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         lock_page(page);
  54
  55         WARN_ON(!PageUptodate(page));
  56
  57         /*
  58          * At least for ext2 with nobh option, we need to wait on writeback
  59          * completing on this page, since we'll remove it from the pagecache.
  60          * Otherwise truncate wont wait on the page, allowing the disk
  61          * blocks to be reused by someone else before we actually wrote our
  62          * data to them. fs corruption ensues.
  63          */
  64         wait_on_page_writeback(page);
  65
  66         if (PagePrivate(page))
  67                 try_to_release_page(page, mapping_gfp_mask(mapping));
  68
  69         if (!remove_mapping(mapping, page)) {
  70                 unlock_page(page);
  71                 return 1;
  72         }
  73
  74         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
  75         return 0;
  76 }
  77
  78 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  79                                         struct pipe_buffer *buf)
  80 {
  81         page_cache_release(buf->page);
  82         buf->page = NULL;
  83         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
  84 }
  85
  86 static void *page_cache_pipe_buf_map(struct file *file,
  87                                      struct pipe_inode_info *info,
  88                                      struct pipe_buffer *buf)
  89 {
  90         struct page *page = buf->page;
  91         int err;
  92
  93         if (!PageUptodate(page)) {
  94                 lock_page(page);
  95
  96                 /*
  97                  * Page got truncated/unhashed. This will cause a 0-byte
  98                  * splice, if this is the first page.
  99                  */
 100                 if (!page->mapping) {
 101                         err = -ENODATA;
 102                         goto error;
 103                 }
 104
 105                 /*
 106                  * Uh oh, read-error from disk.
 107                  */
 108                 if (!PageUptodate(page)) {
 109                         err = -EIO;
 110                         goto error;
 111                 }
 112
 113                 /*
 114                  * Page is ok afterall, fall through to mapping.
 115                  */
 116                 unlock_page(page);
 117         }
 118
 119         return kmap(page);
 120 error:
 121         unlock_page(page);
 122         return ERR_PTR(err);
 123 }
 124
 125 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
 126                                       struct pipe_buffer *buf)
 127 {
 128         kunmap(buf->page);
 129 }
 130
 131 static void page_cache_pipe_buf_get(struct pipe_inode_info *info,
 132                                     struct pipe_buffer *buf)
 133 {
 134         page_cache_get(buf->page);
 135 }
 136
 137 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .map = page_cache_pipe_buf_map,
 140         .unmap = page_cache_pipe_buf_unmap,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = page_cache_pipe_buf_get,
 144 };
 145
 146 /*
 147  * Pipe output worker. This sets up our pipe format with the page cache
 148  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
 149  */
 150 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
 151                             int nr_pages, unsigned long len,
 152                             unsigned int offset, unsigned int flags)
 153 {
 154         int ret, do_wakeup, i;
 155
 156         ret = 0;
 157         do_wakeup = 0;
 158         i = 0;
 159
 160         if (pipe->inode)
 161                 mutex_lock(&pipe->inode->i_mutex);
 162
 163         for (;;) {
 164                 if (!pipe->readers) {
 165                         send_sig(SIGPIPE, current, 0);
 166                         if (!ret)
 167                                 ret = -EPIPE;
 168                         break;
 169                 }
 170
 171                 if (pipe->nrbufs < PIPE_BUFFERS) {
 172                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 173                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 174                         struct page *page = pages[i++];
 175                         unsigned long this_len;
 176
 177                         this_len = PAGE_CACHE_SIZE - offset;
 178                         if (this_len > len)
 179                                 this_len = len;
 180
 181                         buf->page = page;
 182                         buf->offset = offset;
 183                         buf->len = this_len;
 184                         buf->ops = &page_cache_pipe_buf_ops;
 185                         pipe->nrbufs++;
 186                         if (pipe->inode)
 187                                 do_wakeup = 1;
 188
 189                         ret += this_len;
 190                         len -= this_len;
 191                         offset = 0;
 192                         if (!--nr_pages)
 193                                 break;
 194                         if (!len)
 195                                 break;
 196                         if (pipe->nrbufs < PIPE_BUFFERS)
 197                                 continue;
 198
 199                         break;
 200                 }
 201
 202                 if (flags & SPLICE_F_NONBLOCK) {
 203                         if (!ret)
 204                                 ret = -EAGAIN;
 205                         break;
 206                 }
 207
 208                 if (signal_pending(current)) {
 209                         if (!ret)
 210                                 ret = -ERESTARTSYS;
 211                         break;
