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 and fixing the initial implementation
  13  * bugs.
  14  *
  15  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
  16  * Copyright (C) 2005 Linus Torvalds <torvalds@osdl.org>
  17  *
  18  */
  19 #include <linux/fs.h>
  20 #include <linux/file.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/pipe_fs_i.h>
  23 #include <linux/mm_inline.h>
  24 #include <linux/swap.h>
  25 #include <linux/writeback.h>
  26 #include <linux/buffer_head.h>
  27 #include <linux/module.h>
  28 #include <linux/syscalls.h>
  29
  30 /*
  31  * Passed to the actors
  32  */
  33 struct splice_desc {
  34         unsigned int len, total_len;    /* current and remaining length */
  35         unsigned int flags;             /* splice flags */
  36         struct file *file;              /* file to read/write */
  37         loff_t pos;                     /* file position */
  38 };
  39
  40 /*
  41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42  * a vm helper function, it's already simplified quite a bit by the
  43  * addition of remove_mapping(). If success is returned, the caller may
  44  * attempt to reuse this page for another destination.
  45  */
  46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  47                                      struct pipe_buffer *buf)
  48 {
  49         struct page *page = buf->page;
  50         struct address_space *mapping = page_mapping(page);
  51
  52         WARN_ON(!PageLocked(page));
  53         WARN_ON(!PageUptodate(page));
  54
  55         /*
  56          * At least for ext2 with nobh option, we need to wait on writeback
  57          * completing on this page, since we'll remove it from the pagecache.
  58          * Otherwise truncate wont wait on the page, allowing the disk
  59          * blocks to be reused by someone else before we actually wrote our
  60          * data to them. fs corruption ensues.
  61          */
  62         wait_on_page_writeback(page);
  63
  64         if (PagePrivate(page))
  65                 try_to_release_page(page, mapping_gfp_mask(mapping));
  66
  67         if (!remove_mapping(mapping, page))
  68                 return 1;
  69
  70         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
  71         return 0;
  72 }
  73
  74 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  75                                         struct pipe_buffer *buf)
  76 {
  77         page_cache_release(buf->page);
  78         buf->page = NULL;
  79         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
  80 }
  81
  82 static void *page_cache_pipe_buf_map(struct file *file,
  83                                      struct pipe_inode_info *info,
  84                                      struct pipe_buffer *buf)
  85 {
  86         struct page *page = buf->page;
  87         int err;
  88
  89         if (!PageUptodate(page)) {
  90                 lock_page(page);
  91
  92                 /*
  93                  * Page got truncated/unhashed. This will cause a 0-byte
  94                  * splice, if this is the first page
  95                  */
  96                 if (!page->mapping) {
  97                         err = -ENODATA;
  98                         goto error;
  99                 }
 100
 101                 /*
 102                  * uh oh, read-error from disk
 103                  */
 104                 if (!PageUptodate(page)) {
 105                         err = -EIO;
 106                         goto error;
 107                 }
 108
 109                 /*
 110                  * page is ok afterall, fall through to mapping
 111                  */
 112                 unlock_page(page);
 113         }
 114
 115         return kmap(page);
 116 error:
 117         unlock_page(page);
 118         return ERR_PTR(err);
 119 }
 120
 121 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
 122                                       struct pipe_buffer *buf)
 123 {
 124         kunmap(buf->page);
 125 }
 126
 127 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 128         .can_merge = 0,
 129         .map = page_cache_pipe_buf_map,
 130         .unmap = page_cache_pipe_buf_unmap,
 131         .release = page_cache_pipe_buf_release,
 132         .steal = page_cache_pipe_buf_steal,
 133 };
 134
 135 /*
 136  * Pipe output worker. This sets up our pipe format with the page cache
