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@suse.de>
  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
  31 /*
  32  * Passed to the actors
  33  */
  34 struct splice_desc {
  35         unsigned int len, total_len;    /* current and remaining length */
  36         unsigned int flags;             /* splice flags */
  37         struct file *file;              /* file to read/write */
  38         loff_t pos;                     /* file position */
  39 };
  40
  41 /*
  42  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  43  * a vm helper function, it's already simplified quite a bit by the
  44  * addition of remove_mapping(). If success is returned, the caller may
  45  * attempt to reuse this page for another destination.
  46  */
  47 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  48                                      struct pipe_buffer *buf)
  49 {
  50         struct page *page = buf->page;
  51         struct address_space *mapping = page_mapping(page);
  52
  53         WARN_ON(!PageLocked(page));
  54         WARN_ON(!PageUptodate(page));
  55
  56         /*
  57          * At least for ext2 with nobh option, we need to wait on writeback
  58          * completing on this page, since we'll remove it from the pagecache.
  59          * Otherwise truncate wont wait on the page, allowing the disk
  60          * blocks to be reused by someone else before we actually wrote our
  61          * data to them. fs corruption ensues.
  62          */
  63         wait_on_page_writeback(page);
  64
  65         if (PagePrivate(page))
  66                 try_to_release_page(page, mapping_gfp_mask(mapping));
  67
  68         if (!remove_mapping(mapping, page))
  69                 return 1;
  70
  71         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
  72         return 0;
  73 }
  74
  75 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  76                                         struct pipe_buffer *buf)
  77 {
  78         page_cache_release(buf->page);
  79         buf->page = NULL;
  80         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
  81 }
  82
  83 static void *page_cache_pipe_buf_map(struct file *file,
  84                                      struct pipe_inode_info *info,
  85                                      struct pipe_buffer *buf)
  86 {
  87         struct page *page = buf->page;
  88         int err;
  89
  90         if (!PageUptodate(page)) {
  91                 lock_page(page);
  92
  93                 /*
  94                  * Page got truncated/unhashed. This will cause a 0-byte
  95                  * splice, if this is the first page.
  96                  */
  97                 if (!page->mapping) {
  98                         err = -ENODATA;
  99                         goto error;
 100                 }
 101
 102                 /*
 103                  * Uh oh, read-error from disk.
 104                  */
 105                 if (!PageUptodate(page)) {
 106                         err = -EIO;
 107                         goto error;
 108                 }
 109
 110                 /*
 111                  * Page is ok afterall, fall through to mapping.
 112                  */
 113                 unlock_page(page);
 114         }
 115
 116         return kmap(page);
 117 error:
 118         unlock_page(page);
 119         return ERR_PTR(err);
 120 }
 121
 122 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
 123                                       struct pipe_buffer *buf)
 124 {
 125         kunmap(buf->page);
 126 }
 127
 128 static void page_cache_pipe_buf_get(struct pipe_inode_info *info,
 129                                     struct pipe_buffer *buf)
 130 {
 131         page_cache_get(buf->page);
 132 }
 133
 134 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 135         .can_merge = 0,
 136         .map = page_cache_pipe_buf_map,
 137         .unmap = page_cache_pipe_buf_unmap,
 138         .release = page_cache_pipe_buf_release,
 139         .steal = page_cache_pipe_buf_steal,
 140         .get = page_cache_pipe_buf_get,
 141 };
 142
 143 /*
 144  * Pipe output worker. This sets up our pipe format with the page cache
 145  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 146  */
 147 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
 148                             int nr_pages, unsigned long len,
 149                             unsigned int offset, unsigned int flags)
 150 {
 151         int ret, do_wakeup, i;
 152
 153         ret = 0;
 154         do_wakeup = 0;
 155         i = 0;
 156
 157         if (pipe->inode)
 158                 mutex_lock(&pipe->inode->i_mutex);
 159
 160         for (;;) {
 161                 if (!pipe->readers) {
 162                         send_sig(SIGPIPE, current, 0);
 163                         if (!ret)
 164                                 ret = -EPIPE;
 165                         break;
 166                 }
 167
 168                 if (pipe->nrbufs < PIPE_BUFFERS) {
 169                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 170                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 171                         struct page *page = pages[i++];
 172                         unsigned long this_len;
 173
 174                         this_len = PAGE_CACHE_SIZE - offset;
 175                         if (this_len > len)
 176                                 this_len = len;
 177
 178                         buf->page = page;
 179                         buf->offset = offset;
 180                         buf->len = this_len;
 181                         buf->ops = &page_cache_pipe_buf_ops;
 182                         pipe->nrbufs++;
 183                         if (pipe->inode)
 184                                 do_wakeup = 1;
 185
 186                         ret += this_len;
 187                         len -= this_len;
 188                         offset = 0;
 189                         if (!--nr_pages)
 190                                 break;
