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, error;
 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         error = 0;
 264         for (i = 0; i < nr_pages; i++, index++) {
 265 find_page:
 266                 /*
 267                  * lookup the page for this index
 268                  */
 269                 page = find_get_page(mapping, index);
 270                 if (!page) {
 271                         /*
 272                          * If in nonblock mode then dont block on
 273                          * readpage (we've kicked readahead so there
 274                          * will be asynchronous progress):
 275                          */
 276                         if (flags & SPLICE_F_NONBLOCK)
 277                                 break;
 278
 279                         /*
 280                          * page didn't exist, allocate one
 281                          */
 282                         page = page_cache_alloc_cold(mapping);
 283                         if (!page)
 284                                 break;
 285
 286                         error = add_to_page_cache_lru(page, mapping, index,
 287                                                 mapping_gfp_mask(mapping));
 288                         if (unlikely(error)) {
 289                                 page_cache_release(page);
 290                                 break;
 291                         }
 292
 293                         goto readpage;
 294                 }
 295
 296                 /*
 297                  * If the page isn't uptodate, we may need to start io on it
 298                  */
 299                 if (!PageUptodate(page)) {
 300                         lock_page(page);
 301
 302                         /*
 303                          * page was truncated, stop here. if this isn't the
 304                          * first page, we'll just complete what we already
 305                          * added
 306                          */
 307                         if (!page->mapping) {
 308                                 unlock_page(page);
 309                                 page_cache_release(page);
 310                                 break;
 311                         }
 312                         /*
 313                          * page was already under io and is now done, great
 314                          */
 315                         if (PageUptodate(page)) {
 316                                 unlock_page(page);
 317                                 goto fill_it;
 318                         }
 319
 320 readpage:
 321                         /*
 322                          * need to read in the page
 323                          */
 324                         error = mapping->a_ops->readpage(in, page);
 325
 326                         if (unlikely(error)) {
 327                                 page_cache_release(page);
 328                                 if (error == AOP_TRUNCATED_PAGE)
 329                                         goto find_page;
 330                                 break;
 331                         }
 332                 }
 333 fill_it:
 334                 pages[i] = page;
 335         }
 336
 337         if (i)
 338                 return move_to_pipe(pipe, pages, i, offset, len, flags);
 339
 340         return error;
 341 }
 342
 343 /**
 344  * generic_file_splice_read - splice data from file to a pipe
 345  * @in:         file to splice from
 346  * @pipe:       pipe to splice to
 347  * @len:        number of bytes to splice
 348  * @flags:      splice modifier flags
 349  *
 350  * Will read pages from given file and fill them into a pipe.
 351  *
 352  */
 353 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 354                                  size_t len, unsigned int flags)
 355 {
 356         ssize_t spliced;
 357         int ret;
 358
 359         ret = 0;
 360         spliced = 0;
 361
 362         while (len) {
 363                 ret = __generic_file_splice_read(in, pipe, len, flags);
 364
 365                 if (ret <= 0)
 366                         break;
 367
 368                 in->f_pos += ret;
 369                 len -= ret;
 370                 spliced += ret;
 371
 372                 if (!(flags & SPLICE_F_NONBLOCK))
 373                         continue;
 374                 ret = -EAGAIN;
 375                 break;
 376         }
 377
 378         if (spliced)
 379                 return spliced;
 380
 381         return ret;
 382 }
 383
 384 EXPORT_SYMBOL(generic_file_splice_read);
 385
 386 /*
 387  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 388  * using sendpage().
 389  */
 390 static int pipe_to_sendpage(struct pipe_inode_info *info,
 391                             struct pipe_buffer *buf, struct splice_desc *sd)
 392 {
 393         struct file *file = sd->file;
 394         loff_t pos = sd->pos;
 395         unsigned int offset;
 396         ssize_t ret;
 397         void *ptr;
 398         int more;
 399
 400         /*
 401          * sub-optimal, but we are limited by the pipe ->map. we don't
 402          * need a kmap'ed buffer here, we just want to make sure we
 403          * have the page pinned if the pipe page originates from the
 404          * page cache
 405          */
 406         ptr = buf->ops->map(file, info, buf);
 407         if (IS_ERR(ptr))
 408                 return PTR_ERR(ptr);
 409
 410         offset = pos & ~PAGE_CACHE_MASK;
 411         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 412
 413         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 414
 415         buf->ops->unmap(info, buf);
 416         if (ret == sd->len)
 417                 return 0;
 418
 419         return -EIO;
 420 }
 421
 422 /*
 423  * This is a little more tricky than the file -> pipe splicing. There are
 424  * basically three cases:
 425  *
 426  *      - Destination page already exists in the address space and there
 427  *        are users of it. For that case we have no other option that
 428  *        copying the data. Tough luck.
