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, struct pipe_inode_info *pipe,
 235                            size_t len, unsigned int flags)
 236 {
 237         struct address_space *mapping = in->f_mapping;
 238         unsigned int offset, nr_pages;
 239         struct page *pages[PIPE_BUFFERS];
 240         struct page *page;
 241         pgoff_t index;
 242         int i, error;
 243
 244         index = in->f_pos >> PAGE_CACHE_SHIFT;
 245         offset = in->f_pos & ~PAGE_CACHE_MASK;
 246         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 247
 248         if (nr_pages > PIPE_BUFFERS)
 249                 nr_pages = PIPE_BUFFERS;
 250
 251         /*
 252          * initiate read-ahead on this page range. however, don't call into
 253          * read-ahead if this is a non-zero offset (we are likely doing small
 254          * chunk splice and the page is already there) for a single page.
 255          */
 256         if (!offset || nr_pages > 1)
 257                 do_page_cache_readahead(mapping, in, index, nr_pages);
 258
 259         /*
 260          * now fill in the holes
 261          */
 262         error = 0;
 263         for (i = 0; i < nr_pages; i++, index++) {
 264 find_page:
 265                 /*
 266                  * lookup the page for this index
 267                  */
 268                 page = find_get_page(mapping, index);
 269                 if (!page) {
 270                         /*
 271                          * If in nonblock mode then dont block on
 272                          * readpage (we've kicked readahead so there
 273                          * will be asynchronous progress):
 274                          */
 275                         if (flags & SPLICE_F_NONBLOCK)
 276                                 break;
 277
 278                         /*
 279                          * page didn't exist, allocate one
 280                          */
 281                         page = page_cache_alloc_cold(mapping);
 282                         if (!page)
 283                                 break;
 284
 285                         error = add_to_page_cache_lru(page, mapping, index,
 286                                                 mapping_gfp_mask(mapping));
 287                         if (unlikely(error)) {
 288                                 page_cache_release(page);
 289                                 break;
 290                         }
 291
 292                         goto readpage;
 293                 }
 294
 295                 /*
 296                  * If the page isn't uptodate, we may need to start io on it
 297                  */
 298                 if (!PageUptodate(page)) {
 299                         lock_page(page);
 300
 301                         /*
 302                          * page was truncated, stop here. if this isn't the
 303                          * first page, we'll just complete what we already
 304                          * added
 305                          */
 306                         if (!page->mapping) {
 307                                 unlock_page(page);
 308                                 page_cache_release(page);
 309                                 break;
 310                         }
 311                         /*
 312                          * page was already under io and is now done, great
 313                          */
 314                         if (PageUptodate(page)) {
 315                                 unlock_page(page);
 316                                 goto fill_it;
 317                         }
 318
 319 readpage:
 320                         /*
 321                          * need to read in the page
 322                          */
 323                         error = mapping->a_ops->readpage(in, page);
 324
 325                         if (unlikely(error)) {
 326                                 page_cache_release(page);
 327                                 if (error == AOP_TRUNCATED_PAGE)
 328                                         goto find_page;
 329                                 break;
 330                         }
 331                 }
 332 fill_it:
 333                 pages[i] = page;
 334         }
 335
 336         if (i)
 337                 return move_to_pipe(pipe, pages, i, offset, len, flags);
 338
 339         return error;
 340 }
 341
 342 /**
 343  * generic_file_splice_read - splice data from file to a pipe
 344  * @in:         file to splice from
 345  * @pipe:       pipe to splice to
 346  * @len:        number of bytes to splice
 347  * @flags:      splice modifier flags
 348  *
 349  * Will read pages from given file and fill them into a pipe.
 350  */
 351 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 352                                  size_t len, unsigned int flags)
 353 {
 354         ssize_t spliced;
 355         int ret;
 356
 357         ret = 0;
 358         spliced = 0;
 359
 360         while (len) {
 361                 ret = __generic_file_splice_read(in, pipe, len, flags);
 362
 363                 if (ret <= 0)
 364                         break;
 365
 366                 in->f_pos += ret;
 367                 len -= ret;
 368                 spliced += ret;
 369
 370                 if (!(flags & SPLICE_F_NONBLOCK))
 371                         continue;
 372                 ret = -EAGAIN;
 373                 break;
 374         }
 375
 376         if (spliced)
 377                 return spliced;
 378
 379         return ret;
 380 }
 381
 382 EXPORT_SYMBOL(generic_file_splice_read);
 383
 384 /*
 385  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 386  * using sendpage().
