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