SAFE public projects git trees. - safe/jmp/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 void page_cache_pipe_buf_get(struct pipe_inode_info *info,
 129                                     struct pipe_buffer *buf)
 130 {
 131         page_cache_get(buf->page);
 132 }
 133
 134 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 135         .can_merge = 0,
 136         .map = page_cache_pipe_buf_map,
 137         .unmap = page_cache_pipe_buf_unmap,
 138         .release = page_cache_pipe_buf_release,
 139         .steal = page_cache_pipe_buf_steal,
 140         .get = page_cache_pipe_buf_get,
 141 };
 142
 143 /*
 144  * Pipe output worker. This sets up our pipe format with the page cache
 145  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 146  */
 147 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
 148                             int nr_pages, unsigned long len,
 149                             unsigned int offset, unsigned int flags)
 150 {
 151         int ret, do_wakeup, i;
 152
 153         ret = 0;
 154         do_wakeup = 0;
 155         i = 0;
 156
 157         if (pipe->inode)
 158                 mutex_lock(&pipe->inode->i_mutex);
 159
 160         for (;;) {
 161                 if (!pipe->readers) {
 162                         send_sig(SIGPIPE, current, 0);
 163                         if (!ret)
 164                                 ret = -EPIPE;
 165                         break;
 166                 }
 167
 168                 if (pipe->nrbufs < PIPE_BUFFERS) {
 169                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 170                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 171                         struct page *page = pages[i++];
 172                         unsigned long this_len;
 173
 174                         this_len = PAGE_CACHE_SIZE - offset;
 175                         if (this_len > len)
 176                                 this_len = len;
 177
 178                         buf->page = page;
 179                         buf->offset = offset;
 180                         buf->len = this_len;
 181                         buf->ops = &page_cache_pipe_buf_ops;
 182                         pipe->nrbufs++;
 183                         if (pipe->inode)
 184                                 do_wakeup = 1;
 185
 186                         ret += this_len;
 187                         len -= this_len;
 188                         offset = 0;
 189                         if (!--nr_pages)
 190                                 break;
 191                         if (!len)
 192                                 break;
 193                         if (pipe->nrbufs < PIPE_BUFFERS)
 194                                 continue;
 195
 196                         break;
 197                 }
 198
 199                 if (flags & SPLICE_F_NONBLOCK) {
 200                         if (!ret)
 201                                 ret = -EAGAIN;
 202                         break;
 203                 }
 204
 205                 if (signal_pending(current)) {
 206                         if (!ret)
 207                                 ret = -ERESTARTSYS;
 208                         break;
 209                 }
 210
 211                 if (do_wakeup) {
 212                         smp_mb();
 213                         if (waitqueue_active(&pipe->wait))
 214                                 wake_up_interruptible_sync(&pipe->wait);
 215                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 216                         do_wakeup = 0;
 217                 }
 218
 219                 pipe->waiting_writers++;
 220                 pipe_wait(pipe);
 221                 pipe->waiting_writers--;
 222         }
 223
 224         if (pipe->inode)
 225                 mutex_unlock(&pipe->inode->i_mutex);
 226
 227         if (do_wakeup) {
 228                 smp_mb();
 229                 if (waitqueue_active(&pipe->wait))
 230                         wake_up_interruptible(&pipe->wait);
 231                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 232         }
 233
 234         while (i < nr_pages)
 235                 page_cache_release(pages[i++]);
 236
 237         return ret;
 238 }
 239
 240 static int
 241 __generic_file_splice_read(struct file *in, loff_t *ppos,
 242                            struct pipe_inode_info *pipe, size_t len,
 243                            unsigned int flags)
 244 {
 245         struct address_space *mapping = in->f_mapping;
 246         unsigned int loff, offset, nr_pages;
 247         struct page *pages[PIPE_BUFFERS];
 248         struct page *page;
 249         pgoff_t index, end_index;
 250         loff_t isize;
 251         size_t bytes;
 252         int i, error;
 253
 254         index = *ppos >> PAGE_CACHE_SHIFT;
 255         loff = offset = *ppos & ~PAGE_CACHE_MASK;
 256         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 257
 258         if (nr_pages > PIPE_BUFFERS)
 259                 nr_pages = PIPE_BUFFERS;
 260
 261         /*
 262          * Initiate read-ahead on this page range. however, don't call into
 263          * read-ahead if this is a non-zero offset (we are likely doing small
 264          * chunk splice and the page is already there) for a single page.