 191                         if (!len)
 192                                 break;
 193                         if (pipe->nrbufs < PIPE_BUFFERS)
 194                                 continue;
 195
 196                         break;
 197                 }
 198
 199                 if (flags & SPLICE_F_NONBLOCK) {
 200                         if (!ret)
 201                                 ret = -EAGAIN;
 202                         break;
 203                 }
 204
 205                 if (signal_pending(current)) {
 206                         if (!ret)
 207                                 ret = -ERESTARTSYS;
 208                         break;
 209                 }
 210
 211                 if (do_wakeup) {
 212                         smp_mb();
 213                         if (waitqueue_active(&pipe->wait))
 214                                 wake_up_interruptible_sync(&pipe->wait);
 215                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 216                         do_wakeup = 0;
 217                 }
 218
 219                 pipe->waiting_writers++;
 220                 pipe_wait(pipe);
 221                 pipe->waiting_writers--;
 222         }
 223
 224         if (pipe->inode)
 225                 mutex_unlock(&pipe->inode->i_mutex);
 226
 227         if (do_wakeup) {
 228                 smp_mb();
 229                 if (waitqueue_active(&pipe->wait))
 230                         wake_up_interruptible(&pipe->wait);
 231                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 232         }
 233
 234         while (i < nr_pages)
 235                 page_cache_release(pages[i++]);
 236
 237         return ret;
 238 }
 239
 240 static int
 241 __generic_file_splice_read(struct file *in, loff_t *ppos,
 242                            struct pipe_inode_info *pipe, size_t len,
 243                            unsigned int flags)
 244 {
 245         struct address_space *mapping = in->f_mapping;
 246         unsigned int loff, offset, nr_pages;
 247         struct page *pages[PIPE_BUFFERS];
 248         struct page *page;
 249         pgoff_t index, end_index;
 250         loff_t isize;
 251         size_t bytes;
 252         int i, error;
 253
 254         index = *ppos >> PAGE_CACHE_SHIFT;
 255         loff = offset = *ppos & ~PAGE_CACHE_MASK;
 256         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 257
 258         if (nr_pages > PIPE_BUFFERS)
 259                 nr_pages = PIPE_BUFFERS;
 260
 261         /*
 262          * Initiate read-ahead on this page range. however, don't call into
 263          * read-ahead if this is a non-zero offset (we are likely doing small
 264          * chunk splice and the page is already there) for a single page.
 265          */
 266         if (!offset || nr_pages > 1)
 267                 do_page_cache_readahead(mapping, in, index, nr_pages);
 268
 269         /*
 270          * Now fill in the holes:
 271          */
 272         error = 0;
 273         bytes = 0;
 274         for (i = 0; i < nr_pages; i++, index++) {
 275 find_page:
 276                 /*
 277                  * lookup the page for this index
 278                  */
 279                 page = find_get_page(mapping, index);
 280                 if (!page) {
 281                         /*
 282                          * page didn't exist, allocate one
 283                          */
 284                         page = page_cache_alloc_cold(mapping);
 285                         if (!page)
 286                                 break;
 287
 288                         error = add_to_page_cache_lru(page, mapping, index,
 289                                                 mapping_gfp_mask(mapping));
 290                         if (unlikely(error)) {
 291                                 page_cache_release(page);
 292                                 break;
 293                         }
 294
 295                         goto readpage;
 296                 }
 297
 298                 /*
 299                  * If the page isn't uptodate, we may need to start io on it
 300                  */
 301                 if (!PageUptodate(page)) {
 302                         /*
 303                          * If in nonblock mode then dont block on waiting
 304                          * for an in-flight io page
 305                          */
 306                         if (flags & SPLICE_F_NONBLOCK)
 307                                 break;
 308
 309                         lock_page(page);
 310
 311                         /*
 312                          * page was truncated, stop here. if this isn't the
 313                          * first page, we'll just complete what we already
 314                          * added
 315                          */
 316                         if (!page->mapping) {
 317                                 unlock_page(page);
 318                                 page_cache_release(page);
 319                                 break;
 320                         }
 321                         /*
 322                          * page was already under io and is now done, great
 323                          */
 324                         if (PageUptodate(page)) {
 325                                 unlock_page(page);
 326                                 goto fill_it;
 327                         }
 328
 329 readpage:
 330                         /*
 331                          * need to read in the page
 332                          */
 333                         error = mapping->a_ops->readpage(in, page);
 334
 335                         if (unlikely(error)) {
 336                                 page_cache_release(page);
 337                                 if (error == AOP_TRUNCATED_PAGE)
 338                                         goto find_page;
 339                                 break;
 340                         }
 341
 342                         /*
 343                          * i_size must be checked after ->readpage().