 387  */
 388 static int pipe_to_sendpage(struct pipe_inode_info *info,
 389                             struct pipe_buffer *buf, struct splice_desc *sd)
 390 {
 391         struct file *file = sd->file;
 392         loff_t pos = sd->pos;
 393         unsigned int offset;
 394         ssize_t ret;
 395         void *ptr;
 396         int more;
 397
 398         /*
 399          * sub-optimal, but we are limited by the pipe ->map. we don't
 400          * need a kmap'ed buffer here, we just want to make sure we
 401          * have the page pinned if the pipe page originates from the
 402          * page cache
 403          */
 404         ptr = buf->ops->map(file, info, buf);
 405         if (IS_ERR(ptr))
 406                 return PTR_ERR(ptr);
 407
 408         offset = pos & ~PAGE_CACHE_MASK;
 409         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 410
 411         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 412
 413         buf->ops->unmap(info, buf);
 414         if (ret == sd->len)
 415                 return 0;
 416
 417         return -EIO;
 418 }
 419
 420 /*
 421  * This is a little more tricky than the file -> pipe splicing. There are
 422  * basically three cases:
 423  *
 424  *      - Destination page already exists in the address space and there
 425  *        are users of it. For that case we have no other option that
 426  *        copying the data. Tough luck.
 427  *      - Destination page already exists in the address space, but there
 428  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 429  *        through to last case.
 430  *      - Destination page does not exist, we can add the pipe page to
 431  *        the page cache and avoid the copy.
 432  *
 433  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 434  * sd->flags), we attempt to migrate pages from the pipe to the output
 435  * file address space page cache. This is possible if no one else has
 436  * the pipe page referenced outside of the pipe and page cache. If
 437  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 438  * a new page in the output file page cache and fill/dirty that.
 439  */
 440 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 441                         struct splice_desc *sd)
 442 {
 443         struct file *file = sd->file;
 444         struct address_space *mapping = file->f_mapping;
 445         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 446         unsigned int offset;
 447         struct page *page;
 448         pgoff_t index;
 449         char *src;
 450         int ret;
 451
 452         /*
 453          * make sure the data in this buffer is uptodate
 454          */
 455         src = buf->ops->map(file, info, buf);
 456         if (IS_ERR(src))
 457                 return PTR_ERR(src);
 458
 459         index = sd->pos >> PAGE_CACHE_SHIFT;
 460         offset = sd->pos & ~PAGE_CACHE_MASK;
 461
 462         /*
 463          * reuse buf page, if SPLICE_F_MOVE is set
 464          */
 465         if (sd->flags & SPLICE_F_MOVE) {
 466                 /*
 467                  * If steal succeeds, buf->page is now pruned from the vm
 468                  * side (LRU and page cache) and we can reuse it.
 469                  */
 470                 if (buf->ops->steal(info, buf))
 471                         goto find_page;
 472
 473                 /*
 474                  * this will also set the page locked
 475                  */
 476                 page = buf->page;
 477                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 478                         goto find_page;
 479
 480                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 481                         lru_cache_add(page);
 482         } else {
 483 find_page:
 484                 ret = -ENOMEM;
 485                 page = find_or_create_page(mapping, index, gfp_mask);
 486                 if (!page)
 487                         goto out_nomem;
 488
 489                 /*
 490                  * If the page is uptodate, it is also locked. If it isn't
 491                  * uptodate, we can mark it uptodate if we are filling the
 492                  * full page. Otherwise we need to read it in first...
