SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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          * Don't try to 2nd guess the read-ahead logic, call into
 293          * page_cache_readahead() like the page cache reads would do.
 294          */
 295         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 296
 297         /*
 298          * Now fill in the holes:
 299          */
 300         error = 0;
 301         total_len = 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                         /*
 420                          * i_size must be checked after ->readpage().
 421                          */
 422                         isize = i_size_read(mapping->host);
 423                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 424                         if (unlikely(!isize || index > end_index))
 425                                 break;
 426
 427                         /*
 428                          * if this is the last page, see if we need to shrink
 429                          * the length and stop
 430                          */
 431                         if (end_index == index) {
 432                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 433                                 if (total_len + loff > isize)
 434                                         break;
 435                                 /*
 436                                  * force quit after adding this page
 437                                  */
 438                                 len = this_len;
 439                                 this_len = min(this_len, loff);
 440                                 loff = 0;
 441                         }
 442                 }
 443 fill_it:
 444                 partial[page_nr].offset = loff;
 445                 partial[page_nr].len = this_len;
 446                 len -= this_len;
 447                 total_len += this_len;
 448                 loff = 0;
 449                 spd.nr_pages++;
 450                 index++;
 451         }
 452
 453         /*
 454          * Release any pages at the end, if we quit early. 'i' is how far
 455          * we got, 'nr_pages' is how many pages are in the map.
 456          */
 457         while (page_nr < nr_pages)
 458                 page_cache_release(pages[page_nr++]);
 459
 460         if (spd.nr_pages)
 461                 return splice_to_pipe(pipe, &spd);
 462
 463         return error;
 464 }
 465
 466 /**
 467  * generic_file_splice_read - splice data from file to a pipe
 468  * @in:         file to splice from
 469  * @pipe:       pipe to splice to
 470  * @len:        number of bytes to splice
 471  * @flags:      splice modifier flags
 472  *
 473  * Will read pages from given file and fill them into a pipe.
 474  */
 475 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 476                                  struct pipe_inode_info *pipe, size_t len,
 477                                  unsigned int flags)
 478 {
 479         ssize_t spliced;
 480         int ret;
 481         loff_t isize, left;
 482
 483         isize = i_size_read(in->f_mapping->host);
 484         if (unlikely(*ppos >= isize))
 485                 return 0;
 486
 487         left = isize - *ppos;
 488         if (unlikely(left < len))
 489                 len = left;
 490
 491         ret = 0;
 492         spliced = 0;
 493         while (len) {
 494                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 495
 496                 if (ret < 0)
 497                         break;
 498                 else if (!ret) {
 499                         if (spliced)
 500                                 break;
 501                         if (flags & SPLICE_F_NONBLOCK) {
 502                                 ret = -EAGAIN;
 503                                 break;
 504                         }
 505                 }
 506
 507                 *ppos += ret;
 508                 len -= ret;
 509                 spliced += ret;
 510         }
 511
 512         if (spliced)
 513                 return spliced;
 514
 515         return ret;
 516 }
 517
 518 EXPORT_SYMBOL(generic_file_splice_read);
 519
 520 /*
 521  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 522  * using sendpage(). Return the number of bytes sent.
 523  */
 524 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 525                             struct pipe_buffer *buf, struct splice_desc *sd)
 526 {
 527         struct file *file = sd->file;
 528         loff_t pos = sd->pos;
 529         int ret, more;
 530
 531         ret = buf->ops->pin(pipe, buf);
 532         if (!ret) {
 533                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 534
 535                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 536                                            sd->len, &pos, more);
 537         }
 538
 539         return ret;
 540 }
 541
 542 /*
 543  * This is a little more tricky than the file -> pipe splicing. There are
 544  * basically three cases:
 545  *
 546  *      - Destination page already exists in the address space and there
 547  *        are users of it. For that case we have no other option that
 548  *        copying the data. Tough luck.
 549  *      - Destination page already exists in the address space, but there
 550  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 551  *        through to last case.
 552  *      - Destination page does not exist, we can add the pipe page to
 553  *        the page cache and avoid the copy.
 554  *
 555  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 556  * sd->flags), we attempt to migrate pages from the pipe to the output
 557  * file address space page cache. This is possible if no one else has
 558  * the pipe page referenced outside of the pipe and page cache. If
 559  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 560  * a new page in the output file page cache and fill/dirty that.
