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                 if (!pipe->readers) {
 154                         send_sig(SIGPIPE, current, 0);
 155                         if (!ret)
 156                                 ret = -EPIPE;
 157                         break;
 158                 }
 159
 160                 if (pipe->nrbufs < PIPE_BUFFERS) {
 161                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 162                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 163                         struct page *page = pages[i++];
 164                         unsigned long this_len;
 165
 166                         this_len = PAGE_CACHE_SIZE - offset;
 167                         if (this_len > len)
 168                                 this_len = len;
 169
 170                         buf->page = page;
 171                         buf->offset = offset;
 172                         buf->len = this_len;
 173                         buf->ops = &page_cache_pipe_buf_ops;
 174                         pipe->nrbufs++;
 175                         if (pipe->inode)
 176                                 do_wakeup = 1;
 177
 178                         ret += this_len;
 179                         len -= this_len;
 180                         offset = 0;
 181                         if (!--nr_pages)
 182                                 break;
 183                         if (!len)
 184                                 break;
 185                         if (pipe->nrbufs < PIPE_BUFFERS)
 186                                 continue;
 187
 188                         break;
 189                 }
 190
 191                 if (flags & SPLICE_F_NONBLOCK) {
 192                         if (!ret)
 193                                 ret = -EAGAIN;
 194                         break;
 195                 }
 196
 197                 if (signal_pending(current)) {
 198                         if (!ret)
 199                                 ret = -ERESTARTSYS;
 200                         break;
 201                 }
 202
 203                 if (do_wakeup) {
 204                         smp_mb();
 205                         if (waitqueue_active(&pipe->wait))
 206                                 wake_up_interruptible_sync(&pipe->wait);
 207                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 208                         do_wakeup = 0;
 209                 }
 210
 211                 pipe->waiting_writers++;
 212                 pipe_wait(pipe);
 213                 pipe->waiting_writers--;
 214         }
 215
 216         if (pipe->inode)
 217                 mutex_unlock(&pipe->inode->i_mutex);
 218
 219         if (do_wakeup) {
 220                 smp_mb();
 221                 if (waitqueue_active(&pipe->wait))
 222                         wake_up_interruptible(&pipe->wait);
 223                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 224         }
 225
 226         while (i < nr_pages)
 227                 page_cache_release(pages[i++]);
 228
 229         return ret;
 230 }
 231
 232 static int
 233 __generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 234                            size_t len, unsigned int flags)
 235 {
 236         struct address_space *mapping = in->f_mapping;
 237         unsigned int offset, nr_pages;
 238         struct page *pages[PIPE_BUFFERS];
 239         struct page *page;
 240         pgoff_t index;
 241         int i, error;
 242
 243         index = in->f_pos >> PAGE_CACHE_SHIFT;
 244         offset = in->f_pos & ~PAGE_CACHE_MASK;
 245         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 246
 247         if (nr_pages > PIPE_BUFFERS)
 248                 nr_pages = PIPE_BUFFERS;
 249
 250         /*
 251          * initiate read-ahead on this page range. however, don't call into
 252          * read-ahead if this is a non-zero offset (we are likely doing small
 253          * chunk splice and the page is already there) for a single page.