 212                 }
 213
 214                 if (do_wakeup) {
 215                         smp_mb();
 216                         if (waitqueue_active(&pipe->wait))
 217                                 wake_up_interruptible_sync(&pipe->wait);
 218                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 219                         do_wakeup = 0;
 220                 }
 221
 222                 pipe->waiting_writers++;
 223                 pipe_wait(pipe);
 224                 pipe->waiting_writers--;
 225         }
 226
 227         if (pipe->inode)
 228                 mutex_unlock(&pipe->inode->i_mutex);
 229
 230         if (do_wakeup) {
 231                 smp_mb();
 232                 if (waitqueue_active(&pipe->wait))
 233                         wake_up_interruptible(&pipe->wait);
 234                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 235         }
 236
 237         while (i < nr_pages)
 238                 page_cache_release(pages[i++]);
 239
 240         return ret;
 241 }
 242
 243 static int
 244 __generic_file_splice_read(struct file *in, loff_t *ppos,
 245                            struct pipe_inode_info *pipe, size_t len,
 246                            unsigned int flags)
 247 {
 248         struct address_space *mapping = in->f_mapping;
 249         unsigned int loff, offset, nr_pages;
 250         struct page *pages[PIPE_BUFFERS];
 251         struct page *page;
 252         pgoff_t index, end_index;
 253         loff_t isize;
 254         size_t bytes;
 255         int i, error;
 256
 257         index = *ppos >> PAGE_CACHE_SHIFT;
 258         loff = offset = *ppos & ~PAGE_CACHE_MASK;
 259         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 260
 261         if (nr_pages > PIPE_BUFFERS)
 262                 nr_pages = PIPE_BUFFERS;
 263
 264         /*
 265          * Initiate read-ahead on this page range. however, don't call into
 266          * read-ahead if this is a non-zero offset (we are likely doing small
 267          * chunk splice and the page is already there) for a single page.
 268          */
 269         if (!offset || nr_pages > 1)
 270                 do_page_cache_readahead(mapping, in, index, nr_pages);
 271
 272         /*
 273          * Now fill in the holes:
 274          */
 275         error = 0;
 276         bytes = 0;
 277         for (i = 0; i < nr_pages; i++, index++) {
 278                 unsigned int this_len;
 279
 280                 if (!len)
 281                         break;
 282
 283                 /*
 284                  * this_len is the max we'll use from this page
 285                  */
 286                 this_len = min(len, PAGE_CACHE_SIZE - loff);
 287 find_page:
 288                 /*
 289                  * lookup the page for this index
 290                  */
 291                 page = find_get_page(mapping, index);
 292                 if (!page) {
 293                         /*
 294                          * page didn't exist, allocate one
 295                          */
 296                         page = page_cache_alloc_cold(mapping);
 297                         if (!page)
 298                                 break;
 299
 300                         error = add_to_page_cache_lru(page, mapping, index,
 301                                                 mapping_gfp_mask(mapping));
 302                         if (unlikely(error)) {
 303                                 page_cache_release(page);
 304                                 break;
 305                         }
 306
 307                         goto readpage;
 308                 }
 309
 310                 /*
 311                  * If the page isn't uptodate, we may need to start io on it
 312                  */
 313                 if (!PageUptodate(page)) {
 314                         /*
 315                          * If in nonblock mode then dont block on waiting
 316                          * for an in-flight io page
 317                          */
 318                         if (flags & SPLICE_F_NONBLOCK)
 319                                 break;
 320
 321                         lock_page(page);
 322
 323                         /*
 324                          * page was truncated, stop here. if this isn't the
 325                          * first page, we'll just complete what we already
 326                          * added
 327                          */
 328                         if (!page->mapping) {
 329                                 unlock_page(page);
 330                                 page_cache_release(page);
 331                                 break;
 332                         }
 333                         /*
 334                          * page was already under io and is now done, great
 335                          */
 336                         if (PageUptodate(page)) {
 337                                 unlock_page(page);
 338                                 goto fill_it;
 339                         }
 340
 341 readpage:
 342                         /*
 343                          * need to read in the page
 344                          */
 345                         error = mapping->a_ops->readpage(in, page);
 346
 347                         if (unlikely(error)) {
 348                                 page_cache_release(page);
 349                                 if (error == AOP_TRUNCATED_PAGE)
 350                                         goto find_page;
 351                                 break;
 352                         }
 353
 354                         /*
 355                          * i_size must be checked after ->readpage().