 265          */
 266         if (!offset || nr_pages > 1)
 267                 do_page_cache_readahead(mapping, in, index, nr_pages);
 268
 269         /*
 270          * Now fill in the holes:
 271          */
 272         error = 0;
 273         bytes = 0;
 274         for (i = 0; i < nr_pages; i++, index++) {
 275 find_page:
 276                 /*
 277                  * lookup the page for this index
 278                  */
 279                 page = find_get_page(mapping, index);
 280                 if (!page) {
 281                         /*
 282                          * If in nonblock mode then dont block on
 283                          * readpage (we've kicked readahead so there
 284                          * will be asynchronous progress):
 285                          */
 286                         if (flags & SPLICE_F_NONBLOCK)
 287                                 break;
 288
 289                         /*
 290                          * page didn't exist, allocate one
 291                          */
 292                         page = page_cache_alloc_cold(mapping);
 293                         if (!page)
 294                                 break;
 295
 296                         error = add_to_page_cache_lru(page, mapping, index,
 297                                                 mapping_gfp_mask(mapping));
 298                         if (unlikely(error)) {
 299                                 page_cache_release(page);
 300                                 break;
 301                         }
 302
 303                         goto readpage;
 304                 }
 305
 306                 /*
 307                  * If the page isn't uptodate, we may need to start io on it
 308                  */
 309                 if (!PageUptodate(page)) {
 310                         lock_page(page);
 311
 312                         /*
 313                          * page was truncated, stop here. if this isn't the
 314                          * first page, we'll just complete what we already
 315                          * added
 316                          */
 317                         if (!page->mapping) {
 318                                 unlock_page(page);
 319                                 page_cache_release(page);
 320                                 break;
 321                         }
 322                         /*
 323                          * page was already under io and is now done, great
 324                          */
 325                         if (PageUptodate(page)) {
 326                                 unlock_page(page);
 327                                 goto fill_it;
 328                         }
 329
 330 readpage:
 331                         /*
 332                          * need to read in the page
 333                          */
 334                         error = mapping->a_ops->readpage(in, page);
 335
 336                         if (unlikely(error)) {
 337                                 page_cache_release(page);
 338                                 if (error == AOP_TRUNCATED_PAGE)
 339                                         goto find_page;
 340                                 break;
 341                         }
 342
 343                         /*
 344                          * i_size must be checked after ->readpage().
 345                          */
 346                         isize = i_size_read(mapping->host);
 347                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 348                         if (unlikely(!isize || index > end_index)) {
 349                                 page_cache_release(page);
 350                                 break;
 351                         }
 352
 353                         /*
 354                          * if this is the last page, see if we need to shrink
 355                          * the length and stop
 356                          */
 357                         if (end_index == index) {
 358                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 359                                 if (bytes + loff > isize) {
 360                                         page_cache_release(page);
 361                                         break;
 362                                 }
 363                                 /*
 364                                  * force quit after adding this page
 365                                  */
 366                                 nr_pages = i;
 367                         }
 368                 }
 369 fill_it:
 370                 pages[i] = page;
 371                 bytes += PAGE_CACHE_SIZE - loff;
 372                 loff = 0;
 373         }
 374
 375         if (i)
 376                 return move_to_pipe(pipe, pages, i, bytes, offset, flags);
 377
 378         return error;
 379 }
 380
 381 /**
 382  * generic_file_splice_read - splice data from file to a pipe
 383  * @in:         file to splice from
 384  * @pipe:       pipe to splice to
 385  * @len:        number of bytes to splice
 386  * @flags:      splice modifier flags
 387  *
 388  * Will read pages from given file and fill them into a pipe.