 429  *      - Destination page already exists in the address space, but there
 430  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 431  *        through to last case.
 432  *      - Destination page does not exist, we can add the pipe page to
 433  *        the page cache and avoid the copy.
 434  *
 435  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 436  * sd->flags), we attempt to migrate pages from the pipe to the output
 437  * file address space page cache. This is possible if no one else has
 438  * the pipe page referenced outside of the pipe and page cache. If
 439  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 440  * a new page in the output file page cache and fill/dirty that.
 441  */
 442 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 443                         struct splice_desc *sd)
 444 {
 445         struct file *file = sd->file;
 446         struct address_space *mapping = file->f_mapping;
 447         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 448         unsigned int offset;
 449         struct page *page;
 450         pgoff_t index;
 451         char *src;
 452         int ret;
 453
 454         /*
 455          * make sure the data in this buffer is uptodate
 456          */
 457         src = buf->ops->map(file, info, buf);
 458         if (IS_ERR(src))
 459                 return PTR_ERR(src);
 460
 461         index = sd->pos >> PAGE_CACHE_SHIFT;
 462         offset = sd->pos & ~PAGE_CACHE_MASK;
 463
 464         /*
 465          * reuse buf page, if SPLICE_F_MOVE is set
 466          */
 467         if (sd->flags & SPLICE_F_MOVE) {
 468                 /*
 469                  * If steal succeeds, buf->page is now pruned from the vm
 470                  * side (LRU and page cache) and we can reuse it.
 471                  */
 472                 if (buf->ops->steal(info, buf))
 473                         goto find_page;
 474
 475                 /*
 476                  * this will also set the page locked
 477                  */
 478                 page = buf->page;
 479                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 480                         goto find_page;
 481
 482                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 483                         lru_cache_add(page);
 484         } else {
 485 find_page:
 486                 ret = -ENOMEM;
 487                 page = find_or_create_page(mapping, index, gfp_mask);
 488                 if (!page)
 489                         goto out_nomem;
 490
 491                 /*
 492                  * If the page is uptodate, it is also locked. If it isn't
 493                  * uptodate, we can mark it uptodate if we are filling the
 494                  * full page. Otherwise we need to read it in first...
 495                  */
 496                 if (!PageUptodate(page)) {
 497                         if (sd->len < PAGE_CACHE_SIZE) {
 498                                 ret = mapping->a_ops->readpage(file, page);
 499                                 if (unlikely(ret))
 500                                         goto out;
 501
 502                                 lock_page(page);
 503
 504                                 if (!PageUptodate(page)) {
 505                                         /*
 506                                          * page got invalidated, repeat
 507                                          */
 508                                         if (!page->mapping) {
 509                                                 unlock_page(page);
 510                                                 page_cache_release(page);
 511                                                 goto find_page;
 512                                         }
 513                                         ret = -EIO;
 514                                         goto out;
 515                                 }
 516                         } else {
 517                                 WARN_ON(!PageLocked(page));
 518                                 SetPageUptodate(page);
 519                         }
 520                 }
 521         }
 522
 523         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 524         if (ret == AOP_TRUNCATED_PAGE) {
 525                 page_cache_release(page);
 526                 goto find_page;
 527         } else if (ret)
 528                 goto out;
 529
 530         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 531                 char *dst = kmap_atomic(page, KM_USER0);
 532
 533                 memcpy(dst + offset, src + buf->offset, sd->len);
 534                 flush_dcache_page(page);
 535                 kunmap_atomic(dst, KM_USER0);
 536         }
 537
 538         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 539         if (ret == AOP_TRUNCATED_PAGE) {
 540                 page_cache_release(page);
 541                 goto find_page;
 542         } else if (ret)