 493                  */
 494                 if (!PageUptodate(page)) {
 495                         if (sd->len < PAGE_CACHE_SIZE) {
 496                                 ret = mapping->a_ops->readpage(file, page);
 497                                 if (unlikely(ret))
 498                                         goto out;
 499
 500                                 lock_page(page);
 501
 502                                 if (!PageUptodate(page)) {
 503                                         /*
 504                                          * page got invalidated, repeat
 505                                          */
 506                                         if (!page->mapping) {
 507                                                 unlock_page(page);
 508                                                 page_cache_release(page);
 509                                                 goto find_page;
 510                                         }
 511                                         ret = -EIO;
 512                                         goto out;
 513                                 }
 514                         } else {
 515                                 WARN_ON(!PageLocked(page));
 516                                 SetPageUptodate(page);
 517                         }
 518                 }
 519         }
 520
 521         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 522         if (ret == AOP_TRUNCATED_PAGE) {
 523                 page_cache_release(page);
 524                 goto find_page;
 525         } else if (ret)
 526                 goto out;
 527
 528         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 529                 char *dst = kmap_atomic(page, KM_USER0);
 530
 531                 memcpy(dst + offset, src + buf->offset, sd->len);
 532                 flush_dcache_page(page);
 533                 kunmap_atomic(dst, KM_USER0);
 534         }
 535
 536         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 537         if (ret == AOP_TRUNCATED_PAGE) {
 538                 page_cache_release(page);
 539                 goto find_page;
 540         } else if (ret)
 541                 goto out;
 542
 543         mark_page_accessed(page);
 544         balance_dirty_pages_ratelimited(mapping);
 545 out:
 546         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 547                 page_cache_release(page);
 548                 unlock_page(page);
 549         }
 550 out_nomem:
 551         buf->ops->unmap(info, buf);
 552         return ret;
 553 }
 554
 555 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 556                            struct splice_desc *);
 557
 558 /*
 559  * Pipe input worker. Most of this logic works like a regular pipe, the
 560  * key here is the 'actor' worker passed in that actually moves the data
 561  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 562  */
 563 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 564                               size_t len, unsigned int flags,
 565                               splice_actor *actor)
 566 {
 567         int ret, do_wakeup, err;
 568         struct splice_desc sd;
 569
 570         ret = 0;
 571         do_wakeup = 0;
 572
 573         sd.total_len = len;
 574         sd.flags = flags;
 575         sd.file = out;
 576         sd.pos = out->f_pos;
 577
 578         if (pipe->inode)
 579                 mutex_lock(&pipe->inode->i_mutex);
 580
 581         for (;;) {
 582                 if (pipe->nrbufs) {
 583                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 584                         struct pipe_buf_operations *ops = buf->ops;
 585
 586                         sd.len = buf->len;
 587                         if (sd.len > sd.total_len)
 588                                 sd.len = sd.total_len;
 589
 590                         err = actor(pipe, buf, &sd);
 591                         if (err) {
 592                                 if (!ret && err != -ENODATA)
 593                                         ret = err;
 594
 595                                 break;
 596                         }
 597
 598                         ret += sd.len;
 599                         buf->offset += sd.len;
 600                         buf->len -= sd.len;
 601                         if (!buf->len) {
 602                                 buf->ops = NULL;
 603                                 ops->release(pipe, buf);
 604                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 605                                 pipe->nrbufs--;
 606                                 if (pipe->inode)
 607                                         do_wakeup = 1;
 608                         }
 609
 610                         sd.pos += sd.len;
 611                         sd.total_len -= sd.len;
 612                         if (!sd.total_len)
 613                                 break;
 614                 }
 615
 616                 if (pipe->nrbufs)
 617                         continue;
 618                 if (!pipe->writers)
 619                         break;
 620                 if (!pipe->waiting_writers) {
 621                         if (ret)
 622                                 break;
 623                 }
 624
 625                 if (flags & SPLICE_F_NONBLOCK) {
 626                         if (!ret)
 627                                 ret = -EAGAIN;
 628                         break;
 629                 }
 630
 631                 if (signal_pending(current)) {
 632                         if (!ret)
 633                                 ret = -ERESTARTSYS;
 634                         break;
 635                 }
 636
 637                 if (do_wakeup) {
 638                         smp_mb();
 639                         if (waitqueue_active(&pipe->wait))
 640                                 wake_up_interruptible_sync(&pipe->wait);
 641                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 642                         do_wakeup = 0;
 643                 }
 644
 645                 pipe_wait(pipe);
 646         }
 647
 648         if (pipe->inode)
 649                 mutex_unlock(&pipe->inode->i_mutex);
 650
 651         if (do_wakeup) {
 652                 smp_mb();
 653                 if (waitqueue_active(&pipe->wait))
 654                         wake_up_interruptible(&pipe->wait);
 655                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 656         }
 657
 658         out->f_pos = sd.pos;
 659         return ret;
 660
 661 }
 662
 663 /**
 664  * generic_file_splice_write - splice data from a pipe to a file
 665  * @pipe:       pipe info
 666  * @out:        file to write to
 667  * @len:        number of bytes to splice
 668  * @flags:      splice modifier flags
 669  *
 670  * Will either move or copy pages (determined by @flags options) from
 671  * the given pipe inode to the given file.