 137  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 138  */
 139 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
 140                             int nr_pages, unsigned long offset,
 141                             unsigned long len, unsigned int flags)
 142 {
 143         int ret, do_wakeup, i;
 144
 145         ret = 0;
 146         do_wakeup = 0;
 147         i = 0;
 148
 149         if (pipe->inode)
 150                 mutex_lock(&pipe->inode->i_mutex);
 151
 152         for (;;) {
 153                 int bufs;
 154
 155                 if (!pipe->readers) {
 156                         send_sig(SIGPIPE, current, 0);
 157                         if (!ret)
 158                                 ret = -EPIPE;
 159                         break;
 160                 }
 161
 162                 bufs = pipe->nrbufs;
 163                 if (bufs < PIPE_BUFFERS) {
 164                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS - 1);
 165                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 166                         struct page *page = pages[i++];
 167                         unsigned long this_len;
 168
 169                         this_len = PAGE_CACHE_SIZE - offset;
 170                         if (this_len > len)
 171                                 this_len = len;
 172
 173                         buf->page = page;
 174                         buf->offset = offset;
 175                         buf->len = this_len;
 176                         buf->ops = &page_cache_pipe_buf_ops;
 177                         pipe->nrbufs = ++bufs;
 178                         do_wakeup = 1;
 179
 180                         ret += this_len;
 181                         len -= this_len;
 182                         offset = 0;
 183                         if (!--nr_pages)
 184                                 break;
 185                         if (!len)
 186                                 break;
 187                         if (bufs < PIPE_BUFFERS)
 188                                 continue;
 189
 190                         break;
 191                 }
 192
 193                 if (flags & SPLICE_F_NONBLOCK) {
 194                         if (!ret)
 195                                 ret = -EAGAIN;
 196                         break;
 197                 }
 198
 199                 if (signal_pending(current)) {
 200                         if (!ret)
 201                                 ret = -ERESTARTSYS;
 202                         break;
 203                 }
 204
 205                 if (do_wakeup) {
 206                         smp_mb();
 207                         if (waitqueue_active(&pipe->wait))
 208                                 wake_up_interruptible_sync(&pipe->wait);
 209                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 210                         do_wakeup = 0;
 211                 }
 212
 213                 pipe->waiting_writers++;
 214                 pipe_wait(pipe);
 215                 pipe->waiting_writers--;
 216         }
 217
 218         if (pipe->inode)
 219                 mutex_unlock(&pipe->inode->i_mutex);
 220
 221         if (do_wakeup) {
 222                 smp_mb();
 223                 if (waitqueue_active(&pipe->wait))
 224                         wake_up_interruptible(&pipe->wait);
 225                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 226         }
 227
 228         while (i < nr_pages)
 229                 page_cache_release(pages[i++]);
 230
 231         return ret;
 232 }
 233
 234 static int
 235 __generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 236                            size_t len, unsigned int flags)
 237 {
 238         struct address_space *mapping = in->f_mapping;
 239         unsigned int offset, nr_pages;
 240         struct page *pages[PIPE_BUFFERS];
 241         struct page *page;
 242         pgoff_t index;
 243         int i;
 244
 245         index = in->f_pos >> PAGE_CACHE_SHIFT;
 246         offset = in->f_pos & ~PAGE_CACHE_MASK;
 247         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 248
 249         if (nr_pages > PIPE_BUFFERS)
 250                 nr_pages = PIPE_BUFFERS;
 251
 252         /*
 253          * initiate read-ahead on this page range. however, don't call into
 254          * read-ahead if this is a non-zero offset (we are likely doing small
 255          * chunk splice and the page is already there) for a single page.
