err.no Git - linux-2.6/blob - fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@suse.de>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/pipe_fs_i.h>
  24 #include <linux/mm_inline.h>
  25 #include <linux/swap.h>
  26 #include <linux/writeback.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/module.h>
  29 #include <linux/syscalls.h>
  30
  31 /*
  32  * Passed to the actors
  33  */
  34 struct splice_desc {
  35         unsigned int len, total_len;    /* current and remaining length */
  36         unsigned int flags;             /* splice flags */
  37         struct file *file;              /* file to read/write */
  38         loff_t pos;                     /* file position */
  39 };
  40
  41 /*
  42  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  43  * a vm helper function, it's already simplified quite a bit by the
  44  * addition of remove_mapping(). If success is returned, the caller may
  45  * attempt to reuse this page for another destination.
  46  */
  47 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  48                                      struct pipe_buffer *buf)
  49 {
  50         struct page *page = buf->page;
  51         struct address_space *mapping = page_mapping(page);
  52
  53         WARN_ON(!PageLocked(page));
  54         WARN_ON(!PageUptodate(page));
  55
  56         /*
  57          * At least for ext2 with nobh option, we need to wait on writeback
  58          * completing on this page, since we'll remove it from the pagecache.
  59          * Otherwise truncate wont wait on the page, allowing the disk
  60          * blocks to be reused by someone else before we actually wrote our
  61          * data to them. fs corruption ensues.
  62          */
  63         wait_on_page_writeback(page);
  64
  65         if (PagePrivate(page))
  66                 try_to_release_page(page, mapping_gfp_mask(mapping));
  67
  68         if (!remove_mapping(mapping, page))
  69                 return 1;
  70
  71         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
  72         return 0;
  73 }
  74
  75 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  76                                         struct pipe_buffer *buf)
  77 {
  78         page_cache_release(buf->page);
  79         buf->page = NULL;
  80         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
  81 }
  82
  83 static void *page_cache_pipe_buf_map(struct file *file,
  84                                      struct pipe_inode_info *info,
  85                                      struct pipe_buffer *buf)
  86 {
  87         struct page *page = buf->page;
  88         int err;
  89
  90         if (!PageUptodate(page)) {
  91                 lock_page(page);
  92
  93                 /*
  94                  * Page got truncated/unhashed. This will cause a 0-byte
  95                  * splice, if this is the first page.
  96                  */
  97                 if (!page->mapping) {
  98                         err = -ENODATA;
  99                         goto error;
 100                 }
 101
 102                 /*
 103                  * Uh oh, read-error from disk.
 104                  */
 105                 if (!PageUptodate(page)) {
 106                         err = -EIO;
 107                         goto error;
 108                 }
 109
 110                 /*
 111                  * Page is ok afterall, fall through to mapping.
 112                  */
 113                 unlock_page(page);
 114         }
 115
 116         return kmap(page);
 117 error:
 118         unlock_page(page);
 119         return ERR_PTR(err);
 120 }
 121
 122 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
 123                                       struct pipe_buffer *buf)
 124 {
 125         kunmap(buf->page);
 126 }
 127
 128 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 129         .can_merge = 0,
 130         .map = page_cache_pipe_buf_map,
 131         .unmap = page_cache_pipe_buf_unmap,
 132         .release = page_cache_pipe_buf_release,
 133         .steal = page_cache_pipe_buf_steal,
 134 };
 135
 136 /*
 137  * Pipe output worker. This sets up our pipe format with the page cache
 138  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 139  */
 140 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
 141                             int nr_pages, unsigned long offset,
 142                             unsigned long len, unsigned int flags)
 143 {
 144         int ret, do_wakeup, i;
 145
 146         ret = 0;
 147         do_wakeup = 0;
 148         i = 0;
 149
 150         if (pipe->inode)
 151                 mutex_lock(&pipe->inode->i_mutex);
 152
 153         for (;;) {
 154                 if (!pipe->readers) {
 155                         send_sig(SIGPIPE, current, 0);
 156                         if (!ret)
 157                                 ret = -EPIPE;
 158                         break;
 159                 }
 160
 161                 if (pipe->nrbufs < PIPE_BUFFERS) {
 162                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 163                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 164                         struct page *page = pages[i++];
 165                         unsigned long this_len;
 166
 167                         this_len = PAGE_CACHE_SIZE - offset;
 168                         if (this_len > len)
 169                                 this_len = len;
 170
 171                         buf->page = page;
 172                         buf->offset = offset;
 173                         buf->len = this_len;
 174                         buf->ops = &page_cache_pipe_buf_ops;
 175                         pipe->nrbufs++;
 176                         if (pipe->inode)
 177                                 do_wakeup = 1;
 178
 179                         ret += this_len;
 180                         len -= this_len;
 181                         offset = 0;
 182                         if (!--nr_pages)
 183                                 break;
 184                         if (!len)
