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