 561  */
 562 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 563                         struct splice_desc *sd)
 564 {
 565         struct file *file = sd->file;
 566         struct address_space *mapping = file->f_mapping;
 567         unsigned int offset, this_len;
 568         struct page *page;
 569         pgoff_t index;
 570         int ret;
 571
 572         /*
 573          * make sure the data in this buffer is uptodate
 574          */
 575         ret = buf->ops->pin(pipe, buf);
 576         if (unlikely(ret))
 577                 return ret;
 578
 579         index = sd->pos >> PAGE_CACHE_SHIFT;
 580         offset = sd->pos & ~PAGE_CACHE_MASK;
 581
 582         this_len = sd->len;
 583         if (this_len + offset > PAGE_CACHE_SIZE)
 584                 this_len = PAGE_CACHE_SIZE - offset;
 585
 586 find_page:
 587         page = find_lock_page(mapping, index);
 588         if (!page) {
 589                 ret = -ENOMEM;
 590                 page = page_cache_alloc_cold(mapping);
 591                 if (unlikely(!page))
 592                         goto out_ret;
 593
 594                 /*
 595                  * This will also lock the page
 596                  */
 597                 ret = add_to_page_cache_lru(page, mapping, index,
 598                                             GFP_KERNEL);
 599                 if (unlikely(ret))
 600                         goto out;
 601         }
 602
 603         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 604         if (unlikely(ret)) {
 605                 loff_t isize = i_size_read(mapping->host);
 606
 607                 if (ret != AOP_TRUNCATED_PAGE)
 608                         unlock_page(page);
 609                 page_cache_release(page);
 610                 if (ret == AOP_TRUNCATED_PAGE)
 611                         goto find_page;
 612
 613                 /*
 614                  * prepare_write() may have instantiated a few blocks
 615                  * outside i_size.  Trim these off again.
 616                  */
 617                 if (sd->pos + this_len > isize)
 618                         vmtruncate(mapping->host, isize);
 619
 620                 goto out_ret;
 621         }
 622
 623         if (buf->page != page) {
 624                 /*
 625                  * Careful, ->map() uses KM_USER0!
 626                  */
 627                 char *src = buf->ops->map(pipe, buf, 1);
 628                 char *dst = kmap_atomic(page, KM_USER1);
 629
 630                 memcpy(dst + offset, src + buf->offset, this_len);
 631                 flush_dcache_page(page);
 632                 kunmap_atomic(dst, KM_USER1);
 633                 buf->ops->unmap(pipe, buf, src);
 634         }
 635
 636         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 637         if (ret) {
 638                 if (ret == AOP_TRUNCATED_PAGE) {
 639                         page_cache_release(page);
 640                         goto find_page;
 641                 }
 642                 if (ret < 0)
 643                         goto out;
 644                 /*
 645                  * Partial write has happened, so 'ret' already initialized by
 646                  * number of bytes written, Where is nothing we have to do here.
 647                  */
 648         } else
 649                 ret = this_len;
 650         /*
 651          * Return the number of bytes written and mark page as
 652          * accessed, we are now done!