 185                                 break;
 186                         if (pipe->nrbufs < PIPE_BUFFERS)
 187                                 continue;
 188
 189                         break;
 190                 }
 191
 192                 if (flags & SPLICE_F_NONBLOCK) {
 193                         if (!ret)
 194                                 ret = -EAGAIN;
 195                         break;
 196                 }
 197
 198                 if (signal_pending(current)) {
 199                         if (!ret)
 200                                 ret = -ERESTARTSYS;
 201                         break;
 202                 }
 203
 204                 if (do_wakeup) {
 205                         smp_mb();
 206                         if (waitqueue_active(&pipe->wait))
 207                                 wake_up_interruptible_sync(&pipe->wait);
 208                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 209                         do_wakeup = 0;
 210                 }
 211
 212                 pipe->waiting_writers++;
 213                 pipe_wait(pipe);
 214                 pipe->waiting_writers--;
 215         }
 216
 217         if (pipe->inode)
 218                 mutex_unlock(&pipe->inode->i_mutex);
 219
 220         if (do_wakeup) {
 221                 smp_mb();
 222                 if (waitqueue_active(&pipe->wait))
 223                         wake_up_interruptible(&pipe->wait);
 224                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 225         }
 226
 227         while (i < nr_pages)
 228                 page_cache_release(pages[i++]);
 229
 230         return ret;
 231 }
 232
 233 static int
 234 __generic_file_splice_read(struct file *in, loff_t *ppos,
 235                            struct pipe_inode_info *pipe, size_t len,
 236                            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 = *ppos >> PAGE_CACHE_SHIFT;
 246         offset = *ppos & ~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 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 353                                  struct pipe_inode_info *pipe, size_t len,
 354                                  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, ppos, pipe, len, flags);
 364
 365                 if (ret <= 0)
 366                         break;
 367
 368                 *ppos += 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                               loff_t *ppos, 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 = *ppos;
 579
 580         if (pipe->inode)
 581                 mutex_lock(&pipe->inode->i_mutex);
 582
 583         for (;;) {
 584                 if (pipe->nrbufs) {
 585                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 586                         struct pipe_buf_operations *ops = buf->ops;
 587
 588                         sd.len = buf->len;
 589                         if (sd.len > sd.total_len)
 590                                 sd.len = sd.total_len;
 591
 592                         err = actor(pipe, buf, &sd);
 593                         if (err) {
 594                                 if (!ret && err != -ENODATA)
 595                                         ret = err;
 596
 597                                 break;
 598                         }
 599
 600                         ret += sd.len;
 601                         buf->offset += sd.len;
 602                         buf->len -= sd.len;
 603
 604                         if (!buf->len) {
 605                                 buf->ops = NULL;
 606                                 ops->release(pipe, buf);
 607                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 608                                 pipe->nrbufs--;
 609                                 if (pipe->inode)
 610                                         do_wakeup = 1;
 611                         }
 612
 613                         sd.pos += sd.len;
 614                         sd.total_len -= sd.len;
 615                         if (!sd.total_len)
 616                                 break;
 617                 }
 618
 619                 if (pipe->nrbufs)
 620                         continue;
 621                 if (!pipe->writers)
 622                         break;
 623                 if (!pipe->waiting_writers) {
 624                         if (ret)
 625                                 break;
 626                 }
 627
 628                 if (flags & SPLICE_F_NONBLOCK) {
 629                         if (!ret)
 630                                 ret = -EAGAIN;
 631                         break;
 632                 }
 633
 634                 if (signal_pending(current)) {
 635                         if (!ret)
 636                                 ret = -ERESTARTSYS;
 637                         break;
 638                 }
 639
 640                 if (do_wakeup) {
 641                         smp_mb();
 642                         if (waitqueue_active(&pipe->wait))
 643                                 wake_up_interruptible_sync(&pipe->wait);
 644                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 645                         do_wakeup = 0;
 646                 }
 647
 648                 pipe_wait(pipe);
 649         }
 650
 651         if (pipe->inode)
 652                 mutex_unlock(&pipe->inode->i_mutex);
 653
 654         if (do_wakeup) {
 655                 smp_mb();
 656                 if (waitqueue_active(&pipe->wait))
 657                         wake_up_interruptible(&pipe->wait);
 658                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 659         }
 660
 661         return ret;
 662 }
 663
 664 /**
 665  * generic_file_splice_write - splice data from a pipe to a file
 666  * @pipe:       pipe info
 667  * @out:        file to write to
 668  * @len:        number of bytes to splice
 669  * @flags:      splice modifier flags
 670  *
 671  * Will either move or copy pages (determined by @flags options) from
 672  * the given pipe inode to the given file.