 389  */
 390 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 391                                  struct pipe_inode_info *pipe, size_t len,
 392                                  unsigned int flags)
 393 {
 394         ssize_t spliced;
 395         int ret;
 396
 397         ret = 0;
 398         spliced = 0;
 399
 400         while (len) {
 401                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 402
 403                 if (ret <= 0)
 404                         break;
 405
 406                 *ppos += ret;
 407                 len -= ret;
 408                 spliced += ret;
 409
 410                 if (!(flags & SPLICE_F_NONBLOCK))
 411                         continue;
 412                 ret = -EAGAIN;
 413                 break;
 414         }
 415
 416         if (spliced)
 417                 return spliced;
 418
 419         return ret;
 420 }
 421
 422 EXPORT_SYMBOL(generic_file_splice_read);
 423
 424 /*
 425  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 426  * using sendpage().
 427  */
 428 static int pipe_to_sendpage(struct pipe_inode_info *info,
 429                             struct pipe_buffer *buf, struct splice_desc *sd)
 430 {
 431         struct file *file = sd->file;
 432         loff_t pos = sd->pos;
 433         unsigned int offset;
 434         ssize_t ret;
 435         void *ptr;
 436         int more;
 437
 438         /*
 439          * Sub-optimal, but we are limited by the pipe ->map. We don't
 440          * need a kmap'ed buffer here, we just want to make sure we
 441          * have the page pinned if the pipe page originates from the
 442          * page cache.
 443          */
 444         ptr = buf->ops->map(file, info, buf);
 445         if (IS_ERR(ptr))
 446                 return PTR_ERR(ptr);
 447
 448         offset = pos & ~PAGE_CACHE_MASK;
 449         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 450
 451         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 452
 453         buf->ops->unmap(info, buf);
 454         if (ret == sd->len)
 455                 return 0;
 456
 457         return -EIO;
 458 }
 459
 460 /*
 461  * This is a little more tricky than the file -> pipe splicing. There are
 462  * basically three cases:
 463  *
 464  *      - Destination page already exists in the address space and there
 465  *        are users of it. For that case we have no other option that
 466  *        copying the data. Tough luck.
 467  *      - Destination page already exists in the address space, but there
 468  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 469  *        through to last case.
 470  *      - Destination page does not exist, we can add the pipe page to
 471  *        the page cache and avoid the copy.
 472  *
 473  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 474  * sd->flags), we attempt to migrate pages from the pipe to the output
 475  * file address space page cache. This is possible if no one else has
 476  * the pipe page referenced outside of the pipe and page cache. If
 477  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 478  * a new page in the output file page cache and fill/dirty that.
 479  */
 480 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 481                         struct splice_desc *sd)
 482 {
 483         struct file *file = sd->file;
 484         struct address_space *mapping = file->f_mapping;
 485         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 486         unsigned int offset;
 487         struct page *page;
 488         pgoff_t index;
 489         char *src;
 490         int ret;
 491
 492         /*
 493          * make sure the data in this buffer is uptodate
 494          */
 495         src = buf->ops->map(file, info, buf);
 496         if (IS_ERR(src))
 497                 return PTR_ERR(src);
 498
 499         index = sd->pos >> PAGE_CACHE_SHIFT;
 500         offset = sd->pos & ~PAGE_CACHE_MASK;
 501
 502         /*
 503          * Reuse buf page, if SPLICE_F_MOVE is set.
 504          */
 505         if (sd->flags & SPLICE_F_MOVE) {
 506                 /*
 507                  * If steal succeeds, buf->page is now pruned from the vm
 508                  * side (LRU and page cache) and we can reuse it.
 509                  */
 510                 if (buf->ops->steal(info, buf))
 511                         goto find_page;
 512
 513                 /*
 514                  * this will also set the page locked
 515                  */
 516                 page = buf->page;
 517                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 518                         goto find_page;
 519
 520                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 521                         lru_cache_add(page);
 522         } else {
 523 find_page:
 524                 ret = -ENOMEM;
 525                 page = find_or_create_page(mapping, index, gfp_mask);
 526                 if (!page)
 527                         goto out_nomem;
 528
 529                 /*
 530                  * If the page is uptodate, it is also locked. If it isn't
 531                  * uptodate, we can mark it uptodate if we are filling the
 532                  * full page. Otherwise we need to read it in first...