 344                          */
 345                         isize = i_size_read(mapping->host);
 346                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 347                         if (unlikely(!isize || index > end_index)) {
 348                                 page_cache_release(page);
 349                                 break;
 350                         }
 351
 352                         /*
 353                          * if this is the last page, see if we need to shrink
 354                          * the length and stop
 355                          */
 356                         if (end_index == index) {
 357                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 358                                 if (bytes + loff > isize) {
 359                                         page_cache_release(page);
 360                                         break;
 361                                 }
 362                                 /*
 363                                  * force quit after adding this page
 364                                  */
 365                                 nr_pages = i;
 366                         }
 367                 }
 368 fill_it:
 369                 pages[i] = page;
 370                 bytes += PAGE_CACHE_SIZE - loff;
 371                 loff = 0;
 372         }
 373
 374         if (i)
 375                 return move_to_pipe(pipe, pages, i, bytes, offset, flags);
 376
 377         return error;
 378 }
 379
 380 /**
 381  * generic_file_splice_read - splice data from file to a pipe
 382  * @in:         file to splice from
 383  * @pipe:       pipe to splice to
 384  * @len:        number of bytes to splice
 385  * @flags:      splice modifier flags
 386  *
 387  * Will read pages from given file and fill them into a pipe.
 388  */
 389 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 390                                  struct pipe_inode_info *pipe, size_t len,
 391                                  unsigned int flags)
 392 {
 393         ssize_t spliced;
 394         int ret;
 395
 396         ret = 0;
 397         spliced = 0;
 398
 399         while (len) {
 400                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 401
 402                 if (ret < 0)
 403                         break;
 404                 else if (!ret) {
 405                         if (spliced)
 406                                 break;
 407                         if (flags & SPLICE_F_NONBLOCK) {
 408                                 ret = -EAGAIN;
 409                                 break;
 410                         }
 411                 }
 412
 413                 *ppos += ret;
 414                 len -= ret;
 415                 spliced += ret;
 416         }
 417
 418         if (spliced)
 419                 return spliced;
 420
 421         return ret;
 422 }
 423
 424 EXPORT_SYMBOL(generic_file_splice_read);
 425
 426 /*
 427  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 428  * using sendpage().
 429  */
 430 static int pipe_to_sendpage(struct pipe_inode_info *info,
 431                             struct pipe_buffer *buf, struct splice_desc *sd)
 432 {
 433         struct file *file = sd->file;
 434         loff_t pos = sd->pos;
 435         unsigned int offset;
 436         ssize_t ret;
 437         void *ptr;
 438         int more;
 439
 440         /*
 441          * Sub-optimal, but we are limited by the pipe ->map. We don't
 442          * need a kmap'ed buffer here, we just want to make sure we
 443          * have the page pinned if the pipe page originates from the
 444          * page cache.
 445          */
 446         ptr = buf->ops->map(file, info, buf);
 447         if (IS_ERR(ptr))
 448                 return PTR_ERR(ptr);
 449
 450         offset = pos & ~PAGE_CACHE_MASK;
 451         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 452
 453         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 454
 455         buf->ops->unmap(info, buf);
 456         if (ret == sd->len)
 457                 return 0;
 458
 459         return -EIO;
 460 }
 461
 462 /*
 463  * This is a little more tricky than the file -> pipe splicing. There are
 464  * basically three cases:
 465  *
 466  *      - Destination page already exists in the address space and there
 467  *        are users of it. For that case we have no other option that
 468  *        copying the data. Tough luck.
 469  *      - Destination page already exists in the address space, but there
 470  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 471  *        through to last case.