 673  *
 674  */
 675 ssize_t
 676 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 677                           loff_t *ppos, size_t len, unsigned int flags)
 678 {
 679         struct address_space *mapping = out->f_mapping;
 680         ssize_t ret;
 681
 682         ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 683
 684         /*
 685          * If file or inode is SYNC and we actually wrote some data, sync it.
 686          */
 687         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 688             && ret > 0) {
 689                 struct inode *inode = mapping->host;
 690                 int err;
 691
 692                 mutex_lock(&inode->i_mutex);
 693                 err = generic_osync_inode(mapping->host, mapping,
 694                                           OSYNC_METADATA|OSYNC_DATA);
 695                 mutex_unlock(&inode->i_mutex);
 696
 697                 if (err)
 698                         ret = err;
 699         }
 700
 701         return ret;
 702 }
 703
 704 EXPORT_SYMBOL(generic_file_splice_write);
 705
 706 /**
 707  * generic_splice_sendpage - splice data from a pipe to a socket
 708  * @inode:      pipe inode
 709  * @out:        socket to write to
 710  * @len:        number of bytes to splice
 711  * @flags:      splice modifier flags
 712  *
 713  * Will send @len bytes from the pipe to a network socket. No data copying
 714  * is involved.
 715  *
 716  */
 717 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 718                                 loff_t *ppos, size_t len, unsigned int flags)
 719 {
 720         return move_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 721 }
 722
 723 EXPORT_SYMBOL(generic_splice_sendpage);
 724
 725 /*
 726  * Attempt to initiate a splice from pipe to file.
 727  */
 728 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 729                            loff_t *ppos, size_t len, unsigned int flags)
 730 {
 731         int ret;
 732
 733         if (unlikely(!out->f_op || !out->f_op->splice_write))
 734                 return -EINVAL;
 735
 736         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 737                 return -EBADF;
 738
 739         ret = rw_verify_area(WRITE, out, ppos, len);
 740         if (unlikely(ret < 0))
 741                 return ret;
 742
 743         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 744 }
 745
 746 /*
 747  * Attempt to initiate a splice from a file to a pipe.
 748  */
 749 static long do_splice_to(struct file *in, loff_t *ppos,
 750                          struct pipe_inode_info *pipe, size_t len,
 751                          unsigned int flags)
 752 {
 753         loff_t isize, left;
 754         int ret;
 755
 756         if (unlikely(!in->f_op || !in->f_op->splice_read))
 757                 return -EINVAL;
 758
 759         if (unlikely(!(in->f_mode & FMODE_READ)))
 760                 return -EBADF;
 761
 762         ret = rw_verify_area(READ, in, ppos, len);
 763         if (unlikely(ret < 0))
 764                 return ret;
 765
 766         isize = i_size_read(in->f_mapping->host);
 767         if (unlikely(*ppos >= isize))
 768                 return 0;
 769
 770         left = isize - *ppos;
 771         if (unlikely(left < len))
 772                 len = left;
 773
 774         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 775 }
 776
 777 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 778                       size_t len, unsigned int flags)
 779 {
 780         struct pipe_inode_info *pipe;
 781         long ret, bytes;
 782         loff_t out_off;
 783         umode_t i_mode;
 784         int i;
 785
 786         /*
 787          * We require the input being a regular file, as we don't want to
 788          * randomly drop data for eg socket -> socket splicing. Use the
 789          * piped splicing for that!
 790          */
 791         i_mode = in->f_dentry->d_inode->i_mode;
 792         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 793                 return -EINVAL;
 794
 795         /*
 796          * neither in nor out is a pipe, setup an internal pipe attached to
 797          * 'out' and transfer the wanted data from 'in' to 'out' through that
 798          */
 799         pipe = current->splice_pipe;
 800         if (unlikely(!pipe)) {
 801                 pipe = alloc_pipe_info(NULL);
 802                 if (!pipe)
 803                         return -ENOMEM;
 804
 805                 /*
 806                  * We don't have an immediate reader, but we'll read the stuff
 807                  * out of the pipe right after the move_to_pipe(). So set
 808                  * PIPE_READERS appropriately.