 533                  */
 534                 if (!PageUptodate(page)) {
 535                         if (sd->len < PAGE_CACHE_SIZE) {
 536                                 ret = mapping->a_ops->readpage(file, page);
 537                                 if (unlikely(ret))
 538                                         goto out;
 539
 540                                 lock_page(page);
 541
 542                                 if (!PageUptodate(page)) {
 543                                         /*
 544                                          * Page got invalidated, repeat.
 545                                          */
 546                                         if (!page->mapping) {
 547                                                 unlock_page(page);
 548                                                 page_cache_release(page);
 549                                                 goto find_page;
 550                                         }
 551                                         ret = -EIO;
 552                                         goto out;
 553                                 }
 554                         } else {
 555                                 WARN_ON(!PageLocked(page));
 556                                 SetPageUptodate(page);
 557                         }
 558                 }
 559         }
 560
 561         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 562         if (ret == AOP_TRUNCATED_PAGE) {
 563                 page_cache_release(page);
 564                 goto find_page;
 565         } else if (ret)
 566                 goto out;
 567
 568         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 569                 char *dst = kmap_atomic(page, KM_USER0);
 570
 571                 memcpy(dst + offset, src + buf->offset, sd->len);
 572                 flush_dcache_page(page);
 573                 kunmap_atomic(dst, KM_USER0);
 574         }
 575
 576         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 577         if (ret == AOP_TRUNCATED_PAGE) {
 578                 page_cache_release(page);
 579                 goto find_page;
 580         } else if (ret)
 581                 goto out;
 582
 583         mark_page_accessed(page);
 584         balance_dirty_pages_ratelimited(mapping);
 585 out:
 586         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 587                 page_cache_release(page);
 588                 unlock_page(page);
 589         }
 590 out_nomem:
 591         buf->ops->unmap(info, buf);
 592         return ret;
 593 }
 594
 595 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 596                            struct splice_desc *);
 597
 598 /*
 599  * Pipe input worker. Most of this logic works like a regular pipe, the
 600  * key here is the 'actor' worker passed in that actually moves the data
 601  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 602  */
 603 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 604                               loff_t *ppos, size_t len, unsigned int flags,
 605                               splice_actor *actor)
 606 {
 607         int ret, do_wakeup, err;
 608         struct splice_desc sd;
 609
 610         ret = 0;
 611         do_wakeup = 0;
 612
 613         sd.total_len = len;
 614         sd.flags = flags;
 615         sd.file = out;
 616         sd.pos = *ppos;
 617
 618         if (pipe->inode)
 619                 mutex_lock(&pipe->inode->i_mutex);
 620
 621         for (;;) {
 622                 if (pipe->nrbufs) {
 623                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 624                         struct pipe_buf_operations *ops = buf->ops;
 625
 626                         sd.len = buf->len;
 627                         if (sd.len > sd.total_len)
 628                                 sd.len = sd.total_len;
 629
 630                         err = actor(pipe, buf, &sd);
 631                         if (err) {
 632                                 if (!ret && err != -ENODATA)
 633                                         ret = err;
 634
 635                                 break;
 636                         }
 637
 638                         ret += sd.len;
 639                         buf->offset += sd.len;
 640                         buf->len -= sd.len;
 641
 642                         if (!buf->len) {
 643                                 buf->ops = NULL;
 644                                 ops->release(pipe, buf);
 645                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 646                                 pipe->nrbufs--;
 647                                 if (pipe->inode)
 648                                         do_wakeup = 1;
