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         int error, page_nr;
 276         struct splice_pipe_desc spd = {
 277                 .pages = pages,
 278                 .partial = partial,
 279                 .flags = flags,
 280                 .ops = &page_cache_pipe_buf_ops,
 281         };
 282
 283         index = *ppos >> PAGE_CACHE_SHIFT;
 284         loff = *ppos & ~PAGE_CACHE_MASK;
 285         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 286
 287         if (nr_pages > PIPE_BUFFERS)
 288                 nr_pages = PIPE_BUFFERS;
 289
 290         /*
 291          * Don't try to 2nd guess the read-ahead logic, call into
 292          * page_cache_readahead() like the page cache reads would do.
 293          */
 294         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 295
 296         /*
 297          * Now fill in the holes:
 298          */
 299         error = 0;
 300
 301         /*
 302          * Lookup the (hopefully) full range of pages we need.
 303          */
 304         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 305
 306         /*
 307          * If find_get_pages_contig() returned fewer pages than we needed,
 308          * allocate the rest.
 309          */
 310         index += spd.nr_pages;
 311         while (spd.nr_pages < nr_pages) {
 312                 /*
 313                  * Page could be there, find_get_pages_contig() breaks on
 314                  * the first hole.
 315                  */
 316                 page = find_get_page(mapping, index);
 317                 if (!page) {
 318                         /*
 319                          * Make sure the read-ahead engine is notified
 320                          * about this failure.
 321                          */
 322                         handle_ra_miss(mapping, &in->f_ra, index);
 323
 324                         /*
 325                          * page didn't exist, allocate one.
 326                          */
 327                         page = page_cache_alloc_cold(mapping);
 328                         if (!page)
 329                                 break;
 330
 331                         error = add_to_page_cache_lru(page, mapping, index,
 332                                               GFP_KERNEL);
 333                         if (unlikely(error)) {
 334                                 page_cache_release(page);
 335                                 if (error == -EEXIST)
 336                                         continue;
 337                                 break;
 338                         }
 339                         /*
 340                          * add_to_page_cache() locks the page, unlock it
 341                          * to avoid convoluting the logic below even more.
 342                          */
 343                         unlock_page(page);
 344                 }
 345
 346                 pages[spd.nr_pages++] = page;
 347                 index++;
 348         }
 349
 350         /*
 351          * Now loop over the map and see if we need to start IO on any
 352          * pages, fill in the partial map, etc.
 353          */
 354         index = *ppos >> PAGE_CACHE_SHIFT;
 355         nr_pages = spd.nr_pages;
 356         spd.nr_pages = 0;
 357         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 358                 unsigned int this_len;
 359
 360                 if (!len)
 361                         break;
 362
 363                 /*
 364                  * this_len is the max we'll use from this page
 365                  */
 366                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 367                 page = pages[page_nr];
 368
 369                 /*
 370                  * If the page isn't uptodate, we may need to start io on it
 371                  */
 372                 if (!PageUptodate(page)) {
 373                         /*
 374                          * If in nonblock mode then dont block on waiting
 375                          * for an in-flight io page
 376                          */
 377                         if (flags & SPLICE_F_NONBLOCK) {
 378                                 if (TestSetPageLocked(page))
 379                                         break;
 380                         } else
 381                                 lock_page(page);
 382
 383                         /*
 384                          * page was truncated, stop here. if this isn't the
 385                          * first page, we'll just complete what we already
 386                          * added
 387                          */
 388                         if (!page->mapping) {
 389                                 unlock_page(page);
 390                                 break;
 391                         }
 392                         /*
 393                          * page was already under io and is now done, great
 394                          */
 395                         if (PageUptodate(page)) {
 396                                 unlock_page(page);
 397                                 goto fill_it;
 398                         }
 399
 400                         /*
 401                          * need to read in the page
 402                          */
 403                         error = mapping->a_ops->readpage(in, page);
 404                         if (unlikely(error)) {
 405                                 /*
 406                                  * We really should re-lookup the page here,
 407                                  * but it complicates things a lot. Instead
 408                                  * lets just do what we already stored, and
 409                                  * we'll get it the next time we are called.
 410                                  */
 411                                 if (error == AOP_TRUNCATED_PAGE)
 412                                         error = 0;
 413
 414                                 break;
 415                         }
 416                 }
 417 fill_it:
 418                 /*
 419                  * i_size must be checked after PageUptodate.
