git.oblomov.eu Git - 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 #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 (TestSetPageLocked(page)) {
 375                                         error = -EAGAIN;
 376                                         break;
 377                                 }
 378                         } else
 379                                 lock_page(page);
 380
 381                         /*
 382                          * page was truncated, stop here. if this isn't the
 383                          * first page, we'll just complete what we already
 384                          * added
 385                          */
 386                         if (!page->mapping) {
 387                                 unlock_page(page);
 388                                 break;
 389                         }
 390                         /*
 391                          * page was already under io and is now done, great
 392                          */
 393                         if (PageUptodate(page)) {
 394                                 unlock_page(page);
 395                                 goto fill_it;
 396                         }
 397
 398                         /*
 399                          * need to read in the page
 400                          */
 401                         error = mapping->a_ops->readpage(in, page);
 402                         if (unlikely(error)) {
 403                                 /*
 404                                  * We really should re-lookup the page here,
 405                                  * but it complicates things a lot. Instead
 406                                  * lets just do what we already stored, and
 407                                  * we'll get it the next time we are called.
 408                                  */
 409                                 if (error == AOP_TRUNCATED_PAGE)
 410                                         error = 0;
 411
 412                                 break;
 413                         }
 414                 }
 415 fill_it:
 416                 /*
 417                  * i_size must be checked after PageUptodate.
 418                  */
 419                 isize = i_size_read(mapping->host);
 420                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 421                 if (unlikely(!isize || index > end_index))
 422                         break;
 423
 424                 /*
 425                  * if this is the last page, see if we need to shrink
 426                  * the length and stop
 427                  */
 428                 if (end_index == index) {
 429                         unsigned int plen;
 430
 431                         /*
 432                          * max good bytes in this page
 433                          */
 434                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 435                         if (plen <= loff)
 436                                 break;
 437
 438                         /*
 439                          * force quit after adding this page
 440                          */
 441                         this_len = min(this_len, plen - loff);
 442                         len = this_len;
 443                 }
 444
 445                 partial[page_nr].offset = loff;
 446                 partial[page_nr].len = this_len;
 447                 len -= this_len;
 448                 loff = 0;
 449                 spd.nr_pages++;
 450                 index++;
 451         }
 452
 453         /*
 454          * Release any pages at the end, if we quit early. 'page_nr' is how far
 455          * we got, 'nr_pages' is how many pages are in the map.
 456          */
 457         while (page_nr < nr_pages)
 458                 page_cache_release(pages[page_nr++]);
 459         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 460
 461         if (spd.nr_pages)
 462                 return splice_to_pipe(pipe, &spd);
 463
 464         return error;
 465 }
 466
 467 /**
 468  * generic_file_splice_read - splice data from file to a pipe
 469  * @in:         file to splice from
 470  * @ppos:       position in @in
 471  * @pipe:       pipe to splice to
 472  * @len:        number of bytes to splice
 473  * @flags:      splice modifier flags
 474  *
 475  * Description:
 476  *    Will read pages from given file and fill them into a pipe. Can be
 477  *    used as long as the address_space operations for the source implements
 478  *    a readpage() hook.
 479  *
 480  */
 481 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 482                                  struct pipe_inode_info *pipe, size_t len,
 483                                  unsigned int flags)
 484 {
 485         loff_t isize, left;
 486         int ret;
 487
 488         isize = i_size_read(in->f_mapping->host);
 489         if (unlikely(*ppos >= isize))
 490                 return 0;
 491
 492         left = isize - *ppos;
 493         if (unlikely(left < len))
 494                 len = left;
 495
 496         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 497         if (ret > 0)
 498                 *ppos += ret;
 499
 500         return ret;
 501 }
 502
 503 EXPORT_SYMBOL(generic_file_splice_read);
 504
 505 /*
 506  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 507  * using sendpage(). Return the number of bytes sent.
 508  */
 509 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 510                             struct pipe_buffer *buf, struct splice_desc *sd)
 511 {
 512         struct file *file = sd->u.file;
 513         loff_t pos = sd->pos;
 514         int ret, more;
 515
 516         ret = buf->ops->confirm(pipe, buf);
 517         if (!ret) {
 518                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 519
 520                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 521                                            sd->len, &pos, more);
 522         }
 523
 524         return ret;
 525 }
 526
 527 /*
 528  * This is a little more tricky than the file -> pipe splicing. There are
 529  * basically three cases:
 530  *
 531  *      - Destination page already exists in the address space and there
 532  *        are users of it. For that case we have no other option that
 533  *        copying the data. Tough luck.
 534  *      - Destination page already exists in the address space, but there
 535  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 536  *        through to last case.
 537  *      - Destination page does not exist, we can add the pipe page to
 538  *        the page cache and avoid the copy.
 539  *
 540  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 541  * sd->flags), we attempt to migrate pages from the pipe to the output
 542  * file address space page cache. This is possible if no one else has
 543  * the pipe page referenced outside of the pipe and page cache. If
 544  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 545  * a new page in the output file page cache and fill/dirty that.
