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@kernel.dk>
  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/splice.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 #include <linux/uio.h>
  31
  32 /*
  33  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  34  * a vm helper function, it's already simplified quite a bit by the
  35  * addition of remove_mapping(). If success is returned, the caller may
  36  * attempt to reuse this page for another destination.
  37  */
  38 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  39                                      struct pipe_buffer *buf)
  40 {
  41         struct page *page = buf->page;
  42         struct address_space *mapping;
  43
  44         lock_page(page);
  45
  46         mapping = page_mapping(page);
  47         if (mapping) {
  48                 WARN_ON(!PageUptodate(page));
  49
  50                 /*
  51                  * At least for ext2 with nobh option, we need to wait on
  52                  * writeback completing on this page, since we'll remove it
  53                  * from the pagecache.  Otherwise truncate wont wait on the
  54                  * page, allowing the disk blocks to be reused by someone else
  55                  * before we actually wrote our data to them. fs corruption
  56                  * ensues.
  57                  */
  58                 wait_on_page_writeback(page);
  59
  60                 if (PagePrivate(page))
  61                         try_to_release_page(page, GFP_KERNEL);
  62
  63                 /*
  64                  * If we succeeded in removing the mapping, set LRU flag
  65                  * and return good.
  66                  */
  67                 if (remove_mapping(mapping, page)) {
  68                         buf->flags |= PIPE_BUF_FLAG_LRU;
  69                         return 0;
  70                 }
  71         }
  72
  73         /*
  74          * Raced with truncate or failed to remove page from current
  75          * address space, unlock and return failure.
  76          */
  77         unlock_page(page);
  78         return 1;
  79 }
  80
  81 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  82                                         struct pipe_buffer *buf)
  83 {
  84         page_cache_release(buf->page);
  85         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  86 }
  87
  88 static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
  89                                    struct pipe_buffer *buf)
  90 {
  91         struct page *page = buf->page;
  92         int err;
  93
  94         if (!PageUptodate(page)) {
  95                 lock_page(page);
  96
  97                 /*
  98                  * Page got truncated/unhashed. This will cause a 0-byte
  99                  * splice, if this is the first page.
 100                  */
 101                 if (!page->mapping) {
 102                         err = -ENODATA;
 103                         goto error;
 104                 }
 105
 106                 /*
 107                  * Uh oh, read-error from disk.
 108                  */
 109                 if (!PageUptodate(page)) {
 110                         err = -EIO;
 111                         goto error;
 112                 }
 113
 114                 /*
 115                  * Page is ok afterall, we are done.
 116                  */
 117                 unlock_page(page);
 118         }
 119
 120         return 0;
 121 error:
 122         unlock_page(page);
 123         return err;
 124 }
 125
 126 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 127         .can_merge = 0,
 128         .map = generic_pipe_buf_map,
 129         .unmap = generic_pipe_buf_unmap,
 130         .pin = page_cache_pipe_buf_pin,
 131         .release = page_cache_pipe_buf_release,
 132         .steal = page_cache_pipe_buf_steal,
 133         .get = generic_pipe_buf_get,
 134 };
 135
 136 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 137                                     struct pipe_buffer *buf)
 138 {
 139         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 140                 return 1;
 141
 142         buf->flags |= PIPE_BUF_FLAG_LRU;
 143         return generic_pipe_buf_steal(pipe, buf);
 144 }
 145
 146 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 147         .can_merge = 0,
 148         .map = generic_pipe_buf_map,
 149         .unmap = generic_pipe_buf_unmap,
 150         .pin = generic_pipe_buf_pin,
 151         .release = page_cache_pipe_buf_release,
 152         .steal = user_page_pipe_buf_steal,
 153         .get = generic_pipe_buf_get,
 154 };
 155
 156 /*
 157  * Pipe output worker. This fills a pipe with the information contained
 158  * from splice_pipe_desc().
 159  */
 160 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 161                        struct splice_pipe_desc *spd)
 162 {
 163         unsigned int spd_pages = spd->nr_pages;
 164         int ret, do_wakeup, page_nr;
 165
 166         ret = 0;
 167         do_wakeup = 0;
 168         page_nr = 0;
 169
 170         if (pipe->inode)
 171                 mutex_lock(&pipe->inode->i_mutex);
 172
 173         for (;;) {
 174                 if (!pipe->readers) {
 175                         send_sig(SIGPIPE, current, 0);
 176                         if (!ret)
 177                                 ret = -EPIPE;
 178                         break;
 179                 }
 180
 181                 if (pipe->nrbufs < PIPE_BUFFERS) {
 182                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 183                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 184
 185                         buf->page = spd->pages[page_nr];
 186                         buf->offset = spd->partial[page_nr].offset;
 187                         buf->len = spd->partial[page_nr].len;
 188                         buf->private = spd->partial[page_nr].private;
 189                         buf->ops = spd->ops;
 190                         if (spd->flags & SPLICE_F_GIFT)
 191                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 192
 193                         pipe->nrbufs++;
 194                         page_nr++;
 195                         ret += buf->len;
 196
 197                         if (pipe->inode)
 198                                 do_wakeup = 1;
 199
 200                         if (!--spd->nr_pages)
 201                                 break;
 202                         if (pipe->nrbufs < PIPE_BUFFERS)
 203                                 continue;
 204
 205                         break;
 206                 }
 207
 208                 if (spd->flags & SPLICE_F_NONBLOCK) {
 209                         if (!ret)
 210                                 ret = -EAGAIN;
 211                         break;
 212                 }
 213
 214                 if (signal_pending(current)) {
 215                         if (!ret)
 216                                 ret = -ERESTARTSYS;
 217                         break;
 218                 }
 219
 220                 if (do_wakeup) {
 221                         smp_mb();
 222                         if (waitqueue_active(&pipe->wait))
 223                                 wake_up_interruptible_sync(&pipe->wait);
 224                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 225                         do_wakeup = 0;
 226                 }
 227
 228                 pipe->waiting_writers++;
 229                 pipe_wait(pipe);
 230                 pipe->waiting_writers--;
 231         }
 232
 233         if (pipe->inode) {
 234                 mutex_unlock(&pipe->inode->i_mutex);
 235
 236                 if (do_wakeup) {
 237                         smp_mb();
 238                         if (waitqueue_active(&pipe->wait))
 239                                 wake_up_interruptible(&pipe->wait);
 240                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 241                 }
 242         }
 243
 244         while (page_nr < spd_pages)
 245                 page_cache_release(spd->pages[page_nr++]);
 246
 247         return ret;
 248 }
 249
 250 static int
 251 __generic_file_splice_read(struct file *in, loff_t *ppos,
 252                            struct pipe_inode_info *pipe, size_t len,
 253                            unsigned int flags)
 254 {
 255         struct address_space *mapping = in->f_mapping;
 256         unsigned int loff, nr_pages;
 257         struct page *pages[PIPE_BUFFERS];
 258         struct partial_page partial[PIPE_BUFFERS];
 259         struct page *page;
 260         pgoff_t index, end_index;
 261         loff_t isize;
 262         int error, page_nr;
 263         struct splice_pipe_desc spd = {
 264                 .pages = pages,
 265                 .partial = partial,
 266                 .flags = flags,
 267                 .ops = &page_cache_pipe_buf_ops,
 268         };
 269
 270         index = *ppos >> PAGE_CACHE_SHIFT;
 271         loff = *ppos & ~PAGE_CACHE_MASK;
 272         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 273
 274         if (nr_pages > PIPE_BUFFERS)
 275                 nr_pages = PIPE_BUFFERS;
 276
 277         /*
 278          * Don't try to 2nd guess the read-ahead logic, call into
 279          * page_cache_readahead() like the page cache reads would do.