 256          */
 257         if (!offset || nr_pages > 1)
 258                 do_page_cache_readahead(mapping, in, index, nr_pages);
 259
 260         /*
 261          * now fill in the holes
 262          */
 263         for (i = 0; i < nr_pages; i++, index++) {
 264                 /*
 265                  * no page there, look one up / create it
 266                  */
 267                 page = find_or_create_page(mapping, index,
 268                                                    mapping_gfp_mask(mapping));
 269                 if (!page)
 270                         break;
 271
 272                 if (PageUptodate(page))
 273                         unlock_page(page);
 274                 else {
 275                         int error = mapping->a_ops->readpage(in, page);
 276
 277                         if (unlikely(error)) {
 278                                 page_cache_release(page);
 279                                 break;
 280                         }
 281                 }
 282                 pages[i] = page;
 283         }
 284
 285         if (i)
 286                 return move_to_pipe(pipe, pages, i, offset, len, flags);
 287
 288         return 0;
 289 }
 290
 291 /**
 292  * generic_file_splice_read - splice data from file to a pipe
 293  * @in:         file to splice from
 294  * @pipe:       pipe to splice to
 295  * @len:        number of bytes to splice
 296  * @flags:      splice modifier flags
 297  *
 298  * Will read pages from given file and fill them into a pipe.
 299  *
 300  */
 301 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 302                                  size_t len, unsigned int flags)
 303 {
 304         ssize_t spliced;
 305         int ret;
 306
 307         ret = 0;
 308         spliced = 0;
 309
 310         while (len) {
 311                 ret = __generic_file_splice_read(in, pipe, len, flags);
 312
 313                 if (ret <= 0)
 314                         break;
 315
 316                 in->f_pos += ret;
 317                 len -= ret;
 318                 spliced += ret;
 319
 320                 if (!(flags & SPLICE_F_NONBLOCK))
 321                         continue;
 322                 ret = -EAGAIN;
 323                 break;
 324         }
 325
 326         if (spliced)
 327                 return spliced;
 328
 329         return ret;
 330 }
 331
 332 EXPORT_SYMBOL(generic_file_splice_read);
 333
 334 /*
 335  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 336  * using sendpage().
 337  */
 338 static int pipe_to_sendpage(struct pipe_inode_info *info,
 339                             struct pipe_buffer *buf, struct splice_desc *sd)
 340 {
 341         struct file *file = sd->file;
 342         loff_t pos = sd->pos;
 343         unsigned int offset;
 344         ssize_t ret;
 345         void *ptr;
 346         int more;
 347
 348         /*
 349          * sub-optimal, but we are limited by the pipe ->map. we don't
 350          * need a kmap'ed buffer here, we just want to make sure we
 351          * have the page pinned if the pipe page originates from the
 352          * page cache
 353          */
 354         ptr = buf->ops->map(file, info, buf);
 355         if (IS_ERR(ptr))
 356                 return PTR_ERR(ptr);
 357
 358         offset = pos & ~PAGE_CACHE_MASK;
 359         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 360
 361         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 362
 363         buf->ops->unmap(info, buf);
 364         if (ret == sd->len)
 365                 return 0;
 366
 367         return -EIO;
 368 }
 369
 370 /*
 371  * This is a little more tricky than the file -> pipe splicing. There are
 372  * basically three cases:
 373  *
 374  *      - Destination page already exists in the address space and there
 375  *        are users of it. For that case we have no other option that
 376  *        copying the data. Tough luck.
 377  *      - Destination page already exists in the address space, but there
 378  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 379  *        through to last case.
 380  *      - Destination page does not exist, we can add the pipe page to
 381  *        the page cache and avoid the copy.
 382  *
 383  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 384  * sd->flags), we attempt to migrate pages from the pipe to the output
 385  * file address space page cache. This is possible if no one else has
 386  * the pipe page referenced outside of the pipe and page cache. If
 387  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 388  * a new page in the output file page cache and fill/dirty that.
 389  */
 390 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 391                         struct splice_desc *sd)
 392 {
 393         struct file *file = sd->file;
 394         struct address_space *mapping = file->f_mapping;
 395         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 396         unsigned int offset;
 397         struct page *page;
 398         pgoff_t index;
 399         char *src;
 400         int ret;
 401
 402         /*
 403          * make sure the data in this buffer is uptodate
 404          */
 405         src = buf->ops->map(file, info, buf);
 406         if (IS_ERR(src))
 407                 return PTR_ERR(src);
 408
 409         index = sd->pos >> PAGE_CACHE_SHIFT;
 410         offset = sd->pos & ~PAGE_CACHE_MASK;
 411
 412         /*
 413          * reuse buf page, if SPLICE_F_MOVE is set
 414          */
 415         if (sd->flags & SPLICE_F_MOVE) {
 416                 /*
 417                  * If steal succeeds, buf->page is now pruned from the vm
 418                  * side (LRU and page cache) and we can reuse it.