 543                 goto out;
 544
 545         mark_page_accessed(page);
 546         balance_dirty_pages_ratelimited(mapping);
 547 out:
 548         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 549                 page_cache_release(page);
 550                 unlock_page(page);
 551         }
 552 out_nomem:
 553         buf->ops->unmap(info, buf);
 554         return ret;
 555 }
 556
 557 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 558                            struct splice_desc *);
 559
 560 /*
 561  * Pipe input worker. Most of this logic works like a regular pipe, the
 562  * key here is the 'actor' worker passed in that actually moves the data
 563  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 564  */
 565 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 566                               size_t len, unsigned int flags,
 567                               splice_actor *actor)
 568 {
 569         int ret, do_wakeup, err;
 570         struct splice_desc sd;
 571
 572         ret = 0;
 573         do_wakeup = 0;
 574
 575         sd.total_len = len;
 576         sd.flags = flags;
 577         sd.file = out;
 578         sd.pos = out->f_pos;
 579
 580         if (pipe->inode)
 581                 mutex_lock(&pipe->inode->i_mutex);
 582
 583         for (;;) {
 584                 int bufs = pipe->nrbufs;
 585
 586                 if (bufs) {
 587                         int curbuf = pipe->curbuf;
 588                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 589                         struct pipe_buf_operations *ops = buf->ops;
 590
 591                         sd.len = buf->len;
 592                         if (sd.len > sd.total_len)
 593                                 sd.len = sd.total_len;
 594
 595                         err = actor(pipe, buf, &sd);
 596                         if (err) {
 597                                 if (!ret && err != -ENODATA)
 598                                         ret = err;
 599
 600                                 break;
 601                         }
 602
 603                         ret += sd.len;
 604                         buf->offset += sd.len;
 605                         buf->len -= sd.len;
 606                         if (!buf->len) {
 607                                 buf->ops = NULL;
 608                                 ops->release(pipe, buf);
 609                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
 610                                 pipe->curbuf = curbuf;
 611                                 pipe->nrbufs = --bufs;
 612                                 do_wakeup = 1;
 613                         }
 614
 615                         sd.pos += sd.len;
 616                         sd.total_len -= sd.len;
 617                         if (!sd.total_len)
 618                                 break;
 619                 }
 620
 621                 if (bufs)
 622                         continue;
 623                 if (!pipe->writers)
 624                         break;
 625                 if (!pipe->waiting_writers) {
 626                         if (ret)
 627                                 break;
 628                 }
 629
 630                 if (flags & SPLICE_F_NONBLOCK) {
 631                         if (!ret)
 632                                 ret = -EAGAIN;
 633                         break;
 634                 }
 635
 636                 if (signal_pending(current)) {
 637                         if (!ret)
 638                                 ret = -ERESTARTSYS;
 639                         break;
 640                 }
 641
 642                 if (do_wakeup) {
 643                         smp_mb();
 644                         if (waitqueue_active(&pipe->wait))
 645                                 wake_up_interruptible_sync(&pipe->wait);
 646                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 647                         do_wakeup = 0;
 648                 }
 649
 650                 pipe_wait(pipe);
 651         }
 652
 653         if (pipe->inode)
 654                 mutex_unlock(&pipe->inode->i_mutex);
 655
 656         if (do_wakeup) {
 657                 smp_mb();
 658                 if (waitqueue_active(&pipe->wait))
 659                         wake_up_interruptible(&pipe->wait);
 660                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 661         }
 662
 663         mutex_lock(&out->f_mapping->host->i_mutex);
 664         out->f_pos = sd.pos;
 665         mutex_unlock(&out->f_mapping->host->i_mutex);
 666         return ret;
 667
 668 }
 669
 670 /**
 671  * generic_file_splice_write - splice data from a pipe to a file
 672  * @pipe:       pipe info
 673  * @out:        file to write to
 674  * @len:        number of bytes to splice
 675  * @flags:      splice modifier flags
 676  *
 677  * Will either move or copy pages (determined by @flags options) from
 678  * the given pipe inode to the given file.