 356                          */
 357                         isize = i_size_read(mapping->host);
 358                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 359                         if (unlikely(!isize || index > end_index)) {
 360                                 page_cache_release(page);
 361                                 break;
 362                         }
 363
 364                         /*
 365                          * if this is the last page, see if we need to shrink
 366                          * the length and stop
 367                          */
 368                         if (end_index == index) {
 369                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
 370                                 if (bytes + loff > isize) {
 371                                         page_cache_release(page);
 372                                         break;
 373                                 }
 374                                 /*
 375                                  * force quit after adding this page
 376                                  */
 377                                 nr_pages = i;
 378                                 this_len = min(this_len, loff);
 379                         }
 380                 }
 381 fill_it:
 382                 pages[i] = page;
 383                 bytes += this_len;
 384                 len -= this_len;
 385                 loff = 0;
 386         }
 387
 388         if (i)
 389                 return move_to_pipe(pipe, pages, i, bytes, offset, flags);
 390
 391         return error;
 392 }
 393
 394 /**
 395  * generic_file_splice_read - splice data from file to a pipe
 396  * @in:         file to splice from
 397  * @pipe:       pipe to splice to
 398  * @len:        number of bytes to splice
 399  * @flags:      splice modifier flags
 400  *
 401  * Will read pages from given file and fill them into a pipe.
 402  */
 403 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 404                                  struct pipe_inode_info *pipe, size_t len,
 405                                  unsigned int flags)
 406 {
 407         ssize_t spliced;
 408         int ret;
 409
 410         ret = 0;
 411         spliced = 0;
 412
 413         while (len) {
 414                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 415
 416                 if (ret < 0)
 417                         break;
 418                 else if (!ret) {
 419                         if (spliced)
 420                                 break;
 421                         if (flags & SPLICE_F_NONBLOCK) {
 422                                 ret = -EAGAIN;
 423                                 break;
 424                         }
 425                 }
 426
 427                 *ppos += ret;
 428                 len -= ret;
 429                 spliced += ret;
 430         }
 431
 432         if (spliced)
 433                 return spliced;
 434
 435         return ret;
 436 }
 437
 438 EXPORT_SYMBOL(generic_file_splice_read);
 439
 440 /*
 441  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 442  * using sendpage().
 443  */
 444 static int pipe_to_sendpage(struct pipe_inode_info *info,
 445                             struct pipe_buffer *buf, struct splice_desc *sd)
 446 {
 447         struct file *file = sd->file;
 448         loff_t pos = sd->pos;
 449         unsigned int offset;
 450         ssize_t ret;
 451         void *ptr;
 452         int more;
 453
 454         /*
 455          * Sub-optimal, but we are limited by the pipe ->map. We don't
 456          * need a kmap'ed buffer here, we just want to make sure we
 457          * have the page pinned if the pipe page originates from the
 458          * page cache.
 459          */
 460         ptr = buf->ops->map(file, info, buf);
 461         if (IS_ERR(ptr))
 462                 return PTR_ERR(ptr);
 463
 464         offset = pos & ~PAGE_CACHE_MASK;
 465         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 466
 467         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
 468
 469         buf->ops->unmap(info, buf);
 470         if (ret == sd->len)
 471                 return 0;
 472
 473         return -EIO;
 474 }
 475
 476 /*
 477  * This is a little more tricky than the file -> pipe splicing. There are
 478  * basically three cases:
 479  *
 480  *      - Destination page already exists in the address space and there
 481  *        are users of it. For that case we have no other option that
 482  *        copying the data. Tough luck.
 483  *      - Destination page already exists in the address space, but there
 484  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 485  *        through to last case.
 486  *      - Destination page does not exist, we can add the pipe page to
 487  *        the page cache and avoid the copy.
 488  *
 489  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 490  * sd->flags), we attempt to migrate pages from the pipe to the output
 491  * file address space page cache. This is possible if no one else has
 492  * the pipe page referenced outside of the pipe and page cache. If
 493  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 494  * a new page in the output file page cache and fill/dirty that.