 649                         }
 650
 651                         sd.pos += sd.len;
 652                         sd.total_len -= sd.len;
 653                         if (!sd.total_len)
 654                                 break;
 655                 }
 656
 657                 if (pipe->nrbufs)
 658                         continue;
 659                 if (!pipe->writers)
 660                         break;
 661                 if (!pipe->waiting_writers) {
 662                         if (ret)
 663                                 break;
 664                 }
 665
 666                 if (flags & SPLICE_F_NONBLOCK) {
 667                         if (!ret)
 668                                 ret = -EAGAIN;
 669                         break;
 670                 }
 671
 672                 if (signal_pending(current)) {
 673                         if (!ret)
 674                                 ret = -ERESTARTSYS;
 675                         break;
 676                 }
 677
 678                 if (do_wakeup) {
 679                         smp_mb();
 680                         if (waitqueue_active(&pipe->wait))
 681                                 wake_up_interruptible_sync(&pipe->wait);
 682                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 683                         do_wakeup = 0;
 684                 }
 685
 686                 pipe_wait(pipe);
 687         }
 688
 689         if (pipe->inode)
 690                 mutex_unlock(&pipe->inode->i_mutex);
 691
 692         if (do_wakeup) {
 693                 smp_mb();
 694                 if (waitqueue_active(&pipe->wait))
 695                         wake_up_interruptible(&pipe->wait);
 696                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 697         }
 698
 699         return ret;
 700 }
 701
 702 /**
 703  * generic_file_splice_write - splice data from a pipe to a file
 704  * @pipe:       pipe info
 705  * @out:        file to write to
 706  * @len:        number of bytes to splice
 707  * @flags:      splice modifier flags
 708  *
 709  * Will either move or copy pages (determined by @flags options) from
 710  * the given pipe inode to the given file.
 711  *
 712  */
 713 ssize_t
 714 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 715                           loff_t *ppos, size_t len, unsigned int flags)
 716 {
 717         struct address_space *mapping = out->f_mapping;
 718         ssize_t ret;
 719
 720         ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 721
 722         /*
 723          * If file or inode is SYNC and we actually wrote some data, sync it.
 724          */
 725         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
 726             && ret > 0) {
 727                 struct inode *inode = mapping->host;
 728                 int err;
 729
 730                 mutex_lock(&inode->i_mutex);
 731                 err = generic_osync_inode(mapping->host, mapping,
 732                                           OSYNC_METADATA|OSYNC_DATA);
 733                 mutex_unlock(&inode->i_mutex);
 734
 735                 if (err)
 736                         ret = err;
 737         }
 738
 739         return ret;
 740 }
 741
 742 EXPORT_SYMBOL(generic_file_splice_write);
 743
 744 /**
 745  * generic_splice_sendpage - splice data from a pipe to a socket
 746  * @inode:      pipe inode
 747  * @out:        socket to write to
 748  * @len:        number of bytes to splice
 749  * @flags:      splice modifier flags
 750  *
 751  * Will send @len bytes from the pipe to a network socket. No data copying
 752  * is involved.
 753  *
 754  */
 755 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 756                                 loff_t *ppos, size_t len, unsigned int flags)
 757 {
 758         return move_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 759 }
 760
 761 EXPORT_SYMBOL(generic_splice_sendpage);
 762
 763 /*
 764  * Attempt to initiate a splice from pipe to file.
 765  */
 766 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 767                            loff_t *ppos, size_t len, unsigned int flags)
 768 {
 769         int ret;
 770
 771         if (unlikely(!out->f_op || !out->f_op->splice_write))
 772                 return -EINVAL;
 773
 774         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 775                 return -EBADF;
 776
 777         ret = rw_verify_area(WRITE, out, ppos, len);
 778         if (unlikely(ret < 0))
 779                 return ret;
 780
 781         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 782 }
 783
 784 /*
 785  * Attempt to initiate a splice from a file to a pipe.