 420                  */
 421                 isize = i_size_read(mapping->host);
 422                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 423                 if (unlikely(!isize || index > end_index))
 424                         break;
 425
 426                 /*
 427                  * if this is the last page, see if we need to shrink
 428                  * the length and stop
 429                  */
 430                 if (end_index == index) {
 431                         unsigned int plen;
 432
 433                         /*
 434                          * max good bytes in this page
 435                          */
 436                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 437                         if (plen <= loff)
 438                                 break;
 439
 440                         /*
 441                          * force quit after adding this page
 442                          */
 443                         this_len = min(this_len, plen - loff);
 444                         len = this_len;
 445                 }
 446
 447                 partial[page_nr].offset = loff;
 448                 partial[page_nr].len = this_len;
 449                 len -= this_len;
 450                 loff = 0;
 451                 spd.nr_pages++;
 452                 index++;
 453         }
 454
 455         /*
 456          * Release any pages at the end, if we quit early. 'page_nr' is how far
 457          * we got, 'nr_pages' is how many pages are in the map.
 458          */
 459         while (page_nr < nr_pages)
 460                 page_cache_release(pages[page_nr++]);
 461
 462         if (spd.nr_pages)
 463                 return splice_to_pipe(pipe, &spd);
 464
 465         return error;
 466 }
 467
 468 /**
 469  * generic_file_splice_read - splice data from file to a pipe
 470  * @in:         file to splice from
 471  * @pipe:       pipe to splice to
 472  * @len:        number of bytes to splice
 473  * @flags:      splice modifier flags
 474  *
 475  * Will read pages from given file and fill them into a pipe.
 476  */
 477 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 478                                  struct pipe_inode_info *pipe, size_t len,
 479                                  unsigned int flags)
 480 {
 481         ssize_t spliced;
 482         int ret;
 483         loff_t isize, left;
 484
 485         isize = i_size_read(in->f_mapping->host);
 486         if (unlikely(*ppos >= isize))
 487                 return 0;
 488
 489         left = isize - *ppos;
 490         if (unlikely(left < len))
 491                 len = left;
 492
 493         ret = 0;
 494         spliced = 0;
 495         while (len) {
 496                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 497
 498                 if (ret < 0)
 499                         break;
 500                 else if (!ret) {
 501                         if (spliced)
 502                                 break;
 503                         if (flags & SPLICE_F_NONBLOCK) {
 504                                 ret = -EAGAIN;
 505                                 break;
 506                         }
 507                 }
 508
 509                 *ppos += ret;
 510                 len -= ret;
 511                 spliced += ret;
 512         }
 513
 514         if (spliced)
 515                 return spliced;
 516
 517         return ret;
 518 }
 519
 520 EXPORT_SYMBOL(generic_file_splice_read);
 521
 522 /*
 523  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 524  * using sendpage(). Return the number of bytes sent.
 525  */
 526 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 527                             struct pipe_buffer *buf, struct splice_desc *sd)
 528 {
 529         struct file *file = sd->file;
 530         loff_t pos = sd->pos;
 531         int ret, more;
 532
 533         ret = buf->ops->pin(pipe, buf);
 534         if (!ret) {
 535                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 536
 537                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 538                                            sd->len, &pos, more);
 539         }
 540
 541         return ret;
 542 }
 543
 544 /*
 545  * This is a little more tricky than the file -> pipe splicing. There are
 546  * basically three cases:
 547  *
 548  *      - Destination page already exists in the address space and there
 549  *        are users of it. For that case we have no other option that
 550  *        copying the data. Tough luck.
 551  *      - Destination page already exists in the address space, but there
 552  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 553  *        through to last case.
 554  *      - Destination page does not exist, we can add the pipe page to
 555  *        the page cache and avoid the copy.
 556  *
 557  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 558  * sd->flags), we attempt to migrate pages from the pipe to the output
 559  * file address space page cache. This is possible if no one else has
 560  * the pipe page referenced outside of the pipe and page cache. If
 561  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 562  * a new page in the output file page cache and fill/dirty that.
 563  */
 564 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 565                         struct splice_desc *sd)
 566 {
 567         struct file *file = sd->file;
 568         struct address_space *mapping = file->f_mapping;
 569         unsigned int offset, this_len;
 570         struct page *page;
 571         pgoff_t index;
 572         int ret;
 573
 574         /*
 575          * make sure the data in this buffer is uptodate
 576          */
 577         ret = buf->ops->pin(pipe, buf);
 578         if (unlikely(ret))
 579                 return ret;
 580
 581         index = sd->pos >> PAGE_CACHE_SHIFT;
 582         offset = sd->pos & ~PAGE_CACHE_MASK;
 583
 584         this_len = sd->len;
 585         if (this_len + offset > PAGE_CACHE_SIZE)
 586                 this_len = PAGE_CACHE_SIZE - offset;
 587
 588 find_page:
 589         page = find_lock_page(mapping, index);
 590         if (!page) {
 591                 ret = -ENOMEM;
 592                 page = page_cache_alloc_cold(mapping);
 593                 if (unlikely(!page))
 594                         goto out_ret;
 595
 596                 /*
 597                  * This will also lock the page
 598                  */
 599                 ret = add_to_page_cache_lru(page, mapping, index,
 600                                             GFP_KERNEL);
 601                 if (unlikely(ret))
 602                         goto out;
 603         }
 604
 605         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 606         if (unlikely(ret)) {
 607                 loff_t isize = i_size_read(mapping->host);
 608
 609                 if (ret != AOP_TRUNCATED_PAGE)
 610                         unlock_page(page);
 611                 page_cache_release(page);
 612                 if (ret == AOP_TRUNCATED_PAGE)
 613                         goto find_page;
 614
 615                 /*
 616                  * prepare_write() may have instantiated a few blocks
 617                  * outside i_size.  Trim these off again.