 546  */
 547 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 548                         struct splice_desc *sd)
 549 {
 550         struct file *file = sd->u.file;
 551         struct address_space *mapping = file->f_mapping;
 552         unsigned int offset, this_len;
 553         struct page *page;
 554         void *fsdata;
 555         int ret;
 556
 557         /*
 558          * make sure the data in this buffer is uptodate
 559          */
 560         ret = buf->ops->confirm(pipe, buf);
 561         if (unlikely(ret))
 562                 return ret;
 563
 564         offset = sd->pos & ~PAGE_CACHE_MASK;
 565
 566         this_len = sd->len;
 567         if (this_len + offset > PAGE_CACHE_SIZE)
 568                 this_len = PAGE_CACHE_SIZE - offset;
 569
 570         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 571                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 572         if (unlikely(ret))
 573                 goto out;
 574
 575         if (buf->page != page) {
 576                 /*
 577                  * Careful, ->map() uses KM_USER0!
 578                  */
 579                 char *src = buf->ops->map(pipe, buf, 1);
 580                 char *dst = kmap_atomic(page, KM_USER1);
 581
 582                 memcpy(dst + offset, src + buf->offset, this_len);
 583                 flush_dcache_page(page);
 584                 kunmap_atomic(dst, KM_USER1);
 585                 buf->ops->unmap(pipe, buf, src);
 586         }
 587         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 588                                 page, fsdata);
 589 out:
 590         return ret;
 591 }
 592
 593 /**
 594  * __splice_from_pipe - splice data from a pipe to given actor
 595  * @pipe:       pipe to splice from
 596  * @sd:         information to @actor
 597  * @actor:      handler that splices the data
 598  *
 599  * Description:
 600  *    This function does little more than loop over the pipe and call
 601  *    @actor to do the actual moving of a single struct pipe_buffer to
 602  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 603  *    pipe_to_user.
 604  *
 605  */
 606 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 607                            splice_actor *actor)
 608 {
 609         int ret, do_wakeup, err;
 610
 611         ret = 0;
 612         do_wakeup = 0;
 613
 614         for (;;) {
 615                 if (pipe->nrbufs) {
 616                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 617                         const struct pipe_buf_operations *ops = buf->ops;
 618
 619                         sd->len = buf->len;
 620                         if (sd->len > sd->total_len)
 621                                 sd->len = sd->total_len;
 622
 623                         err = actor(pipe, buf, sd);
 624                         if (err <= 0) {
 625                                 if (!ret && err != -ENODATA)
 626                                         ret = err;
 627
 628                                 break;
 629                         }
 630
 631                         ret += err;
 632                         buf->offset += err;
 633                         buf->len -= err;
 634
 635                         sd->len -= err;
 636                         sd->pos += err;
 637                         sd->total_len -= err;
 638                         if (sd->len)
 639                                 continue;
 640
 641                         if (!buf->len) {
 642                                 buf->ops = NULL;
 643                                 ops->release(pipe, buf);
 644                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 645                                 pipe->nrbufs--;
 646                                 if (pipe->inode)
 647                                         do_wakeup = 1;
 648                         }
 649
 650                         if (!sd->total_len)
 651                                 break;
 652                 }
 653
 654                 if (pipe->nrbufs)
 655                         continue;
 656                 if (!pipe->writers)
 657                         break;
 658                 if (!pipe->waiting_writers) {
 659                         if (ret)
 660                                 break;
 661                 }
 662
 663                 if (sd->flags & SPLICE_F_NONBLOCK) {
 664                         if (!ret)
 665                                 ret = -EAGAIN;
 666                         break;
 667                 }
 668
 669                 if (signal_pending(current)) {
 670                         if (!ret)
 671                                 ret = -ERESTARTSYS;
 672                         break;
 673                 }
 674
 675                 if (do_wakeup) {
 676                         smp_mb();
 677                         if (waitqueue_active(&pipe->wait))
 678                                 wake_up_interruptible_sync(&pipe->wait);
 679                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 680                         do_wakeup = 0;
 681                 }
 682
 683                 pipe_wait(pipe);
 684         }
 685
 686         if (do_wakeup) {
 687                 smp_mb();
 688                 if (waitqueue_active(&pipe->wait))
 689                         wake_up_interruptible(&pipe->wait);
 690                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 691         }
 692
 693         return ret;
 694 }
 695 EXPORT_SYMBOL(__splice_from_pipe);
 696
 697 /**
 698  * splice_from_pipe - splice data from a pipe to a file
 699  * @pipe:       pipe to splice from
 700  * @out:        file to splice to
 701  * @ppos:       position in @out
 702  * @len:        how many bytes to splice
 703  * @flags:      splice modifier flags
 704  * @actor:      handler that splices the data
 705  *
 706  * Description:
 707  *    See __splice_from_pipe. This function locks the input and output inodes,
 708  *    otherwise it's identical to __splice_from_pipe().