 419                  */
 420                 if (buf->ops->steal(info, buf))
 421                         goto find_page;
 422
 423                 /*
 424                  * this will also set the page locked
 425                  */
 426                 page = buf->page;
 427                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 428                         goto find_page;
 429
 430                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 431                         lru_cache_add(page);
 432         } else {
 433 find_page:
 434                 ret = -ENOMEM;
 435                 page = find_or_create_page(mapping, index, gfp_mask);
 436                 if (!page)
 437                         goto out_nomem;
 438
 439                 /*
 440                  * If the page is uptodate, it is also locked. If it isn't
 441                  * uptodate, we can mark it uptodate if we are filling the
 442                  * full page. Otherwise we need to read it in first...
 443                  */
 444                 if (!PageUptodate(page)) {
 445                         if (sd->len < PAGE_CACHE_SIZE) {
 446                                 ret = mapping->a_ops->readpage(file, page);
 447                                 if (unlikely(ret))
 448                                         goto out;
 449
 450                                 lock_page(page);
 451
 452                                 if (!PageUptodate(page)) {
 453                                         /*
 454                                          * page got invalidated, repeat
 455                                          */
 456                                         if (!page->mapping) {
 457                                                 unlock_page(page);
 458                                                 page_cache_release(page);
 459                                                 goto find_page;
 460                                         }
 461                                         ret = -EIO;
 462                                         goto out;
 463                                 }
 464                         } else {
 465                                 WARN_ON(!PageLocked(page));
 466                                 SetPageUptodate(page);
 467                         }
 468                 }
 469         }
 470
 471         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 472         if (ret == AOP_TRUNCATED_PAGE) {
 473                 page_cache_release(page);
 474                 goto find_page;
 475         } else if (ret)
 476                 goto out;
 477
 478         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 479                 char *dst = kmap_atomic(page, KM_USER0);
 480
 481                 memcpy(dst + offset, src + buf->offset, sd->len);
 482                 flush_dcache_page(page);
 483                 kunmap_atomic(dst, KM_USER0);
 484         }
 485
 486         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 487         if (ret == AOP_TRUNCATED_PAGE) {
 488                 page_cache_release(page);
 489                 goto find_page;
 490         } else if (ret)
 491                 goto out;
 492
 493         mark_page_accessed(page);
 494         balance_dirty_pages_ratelimited(mapping);
 495 out:
 496         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 497                 page_cache_release(page);
 498                 unlock_page(page);
 499         }
 500 out_nomem:
 501         buf->ops->unmap(info, buf);
 502         return ret;
 503 }
 504
 505 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 506                            struct splice_desc *);
 507
 508 /*
 509  * Pipe input worker. Most of this logic works like a regular pipe, the
 510  * key here is the 'actor' worker passed in that actually moves the data
 511  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 512  */
 513 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 514                               size_t len, unsigned int flags,
 515                               splice_actor *actor)
 516 {
 517         int ret, do_wakeup, err;
 518         struct splice_desc sd;
 519
 520         ret = 0;
 521         do_wakeup = 0;
 522
 523         sd.total_len = len;
 524         sd.flags = flags;
 525         sd.file = out;
 526         sd.pos = out->f_pos;
 527
 528         if (pipe->inode)
 529                 mutex_lock(&pipe->inode->i_mutex);
 530
 531         for (;;) {
 532                 int bufs = pipe->nrbufs;
 533
 534                 if (bufs) {
 535                         int curbuf = pipe->curbuf;
 536                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 537                         struct pipe_buf_operations *ops = buf->ops;
 538
 539                         sd.len = buf->len;
 540                         if (sd.len > sd.total_len)
 541                                 sd.len = sd.total_len;
 542