 280          */
 281         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
 282
 283         /*
 284          * Now fill in the holes:
 285          */
 286         error = 0;
 287
 288         /*
 289          * Lookup the (hopefully) full range of pages we need.
 290          */
 291         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 292
 293         /*
 294          * If find_get_pages_contig() returned fewer pages than we needed,
 295          * allocate the rest.
 296          */
 297         index += spd.nr_pages;
 298         while (spd.nr_pages < nr_pages) {
 299                 /*
 300                  * Page could be there, find_get_pages_contig() breaks on
 301                  * the first hole.
 302                  */
 303                 page = find_get_page(mapping, index);
 304                 if (!page) {
 305                         /*
 306                          * Make sure the read-ahead engine is notified
 307                          * about this failure.
 308                          */
 309                         handle_ra_miss(mapping, &in->f_ra, index);
 310
 311                         /*
 312                          * page didn't exist, allocate one.
 313                          */
 314                         page = page_cache_alloc_cold(mapping);
 315                         if (!page)
 316                                 break;
 317
 318                         error = add_to_page_cache_lru(page, mapping, index,
 319                                               GFP_KERNEL);
 320                         if (unlikely(error)) {
 321                                 page_cache_release(page);
 322                                 if (error == -EEXIST)
 323                                         continue;
 324                                 break;
 325                         }
 326                         /*
 327                          * add_to_page_cache() locks the page, unlock it
 328                          * to avoid convoluting the logic below even more.
 329                          */
 330                         unlock_page(page);
 331                 }
 332
 333                 pages[spd.nr_pages++] = page;
 334                 index++;
 335         }
 336
 337         /*
 338          * Now loop over the map and see if we need to start IO on any
 339          * pages, fill in the partial map, etc.
 340          */
 341         index = *ppos >> PAGE_CACHE_SHIFT;
 342         nr_pages = spd.nr_pages;
 343         spd.nr_pages = 0;
 344         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 345                 unsigned int this_len;
 346
 347                 if (!len)
 348                         break;
 349
 350                 /*
 351                  * this_len is the max we'll use from this page
 352                  */
 353                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 354                 page = pages[page_nr];
 355
 356                 /*
 357                  * If the page isn't uptodate, we may need to start io on it
 358                  */
 359                 if (!PageUptodate(page)) {
 360                         /*
 361                          * If in nonblock mode then dont block on waiting
 362                          * for an in-flight io page
 363                          */
 364                         if (flags & SPLICE_F_NONBLOCK) {
 365                                 if (TestSetPageLocked(page))
 366                                         break;
 367                         } else
 368                                 lock_page(page);
 369
 370                         /*
 371                          * page was truncated, stop here. if this isn't the
 372                          * first page, we'll just complete what we already
 373                          * added
 374                          */
 375                         if (!page->mapping) {
 376                                 unlock_page(page);
 377                                 break;
 378                         }
 379                         /*
 380                          * page was already under io and is now done, great
 381                          */
 382                         if (PageUptodate(page)) {
 383                                 unlock_page(page);
 384                                 goto fill_it;
 385                         }
 386
 387                         /*
 388                          * need to read in the page
 389                          */
 390                         error = mapping->a_ops->readpage(in, page);
 391                         if (unlikely(error)) {
 392                                 /*
 393                                  * We really should re-lookup the page here,
 394                                  * but it complicates things a lot. Instead
 395                                  * lets just do what we already stored, and
 396                                  * we'll get it the next time we are called.
 397                                  */
 398                                 if (error == AOP_TRUNCATED_PAGE)
 399                                         error = 0;
 400
 401                                 break;
 402                         }
 403                 }
 404 fill_it:
 405                 /*
 406                  * i_size must be checked after PageUptodate.
 407                  */
 408                 isize = i_size_read(mapping->host);
 409                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 410                 if (unlikely(!isize || index > end_index))
 411                         break;
 412
 413                 /*
 414                  * if this is the last page, see if we need to shrink
 415                  * the length and stop
 416                  */
 417                 if (end_index == index) {
 418                         unsigned int plen;
 419
 420                         /*
 421                          * max good bytes in this page
 422                          */
 423                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 424                         if (plen <= loff)
 425                                 break;
 426
 427                         /*
 428                          * force quit after adding this page
 429                          */
 430                         this_len = min(this_len, plen - loff);
 431                         len = this_len;
 432                 }
 433
 434                 partial[page_nr].offset = loff;
 435                 partial[page_nr].len = this_len;
 436                 len -= this_len;
 437                 loff = 0;
 438                 spd.nr_pages++;
 439                 index++;
 440         }
 441
 442         /*
 443          * Release any pages at the end, if we quit early. 'page_nr' is how far
 444          * we got, 'nr_pages' is how many pages are in the map.