 709  *
 710  */
 711 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 712                          loff_t *ppos, size_t len, unsigned int flags,
 713                          splice_actor *actor)
 714 {
 715         ssize_t ret;
 716         struct inode *inode = out->f_mapping->host;
 717         struct splice_desc sd = {
 718                 .total_len = len,
 719                 .flags = flags,
 720                 .pos = *ppos,
 721                 .u.file = out,
 722         };
 723
 724         /*
 725          * The actor worker might be calling ->prepare_write and
 726          * ->commit_write. Most of the time, these expect i_mutex to
 727          * be held. Since this may result in an ABBA deadlock with
 728          * pipe->inode, we have to order lock acquiry here.
 729          */
 730         inode_double_lock(inode, pipe->inode);
 731         ret = __splice_from_pipe(pipe, &sd, actor);
 732         inode_double_unlock(inode, pipe->inode);
 733
 734         return ret;
 735 }
 736
 737 /**
 738  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
 739  * @pipe:       pipe info
 740  * @out:        file to write to
 741  * @ppos:       position in @out
 742  * @len:        number of bytes to splice
 743  * @flags:      splice modifier flags
 744  *
 745  * Description:
 746  *    Will either move or copy pages (determined by @flags options) from
 747  *    the given pipe inode to the given file. The caller is responsible
 748  *    for acquiring i_mutex on both inodes.
 749  *
 750  */
 751 ssize_t
 752 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
 753                                  loff_t *ppos, size_t len, unsigned int flags)
 754 {
 755         struct address_space *mapping = out->f_mapping;
 756         struct inode *inode = mapping->host;
 757         struct splice_desc sd = {
 758                 .total_len = len,
 759                 .flags = flags,
 760                 .pos = *ppos,
 761                 .u.file = out,
 762         };
 763         ssize_t ret;
 764         int err;
 765
 766         err = remove_suid(out->f_path.dentry);
 767         if (unlikely(err))
 768                 return err;
 769
 770         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
 771         if (ret > 0) {
 772                 unsigned long nr_pages;
 773
 774                 *ppos += ret;
 775                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 776
 777                 /*
 778                  * If file or inode is SYNC and we actually wrote some data,
 779                  * sync it.
 780                  */
 781                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 782                         err = generic_osync_inode(inode, mapping,
 783                                                   OSYNC_METADATA|OSYNC_DATA);
 784
 785                         if (err)
 786                                 ret = err;
 787                 }
 788                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 789         }
 790
 791         return ret;
 792 }
 793
 794 EXPORT_SYMBOL(generic_file_splice_write_nolock);
 795
 796 /**
 797  * generic_file_splice_write - splice data from a pipe to a file
 798  * @pipe:       pipe info
 799  * @out:        file to write to
 800  * @ppos:       position in @out
 801  * @len:        number of bytes to splice
 802  * @flags:      splice modifier flags
 803  *
 804  * Description:
 805  *    Will either move or copy pages (determined by @flags options) from
 806  *    the given pipe inode to the given file.
 807  *
 808  */
 809 ssize_t
 810 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 811                           loff_t *ppos, size_t len, unsigned int flags)
 812 {
 813         struct address_space *mapping = out->f_mapping;
 814         struct inode *inode = mapping->host;
 815         struct splice_desc sd = {
 816                 .total_len = len,
 817                 .flags = flags,
 818                 .pos = *ppos,
 819                 .u.file = out,
 820         };
 821         ssize_t ret;
 822
 823         inode_double_lock(inode, pipe->inode);
 824         ret = remove_suid(out->f_path.dentry);
 825         if (likely(!ret))
 826                 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
 827         inode_double_unlock(inode, pipe->inode);
 828         if (ret > 0) {
 829                 unsigned long nr_pages;
 830
 831                 *ppos += ret;
 832                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 833
 834                 /*
 835                  * If file or inode is SYNC and we actually wrote some data,
 836                  * sync it.
 837                  */
 838                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
 839                         int err;
 840
 841                         mutex_lock(&inode->i_mutex);
 842                         err = generic_osync_inode(inode, mapping,
 843                                                   OSYNC_METADATA|OSYNC_DATA);
 844                         mutex_unlock(&inode->i_mutex);
 845
 846                         if (err)
 847                                 ret = err;
 848                 }
 849                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 850         }
 851
 852         return ret;
 853 }
 854
 855 EXPORT_SYMBOL(generic_file_splice_write);
 856
 857 /**
 858  * generic_splice_sendpage - splice data from a pipe to a socket
 859  * @pipe:       pipe to splice from
 860  * @out:        socket to write to
 861  * @ppos:       position in @out
 862  * @len:        number of bytes to splice
 863  * @flags:      splice modifier flags
 864  *
 865  * Description:
 866  *    Will send @len bytes from the pipe to a network socket. No data copying
 867  *    is involved.
 868  *
 869  */
 870 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 871                                 loff_t *ppos, size_t len, unsigned int flags)
 872 {
 873         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 874 }
 875
 876 EXPORT_SYMBOL(generic_splice_sendpage);
 877
 878 /*
 879  * Attempt to initiate a splice from pipe to file.
