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 #include <linux/security.h>
  32
  33 /*
  34  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  35  * a vm helper function, it's already simplified quite a bit by the
  36  * addition of remove_mapping(). If success is returned, the caller may
  37  * attempt to reuse this page for another destination.
  38  */
  39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  40                                      struct pipe_buffer *buf)
  41 {
  42         struct page *page = buf->page;
  43         struct address_space *mapping;
  44
  45         lock_page(page);
  46
  47         mapping = page_mapping(page);
  48         if (mapping) {
  49                 WARN_ON(!PageUptodate(page));
  50
  51                 /*
  52                  * At least for ext2 with nobh option, we need to wait on
  53                  * writeback completing on this page, since we'll remove it
  54                  * from the pagecache.  Otherwise truncate wont wait on the
  55                  * page, allowing the disk blocks to be reused by someone else
  56                  * before we actually wrote our data to them. fs corruption
  57                  * ensues.
  58                  */
  59                 wait_on_page_writeback(page);
  60
  61                 if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
  62                         goto out_unlock;
  63
  64                 /*
  65                  * If we succeeded in removing the mapping, set LRU flag
  66                  * and return good.
  67                  */
  68                 if (remove_mapping(mapping, page)) {
  69                         buf->flags |= PIPE_BUF_FLAG_LRU;
  70                         return 0;
  71                 }
  72         }
  73
  74         /*
  75          * Raced with truncate or failed to remove page from current
  76          * address space, unlock and return failure.
  77          */
  78 out_unlock:
  79         unlock_page(page);
  80         return 1;
  81 }
  82
  83 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  84                                         struct pipe_buffer *buf)
  85 {
  86         page_cache_release(buf->page);
  87         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  88 }
  89
  90 /*
  91  * Check whether the contents of buf is OK to access. Since the content
  92  * is a page cache page, IO may be in flight.
  93  */
  94 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  95                                        struct pipe_buffer *buf)
  96 {
  97         struct page *page = buf->page;
  98         int err;
  99
 100         if (!PageUptodate(page)) {
 101                 lock_page(page);
 102
 103                 /*
 104                  * Page got truncated/unhashed. This will cause a 0-byte
 105                  * splice, if this is the first page.
 106                  */
 107                 if (!page->mapping) {
 108                         err = -ENODATA;
 109                         goto error;
 110                 }
 111
 112                 /*
 113                  * Uh oh, read-error from disk.
 114                  */
 115                 if (!PageUptodate(page)) {
 116                         err = -EIO;
 117                         goto error;
 118                 }
 119
 120                 /*
 121                  * Page is ok afterall, we are done.
 122                  */
 123                 unlock_page(page);
 124         }
 125
 126         return 0;
 127 error:
 128         unlock_page(page);
 129         return err;
 130 }
 131
 132 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 133         .can_merge = 0,
 134         .map = generic_pipe_buf_map,
 135         .unmap = generic_pipe_buf_unmap,
 136         .confirm = page_cache_pipe_buf_confirm,
 137         .release = page_cache_pipe_buf_release,
 138         .steal = page_cache_pipe_buf_steal,
 139         .get = generic_pipe_buf_get,
 140 };
 141
 142 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 143                                     struct pipe_buffer *buf)
 144 {
 145         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 146                 return 1;
 147
 148         buf->flags |= PIPE_BUF_FLAG_LRU;
 149         return generic_pipe_buf_steal(pipe, buf);
 150 }
 151
 152 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 153         .can_merge = 0,
 154         .map = generic_pipe_buf_map,
 155         .unmap = generic_pipe_buf_unmap,
 156         .confirm = generic_pipe_buf_confirm,
 157         .release = page_cache_pipe_buf_release,
 158         .steal = user_page_pipe_buf_steal,
 159         .get = generic_pipe_buf_get,
 160 };
 161
 162 /**
 163  * splice_to_pipe - fill passed data into a pipe
 164  * @pipe:       pipe to fill
 165  * @spd:        data to fill
 166  *
 167  * Description:
 168  *    @spd contains a map of pages and len/offset tuples, along with
 169  *    the struct pipe_buf_operations associated with these pages. This
 170  *    function will link that data to the pipe.
 171  *
 172  */
 173 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 174                        struct splice_pipe_desc *spd)
 175 {
 176         unsigned int spd_pages = spd->nr_pages;
 177         int ret, do_wakeup, page_nr;
 178
 179         ret = 0;
 180         do_wakeup = 0;
 181         page_nr = 0;
 182
 183         if (pipe->inode)
 184                 mutex_lock(&pipe->inode->i_mutex);
 185
 186         for (;;) {
 187                 if (!pipe->readers) {
 188                         send_sig(SIGPIPE, current, 0);
 189                         if (!ret)
 190                                 ret = -EPIPE;
 191                         break;
 192                 }
 193
 194                 if (pipe->nrbufs < PIPE_BUFFERS) {
 195                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 196                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 197
 198                         buf->page = spd->pages[page_nr];
 199                         buf->offset = spd->partial[page_nr].offset;
 200                         buf->len = spd->partial[page_nr].len;
 201                         buf->private = spd->partial[page_nr].private;
 202                         buf->ops = spd->ops;
 203                         if (spd->flags & SPLICE_F_GIFT)
 204                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 205
 206                         pipe->nrbufs++;
 207                         page_nr++;
 208                         ret += buf->len;
 209
 210                         if (pipe->inode)
 211                                 do_wakeup = 1;
 212
 213                         if (!--spd->nr_pages)
 214                                 break;
 215                         if (pipe->nrbufs < PIPE_BUFFERS)
 216                                 continue;
 217
 218                         break;
 219                 }
 220
 221                 if (spd->flags & SPLICE_F_NONBLOCK) {
 222                         if (!ret)
 223                                 ret = -EAGAIN;
 224                         break;
 225                 }
 226
 227                 if (signal_pending(current)) {
 228                         if (!ret)
 229                                 ret = -ERESTARTSYS;
 230                         break;
 231                 }
 232
 233                 if (do_wakeup) {
 234                         smp_mb();
 235                         if (waitqueue_active(&pipe->wait))
 236                                 wake_up_interruptible_sync(&pipe->wait);
 237                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 238                         do_wakeup = 0;
 239                 }
 240
 241                 pipe->waiting_writers++;
 242                 pipe_wait(pipe);
 243                 pipe->waiting_writers--;
 244         }
 245
 246         if (pipe->inode) {
 247                 mutex_unlock(&pipe->inode->i_mutex);
 248
 249                 if (do_wakeup) {
 250                         smp_mb();
 251                         if (waitqueue_active(&pipe->wait))
 252                                 wake_up_interruptible(&pipe->wait);
 253                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 254                 }
 255         }
 256
 257         while (page_nr < spd_pages)
 258                 spd->spd_release(spd, page_nr++);
 259
 260         return ret;
 261 }
 262
 263 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 264 {
 265         page_cache_release(spd->pages[i]);
 266 }
 267
 268 static int
 269 __generic_file_splice_read(struct file *in, loff_t *ppos,
 270                            struct pipe_inode_info *pipe, size_t len,
 271                            unsigned int flags)
 272 {
 273         struct address_space *mapping = in->f_mapping;
 274         unsigned int loff, nr_pages, req_pages;
 275         struct page *pages[PIPE_BUFFERS];
 276         struct partial_page partial[PIPE_BUFFERS];
 277         struct page *page;
 278         pgoff_t index, end_index;
 279         loff_t isize;
 280         int error, page_nr;
 281         struct splice_pipe_desc spd = {
 282                 .pages = pages,
 283                 .partial = partial,
 284                 .flags = flags,
 285                 .ops = &page_cache_pipe_buf_ops,
 286                 .spd_release = spd_release_page,
 287         };
 288
 289         index = *ppos >> PAGE_CACHE_SHIFT;
 290         loff = *ppos & ~PAGE_CACHE_MASK;
 291         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 292         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
 293
 294         /*
 295          * Lookup the (hopefully) full range of pages we need.