 786  */
 787 static long do_splice_to(struct file *in, loff_t *ppos,
 788                          struct pipe_inode_info *pipe, size_t len,
 789                          unsigned int flags)
 790 {
 791         loff_t isize, left;
 792         int ret;
 793
 794         if (unlikely(!in->f_op || !in->f_op->splice_read))
 795                 return -EINVAL;
 796
 797         if (unlikely(!(in->f_mode & FMODE_READ)))
 798                 return -EBADF;
 799
 800         ret = rw_verify_area(READ, in, ppos, len);
 801         if (unlikely(ret < 0))
 802                 return ret;
 803
 804         isize = i_size_read(in->f_mapping->host);
 805         if (unlikely(*ppos >= isize))
 806                 return 0;
 807
 808         left = isize - *ppos;
 809         if (unlikely(left < len))
 810                 len = left;
 811
 812         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 813 }
 814
 815 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 816                       size_t len, unsigned int flags)
 817 {
 818         struct pipe_inode_info *pipe;
 819         long ret, bytes;
 820         loff_t out_off;
 821         umode_t i_mode;
 822         int i;
 823
 824         /*
 825          * We require the input being a regular file, as we don't want to
 826          * randomly drop data for eg socket -> socket splicing. Use the
 827          * piped splicing for that!
 828          */
 829         i_mode = in->f_dentry->d_inode->i_mode;
 830         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 831                 return -EINVAL;
 832
 833         /*
 834          * neither in nor out is a pipe, setup an internal pipe attached to
 835          * 'out' and transfer the wanted data from 'in' to 'out' through that
 836          */
 837         pipe = current->splice_pipe;
 838         if (unlikely(!pipe)) {
 839                 pipe = alloc_pipe_info(NULL);
 840                 if (!pipe)
 841                         return -ENOMEM;
 842
 843                 /*
 844                  * We don't have an immediate reader, but we'll read the stuff
 845                  * out of the pipe right after the move_to_pipe(). So set
 846                  * PIPE_READERS appropriately.
 847                  */
 848                 pipe->readers = 1;
 849
 850                 current->splice_pipe = pipe;
 851         }
 852
 853         /*
 854          * Do the splice.
 855          */
 856         ret = 0;
 857         bytes = 0;
 858         out_off = 0;
 859
 860         while (len) {
 861                 size_t read_len, max_read_len;
 862
 863                 /*
 864                  * Do at most PIPE_BUFFERS pages worth of transfer:
 865                  */
 866                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 867
 868                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 869                 if (unlikely(ret < 0))
 870                         goto out_release;
 871
 872                 read_len = ret;
 873
 874                 /*
 875                  * NOTE: nonblocking mode only applies to the input. We
 876                  * must not do the output in nonblocking mode as then we
 877                  * could get stuck data in the internal pipe:
 878                  */
 879                 ret = do_splice_from(pipe, out, &out_off, read_len,
 880                                      flags & ~SPLICE_F_NONBLOCK);
 881                 if (unlikely(ret < 0))
 882                         goto out_release;
 883
 884                 bytes += ret;
 885                 len -= ret;
 886
 887                 /*
 888                  * In nonblocking mode, if we got back a short read then
 889                  * that was due to either an IO error or due to the
 890                  * pagecache entry not being there. In the IO error case
 891                  * the _next_ splice attempt will produce a clean IO error
 892                  * return value (not a short read), so in both cases it's
 893                  * correct to break out of the loop here:
 894                  */
 895                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 896                         break;
 897         }
 898
 899         pipe->nrbufs = pipe->curbuf = 0;
 900
 901         return bytes;
 902
 903 out_release:
 904         /*
 905          * If we did an incomplete transfer we must release
 906          * the pipe buffers in question:
 907          */
 908         for (i = 0; i < PIPE_BUFFERS; i++) {
 909                 struct pipe_buffer *buf = pipe->bufs + i;
 910
 911                 if (buf->ops) {
 912                         buf->ops->release(pipe, buf);
 913                         buf->ops = NULL;
 914                 }
 915         }
 916         pipe->nrbufs = pipe->curbuf = 0;
 917
 918         /*
 919          * If we transferred some data, return the number of bytes:
 920          */
 921         if (bytes > 0)
 922                 return bytes;
 923
 924         return ret;
 925 }
 926
 927 EXPORT_SYMBOL(do_splice_direct);
 928
 929 /*
 930  * Determine where to splice to/from.