 653          */
 654         mark_page_accessed(page);
 655 out:
 656         page_cache_release(page);
 657         unlock_page(page);
 658 out_ret:
 659         return ret;
 660 }
 661
 662 /*
 663  * Pipe input worker. Most of this logic works like a regular pipe, the
 664  * key here is the 'actor' worker passed in that actually moves the data
 665  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 666  */
 667 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 668                            struct file *out, loff_t *ppos, size_t len,
 669                            unsigned int flags, splice_actor *actor)
 670 {
 671         int ret, do_wakeup, err;
 672         struct splice_desc sd;
 673
 674         ret = 0;
 675         do_wakeup = 0;
 676
 677         sd.total_len = len;
 678         sd.flags = flags;
 679         sd.file = out;
 680         sd.pos = *ppos;
 681
 682         for (;;) {
 683                 if (pipe->nrbufs) {
 684                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 685                         const struct pipe_buf_operations *ops = buf->ops;
 686
 687                         sd.len = buf->len;
 688                         if (sd.len > sd.total_len)
 689                                 sd.len = sd.total_len;
 690
 691                         err = actor(pipe, buf, &sd);
 692                         if (err <= 0) {
 693                                 if (!ret && err != -ENODATA)
 694                                         ret = err;
 695
 696                                 break;
 697                         }
 698
 699                         ret += err;
 700                         buf->offset += err;
 701                         buf->len -= err;
 702
 703                         sd.len -= err;
 704                         sd.pos += err;
 705                         sd.total_len -= err;
 706                         if (sd.len)
 707                                 continue;
 708
 709                         if (!buf->len) {
 710                                 buf->ops = NULL;
 711                                 ops->release(pipe, buf);
 712                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 713                                 pipe->nrbufs--;
 714                                 if (pipe->inode)
 715                                         do_wakeup = 1;
 716                         }
 717
 718                         if (!sd.total_len)
 719                                 break;
 720                 }
 721
 722                 if (pipe->nrbufs)
 723                         continue;
 724                 if (!pipe->writers)
 725                         break;
 726                 if (!pipe->waiting_writers) {
 727                         if (ret)
 728                                 break;
 729                 }
 730
 731                 if (flags & SPLICE_F_NONBLOCK) {
 732                         if (!ret)
 733                                 ret = -EAGAIN;
 734                         break;
 735                 }
 736
 737                 if (signal_pending(current)) {
 738                         if (!ret)
 739                                 ret = -ERESTARTSYS;
 740                         break;
 741                 }
 742
 743                 if (do_wakeup) {
 744                         smp_mb();
 745                         if (waitqueue_active(&pipe->wait))
 746                                 wake_up_interruptible_sync(&pipe->wait);
 747                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 748                         do_wakeup = 0;
 749                 }
 750
 751                 pipe_wait(pipe);
 752         }
 753
 754         if (do_wakeup) {
 755                 smp_mb();
 756                 if (waitqueue_active(&pipe->wait))
 757                         wake_up_interruptible(&pipe->wait);
 758                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 759         }
 760
 761         return ret;
 762 }
 763 EXPORT_SYMBOL(__splice_from_pipe);
 764
 765 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 766                          loff_t *ppos, size_t len, unsigned int flags,
 767                          splice_actor *actor)
 768 {
 769         ssize_t ret;
 770         struct inode *inode = out->f_mapping->host;
 771
 772         /*
 773          * The actor worker might be calling ->prepare_write and
 774          * ->commit_write. Most of the time, these expect i_mutex to
 775          * be held. Since this may result in an ABBA deadlock with
 776          * pipe->inode, we have to order lock acquiry here.
 777          */
 778         inode_double_lock(inode, pipe->inode);
 779         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 780         inode_double_unlock(inode, pipe->inode);
 781
 782         return ret;
 783 }
 784
 785 /**
 786  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 787  * @pipe:       pipe info
 788  * @out:        file to write to
 789  * @len:        number of bytes to splice
 790  * @flags:      splice modifier flags
 791  *
 792  * Will either move or copy pages (determined by @flags options) from
 793  * the given pipe inode to the given file. The caller is responsible
 794  * for acquiring i_mutex on both inodes.
 795  *
 796  */
 797 ssize_t
 798 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 799                                  loff_t *ppos, size_t len, unsigned int flags)
 800 {
 801         struct address_space *mapping = out->f_mapping;
 802         struct inode *inode = mapping->host;
 803         ssize_t ret;
 804         int err;
 805
 806         err = remove_suid(out->f_path.dentry);
 807         if (unlikely(err))
 808                 return err;
 809
 810         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 811         if (ret > 0) {
 812                 *ppos += ret;
 813
 814                 /*
 815                  * If file or inode is SYNC and we actually wrote some data,
 816                  * sync it.
 817                  */
 818                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 819                         err = generic_osync_inode(inode, mapping,
 820                                                   OSYNC_METADATA|OSYNC_DATA);
 821
 822                         if (err)
 823                                 ret = err;
 824                 }
 825                 balance_dirty_pages_ratelimited(mapping);
 826         }
 827
 828         return ret;
 829 }
 830
 831 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 832
 833 /**
 834  * generic_file_splice_write - splice data from a pipe to a file
 835  * @pipe:       pipe info
 836  * @out:        file to write to
 837  * @len:        number of bytes to splice
 838  * @flags:      splice modifier flags
 839  *
 840  * Will either move or copy pages (determined by @flags options) from
 841  * the given pipe inode to the given file.