 543                         err = actor(pipe, buf, &sd);
 544                         if (err) {
 545                                 if (!ret && err != -ENODATA)
 546                                         ret = err;
 547
 548                                 break;
 549                         }
 550
 551                         ret += sd.len;
 552                         buf->offset += sd.len;
 553                         buf->len -= sd.len;
 554                         if (!buf->len) {
 555                                 buf->ops = NULL;
 556                                 ops->release(pipe, buf);
 557                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
 558                                 pipe->curbuf = curbuf;
 559                                 pipe->nrbufs = --bufs;
 560                                 do_wakeup = 1;
 561                         }
 562
 563                         sd.pos += sd.len;
 564                         sd.total_len -= sd.len;
 565                         if (!sd.total_len)
 566                                 break;
 567                 }
 568
 569                 if (bufs)
 570                         continue;
 571                 if (!pipe->writers)
 572                         break;
 573                 if (!pipe->waiting_writers) {
 574                         if (ret)
 575                                 break;
 576                 }
 577
 578                 if (flags & SPLICE_F_NONBLOCK) {
 579                         if (!ret)
 580                                 ret = -EAGAIN;
 581                         break;
 582                 }
 583
 584                 if (signal_pending(current)) {
 585                         if (!ret)
 586                                 ret = -ERESTARTSYS;
 587                         break;
 588                 }
 589
 590                 if (do_wakeup) {
 591                         smp_mb();
 592                         if (waitqueue_active(&pipe->wait))
 593                                 wake_up_interruptible_sync(&pipe->wait);
 594                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 595                         do_wakeup = 0;
 596                 }
 597
 598                 pipe_wait(pipe);
 599         }
 600
 601         if (pipe->inode)
 602                 mutex_unlock(&pipe->inode->i_mutex);
 603
 604         if (do_wakeup) {
 605                 smp_mb();
 606                 if (waitqueue_active(&pipe->wait))
 607                         wake_up_interruptible(&pipe->wait);
 608                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 609         }
 610
 611         mutex_lock(&out->f_mapping->host->i_mutex);
 612         out->f_pos = sd.pos;
 613         mutex_unlock(&out->f_mapping->host->i_mutex);
 614         return ret;
 615
 616 }
 617
 618 /**
 619  * generic_file_splice_write - splice data from a pipe to a file
 620  * @pipe:       pipe info
 621  * @out:        file to write to
 622  * @len:        number of bytes to splice
 623  * @flags:      splice modifier flags
 624  *
 625  * Will either move or copy pages (determined by @flags options) from
 626  * the given pipe inode to the given file.
 627  *
 628  */
 629 ssize_t
 630 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 631                           size_t len, unsigned int flags)
 632 {
 633         struct address_space *mapping = out->f_mapping;
 634         ssize_t ret;
 635
 636         ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
 637
 638         /*
 639          * if file or inode is SYNC and we actually wrote some data, sync it
 640          */
 641         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 642             && ret > 0) {
 643                 struct inode *inode = mapping->host;
 644                 int err;
 645
 646                 mutex_lock(&inode->i_mutex);
 647                 err = generic_osync_inode(mapping->host, mapping,
 648                                                 OSYNC_METADATA|OSYNC_DATA);
 649                 mutex_unlock(&inode->i_mutex);
 650
 651                 if (err)
 652                         ret = err;
 653         }
 654
 655         return ret;
 656 }
 657
 658 EXPORT_SYMBOL(generic_file_splice_write);
 659
 660 /**
 661  * generic_splice_sendpage - splice data from a pipe to a socket
 662  * @inode:      pipe inode
 663  * @out:        socket to write to
 664  * @len:        number of bytes to splice
 665  * @flags:      splice modifier flags
 666  *
 667  * Will send @len bytes from the pipe to a network socket. No data copying
 668  * is involved.
 669  *
 670  */
 671 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 672                                 size_t len, unsigned int flags)
 673 {
 674         return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
 675 }
 676
 677 EXPORT_SYMBOL(generic_splice_sendpage);
 678
 679 /*
 680  * Attempt to initiate a splice from pipe to file.