 296          */
 297         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
 298         index += spd.nr_pages;
 299
 300         /*
 301          * If find_get_pages_contig() returned fewer pages than we needed,
 302          * readahead/allocate the rest and fill in the holes.
 303          */
 304         if (spd.nr_pages < nr_pages)
 305                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 306                                 index, req_pages - spd.nr_pages);
 307
 308         error = 0;
 309         while (spd.nr_pages < nr_pages) {
 310                 /*
 311                  * Page could be there, find_get_pages_contig() breaks on
 312                  * the first hole.
 313                  */
 314                 page = find_get_page(mapping, index);
 315                 if (!page) {
 316                         /*
 317                          * page didn't exist, allocate one.
 318                          */
 319                         page = page_cache_alloc_cold(mapping);
 320                         if (!page)
 321                                 break;
 322
 323                         error = add_to_page_cache_lru(page, mapping, index,
 324                                                 mapping_gfp_mask(mapping));
 325                         if (unlikely(error)) {
 326                                 page_cache_release(page);
 327                                 if (error == -EEXIST)
 328                                         continue;
 329                                 break;
 330                         }
 331                         /*
 332                          * add_to_page_cache() locks the page, unlock it
 333                          * to avoid convoluting the logic below even more.
 334                          */
 335                         unlock_page(page);
 336                 }
 337
 338                 pages[spd.nr_pages++] = page;
 339                 index++;
 340         }
 341
 342         /*
 343          * Now loop over the map and see if we need to start IO on any
 344          * pages, fill in the partial map, etc.
 345          */
 346         index = *ppos >> PAGE_CACHE_SHIFT;
 347         nr_pages = spd.nr_pages;
 348         spd.nr_pages = 0;
 349         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 350                 unsigned int this_len;
 351
 352                 if (!len)
 353                         break;
 354
 355                 /*
 356                  * this_len is the max we'll use from this page
 357                  */
 358                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 359                 page = pages[page_nr];
 360
 361                 if (PageReadahead(page))
 362                         page_cache_async_readahead(mapping, &in->f_ra, in,
 363                                         page, index, req_pages - page_nr);
 364
 365                 /*
 366                  * If the page isn't uptodate, we may need to start io on it
 367                  */
 368                 if (!PageUptodate(page)) {
 369                         /*
 370                          * If in nonblock mode then dont block on waiting
 371                          * for an in-flight io page
 372                          */
 373                         if (flags & SPLICE_F_NONBLOCK) {
 374                                 if (!trylock_page(page)) {
 375                                         error = -EAGAIN;
 376                                         break;
 377                                 }
 378                         } else
 379                                 lock_page(page);
 380
 381                         /*
 382                          * Page was truncated, or invalidated by the
 383                          * filesystem.  Redo the find/create, but this time the
 384                          * page is kept locked, so there's no chance of another
 385                          * race with truncate/invalidate.
 386                          */
 387                         if (!page->mapping) {
 388                                 unlock_page(page);
 389                                 page = find_or_create_page(mapping, index,
 390                                                 mapping_gfp_mask(mapping));
 391
 392                                 if (!page) {
 393                                         error = -ENOMEM;
 394                                         break;
 395                                 }
 396                                 page_cache_release(pages[page_nr]);
 397                                 pages[page_nr] = page;
 398                         }
 399                         /*
 400                          * page was already under io and is now done, great
 401                          */
 402                         if (PageUptodate(page)) {
 403                                 unlock_page(page);
 404                                 goto fill_it;
 405                         }
 406
 407                         /*
 408                          * need to read in the page
 409                          */
 410                         error = mapping->a_ops->readpage(in, page);
 411                         if (unlikely(error)) {
 412                                 /*
 413                                  * We really should re-lookup the page here,
 414                                  * but it complicates things a lot. Instead
 415                                  * lets just do what we already stored, and
 416                                  * we'll get it the next time we are called.
 417                                  */
 418                                 if (error == AOP_TRUNCATED_PAGE)
 419                                         error = 0;
 420
 421                                 break;
 422                         }
 423                 }
 424 fill_it:
 425                 /*
 426                  * i_size must be checked after PageUptodate.
 427                  */
 428                 isize = i_size_read(mapping->host);
 429                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 430                 if (unlikely(!isize || index > end_index))
 431                         break;
 432
 433                 /*
 434                  * if this is the last page, see if we need to shrink
 435                  * the length and stop
 436                  */
 437                 if (end_index == index) {
 438                         unsigned int plen;
 439
 440                         /*
 441                          * max good bytes in this page
 442                          */
 443                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 444                         if (plen <= loff)
 445                                 break;
 446
 447                         /*
 448                          * force quit after adding this page
 449                          */
 450                         this_len = min(this_len, plen - loff);
 451                         len = this_len;
 452                 }
 453
 454                 partial[page_nr].offset = loff;
 455                 partial[page_nr].len = this_len;
 456                 len -= this_len;
 457                 loff = 0;
 458                 spd.nr_pages++;
 459                 index++;
 460         }
 461
 462         /*
 463          * Release any pages at the end, if we quit early. 'page_nr' is how far
 464          * we got, 'nr_pages' is how many pages are in the map.