 618                  */
 619                 if (sd->pos + this_len > isize)
 620                         vmtruncate(mapping->host, isize);
 621
 622                 goto out_ret;
 623         }
 624
 625         if (buf->page != page) {
 626                 /*
 627                  * Careful, ->map() uses KM_USER0!
 628                  */
 629                 char *src = buf->ops->map(pipe, buf, 1);
 630                 char *dst = kmap_atomic(page, KM_USER1);
 631
 632                 memcpy(dst + offset, src + buf->offset, this_len);
 633                 flush_dcache_page(page);
 634                 kunmap_atomic(dst, KM_USER1);
 635                 buf->ops->unmap(pipe, buf, src);
 636         }
 637
 638         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 639         if (ret) {
 640                 if (ret == AOP_TRUNCATED_PAGE) {
 641                         page_cache_release(page);
 642                         goto find_page;
 643                 }
 644                 if (ret < 0)
 645                         goto out;
 646                 /*
 647                  * Partial write has happened, so 'ret' already initialized by
 648                  * number of bytes written, Where is nothing we have to do here.
 649                  */
 650         } else
 651                 ret = this_len;
 652         /*
 653          * Return the number of bytes written and mark page as
 654          * accessed, we are now done!
 655          */
 656         mark_page_accessed(page);
 657 out:
 658         page_cache_release(page);
 659         unlock_page(page);
 660 out_ret:
 661         return ret;
 662 }
 663
 664 /*
 665  * Pipe input worker. Most of this logic works like a regular pipe, the
 666  * key here is the 'actor' worker passed in that actually moves the data
 667  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 668  */
 669 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
 670                            struct file *out, loff_t *ppos, size_t len,
 671                            unsigned int flags, splice_actor *actor)
 672 {
 673         int ret, do_wakeup, err;
 674         struct splice_desc sd;
 675
 676         ret = 0;
 677         do_wakeup = 0;
 678
 679         sd.total_len = len;
 680         sd.flags = flags;
 681         sd.file = out;
 682         sd.pos = *ppos;
 683
 684         for (;;) {
 685                 if (pipe->nrbufs) {
 686                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 687                         const struct pipe_buf_operations *ops = buf->ops;
 688
 689                         sd.len = buf->len;
 690                         if (sd.len > sd.total_len)
 691                                 sd.len = sd.total_len;
 692
 693                         err = actor(pipe, buf, &sd);
 694                         if (err <= 0) {
 695                                 if (!ret && err != -ENODATA)
 696                                         ret = err;
 697
 698                                 break;
 699                         }
 700
 701                         ret += err;
 702                         buf->offset += err;
 703                         buf->len -= err;
 704
 705                         sd.len -= err;
 706                         sd.pos += err;
 707                         sd.total_len -= err;
 708                         if (sd.len)
 709                                 continue;
 710
 711                         if (!buf->len) {
 712                                 buf->ops = NULL;
 713                                 ops->release(pipe, buf);
 714                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 715                                 pipe->nrbufs--;
 716                                 if (pipe->inode)
 717                                         do_wakeup = 1;
 718                         }
 719
 720                         if (!sd.total_len)
 721                                 break;
 722                 }
 723
 724                 if (pipe->nrbufs)
 725                         continue;
 726                 if (!pipe->writers)
 727                         break;
 728                 if (!pipe->waiting_writers) {
 729                         if (ret)
 730                                 break;
 731                 }
 732
 733                 if (flags & SPLICE_F_NONBLOCK) {
 734                         if (!ret)
 735                                 ret = -EAGAIN;
 736                         break;
 737                 }
 738
 739                 if (signal_pending(current)) {
 740                         if (!ret)
 741                                 ret = -ERESTARTSYS;
 742                         break;
 743                 }
 744
 745                 if (do_wakeup) {
 746                         smp_mb();
 747                         if (waitqueue_active(&pipe->wait))
 748                                 wake_up_interruptible_sync(&pipe->wait);
 749                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 750                         do_wakeup = 0;
 751                 }
 752
 753                 pipe_wait(pipe);
 754         }
 755
 756         if (do_wakeup) {
 757                 smp_mb();
 758                 if (waitqueue_active(&pipe->wait))
 759                         wake_up_interruptible(&pipe->wait);
 760                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 761         }
 762
 763         return ret;
 764 }
 765 EXPORT_SYMBOL(__splice_from_pipe);
 766
 767 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 768                          loff_t *ppos, size_t len, unsigned int flags,
 769                          splice_actor *actor)
 770 {
 771         ssize_t ret;
 772         struct inode *inode = out->f_mapping->host;
 773
 774         /*
 775          * The actor worker might be calling ->prepare_write and
 776          * ->commit_write. Most of the time, these expect i_mutex to
 777          * be held. Since this may result in an ABBA deadlock with
 778          * pipe->inode, we have to order lock acquiry here.