 842  *
 843  */
 844 ssize_t
 845 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 846                           loff_t *ppos, size_t len, unsigned int flags)
 847 {
 848         struct address_space *mapping = out->f_mapping;
 849         struct inode *inode = mapping->host;
 850         ssize_t ret;
 851         int err;
 852
 853         err = should_remove_suid(out->f_path.dentry);
 854         if (unlikely(err)) {
 855                 mutex_lock(&inode->i_mutex);
 856                 err = __remove_suid(out->f_path.dentry, err);
 857                 mutex_unlock(&inode->i_mutex);
 858                 if (err)
 859                         return err;
 860         }
 861
 862         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 863         if (ret > 0) {
 864                 *ppos += ret;
 865
 866                 /*
 867                  * If file or inode is SYNC and we actually wrote some data,
 868                  * sync it.
 869                  */
 870                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 871                         mutex_lock(&inode->i_mutex);
 872                         err = generic_osync_inode(inode, mapping,
 873                                                   OSYNC_METADATA|OSYNC_DATA);
 874                         mutex_unlock(&inode->i_mutex);
 875
 876                         if (err)
 877                                 ret = err;
 878                 }
 879                 balance_dirty_pages_ratelimited(mapping);
 880         }
 881
 882         return ret;
 883 }
 884
 885 EXPORT_SYMBOL(generic_file_splice_write);
 886
 887 /**
 888  * generic_splice_sendpage - splice data from a pipe to a socket
 889  * @inode:      pipe inode
 890  * @out:        socket to write to
 891  * @len:        number of bytes to splice
 892  * @flags:      splice modifier flags
 893  *
 894  * Will send @len bytes from the pipe to a network socket. No data copying
 895  * is involved.
 896  *
 897  */
 898 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 899                                 loff_t *ppos, size_t len, unsigned int flags)
 900 {
 901         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 902 }
 903
 904 EXPORT_SYMBOL(generic_splice_sendpage);
 905
 906 /*
 907  * Attempt to initiate a splice from pipe to file.
 908  */
 909 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 910                            loff_t *ppos, size_t len, unsigned int flags)
 911 {
 912         int ret;
 913
 914         if (unlikely(!out->f_op || !out->f_op->splice_write))
 915                 return -EINVAL;
 916
 917         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 918                 return -EBADF;
 919
 920         ret = rw_verify_area(WRITE, out, ppos, len);
 921         if (unlikely(ret < 0))
 922                 return ret;
 923
 924         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 925 }
 926
 927 /*
 928  * Attempt to initiate a splice from a file to a pipe.
 929  */
 930 static long do_splice_to(struct file *in, loff_t *ppos,
 931                          struct pipe_inode_info *pipe, size_t len,
 932                          unsigned int flags)
 933 {
 934         int ret;
 935
 936         if (unlikely(!in->f_op || !in->f_op->splice_read))
 937                 return -EINVAL;
 938
 939         if (unlikely(!(in->f_mode & FMODE_READ)))
 940                 return -EBADF;
 941
 942         ret = rw_verify_area(READ, in, ppos, len);
 943         if (unlikely(ret < 0))
 944                 return ret;
 945
 946         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 947 }
 948
 949 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 950                       size_t len, unsigned int flags)
 951 {
 952         struct pipe_inode_info *pipe;
 953         long ret, bytes;
 954         loff_t out_off;
 955         umode_t i_mode;
 956         int i;
 957
 958         /*
 959          * We require the input being a regular file, as we don't want to
 960          * randomly drop data for eg socket -> socket splicing. Use the
 961          * piped splicing for that!
 962          */
 963         i_mode = in->f_path.dentry->d_inode->i_mode;
 964         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 965                 return -EINVAL;
 966
 967         /*
 968          * neither in nor out is a pipe, setup an internal pipe attached to
 969          * 'out' and transfer the wanted data from 'in' to 'out' through that
 970          */
 971         pipe = current->splice_pipe;
 972         if (unlikely(!pipe)) {
 973                 pipe = alloc_pipe_info(NULL);
 974                 if (!pipe)
 975                         return -ENOMEM;
 976
 977                 /*
 978                  * We don't have an immediate reader, but we'll read the stuff
 979                  * out of the pipe right after the splice_to_pipe(). So set
 980                  * PIPE_READERS appropriately.