 495  */
 496 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 497                         struct splice_desc *sd)
 498 {
 499         struct file *file = sd->file;
 500         struct address_space *mapping = file->f_mapping;
 501         gfp_t gfp_mask = mapping_gfp_mask(mapping);
 502         unsigned int offset;
 503         struct page *page;
 504         pgoff_t index;
 505         char *src;
 506         int ret;
 507
 508         /*
 509          * make sure the data in this buffer is uptodate
 510          */
 511         src = buf->ops->map(file, info, buf);
 512         if (IS_ERR(src))
 513                 return PTR_ERR(src);
 514
 515         index = sd->pos >> PAGE_CACHE_SHIFT;
 516         offset = sd->pos & ~PAGE_CACHE_MASK;
 517
 518         /*
 519          * Reuse buf page, if SPLICE_F_MOVE is set.
 520          */
 521         if (sd->flags & SPLICE_F_MOVE) {
 522                 /*
 523                  * If steal succeeds, buf->page is now pruned from the vm
 524                  * side (LRU and page cache) and we can reuse it. The page
 525                  * will also be looked on successful return.
 526                  */
 527                 if (buf->ops->steal(info, buf))
 528                         goto find_page;
 529
 530                 page = buf->page;
 531                 if (add_to_page_cache(page, mapping, index, gfp_mask))
 532                         goto find_page;
 533
 534                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 535                         lru_cache_add(page);
 536         } else {
 537 find_page:
 538                 page = find_lock_page(mapping, index);
 539                 if (!page) {
 540                         ret = -ENOMEM;
 541                         page = page_cache_alloc_cold(mapping);
 542                         if (unlikely(!page))
 543                                 goto out_nomem;
 544
 545                         /*
 546                          * This will also lock the page
 547                          */
 548                         ret = add_to_page_cache_lru(page, mapping, index,
 549                                                     gfp_mask);
 550                         if (unlikely(ret))
 551                                 goto out;
 552                 }
 553
 554                 /*
 555                  * We get here with the page locked. If the page is also
 556                  * uptodate, we don't need to do more. If it isn't, we
 557                  * may need to bring it in if we are not going to overwrite
 558                  * the full page.
 559                  */
 560                 if (!PageUptodate(page)) {
 561                         if (sd->len < PAGE_CACHE_SIZE) {
 562                                 ret = mapping->a_ops->readpage(file, page);
 563                                 if (unlikely(ret))
 564                                         goto out;
 565
 566                                 lock_page(page);
 567
 568                                 if (!PageUptodate(page)) {
 569                                         /*
 570                                          * Page got invalidated, repeat.
 571                                          */
 572                                         if (!page->mapping) {
 573                                                 unlock_page(page);
 574                                                 page_cache_release(page);
 575                                                 goto find_page;
 576                                         }
 577                                         ret = -EIO;
 578                                         goto out;
 579                                 }
 580                         } else
 581                                 SetPageUptodate(page);
 582                 }
 583         }
 584
 585         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 586         if (ret == AOP_TRUNCATED_PAGE) {
 587                 page_cache_release(page);
 588                 goto find_page;
 589         } else if (ret)
 590                 goto out;
 591
 592         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
 593                 char *dst = kmap_atomic(page, KM_USER0);
 594
 595                 memcpy(dst + offset, src + buf->offset, sd->len);
 596                 flush_dcache_page(page);
 597                 kunmap_atomic(dst, KM_USER0);
 598         }
 599
 600         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 601         if (ret == AOP_TRUNCATED_PAGE) {
 602                 page_cache_release(page);
 603                 goto find_page;
 604         } else if (ret)
 605                 goto out;
 606
 607         mark_page_accessed(page);
 608         balance_dirty_pages_ratelimited(mapping);
 609 out:
 610         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN))
 611                 page_cache_release(page);
 612
 613         unlock_page(page);
 614 out_nomem:
 615         buf->ops->unmap(info, buf);
 616         return ret;
 617 }
 618
 619 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 620                            struct splice_desc *);
 621
 622 /*
 623  * Pipe input worker. Most of this logic works like a regular pipe, the
 624  * key here is the 'actor' worker passed in that actually moves the data
 625  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 626  */
 627 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 628                               loff_t *ppos, size_t len, unsigned int flags,
 629                               splice_actor *actor)
 630 {
 631         int ret, do_wakeup, err;
 632         struct splice_desc sd;