 779          */
 780         inode_double_lock(inode, pipe->inode);
 781         ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
 782         inode_double_unlock(inode, pipe->inode);
 783
 784         return ret;
 785 }
 786
 787 /**
 788  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 789  * @pipe:       pipe info
 790  * @out:        file to write to
 791  * @len:        number of bytes to splice
 792  * @flags:      splice modifier flags
 793  *
 794  * Will either move or copy pages (determined by @flags options) from
 795  * the given pipe inode to the given file. The caller is responsible
 796  * for acquiring i_mutex on both inodes.
 797  *
 798  */
 799 ssize_t
 800 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 801                                  loff_t *ppos, size_t len, unsigned int flags)
 802 {
 803         struct address_space *mapping = out->f_mapping;
 804         struct inode *inode = mapping->host;
 805         ssize_t ret;
 806         int err;
 807
 808         err = remove_suid(out->f_path.dentry);
 809         if (unlikely(err))
 810                 return err;
 811
 812         ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 813         if (ret > 0) {
 814                 unsigned long nr_pages;
 815
 816                 *ppos += ret;
 817                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 818
 819                 /*
 820                  * If file or inode is SYNC and we actually wrote some data,
 821                  * sync it.
 822                  */
 823                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 824                         err = generic_osync_inode(inode, mapping,
 825                                                   OSYNC_METADATA|OSYNC_DATA);
 826
 827                         if (err)
 828                                 ret = err;
 829                 }
 830                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 831         }
 832
 833         return ret;
 834 }
 835
 836 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 837
 838 /**
 839  * generic_file_splice_write - splice data from a pipe to a file
 840  * @pipe:       pipe info
 841  * @out:        file to write to
 842  * @len:        number of bytes to splice
 843  * @flags:      splice modifier flags
 844  *
 845  * Will either move or copy pages (determined by @flags options) from
 846  * the given pipe inode to the given file.
 847  *
 848  */
 849 ssize_t
 850 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 851                           loff_t *ppos, size_t len, unsigned int flags)
 852 {
 853         struct address_space *mapping = out->f_mapping;
 854         struct inode *inode = mapping->host;
 855         ssize_t ret;
 856         int err;
 857
 858         err = should_remove_suid(out->f_path.dentry);
 859         if (unlikely(err)) {
 860                 mutex_lock(&inode->i_mutex);
 861                 err = __remove_suid(out->f_path.dentry, err);
 862                 mutex_unlock(&inode->i_mutex);
 863                 if (err)
 864                         return err;
 865         }
 866
 867         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 868         if (ret > 0) {
 869                 unsigned long nr_pages;
 870
 871                 *ppos += ret;
 872                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 873
 874                 /*
 875                  * If file or inode is SYNC and we actually wrote some data,
 876                  * sync it.
 877                  */
 878                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 879                         mutex_lock(&inode->i_mutex);
 880                         err = generic_osync_inode(inode, mapping,
 881                                                   OSYNC_METADATA|OSYNC_DATA);
 882                         mutex_unlock(&inode->i_mutex);
 883
 884                         if (err)
 885                                 ret = err;
 886                 }
 887                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 888         }
 889
 890         return ret;
 891 }
 892
 893 EXPORT_SYMBOL(generic_file_splice_write);
 894
 895 /**
 896  * generic_splice_sendpage - splice data from a pipe to a socket
 897  * @inode:      pipe inode
 898  * @out:        socket to write to
 899  * @len:        number of bytes to splice
 900  * @flags:      splice modifier flags
 901  *
 902  * Will send @len bytes from the pipe to a network socket. No data copying
 903  * is involved.
 904  *
 905  */
 906 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 907                                 loff_t *ppos, size_t len, unsigned int flags)
 908 {
 909         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 910 }
 911
 912 EXPORT_SYMBOL(generic_splice_sendpage);
 913
 914 /*
 915  * Attempt to initiate a splice from pipe to file.
 916  */
 917 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 918                            loff_t *ppos, size_t len, unsigned int flags)
 919 {
 920         int ret;
 921
 922         if (unlikely(!out->f_op || !out->f_op->splice_write))
 923                 return -EINVAL;
 924
 925         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 926                 return -EBADF;
 927
 928         ret = rw_verify_area(WRITE, out, ppos, len);
 929         if (unlikely(ret < 0))
 930                 return ret;
 931
 932         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 933 }
 934
 935 /*
 936  * Attempt to initiate a splice from a file to a pipe.