 931  */
 932 static long do_splice(struct file *in, loff_t __user *off_in,
 933                       struct file *out, loff_t __user *off_out,
 934                       size_t len, unsigned int flags)
 935 {
 936         struct pipe_inode_info *pipe;
 937         loff_t offset, *off;
 938
 939         pipe = in->f_dentry->d_inode->i_pipe;
 940         if (pipe) {
 941                 if (off_in)
 942                         return -ESPIPE;
 943                 if (off_out) {
 944                         if (out->f_op->llseek == no_llseek)
 945                                 return -EINVAL;
 946                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
 947                                 return -EFAULT;
 948                         off = &offset;
 949                 } else
 950                         off = &out->f_pos;
 951
 952                 return do_splice_from(pipe, out, off, len, flags);
 953         }
 954
 955         pipe = out->f_dentry->d_inode->i_pipe;
 956         if (pipe) {
 957                 if (off_out)
 958                         return -ESPIPE;
 959                 if (off_in) {
 960                         if (in->f_op->llseek == no_llseek)
 961                                 return -EINVAL;
 962                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
 963                                 return -EFAULT;
 964                         off = &offset;
 965                 } else
 966                         off = &in->f_pos;
 967
 968                 return do_splice_to(in, off, pipe, len, flags);
 969         }
 970
 971         return -EINVAL;
 972 }
 973
 974 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 975                            int fd_out, loff_t __user *off_out,
 976                            size_t len, unsigned int flags)
 977 {
 978         long error;
 979         struct file *in, *out;
 980         int fput_in, fput_out;
 981
 982         if (unlikely(!len))
 983                 return 0;
 984
 985         error = -EBADF;
 986         in = fget_light(fd_in, &fput_in);
 987         if (in) {
 988                 if (in->f_mode & FMODE_READ) {
 989                         out = fget_light(fd_out, &fput_out);
 990                         if (out) {
 991                                 if (out->f_mode & FMODE_WRITE)
 992                                         error = do_splice(in, off_in,
 993                                                           out, off_out,
 994                                                           len, flags);
 995                                 fput_light(out, fput_out);
 996                         }
 997                 }
 998
 999                 fput_light(in, fput_in);
1000         }
1001
1002         return error;
1003 }
1004
1005 /*
1006  * Link contents of ipipe to opipe.
1007  */
1008 static int link_pipe(struct pipe_inode_info *ipipe,
1009                      struct pipe_inode_info *opipe,
1010                      size_t len, unsigned int flags)
1011 {
1012         struct pipe_buffer *ibuf, *obuf;
1013         int ret = 0, do_wakeup = 0, i;
1014
1015         /*
1016          * Potential ABBA deadlock, work around it by ordering lock
1017          * grabbing by inode address. Otherwise two different processes
1018          * could deadlock (one doing tee from A -> B, the other from B -> A).
1019          */
1020         if (ipipe->inode < opipe->inode) {
1021                 mutex_lock(&ipipe->inode->i_mutex);
1022                 mutex_lock(&opipe->inode->i_mutex);
1023         } else {
1024                 mutex_lock(&opipe->inode->i_mutex);
1025                 mutex_lock(&ipipe->inode->i_mutex);
1026         }
1027
1028         for (i = 0;; i++) {
1029                 if (!opipe->readers) {
1030                         send_sig(SIGPIPE, current, 0);
1031                         if (!ret)
1032                                 ret = -EPIPE;
1033                         break;
1034                 }
1035                 if (ipipe->nrbufs - i) {
1036                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1037
1038                         /*
1039                          * If we have room, fill this buffer
1040                          */
1041                         if (opipe->nrbufs < PIPE_BUFFERS) {
1042                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1043
1044                                 /*
1045                                  * Get a reference to this pipe buffer,
1046                                  * so we can copy the contents over.
