             flags & SPLICE_F_GIFT);
1298         if (spd.nr_pages <= 0)
1299                 return spd.nr_pages;
1300
1301         return splice_to_pipe(pipe, &spd);
1302 }
1303
1304 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1305                              unsigned long nr_segs, unsigned int flags)
1306 {
1307         struct file *file;
1308         long error;
1309         int fput;
1310
1311         error = -EBADF;
1312         file = fget_light(fd, &fput);
1313         if (file) {
1314                 if (file->f_mode & FMODE_WRITE)
1315                         error = do_vmsplice(file, iov, nr_segs, flags);
1316
1317                 fput_light(file, fput);
1318         }
1319
1320         return error;
1321 }
1322
1323 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1324                            int fd_out, loff_t __user *off_out,
1325                            size_t len, unsigned int flags)
1326 {
1327         long error;
1328         struct file *in, *out;
1329         int fput_in, fput_out;
1330
1331         if (unlikely(!len))
1332                 return 0;
1333
1334         error = -EBADF;
1335         in = fget_light(fd_in, &fput_in);
1336         if (in) {
1337                 if (in->f_mode & FMODE_READ) {
1338                         out = fget_light(fd_out, &fput_out);
1339                         if (out) {
1340                                 if (out->f_mode & FMODE_WRITE)
1341                                         error = do_splice(in, off_in,
1342                                                           out, off_out,
1343                                                           len, flags);
1344                                 fput_light(out, fput_out);
1345                         }
1346                 }
1347
1348                 fput_light(in, fput_in);
1349         }
1350
1351         return error;
1352 }
1353
1354 /*
1355  * Make sure there's data to read. Wait for input if we can, otherwise
1356  * return an appropriate error.
1357  */
1358 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1359 {
1360         int ret;
1361
1362         /*
1363          * Check ->nrbufs without the inode lock first. This function
1364          * is speculative anyways, so missing one is ok.
1365          */
1366         if (pipe->nrbufs)
1367                 return 0;
1368
1369         ret = 0;
1370         mutex_lock(&pipe->inode->i_mutex);
1371
1372         while (!pipe->nrbufs) {
1373                 if (signal_pending(current)) {
1374                         ret = -ERESTARTSYS;
1375                         break;
1376                 }
1377                 if (!pipe->writers)
1378                         break;
1379                 if (!pipe->waiting_writers) {
1380                         if (flags & SPLICE_F_NONBLOCK) {
1381                                 ret = -EAGAIN;
1382                                 break;
1383                         }
1384                 }
1385                 pipe_wait(pipe);
1386         }
1387
1388         mutex_unlock(&pipe->inode->i_mutex);
1389         return ret;
1390 }
1391
1392 /*
1393  * Make sure there's writeable room. Wait for room if we can, otherwise
1394  * return an appropriate error.
1395  */
1396 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1397 {
1398         int ret;
1399
1400         /*
1401          * Check ->nrbufs without the inode lock first. This function
1402          * is speculative anyways, so missing one is ok.
1403          */
1404         if (pipe->nrbufs < PIPE_BUFFERS)
1405                 return 0;
1406
1407         ret = 0;
1408         mutex_lock(&pipe->inode->i_mutex);
1409
1410         while (pipe->nrbufs >= PIPE_BUFFERS) {
1411                 if (!pipe->readers) {
1412                         send_sig(SIGPIPE, current, 0);
1413                         ret = -EPIPE;
1414                         break;
1415                 }
1416                 if (flags & SPLICE_F_NONBLOCK) {
1417                         ret = -EAGAIN;
1418                         break;
1419                 }
1420                 if (signal_pending(current)) {
1421                         ret = -ERESTARTSYS;
1422                         break;
1423                 }
1424                 pipe->waiting_writers++;
1425                 pipe_wait(pipe);
1426                 pipe->waiting_writers--;
1427         }
1428
1429         mutex_unlock(&pipe->inode->i_mutex);
1430         return ret;
1431 }
1432
1433 /*
1434  * Link contents of ipipe to opipe.