 937  */
 938 static long do_splice_to(struct file *in, loff_t *ppos,
 939                          struct pipe_inode_info *pipe, size_t len,
 940                          unsigned int flags)
 941 {
 942         int ret;
 943
 944         if (unlikely(!in->f_op || !in->f_op->splice_read))
 945                 return -EINVAL;
 946
 947         if (unlikely(!(in->f_mode & FMODE_READ)))
 948                 return -EBADF;
 949
 950         ret = rw_verify_area(READ, in, ppos, len);
 951         if (unlikely(ret < 0))
 952                 return ret;
 953
 954         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 955 }
 956
 957 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 958                       size_t len, unsigned int flags)
 959 {
 960         struct pipe_inode_info *pipe;
 961         long ret, bytes;
 962         loff_t out_off;
 963         umode_t i_mode;
 964         int i;
 965
 966         /*
 967          * We require the input being a regular file, as we don't want to
 968          * randomly drop data for eg socket -> socket splicing. Use the
 969          * piped splicing for that!
 970          */
 971         i_mode = in->f_path.dentry->d_inode->i_mode;
 972         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 973                 return -EINVAL;
 974
 975         /*
 976          * neither in nor out is a pipe, setup an internal pipe attached to
 977          * 'out' and transfer the wanted data from 'in' to 'out' through that
 978          */
 979         pipe = current->splice_pipe;
 980         if (unlikely(!pipe)) {
 981                 pipe = alloc_pipe_info(NULL);
 982                 if (!pipe)
 983                         return -ENOMEM;
 984
 985                 /*
 986                  * We don't have an immediate reader, but we'll read the stuff
 987                  * out of the pipe right after the splice_to_pipe(). So set
 988                  * PIPE_READERS appropriately.
 989                  */
 990                 pipe->readers = 1;
 991
 992                 current->splice_pipe = pipe;
 993         }
 994
 995         /*
 996          * Do the splice.
 997          */
 998         ret = 0;
 999         bytes = 0;
1000         out_off = 0;
1001
1002         while (len) {
1003                 size_t read_len, max_read_len;
1004
1005                 /*
1006                  * Do at most PIPE_BUFFERS pages worth of transfer:
1007                  */
1008                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1009
1010                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1011                 if (unlikely(ret < 0))
1012                         goto out_release;
1013
1014                 read_len = ret;
1015
1016                 /*
1017                  * NOTE: nonblocking mode only applies to the input. We
1018                  * must not do the output in nonblocking mode as then we
1019                  * could get stuck data in the internal pipe:
1020                  */
1021                 ret = do_splice_from(pipe, out, &out_off, read_len,
1022                                      flags & ~SPLICE_F_NONBLOCK);
1023                 if (unlikely(ret < 0))
1024                         goto out_release;
1025
1026                 bytes += ret;
1027                 len -= ret;
1028
1029                 /*
1030                  * In nonblocking mode, if we got back a short read then
1031                  * that was due to either an IO error or due to the
1032                  * pagecache entry not being there. In the IO error case
1033                  * the _next_ splice attempt will produce a clean IO error
1034                  * return value (not a short read), so in both cases it's
1035                  * correct to break out of the loop here:
1036                  */
1037                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1038                         break;
1039         }
1040
1041         pipe->nrbufs = pipe->curbuf = 0;
1042
1043         return bytes;
1044
1045 out_release:
1046         /*
1047          * If we did an incomplete transfer we must release
1048          * the pipe buffers in question:
1049          */
1050         for (i = 0; i < PIPE_BUFFERS; i++) {
1051                 struct pipe_buffer *buf = pipe->bufs + i;
1052
1053                 if (buf->ops) {
1054                         buf->ops->release(pipe, buf);
1055                         buf->ops = NULL;
1056                 }
1057         }
1058         pipe->nrbufs = pipe->curbuf = 0;
1059
1060         /*
1061          * If we transferred some data, return the number of bytes:
1062          */
1063         if (bytes > 0)
1064                 return bytes;
1065
1066         return ret;
1067 }
1068
1069 /*
1070  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1071  * location, so checking ->i_pipe is not enough to verify that this is a
1072  * pipe.
1073  */
1074 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1075 {
1076         if (S_ISFIFO(inode->i_mode))
1077                 return inode->i_pipe;
1078
1079         return NULL;
1080 }
1081
1082 /*
1083  * Determine where to splice to/from.