 465          */
 466         while (page_nr < nr_pages)
 467                 page_cache_release(pages[page_nr++]);
 468         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 469
 470         if (spd.nr_pages)
 471                 return splice_to_pipe(pipe, &spd);
 472
 473         return error;
 474 }
 475
 476 /**
 477  * generic_file_splice_read - splice data from file to a pipe
 478  * @in:         file to splice from
 479  * @ppos:       position in @in
 480  * @pipe:       pipe to splice to
 481  * @len:        number of bytes to splice
 482  * @flags:      splice modifier flags
 483  *
 484  * Description:
 485  *    Will read pages from given file and fill them into a pipe. Can be
 486  *    used as long as the address_space operations for the source implements
 487  *    a readpage() hook.
 488  *
 489  */
 490 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 491                                  struct pipe_inode_info *pipe, size_t len,
 492                                  unsigned int flags)
 493 {
 494         loff_t isize, left;
 495         int ret;
 496
 497         isize = i_size_read(in->f_mapping->host);
 498         if (unlikely(*ppos >= isize))
 499                 return 0;
 500
 501         left = isize - *ppos;
 502         if (unlikely(left < len))
 503                 len = left;
 504
 505         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 506         if (ret > 0)
 507                 *ppos += ret;
 508
 509         return ret;
 510 }
 511
 512 EXPORT_SYMBOL(generic_file_splice_read);
 513
 514 /*
 515  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 516  * using sendpage(). Return the number of bytes sent.
 517  */
 518 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 519                             struct pipe_buffer *buf, struct splice_desc *sd)
 520 {
 521         struct file *file = sd->u.file;
 522         loff_t pos = sd->pos;
 523         int ret, more;
 524
 525         ret = buf->ops->confirm(pipe, buf);
 526         if (!ret) {
 527                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 528
 529                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 530                                            sd->len, &pos, more);
 531         }
 532
 533         return ret;
 534 }
 535
 536 /*
 537  * This is a little more tricky than the file -> pipe splicing. There are
 538  * basically three cases:
 539  *
 540  *      - Destination page already exists in the address space and there
 541  *        are users of it. For that case we have no other option that
 542  *        copying the data. Tough luck.
 543  *      - Destination page already exists in the address space, but there
 544  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 545  *        through to last case.
 546  *      - Destination page does not exist, we can add the pipe page to
 547  *        the page cache and avoid the copy.
 548  *
 549  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 550  * sd->flags), we attempt to migrate pages from the pipe to the output
 551  * file address space page cache. This is possible if no one else has
 552  * the pipe page referenced outside of the pipe and page cache. If
 553  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 554  * a new page in the output file page cache and fill/dirty that.
 555  */
 556 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 557                  struct splice_desc *sd)
 558 {
 559         struct file *file = sd->u.file;
 560         struct address_space *mapping = file->f_mapping;
 561         unsigned int offset, this_len;
 562         struct page *page;
 563         void *fsdata;
 564         int ret;
 565
 566         /*
 567          * make sure the data in this buffer is uptodate
 568          */
 569         ret = buf->ops->confirm(pipe, buf);
 570         if (unlikely(ret))
 571                 return ret;
 572
 573         offset = sd->pos & ~PAGE_CACHE_MASK;
 574
 575         this_len = sd->len;
 576         if (this_len + offset > PAGE_CACHE_SIZE)
 577                 this_len = PAGE_CACHE_SIZE - offset;
 578
 579         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 580                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 581         if (unlikely(ret))
 582                 goto out;
 583
 584         if (buf->page != page) {
 585                 /*
 586                  * Careful, ->map() uses KM_USER0!
 587                  */
 588                 char *src = buf->ops->map(pipe, buf, 1);
 589                 char *dst = kmap_atomic(page, KM_USER1);
 590
 591                 memcpy(dst + offset, src + buf->offset, this_len);
 592                 flush_dcache_page(page);
 593                 kunmap_atomic(dst, KM_USER1);
 594                 buf->ops->unmap(pipe, buf, src);
 595         }
 596         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 597                                 page, fsdata);
 598 out:
 599         return ret;
 600 }
 601 EXPORT_SYMBOL(pipe_to_file);
 602
 603 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 604 {
 605         smp_mb();
 606         if (waitqueue_active(&pipe->wait))
 607                 wake_up_interruptible(&pipe->wait);
 608         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 609 }
 610
 611 /**
 612  * splice_from_pipe_feed - feed available data from a pipe to a file
 613  * @pipe:       pipe to splice from
 614  * @sd:         information to @actor
 615  * @actor:      handler that splices the data
 616  *
 617  * Description:
 618
 619  *    This function loops over the pipe and calls @actor to do the
 620  *    actual moving of a single struct pipe_buffer to the desired
 621  *    destination.  It returns when there's no more buffers left in
 622  *    the pipe or if the requested number of bytes (@sd->total_len)
 623  *    have been copied.  It returns a positive number (one) if the
 624  *    pipe needs to be filled with more data, zero if the required
 625  *    number of bytes have been copied and -errno on error.
 626  *
 627  *    This, together with splice_from_pipe_{begin,end,next}, may be
 628  *    used to implement the functionality of __splice_from_pipe() when
 629  *    locking is required around copying the pipe buffers to the
 630  *    destination.
 631  */
 632 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 633                           splice_actor *actor)
 634 {
 635         int ret;
 636
 637         while (pipe->nrbufs) {
 638                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 639                 const struct pipe_buf_operations *ops = buf->ops;
 640
 641                 sd->len = buf->len;
 642                 if (sd->len > sd->total_len)
 643                         sd->len = sd->total_len;
 644
 645                 ret = actor(pipe, buf, sd);
 646                 if (ret <= 0) {
 647                         if (ret == -ENODATA)
 648                                 ret = 0;
 649                         return ret;
 650                 }
 651                 buf->offset += ret;
 652                 buf->len -= ret;
 653
 654                 sd->num_spliced += ret;
 655                 sd->len -= ret;
 656                 sd->pos += ret;
 657                 sd->total_len -= ret;
 658
 659                 if (!buf->len) {
 660                         buf->ops = NULL;
 661                         ops->release(pipe, buf);
 662                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 663                         pipe->nrbufs--;
 664                         if (pipe->inode)
 665                                 sd->need_wakeup = true;
 666                 }
 667
 668                 if (!sd->total_len)
 669                         return 0;
 670         }
 671
 672         return 1;
 673 }
 674 EXPORT_SYMBOL(splice_from_pipe_feed);
 675
 676 /**
 677  * splice_from_pipe_next - wait for some data to splice from
 678  * @pipe:       pipe to splice from
 679  * @sd:         information about the splice operation
 680  *
 681  * Description:
 682  *    This function will wait for some data and return a positive
 683  *    value (one) if pipe buffers are available.  It will return zero
 684  *    or -errno if no more data needs to be spliced.