 445          */
 446         while (page_nr < nr_pages)
 447                 page_cache_release(pages[page_nr++]);
 448
 449         if (spd.nr_pages)
 450                 return splice_to_pipe(pipe, &spd);
 451
 452         return error;
 453 }
 454
 455 /**
 456  * generic_file_splice_read - splice data from file to a pipe
 457  * @in:         file to splice from
 458  * @pipe:       pipe to splice to
 459  * @len:        number of bytes to splice
 460  * @flags:      splice modifier flags
 461  *
 462  * Will read pages from given file and fill them into a pipe.
 463  */
 464 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 465                                  struct pipe_inode_info *pipe, size_t len,
 466                                  unsigned int flags)
 467 {
 468         ssize_t spliced;
 469         int ret;
 470         loff_t isize, left;
 471
 472         isize = i_size_read(in->f_mapping->host);
 473         if (unlikely(*ppos >= isize))
 474                 return 0;
 475
 476         left = isize - *ppos;
 477         if (unlikely(left < len))
 478                 len = left;
 479
 480         ret = 0;
 481         spliced = 0;
 482         while (len) {
 483                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 484
 485                 if (ret < 0)
 486                         break;
 487                 else if (!ret) {
 488                         if (spliced)
 489                                 break;
 490                         if (flags & SPLICE_F_NONBLOCK) {
 491                                 ret = -EAGAIN;
 492                                 break;
 493                         }
 494                 }
 495
 496                 *ppos += ret;
 497                 len -= ret;
 498                 spliced += ret;
 499         }
 500
 501         if (spliced)
 502                 return spliced;
 503
 504         return ret;
 505 }
 506
 507 EXPORT_SYMBOL(generic_file_splice_read);
 508
 509 /*
 510  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 511  * using sendpage(). Return the number of bytes sent.
 512  */
 513 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 514                             struct pipe_buffer *buf, struct splice_desc *sd)
 515 {
 516         struct file *file = sd->u.file;
 517         loff_t pos = sd->pos;
 518         int ret, more;
 519
 520         ret = buf->ops->pin(pipe, buf);
 521         if (!ret) {
 522                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 523
 524                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 525                                            sd->len, &pos, more);
 526         }
 527
 528         return ret;
 529 }
 530
 531 /*
 532  * This is a little more tricky than the file -> pipe splicing. There are
 533  * basically three cases:
 534  *
 535  *      - Destination page already exists in the address space and there
 536  *        are users of it. For that case we have no other option that
 537  *        copying the data. Tough luck.
 538  *      - Destination page already exists in the address space, but there
 539  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 540  *        through to last case.
 541  *      - Destination page does not exist, we can add the pipe page to
 542  *        the page cache and avoid the copy.
 543  *
 544  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 545  * sd->flags), we attempt to migrate pages from the pipe to the output
 546  * file address space page cache. This is possible if no one else has
 547  * the pipe page referenced outside of the pipe and page cache. If
 548  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 549  * a new page in the output file page cache and fill/dirty that.
 550  */
 551 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 552                         struct splice_desc *sd)
 553 {
 554         struct file *file = sd->u.file;
 555         struct address_space *mapping = file->f_mapping;
 556         unsigned int offset, this_len;
 557         struct page *page;
 558         pgoff_t index;
 559         int ret;
 560
 561         /*
 562          * make sure the data in this buffer is uptodate
 563          */
 564         ret = buf->ops->pin(pipe, buf);
 565         if (unlikely(ret))
 566                 return ret;
 567
 568         index = sd->pos >> PAGE_CACHE_SHIFT;
 569         offset = sd->pos & ~PAGE_CACHE_MASK;
 570
 571         this_len = sd->len;
 572         if (this_len + offset > PAGE_CACHE_SIZE)
 573                 this_len = PAGE_CACHE_SIZE - offset;
 574
 575 find_page:
 576         page = find_lock_page(mapping, index);
 577         if (!page) {
 578                 ret = -ENOMEM;
 579                 page = page_cache_alloc_cold(mapping);
 580                 if (unlikely(!page))
 581                         goto out_ret;
 582
 583                 /*
 584                  * This will also lock the page
 585                  */
 586                 ret = add_to_page_cache_lru(page, mapping, index,
 587                                             GFP_KERNEL);
 588                 if (unlikely(ret))
 589                         goto out;
 590         }
 591
 592         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
 593         if (unlikely(ret)) {
 594                 loff_t isize = i_size_read(mapping->host);
 595
 596                 if (ret != AOP_TRUNCATED_PAGE)
 597                         unlock_page(page);
 598                 page_cache_release(page);
 599                 if (ret == AOP_TRUNCATED_PAGE)
 600                         goto find_page;
 601
 602                 /*
 603                  * prepare_write() may have instantiated a few blocks
 604                  * outside i_size.  Trim these off again.
 605                  */
 606                 if (sd->pos + this_len > isize)
 607                         vmtruncate(mapping->host, isize);
 608
 609                 goto out_ret;
 610         }
 611
 612         if (buf->page != page) {
 613                 /*
 614                  * Careful, ->map() uses KM_USER0!
 615                  */
 616                 char *src = buf->ops->map(pipe, buf, 1);
 617                 char *dst = kmap_atomic(page, KM_USER1);
 618
 619                 memcpy(dst + offset, src + buf->offset, this_len);
 620                 flush_dcache_page(page);
 621                 kunmap_atomic(dst, KM_USER1);
 622                 buf->ops->unmap(pipe, buf, src);
 623         }
 624
 625         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
 626         if (ret) {
 627                 if (ret == AOP_TRUNCATED_PAGE) {
 628                         page_cache_release(page);
 629                         goto find_page;
 630                 }
 631                 if (ret < 0)
 632                         goto out;
 633                 /*
 634                  * Partial write has happened, so 'ret' already initialized by
 635                  * number of bytes written, Where is nothing we have to do here.
 636                  */
 637         } else
 638                 ret = this_len;
 639         /*
 640          * Return the number of bytes written and mark page as
 641          * accessed, we are now done!