1084  */
1085 static long do_splice(struct file *in, loff_t __user *off_in,
1086                       struct file *out, loff_t __user *off_out,
1087                       size_t len, unsigned int flags)
1088 {
1089         struct pipe_inode_info *pipe;
1090         loff_t offset, *off;
1091         long ret;
1092
1093         pipe = pipe_info(in->f_path.dentry->d_inode);
1094         if (pipe) {
1095                 if (off_in)
1096                         return -ESPIPE;
1097                 if (off_out) {
1098                         if (out->f_op->llseek == no_llseek)
1099                                 return -EINVAL;
1100                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1101                                 return -EFAULT;
1102                         off = &offset;
1103                 } else
1104                         off = &out->f_pos;
1105
1106                 ret = do_splice_from(pipe, out, off, len, flags);
1107
1108                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1109                         ret = -EFAULT;
1110
1111                 return ret;
1112         }
1113
1114         pipe = pipe_info(out->f_path.dentry->d_inode);
1115         if (pipe) {
1116                 if (off_out)
1117                         return -ESPIPE;
1118                 if (off_in) {
1119                         if (in->f_op->llseek == no_llseek)
1120                                 return -EINVAL;
1121                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1122                                 return -EFAULT;
1123                         off = &offset;
1124                 } else
1125                         off = &in->f_pos;
1126
1127                 ret = do_splice_to(in, off, pipe, len, flags);
1128
1129                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1130                         ret = -EFAULT;
1131
1132                 return ret;
1133         }
1134
1135         return -EINVAL;
1136 }
1137
1138 /*
1139  * Map an iov into an array of pages and offset/length tupples. With the
1140  * partial_page structure, we can map several non-contiguous ranges into
1141  * our ones pages[] map instead of splitting that operation into pieces.
1142  * Could easily be exported as a generic helper for other users, in which
1143  * case one would probably want to add a 'max_nr_pages' parameter as well.
1144  */
1145 static int get_iovec_page_array(const struct iovec __user *iov,
1146                                 unsigned int nr_vecs, struct page **pages,
1147                                 struct partial_page *partial, int aligned)
1148 {
1149         int buffers = 0, error = 0;
1150
1151         /*
1152          * It's ok to take the mmap_sem for reading, even
1153          * across a "get_user()".
1154          */
1155         down_read(&current->mm->mmap_sem);
1156
1157         while (nr_vecs) {
1158                 unsigned long off, npages;
1159                 void __user *base;
1160                 size_t len;
1161                 int i;
1162
1163                 /*
1164                  * Get user address base and length for this iovec.
1165                  */
1166                 error = get_user(base, &iov->iov_base);
1167                 if (unlikely(error))
1168                         break;
1169                 error = get_user(len, &iov->iov_len);
1170                 if (unlikely(error))
1171                         break;
1172
1173                 /*
1174                  * Sanity check this iovec. 0 read succeeds.
1175                  */
1176                 if (unlikely(!len))
1177                         break;
1178                 error = -EFAULT;
1179                 if (unlikely(!base))
1180                         break;
1181
1182                 /*
1183                  * Get this base offset and number of pages, then map
1184                  * in the user pages.
1185                  */
1186                 off = (unsigned long) base & ~PAGE_MASK;
1187
1188                 /*
1189                  * If asked for alignment, the offset must be zero and the
1190                  * length a multiple of the PAGE_SIZE.
1191                  */
1192                 error = -EINVAL;
1193                 if (aligned && (off || len & ~PAGE_MASK))
1194                         break;
1195
1196                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1197                 if (npages > PIPE_BUFFERS - buffers)
1198                         npages = PIPE_BUFFERS - buffers;
1199
1200                 error = get_user_pages(current, current->mm,
1201                                        (unsigned long) base, npages, 0, 0,
1202                                        &pages[buffers], NULL);
1203
1204                 if (unlikely(error <= 0))
1205                         break;
1206
1207                 /*
1208                  * Fill this contiguous range into the partial page map.
1209                  */
1210                 for (i = 0; i < error; i++) {
1211                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1212
1213                         partial[buffers].offset = off;
1214                         partial[buffers].len = plen;
1215
1216                         off = 0;
1217                         len -= plen;
1218                         buffers++;
1219                 }
1220
1221                 /*
1222                  * We didn't complete this iov, stop here since it probably
1223                  * means we have to move some of this into a pipe to
1224                  * be able to continue.
1225                  */
1226                 if (len)
1227                         break;
1228
1229                 /*
1230                  * Don't continue if we mapped fewer pages than we asked for,
1231                  * or if we mapped the max number of pages that we have
1232                  * room for.
1233                  */
1234                 if (error < npages || buffers == PIPE_BUFFERS)
1235                         break;
1236
1237                 nr_vecs--;
1238                 iov++;
1239         }
1240
1241         up_read(&current->mm->mmap_sem);
1242
1243         if (buffers)
1244                 return buffers;
1245
1246         return error;
1247 }
1248
1249 /*
1250  * vmsplice splices a user address range into a pipe. It can be thought of
1251  * as splice-from-memory, where the regular splice is splice-from-file (or
1252  * to file). In both cases the output is a pipe, naturally.