 685  */
 686 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 687 {
 688         while (!pipe->nrbufs) {
 689                 if (!pipe->writers)
 690                         return 0;
 691
 692                 if (!pipe->waiting_writers && sd->num_spliced)
 693                         return 0;
 694
 695                 if (sd->flags & SPLICE_F_NONBLOCK)
 696                         return -EAGAIN;
 697
 698                 if (signal_pending(current))
 699                         return -ERESTARTSYS;
 700
 701                 if (sd->need_wakeup) {
 702                         wakeup_pipe_writers(pipe);
 703                         sd->need_wakeup = false;
 704                 }
 705
 706                 pipe_wait(pipe);
 707         }
 708
 709         return 1;
 710 }
 711 EXPORT_SYMBOL(splice_from_pipe_next);
 712
 713 /**
 714  * splice_from_pipe_begin - start splicing from pipe
 715  * @pipe:       pipe to splice from
 716  *
 717  * Description:
 718  *    This function should be called before a loop containing
 719  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 720  *    initialize the necessary fields of @sd.
 721  */
 722 void splice_from_pipe_begin(struct splice_desc *sd)
 723 {
 724         sd->num_spliced = 0;
 725         sd->need_wakeup = false;
 726 }
 727 EXPORT_SYMBOL(splice_from_pipe_begin);
 728
 729 /**
 730  * splice_from_pipe_end - finish splicing from pipe
 731  * @pipe:       pipe to splice from
 732  * @sd:         information about the splice operation
 733  *
 734  * Description:
 735  *    This function will wake up pipe writers if necessary.  It should
 736  *    be called after a loop containing splice_from_pipe_next() and
 737  *    splice_from_pipe_feed().
 738  */
 739 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 740 {
 741         if (sd->need_wakeup)
 742                 wakeup_pipe_writers(pipe);
 743 }
 744 EXPORT_SYMBOL(splice_from_pipe_end);
 745
 746 /**
 747  * __splice_from_pipe - splice data from a pipe to given actor
 748  * @pipe:       pipe to splice from
 749  * @sd:         information to @actor
 750  * @actor:      handler that splices the data
 751  *
 752  * Description:
 753  *    This function does little more than loop over the pipe and call
 754  *    @actor to do the actual moving of a single struct pipe_buffer to
 755  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 756  *    pipe_to_user.
 757  *
 758  */
 759 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 760                            splice_actor *actor)
 761 {
 762         int ret;
 763
 764         splice_from_pipe_begin(sd);
 765         do {
 766                 ret = splice_from_pipe_next(pipe, sd);
 767                 if (ret > 0)
 768                         ret = splice_from_pipe_feed(pipe, sd, actor);
 769         } while (ret > 0);
 770         splice_from_pipe_end(pipe, sd);
 771
 772         return sd->num_spliced ? sd->num_spliced : ret;
 773 }
 774 EXPORT_SYMBOL(__splice_from_pipe);
 775
 776 /**
 777  * splice_from_pipe - splice data from a pipe to a file
 778  * @pipe:       pipe to splice from
 779  * @out:        file to splice to
 780  * @ppos:       position in @out
 781  * @len:        how many bytes to splice
 782  * @flags:      splice modifier flags
 783  * @actor:      handler that splices the data
 784  *
 785  * Description:
 786  *    See __splice_from_pipe. This function locks the pipe inode,
 787  *    otherwise it's identical to __splice_from_pipe().
 788  *
 789  */
 790 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 791                          loff_t *ppos, size_t len, unsigned int flags,
 792                          splice_actor *actor)
 793 {
 794         ssize_t ret;
 795         struct splice_desc sd = {
 796                 .total_len = len,
 797                 .flags = flags,
 798                 .pos = *ppos,
 799                 .u.file = out,
 800         };
 801
 802         if (pipe->inode)
 803                 mutex_lock(&pipe->inode->i_mutex);
 804         ret = __splice_from_pipe(pipe, &sd, actor);
 805         if (pipe->inode)
 806                 mutex_unlock(&pipe->inode->i_mutex);
 807
 808         return ret;
 809 }
 810
 811 /**
 812  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 813  * @pipe:       pipe info
 814  * @out:        file to write to
 815  * @ppos:       position in @out
 816  * @len:        number of bytes to splice
 817  * @flags:      splice modifier flags
 818  *
 819  * Description:
 820  *    Will either move or copy pages (determined by @flags options) from
 821  *    the given pipe inode to the given file. The caller is responsible
 822  *    for acquiring i_mutex on both inodes.
 823  *
 824  */
 825 ssize_t
 826 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 827                                  loff_t *ppos, size_t len, unsigned int flags)
 828 {
 829         struct address_space *mapping = out->f_mapping;
 830         struct inode *inode = mapping->host;
 831         struct splice_desc sd = {
 832                 .total_len = len,
 833                 .flags = flags,
 834                 .pos = *ppos,
 835                 .u.file = out,
 836         };
 837         ssize_t ret;
 838         int err;
 839
 840         err = file_remove_suid(out);
 841         if (unlikely(err))
 842                 return err;
 843
 844         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
 845         if (ret > 0) {
 846                 unsigned long nr_pages;
 847
 848                 *ppos += ret;
 849                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 850
 851                 /*
 852                  * If file or inode is SYNC and we actually wrote some data,
 853                  * sync it.
 854                  */
 855                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 856                         err = generic_osync_inode(inode, mapping,
 857                                                   OSYNC_METADATA|OSYNC_DATA);
 858
 859                         if (err)
 860                                 ret = err;
 861                 }
 862                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 863         }
 864
 865         return ret;
 866 }
 867
 868 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 869
 870 /**
 871  * generic_file_splice_write - splice data from a pipe to a file
 872  * @pipe:       pipe info
 873  * @out:        file to write to
 874  * @ppos:       position in @out
 875  * @len:        number of bytes to splice
 876  * @flags:      splice modifier flags
 877  *
 878  * Description:
 879  *    Will either move or copy pages (determined by @flags options) from
 880  *    the given pipe inode to the given file.