 880  */
 881 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 882                            loff_t *ppos, size_t len, unsigned int flags)
 883 {
 884         int ret;
 885
 886         if (unlikely(!out->f_op || !out->f_op->splice_write))
 887                 return -EINVAL;
 888
 889         if (unlikely(!(out->f_mode & FMODE_WRITE)))
 890                 return -EBADF;
 891
 892         ret = rw_verify_area(WRITE, out, ppos, len);
 893         if (unlikely(ret < 0))
 894                 return ret;
 895
 896         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 897 }
 898
 899 /*
 900  * Attempt to initiate a splice from a file to a pipe.
 901  */
 902 static long do_splice_to(struct file *in, loff_t *ppos,
 903                          struct pipe_inode_info *pipe, size_t len,
 904                          unsigned int flags)
 905 {
 906         int ret;
 907
 908         if (unlikely(!in->f_op || !in->f_op->splice_read))
 909                 return -EINVAL;
 910
 911         if (unlikely(!(in->f_mode & FMODE_READ)))
 912                 return -EBADF;
 913
 914         ret = rw_verify_area(READ, in, ppos, len);
 915         if (unlikely(ret < 0))
 916                 return ret;
 917
 918         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 919 }
 920
 921 /**
 922  * splice_direct_to_actor - splices data directly between two non-pipes
 923  * @in:         file to splice from
 924  * @sd:         actor information on where to splice to
 925  * @actor:      handles the data splicing
 926  *
 927  * Description:
 928  *    This is a special case helper to splice directly between two
 929  *    points, without requiring an explicit pipe. Internally an allocated
 930  *    pipe is cached in the process, and reused during the lifetime of
 931  *    that process.
 932  *
 933  */
 934 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 935                                splice_direct_actor *actor)
 936 {
 937         struct pipe_inode_info *pipe;
 938         long ret, bytes;
 939         umode_t i_mode;
 940         size_t len;
 941         int i, flags;
 942
 943         /*
 944          * We require the input being a regular file, as we don't want to
 945          * randomly drop data for eg socket -> socket splicing. Use the
 946          * piped splicing for that!
 947          */
 948         i_mode = in->f_path.dentry->d_inode->i_mode;
 949         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 950                 return -EINVAL;
 951
 952         /*
 953          * neither in nor out is a pipe, setup an internal pipe attached to
 954          * 'out' and transfer the wanted data from 'in' to 'out' through that
 955          */
 956         pipe = current->splice_pipe;
 957         if (unlikely(!pipe)) {
 958                 pipe = alloc_pipe_info(NULL);
 959                 if (!pipe)
 960                         return -ENOMEM;
 961
 962                 /*
 963                  * We don't have an immediate reader, but we'll read the stuff
 964                  * out of the pipe right after the splice_to_pipe(). So set
 965                  * PIPE_READERS appropriately.
 966                  */
 967                 pipe->readers = 1;
 968
 969                 current->splice_pipe = pipe;
 970         }
 971
 972         /*
 973          * Do the splice.
 974          */
 975         ret = 0;
 976         bytes = 0;
 977         len = sd->total_len;
 978         flags = sd->flags;
 979
 980         /*
 981          * Don't block on output, we have to drain the direct pipe.
 982          */
 983         sd->flags &= ~SPLICE_F_NONBLOCK;
 984
 985         while (len) {
 986                 size_t read_len;
 987                 loff_t pos = sd->pos, prev_pos = pos;
 988
 989                 ret = do_splice_to(in, &pos, pipe, len, flags);
 990                 if (unlikely(ret <= 0))
 991                         goto out_release;
 992
 993                 read_len = ret;
 994                 sd->total_len = read_len;
 995
 996                 /*
 997                  * NOTE: nonblocking mode only applies to the input. We
 998                  * must not do the output in nonblocking mode as then we
 999                  * could get stuck data in the internal pipe:
1000                  */
1001                 ret = actor(pipe, sd);
1002                 if (unlikely(ret <= 0)) {
1003                         sd->pos = prev_pos;
1004                         goto out_release;
1005                 }
1006
1007                 bytes += ret;
1008                 len -= ret;
1009                 sd->pos = pos;
1010
1011                 if (ret < read_len) {
1012                         sd->pos = prev_pos + ret;
1013                         goto out_release;
1014                 }
1015         }
1016
1017 done:
1018         pipe->nrbufs = pipe->curbuf = 0;
1019         file_accessed(in);
1020         return bytes;
1021
1022 out_release:
1023         /*
1024          * If we did an incomplete transfer we must release
1025          * the pipe buffers in question:
1026          */
1027         for (i = 0; i < PIPE_BUFFERS; i++) {
1028                 struct pipe_buffer *buf = pipe->bufs + i;
1029
1030                 if (buf->ops) {
1031                         buf->ops->release(pipe, buf);
1032                         buf->ops = NULL;
1033                 }
1034         }
1035
1036         if (!bytes)
1037                 bytes = ret;
1038
1039         goto done;
1040 }
1041 EXPORT_SYMBOL(splice_direct_to_actor);
1042
1043 static int direct_splice_actor(struct pipe_inode_info *pipe,
1044                                struct splice_desc *sd)
1045 {
1046         struct file *file = sd->u.file;
1047
1048         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1049 }
1050
1051 /**
1052  * do_splice_direct - splices data directly between two files
1053  * @in:         file to splice from
1054  * @ppos:       input file offset
1055  * @out:        file to splice to
1056  * @len:        number of bytes to splice
1057  * @flags:      splice modifier flags
1058  *
1059  * Description:
1060  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1061  *    doing it in the application would incur an extra system call
1062  *    (splice in + splice out, as compared to just sendfile()). So this helper
1063  *    can splice directly through a process-private pipe.