1253  *
1254  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1255  * not the other way around. Splicing from user memory is a simple operation
1256  * that can be supported without any funky alignment restrictions or nasty
1257  * vm tricks. We simply map in the user memory and fill them into a pipe.
1258  * The reverse isn't quite as easy, though. There are two possible solutions
1259  * for that:
1260  *
1261  *      - memcpy() the data internally, at which point we might as well just
1262  *        do a regular read() on the buffer anyway.
1263  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1264  *        has restriction limitations on both ends of the pipe).
1265  *
1266  * Alas, it isn't here.
1267  *
1268  */
1269 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1270                         unsigned long nr_segs, unsigned int flags)
1271 {
1272         struct pipe_inode_info *pipe;
1273         struct page *pages[PIPE_BUFFERS];
1274         struct partial_page partial[PIPE_BUFFERS];
1275         struct splice_pipe_desc spd = {
1276                 .pages = pages,
1277                 .partial = partial,
1278                 .flags = flags,
1279                 .ops = &user_page_pipe_buf_ops,
1280         };
1281
1282         pipe = pipe_info(file->f_path.dentry->d_inode);
1283         if (!pipe)
1284                 return -EBADF;
1285         if (unlikely(nr_segs > UIO_MAXIOV))
1286                 return -EINVAL;
1287         else if (unlikely(!nr_segs))
1288                 return 0;
1289
1290         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1291                                             flags & SPLICE_F_GIFT);
1292         if (spd.nr_pages <= 0)
1293                 return spd.nr_pages;
1294
1295         return splice_to_pipe(pipe, &spd);
1296 }
1297
1298 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1299                              unsigned long nr_segs, unsigned int flags)
1300 {
1301         struct file *file;
1302         long error;
1303         int fput;
1304
1305         error = -EBADF;
1306         file = fget_light(fd, &fput);
1307         if (file) {
1308                 if (file->f_mode & FMODE_WRITE)
1309                         error = do_vmsplice(file, iov, nr_segs, flags);
1310
1311                 fput_light(file, fput);
1312         }
1313
1314         return error;
1315 }
1316
1317 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1318                            int fd_out, loff_t __user *off_out,
1319                            size_t len, unsigned int flags)
1320 {
1321         long error;
1322         struct file *in, *out;
1323         int fput_in, fput_out;
1324
1325         if (unlikely(!len))
1326                 return 0;
1327
1328         error = -EBADF;
1329         in = fget_light(fd_in, &fput_in);
1330         if (in) {
1331                 if (in->f_mode & FMODE_READ) {
1332                         out = fget_light(fd_out, &fput_out);
1333                         if (out) {
1334                                 if (out->f_mode & FMODE_WRITE)
1335                                         error = do_splice(in, off_in,
1336                                                           out, off_out,
1337                                                           len, flags);
1338                                 fput_light(out, fput_out);
1339                         }
1340                 }
1341
1342                 fput_light(in, fput_in);
1343         }
1344
1345         return error;
1346 }
1347
1348 /*
1349  * Make sure there's data to read. Wait for input if we can, otherwise
1350  * return an appropriate error.
1351  */
1352 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1353 {
1354         int ret;
1355
1356         /*
1357          * Check ->nrbufs without the inode lock first. This function
1358          * is speculative anyways, so missing one is ok.
1359          */
1360         if (pipe->nrbufs)
1361                 return 0;
1362
1363         ret = 0;
1364         mutex_lock(&pipe->inode->i_mutex);
1365
1366         while (!pipe->nrbufs) {
1367                 if (signal_pending(current)) {
1368                         ret = -ERESTARTSYS;
1369                         break;
1370                 }
1371                 if (!pipe->writers)
1372                         break;
1373                 if (!pipe->waiting_writers) {
1374                         if (flags & SPLICE_F_NONBLOCK) {
1375                                 ret = -EAGAIN;
1376                                 break;
1377                         }
1378                 }
1379                 pipe_wait(pipe);
1380         }
1381
1382         mutex_unlock(&pipe->inode->i_mutex);
1383         return ret;
1384 }
1385
1386 /*
1387  * Make sure there's writeable room. Wait for room if we can, otherwise
1388  * return an appropriate error.
1389  */
1390 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1391 {
1392         int ret;
1393
1394         /*
1395          * Check ->nrbufs without the inode lock first. This function
1396          * is speculative anyways, so missing one is ok.