 633
 634         ret = 0;
 635         do_wakeup = 0;
 636
 637         sd.total_len = len;
 638         sd.flags = flags;
 639         sd.file = out;
 640         sd.pos = *ppos;
 641
 642         if (pipe->inode)
 643                 mutex_lock(&pipe->inode->i_mutex);
 644
 645         for (;;) {
 646                 if (pipe->nrbufs) {
 647                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 648                         struct pipe_buf_operations *ops = buf->ops;
 649
 650                         sd.len = buf->len;
 651                         if (sd.len > sd.total_len)
 652                                 sd.len = sd.total_len;
 653
 654                         err = actor(pipe, buf, &sd);
 655                         if (err) {
 656                                 if (!ret && err != -ENODATA)
 657                                         ret = err;
 658
 659                                 break;
 660                         }
 661
 662                         ret += sd.len;
 663                         buf->offset += sd.len;
 664                         buf->len -= sd.len;
 665
 666                         if (!buf->len) {
 667                                 buf->ops = NULL;
 668                                 ops->release(pipe, buf);
 669                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 670                                 pipe->nrbufs--;
 671                                 if (pipe->inode)
 672                                         do_wakeup = 1;
 673                         }
 674
 675                         sd.pos += sd.len;
 676                         sd.total_len -= sd.len;
 677                         if (!sd.total_len)
 678                                 break;
 679                 }
 680
 681                 if (pipe->nrbufs)
 682                         continue;
 683                 if (!pipe->writers)
 684                         break;
 685                 if (!pipe->waiting_writers) {
 686                         if (ret)
 687                                 break;
 688                 }
 689
 690                 if (flags & SPLICE_F_NONBLOCK) {
 691                         if (!ret)
 692                                 ret = -EAGAIN;
 693                         break;
 694                 }
 695
 696                 if (signal_pending(current)) {
 697                         if (!ret)
 698                                 ret = -ERESTARTSYS;
 699                         break;
 700                 }
 701
 702                 if (do_wakeup) {
 703                         smp_mb();
 704                         if (waitqueue_active(&pipe->wait))
 705                                 wake_up_interruptible_sync(&pipe->wait);
 706                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 707                         do_wakeup = 0;
 708                 }
 709
 710                 pipe_wait(pipe);
 711         }
 712
 713         if (pipe->inode)
 714                 mutex_unlock(&pipe->inode->i_mutex);
 715
 716         if (do_wakeup) {
 717                 smp_mb();
 718                 if (waitqueue_active(&pipe->wait))
 719                         wake_up_interruptible(&pipe->wait);
 720                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 721         }
 722
 723         return ret;
 724 }
 725
 726 /**
 727  * generic_file_splice_write - splice data from a pipe to a file
 728  * @pipe:       pipe info
 729  * @out:        file to write to
 730  * @len:        number of bytes to splice
 731  * @flags:      splice modifier flags
 732  *
 733  * Will either move or copy pages (determined by @flags options) from
 734  * the given pipe inode to the given file.
 735  *
 736  */
 737 ssize_t
 738 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 739                           loff_t *ppos, size_t len, unsigned int flags)
 740 {
 741         struct address_space *mapping = out->f_mapping;
 742         ssize_t ret;
 743
 744         ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 745         if (ret > 0) {
 746                 struct inode *inode = mapping->host;
 747
 748                 *ppos += ret;
 749
 750                 /*
 751                  * If file or inode is SYNC and we actually wrote some data,
 752                  * sync it.
 753                  */
 754                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 755                         int err;
 756
 757                         mutex_lock(&inode->i_mutex);
 758                         err = generic_osync_inode(inode, mapping,
 759                                                   OSYNC_METADATA|OSYNC_DATA);
 760                         mutex_unlock(&inode->i_mutex);
 761
 762                         if (err)
 763                                 ret = err;
 764                 }
 765         }
 766
 767         return ret;
 768 }
 769
 770 EXPORT_SYMBOL(generic_file_splice_write);
 771
 772 /**
 773  * generic_splice_sendpage - splice data from a pipe to a socket
 774  * @inode:      pipe inode
 775  * @out:        socket to write to
 776  * @len:        number of bytes to splice
 777  * @flags:      splice modifier flags
 778  *
 779  * Will send @len bytes from the pipe to a network socket. No data copying
 780  * is involved.
 781  *
 782  */
 783 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 784                                 loff_t *ppos, size_t len, unsigned int flags)
 785 {
 786         return move_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 787 }
 788
 789 EXPORT_SYMBOL(generic_splice_sendpage);
 790
 791 /*
 792  * Attempt to initiate a splice from pipe to file.