 472  *      - Destination page does not exist, we can add the pipe page to
 473  *        the page cache and avoid the copy.
 474  *
 475  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 476  * sd->flags), we attempt to migrate pages from the pipe to the output
 477  * file address space page cache. This is possible if no one else has
 478  * the pipe page referenced outside of the pipe and page cache. If
 479  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 480  * a new page in the output file page cache and fill/dirty that.
 481  */
 482 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 483                         struct splice_desc *sd)
 484 {
 485         struct file *file = sd->file;
 486         struct address_space *mapping = file->f_mapping;
 487         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 488         unsigned int offset;
 489         struct page *page;
 490         pgoff_t index;
 491         char *src;
 492         int ret;
 493
 494         /*
 495          * make sure the data in this buffer is uptodate
 496          */
 497         src = buf->ops->map(file, info, buf);
 498         if (IS_ERR(src))
 499                 return PTR_ERR(src);
 500
 501         index = sd->pos >> PAGE_CACHE_SHIFT;
 502         offset = sd->pos & ~PAGE_CACHE_MASK;
 503
 504         /*
 505          * Reuse buf page, if SPLICE_F_MOVE is set.
 506          */
 507         if (sd->flags & SPLICE_F_MOVE) {
 508                 /*
 509                  * If steal succeeds, buf->page is now pruned from the vm
 510                  * side (LRU and page cache) and we can reuse it.
 511                  */
 512                 if (buf->ops->steal(info, buf))
 513                         goto find_page;
 514
 515                 /*
 516                  * this will also set the page locked
 517                  */
 518                 page = buf->page;
 519                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 520                         goto find_page;
 521
 522                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 523                         lru_cache_add(page);
 524         } else {
 525 find_page:
 526                 ret = -ENOMEM;
 527                 page = find_or_create_page(mapping, index, gfp_mask);
 528                 if (!page)
 529                         goto out_nomem;
 530
 531                 /*
 532                  * If the page is uptodate, it is also locked. If it isn't
 533                  * uptodate, we can mark it uptodate if we are filling the
 534                  * full page. Otherwise we need to read it in first...
 535                  */
 536                 if (!PageUptodate(page)) {
 537                         if (sd->len < PAGE_CACHE_SIZE) {
 538                                 ret = mapping->a_ops->readpage(file, page);
 539                                 if (unlikely(ret))
 540                                         goto out;
 541
 542                                 lock_page(page);
 543
 544                                 if (!PageUptodate(page)) {
 545                                         /*
 546                                          * Page got invalidated, repeat.
 547                                          */
 548                                         if (!page->mapping) {
 549                                                 unlock_page(page);
 550                                                 page_cache_release(page);
 551                                                 goto find_page;
 552                                         }
 553                                         ret = -EIO;
 554                                         goto out;
 555                                 }
 556                         } else {
 557                                 WARN_ON(!PageLocked(page));
 558                                 SetPageUptodate(page);
 559                         }
 560                 }
 561         }
 562
 563         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 564         if (ret == AOP_TRUNCATED_PAGE) {
 565                 page_cache_release(page);
 566                 goto find_page;
 567         } else if (ret)
 568                 goto out;
 569
 570         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 571                 char *dst = kmap_atomic(page, KM_USER0);
 572
 573                 memcpy(dst + offset, src + buf->offset, sd->len);
 574                 flush_dcache_page(page);
 575                 kunmap_atomic(dst, KM_USER0);
 576         }
 577
 578         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 579         if (ret == AOP_TRUNCATED_PAGE) {
 580                 page_cache_release(page);
 581                 goto find_page;
 582         } else if (ret)
 583                 goto out;
 584
 585         mark_page_accessed(page);
 586         balance_dirty_pages_ratelimited(mapping);
 587 out:
 588         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 589                 page_cache_release(page);
 590                 unlock_page(page);
 591         }
 592 out_nomem:
 593         buf->ops->unmap(info, buf);
 594         return ret;
 595 }
 596
 597 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 598                            struct splice_desc *);
 599
 600 /*
 601  * Pipe input worker. Most of this logic works like a regular pipe, the
 602  * key here is the 'actor' worker passed in that actually moves the data