 254          */
 255         if (!offset || nr_pages > 1)
 256                 do_page_cache_readahead(mapping, in, index, nr_pages);
 257
 258         /*
 259          * now fill in the holes
 260          */
 261         error = 0;
 262         for (i = 0; i < nr_pages; i++, index++) {
 263 find_page:
 264                 /*
 265                  * lookup the page for this index
 266                  */
 267                 page = find_get_page(mapping, index);
 268                 if (!page) {
 269                         /*
 270                          * If in nonblock mode then dont block on
 271                          * readpage (we've kicked readahead so there
 272                          * will be asynchronous progress):
 273                          */
 274                         if (flags & SPLICE_F_NONBLOCK)
 275                                 break;
 276
 277                         /*
 278                          * page didn't exist, allocate one
 279                          */
 280                         page = page_cache_alloc_cold(mapping);
 281                         if (!page)
 282                                 break;
 283
 284                         error = add_to_page_cache_lru(page, mapping, index,
 285                                                 mapping_gfp_mask(mapping));
 286                         if (unlikely(error)) {
 287                                 page_cache_release(page);
 288                                 break;
 289                         }
 290
 291                         goto readpage;
 292                 }
 293
 294                 /*
 295                  * If the page isn't uptodate, we may need to start io on it
 296                  */
 297                 if (!PageUptodate(page)) {
 298                         lock_page(page);
 299
 300                         /*
 301                          * page was truncated, stop here. if this isn't the
 302                          * first page, we'll just complete what we already
 303                          * added
 304                          */
 305                         if (!page->mapping) {
 306                                 unlock_page(page);
 307                                 page_cache_release(page);
 308                                 break;
 309                         }
 310                         /*
 311                          * page was already under io and is now done, great
 312                          */
 313                         if (PageUptodate(page)) {
 314                                 unlock_page(page);
 315                                 goto fill_it;
 316                         }
 317
 318 readpage:
 319                         /*
 320                          * need to read in the page
 321                          */
 322                         error = mapping->a_ops->readpage(in, page);
 323
 324                         if (unlikely(error)) {
 325                                 page_cache_release(page);
 326                                 if (error == AOP_TRUNCATED_PAGE)
 327                                         goto find_page;
 328                                 break;
 329                         }
 330                 }
 331 fill_it:
 332                 pages[i] = page;
 333         }
 334
 335         if (i)
 336                 return move_to_pipe(pipe, pages, i, offset, len, flags);
 337
 338         return error;
 339 }
 340
 341 /**
 342  * generic_file_splice_read - splice data from file to a pipe
 343  * @in:         file to splice from
 344  * @pipe:       pipe to splice to
 345  * @len:        number of bytes to splice
 346  * @flags:      splice modifier flags
 347  *
 348  * Will read pages from given file and fill them into a pipe.
 349  */
 350 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
 351                                  size_t len, unsigned int flags)
 352 {
 353         ssize_t spliced;
 354         int ret;
 355
 356         ret = 0;
 357         spliced = 0;
 358
 359         while (len) {
 360                 ret = __generic_file_splice_read(in, pipe, len, flags);
 361
 362                 if (ret <= 0)
 363                         break;
 364
 365                 in->f_pos += ret;
 366                 len -= ret;
 367                 spliced += ret;
 368
 369                 if (!(flags & SPLICE_F_NONBLOCK))
 370                         continue;
 371                 ret = -EAGAIN;
 372                 break;
 373         }
 374
 375         if (spliced)
 376                 return spliced;
 377
 378         return ret;
 379 }
 380
 381 EXPORT_SYMBOL(generic_file_splice_read);
 382
 383 /*
 384  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 385  * using sendpage().
 386  */
 387 static int pipe_to_sendpage(struct pipe_inode_info *info,
 388                             struct pipe_buffer *buf, struct splice_desc *sd)
 389 {
 390         struct file *file = sd->file;
 391         loff_t pos = sd->pos;
 392         unsigned int offset;
 393         ssize_t ret;
 394         void *ptr;
 395         int more;
 396
 397         /*
 398          * sub-optimal, but we are limited by the pipe ->map. we don't
 399          * need a kmap'ed buffer here, we just want to make sure we
 400          * have the page pinned if the pipe page originates from the
 401          * page cache
 402          */
 403         ptr = buf->ops->map(file, info, buf);
 404         if (IS_ERR(ptr))
 405                 return PTR_ERR(ptr);
 406
 407         offset = pos & ~PAGE_CACHE_MASK;
 408         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 409
 410         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 411
 412         buf->ops->unmap(info, buf);
 413         if (ret == sd->len)
 414                 return 0;
 415
 416         return -EIO;
 417 }
 418
 419 /*
 420  * This is a little more tricky than the file -> pipe splicing. There are
 421  * basically three cases:
 422  *
 423  *      - Destination page already exists in the address space and there
 424  *        are users of it. For that case we have no other option that
 425  *        copying the data. Tough luck.
 426  *      - Destination page already exists in the address space, but there
 427  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 428  *        through to last case.
 429  *      - Destination page does not exist, we can add the pipe page to
 430  *        the page cache and avoid the copy.
 431  *
 432  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 433  * sd->flags), we attempt to migrate pages from the pipe to the output
 434  * file address space page cache. This is possible if no one else has
 435  * the pipe page referenced outside of the pipe and page cache. If
 436  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 437  * a new page in the output file page cache and fill/dirty that.