 881  *
 882  */
 883 ssize_t
 884 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 885                           loff_t *ppos, size_t len, unsigned int flags)
 886 {
 887         struct address_space *mapping = out->f_mapping;
 888         struct inode *inode = mapping->host;
 889         struct splice_desc sd = {
 890                 .total_len = len,
 891                 .flags = flags,
 892                 .pos = *ppos,
 893                 .u.file = out,
 894         };
 895         ssize_t ret;
 896
 897         if (pipe->inode)
 898                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
 899
 900         splice_from_pipe_begin(&sd);
 901         do {
 902                 ret = splice_from_pipe_next(pipe, &sd);
 903                 if (ret <= 0)
 904                         break;
 905
 906                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
 907                 ret = file_remove_suid(out);
 908                 if (!ret)
 909                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
 910                 mutex_unlock(&inode->i_mutex);
 911         } while (ret > 0);
 912         splice_from_pipe_end(pipe, &sd);
 913
 914         if (pipe->inode)
 915                 mutex_unlock(&pipe->inode->i_mutex);
 916
 917         if (sd.num_spliced)
 918                 ret = sd.num_spliced;
 919
 920         if (ret > 0) {
 921                 unsigned long nr_pages;
 922
 923                 *ppos += ret;
 924                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 925
 926                 /*
 927                  * If file or inode is SYNC and we actually wrote some data,
 928                  * sync it.
 929                  */
 930                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 931                         int err;
 932
 933                         mutex_lock(&inode->i_mutex);
 934                         err = generic_osync_inode(inode, mapping,
 935                                                   OSYNC_METADATA|OSYNC_DATA);
 936                         mutex_unlock(&inode->i_mutex);
 937
 938                         if (err)
 939                                 ret = err;
 940                 }
 941                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 942         }
 943
 944         return ret;
 945 }
 946
 947 EXPORT_SYMBOL(generic_file_splice_write);
 948
 949 /**
 950  * generic_splice_sendpage - splice data from a pipe to a socket
 951  * @pipe:       pipe to splice from
 952  * @out:        socket to write to
 953  * @ppos:       position in @out
 954  * @len:        number of bytes to splice
 955  * @flags:      splice modifier flags
 956  *
 957  * Description:
 958  *    Will send @len bytes from the pipe to a network socket. No data copying
 959  *    is involved.
 960  *
 961  */
 962 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 963                                 loff_t *ppos, size_t len, unsigned int flags)
 964 {
 965         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 966 }
 967
 968 EXPORT_SYMBOL(generic_splice_sendpage);
 969
 970 /*
 971  * Attempt to initiate a splice from pipe to file.
 972  */
 973 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 974                            loff_t *ppos, size_t len, unsigned int flags)
 975 {
 976         int ret;
 977
 978         if (unlikely(!out->f_op || !out->f_op->splice_write))
 979                 return -EINVAL;
 980
 981         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 982                 return -EBADF;
 983
 984         if (unlikely(out->f_flags & O_APPEND))
 985                 return -EINVAL;
 986
 987         ret = rw_verify_area(WRITE, out, ppos, len);
 988         if (unlikely(ret < 0))
 989                 return ret;
 990
 991         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 992 }
 993
 994 /*
 995  * Attempt to initiate a splice from a file to a pipe.
 996  */
 997 static long do_splice_to(struct file *in, loff_t *ppos,
 998                          struct pipe_inode_info *pipe, size_t len,
 999                          unsigned int flags)
1000 {
1001         int ret;
1002
1003         if (unlikely(!in->f_op || !in->f_op->splice_read))
1004                 return -EINVAL;
1005
1006         if (unlikely(!(in->f_mode & FMODE_READ)))
1007                 return -EBADF;
1008
1009         ret = rw_verify_area(READ, in, ppos, len);
1010         if (unlikely(ret < 0))
1011                 return ret;
1012
1013         return in->f_op->splice_read(in, ppos, pipe, len, flags);
1014 }
1015
1016 /**
1017  * splice_direct_to_actor - splices data directly between two non-pipes
1018  * @in:         file to splice from
1019  * @sd:         actor information on where to splice to
1020  * @actor:      handles the data splicing
1021  *
1022  * Description:
1023  *    This is a special case helper to splice directly between two
1024  *    points, without requiring an explicit pipe. Internally an allocated
1025  *    pipe is cached in the process, and reused during the lifetime of
1026  *    that process.
1027  *
1028  */
1029 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1030                                splice_direct_actor *actor)
1031 {
1032         struct pipe_inode_info *pipe;
1033         long ret, bytes;
1034         umode_t i_mode;
1035         size_t len;
1036         int i, flags;
1037
1038         /*
1039          * We require the input being a regular file, as we don't want to
1040          * randomly drop data for eg socket -> socket splicing. Use the
1041          * piped splicing for that!
1042          */
1043         i_mode = in->f_path.dentry->d_inode->i_mode;
1044         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1045                 return -EINVAL;
1046
1047         /*
1048          * neither in nor out is a pipe, setup an internal pipe attached to
1049          * 'out' and transfer the wanted data from 'in' to 'out' through that
1050          */
1051         pipe = current->splice_pipe;
1052         if (unlikely(!pipe)) {
1053                 pipe = alloc_pipe_info(NULL);
1054                 if (!pipe)
1055                         return -ENOMEM;
1056
1057                 /*
1058                  * We don't have an immediate reader, but we'll read the stuff
1059                  * out of the pipe right after the splice_to_pipe(). So set
1060                  * PIPE_READERS appropriately.
1061                  */
1062                 pipe->readers = 1;
1063
1064                 current->splice_pipe = pipe;
1065         }
1066
1067         /*
1068          * Do the splice.
1069          */
1070         ret = 0;
1071         bytes = 0;
1072         len = sd->total_len;
1073         flags = sd->flags;
1074
1075         /*
1076          * Don't block on output, we have to drain the direct pipe.