 793  */
 794 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 795                            loff_t *ppos, size_t len, unsigned int flags)
 796 {
 797         int ret;
 798
 799         if (unlikely(!out->f_op || !out->f_op->splice_write))
 800                 return -EINVAL;
 801
 802         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 803                 return -EBADF;
 804
 805         ret = rw_verify_area(WRITE, out, ppos, len);
 806         if (unlikely(ret < 0))
 807                 return ret;
 808
 809         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 810 }
 811
 812 /*
 813  * Attempt to initiate a splice from a file to a pipe.
 814  */
 815 static long do_splice_to(struct file *in, loff_t *ppos,
 816                          struct pipe_inode_info *pipe, size_t len,
 817                          unsigned int flags)
 818 {
 819         loff_t isize, left;
 820         int ret;
 821
 822         if (unlikely(!in->f_op || !in->f_op->splice_read))
 823                 return -EINVAL;
 824
 825         if (unlikely(!(in->f_mode & FMODE_READ)))
 826                 return -EBADF;
 827
 828         ret = rw_verify_area(READ, in, ppos, len);
 829         if (unlikely(ret < 0))
 830                 return ret;
 831
 832         isize = i_size_read(in->f_mapping->host);
 833         if (unlikely(*ppos >= isize))
 834                 return 0;
 835
 836         left = isize - *ppos;
 837         if (unlikely(left < len))
 838                 len = left;
 839
 840         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 841 }
 842
 843 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 844                       size_t len, unsigned int flags)
 845 {
 846         struct pipe_inode_info *pipe;
 847         long ret, bytes;
 848         loff_t out_off;
 849         umode_t i_mode;
 850         int i;
 851
 852         /*
 853          * We require the input being a regular file, as we don't want to
 854          * randomly drop data for eg socket -> socket splicing. Use the
 855          * piped splicing for that!
 856          */
 857         i_mode = in->f_dentry->d_inode->i_mode;
 858         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 859                 return -EINVAL;
 860
 861         /*
 862          * neither in nor out is a pipe, setup an internal pipe attached to
 863          * 'out' and transfer the wanted data from 'in' to 'out' through that
 864          */
 865         pipe = current->splice_pipe;
 866         if (unlikely(!pipe)) {
 867                 pipe = alloc_pipe_info(NULL);
 868                 if (!pipe)
 869                         return -ENOMEM;
 870
 871                 /*
 872                  * We don't have an immediate reader, but we'll read the stuff
 873                  * out of the pipe right after the move_to_pipe(). So set
 874                  * PIPE_READERS appropriately.
 875                  */
 876                 pipe->readers = 1;
 877
 878                 current->splice_pipe = pipe;
 879         }
 880
 881         /*
 882          * Do the splice.
 883          */
 884         ret = 0;
 885         bytes = 0;
 886         out_off = 0;
 887
 888         while (len) {
 889                 size_t read_len, max_read_len;
 890
 891                 /*
 892                  * Do at most PIPE_BUFFERS pages worth of transfer:
 893                  */
 894                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
 895
 896                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
 897                 if (unlikely(ret < 0))
 898                         goto out_release;
 899
 900                 read_len = ret;
 901
 902                 /*
 903                  * NOTE: nonblocking mode only applies to the input. We
 904                  * must not do the output in nonblocking mode as then we
 905                  * could get stuck data in the internal pipe:
 906                  */
 907                 ret = do_splice_from(pipe, out, &out_off, read_len,
 908                                      flags & ~SPLICE_F_NONBLOCK);
 909                 if (unlikely(ret < 0))
 910                         goto out_release;
 911
 912                 bytes += ret;
 913                 len -= ret;
 914
 915                 /*
 916                  * In nonblocking mode, if we got back a short read then
 917                  * that was due to either an IO error or due to the
 918                  * pagecache entry not being there. In the IO error case
 919                  * the _next_ splice attempt will produce a clean IO error
 920                  * return value (not a short read), so in both cases it's
 921                  * correct to break out of the loop here:
 922                  */
 923                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
 924                         break;
 925         }
 926
 927         pipe->nrbufs = pipe->curbuf = 0;
 928
 929         return bytes;
 930
 931 out_release:
 932         /*
 933          * If we did an incomplete transfer we must release
 934          * the pipe buffers in question:
 935          */
 936         for (i = 0; i < PIPE_BUFFERS; i++) {
 937                 struct pipe_buffer *buf = pipe->bufs + i;
 938
 939                 if (buf->ops) {
 940                         buf->ops->release(pipe, buf);
 941                         buf->ops = NULL;
 942                 }
 943         }
 944         pipe->nrbufs = pipe->curbuf = 0;
 945
 946         /*
 947          * If we transferred some data, return the number of bytes:
 948          */
 949         if (bytes > 0)
 950                 return bytes;
 951
 952         return ret;
 953 }
 954
 955 EXPORT_SYMBOL(do_splice_direct);
 956
 957 /*
 958  * Determine where to splice to/from.