 603  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 604  */
 605 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 606                               loff_t *ppos, size_t len, unsigned int flags,
 607                               splice_actor *actor)
 608 {
 609         int ret, do_wakeup, err;
 610         struct splice_desc sd;
 611
 612         ret = 0;
 613         do_wakeup = 0;
 614
 615         sd.total_len = len;
 616         sd.flags = flags;
 617         sd.file = out;
 618         sd.pos = *ppos;
 619
 620         if (pipe->inode)
 621                 mutex_lock(&pipe->inode->i_mutex);
 622
 623         for (;;) {
 624                 if (pipe->nrbufs) {
 625                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 626                         struct pipe_buf_operations *ops = buf->ops;
 627
 628                         sd.len = buf->len;
 629                         if (sd.len > sd.total_len)
 630                                 sd.len = sd.total_len;
 631
 632                         err = actor(pipe, buf, &sd);
 633                         if (err) {
 634                                 if (!ret && err != -ENODATA)
 635                                         ret = err;
 636
 637                                 break;
 638                         }
 639
 640                         ret += sd.len;
 641                         buf->offset += sd.len;
 642                         buf->len -= sd.len;
 643
 644                         if (!buf->len) {
 645                                 buf->ops = NULL;
 646                                 ops->release(pipe, buf);
 647                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 648                                 pipe->nrbufs--;
 649                                 if (pipe->inode)
 650                                         do_wakeup = 1;
 651                         }
 652
 653                         sd.pos += sd.len;
 654                         sd.total_len -= sd.len;
 655                         if (!sd.total_len)
 656                                 break;
 657                 }
 658
 659                 if (pipe->nrbufs)
 660                         continue;
 661                 if (!pipe->writers)
 662                         break;
 663                 if (!pipe->waiting_writers) {
 664                         if (ret)
 665                                 break;
 666                 }
 667
 668                 if (flags & SPLICE_F_NONBLOCK) {
 669                         if (!ret)
 670                                 ret = -EAGAIN;
 671                         break;
 672                 }
 673
 674                 if (signal_pending(current)) {
 675                         if (!ret)
 676                                 ret = -ERESTARTSYS;
 677                         break;
 678                 }
 679
 680                 if (do_wakeup) {
 681                         smp_mb();
 682                         if (waitqueue_active(&pipe->wait))
 683                                 wake_up_interruptible_sync(&pipe->wait);
 684                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 685                         do_wakeup = 0;
 686                 }
 687
 688                 pipe_wait(pipe);
 689         }
 690
 691         if (pipe->inode)
 692                 mutex_unlock(&pipe->inode->i_mutex);
 693
 694         if (do_wakeup) {
 695                 smp_mb();
 696                 if (waitqueue_active(&pipe->wait))
 697                         wake_up_interruptible(&pipe->wait);
 698                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 699         }
 700
 701         return ret;
 702 }
 703
 704 /**
 705  * generic_file_splice_write - splice data from a pipe to a file
 706  * @pipe:       pipe info
 707  * @out:        file to write to
 708  * @len:        number of bytes to splice
 709  * @flags:      splice modifier flags
 710  *
 711  * Will either move or copy pages (determined by @flags options) from
 712  * the given pipe inode to the given file.
 713  *
 714  */
 715 ssize_t
 716 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 717                           loff_t *ppos, size_t len, unsigned int flags)
 718 {
 719         struct address_space *mapping = out->f_mapping;
 720         ssize_t ret;
 721
 722         ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 723
 724         /*
 725          * If file or inode is SYNC and we actually wrote some data, sync it.
 726          */
 727         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 728             && ret > 0) {
 729                 struct inode *inode = mapping->host;
 730                 int err;
 731
 732                 mutex_lock(&inode->i_mutex);
 733                 err = generic_osync_inode(mapping->host, mapping,
 734                                           OSYNC_METADATA|OSYNC_DATA);
 735                 mutex_unlock(&inode->i_mutex);
 736
 737                 if (err)
 738                         ret = err;
 739         }
 740
 741         return ret;
 742 }
 743
 744 EXPORT_SYMBOL(generic_file_splice_write);
 745
 746 /**
 747  * generic_splice_sendpage - splice data from a pipe to a socket
 748  * @inode:      pipe inode
 749  * @out:        socket to write to
 750  * @len:        number of bytes to splice
 751  * @flags:      splice modifier flags
 752  *
 753  * Will send @len bytes from the pipe to a network socket. No data copying
 754  * is involved.