 681  */
 682 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 683                            loff_t __user *off_out, size_t len,
 684                            unsigned int flags)
 685 {
 686         loff_t pos;
 687         int ret;
 688
 689         if (!out->f_op || !out->f_op->splice_write)
 690                 return -EINVAL;
 691
 692         if (!(out->f_mode & FMODE_WRITE))
 693                 return -EBADF;
 694
 695         if (off_out && copy_from_user(&out->f_pos, off_out, sizeof(loff_t)))
 696                 return -EFAULT;
 697
 698         pos = out->f_pos;
 699
 700         ret = rw_verify_area(WRITE, out, &pos, len);
 701         if (unlikely(ret < 0))
 702                 return ret;
 703
 704         return out->f_op->splice_write(pipe, out, len, flags);
 705 }
 706
 707 /*
 708  * Attempt to initiate a splice from a file to a pipe.
 709  */
 710 static long do_splice_to(struct file *in, loff_t __user *off_in,
 711                          struct pipe_inode_info *pipe, size_t len,
 712                          unsigned int flags)
 713 {
 714         loff_t pos, isize, left;
 715         int ret;
 716
 717         if (!in->f_op || !in->f_op->splice_read)
 718                 return -EINVAL;
 719
 720         if (!(in->f_mode & FMODE_READ))
 721                 return -EBADF;
 722
 723         if (off_in && copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
 724                 return -EFAULT;
 725
 726         pos = in->f_pos;
 727
 728         ret = rw_verify_area(READ, in, &pos, len);
 729         if (unlikely(ret < 0))
 730                 return ret;
 731
 732         isize = i_size_read(in->f_mapping->host);
 733         if (unlikely(in->f_pos >= isize))
 734                 return 0;
 735
 736         left = isize - in->f_pos;
 737         if (left < len)
 738                 len = left;
 739
 740         return in->f_op->splice_read(in, pipe, len, flags);
 741 }
 742
 743 /*
 744  * Determine where to splice to/from.
 745  */
 746 static long do_splice(struct file *in, loff_t __user *off_in,
 747                       struct file *out, loff_t __user *off_out,
 748                       size_t len, unsigned int flags)
 749 {
 750         struct pipe_inode_info *pipe;
 751
 752         if (off_out && out->f_op->llseek == no_llseek)
 753                 return -EINVAL;
 754         if (off_in && in->f_op->llseek == no_llseek)
 755                 return -EINVAL;
 756
 757         pipe = in->f_dentry->d_inode->i_pipe;
 758         if (pipe) {
 759                 if (off_in)
 760                         return -ESPIPE;
 761
 762                 return do_splice_from(pipe, out, off_out, len, flags);
 763         }
 764
 765         pipe = out->f_dentry->d_inode->i_pipe;
 766         if (pipe) {
 767                 if (off_out)
 768                         return -ESPIPE;
 769
 770                 return do_splice_to(in, off_in, pipe, len, flags);
 771         }
 772
 773         return -EINVAL;
 774 }
 775
 776 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 777                            int fd_out, loff_t __user *off_out,
 778                            size_t len, unsigned int flags)
 779 {
 780         long error;
 781         struct file *in, *out;
 782         int fput_in, fput_out;
 783
 784         if (unlikely(!len))
 785                 return 0;
 786
 787         error = -EBADF;
 788         in = fget_light(fd_in, &fput_in);
 789         if (in) {
 790                 if (in->f_mode & FMODE_READ) {
 791                         out = fget_light(fd_out, &fput_out);
 792                         if (out) {
 793                                 if (out->f_mode & FMODE_WRITE)
 794                                         error = do_splice(in, off_in,
 795                                                           out, off_out,
 796                                                           len, flags);
 797                                 fput_light(out, fput_out);
 798                         }
 799                 }
 800
 801                 fput_light(in, fput_in);
 802         }
 803
 804         return error;
 805 }