 959  */
 960 static long do_splice(struct file *in, loff_t __user *off_in,
 961                       struct file *out, loff_t __user *off_out,
 962                       size_t len, unsigned int flags)
 963 {
 964         struct pipe_inode_info *pipe;
 965         loff_t offset, *off;
 966         long ret;
 967
 968         pipe = in->f_dentry->d_inode->i_pipe;
 969         if (pipe) {
 970                 if (off_in)
 971                         return -ESPIPE;
 972                 if (off_out) {
 973                         if (out->f_op->llseek == no_llseek)
 974                                 return -EINVAL;
 975                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
 976                                 return -EFAULT;
 977                         off = &offset;
 978                 } else
 979                         off = &out->f_pos;
 980
 981                 ret = do_splice_from(pipe, out, off, len, flags);
 982
 983                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
 984                         ret = -EFAULT;
 985
 986                 return ret;
 987         }
 988
 989         pipe = out->f_dentry->d_inode->i_pipe;
 990         if (pipe) {
 991                 if (off_out)
 992                         return -ESPIPE;
 993                 if (off_in) {
 994                         if (in->f_op->llseek == no_llseek)
 995                                 return -EINVAL;
 996                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
 997                                 return -EFAULT;
 998                         off = &offset;
 999                 } else
1000                         off = &in->f_pos;
1001
1002                 ret = do_splice_to(in, off, pipe, len, flags);
1003
1004                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1005                         ret = -EFAULT;
1006
1007                 return ret;
1008         }
1009
1010         return -EINVAL;
1011 }
1012
1013 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1014                            int fd_out, loff_t __user *off_out,
1015                            size_t len, unsigned int flags)
1016 {
1017         long error;
1018         struct file *in, *out;
1019         int fput_in, fput_out;
1020
1021         if (unlikely(!len))
1022                 return 0;
1023
1024         error = -EBADF;
1025         in = fget_light(fd_in, &fput_in);
1026         if (in) {
1027                 if (in->f_mode & FMODE_READ) {
1028                         out = fget_light(fd_out, &fput_out);
1029                         if (out) {
1030                                 if (out->f_mode & FMODE_WRITE)
1031                                         error = do_splice(in, off_in,
1032                                                           out, off_out,
1033                                                           len, flags);
1034                                 fput_light(out, fput_out);
1035                         }
1036                 }
1037
1038                 fput_light(in, fput_in);
1039         }
1040
1041         return error;
1042 }
1043
1044 /*
1045  * Link contents of ipipe to opipe.
1046  */
1047 static int link_pipe(struct pipe_inode_info *ipipe,
1048                      struct pipe_inode_info *opipe,
1049                      size_t len, unsigned int flags)
1050 {
1051         struct pipe_buffer *ibuf, *obuf;
1052         int ret, do_wakeup, i, ipipe_first;
1053
1054         ret = do_wakeup = ipipe_first = 0;
1055
1056         /*
1057          * Potential ABBA deadlock, work around it by ordering lock
1058          * grabbing by inode address. Otherwise two different processes
1059          * could deadlock (one doing tee from A -> B, the other from B -> A).
1060          */
1061         if (ipipe->inode < opipe->inode) {
1062                 ipipe_first = 1;
1063                 mutex_lock(&ipipe->inode->i_mutex);
1064                 mutex_lock(&opipe->inode->i_mutex);
1065         } else {
1066                 mutex_lock(&opipe->inode->i_mutex);
1067                 mutex_lock(&ipipe->inode->i_mutex);
1068         }
1069
1070         for (i = 0;; i++) {
1071                 if (!opipe->readers) {
1072                         send_sig(SIGPIPE, current, 0);
1073                         if (!ret)
1074                                 ret = -EPIPE;
1075                         break;
1076                 }
1077                 if (ipipe->nrbufs - i) {
1078                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1079
1080                         /*
1081                          * If we have room, fill this buffer
1082                          */
1083                         if (opipe->nrbufs < PIPE_BUFFERS) {
1084                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1085
1086                                 /*
1087                                  * Get a reference to this pipe buffer,
1088                                  * so we can copy the contents over.