 755  *
 756  */
 757 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 758                                 loff_t *ppos, size_t len, unsigned int flags)
 759 {
 760         return move_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 761 }
 762
 763 EXPORT_SYMBOL(generic_splice_sendpage);
 764
 765 /*
 766  * Attempt to initiate a splice from pipe to file.
 767  */
 768 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 769                            loff_t *ppos, size_t len, unsigned int flags)
 770 {
 771         int ret;
 772
 773         if (unlikely(!out->f_op || !out->f_op->splice_write))
 774                 return -EINVAL;
 775
 776         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 777                 return -EBADF;
 778
 779         ret = rw_verify_area(WRITE, out, ppos, len);
 780         if (unlikely(ret < 0))
 781                 return ret;
 782
 783         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 784 }
 785
 786 /*
 787  * Attempt to initiate a splice from a file to a pipe.
 788  */
 789 static long do_splice_to(struct file *in, loff_t *ppos,
 790                          struct pipe_inode_info *pipe, size_t len,
 791                          unsigned int flags)
 792 {
 793         loff_t isize, left;
 794         int ret;
 795
 796         if (unlikely(!in->f_op || !in->f_op->splice_read))
 797                 return -EINVAL;
 798
 799         if (unlikely(!(in->f_mode & FMODE_READ)))
 800                 return -EBADF;
 801
 802         ret = rw_verify_area(READ, in, ppos, len);
 803         if (unlikely(ret < 0))
 804                 return ret;
 805
 806         isize = i_size_read(in->f_mapping->host);
 807         if (unlikely(*ppos >= isize))
 808                 return 0;
 809
 810         left = isize - *ppos;
 811         if (unlikely(left < len))
 812                 len = left;
 813
 814         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 815 }
 816
 817 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 818                       size_t len, unsigned int flags)
 819 {
 820         struct pipe_inode_info *pipe;
 821         long ret, bytes;
 822         loff_t out_off;
 823         umode_t i_mode;
 824         int i;
 825
 826         /*
 827          * We require the input being a regular file, as we don't want to
 828          * randomly drop data for eg socket -> socket splicing. Use the
 829          * piped splicing for that!
 830          */
 831         i_mode = in->f_dentry->d_inode->i_mode;
 832         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 833                 return -EINVAL;
 834
 835         /*
 836          * neither in nor out is a pipe, setup an internal pipe attached to
 837          * 'out' and transfer the wanted data from 'in' to 'out' through that
 838          */
 839         pipe = current->splice_pipe;
 840         if (unlikely(!pipe)) {
 841                 pipe = alloc_pipe_info(NULL);
 842                 if (!pipe)
 843                         return -ENOMEM;
 844
 845                 /*
 846                  * We don't have an immediate reader, but we'll read the stuff
 847                  * out of the pipe right after the move_to_pipe(). So set
 848                  * PIPE_READERS appropriately.
 849                  */
 850                 pipe->readers = 1;
 851
 852                 current->splice_pipe = pipe;
 853         }
 854
 855         /*
 856          * Do the splice.
 857          */
 858         ret = 0;
 859         bytes = 0;
 860         out_off = 0;
 861
 862         while (len) {
 863                 size_t read_len, max_read_len;
 864
 865                 /*
 866                  * Do at most PIPE_BUFFERS pages worth of transfer:
 867                  */
 868                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 869
 870                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 871                 if (unlikely(ret < 0))
 872                         goto out_release;
 873
 874                 read_len = ret;
 875
 876                 /*
 877                  * NOTE: nonblocking mode only applies to the input. We
 878                  * must not do the output in nonblocking mode as then we
 879                  * could get stuck data in the internal pipe:
 880                  */
 881                 ret = do_splice_from(pipe, out, &out_off, read_len,
 882                                      flags & ~SPLICE_F_NONBLOCK);
 883                 if (unlikely(ret < 0))
 884                         goto out_release;
 885
 886                 bytes += ret;
 887                 len -= ret;
 888
 889                 /*
 890                  * In nonblocking mode, if we got back a short read then
 891                  * that was due to either an IO error or due to the
 892                  * pagecache entry not being there. In the IO error case
 893                  * the _next_ splice attempt will produce a clean IO error
 894                  * return value (not a short read), so in both cases it's
 895                  * correct to break out of the loop here:
 896                  */
 897                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 898                         break;
 899         }
 900
 901         pipe->nrbufs = pipe->curbuf = 0;
 902
 903         return bytes;
 904
 905 out_release:
 906         /*
 907          * If we did an incomplete transfer we must release
 908          * the pipe buffers in question:
 909          */
 910         for (i = 0; i < PIPE_BUFFERS; i++) {
 911                 struct pipe_buffer *buf = pipe->bufs + i;
 912
 913                 if (buf->ops) {
 914                         buf->ops->release(pipe, buf);
 915                         buf->ops = NULL;
 916                 }
 917         }
 918         pipe->nrbufs = pipe->curbuf = 0;
 919
 920         /*
 921          * If we transferred some data, return the number of bytes:
 922          */
 923         if (bytes > 0)
 924                 return bytes;
 925
 926         return ret;
 927 }
 928
 929 EXPORT_SYMBOL(do_splice_direct);
 930
 931 /*
 932  * Determine where to splice to/from.