 981                  */
 982                 pipe->readers = 1;
 983
 984                 current->splice_pipe = pipe;
 985         }
 986
 987         /*
 988          * Do the splice.
 989          */
 990         ret = 0;
 991         bytes = 0;
 992         out_off = 0;
 993
 994         while (len) {
 995                 size_t read_len, max_read_len;
 996
 997                 /*
 998                  * Do at most PIPE_BUFFERS pages worth of transfer:
 999                  */
1000                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1001
1002                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1003                 if (unlikely(ret < 0))
1004                         goto out_release;
1005
1006                 read_len = ret;
1007
1008                 /*
1009                  * NOTE: nonblocking mode only applies to the input. We
1010                  * must not do the output in nonblocking mode as then we
1011                  * could get stuck data in the internal pipe:
1012                  */
1013                 ret = do_splice_from(pipe, out, &out_off, read_len,
1014                                      flags & ~SPLICE_F_NONBLOCK);
1015                 if (unlikely(ret < 0))
1016                         goto out_release;
1017
1018                 bytes += ret;
1019                 len -= ret;
1020
1021                 /*
1022                  * In nonblocking mode, if we got back a short read then
1023                  * that was due to either an IO error or due to the
1024                  * pagecache entry not being there. In the IO error case
1025                  * the _next_ splice attempt will produce a clean IO error
1026                  * return value (not a short read), so in both cases it's
1027                  * correct to break out of the loop here:
1028                  */
1029                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1030                         break;
1031         }
1032
1033         pipe->nrbufs = pipe->curbuf = 0;
1034
1035         return bytes;
1036
1037 out_release:
1038         /*
1039          * If we did an incomplete transfer we must release
1040          * the pipe buffers in question:
1041          */
1042         for (i = 0; i < PIPE_BUFFERS; i++) {
1043                 struct pipe_buffer *buf = pipe->bufs + i;
1044
1045                 if (buf->ops) {
1046                         buf->ops->release(pipe, buf);
1047                         buf->ops = NULL;
1048                 }
1049         }
1050         pipe->nrbufs = pipe->curbuf = 0;
1051
1052         /*
1053          * If we transferred some data, return the number of bytes:
1054          */
1055         if (bytes > 0)
1056                 return bytes;
1057
1058         return ret;
1059 }
1060
1061 /*
1062  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1063  * location, so checking ->i_pipe is not enough to verify that this is a
1064  * pipe.
1065  */
1066 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1067 {
1068         if (S_ISFIFO(inode->i_mode))
1069                 return inode->i_pipe;
1070
1071         return NULL;
1072 }
1073
1074 /*
1075  * Determine where to splice to/from.
1076  */
1077 static long do_splice(struct file *in, loff_t __user *off_in,
1078                       struct file *out, loff_t __user *off_out,
1079                       size_t len, unsigned int flags)
1080 {
1081         struct pipe_inode_info *pipe;
1082         loff_t offset, *off;
1083         long ret;
1084
1085         pipe = pipe_info(in->f_path.dentry->d_inode);
1086         if (pipe) {
1087                 if (off_in)
1088                         return -ESPIPE;
1089                 if (off_out) {
1090                         if (out->f_op->llseek == no_llseek)
1091                                 return -EINVAL;
1092                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1093                                 return -EFAULT;
1094                         off = &offset;
1095                 } else
1096                         off = &out->f_pos;
1097
1098                 ret = do_splice_from(pipe, out, off, len, flags);
1099
1100                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1101                         ret = -EFAULT;
1102
1103                 return ret;
1104         }
1105
1106         pipe = pipe_info(out->f_path.dentry->d_inode);
1107         if (pipe) {
1108                 if (off_out)
1109                         return -ESPIPE;
1110                 if (off_in) {
1111                         if (in->f_op->llseek == no_llseek)
1112                                 return -EINVAL;
1113                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1114                                 return -EFAULT;
1115                         off = &offset;
1116                 } else
1117                         off = &in->f_pos;
1118
1119                 ret = do_splice_to(in, off, pipe, len, flags);
1120
1121                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1122                         ret = -EFAULT;
1123
1124                 return ret;
1125         }
1126
1127         return -EINVAL;
1128 }
1129
1130 /*
1131  * Map an iov into an array of pages and offset/length tupples. With the
1132  * partial_page structure, we can map several non-contiguous ranges into
1133  * our ones pages[] map instead of splitting that operation into pieces.