 642          */
 643         mark_page_accessed(page);
 644 out:
 645         page_cache_release(page);
 646         unlock_page(page);
 647 out_ret:
 648         return ret;
 649 }
 650
 651 /*
 652  * Pipe input worker. Most of this logic works like a regular pipe, the
 653  * key here is the 'actor' worker passed in that actually moves the data
 654  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
 655  */
 656 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 657                            splice_actor *actor)
 658 {
 659         int ret, do_wakeup, err;
 660
 661         ret = 0;
 662         do_wakeup = 0;
 663
 664         for (;;) {
 665                 if (pipe->nrbufs) {
 666                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 667                         const struct pipe_buf_operations *ops = buf->ops;
 668
 669                         sd->len = buf->len;
 670                         if (sd->len > sd->total_len)
 671                                 sd->len = sd->total_len;
 672
 673                         err = actor(pipe, buf, sd);
 674                         if (err <= 0) {
 675                                 if (!ret && err != -ENODATA)
 676                                         ret = err;
 677
 678                                 break;
 679                         }
 680
 681                         ret += err;
 682                         buf->offset += err;
 683                         buf->len -= err;
 684
 685                         sd->len -= err;
 686                         sd->pos += err;
 687                         sd->total_len -= err;
 688                         if (sd->len)
 689                                 continue;
 690
 691                         if (!buf->len) {
 692                                 buf->ops = NULL;
 693                                 ops->release(pipe, buf);
 694                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 695                                 pipe->nrbufs--;
 696                                 if (pipe->inode)
 697                                         do_wakeup = 1;
 698                         }
 699
 700                         if (!sd->total_len)
 701                                 break;
 702                 }
 703
 704                 if (pipe->nrbufs)
 705                         continue;
 706                 if (!pipe->writers)
 707                         break;
 708                 if (!pipe->waiting_writers) {
 709                         if (ret)
 710                                 break;
 711                 }
 712
 713                 if (sd->flags & SPLICE_F_NONBLOCK) {
 714                         if (!ret)
 715                                 ret = -EAGAIN;
 716                         break;
 717                 }
 718
 719                 if (signal_pending(current)) {
 720                         if (!ret)
 721                                 ret = -ERESTARTSYS;
 722                         break;
 723                 }
 724
 725                 if (do_wakeup) {
 726                         smp_mb();
 727                         if (waitqueue_active(&pipe->wait))
 728                                 wake_up_interruptible_sync(&pipe->wait);
 729                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 730                         do_wakeup = 0;
 731                 }
 732
 733                 pipe_wait(pipe);
 734         }
 735
 736         if (do_wakeup) {
 737                 smp_mb();
 738                 if (waitqueue_active(&pipe->wait))
 739                         wake_up_interruptible(&pipe->wait);
 740                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 741         }
 742
 743         return ret;
 744 }
 745 EXPORT_SYMBOL(__splice_from_pipe);
 746
 747 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 748                          loff_t *ppos, size_t len, unsigned int flags,
 749                          splice_actor *actor)
 750 {
 751         ssize_t ret;
 752         struct inode *inode = out->f_mapping->host;
 753         struct splice_desc sd = {
 754                 .total_len = len,
 755                 .flags = flags,
 756                 .pos = *ppos,
 757                 .u.file = out,
 758         };
 759
 760         /*
 761          * The actor worker might be calling ->prepare_write and
 762          * ->commit_write. Most of the time, these expect i_mutex to
 763          * be held. Since this may result in an ABBA deadlock with
 764          * pipe->inode, we have to order lock acquiry here.
 765          */
 766         inode_double_lock(inode, pipe->inode);
 767         ret = __splice_from_pipe(pipe, &sd, actor);
 768         inode_double_unlock(inode, pipe->inode);
 769
 770         return ret;
 771 }
 772
 773 /**
 774  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 775  * @pipe:       pipe info
 776  * @out:        file to write to
 777  * @len:        number of bytes to splice
 778  * @flags:      splice modifier flags
 779  *
 780  * Will either move or copy pages (determined by @flags options) from
 781  * the given pipe inode to the given file. The caller is responsible
 782  * for acquiring i_mutex on both inodes.
 783  *
 784  */
 785 ssize_t
 786 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 787                                  loff_t *ppos, size_t len, unsigned int flags)
 788 {
 789         struct address_space *mapping = out->f_mapping;
 790         struct inode *inode = mapping->host;
 791         struct splice_desc sd = {
 792                 .total_len = len,
 793                 .flags = flags,
 794                 .pos = *ppos,
 795                 .u.file = out,
 796         };
 797         ssize_t ret;
 798         int err;
 799
 800         err = remove_suid(out->f_path.dentry);
 801         if (unlikely(err))
 802                 return err;
 803
 804         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
 805         if (ret > 0) {
 806                 unsigned long nr_pages;
 807
 808                 *ppos += ret;
 809                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 810
 811                 /*
 812                  * If file or inode is SYNC and we actually wrote some data,
 813                  * sync it.
 814                  */
 815                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 816                         err = generic_osync_inode(inode, mapping,
 817                                                   OSYNC_METADATA|OSYNC_DATA);
 818
 819                         if (err)
 820                                 ret = err;
 821                 }
 822                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 823         }
 824
 825         return ret;
 826 }
 827
 828 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 829
 830 /**
 831  * generic_file_splice_write - splice data from a pipe to a file
 832  * @pipe:       pipe info
 833  * @out:        file to write to
 834  * @len:        number of bytes to splice
 835  * @flags:      splice modifier flags
 836  *
 837  * Will either move or copy pages (determined by @flags options) from
 838  * the given pipe inode to the given file.
 839  *
 840  */
 841 ssize_t
 842 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 843                           loff_t *ppos, size_t len, unsigned int flags)
 844 {
 845         struct address_space *mapping = out->f_mapping;
 846         struct inode *inode = mapping->host;
 847         ssize_t ret;
 848         int err;
 849
 850         err = should_remove_suid(out->f_path.dentry);
 851         if (unlikely(err)) {
 852                 mutex_lock(&inode->i_mutex);
 853                 err = __remove_suid(out->f_path.dentry, err);
 854                 mutex_unlock(&inode->i_mutex);
 855                 if (err)
 856                         return err;
 857         }
 858
 859         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
 860         if (ret > 0) {
 861                 unsigned long nr_pages;
 862
 863                 *ppos += ret;
 864                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 865
 866                 /*
 867                  * If file or inode is SYNC and we actually wrote some data,
 868                  * sync it.