1064  *
1065  */
1066 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1067                       size_t len, unsigned int flags)
1068 {
1069         struct splice_desc sd = {
1070                 .len            = len,
1071                 .total_len      = len,
1072                 .flags          = flags,
1073                 .pos            = *ppos,
1074                 .u.file         = out,
1075         };
1076         long ret;
1077
1078         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1079         if (ret > 0)
1080                 *ppos = sd.pos;
1081
1082         return ret;
1083 }
1084
1085 /*
1086  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1087  * location, so checking ->i_pipe is not enough to verify that this is a
1088  * pipe.
1089  */
1090 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1091 {
1092         if (S_ISFIFO(inode->i_mode))
1093                 return inode->i_pipe;
1094
1095         return NULL;
1096 }
1097
1098 /*
1099  * Determine where to splice to/from.
1100  */
1101 static long do_splice(struct file *in, loff_t __user *off_in,
1102                       struct file *out, loff_t __user *off_out,
1103                       size_t len, unsigned int flags)
1104 {
1105         struct pipe_inode_info *pipe;
1106         loff_t offset, *off;
1107         long ret;
1108
1109         pipe = pipe_info(in->f_path.dentry->d_inode);
1110         if (pipe) {
1111                 if (off_in)
1112                         return -ESPIPE;
1113                 if (off_out) {
1114                         if (out->f_op->llseek == no_llseek)
1115                                 return -EINVAL;
1116                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1117                                 return -EFAULT;
1118                         off = &offset;
1119                 } else
1120                         off = &out->f_pos;
1121
1122                 ret = do_splice_from(pipe, out, off, len, flags);
1123
1124                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1125                         ret = -EFAULT;
1126
1127                 return ret;
1128         }
1129
1130         pipe = pipe_info(out->f_path.dentry->d_inode);
1131         if (pipe) {
1132                 if (off_out)
1133                         return -ESPIPE;
1134                 if (off_in) {
1135                         if (in->f_op->llseek == no_llseek)
1136                                 return -EINVAL;
1137                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1138                                 return -EFAULT;
1139                         off = &offset;
1140                 } else
1141                         off = &in->f_pos;
1142
1143                 ret = do_splice_to(in, off, pipe, len, flags);
1144
1145                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1146                         ret = -EFAULT;
1147
1148                 return ret;
1149         }
1150
1151         return -EINVAL;
1152 }
1153
1154 /*
1155  * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1156  * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1157  * for writing) and page faulting on the user memory pointed to by src.
1158  * This assumes that we will very rarely hit the partial != 0 path, or this
1159  * will not be a win.
1160  */
1161 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1162 {
1163         int partial;
1164
1165         if (!access_ok(VERIFY_READ, src, n))
1166                 return -EFAULT;
1167
1168         pagefault_disable();
1169         partial = __copy_from_user_inatomic(dst, src, n);
1170         pagefault_enable();
1171
1172         /*
1173          * Didn't copy everything, drop the mmap_sem and do a faulting copy
1174          */
1175         if (unlikely(partial)) {
1176                 up_read(&current->mm->mmap_sem);
1177                 partial = copy_from_user(dst, src, n);
1178                 down_read(&current->mm->mmap_sem);
1179         }
1180
1181         return partial;
1182 }
1183
1184 /*
1185  * Map an iov into an array of pages and offset/length tupples. With the
1186  * partial_page structure, we can map several non-contiguous ranges into
1187  * our ones pages[] map instead of splitting that operation into pieces.
1188  * Could easily be exported as a generic helper for other users, in which
1189  * case one would probably want to add a 'max_nr_pages' parameter as well.
1190  */
1191 static int get_iovec_page_array(const struct iovec __user *iov,
1192                                 unsigned int nr_vecs, struct page **pages,
1193                                 struct partial_page *partial, int aligned)
1194 {
1195         int buffers = 0, error = 0;
1196
1197         down_read(&current->mm->mmap_sem);
1198
1199         while (nr_vecs) {
1200                 unsigned long off, npages;
1201                 struct iovec entry;
1202                 void __user *base;
1203                 size_t len;
1204                 int i;
1205
1206                 error = -EFAULT;
1207                 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1208                         break;
1209
1210                 base = entry.iov_base;
1211                 len = entry.iov_len;
1212
1213                 /*
1214                  * Sanity check this iovec. 0 read succeeds.