 679  *
 680  */
 681 ssize_t
 682 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 683                           size_t len, unsigned int flags)
 684 {
 685         struct address_space *mapping = out->f_mapping;
 686         ssize_t ret;
 687
 688         ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
 689
 690         /*
 691          * if file or inode is SYNC and we actually wrote some data, sync it
 692          */
 693         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 694             && ret > 0) {
 695                 struct inode *inode = mapping->host;
 696                 int err;
 697
 698                 mutex_lock(&inode->i_mutex);
 699                 err = generic_osync_inode(mapping->host, mapping,
 700                                                 OSYNC_METADATA|OSYNC_DATA);
 701                 mutex_unlock(&inode->i_mutex);
 702
 703                 if (err)
 704                         ret = err;
 705         }
 706
 707         return ret;
 708 }
 709
 710 EXPORT_SYMBOL(generic_file_splice_write);
 711
 712 /**
 713  * generic_splice_sendpage - splice data from a pipe to a socket
 714  * @inode:      pipe inode
 715  * @out:        socket to write to
 716  * @len:        number of bytes to splice
 717  * @flags:      splice modifier flags
 718  *
 719  * Will send @len bytes from the pipe to a network socket. No data copying
 720  * is involved.
 721  *
 722  */
 723 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 724                                 size_t len, unsigned int flags)
 725 {
 726         return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
 727 }
 728
 729 EXPORT_SYMBOL(generic_splice_sendpage);
 730
 731 /*
 732  * Attempt to initiate a splice from pipe to file.
 733  */
 734 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 735                            size_t len, unsigned int flags)
 736 {
 737         loff_t pos;
 738         int ret;
 739
 740         if (!out->f_op || !out->f_op->splice_write)
 741                 return -EINVAL;
 742
 743         if (!(out->f_mode & FMODE_WRITE))
 744                 return -EBADF;
 745
 746         pos = out->f_pos;
 747
 748         ret = rw_verify_area(WRITE, out, &pos, len);
 749         if (unlikely(ret < 0))
 750                 return ret;
 751
 752         return out->f_op->splice_write(pipe, out, len, flags);
 753 }
 754
 755 /*
 756  * Attempt to initiate a splice from a file to a pipe.
 757  */
 758 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
 759                          size_t len, unsigned int flags)
 760 {
 761         loff_t pos, isize, left;
 762         int ret;
 763
 764         if (!in->f_op || !in->f_op->splice_read)
 765                 return -EINVAL;
 766
 767         if (!(in->f_mode & FMODE_READ))
 768                 return -EBADF;
 769
 770         pos = in->f_pos;
 771
 772         ret = rw_verify_area(READ, in, &pos, len);
 773         if (unlikely(ret < 0))
 774                 return ret;
 775
 776         isize = i_size_read(in->f_mapping->host);
 777         if (unlikely(in->f_pos >= isize))
 778                 return 0;
 779
 780         left = isize - in->f_pos;
 781         if (left < len)
 782                 len = left;
 783
 784         return in->f_op->splice_read(in, pipe, len, flags);
 785 }
 786
 787 long do_splice_direct(struct file *in, struct file *out, size_t len,
 788                       unsigned int flags)
 789 {
 790         struct pipe_inode_info *pipe;
 791         long ret, bytes;
 792         umode_t i_mode;
 793         int i;
 794
 795         /*
 796          * We require the input being a regular file, as we don't want to
 797          * randomly drop data for eg socket -> socket splicing. Use the
 798          * piped splicing for that!
 799          */
 800         i_mode = in->f_dentry->d_inode->i_mode;
 801         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 802                 return -EINVAL;
 803
 804         /*
 805          * neither in nor out is a pipe, setup an internal pipe attached to
 806          * 'out' and transfer the wanted data from 'in' to 'out' through that
 807          */
 808         pipe = current->splice_pipe;
 809         if (!pipe) {
 810                 pipe = alloc_pipe_info(NULL);
 811                 if (!pipe)
 812                         return -ENOMEM;
 813
 814                 /*
 815                  * We don't have an immediate reader, but we'll read the stuff
 816                  * out of the pipe right after the move_to_pipe(). So set
 817                  * PIPE_READERS appropriately.