 869                  */
 870                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 871                         mutex_lock(&inode->i_mutex);
 872                         err = generic_osync_inode(inode, mapping,
 873                                                   OSYNC_METADATA|OSYNC_DATA);
 874                         mutex_unlock(&inode->i_mutex);
 875
 876                         if (err)
 877                                 ret = err;
 878                 }
 879                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 880         }
 881
 882         return ret;
 883 }
 884
 885 EXPORT_SYMBOL(generic_file_splice_write);
 886
 887 /**
 888  * generic_splice_sendpage - splice data from a pipe to a socket
 889  * @inode:      pipe inode
 890  * @out:        socket to write to
 891  * @len:        number of bytes to splice
 892  * @flags:      splice modifier flags
 893  *
 894  * Will send @len bytes from the pipe to a network socket. No data copying
 895  * is involved.
 896  *
 897  */
 898 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 899                                 loff_t *ppos, size_t len, unsigned int flags)
 900 {
 901         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 902 }
 903
 904 EXPORT_SYMBOL(generic_splice_sendpage);
 905
 906 /*
 907  * Attempt to initiate a splice from pipe to file.
 908  */
 909 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 910                            loff_t *ppos, size_t len, unsigned int flags)
 911 {
 912         int ret;
 913
 914         if (unlikely(!out->f_op || !out->f_op->splice_write))
 915                 return -EINVAL;
 916
 917         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 918                 return -EBADF;
 919
 920         ret = rw_verify_area(WRITE, out, ppos, len);
 921         if (unlikely(ret < 0))
 922                 return ret;
 923
 924         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 925 }
 926
 927 /*
 928  * Attempt to initiate a splice from a file to a pipe.
 929  */
 930 static long do_splice_to(struct file *in, loff_t *ppos,
 931                          struct pipe_inode_info *pipe, size_t len,
 932                          unsigned int flags)
 933 {
 934         int ret;
 935
 936         if (unlikely(!in->f_op || !in->f_op->splice_read))
 937                 return -EINVAL;
 938
 939         if (unlikely(!(in->f_mode & FMODE_READ)))
 940                 return -EBADF;
 941
 942         ret = rw_verify_area(READ, in, ppos, len);
 943         if (unlikely(ret < 0))
 944                 return ret;
 945
 946         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 947 }
 948
 949 /*
 950  * Splices from an input file to an actor, using a 'direct' pipe.
 951  */
 952 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 953                                splice_direct_actor *actor)
 954 {
 955         struct pipe_inode_info *pipe;
 956         long ret, bytes;
 957         umode_t i_mode;
 958         size_t len;
 959         int i, flags;
 960
 961         /*
 962          * We require the input being a regular file, as we don't want to
 963          * randomly drop data for eg socket -> socket splicing. Use the
 964          * piped splicing for that!
 965          */
 966         i_mode = in->f_path.dentry->d_inode->i_mode;
 967         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 968                 return -EINVAL;
 969
 970         /*
 971          * neither in nor out is a pipe, setup an internal pipe attached to
 972          * 'out' and transfer the wanted data from 'in' to 'out' through that
 973          */
 974         pipe = current->splice_pipe;
 975         if (unlikely(!pipe)) {
 976                 pipe = alloc_pipe_info(NULL);
 977                 if (!pipe)
 978                         return -ENOMEM;
 979
 980                 /*
 981                  * We don't have an immediate reader, but we'll read the stuff
 982                  * out of the pipe right after the splice_to_pipe(). So set
 983                  * PIPE_READERS appropriately.
 984                  */
 985                 pipe->readers = 1;
 986
 987                 current->splice_pipe = pipe;
 988         }
 989
 990         /*
 991          * Do the splice.
 992          */
 993         ret = 0;
 994         bytes = 0;
 995         len = sd->total_len;
 996         flags = sd->flags;
 997
 998         /*
 999          * Don't block on output, we have to drain the direct pipe.
1000          */
1001         sd->flags &= ~SPLICE_F_NONBLOCK;
1002
1003         while (len) {
1004                 size_t read_len, max_read_len;
1005
1006                 /*
1007                  * Do at most PIPE_BUFFERS pages worth of transfer:
1008                  */
1009                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1010
1011                 ret = do_splice_to(in, &sd->pos, pipe, max_read_len, flags);
1012                 if (unlikely(ret < 0))
1013                         goto out_release;
1014
1015                 read_len = ret;
1016                 sd->total_len = read_len;
1017
1018                 /*
1019                  * NOTE: nonblocking mode only applies to the input. We
1020                  * must not do the output in nonblocking mode as then we
1021                  * could get stuck data in the internal pipe:
1022                  */
1023                 ret = actor(pipe, sd);
1024                 if (unlikely(ret < 0))
1025                         goto out_release;
1026
1027                 bytes += ret;
1028                 len -= ret;
1029
1030                 /*
1031                  * In nonblocking mode, if we got back a short read then
1032                  * that was due to either an IO error or due to the
1033                  * pagecache entry not being there. In the IO error case
1034                  * the _next_ splice attempt will produce a clean IO error
1035                  * return value (not a short read), so in both cases it's
1036                  * correct to break out of the loop here:
1037                  */
1038                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1039                         break;
1040         }
1041
1042         pipe->nrbufs = pipe->curbuf = 0;
1043
1044         return bytes;
1045
1046 out_release:
1047         /*
1048          * If we did an incomplete transfer we must release
1049          * the pipe buffers in question:
1050          */
1051         for (i = 0; i < PIPE_BUFFERS; i++) {
1052                 struct pipe_buffer *buf = pipe->bufs + i;
1053
1054                 if (buf->ops) {
1055                         buf->ops->release(pipe, buf);
1056                         buf->ops = NULL;
1057                 }
1058         }
1059         pipe->nrbufs = pipe->curbuf = 0;
1060
1061         /*
1062          * If we transferred some data, return the number of bytes:
1063          */
1064         if (bytes > 0)
1065                 return bytes;
1066
1067         return ret;
1068
1069 }
1070 EXPORT_SYMBOL(splice_direct_to_actor);
1071
1072 static int direct_splice_actor(struct pipe_inode_info *pipe,
1073                                struct splice_desc *sd)
1074 {
1075         struct file *file = sd->u.file;
1076
1077         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1078 }
1079
1080 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1081                       size_t len, unsigned int flags)
1082 {
1083         struct splice_desc sd = {
1084                 .len            = len,
1085                 .total_len      = len,
1086                 .flags          = flags,
1087                 .pos            = *ppos,
1088                 .u.file         = out,
1089         };
1090         size_t ret;
1091
1092         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1093         *ppos = sd.pos;
1094         return ret;
1095 }
1096
1097 /*
1098  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1099  * location, so checking ->i_pipe is not enough to verify that this is a
1100  * pipe.