1134  * Could easily be exported as a generic helper for other users, in which
1135  * case one would probably want to add a 'max_nr_pages' parameter as well.
1136  */
1137 static int get_iovec_page_array(const struct iovec __user *iov,
1138                                 unsigned int nr_vecs, struct page **pages,
1139                                 struct partial_page *partial, int aligned)
1140 {
1141         int buffers = 0, error = 0;
1142
1143         /*
1144          * It's ok to take the mmap_sem for reading, even
1145          * across a "get_user()".
1146          */
1147         down_read(&current->mm->mmap_sem);
1148
1149         while (nr_vecs) {
1150                 unsigned long off, npages;
1151                 void __user *base;
1152                 size_t len;
1153                 int i;
1154
1155                 /*
1156                  * Get user address base and length for this iovec.
1157                  */
1158                 error = get_user(base, &iov->iov_base);
1159                 if (unlikely(error))
1160                         break;
1161                 error = get_user(len, &iov->iov_len);
1162                 if (unlikely(error))
1163                         break;
1164
1165                 /*
1166                  * Sanity check this iovec. 0 read succeeds.
1167                  */
1168                 if (unlikely(!len))
1169                         break;
1170                 error = -EFAULT;
1171                 if (unlikely(!base))
1172                         break;
1173
1174                 /*
1175                  * Get this base offset and number of pages, then map
1176                  * in the user pages.
1177                  */
1178                 off = (unsigned long) base & ~PAGE_MASK;
1179
1180                 /*
1181                  * If asked for alignment, the offset must be zero and the
1182                  * length a multiple of the PAGE_SIZE.
1183                  */
1184                 error = -EINVAL;
1185                 if (aligned && (off || len & ~PAGE_MASK))
1186                         break;
1187
1188                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1189                 if (npages > PIPE_BUFFERS - buffers)
1190                         npages = PIPE_BUFFERS - buffers;
1191
1192                 error = get_user_pages(current, current->mm,
1193                                        (unsigned long) base, npages, 0, 0,
1194                                        &pages[buffers], NULL);
1195
1196                 if (unlikely(error <= 0))
1197                         break;
1198
1199                 /*
1200                  * Fill this contiguous range into the partial page map.
1201                  */
1202                 for (i = 0; i < error; i++) {
1203                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1204
1205                         partial[buffers].offset = off;
1206                         partial[buffers].len = plen;
1207
1208                         off = 0;
1209                         len -= plen;
1210                         buffers++;
1211                 }
1212
1213                 /*
1214                  * We didn't complete this iov, stop here since it probably
1215                  * means we have to move some of this into a pipe to
1216                  * be able to continue.
1217                  */
1218                 if (len)
1219                         break;
1220
1221                 /*
1222                  * Don't continue if we mapped fewer pages than we asked for,
1223                  * or if we mapped the max number of pages that we have
1224                  * room for.
1225                  */
1226                 if (error < npages || buffers == PIPE_BUFFERS)
1227                         break;
1228
1229                 nr_vecs--;
1230                 iov++;
1231         }
1232
1233         up_read(&current->mm->mmap_sem);
1234
1235         if (buffers)
1236                 return buffers;
1237
1238         return error;
1239 }
1240
1241 /*
1242  * vmsplice splices a user address range into a pipe. It can be thought of
1243  * as splice-from-memory, where the regular splice is splice-from-file (or
1244  * to file). In both cases the output is a pipe, naturally.
1245  *
1246  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1247  * not the other way around. Splicing from user memory is a simple operation
1248  * that can be supported without any funky alignment restrictions or nasty
1249  * vm tricks. We simply map in the user memory and fill them into a pipe.
1250  * The reverse isn't quite as easy, though. There are two possible solutions
1251  * for that:
1252  *
1253  *      - memcpy() the data internally, at which point we might as well just
1254  *        do a regular read() on the buffer anyway.
1255  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1256  *        has restriction limitations on both ends of the pipe).
1257  *
1258  * Alas, it isn't here.