1077          */
1078         sd->flags &= ~SPLICE_F_NONBLOCK;
1079
1080         while (len) {
1081                 size_t read_len;
1082                 loff_t pos = sd->pos, prev_pos = pos;
1083
1084                 ret = do_splice_to(in, &pos, pipe, len, flags);
1085                 if (unlikely(ret <= 0))
1086                         goto out_release;
1087
1088                 read_len = ret;
1089                 sd->total_len = read_len;
1090
1091                 /*
1092                  * NOTE: nonblocking mode only applies to the input. We
1093                  * must not do the output in nonblocking mode as then we
1094                  * could get stuck data in the internal pipe:
1095                  */
1096                 ret = actor(pipe, sd);
1097                 if (unlikely(ret <= 0)) {
1098                         sd->pos = prev_pos;
1099                         goto out_release;
1100                 }
1101
1102                 bytes += ret;
1103                 len -= ret;
1104                 sd->pos = pos;
1105
1106                 if (ret < read_len) {
1107                         sd->pos = prev_pos + ret;
1108                         goto out_release;
1109                 }
1110         }
1111
1112 done:
1113         pipe->nrbufs = pipe->curbuf = 0;
1114         file_accessed(in);
1115         return bytes;
1116
1117 out_release:
1118         /*
1119          * If we did an incomplete transfer we must release
1120          * the pipe buffers in question:
1121          */
1122         for (i = 0; i < PIPE_BUFFERS; i++) {
1123                 struct pipe_buffer *buf = pipe->bufs + i;
1124
1125                 if (buf->ops) {
1126                         buf->ops->release(pipe, buf);
1127                         buf->ops = NULL;
1128                 }
1129         }
1130
1131         if (!bytes)
1132                 bytes = ret;
1133
1134         goto done;
1135 }
1136 EXPORT_SYMBOL(splice_direct_to_actor);
1137
1138 static int direct_splice_actor(struct pipe_inode_info *pipe,
1139                                struct splice_desc *sd)
1140 {
1141         struct file *file = sd->u.file;
1142
1143         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1144 }
1145
1146 /**
1147  * do_splice_direct - splices data directly between two files
1148  * @in:         file to splice from
1149  * @ppos:       input file offset
1150  * @out:        file to splice to
1151  * @len:        number of bytes to splice
1152  * @flags:      splice modifier flags
1153  *
1154  * Description:
1155  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1156  *    doing it in the application would incur an extra system call
1157  *    (splice in + splice out, as compared to just sendfile()). So this helper
1158  *    can splice directly through a process-private pipe.
1159  *
1160  */
1161 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1162                       size_t len, unsigned int flags)
1163 {
1164         struct splice_desc sd = {
1165                 .len            = len,
1166                 .total_len      = len,
1167                 .flags          = flags,
1168                 .pos            = *ppos,
1169                 .u.file         = out,
1170         };
1171         long ret;
1172
1173         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1174         if (ret > 0)
1175                 *ppos = sd.pos;
1176
1177         return ret;
1178 }
1179
1180 /*
1181  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1182  * location, so checking ->i_pipe is not enough to verify that this is a
1183  * pipe.
1184  */
1185 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1186 {
1187         if (S_ISFIFO(inode->i_mode))
1188                 return inode->i_pipe;
1189
1190         return NULL;
1191 }
1192
1193 /*
1194  * Determine where to splice to/from.
1195  */
1196 static long do_splice(struct file *in, loff_t __user *off_in,
1197                       struct file *out, loff_t __user *off_out,
1198                       size_t len, unsigned int flags)
1199 {
1200         struct pipe_inode_info *pipe;
1201         loff_t offset, *off;
1202         long ret;
1203
1204         pipe = pipe_info(in->f_path.dentry->d_inode);
1205         if (pipe) {
1206                 if (off_in)
1207                         return -ESPIPE;
1208                 if (off_out) {
1209                         if (out->f_op->llseek == no_llseek)
1210                                 return -EINVAL;
1211                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1212                                 return -EFAULT;
1213                         off = &offset;
1214                 } else
1215                         off = &out->f_pos;
1216
1217                 ret = do_splice_from(pipe, out, off, len, flags);
1218
1219                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1220                         ret = -EFAULT;
1221
1222                 return ret;
1223         }
1224
1225         pipe = pipe_info(out->f_path.dentry->d_inode);
1226         if (pipe) {
1227                 if (off_out)
1228                         return -ESPIPE;
1229                 if (off_in) {
1230                         if (in->f_op->llseek == no_llseek)
1231                                 return -EINVAL;
1232                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1233                                 return -EFAULT;
1234                         off = &offset;
1235                 } else
1236                         off = &in->f_pos;
1237
1238                 ret = do_splice_to(in, off, pipe, len, flags);
1239
1240                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1241                         ret = -EFAULT;
1242
1243                 return ret;
1244         }
1245
1246         return -EINVAL;
1247 }
1248
1249 /*
1250  * Map an iov into an array of pages and offset/length tupples. With the
1251  * partial_page structure, we can map several non-contiguous ranges into
1252  * our ones pages[] map instead of splitting that operation into pieces.
1253  * Could easily be exported as a generic helper for other users, in which
1254  * case one would probably want to add a 'max_nr_pages' parameter as well.
1255  */
1256 static int get_iovec_page_array(const struct iovec __user *iov,
1257                                 unsigned int nr_vecs, struct page **pages,
1258                                 struct partial_page *partial, int aligned)
1259 {
1260         int buffers = 0, error = 0;
1261
1262         while (nr_vecs) {
1263                 unsigned long off, npages;
1264                 struct iovec entry;
1265                 void __user *base;
1266                 size_t len;
1267                 int i;
1268
1269                 error = -EFAULT;
1270                 if (copy_from_user(&entry, iov, sizeof(entry)))
1271                         break;
1272
1273                 base = entry.iov_base;
1274                 len = entry.iov_len;
1275
1276                 /*
1277                  * Sanity check this iovec. 0 read succeeds.
1278                  */
1279                 error = 0;
1280                 if (unlikely(!len))
1281                         break;
1282                 error = -EFAULT;
1283                 if (!access_ok(VERIFY_READ, base, len))
1284                         break;
1285
1286                 /*
1287                  * Get this base offset and number of pages, then map
1288                  * in the user pages.
1289                  */
1290                 off = (unsigned long) base & ~PAGE_MASK;
1291
1292                 /*
1293                  * If asked for alignment, the offset must be zero and the
1294                  * length a multiple of the PAGE_SIZE.
1295                  */
1296                 error = -EINVAL;
1297                 if (aligned && (off || len & ~PAGE_MASK))
1298                         break;
1299
1300                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1301                 if (npages > PIPE_BUFFERS - buffers)
1302                         npages = PIPE_BUFFERS - buffers;
1303
1304                 error = get_user_pages_fast((unsigned long)base, npages,
1305                                         0, &pages[buffers]);
1306
1307                 if (unlikely(error <= 0))
1308                         break;
1309
1310                 /*
1311                  * Fill this contiguous range into the partial page map.
1312                  */
1313                 for (i = 0; i < error; i++) {
1314                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1315
1316                         partial[buffers].offset = off;
1317                         partial[buffers].len = plen;
1318
1319                         off = 0;
1320                         len -= plen;
1321                         buffers++;
1322                 }
1323
1324                 /*
1325                  * We didn't complete this iov, stop here since it probably
1326                  * means we have to move some of this into a pipe to
1327                  * be able to continue.
1328                  */
1329                 if (len)
1330                         break;
1331
1332                 /*
1333                  * Don't continue if we mapped fewer pages than we asked for,
1334                  * or if we mapped the max number of pages that we have
1335                  * room for.