1101  */
1102 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1103 {
1104         if (S_ISFIFO(inode->i_mode))
1105                 return inode->i_pipe;
1106
1107         return NULL;
1108 }
1109
1110 /*
1111  * Determine where to splice to/from.
1112  */
1113 static long do_splice(struct file *in, loff_t __user *off_in,
1114                       struct file *out, loff_t __user *off_out,
1115                       size_t len, unsigned int flags)
1116 {
1117         struct pipe_inode_info *pipe;
1118         loff_t offset, *off;
1119         long ret;
1120
1121         pipe = pipe_info(in->f_path.dentry->d_inode);
1122         if (pipe) {
1123                 if (off_in)
1124                         return -ESPIPE;
1125                 if (off_out) {
1126                         if (out->f_op->llseek == no_llseek)
1127                                 return -EINVAL;
1128                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1129                                 return -EFAULT;
1130                         off = &offset;
1131                 } else
1132                         off = &out->f_pos;
1133
1134                 ret = do_splice_from(pipe, out, off, len, flags);
1135
1136                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1137                         ret = -EFAULT;
1138
1139                 return ret;
1140         }
1141
1142         pipe = pipe_info(out->f_path.dentry->d_inode);
1143         if (pipe) {
1144                 if (off_out)
1145                         return -ESPIPE;
1146                 if (off_in) {
1147                         if (in->f_op->llseek == no_llseek)
1148                                 return -EINVAL;
1149                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1150                                 return -EFAULT;
1151                         off = &offset;
1152                 } else
1153                         off = &in->f_pos;
1154
1155                 ret = do_splice_to(in, off, pipe, len, flags);
1156
1157                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1158                         ret = -EFAULT;
1159
1160                 return ret;
1161         }
1162
1163         return -EINVAL;
1164 }
1165
1166 /*
1167  * Map an iov into an array of pages and offset/length tupples. With the
1168  * partial_page structure, we can map several non-contiguous ranges into
1169  * our ones pages[] map instead of splitting that operation into pieces.
1170  * Could easily be exported as a generic helper for other users, in which
1171  * case one would probably want to add a 'max_nr_pages' parameter as well.
1172  */
1173 static int get_iovec_page_array(const struct iovec __user *iov,
1174                                 unsigned int nr_vecs, struct page **pages,
1175                                 struct partial_page *partial, int aligned)
1176 {
1177         int buffers = 0, error = 0;
1178
1179         /*
1180          * It's ok to take the mmap_sem for reading, even
1181          * across a "get_user()".
1182          */
1183         down_read(&current->mm->mmap_sem);
1184
1185         while (nr_vecs) {
1186                 unsigned long off, npages;
1187                 void __user *base;
1188                 size_t len;
1189                 int i;
1190
1191                 /*
1192                  * Get user address base and length for this iovec.
1193                  */
1194                 error = get_user(base, &iov->iov_base);
1195                 if (unlikely(error))
1196                         break;
1197                 error = get_user(len, &iov->iov_len);
1198                 if (unlikely(error))
1199                         break;
1200
1201                 /*
1202                  * Sanity check this iovec. 0 read succeeds.
1203                  */
1204                 if (unlikely(!len))
1205                         break;
1206                 error = -EFAULT;
1207                 if (unlikely(!base))
1208                         break;
1209
1210                 /*
1211                  * Get this base offset and number of pages, then map
1212                  * in the user pages.
1213                  */
1214                 off = (unsigned long) base & ~PAGE_MASK;
1215
1216                 /*
1217                  * If asked for alignment, the offset must be zero and the
1218                  * length a multiple of the PAGE_SIZE.
1219                  */
1220                 error = -EINVAL;
1221                 if (aligned && (off || len & ~PAGE_MASK))
1222                         break;
1223
1224                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1225                 if (npages > PIPE_BUFFERS - buffers)
1226                         npages = PIPE_BUFFERS - buffers;
1227
1228                 error = get_user_pages(current, current->mm,
1229                                        (unsigned long) base, npages, 0, 0,
1230                                        &pages[buffers], NULL);
1231
1232                 if (unlikely(error <= 0))
1233                         break;
1234
1235                 /*
1236                  * Fill this contiguous range into the partial page map.
1237                  */
1238                 for (i = 0; i < error; i++) {
1239                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1240
1241                         partial[buffers].offset = off;
1242                         partial[buffers].len = plen;
1243
1244                         off = 0;
1245                         len -= plen;
1246                         buffers++;
1247                 }
1248
1249                 /*
1250                  * We didn't complete this iov, stop here since it probably
1251                  * means we have to move some of this into a pipe to
1252                  * be able to continue.
1253                  */
1254                 if (len)
1255                         break;
1256
1257                 /*
1258                  * Don't continue if we mapped fewer pages than we asked for,
1259                  * or if we mapped the max number of pages that we have
1260                  * room for.