1089                                  */
1090                                 ibuf->ops->get(ipipe, ibuf);
1091
1092                                 obuf = opipe->bufs + nbuf;
1093                                 *obuf = *ibuf;
1094
1095                                 if (obuf->len > len)
1096                                         obuf->len = len;
1097
1098                                 opipe->nrbufs++;
1099                                 do_wakeup = 1;
1100                                 ret += obuf->len;
1101                                 len -= obuf->len;
1102
1103                                 if (!len)
1104                                         break;
1105                                 if (opipe->nrbufs < PIPE_BUFFERS)
1106                                         continue;
1107                         }
1108
1109                         /*
1110                          * We have input available, but no output room.
1111                          * If we already copied data, return that. If we
1112                          * need to drop the opipe lock, it must be ordered
1113                          * last to avoid deadlocks.
1114                          */
1115                         if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
1116                                 if (!ret)
1117                                         ret = -EAGAIN;
1118                                 break;
1119                         }
1120                         if (signal_pending(current)) {
1121                                 if (!ret)
1122                                         ret = -ERESTARTSYS;
1123                                 break;
1124                         }
1125                         if (do_wakeup) {
1126                                 smp_mb();
1127                                 if (waitqueue_active(&opipe->wait))
1128                                         wake_up_interruptible(&opipe->wait);
1129                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1130                                 do_wakeup = 0;
1131                         }
1132
1133                         opipe->waiting_writers++;
1134                         pipe_wait(opipe);
1135                         opipe->waiting_writers--;
1136                         continue;
1137                 }
1138
1139                 /*
1140                  * No input buffers, do the usual checks for available
1141                  * writers and blocking and wait if necessary
1142                  */
1143                 if (!ipipe->writers)
1144                         break;
1145                 if (!ipipe->waiting_writers) {
1146                         if (ret)
1147                                 break;
1148                 }
1149                 /*
1150                  * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1151                  * with another process, we can only safely do that if
1152                  * the ipipe lock is ordered last.
1153                  */
1154                 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
1155                         if (!ret)
1156                                 ret = -EAGAIN;
1157                         break;
1158                 }
1159                 if (signal_pending(current)) {
1160                         if (!ret)
1161                                 ret = -ERESTARTSYS;
1162                         break;
1163                 }
1164
1165                 if (waitqueue_active(&ipipe->wait))
1166                         wake_up_interruptible_sync(&ipipe->wait);
1167                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1168
1169                 pipe_wait(ipipe);
1170         }
1171
1172         mutex_unlock(&ipipe->inode->i_mutex);
1173         mutex_unlock(&opipe->inode->i_mutex);
1174
1175         if (do_wakeup) {
1176                 smp_mb();
1177                 if (waitqueue_active(&opipe->wait))
1178                         wake_up_interruptible(&opipe->wait);
1179                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1180         }
1181
1182         return ret;
1183 }
1184
1185 /*
1186  * This is a tee(1) implementation that works on pipes. It doesn't copy
1187  * any data, it simply references the 'in' pages on the 'out' pipe.
1188  * The 'flags' used are the SPLICE_F_* variants, currently the only
1189  * applicable one is SPLICE_F_NONBLOCK.
1190  */
1191 static long do_tee(struct file *in, struct file *out, size_t len,
1192                    unsigned int flags)
1193 {
1194         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1195         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1196
1197         /*
1198          * Link ipipe to the two output pipes, consuming as we go along.
1199          */
1200         if (ipipe && opipe)
1201                 return link_pipe(ipipe, opipe, len, flags);
1202
1203         return -EINVAL;
1204 }
1205
1206 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1207 {
1208         struct file *in;
1209         int error, fput_in;
1210
1211         if (unlikely(!len))
1212                 return 0;
1213
1214         error = -EBADF;
1215         in = fget_light(fdin, &fput_in);
1216         if (in) {
1217                 if (in->f_mode & FMODE_READ) {
1218                         int fput_out;
1219                         struct file *out = fget_light(fdout, &fput_out);
1220
1221                         if (out) {
1222                                 if (out->f_mode & FMODE_WRITE)
1223                                         error = do_tee(in, out, len, flags);
1224                                 fput_light(out, fput_out);
1225                         }
1226                 }
1227                 fput_light(in, fput_in);
1228         }
1229
1230         return error;
1231 }