1336                  */
1337                 if (error < npages || buffers == PIPE_BUFFERS)
1338                         break;
1339
1340                 nr_vecs--;
1341                 iov++;
1342         }
1343
1344         if (buffers)
1345                 return buffers;
1346
1347         return error;
1348 }
1349
1350 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1351                         struct splice_desc *sd)
1352 {
1353         char *src;
1354         int ret;
1355
1356         ret = buf->ops->confirm(pipe, buf);
1357         if (unlikely(ret))
1358                 return ret;
1359
1360         /*
1361          * See if we can use the atomic maps, by prefaulting in the
1362          * pages and doing an atomic copy
1363          */
1364         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1365                 src = buf->ops->map(pipe, buf, 1);
1366                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1367                                                         sd->len);
1368                 buf->ops->unmap(pipe, buf, src);
1369                 if (!ret) {
1370                         ret = sd->len;
1371                         goto out;
1372                 }
1373         }
1374
1375         /*
1376          * No dice, use slow non-atomic map and copy
1377          */
1378         src = buf->ops->map(pipe, buf, 0);
1379
1380         ret = sd->len;
1381         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1382                 ret = -EFAULT;
1383
1384         buf->ops->unmap(pipe, buf, src);
1385 out:
1386         if (ret > 0)
1387                 sd->u.userptr += ret;
1388         return ret;
1389 }
1390
1391 /*
1392  * For lack of a better implementation, implement vmsplice() to userspace
1393  * as a simple copy of the pipes pages to the user iov.
1394  */
1395 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1396                              unsigned long nr_segs, unsigned int flags)
1397 {
1398         struct pipe_inode_info *pipe;
1399         struct splice_desc sd;
1400         ssize_t size;
1401         int error;
1402         long ret;
1403
1404         pipe = pipe_info(file->f_path.dentry->d_inode);
1405         if (!pipe)
1406                 return -EBADF;
1407
1408         if (pipe->inode)
1409                 mutex_lock(&pipe->inode->i_mutex);
1410
1411         error = ret = 0;
1412         while (nr_segs) {
1413                 void __user *base;
1414                 size_t len;
1415
1416                 /*
1417                  * Get user address base and length for this iovec.
1418                  */
1419                 error = get_user(base, &iov->iov_base);
1420                 if (unlikely(error))
1421                         break;
1422                 error = get_user(len, &iov->iov_len);
1423                 if (unlikely(error))
1424                         break;
1425
1426                 /*
1427                  * Sanity check this iovec. 0 read succeeds.
1428                  */
1429                 if (unlikely(!len))
1430                         break;
1431                 if (unlikely(!base)) {
1432                         error = -EFAULT;
1433                         break;
1434                 }
1435
1436                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1437                         error = -EFAULT;
1438                         break;
1439                 }
1440
1441                 sd.len = 0;
1442                 sd.total_len = len;
1443                 sd.flags = flags;
1444                 sd.u.userptr = base;
1445                 sd.pos = 0;
1446
1447                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1448                 if (size < 0) {
1449                         if (!ret)
1450                                 ret = size;
1451
1452                         break;
1453                 }
1454
1455                 ret += size;
1456
1457                 if (size < len)
1458                         break;
1459
1460                 nr_segs--;
1461                 iov++;
1462         }
1463
1464         if (pipe->inode)
1465                 mutex_unlock(&pipe->inode->i_mutex);
1466
1467         if (!ret)
1468                 ret = error;
1469
1470         return ret;
1471 }
1472
1473 /*
1474  * vmsplice splices a user address range into a pipe. It can be thought of
1475  * as splice-from-memory, where the regular splice is splice-from-file (or
1476  * to file). In both cases the output is a pipe, naturally.
1477  */
1478 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1479                              unsigned long nr_segs, unsigned int flags)
1480 {
1481         struct pipe_inode_info *pipe;
1482         struct page *pages[PIPE_BUFFERS];
1483         struct partial_page partial[PIPE_BUFFERS];
1484         struct splice_pipe_desc spd = {
1485                 .pages = pages,
1486                 .partial = partial,
1487                 .flags = flags,
1488                 .ops = &user_page_pipe_buf_ops,
1489                 .spd_release = spd_release_page,
1490         };
1491
1492         pipe = pipe_info(file->f_path.dentry->d_inode);
1493         if (!pipe)
1494                 return -EBADF;
1495
1496         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1497                                             flags & SPLICE_F_GIFT);
1498         if (spd.nr_pages <= 0)
1499                 return spd.nr_pages;
1500
1501         return splice_to_pipe(pipe, &spd);
1502 }
1503
1504 /*
1505  * Note that vmsplice only really supports true splicing _from_ user memory
1506  * to a pipe, not the other way around. Splicing from user memory is a simple
1507  * operation that can be supported without any funky alignment restrictions
1508  * or nasty vm tricks. We simply map in the user memory and fill them into
1509  * a pipe. The reverse isn't quite as easy, though. There are two possible
1510  * solutions for that:
1511  *
1512  *      - memcpy() the data internally, at which point we might as well just
1513  *        do a regular read() on the buffer anyway.
1514  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1515  *        has restriction limitations on both ends of the pipe).
1516  *
1517  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1518  *
1519  */
1520 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1521                 unsigned long, nr_segs, unsigned int, flags)
1522 {
1523         struct file *file;
1524         long error;
1525         int fput;
1526
1527         if (unlikely(nr_segs > UIO_MAXIOV))
1528                 return -EINVAL;
1529         else if (unlikely(!nr_segs))
1530                 return 0;
1531
1532         error = -EBADF;
1533         file = fget_light(fd, &fput);
1534         if (file) {
1535                 if (file->f_mode & FMODE_WRITE)
1536                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1537                 else if (file->f_mode & FMODE_READ)
1538                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1539
1540                 fput_light(file, fput);
1541         }
1542
1543         return error;
1544 }
1545
1546 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1547                 int, fd_out, loff_t __user *, off_out,
1548                 size_t, len, unsigned int, flags)
1549 {
1550         long error;
1551         struct file *in, *out;
1552         int fput_in, fput_out;
1553
1554         if (unlikely(!len))
1555                 return 0;
1556
1557         error = -EBADF;
1558         in = fget_light(fd_in, &fput_in);
1559         if (in) {
1560                 if (in->f_mode & FMODE_READ) {
1561                         out = fget_light(fd_out, &fput_out);
1562                         if (out) {
1563                                 if (out->f_mode & FMODE_WRITE)
1564                                         error = do_splice(in, off_in,
1565                                                           out, off_out,
1566                                                           len, flags);
1567                                 fput_light(out, fput_out);
1568                         }
1569                 }
1570
1571                 fput_light(in, fput_in);
1572         }
1573
1574         return error;
1575 }
1576
1577 /*
1578  * Make sure there's data to read. Wait for input if we can, otherwise
1579  * return an appropriate error.