1261                  */
1262                 if (error < npages || buffers == PIPE_BUFFERS)
1263                         break;
1264
1265                 nr_vecs--;
1266                 iov++;
1267         }
1268
1269         up_read(&current->mm->mmap_sem);
1270
1271         if (buffers)
1272                 return buffers;
1273
1274         return error;
1275 }
1276
1277 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1278                         struct splice_desc *sd)
1279 {
1280         char *src;
1281         int ret;
1282
1283         ret = buf->ops->pin(pipe, buf);
1284         if (unlikely(ret))
1285                 return ret;
1286
1287         /*
1288          * See if we can use the atomic maps, by prefaulting in the
1289          * pages and doing an atomic copy
1290          */
1291         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1292                 src = buf->ops->map(pipe, buf, 1);
1293                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1294                                                         sd->len);
1295                 buf->ops->unmap(pipe, buf, src);
1296                 if (!ret) {
1297                         ret = sd->len;
1298                         goto out;
1299                 }
1300         }
1301
1302         /*
1303          * No dice, use slow non-atomic map and copy
1304          */
1305         src = buf->ops->map(pipe, buf, 0);
1306
1307         ret = sd->len;
1308         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1309                 ret = -EFAULT;
1310
1311 out:
1312         if (ret > 0)
1313                 sd->u.userptr += ret;
1314         buf->ops->unmap(pipe, buf, src);
1315         return ret;
1316 }
1317
1318 /*
1319  * For lack of a better implementation, implement vmsplice() to userspace
1320  * as a simple copy of the pipes pages to the user iov.
1321  */
1322 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1323                              unsigned long nr_segs, unsigned int flags)
1324 {
1325         struct pipe_inode_info *pipe;
1326         struct splice_desc sd;
1327         ssize_t size;
1328         int error;
1329         long ret;
1330
1331         pipe = pipe_info(file->f_path.dentry->d_inode);
1332         if (!pipe)
1333                 return -EBADF;
1334
1335         if (pipe->inode)
1336                 mutex_lock(&pipe->inode->i_mutex);
1337
1338         error = ret = 0;
1339         while (nr_segs) {
1340                 void __user *base;
1341                 size_t len;
1342
1343                 /*
1344                  * Get user address base and length for this iovec.
1345                  */
1346                 error = get_user(base, &iov->iov_base);
1347                 if (unlikely(error))
1348                         break;
1349                 error = get_user(len, &iov->iov_len);
1350                 if (unlikely(error))
1351                         break;
1352
1353                 /*
1354                  * Sanity check this iovec. 0 read succeeds.
1355                  */
1356                 if (unlikely(!len))
1357                         break;
1358                 if (unlikely(!base)) {
1359                         error = -EFAULT;
1360                         break;
1361                 }
1362
1363                 sd.len = 0;
1364                 sd.total_len = len;
1365                 sd.flags = flags;
1366                 sd.u.userptr = base;
1367                 sd.pos = 0;
1368
1369                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1370                 if (size < 0) {
1371                         if (!ret)
1372                                 ret = size;
1373
1374                         break;
1375                 }
1376
1377                 ret += size;
1378
1379                 if (size < len)
1380                         break;
1381
1382                 nr_segs--;
1383                 iov++;
1384         }
1385
1386         if (pipe->inode)
1387                 mutex_unlock(&pipe->inode->i_mutex);
1388
1389         if (!ret)
1390                 ret = error;
1391
1392         return ret;
1393 }
1394
1395 /*
1396  * vmsplice splices a user address range into a pipe. It can be thought of
1397  * as splice-from-memory, where the regular splice is splice-from-file (or
1398  * to file). In both cases the output is a pipe, naturally.
1399  */
1400 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1401                              unsigned long nr_segs, unsigned int flags)
1402 {
1403         struct pipe_inode_info *pipe;
1404         struct page *pages[PIPE_BUFFERS];
1405         struct partial_page partial[PIPE_BUFFERS];
1406         struct splice_pipe_desc spd = {
1407                 .pages = pages,
1408                 .partial = partial,
1409                 .flags = flags,
1410                 .ops = &user_page_pipe_buf_ops,
1411         };
1412
1413         pipe = pipe_info(file->f_path.dentry->d_inode);
1414         if (!pipe)
1415                 return -EBADF;
1416
1417         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1418                                             flags & SPLICE_F_GIFT);
1419         if (spd.nr_pages <= 0)
1420                 return spd.nr_pages;
1421
1422         return splice_to_pipe(pipe, &spd);
1423 }
1424
1425 /*
1426  * Note that vmsplice only really supports true splicing _from_ user memory
1427  * to a pipe, not the other way around. Splicing from user memory is a simple
1428  * operation that can be supported without any funky alignment restrictions
1429  * or nasty vm tricks. We simply map in the user memory and fill them into
1430  * a pipe. The reverse isn't quite as easy, though. There are two possible
1431  * solutions for that:
1432  *
1433  *      - memcpy() the data internally, at which point we might as well just
1434  *        do a regular read() on the buffer anyway.
1435  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1436  *        has restriction limitations on both ends of the pipe).
1437  *
1438  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1439  *
1440  */
1441 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1442                              unsigned long nr_segs, unsigned int flags)
1443 {
1444         struct file *file;
1445         long error;
1446         int fput;
1447
1448         if (unlikely(nr_segs > UIO_MAXIOV))
1449                 return -EINVAL;
1450         else if (unlikely(!nr_segs))
1451                 return 0;
1452
1453         error = -EBADF;
1454         file = fget_light(fd, &fput);
1455         if (file) {
1456                 if (file->f_mode & FMODE_WRITE)
1457                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1458                 else if (file->f_mode & FMODE_READ)
1459                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1460
1461                 fput_light(file, fput);
1462         }
1463
1464         return error;
1465 }
1466
1467 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1468                            int fd_out, loff_t __user *off_out,
1469                            size_t len, unsigned int flags)
1470 {
1471         long error;
1472         struct file *in, *out;
1473         int fput_in, fput_out;
1474
1475         if (unlikely(!len))
1476                 return 0;
1477
1478         error = -EBADF;
1479         in = fget_light(fd_in, &fput_in);
1480         if (in) {
1481                 if (in->f_mode & FMODE_READ) {
1482                         out = fget_light(fd_out, &fput_out);
1483                         if (out) {
1484                                 if (out->f_mode & FMODE_WRITE)
1485                                         error = do_splice(in, off_in,
1486                                                           out, off_out,
1487                                                           len, flags);
1488                                 fput_light(out, fput_out);
1489                         }
1490                 }
1491
1492                 fput_light(in, fput_in);
1493         }
1494
1495         return error;
1496 }
1497
1498 /*
1499  * Make sure there's data to read. Wait for input if we can, otherwise
1500  * return an appropriate error.
1501  */
1502 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1503 {
1504         int ret;
1505
1506         /*
1507          * Check ->nrbufs without the inode lock first. This function
1508          * is speculative anyways, so missing one is ok.