 438  */
 439 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 440                         struct splice_desc *sd)
 441 {
 442         struct file *file = sd->file;
 443         struct address_space *mapping = file->f_mapping;
 444         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 445         unsigned int offset;
 446         struct page *page;
 447         pgoff_t index;
 448         char *src;
 449         int ret;
 450
 451         /*
 452          * make sure the data in this buffer is uptodate
 453          */
 454         src = buf->ops->map(file, info, buf);
 455         if (IS_ERR(src))
 456                 return PTR_ERR(src);
 457
 458         index = sd->pos >> PAGE_CACHE_SHIFT;
 459         offset = sd->pos & ~PAGE_CACHE_MASK;
 460
 461         /*
 462          * reuse buf page, if SPLICE_F_MOVE is set
 463          */
 464         if (sd->flags & SPLICE_F_MOVE) {
 465                 /*
 466                  * If steal succeeds, buf->page is now pruned from the vm
 467                  * side (LRU and page cache) and we can reuse it.
 468                  */
 469                 if (buf->ops->steal(info, buf))
 470                         goto find_page;
 471
 472                 /*
 473                  * this will also set the page locked
 474                  */
 475                 page = buf->page;
 476                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 477                         goto find_page;
 478
 479                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 480                         lru_cache_add(page);
 481         } else {
 482 find_page:
 483                 ret = -ENOMEM;
 484                 page = find_or_create_page(mapping, index, gfp_mask);
 485                 if (!page)
 486                         goto out_nomem;
 487
 488                 /*
 489                  * If the page is uptodate, it is also locked. If it isn't
 490                  * uptodate, we can mark it uptodate if we are filling the
 491                  * full page. Otherwise we need to read it in first...
 492                  */
 493                 if (!PageUptodate(page)) {
 494                         if (sd->len < PAGE_CACHE_SIZE) {
 495                                 ret = mapping->a_ops->readpage(file, page);
 496                                 if (unlikely(ret))
 497                                         goto out;
 498
 499                                 lock_page(page);
 500
 501                                 if (!PageUptodate(page)) {
 502                                         /*
 503                                          * page got invalidated, repeat
 504                                          */
 505                                         if (!page->mapping) {
 506                                                 unlock_page(page);
 507                                                 page_cache_release(page);
 508                                                 goto find_page;
 509                                         }
 510                                         ret = -EIO;
 511                                         goto out;
 512                                 }
 513                         } else {
 514                                 WARN_ON(!PageLocked(page));
 515                                 SetPageUptodate(page);
 516                         }
 517                 }
 518         }
 519
 520         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 521         if (ret == AOP_TRUNCATED_PAGE) {
 522                 page_cache_release(page);
 523                 goto find_page;
 524         } else if (ret)
 525                 goto out;
 526
 527         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 528                 char *dst = kmap_atomic(page, KM_USER0);
 529
 530                 memcpy(dst + offset, src + buf->offset, sd->len);
 531                 flush_dcache_page(page);
 532                 kunmap_atomic(dst, KM_USER0);
 533         }
 534
 535         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 536         if (ret == AOP_TRUNCATED_PAGE) {
 537                 page_cache_release(page);
 538                 goto find_page;
 539         } else if (ret)
 540                 goto out;
 541
 542         mark_page_accessed(page);
 543         balance_dirty_pages_ratelimited(mapping);
 544 out:
 545         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 546                 page_cache_release(page);
 547                 unlock_page(page);
 548         }
 549 out_nomem:
 550         buf->ops->unmap(info, buf);
 551         return ret;
 552 }
 553
 554 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 555                            struct splice_desc *);
 556
 557 /*
 558  * Pipe input worker. Most of this logic works like a regular pipe, the
 559  * key here is the 'actor' worker passed in that actually moves the data
 560  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 561  */
 562 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 563                               size_t len, unsigned int flags,
 564                               splice_actor *actor)
 565 {
 566         int ret, do_wakeup, err;
 567         struct splice_desc sd;
 568
 569         ret = 0;
 570         do_wakeup = 0;
 571
 572         sd.total_len = len;
 573         sd.flags = flags;