1259  *
1260  */
1261 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1262                         unsigned long nr_segs, unsigned int flags)
1263 {
1264         struct pipe_inode_info *pipe;
1265         struct page *pages[PIPE_BUFFERS];
1266         struct partial_page partial[PIPE_BUFFERS];
1267         struct splice_pipe_desc spd = {
1268                 .pages = pages,
1269                 .partial = partial,
1270                 .flags = flags,
1271                 .ops = &user_page_pipe_buf_ops,
1272         };
1273
1274         pipe = pipe_info(file->f_path.dentry->d_inode);
1275         if (!pipe)
1276                 return -EBADF;
1277         if (unlikely(nr_segs > UIO_MAXIOV))
1278                 return -EINVAL;
1279         else if (unlikely(!nr_segs))
1280                 return 0;
1281
1282         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1283                                             flags & SPLICE_F_GIFT);
1284         if (spd.nr_pages <= 0)
1285                 return spd.nr_pages;
1286
1287         return splice_to_pipe(pipe, &spd);
1288 }
1289
1290 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1291                              unsigned long nr_segs, unsigned int flags)
1292 {
1293         struct file *file;
1294         long error;
1295         int fput;
1296
1297         error = -EBADF;
1298         file = fget_light(fd, &fput);
1299         if (file) {
1300                 if (file->f_mode & FMODE_WRITE)
1301                         error = do_vmsplice(file, iov, nr_segs, flags);
1302
1303                 fput_light(file, fput);
1304         }
1305
1306         return error;
1307 }
1308
1309 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1310                            int fd_out, loff_t __user *off_out,
1311                            size_t len, unsigned int flags)
1312 {
1313         long error;
1314         struct file *in, *out;
1315         int fput_in, fput_out;
1316
1317         if (unlikely(!len))
1318                 return 0;
1319
1320         error = -EBADF;
1321         in = fget_light(fd_in, &fput_in);
1322         if (in) {
1323                 if (in->f_mode & FMODE_READ) {
1324                         out = fget_light(fd_out, &fput_out);
1325                         if (out) {
1326                                 if (out->f_mode & FMODE_WRITE)
1327                                         error = do_splice(in, off_in,
1328                                                           out, off_out,
1329                                                           len, flags);
1330                                 fput_light(out, fput_out);
1331                         }
1332                 }
1333
1334                 fput_light(in, fput_in);
1335         }
1336
1337         return error;
1338 }
1339
1340 /*
1341  * Make sure there's data to read. Wait for input if we can, otherwise
1342  * return an appropriate error.
1343  */
1344 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1345 {
1346         int ret;
1347
1348         /*
1349          * Check ->nrbufs without the inode lock first. This function
1350          * is speculative anyways, so missing one is ok.
1351          */
1352         if (pipe->nrbufs)
1353                 return 0;
1354
1355         ret = 0;
1356         mutex_lock(&pipe->inode->i_mutex);
1357
1358         while (!pipe->nrbufs) {
1359                 if (signal_pending(current)) {
1360                         ret = -ERESTARTSYS;
1361                         break;
1362                 }
1363                 if (!pipe->writers)
1364                         break;
1365                 if (!pipe->waiting_writers) {
1366                         if (flags & SPLICE_F_NONBLOCK) {
1367                                 ret = -EAGAIN;
1368                                 break;
1369                         }
1370                 }
1371                 pipe_wait(pipe);
1372         }
1373
1374         mutex_unlock(&pipe->inode->i_mutex);
1375         return ret;
1376 }
1377
1378 /*
1379  * Make sure there's writeable room. Wait for room if we can, otherwise
1380  * return an appropriate error.
1381  */
1382 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1383 {
1384         int ret;
1385
1386         /*
1387          * Check ->nrbufs without the inode lock first. This function
1388          * is speculative anyways, so missing one is ok.
1389          */
1390         if (pipe->nrbufs < PIPE_BUFFERS)
1391                 return 0;
1392
1393         ret = 0;
1394         mutex_lock(&pipe->inode->i_mutex);
1395
1396         while (pipe->nrbufs >= PIPE_BUFFERS) {
1397                 if (!pipe->readers) {
1398                         send_sig(SIGPIPE, current, 0);
1399                         ret = -EPIPE;
1400                         break;
1401                 }
1402                 if (flags & SPLICE_F_NONBLOCK) {
1403                         ret = -EAGAIN;
1404                         break;
1405                 }
1406                 if (signal_pending(current)) {
1407                         ret = -ERESTARTSYS;
1408                         break;
1409                 }
1410                 pipe->waiting_writers++;
1411                 pipe_wait(pipe);
1412                 pipe->waiting_writers--;
1413         }
1414
1415         mutex_unlock(&pipe->inode->i_mutex);
1416         return ret;
1417 }
1418
1419 /*
1420  * Link contents of ipipe to opipe.