 809                  */
 810                 pipe->readers = 1;
 811
 812                 current->splice_pipe = pipe;
 813         }
 814
 815         /*
 816          * Do the splice.
 817          */
 818         ret = 0;
 819         bytes = 0;
 820         out_off = 0;
 821
 822         while (len) {
 823                 size_t read_len, max_read_len;
 824
 825                 /*
 826                  * Do at most PIPE_BUFFERS pages worth of transfer:
 827                  */
 828                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 829
 830                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 831                 if (unlikely(ret < 0))
 832                         goto out_release;
 833
 834                 read_len = ret;
 835
 836                 /*
 837                  * NOTE: nonblocking mode only applies to the input. We
 838                  * must not do the output in nonblocking mode as then we
 839                  * could get stuck data in the internal pipe:
 840                  */
 841                 ret = do_splice_from(pipe, out, &out_off, read_len,
 842                                      flags & ~SPLICE_F_NONBLOCK);
 843                 if (unlikely(ret < 0))
 844                         goto out_release;
 845
 846                 bytes += ret;
 847                 len -= ret;
 848
 849                 /*
 850                  * In nonblocking mode, if we got back a short read then
 851                  * that was due to either an IO error or due to the
 852                  * pagecache entry not being there. In the IO error case
 853                  * the _next_ splice attempt will produce a clean IO error
 854                  * return value (not a short read), so in both cases it's
 855                  * correct to break out of the loop here:
 856                  */
 857                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 858                         break;
 859         }
 860
 861         pipe->nrbufs = pipe->curbuf = 0;
 862
 863         return bytes;
 864
 865 out_release:
 866         /*
 867          * If we did an incomplete transfer we must release
 868          * the pipe buffers in question:
 869          */
 870         for (i = 0; i < PIPE_BUFFERS; i++) {
 871                 struct pipe_buffer *buf = pipe->bufs + i;
 872
 873                 if (buf->ops) {
 874                         buf->ops->release(pipe, buf);
 875                         buf->ops = NULL;
 876                 }
 877         }
 878         pipe->nrbufs = pipe->curbuf = 0;
 879
 880         /*
 881          * If we transferred some data, return the number of bytes:
 882          */
 883         if (bytes > 0)
 884                 return bytes;
 885
 886         return ret;
 887 }
 888
 889 EXPORT_SYMBOL(do_splice_direct);
 890
 891 /*
 892  * Determine where to splice to/from.
 893  */
 894 static long do_splice(struct file *in, loff_t __user *off_in,
 895                       struct file *out, loff_t __user *off_out,
 896                       size_t len, unsigned int flags)
 897 {
 898         struct pipe_inode_info *pipe;
 899         loff_t offset, *off;
 900
 901         pipe = in->f_dentry->d_inode->i_pipe;
 902         if (pipe) {
 903                 if (off_in)
 904                         return -ESPIPE;
 905                 if (off_out) {
 906                         if (out->f_op->llseek == no_llseek)
 907                                 return -EINVAL;
 908                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
 909                                 return -EFAULT;
 910                         off = &offset;
 911                 } else
 912                         off = &out->f_pos;
 913
 914                 return do_splice_from(pipe, out, off, len, flags);
 915         }
 916
 917         pipe = out->f_dentry->d_inode->i_pipe;
 918         if (pipe) {
 919                 if (off_out)
 920                         return -ESPIPE;
 921                 if (off_in) {
 922                         if (in->f_op->llseek == no_llseek)
 923                                 return -EINVAL;
 924                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
 925                                 return -EFAULT;
 926                         off = &offset;
 927                 } else
 928                         off = &in->f_pos;
 929
 930                 return do_splice_to(in, off, pipe, len, flags);
 931         }
 932
 933         return -EINVAL;
 934 }
 935
 936 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 937                            int fd_out, loff_t __user *off_out,
 938                            size_t len, unsigned int flags)
 939 {
 940         long error;
 941         struct file *in, *out;
 942         int fput_in, fput_out;
 943
 944         if (unlikely(!len))
 945                 return 0;
 946
 947         error = -EBADF;
 948         in = fget_light(fd_in, &fput_in);
 949         if (in) {
 950                 if (in->f_mode & FMODE_READ) {
 951                         out = fget_light(fd_out, &fput_out);
 952                         if (out) {
 953                                 if (out->f_mode & FMODE_WRITE)
 954                                         error = do_splice(in, off_in,
 955                                                           out, off_out,
 956                                                           len, flags);
 957                                 fput_light(out, fput_out);
 958                         }
 959                 }
 960
 961                 fput_light(in, fput_in);
 962         }
 963
 964         return error;
 965 }