1509          */
1510         if (pipe->nrbufs)
1511                 return 0;
1512
1513         ret = 0;
1514         mutex_lock(&pipe->inode->i_mutex);
1515
1516         while (!pipe->nrbufs) {
1517                 if (signal_pending(current)) {
1518                         ret = -ERESTARTSYS;
1519                         break;
1520                 }
1521                 if (!pipe->writers)
1522                         break;
1523                 if (!pipe->waiting_writers) {
1524                         if (flags & SPLICE_F_NONBLOCK) {
1525                                 ret = -EAGAIN;
1526                                 break;
1527                         }
1528                 }
1529                 pipe_wait(pipe);
1530         }
1531
1532         mutex_unlock(&pipe->inode->i_mutex);
1533         return ret;
1534 }
1535
1536 /*
1537  * Make sure there's writeable room. Wait for room if we can, otherwise
1538  * return an appropriate error.
1539  */
1540 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1541 {
1542         int ret;
1543
1544         /*
1545          * Check ->nrbufs without the inode lock first. This function
1546          * is speculative anyways, so missing one is ok.
1547          */
1548         if (pipe->nrbufs < PIPE_BUFFERS)
1549                 return 0;
1550
1551         ret = 0;
1552         mutex_lock(&pipe->inode->i_mutex);
1553
1554         while (pipe->nrbufs >= PIPE_BUFFERS) {
1555                 if (!pipe->readers) {
1556                         send_sig(SIGPIPE, current, 0);
1557                         ret = -EPIPE;
1558                         break;
1559                 }
1560                 if (flags & SPLICE_F_NONBLOCK) {
1561                         ret = -EAGAIN;
1562                         break;
1563                 }
1564                 if (signal_pending(current)) {
1565                         ret = -ERESTARTSYS;
1566                         break;
1567                 }
1568                 pipe->waiting_writers++;
1569                 pipe_wait(pipe);
1570                 pipe->waiting_writers--;
1571         }
1572
1573         mutex_unlock(&pipe->inode->i_mutex);
1574         return ret;
1575 }
1576
1577 /*
1578  * Link contents of ipipe to opipe.
1579  */
1580 static int link_pipe(struct pipe_inode_info *ipipe,
1581                      struct pipe_inode_info *opipe,
1582                      size_t len, unsigned int flags)
1583 {
1584         struct pipe_buffer *ibuf, *obuf;
1585         int ret = 0, i = 0, nbuf;
1586
1587         /*
1588          * Potential ABBA deadlock, work around it by ordering lock
1589          * grabbing by inode address. Otherwise two different processes
1590          * could deadlock (one doing tee from A -> B, the other from B -> A).
1591          */
1592         inode_double_lock(ipipe->inode, opipe->inode);
1593
1594         do {
1595                 if (!opipe->readers) {
1596                         send_sig(SIGPIPE, current, 0);
1597                         if (!ret)
1598                                 ret = -EPIPE;
1599                         break;
1600                 }
1601
1602                 /*
1603                  * If we have iterated all input buffers or ran out of
1604                  * output room, break.
1605                  */
1606                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1607                         break;
1608
1609                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1610                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1611
1612                 /*
1613                  * Get a reference to this pipe buffer,
1614                  * so we can copy the contents over.
1615                  */
1616                 ibuf->ops->get(ipipe, ibuf);
1617
1618                 obuf = opipe->bufs + nbuf;
1619                 *obuf = *ibuf;
1620
1621                 /*
1622                  * Don't inherit the gift flag, we need to
1623                  * prevent multiple steals of this page.
1624                  */
1625                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1626
1627                 if (obuf->len > len)
1628                         obuf->len = len;
1629
1630                 opipe->nrbufs++;
1631                 ret += obuf->len;
1632                 len -= obuf->len;
1633                 i++;
1634         } while (len);
1635
1636         inode_double_unlock(ipipe->inode, opipe->inode);
1637
1638         /*
1639          * If we put data in the output pipe, wakeup any potential readers.
1640          */
1641         if (ret > 0) {
1642                 smp_mb();
1643                 if (waitqueue_active(&opipe->wait))
1644                         wake_up_interruptible(&opipe->wait);
1645                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1646         }
1647
1648         return ret;
1649 }
1650
1651 /*
1652  * This is a tee(1) implementation that works on pipes. It doesn't copy
1653  * any data, it simply references the 'in' pages on the 'out' pipe.
1654  * The 'flags' used are the SPLICE_F_* variants, currently the only
1655  * applicable one is SPLICE_F_NONBLOCK.
1656  */
1657 static long do_tee(struct file *in, struct file *out, size_t len,
1658                    unsigned int flags)
1659 {
1660         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1661         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1662         int ret = -EINVAL;
1663
1664         /*
1665          * Duplicate the contents of ipipe to opipe without actually
1666          * copying the data.
1667          */
1668         if (ipipe && opipe && ipipe != opipe) {
1669                 /*
1670                  * Keep going, unless we encounter an error. The ipipe/opipe
1671                  * ordering doesn't really matter.
1672                  */
1673                 ret = link_ipipe_prep(ipipe, flags);
1674                 if (!ret) {
1675                         ret = link_opipe_prep(opipe, flags);
1676                         if (!ret) {
1677                                 ret = link_pipe(ipipe, opipe, len, flags);
1678                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1679                                         ret = -EAGAIN;
1680                         }
1681                 }
1682         }
1683
1684         return ret;
1685 }
1686
1687 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1688 {
1689         struct file *in;
1690         int error, fput_in;
1691
1692         if (unlikely(!len))
1693                 return 0;
1694
1695         error = -EBADF;
1696         in = fget_light(fdin, &fput_in);
1697         if (in) {
1698                 if (in->f_mode & FMODE_READ) {
1699                         int fput_out;
1700                         struct file *out = fget_light(fdout, &fput_out);
1701
1702                         if (out) {
1703                                 if (out->f_mode & FMODE_WRITE)
1704                                         error = do_tee(in, out, len, flags);
1705                                 fput_light(out, fput_out);
1706                         }
1707                 }
1708                 fput_light(in, fput_in);
1709         }
1710
1711         return error;
1712 }