1047                                  */
1048                                 ibuf->ops->get(ipipe, ibuf);
1049
1050                                 obuf = opipe->bufs + nbuf;
1051                                 *obuf = *ibuf;
1052
1053                                 if (obuf->len > len)
1054                                         obuf->len = len;
1055
1056                                 opipe->nrbufs++;
1057                                 do_wakeup = 1;
1058                                 ret += obuf->len;
1059                                 len -= obuf->len;
1060
1061                                 if (!len)
1062                                         break;
1063                                 if (opipe->nrbufs < PIPE_BUFFERS)
1064                                         continue;
1065                         }
1066
1067                         /*
1068                          * We have input available, but no output room.
1069                          * If we already copied data, return that.
1070                          */
1071                         if (flags & SPLICE_F_NONBLOCK) {
1072                                 if (!ret)
1073                                         ret = -EAGAIN;
1074                                 break;
1075                         }
1076                         if (signal_pending(current)) {
1077                                 if (!ret)
1078                                         ret = -ERESTARTSYS;
1079                                 break;
1080                         }
1081                         if (do_wakeup) {
1082                                 smp_mb();
1083                                 if (waitqueue_active(&opipe->wait))
1084                                         wake_up_interruptible(&opipe->wait);
1085                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1086                                 do_wakeup = 0;
1087                         }
1088
1089                         opipe->waiting_writers++;
1090                         pipe_wait(opipe);
1091                         opipe->waiting_writers--;
1092                         continue;
1093                 }
1094
1095                 /*
1096                  * No input buffers, do the usual checks for available
1097                  * writers and blocking and wait if necessary
1098                  */
1099                 if (!ipipe->writers)
1100                         break;
1101                 if (!ipipe->waiting_writers) {
1102                         if (ret)
1103                                 break;
1104                 }
1105                 if (flags & SPLICE_F_NONBLOCK) {
1106                         if (!ret)
1107                                 ret = -EAGAIN;
1108                         break;
1109                 }
1110                 if (signal_pending(current)) {
1111                         if (!ret)
1112                                 ret = -ERESTARTSYS;
1113                         break;
1114                 }
1115
1116                 if (waitqueue_active(&ipipe->wait))
1117                         wake_up_interruptible_sync(&ipipe->wait);
1118                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1119
1120                 pipe_wait(ipipe);
1121         }
1122
1123         mutex_unlock(&ipipe->inode->i_mutex);
1124         mutex_unlock(&opipe->inode->i_mutex);
1125
1126         if (do_wakeup) {
1127                 smp_mb();
1128                 if (waitqueue_active(&opipe->wait))
1129                         wake_up_interruptible(&opipe->wait);
1130                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1131         }
1132
1133         return ret;
1134 }
1135
1136 /*
1137  * This is a tee(1) implementation that works on pipes. It doesn't copy
1138  * any data, it simply references the 'in' pages on the 'out' pipe.
1139  * The 'flags' used are the SPLICE_F_* variants, currently the only
1140  * applicable one is SPLICE_F_NONBLOCK.
1141  */
1142 static long do_tee(struct file *in, struct file *out, size_t len,
1143                    unsigned int flags)
1144 {
1145         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1146         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1147
1148         /*
1149          * Link ipipe to the two output pipes, consuming as we go along.
1150          */
1151         if (ipipe && opipe)
1152                 return link_pipe(ipipe, opipe, len, flags);
1153
1154         return -EINVAL;
1155 }
1156
1157 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1158 {
1159         struct file *in;
1160         int error, fput_in;
1161
1162         if (unlikely(!len))
1163                 return 0;
1164
1165         error = -EBADF;
1166         in = fget_light(fdin, &fput_in);
1167         if (in) {
1168                 if (in->f_mode & FMODE_READ) {
1169                         int fput_out;
1170                         struct file *out = fget_light(fdout, &fput_out);
1171
1172                         if (out) {
1173                                 if (out->f_mode & FMODE_WRITE)
1174                                         error = do_tee(in, out, len, flags);
1175                                 fput_light(out, fput_out);
1176                         }
1177                 }
1178                 fput_light(in, fput_in);
1179         }
1180
1181         return error;
1182 }