1215                  */
1216                 error = 0;
1217                 if (unlikely(!len))
1218                         break;
1219                 error = -EFAULT;
1220                 if (!access_ok(VERIFY_READ, base, len))
1221                         break;
1222
1223                 /*
1224                  * Get this base offset and number of pages, then map
1225                  * in the user pages.
1226                  */
1227                 off = (unsigned long) base & ~PAGE_MASK;
1228
1229                 /*
1230                  * If asked for alignment, the offset must be zero and the
1231                  * length a multiple of the PAGE_SIZE.
1232                  */
1233                 error = -EINVAL;
1234                 if (aligned && (off || len & ~PAGE_MASK))
1235                         break;
1236
1237                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1238                 if (npages > PIPE_BUFFERS - buffers)
1239                         npages = PIPE_BUFFERS - buffers;
1240
1241                 error = get_user_pages(current, current->mm,
1242                                        (unsigned long) base, npages, 0, 0,
1243                                        &pages[buffers], NULL);
1244
1245                 if (unlikely(error <= 0))
1246                         break;
1247
1248                 /*
1249                  * Fill this contiguous range into the partial page map.
1250                  */
1251                 for (i = 0; i < error; i++) {
1252                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1253
1254                         partial[buffers].offset = off;
1255                         partial[buffers].len = plen;
1256
1257                         off = 0;
1258                         len -= plen;
1259                         buffers++;
1260                 }
1261
1262                 /*
1263                  * We didn't complete this iov, stop here since it probably
1264                  * means we have to move some of this into a pipe to
1265                  * be able to continue.
1266                  */
1267                 if (len)
1268                         break;
1269
1270                 /*
1271                  * Don't continue if we mapped fewer pages than we asked for,
1272                  * or if we mapped the max number of pages that we have
1273                  * room for.
1274                  */
1275                 if (error < npages || buffers == PIPE_BUFFERS)
1276                         break;
1277
1278                 nr_vecs--;
1279                 iov++;
1280         }
1281
1282         up_read(&current->mm->mmap_sem);
1283
1284         if (buffers)
1285                 return buffers;
1286
1287         return error;
1288 }
1289
1290 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1291                         struct splice_desc *sd)
1292 {
1293         char *src;
1294         int ret;
1295
1296         ret = buf->ops->confirm(pipe, buf);
1297         if (unlikely(ret))
1298                 return ret;
1299
1300         /*
1301          * See if we can use the atomic maps, by prefaulting in the
1302          * pages and doing an atomic copy
1303          */
1304         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1305                 src = buf->ops->map(pipe, buf, 1);
1306                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1307                                                         sd->len);
1308                 buf->ops->unmap(pipe, buf, src);
1309                 if (!ret) {
1310                         ret = sd->len;
1311                         goto out;
1312                 }
1313         }
1314
1315         /*
1316          * No dice, use slow non-atomic map and copy
1317          */
1318         src = buf->ops->map(pipe, buf, 0);
1319
1320         ret = sd->len;
1321         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1322                 ret = -EFAULT;
1323
1324         buf->ops->unmap(pipe, buf, src);
1325 out:
1326         if (ret > 0)
1327                 sd->u.userptr += ret;
1328         return ret;
1329 }
1330
1331 /*
1332  * For lack of a better implementation, implement vmsplice() to userspace
1333  * as a simple copy of the pipes pages to the user iov.
1334  */
1335 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1336                              unsigned long nr_segs, unsigned int flags)
1337 {
1338         struct pipe_inode_info *pipe;
1339         struct splice_desc sd;
1340         ssize_t size;
1341         int error;
1342         long ret;
1343
1344         pipe = pipe_info(file->f_path.dentry->d_inode);
1345         if (!pipe)
1346                 return -EBADF;
1347
1348         if (pipe->inode)
1349                 mutex_lock(&pipe->inode->i_mutex);
1350
1351         error = ret = 0;
1352         while (nr_segs) {
1353                 void __user *base;
1354                 size_t len;
1355
1356                 /*
1357                  * Get user address base and length for this iovec.
1358                  */
1359                 error = get_user(base, &iov->iov_base);
1360                 if (unlikely(error))
1361                         break;
1362                 error = get_user(len, &iov->iov_len);
1363                 if (unlikely(error))
1364                         break;
1365
1366                 /*
1367                  * Sanity check this iovec. 0 read succeeds.
1368                  */
1369                 if (unlikely(!len))
1370                         break;
1371                 if (unlikely(!base)) {
1372                         error = -EFAULT;
1373                         break;
1374                 }
1375
1376                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1377                         error = -EFAULT;
1378                         break;
1379                 }
1380
1381                 sd.len = 0;
1382                 sd.total_len = len;
1383                 sd.flags = flags;
1384                 sd.u.userptr = base;
1385                 sd.pos = 0;
1386
1387                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1388                 if (size < 0) {
1389                         if (!ret)
1390                                 ret = size;
1391
1392                         break;
1393                 }
1394
1395                 ret += size;
1396
1397                 if (size < len)
1398                         break;
1399
1400                 nr_segs--;
1401                 iov++;
1402         }
1403
1404         if (pipe->inode)
1405                 mutex_unlock(&pipe->inode->i_mutex);
1406
1407         if (!ret)
1408                 ret = error;
1409
1410         return ret;
1411 }
1412
1413 /*
1414  * vmsplice splices a user address range into a pipe. It can be thought of
1415  * as splice-from-memory, where the regular splice is splice-from-file (or
1416  * to file). In both cases the output is a pipe, naturally.
