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