 933  */
 934 static long do_splice(struct file *in, loff_t __user *off_in,
 935                       struct file *out, loff_t __user *off_out,
 936                       size_t len, unsigned int flags)
 937 {
 938         struct pipe_inode_info *pipe;
 939         loff_t offset, *off;
 940
 941         pipe = in->f_dentry->d_inode->i_pipe;
 942         if (pipe) {
 943                 if (off_in)
 944                         return -ESPIPE;
 945                 if (off_out) {
 946                         if (out->f_op->llseek == no_llseek)
 947                                 return -EINVAL;
 948                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
 949                                 return -EFAULT;
 950                         off = &offset;
 951                 } else
 952                         off = &out->f_pos;
 953
 954                 return do_splice_from(pipe, out, off, len, flags);
 955         }
 956
 957         pipe = out->f_dentry->d_inode->i_pipe;
 958         if (pipe) {
 959                 if (off_out)
 960                         return -ESPIPE;
 961                 if (off_in) {
 962                         if (in->f_op->llseek == no_llseek)
 963                                 return -EINVAL;
 964                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
 965                                 return -EFAULT;
 966                         off = &offset;
 967                 } else
 968                         off = &in->f_pos;
 969
 970                 return do_splice_to(in, off, pipe, len, flags);
 971         }
 972
 973         return -EINVAL;
 974 }
 975
 976 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 977                            int fd_out, loff_t __user *off_out,
 978                            size_t len, unsigned int flags)
 979 {
 980         long error;
 981         struct file *in, *out;
 982         int fput_in, fput_out;
 983
 984         if (unlikely(!len))
 985                 return 0;
 986
 987         error = -EBADF;
 988         in = fget_light(fd_in, &fput_in);
 989         if (in) {
 990                 if (in->f_mode & FMODE_READ) {
 991                         out = fget_light(fd_out, &fput_out);
 992                         if (out) {
 993                                 if (out->f_mode & FMODE_WRITE)
 994                                         error = do_splice(in, off_in,
 995                                                           out, off_out,
 996                                                           len, flags);
 997                                 fput_light(out, fput_out);
 998                         }
 999                 }
1000
1001                 fput_light(in, fput_in);
1002         }
1003
1004         return error;
1005 }
1006
1007 /*
1008  * Link contents of ipipe to opipe.
1009  */
1010 static int link_pipe(struct pipe_inode_info *ipipe,
1011                      struct pipe_inode_info *opipe,
1012                      size_t len, unsigned int flags)
1013 {
1014         struct pipe_buffer *ibuf, *obuf;
1015         int ret = 0, do_wakeup = 0, i;
1016
1017         /*
1018          * Potential ABBA deadlock, work around it by ordering lock
1019          * grabbing by inode address. Otherwise two different processes
1020          * could deadlock (one doing tee from A -> B, the other from B -> A).
1021          */
1022         if (ipipe->inode < opipe->inode) {
1023                 mutex_lock(&ipipe->inode->i_mutex);
1024                 mutex_lock(&opipe->inode->i_mutex);
1025         } else {
1026                 mutex_lock(&opipe->inode->i_mutex);
1027                 mutex_lock(&ipipe->inode->i_mutex);
1028         }
1029
1030         for (i = 0;; i++) {
1031                 if (!opipe->readers) {
1032                         send_sig(SIGPIPE, current, 0);
1033                         if (!ret)
1034                                 ret = -EPIPE;
1035                         break;
1036                 }
1037                 if (ipipe->nrbufs - i) {
1038                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1039
1040                         /*
1041                          * If we have room, fill this buffer
1042                          */
1043                         if (opipe->nrbufs < PIPE_BUFFERS) {
1044                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1045
1046                                 /*
1047                                  * Get a reference to this pipe buffer,
1048                                  * so we can copy the contents over.