 818                  */
 819                 pipe->readers = 1;
 820
 821                 current->splice_pipe = pipe;
 822         }
 823
 824         /*
 825          * do the splice
 826          */
 827         ret = 0;
 828         bytes = 0;
 829
 830         while (len) {
 831                 size_t read_len, max_read_len;
 832
 833                 /*
 834                  * Do at most PIPE_BUFFERS pages worth of transfer:
 835                  */
 836                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 837
 838                 ret = do_splice_to(in, pipe, max_read_len, flags);
 839                 if (unlikely(ret < 0))
 840                         goto out_release;
 841
 842                 read_len = ret;
 843
 844                 /*
 845                  * NOTE: nonblocking mode only applies to the input. We
 846                  * must not do the output in nonblocking mode as then we
 847                  * could get stuck data in the internal pipe:
 848                  */
 849                 ret = do_splice_from(pipe, out, read_len,
 850                                      flags & ~SPLICE_F_NONBLOCK);
 851                 if (unlikely(ret < 0))
 852                         goto out_release;
 853
 854                 bytes += ret;
 855                 len -= ret;
 856
 857                 /*
 858                  * In nonblocking mode, if we got back a short read then
 859                  * that was due to either an IO error or due to the
 860                  * pagecache entry not being there. In the IO error case
 861                  * the _next_ splice attempt will produce a clean IO error
 862                  * return value (not a short read), so in both cases it's
 863                  * correct to break out of the loop here:
 864                  */
 865                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 866                         break;
 867         }
 868
 869         pipe->nrbufs = pipe->curbuf = 0;
 870
 871         return bytes;
 872
 873 out_release:
 874         /*
 875          * If we did an incomplete transfer we must release
 876          * the pipe buffers in question:
 877          */
 878         for (i = 0; i < PIPE_BUFFERS; i++) {
 879                 struct pipe_buffer *buf = pipe->bufs + i;
 880
 881                 if (buf->ops) {
 882                         buf->ops->release(pipe, buf);
 883                         buf->ops = NULL;
 884                 }
 885         }
 886         pipe->nrbufs = pipe->curbuf = 0;
 887
 888         /*
 889          * If we transferred some data, return the number of bytes:
 890          */
 891         if (bytes > 0)
 892                 return bytes;
 893
 894         return ret;
 895 }
 896
 897 EXPORT_SYMBOL(do_splice_direct);
 898
 899 /*
 900  * Determine where to splice to/from.
 901  */
 902 static long do_splice(struct file *in, loff_t __user *off_in,
 903                       struct file *out, loff_t __user *off_out,
 904                       size_t len, unsigned int flags)
 905 {
 906         struct pipe_inode_info *pipe;
 907
 908         pipe = in->f_dentry->d_inode->i_pipe;
 909         if (pipe) {
 910                 if (off_in)
 911                         return -ESPIPE;
 912                 if (off_out) {
 913                         if (out->f_op->llseek == no_llseek)
 914                                 return -EINVAL;
 915                         if (copy_from_user(&out->f_pos, off_out,
 916                                            sizeof(loff_t)))
 917                                 return -EFAULT;
 918                 }
 919
 920                 return do_splice_from(pipe, out, len, flags);
 921         }
 922
 923         pipe = out->f_dentry->d_inode->i_pipe;
 924         if (pipe) {
 925                 if (off_out)
 926                         return -ESPIPE;
 927                 if (off_in) {
 928                         if (in->f_op->llseek == no_llseek)
 929                                 return -EINVAL;
 930                         if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
 931                                 return -EFAULT;
 932                 }
 933
 934                 return do_splice_to(in, pipe, len, flags);
 935         }
 936
 937         return -EINVAL;
 938 }
 939
 940 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 941                            int fd_out, loff_t __user *off_out,
 942                            size_t len, unsigned int flags)
 943 {
 944         long error;
 945         struct file *in, *out;
 946         int fput_in, fput_out;
 947
 948         if (unlikely(!len))
 949                 return 0;
 950
 951         error = -EBADF;
 952         in = fget_light(fd_in, &fput_in);
 953         if (in) {
 954                 if (in->f_mode & FMODE_READ) {
 955                         out = fget_light(fd_out, &fput_out);
 956                         if (out) {
 957                                 if (out->f_mode & FMODE_WRITE)
 958                                         error = do_splice(in, off_in,
 959                                                           out, off_out,
 960                                                           len, flags);
 961                                 fput_light(out, fput_out);
 962                         }
 963                 }
 964
 965                 fput_light(in, fput_in);
 966         }
 967
 968         return error;
 969 }