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