 574         sd.file = out;
 575         sd.pos = out->f_pos;
 576
 577         if (pipe->inode)
 578                 mutex_lock(&pipe->inode->i_mutex);
 579
 580         for (;;) {
 581                 if (pipe->nrbufs) {
 582                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 583                         struct pipe_buf_operations *ops = buf->ops;
 584
 585                         sd.len = buf->len;
 586                         if (sd.len > sd.total_len)
 587                                 sd.len = sd.total_len;
 588
 589                         err = actor(pipe, buf, &sd);
 590                         if (err) {
 591                                 if (!ret && err != -ENODATA)
 592                                         ret = err;
 593
 594                                 break;
 595                         }
 596
 597                         ret += sd.len;
 598                         buf->offset += sd.len;
 599                         buf->len -= sd.len;
 600                         if (!buf->len) {
 601                                 buf->ops = NULL;
 602                                 ops->release(pipe, buf);
 603                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 604                                 pipe->nrbufs--;
 605                                 if (pipe->inode)
 606                                         do_wakeup = 1;
 607                         }
 608
 609                         sd.pos += sd.len;
 610                         sd.total_len -= sd.len;
 611                         if (!sd.total_len)
 612                                 break;
 613                 }
 614
 615                 if (pipe->nrbufs)
 616                         continue;
 617                 if (!pipe->writers)
 618                         break;
 619                 if (!pipe->waiting_writers) {
 620                         if (ret)
 621                                 break;
 622                 }
 623
 624                 if (flags & SPLICE_F_NONBLOCK) {
 625                         if (!ret)
 626                                 ret = -EAGAIN;
 627                         break;
 628                 }
 629
 630                 if (signal_pending(current)) {
 631                         if (!ret)
 632                                 ret = -ERESTARTSYS;
 633                         break;
 634                 }
 635
 636                 if (do_wakeup) {
 637                         smp_mb();
 638                         if (waitqueue_active(&pipe->wait))
 639                                 wake_up_interruptible_sync(&pipe->wait);
 640                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 641                         do_wakeup = 0;
 642                 }
 643
 644                 pipe_wait(pipe);
 645         }
 646
 647         if (pipe->inode)
 648                 mutex_unlock(&pipe->inode->i_mutex);
 649
 650         if (do_wakeup) {
 651                 smp_mb();
 652                 if (waitqueue_active(&pipe->wait))
 653                         wake_up_interruptible(&pipe->wait);
 654                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 655         }
 656
 657         out->f_pos = sd.pos;
 658         return ret;
 659
 660 }
 661
 662 /**
 663  * generic_file_splice_write - splice data from a pipe to a file
 664  * @pipe:       pipe info
 665  * @out:        file to write to
 666  * @len:        number of bytes to splice
 667  * @flags:      splice modifier flags
 668  *
 669  * Will either move or copy pages (determined by @flags options) from
 670  * the given pipe inode to the given file.
 671  *
 672  */
 673 ssize_t
 674 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 675                           size_t len, unsigned int flags)
 676 {
 677         struct address_space *mapping = out->f_mapping;
 678         ssize_t ret;
 679
 680         ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
 681
 682         /*
 683          * if file or inode is SYNC and we actually wrote some data, sync it
 684          */
 685         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 686             && ret > 0) {
 687                 struct inode *inode = mapping->host;
 688                 int err;
 689
 690                 mutex_lock(&inode->i_mutex);
 691                 err = generic_osync_inode(mapping->host, mapping,
 692                                           OSYNC_METADATA|OSYNC_DATA);
 693                 mutex_unlock(&inode->i_mutex);
 694
 695                 if (err)
 696                         ret = err;
 697         }
 698
 699         return ret;
 700 }
 701
 702 EXPORT_SYMBOL(generic_file_splice_write);
 703
 704 /**
 705  * generic_splice_sendpage - splice data from a pipe to a socket
 706  * @inode:      pipe inode
 707  * @out:        socket to write to
 708  * @len:        number of bytes to splice
 709  * @flags:      splice modifier flags
 710  *
 711  * Will send @len bytes from the pipe to a network socket. No data copying
 712  * is involved.
 713  *
 714  */
 715 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 716                                 size_t len, unsigned int flags)
 717 {
 718         return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
 719 }
 720
 721 EXPORT_SYMBOL(generic_splice_sendpage);
 722
 723 /*
 724  * Attempt to initiate a splice from pipe to file.