 672  *
 673  */
 674 ssize_t
 675 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 676                           size_t len, unsigned int flags)
 677 {
 678         struct address_space *mapping = out->f_mapping;
 679         ssize_t ret;
 680
 681         ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
 682
 683         /*
 684          * if file or inode is SYNC and we actually wrote some data, sync it
 685          */
 686         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 687             && ret > 0) {
 688                 struct inode *inode = mapping->host;
 689                 int err;
 690
 691                 mutex_lock(&inode->i_mutex);
 692                 err = generic_osync_inode(mapping->host, mapping,
 693                                           OSYNC_METADATA|OSYNC_DATA);
 694                 mutex_unlock(&inode->i_mutex);
 695
 696                 if (err)
 697                         ret = err;
 698         }
 699
 700         return ret;
 701 }
 702
 703 EXPORT_SYMBOL(generic_file_splice_write);
 704
 705 /**
 706  * generic_splice_sendpage - splice data from a pipe to a socket
 707  * @inode:      pipe inode
 708  * @out:        socket to write to
 709  * @len:        number of bytes to splice
 710  * @flags:      splice modifier flags
 711  *
 712  * Will send @len bytes from the pipe to a network socket. No data copying
 713  * is involved.
 714  *
 715  */
 716 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 717                                 size_t len, unsigned int flags)
 718 {
 719         return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
 720 }
 721
 722 EXPORT_SYMBOL(generic_splice_sendpage);
 723
 724 /*
 725  * Attempt to initiate a splice from pipe to file.
 726  */
 727 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 728                            size_t len, unsigned int flags)
 729 {
 730         loff_t pos;
 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         pos = out->f_pos;
 740
 741         ret = rw_verify_area(WRITE, out, &pos, len);
 742         if (unlikely(ret < 0))
 743                 return ret;
 744
 745         return out->f_op->splice_write(pipe, out, len, flags);
 746 }
 747
 748 /*
 749  * Attempt to initiate a splice from a file to a pipe.
 750  */
 751 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
 752                          size_t len, unsigned int flags)
 753 {
 754         loff_t pos, isize, left;
 755         int ret;
 756
 757         if (unlikely(!in->f_op || !in->f_op->splice_read))
 758                 return -EINVAL;
 759
 760         if (unlikely(!(in->f_mode & FMODE_READ)))
 761                 return -EBADF;
 762
 763         pos = in->f_pos;
 764
 765         ret = rw_verify_area(READ, in, &pos, len);
 766         if (unlikely(ret < 0))
 767                 return ret;
 768
 769         isize = i_size_read(in->f_mapping->host);
 770         if (unlikely(in->f_pos >= isize))
 771                 return 0;
 772
 773         left = isize - in->f_pos;
 774         if (unlikely(left < len))
 775                 len = left;
 776
 777         return in->f_op->splice_read(in, pipe, len, flags);
 778 }
 779
 780 long do_splice_direct(struct file *in, struct file *out, size_t len,
 781                       unsigned int flags)
 782 {
 783         struct pipe_inode_info *pipe;
 784         long ret, bytes;
 785         umode_t i_mode;
 786         int i;
 787
 788         /*
 789          * We require the input being a regular file, as we don't want to
 790          * randomly drop data for eg socket -> socket splicing. Use the
 791          * piped splicing for that!
 792          */
 793         i_mode = in->f_dentry->d_inode->i_mode;
 794         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 795                 return -EINVAL;
 796
 797         /*
 798          * neither in nor out is a pipe, setup an internal pipe attached to
 799          * 'out' and transfer the wanted data from 'in' to 'out' through that
 800          */
 801         pipe = current->splice_pipe;
 802         if (unlikely(!pipe)) {
 803                 pipe = alloc_pipe_info(NULL);
 804                 if (!pipe)
 805                         return -ENOMEM;
 806
 807                 /*
 808                  * We don't have an immediate reader, but we'll read the stuff
 809                  * out of the pipe right after the move_to_pipe(). So set
 810                  * PIPE_READERS appropriately.