1417  */
1418 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1419                              unsigned long nr_segs, unsigned int flags)
1420 {
1421         struct pipe_inode_info *pipe;
1422         struct page *pages[PIPE_BUFFERS];
1423         struct partial_page partial[PIPE_BUFFERS];
1424         struct splice_pipe_desc spd = {
1425                 .pages = pages,
1426                 .partial = partial,
1427                 .flags = flags,
1428                 .ops = &user_page_pipe_buf_ops,
1429                 .spd_release = spd_release_page,
1430         };
1431
1432         pipe = pipe_info(file->f_path.dentry->d_inode);
1433         if (!pipe)
1434                 return -EBADF;
1435
1436         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1437                                             flags & SPLICE_F_GIFT);
1438         if (spd.nr_pages <= 0)
1439                 return spd.nr_pages;
1440
1441         return splice_to_pipe(pipe, &spd);
1442 }
1443
1444 /*
1445  * Note that vmsplice only really supports true splicing _from_ user memory
1446  * to a pipe, not the other way around. Splicing from user memory is a simple
1447  * operation that can be supported without any funky alignment restrictions
1448  * or nasty vm tricks. We simply map in the user memory and fill them into
1449  * a pipe. The reverse isn't quite as easy, though. There are two possible
1450  * solutions for that:
1451  *
1452  *      - memcpy() the data internally, at which point we might as well just
1453  *        do a regular read() on the buffer anyway.
1454  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1455  *        has restriction limitations on both ends of the pipe).
1456  *
1457  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1458  *
1459  */
1460 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1461                              unsigned long nr_segs, unsigned int flags)
1462 {
1463         struct file *file;
1464         long error;
1465         int fput;
1466
1467         if (unlikely(nr_segs > UIO_MAXIOV))
1468                 return -EINVAL;
1469         else if (unlikely(!nr_segs))
1470                 return 0;
1471
1472         error = -EBADF;
1473         file = fget_light(fd, &fput);
1474         if (file) {
1475                 if (file->f_mode & FMODE_WRITE)
1476                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1477                 else if (file->f_mode & FMODE_READ)
1478                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1479
1480                 fput_light(file, fput);
1481         }
1482
1483         return error;
1484 }
1485
1486 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1487                            int fd_out, loff_t __user *off_out,
1488                            size_t len, unsigned int flags)
1489 {
1490         long error;
1491         struct file *in, *out;
1492         int fput_in, fput_out;
1493
1494         if (unlikely(!len))
1495                 return 0;
1496
1497         error = -EBADF;
1498         in = fget_light(fd_in, &fput_in);
1499         if (in) {
1500                 if (in->f_mode & FMODE_READ) {
1501                         out = fget_light(fd_out, &fput_out);
1502                         if (out) {
1503                                 if (out->f_mode & FMODE_WRITE)
1504                                         error = do_splice(in, off_in,
1505                                                           out, off_out,
1506                                                           len, flags);
1507                                 fput_light(out, fput_out);
1508                         }
1509                 }
1510
1511                 fput_light(in, fput_in);
1512         }
1513
1514         return error;
1515 }
1516
1517 /*
1518  * Make sure there's data to read. Wait for input if we can, otherwise
1519  * return an appropriate error.
1520  */
1521 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1522 {
1523         int ret;
1524
1525         /*
1526          * Check ->nrbufs without the inode lock first. This function
1527          * is speculative anyways, so missing one is ok.
1528          */
1529         if (pipe->nrbufs)
1530                 return 0;
1531
1532         ret = 0;
1533         mutex_lock(&pipe->inode->i_mutex);
1534
1535         while (!pipe->nrbufs) {
1536                 if (signal_pending(current)) {
1537                         ret = -ERESTARTSYS;
1538                         break;
1539                 }
1540                 if (!pipe->writers)
1541                         break;
1542                 if (!pipe->waiting_writers) {
1543                         if (flags & SPLICE_F_NONBLOCK) {
1544                                 ret = -EAGAIN;
1545                                 break;
1546                         }
1547                 }
1548                 pipe_wait(pipe);
1549         }
1550
1551         mutex_unlock(&pipe->inode->i_mutex);
1552         return ret;
1553 }
1554
1555 /*
1556  * Make sure there's writeable room. Wait for room if we can, otherwise
1557  * return an appropriate error.
1558  */
1559 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1560 {
1561         int ret;
1562
1563         /*
1564          * Check ->nrbufs without the inode lock first. This function
1565          * is speculative anyways, so missing one is ok.