1421  */
1422 static int link_pipe(struct pipe_inode_info *ipipe,
1423                      struct pipe_inode_info *opipe,
1424                      size_t len, unsigned int flags)
1425 {
1426         struct pipe_buffer *ibuf, *obuf;
1427         int ret = 0, i = 0, nbuf;
1428
1429         /*
1430          * Potential ABBA deadlock, work around it by ordering lock
1431          * grabbing by inode address. Otherwise two different processes
1432          * could deadlock (one doing tee from A -> B, the other from B -> A).
1433          */
1434         inode_double_lock(ipipe->inode, opipe->inode);
1435
1436         do {
1437                 if (!opipe->readers) {
1438                         send_sig(SIGPIPE, current, 0);
1439                         if (!ret)
1440                                 ret = -EPIPE;
1441                         break;
1442                 }
1443
1444                 /*
1445                  * If we have iterated all input buffers or ran out of
1446                  * output room, break.
1447                  */
1448                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1449                         break;
1450
1451                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1452                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1453
1454                 /*
1455                  * Get a reference to this pipe buffer,
1456                  * so we can copy the contents over.
1457                  */
1458                 ibuf->ops->get(ipipe, ibuf);
1459
1460                 obuf = opipe->bufs + nbuf;
1461                 *obuf = *ibuf;
1462
1463                 /*
1464                  * Don't inherit the gift flag, we need to
1465                  * prevent multiple steals of this page.
1466                  */
1467                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1468
1469                 if (obuf->len > len)
1470                         obuf->len = len;
1471
1472                 opipe->nrbufs++;
1473                 ret += obuf->len;
1474                 len -= obuf->len;
1475                 i++;
1476         } while (len);
1477
1478         inode_double_unlock(ipipe->inode, opipe->inode);
1479
1480         /*
1481          * If we put data in the output pipe, wakeup any potential readers.
1482          */
1483         if (ret > 0) {
1484                 smp_mb();
1485                 if (waitqueue_active(&opipe->wait))
1486                         wake_up_interruptible(&opipe->wait);
1487                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1488         }
1489
1490         return ret;
1491 }
1492
1493 /*
1494  * This is a tee(1) implementation that works on pipes. It doesn't copy
1495  * any data, it simply references the 'in' pages on the 'out' pipe.
1496  * The 'flags' used are the SPLICE_F_* variants, currently the only
1497  * applicable one is SPLICE_F_NONBLOCK.
1498  */
1499 static long do_tee(struct file *in, struct file *out, size_t len,
1500                    unsigned int flags)
1501 {
1502         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1503         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1504         int ret = -EINVAL;
1505
1506         /*
1507          * Duplicate the contents of ipipe to opipe without actually
1508          * copying the data.
1509          */
1510         if (ipipe && opipe && ipipe != opipe) {
1511                 /*
1512                  * Keep going, unless we encounter an error. The ipipe/opipe
1513                  * ordering doesn't really matter.
1514                  */
1515                 ret = link_ipipe_prep(ipipe, flags);
1516                 if (!ret) {
1517                         ret = link_opipe_prep(opipe, flags);
1518                         if (!ret) {
1519                                 ret = link_pipe(ipipe, opipe, len, flags);
1520                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1521                                         ret = -EAGAIN;
1522                         }
1523                 }
1524         }
1525
1526         return ret;
1527 }
1528
1529 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1530 {
1531         struct file *in;
1532         int error, fput_in;
1533
1534         if (unlikely(!len))
1535                 return 0;
1536
1537         error = -EBADF;
1538         in = fget_light(fdin, &fput_in);
1539         if (in) {
1540                 if (in->f_mode & FMODE_READ) {
1541                         int fput_out;
1542                         struct file *out = fget_light(fdout, &fput_out);
1543
1544                         if (out) {
1545                                 if (out->f_mode & FMODE_WRITE)
1546                                         error = do_tee(in, out, len, flags);
1547                                 fput_light(out, fput_out);
1548                         }
1549                 }
1550                 fput_light(in, fput_in);
1551         }
1552
1553         return error;
1554 }