1397          */
1398         if (pipe->nrbufs < PIPE_BUFFERS)
1399                 return 0;
1400
1401         ret = 0;
1402         mutex_lock(&pipe->inode->i_mutex);
1403
1404         while (pipe->nrbufs >= PIPE_BUFFERS) {
1405                 if (!pipe->readers) {
1406                         send_sig(SIGPIPE, current, 0);
1407                         ret = -EPIPE;
1408                         break;
1409                 }
1410                 if (flags & SPLICE_F_NONBLOCK) {
1411                         ret = -EAGAIN;
1412                         break;
1413                 }
1414                 if (signal_pending(current)) {
1415                         ret = -ERESTARTSYS;
1416                         break;
1417                 }
1418                 pipe->waiting_writers++;
1419                 pipe_wait(pipe);
1420                 pipe->waiting_writers--;
1421         }
1422
1423         mutex_unlock(&pipe->inode->i_mutex);
1424         return ret;
1425 }
1426
1427 /*
1428  * Link contents of ipipe to opipe.
1429  */
1430 static int link_pipe(struct pipe_inode_info *ipipe,
1431                      struct pipe_inode_info *opipe,
1432                      size_t len, unsigned int flags)
1433 {
1434         struct pipe_buffer *ibuf, *obuf;
1435         int ret = 0, i = 0, nbuf;
1436
1437         /*
1438          * Potential ABBA deadlock, work around it by ordering lock
1439          * grabbing by inode address. Otherwise two different processes
1440          * could deadlock (one doing tee from A -> B, the other from B -> A).
1441          */
1442         inode_double_lock(ipipe->inode, opipe->inode);
1443
1444         do {
1445                 if (!opipe->readers) {
1446                         send_sig(SIGPIPE, current, 0);
1447                         if (!ret)
1448                                 ret = -EPIPE;
1449                         break;
1450                 }
1451
1452                 /*
1453                  * If we have iterated all input buffers or ran out of
1454                  * output room, break.
1455                  */
1456                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1457                         break;
1458
1459                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1460                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1461
1462                 /*
1463                  * Get a reference to this pipe buffer,
1464                  * so we can copy the contents over.
1465                  */
1466                 ibuf->ops->get(ipipe, ibuf);
1467
1468                 obuf = opipe->bufs + nbuf;
1469                 *obuf = *ibuf;
1470
1471                 /*
1472                  * Don't inherit the gift flag, we need to
1473                  * prevent multiple steals of this page.
1474                  */
1475                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1476
1477                 if (obuf->len > len)
1478                         obuf->len = len;
1479
1480                 opipe->nrbufs++;
1481                 ret += obuf->len;
1482                 len -= obuf->len;
1483                 i++;
1484         } while (len);
1485
1486         inode_double_unlock(ipipe->inode, opipe->inode);
1487
1488         /*
1489          * If we put data in the output pipe, wakeup any potential readers.
1490          */
1491         if (ret > 0) {
1492                 smp_mb();
1493                 if (waitqueue_active(&opipe->wait))
1494                         wake_up_interruptible(&opipe->wait);
1495                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1496         }
1497
1498         return ret;
1499 }
1500
1501 /*
1502  * This is a tee(1) implementation that works on pipes. It doesn't copy
1503  * any data, it simply references the 'in' pages on the 'out' pipe.
1504  * The 'flags' used are the SPLICE_F_* variants, currently the only
1505  * applicable one is SPLICE_F_NONBLOCK.
1506  */
1507 static long do_tee(struct file *in, struct file *out, size_t len,
1508                    unsigned int flags)
1509 {
1510         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1511         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1512         int ret = -EINVAL;
1513
1514         /*
1515          * Duplicate the contents of ipipe to opipe without actually
1516          * copying the data.
1517          */
1518         if (ipipe && opipe && ipipe != opipe) {
1519                 /*
1520                  * Keep going, unless we encounter an error. The ipipe/opipe
1521                  * ordering doesn't really matter.
1522                  */
1523                 ret = link_ipipe_prep(ipipe, flags);
1524                 if (!ret) {
1525                         ret = link_opipe_prep(opipe, flags);
1526                         if (!ret) {
1527                                 ret = link_pipe(ipipe, opipe, len, flags);
1528                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1529                                         ret = -EAGAIN;
1530                         }
1531                 }
1532         }
1533
1534         return ret;
1535 }
1536
1537 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1538 {
1539         struct file *in;
1540         int error, fput_in;
1541
1542         if (unlikely(!len))
1543                 return 0;
1544
1545         error = -EBADF;
1546         in = fget_light(fdin, &fput_in);
1547         if (in) {
1548                 if (in->f_mode & FMODE_READ) {
1549                         int fput_out;
1550                         struct file *out = fget_light(fdout, &fput_out);
1551
1552                         if (out) {
1553                                 if (out->f_mode & FMODE_WRITE)
1554                                         error = do_tee(in, out, len, flags);
1555                                 fput_light(out, fput_out);
1556                         }
1557                 }
1558                 fput_light(in, fput_in);
1559         }
1560
1561         return error;
1562 }