1435  */
1436 static int link_pipe(struct pipe_inode_info *ipipe,
1437                      struct pipe_inode_info *opipe,
1438                      size_t len, unsigned int flags)
1439 {
1440         struct pipe_buffer *ibuf, *obuf;
1441         int ret = 0, i = 0, nbuf;
1442
1443         /*
1444          * Potential ABBA deadlock, work around it by ordering lock
1445          * grabbing by inode address. Otherwise two different processes
1446          * could deadlock (one doing tee from A -> B, the other from B -> A).
1447          */
1448         inode_double_lock(ipipe->inode, opipe->inode);
1449
1450         do {
1451                 if (!opipe->readers) {
1452                         send_sig(SIGPIPE, current, 0);
1453                         if (!ret)
1454                                 ret = -EPIPE;
1455                         break;
1456                 }
1457
1458                 /*
1459                  * If we have iterated all input buffers or ran out of
1460                  * output room, break.
1461                  */
1462                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1463                         break;
1464
1465                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1466                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1467
1468                 /*
1469                  * Get a reference to this pipe buffer,
1470                  * so we can copy the contents over.
1471                  */
1472                 ibuf->ops->get(ipipe, ibuf);
1473
1474                 obuf = opipe->bufs + nbuf;
1475                 *obuf = *ibuf;
1476
1477                 /*
1478                  * Don't inherit the gift flag, we need to
1479                  * prevent multiple steals of this page.
1480                  */
1481                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1482
1483                 if (obuf->len > len)
1484                         obuf->len = len;
1485
1486                 opipe->nrbufs++;
1487                 ret += obuf->len;
1488                 len -= obuf->len;
1489                 i++;
1490         } while (len);
1491
1492         inode_double_unlock(ipipe->inode, opipe->inode);
1493
1494         /*
1495          * If we put data in the output pipe, wakeup any potential readers.
1496          */
1497         if (ret > 0) {
1498                 smp_mb();
1499                 if (waitqueue_active(&opipe->wait))
1500                         wake_up_interruptible(&opipe->wait);
1501                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1502         }
1503
1504         return ret;
1505 }
1506
1507 /*
1508  * This is a tee(1) implementation that works on pipes. It doesn't copy
1509  * any data, it simply references the 'in' pages on the 'out' pipe.
1510  * The 'flags' used are the SPLICE_F_* variants, currently the only
1511  * applicable one is SPLICE_F_NONBLOCK.
1512  */
1513 static long do_tee(struct file *in, struct file *out, size_t len,
1514                    unsigned int flags)
1515 {
1516         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1517         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1518         int ret = -EINVAL;
1519
1520         /*
1521          * Duplicate the contents of ipipe to opipe without actually
1522          * copying the data.
1523          */
1524         if (ipipe && opipe && ipipe != opipe) {
1525                 /*
1526                  * Keep going, unless we encounter an error. The ipipe/opipe
1527                  * ordering doesn't really matter.
1528                  */
1529                 ret = link_ipipe_prep(ipipe, flags);
1530                 if (!ret) {
1531                         ret = link_opipe_prep(opipe, flags);
1532                         if (!ret) {
1533                                 ret = link_pipe(ipipe, opipe, len, flags);
1534                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1535                                         ret = -EAGAIN;
1536                         }
1537                 }
1538         }
1539
1540         return ret;
1541 }
1542
1543 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1544 {
1545         struct file *in;
1546         int error, fput_in;
1547
1548         if (unlikely(!len))
1549                 return 0;
1550
1551         error = -EBADF;
1552         in = fget_light(fdin, &fput_in);
1553         if (in) {
1554                 if (in->f_mode & FMODE_READ) {
1555                         int fput_out;
1556                         struct file *out = fget_light(fdout, &fput_out);
1557
1558                         if (out) {
1559                                 if (out->f_mode & FMODE_WRITE)
1560                                         error = do_tee(in, out, len, flags);
1561                                 fput_light(out, fput_out);
1562                         }
1563                 }
1564                 fput_light(in, fput_in);
1565         }
1566
1567         return error;
1568 }