 725  */
 726 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 727                            size_t len, unsigned int flags)
 728 {
 729         loff_t pos;
 730         int ret;
 731
 732         if (unlikely(!out->f_op || !out->f_op->splice_write))
 733                 return -EINVAL;
 734
 735         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 736                 return -EBADF;
 737
 738         pos = out->f_pos;
 739
 740         ret = rw_verify_area(WRITE, out, &pos, len);
 741         if (unlikely(ret < 0))
 742                 return ret;
 743
 744         return out->f_op->splice_write(pipe, out, len, flags);
 745 }
 746
 747 /*
 748  * Attempt to initiate a splice from a file to a pipe.
 749  */
 750 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
 751                          size_t len, unsigned int flags)
 752 {
 753         loff_t pos, 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         pos = in->f_pos;
 763
 764         ret = rw_verify_area(READ, in, &pos, len);
 765         if (unlikely(ret < 0))
 766                 return ret;
 767
 768         isize = i_size_read(in->f_mapping->host);
 769         if (unlikely(in->f_pos >= isize))
 770                 return 0;
 771
 772         left = isize - in->f_pos;
 773         if (unlikely(left < len))
 774                 len = left;
 775
 776         return in->f_op->splice_read(in, pipe, len, flags);
 777 }
 778
 779 long do_splice_direct(struct file *in, struct file *out, size_t len,
 780                       unsigned int flags)
 781 {
 782         struct pipe_inode_info *pipe;
 783         long ret, bytes;
 784         umode_t i_mode;
 785         int i;
 786
 787         /*
 788          * We require the input being a regular file, as we don't want to
 789          * randomly drop data for eg socket -> socket splicing. Use the
 790          * piped splicing for that!
 791          */
 792         i_mode = in->f_dentry->d_inode->i_mode;
 793         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 794                 return -EINVAL;
 795
 796         /*
 797          * neither in nor out is a pipe, setup an internal pipe attached to
 798          * 'out' and transfer the wanted data from 'in' to 'out' through that
 799          */
 800         pipe = current->splice_pipe;
 801         if (unlikely(!pipe)) {
 802                 pipe = alloc_pipe_info(NULL);
 803                 if (!pipe)
 804                         return -ENOMEM;
 805
 806                 /*
 807                  * We don't have an immediate reader, but we'll read the stuff
 808                  * out of the pipe right after the move_to_pipe(). So set
 809                  * PIPE_READERS appropriately.
 810                  */
 811                 pipe->readers = 1;
 812
 813                 current->splice_pipe = pipe;
 814         }
 815
 816         /*
 817          * do the splice
 818          */
 819         ret = 0;
 820         bytes = 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, 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, 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
 900         pipe = in->f_dentry->d_inode->i_pipe;
 901         if (pipe) {
 902                 if (off_in)
 903                         return -ESPIPE;
 904                 if (off_out) {
 905                         if (out->f_op->llseek == no_llseek)
 906                                 return -EINVAL;
 907                         if (copy_from_user(&out->f_pos, off_out,
 908                                            sizeof(loff_t)))
 909                                 return -EFAULT;
 910                 }
 911
 912                 return do_splice_from(pipe, out, len, flags);
 913         }
 914
 915         pipe = out->f_dentry->d_inode->i_pipe;
 916         if (pipe) {
 917                 if (off_out)
 918                         return -ESPIPE;
 919                 if (off_in) {
 920                         if (in->f_op->llseek == no_llseek)
 921                                 return -EINVAL;
 922                         if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
 923                                 return -EFAULT;
 924                 }
 925
 926                 return do_splice_to(in, pipe, len, flags);
 927         }
 928
 929         return -EINVAL;
 930 }
 931
 932 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 933                            int fd_out, loff_t __user *off_out,
 934                            size_t len, unsigned int flags)
 935 {
 936         long error;
 937         struct file *in, *out;
 938         int fput_in, fput_out;
 939
 940         if (unlikely(!len))
 941                 return 0;
 942
 943         error = -EBADF;
 944         in = fget_light(fd_in, &fput_in);
 945         if (in) {
 946                 if (in->f_mode & FMODE_READ) {
 947                         out = fget_light(fd_out, &fput_out);
 948                         if (out) {
 949                                 if (out->f_mode & FMODE_WRITE)
 950                                         error = do_splice(in, off_in,
 951                                                           out, off_out,
 952                                                           len, flags);
 953                                 fput_light(out, fput_out);
 954                         }
 955                 }
 956
 957                 fput_light(in, fput_in);
 958         }
 959
 960         return error;
 961 }