1566          */
1567         if (pipe->nrbufs < PIPE_BUFFERS)
1568                 return 0;
1569
1570         ret = 0;
1571         mutex_lock(&pipe->inode->i_mutex);
1572
1573         while (pipe->nrbufs >= PIPE_BUFFERS) {
1574                 if (!pipe->readers) {
1575                         send_sig(SIGPIPE, current, 0);
1576                         ret = -EPIPE;
1577                         break;
1578                 }
1579                 if (flags & SPLICE_F_NONBLOCK) {
1580                         ret = -EAGAIN;
1581                         break;
1582                 }
1583                 if (signal_pending(current)) {
1584                         ret = -ERESTARTSYS;
1585                         break;
1586                 }
1587                 pipe->waiting_writers++;
1588                 pipe_wait(pipe);
1589                 pipe->waiting_writers--;
1590         }
1591
1592         mutex_unlock(&pipe->inode->i_mutex);
1593         return ret;
1594 }
1595
1596 /*
1597  * Link contents of ipipe to opipe.
1598  */
1599 static int link_pipe(struct pipe_inode_info *ipipe,
1600                      struct pipe_inode_info *opipe,
1601                      size_t len, unsigned int flags)
1602 {
1603         struct pipe_buffer *ibuf, *obuf;
1604         int ret = 0, i = 0, nbuf;
1605
1606         /*
1607          * Potential ABBA deadlock, work around it by ordering lock
1608          * grabbing by inode address. Otherwise two different processes
1609          * could deadlock (one doing tee from A -> B, the other from B -> A).
1610          */
1611         inode_double_lock(ipipe->inode, opipe->inode);
1612
1613         do {
1614                 if (!opipe->readers) {
1615                         send_sig(SIGPIPE, current, 0);
1616                         if (!ret)
1617                                 ret = -EPIPE;
1618                         break;
1619                 }
1620
1621                 /*
1622                  * If we have iterated all input buffers or ran out of
1623                  * output room, break.
1624                  */
1625                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1626                         break;
1627
1628                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1629                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1630
1631                 /*
1632                  * Get a reference to this pipe buffer,
1633                  * so we can copy the contents over.
1634                  */
1635                 ibuf->ops->get(ipipe, ibuf);
1636
1637                 obuf = opipe->bufs + nbuf;
1638                 *obuf = *ibuf;
1639
1640                 /*
1641                  * Don't inherit the gift flag, we need to
1642                  * prevent multiple steals of this page.
1643                  */
1644                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1645
1646                 if (obuf->len > len)
1647                         obuf->len = len;
1648
1649                 opipe->nrbufs++;
1650                 ret += obuf->len;
1651                 len -= obuf->len;
1652                 i++;
1653         } while (len);
1654
1655         /*
1656          * return EAGAIN if we have the potential of some data in the
1657          * future, otherwise just return 0
1658          */
1659         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1660                 ret = -EAGAIN;
1661
1662         inode_double_unlock(ipipe->inode, opipe->inode);
1663
1664         /*
1665          * If we put data in the output pipe, wakeup any potential readers.
1666          */
1667         if (ret > 0) {
1668                 smp_mb();
1669                 if (waitqueue_active(&opipe->wait))
1670                         wake_up_interruptible(&opipe->wait);
1671                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1672         }
1673
1674         return ret;
1675 }
1676
1677 /*
1678  * This is a tee(1) implementation that works on pipes. It doesn't copy
1679  * any data, it simply references the 'in' pages on the 'out' pipe.
1680  * The 'flags' used are the SPLICE_F_* variants, currently the only
1681  * applicable one is SPLICE_F_NONBLOCK.
1682  */
1683 static long do_tee(struct file *in, struct file *out, size_t len,
1684                    unsigned int flags)
1685 {
1686         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1687         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1688         int ret = -EINVAL;
1689
1690         /*
1691          * Duplicate the contents of ipipe to opipe without actually
1692          * copying the data.
1693          */
1694         if (ipipe && opipe && ipipe != opipe) {
1695                 /*
1696                  * Keep going, unless we encounter an error. The ipipe/opipe
1697                  * ordering doesn't really matter.
1698                  */
1699                 ret = link_ipipe_prep(ipipe, flags);
1700                 if (!ret) {
1701                         ret = link_opipe_prep(opipe, flags);
1702                         if (!ret)
1703                                 ret = link_pipe(ipipe, opipe, len, flags);
1704                 }
1705         }
1706
1707         return ret;
1708 }
1709
1710 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1711 {
1712         struct file *in;
1713         int error, fput_in;
1714
1715         if (unlikely(!len))
1716                 return 0;
1717
1718         error = -EBADF;
1719         in = fget_light(fdin, &fput_in);
1720         if (in) {
1721                 if (in->f_mode & FMODE_READ) {
1722                         int fput_out;
1723                         struct file *out = fget_light(fdout, &fput_out);
1724
1725                         if (out) {
1726                                 if (out->f_mode & FMODE_WRITE)
1727                                         error = do_tee(in, out, len, flags);
1728                                 fput_light(out, fput_out);
1729                         }
1730                 }
1731                 fput_light(in, fput_in);
1732         }
1733
1734         return error;
1735 }