1580  */
1581 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1582 {
1583         int ret;
1584
1585         /*
1586          * Check ->nrbufs without the inode lock first. This function
1587          * is speculative anyways, so missing one is ok.
1588          */
1589         if (pipe->nrbufs)
1590                 return 0;
1591
1592         ret = 0;
1593         mutex_lock(&pipe->inode->i_mutex);
1594
1595         while (!pipe->nrbufs) {
1596                 if (signal_pending(current)) {
1597                         ret = -ERESTARTSYS;
1598                         break;
1599                 }
1600                 if (!pipe->writers)
1601                         break;
1602                 if (!pipe->waiting_writers) {
1603                         if (flags & SPLICE_F_NONBLOCK) {
1604                                 ret = -EAGAIN;
1605                                 break;
1606                         }
1607                 }
1608                 pipe_wait(pipe);
1609         }
1610
1611         mutex_unlock(&pipe->inode->i_mutex);
1612         return ret;
1613 }
1614
1615 /*
1616  * Make sure there's writeable room. Wait for room if we can, otherwise
1617  * return an appropriate error.
1618  */
1619 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1620 {
1621         int ret;
1622
1623         /*
1624          * Check ->nrbufs without the inode lock first. This function
1625          * is speculative anyways, so missing one is ok.
1626          */
1627         if (pipe->nrbufs < PIPE_BUFFERS)
1628                 return 0;
1629
1630         ret = 0;
1631         mutex_lock(&pipe->inode->i_mutex);
1632
1633         while (pipe->nrbufs >= PIPE_BUFFERS) {
1634                 if (!pipe->readers) {
1635                         send_sig(SIGPIPE, current, 0);
1636                         ret = -EPIPE;
1637                         break;
1638                 }
1639                 if (flags & SPLICE_F_NONBLOCK) {
1640                         ret = -EAGAIN;
1641                         break;
1642                 }
1643                 if (signal_pending(current)) {
1644                         ret = -ERESTARTSYS;
1645                         break;
1646                 }
1647                 pipe->waiting_writers++;
1648                 pipe_wait(pipe);
1649                 pipe->waiting_writers--;
1650         }
1651
1652         mutex_unlock(&pipe->inode->i_mutex);
1653         return ret;
1654 }
1655
1656 /*
1657  * Link contents of ipipe to opipe.
1658  */
1659 static int link_pipe(struct pipe_inode_info *ipipe,
1660                      struct pipe_inode_info *opipe,
1661                      size_t len, unsigned int flags)
1662 {
1663         struct pipe_buffer *ibuf, *obuf;
1664         int ret = 0, i = 0, nbuf;
1665
1666         /*
1667          * Potential ABBA deadlock, work around it by ordering lock
1668          * grabbing by inode address. Otherwise two different processes
1669          * could deadlock (one doing tee from A -> B, the other from B -> A).
1670          */
1671         inode_double_lock(ipipe->inode, opipe->inode);
1672
1673         do {
1674                 if (!opipe->readers) {
1675                         send_sig(SIGPIPE, current, 0);
1676                         if (!ret)
1677                                 ret = -EPIPE;
1678                         break;
1679                 }
1680
1681                 /*
1682                  * If we have iterated all input buffers or ran out of
1683                  * output room, break.
1684                  */
1685                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1686                         break;
1687
1688                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1689                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1690
1691                 /*
1692                  * Get a reference to this pipe buffer,
1693                  * so we can copy the contents over.
1694                  */
1695                 ibuf->ops->get(ipipe, ibuf);
1696
1697                 obuf = opipe->bufs + nbuf;
1698                 *obuf = *ibuf;
1699
1700                 /*
1701                  * Don't inherit the gift flag, we need to
1702                  * prevent multiple steals of this page.
1703                  */
1704                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1705
1706                 if (obuf->len > len)
1707                         obuf->len = len;
1708
1709                 opipe->nrbufs++;
1710                 ret += obuf->len;
1711                 len -= obuf->len;
1712                 i++;
1713         } while (len);
1714
1715         /*
1716          * return EAGAIN if we have the potential of some data in the
1717          * future, otherwise just return 0
1718          */
1719         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1720                 ret = -EAGAIN;
1721
1722         inode_double_unlock(ipipe->inode, opipe->inode);
1723
1724         /*
1725          * If we put data in the output pipe, wakeup any potential readers.
1726          */
1727         if (ret > 0) {
1728                 smp_mb();
1729                 if (waitqueue_active(&opipe->wait))
1730                         wake_up_interruptible(&opipe->wait);
1731                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1732         }
1733
1734         return ret;
1735 }
1736
1737 /*
1738  * This is a tee(1) implementation that works on pipes. It doesn't copy
1739  * any data, it simply references the 'in' pages on the 'out' pipe.
1740  * The 'flags' used are the SPLICE_F_* variants, currently the only
1741  * applicable one is SPLICE_F_NONBLOCK.
1742  */
1743 static long do_tee(struct file *in, struct file *out, size_t len,
1744                    unsigned int flags)
1745 {
1746         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1747         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1748         int ret = -EINVAL;
1749
1750         /*
1751          * Duplicate the contents of ipipe to opipe without actually
1752          * copying the data.
1753          */
1754         if (ipipe && opipe && ipipe != opipe) {
1755                 /*
1756                  * Keep going, unless we encounter an error. The ipipe/opipe
1757                  * ordering doesn't really matter.
1758                  */
1759                 ret = link_ipipe_prep(ipipe, flags);
1760                 if (!ret) {
1761                         ret = link_opipe_prep(opipe, flags);
1762                         if (!ret)
1763                                 ret = link_pipe(ipipe, opipe, len, flags);
1764                 }
1765         }
1766
1767         return ret;
1768 }
1769
1770 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1771 {
1772         struct file *in;
1773         int error, fput_in;
1774
1775         if (unlikely(!len))
1776                 return 0;
1777
1778         error = -EBADF;
1779         in = fget_light(fdin, &fput_in);
1780         if (in) {
1781                 if (in->f_mode & FMODE_READ) {
1782                         int fput_out;
1783                         struct file *out = fget_light(fdout, &fput_out);
1784
1785                         if (out) {
1786                                 if (out->f_mode & FMODE_WRITE)
1787                                         error = do_tee(in, out, len, flags);
1788                                 fput_light(out, fput_out);
1789                         }
1790                 }
1791                 fput_light(in, fput_in);
1792         }
1793
1794         return error;
1795 }