 811                  */
 812                 pipe->readers = 1;
 813
 814                 current->splice_pipe = pipe;
 815         }
 816
 817         /*
 818          * do the splice
 819          */
 820         ret = 0;
 821         bytes = 0;
 822
 823         while (len) {
 824                 size_t read_len, max_read_len;
 825
 826                 /*
 827                  * Do at most PIPE_BUFFERS pages worth of transfer:
 828                  */
 829                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 830
 831                 ret = do_splice_to(in, pipe, max_read_len, flags);
 832                 if (unlikely(ret < 0))
 833                         goto out_release;
 834
 835                 read_len = ret;
 836
 837                 /*
 838                  * NOTE: nonblocking mode only applies to the input. We
 839                  * must not do the output in nonblocking mode as then we
 840                  * could get stuck data in the internal pipe:
 841                  */
 842                 ret = do_splice_from(pipe, out, read_len,
 843                                      flags & ~SPLICE_F_NONBLOCK);
 844                 if (unlikely(ret < 0))
 845                         goto out_release;
 846
 847                 bytes += ret;
 848                 len -= ret;
 849
 850                 /*
 851                  * In nonblocking mode, if we got back a short read then
 852                  * that was due to either an IO error or due to the
 853                  * pagecache entry not being there. In the IO error case
 854                  * the _next_ splice attempt will produce a clean IO error
 855                  * return value (not a short read), so in both cases it's
 856                  * correct to break out of the loop here:
 857                  */
 858                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 859                         break;
 860         }
 861
 862         pipe->nrbufs = pipe->curbuf = 0;
 863
 864         return bytes;
 865
 866 out_release:
 867         /*
 868          * If we did an incomplete transfer we must release
 869          * the pipe buffers in question:
 870          */
 871         for (i = 0; i < PIPE_BUFFERS; i++) {
 872                 struct pipe_buffer *buf = pipe->bufs + i;
 873
 874                 if (buf->ops) {
 875                         buf->ops->release(pipe, buf);
 876                         buf->ops = NULL;
 877                 }
 878         }
 879         pipe->nrbufs = pipe->curbuf = 0;
 880
 881         /*
 882          * If we transferred some data, return the number of bytes:
 883          */
 884         if (bytes > 0)
 885                 return bytes;
 886
 887         return ret;
 888 }
 889
 890 EXPORT_SYMBOL(do_splice_direct);
 891
 892 /*
 893  * Determine where to splice to/from.
 894  */
 895 static long do_splice(struct file *in, loff_t __user *off_in,
 896                       struct file *out, loff_t __user *off_out,
 897                       size_t len, unsigned int flags)
 898 {
 899         struct pipe_inode_info *pipe;
 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(&out->f_pos, off_out,
 909                                            sizeof(loff_t)))
 910                                 return -EFAULT;
 911                 }
 912
 913                 return do_splice_from(pipe, out, len, flags);
 914         }
 915
 916         pipe = out->f_dentry->d_inode->i_pipe;
 917         if (pipe) {
 918                 if (off_out)
 919                         return -ESPIPE;
 920                 if (off_in) {
 921                         if (in->f_op->llseek == no_llseek)
 922                                 return -EINVAL;
 923                         if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
 924                                 return -EFAULT;
 925                 }
 926
 927                 return do_splice_to(in, pipe, len, flags);
 928         }
 929
 930         return -EINVAL;
 931 }
 932
 933 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 934                            int fd_out, loff_t __user *off_out,
 935                            size_t len, unsigned int flags)
 936 {
 937         long error;
 938         struct file *in, *out;
 939         int fput_in, fput_out;
 940
 941         if (unlikely(!len))
 942                 return 0;
 943
 944         error = -EBADF;
 945         in = fget_light(fd_in, &fput_in);
 946         if (in) {
 947                 if (in->f_mode & FMODE_READ) {
 948                         out = fget_light(fd_out, &fput_out);
 949                         if (out) {
 950                                 if (out->f_mode & FMODE_WRITE)
 951                                         error = do_splice(in, off_in,
 952                                                           out, off_out,
 953                                                           len, flags);
 954                                 fput_light(out, fput_out);
 955                         }
 956                 }
 957
 958                 fput_light(in, fput_in);
 959         }
 960
 961         return error;
 962 }