1049                                  */
1050                                 ibuf->ops->get(ipipe, ibuf);
1051
1052                                 obuf = opipe->bufs + nbuf;
1053                                 *obuf = *ibuf;
1054
1055                                 if (obuf->len > len)
1056                                         obuf->len = len;
1057
1058                                 opipe->nrbufs++;
1059                                 do_wakeup = 1;
1060                                 ret += obuf->len;
1061                                 len -= obuf->len;
1062
1063                                 if (!len)
1064                                         break;
1065                                 if (opipe->nrbufs < PIPE_BUFFERS)
1066                                         continue;
1067                         }
1068
1069                         /*
1070                          * We have input available, but no output room.
1071                          * If we already copied data, return that.
1072                          */
1073                         if (flags & SPLICE_F_NONBLOCK) {
1074                                 if (!ret)
1075                                         ret = -EAGAIN;
1076                                 break;
1077                         }
1078                         if (signal_pending(current)) {
1079                                 if (!ret)
1080                                         ret = -ERESTARTSYS;
1081                                 break;
1082                         }
1083                         if (do_wakeup) {
1084                                 smp_mb();
1085                                 if (waitqueue_active(&opipe->wait))
1086                                         wake_up_interruptible(&opipe->wait);
1087                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1088                                 do_wakeup = 0;
1089                         }
1090
1091                         opipe->waiting_writers++;
1092                         pipe_wait(opipe);
1093                         opipe->waiting_writers--;
1094                         continue;
1095                 }
1096
1097                 /*
1098                  * No input buffers, do the usual checks for available
1099                  * writers and blocking and wait if necessary
1100                  */
1101                 if (!ipipe->writers)
1102                         break;
1103                 if (!ipipe->waiting_writers) {
1104                         if (ret)
1105                                 break;
1106                 }
1107                 if (flags & SPLICE_F_NONBLOCK) {
1108                         if (!ret)
1109                                 ret = -EAGAIN;
1110                         break;
1111                 }
1112                 if (signal_pending(current)) {
1113                         if (!ret)
1114                                 ret = -ERESTARTSYS;
1115                         break;
1116                 }
1117
1118                 if (waitqueue_active(&ipipe->wait))
1119                         wake_up_interruptible_sync(&ipipe->wait);
1120                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1121
1122                 pipe_wait(ipipe);
1123         }
1124
1125         mutex_unlock(&ipipe->inode->i_mutex);
1126         mutex_unlock(&opipe->inode->i_mutex);
1127
1128         if (do_wakeup) {
1129                 smp_mb();
1130                 if (waitqueue_active(&opipe->wait))
1131                         wake_up_interruptible(&opipe->wait);
1132                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1133         }
1134
1135         return ret;
1136 }
1137
1138 /*
1139  * This is a tee(1) implementation that works on pipes. It doesn't copy
1140  * any data, it simply references the 'in' pages on the 'out' pipe.
1141  * The 'flags' used are the SPLICE_F_* variants, currently the only
1142  * applicable one is SPLICE_F_NONBLOCK.
1143  */
1144 static long do_tee(struct file *in, struct file *out, size_t len,
1145                    unsigned int flags)
1146 {
1147         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1148         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1149
1150         /*
1151          * Link ipipe to the two output pipes, consuming as we go along.
1152          */
1153         if (ipipe && opipe)
1154                 return link_pipe(ipipe, opipe, len, flags);
1155
1156         return -EINVAL;
1157 }
1158
1159 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1160 {
1161         struct file *in;
1162         int error, fput_in;
1163
1164         if (unlikely(!len))
1165                 return 0;
1166
1167         error = -EBADF;
1168         in = fget_light(fdin, &fput_in);
1169         if (in) {
1170                 if (in->f_mode & FMODE_READ) {
1171                         int fput_out;
1172                         struct file *out = fget_light(fdout, &fput_out);
1173
1174                         if (out) {
1175                                 if (out->f_mode & FMODE_WRITE)
1176                                         error = do_tee(in, out, len, flags);
1177                                 fput_light(out, fput_out);
1178                         }
1179                 }
1180                 fput_light(in, fput_in);
1181         }
1182
1183         return error;
1184 }