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/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/module.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33
  34 /*
  35  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  36  * a vm helper function, it's already simplified quite a bit by the
  37  * addition of remove_mapping(). If success is returned, the caller may
  38  * attempt to reuse this page for another destination.
  39  */
  40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  41                                      struct pipe_buffer *buf)
  42 {
  43         struct page *page = buf->page;
  44         struct address_space *mapping;
  45
  46         lock_page(page);
  47
  48         mapping = page_mapping(page);
  49         if (mapping) {
  50                 WARN_ON(!PageUptodate(page));
  51
  52                 /*
  53                  * At least for ext2 with nobh option, we need to wait on
  54                  * writeback completing on this page, since we'll remove it
  55                  * from the pagecache.  Otherwise truncate wont wait on the
  56                  * page, allowing the disk blocks to be reused by someone else
  57                  * before we actually wrote our data to them. fs corruption
  58                  * ensues.
  59                  */
  60                 wait_on_page_writeback(page);
  61
  62                 if (page_has_private(page) &&
  63                     !try_to_release_page(page, GFP_KERNEL))
  64                         goto out_unlock;
  65
  66                 /*
  67                  * If we succeeded in removing the mapping, set LRU flag
  68                  * and return good.
  69                  */
  70                 if (remove_mapping(mapping, page)) {
  71                         buf->flags |= PIPE_BUF_FLAG_LRU;
  72                         return 0;
  73                 }
  74         }
  75
  76         /*
  77          * Raced with truncate or failed to remove page from current
  78          * address space, unlock and return failure.
  79          */
  80 out_unlock:
  81         unlock_page(page);
  82         return 1;
  83 }
  84
  85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  86                                         struct pipe_buffer *buf)
  87 {
  88         page_cache_release(buf->page);
  89         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  90 }
  91
  92 /*
  93  * Check whether the contents of buf is OK to access. Since the content
  94  * is a page cache page, IO may be in flight.
  95  */
  96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  97                                        struct pipe_buffer *buf)
  98 {
  99         struct page *page = buf->page;
 100         int err;
 101
 102         if (!PageUptodate(page)) {
 103                 lock_page(page);
 104
 105                 /*
 106                  * Page got truncated/unhashed. This will cause a 0-byte
 107                  * splice, if this is the first page.
 108                  */
 109                 if (!page->mapping) {
 110                         err = -ENODATA;
 111                         goto error;
 112                 }
 113
 114                 /*
 115                  * Uh oh, read-error from disk.
 116                  */
 117                 if (!PageUptodate(page)) {
 118                         err = -EIO;
 119                         goto error;
 120                 }
 121
 122                 /*
 123                  * Page is ok afterall, we are done.
 124                  */
 125                 unlock_page(page);
 126         }
 127
 128         return 0;
 129 error:
 130         unlock_page(page);
 131         return err;
 132 }
 133
 134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 135         .can_merge = 0,
 136         .map = generic_pipe_buf_map,
 137         .unmap = generic_pipe_buf_unmap,
 138         .confirm = page_cache_pipe_buf_confirm,
 139         .release = page_cache_pipe_buf_release,
 140         .steal = page_cache_pipe_buf_steal,
 141         .get = generic_pipe_buf_get,
 142 };
 143
 144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 145                                     struct pipe_buffer *buf)
 146 {
 147         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 148                 return 1;
 149
 150         buf->flags |= PIPE_BUF_FLAG_LRU;
 151         return generic_pipe_buf_steal(pipe, buf);
 152 }
 153
 154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 155         .can_merge = 0,
 156         .map = generic_pipe_buf_map,
 157         .unmap = generic_pipe_buf_unmap,
 158         .confirm = generic_pipe_buf_confirm,
 159         .release = page_cache_pipe_buf_release,
 160         .steal = user_page_pipe_buf_steal,
 161         .get = generic_pipe_buf_get,
 162 };
 163
 164 /**
 165  * splice_to_pipe - fill passed data into a pipe
 166  * @pipe:       pipe to fill
 167  * @spd:        data to fill
 168  *
 169  * Description:
 170  *    @spd contains a map of pages and len/offset tuples, along with
 171  *    the struct pipe_buf_operations associated with these pages. This
 172  *    function will link that data to the pipe.
 173  *
 174  */
 175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 176                        struct splice_pipe_desc *spd)
 177 {
 178         unsigned int spd_pages = spd->nr_pages;
 179         int ret, do_wakeup, page_nr;
 180
 181         ret = 0;
 182         do_wakeup = 0;
 183         page_nr = 0;
 184
 185         pipe_lock(pipe);
 186
 187         for (;;) {
 188                 if (!pipe->readers) {
 189                         send_sig(SIGPIPE, current, 0);
 190                         if (!ret)
 191                                 ret = -EPIPE;
 192                         break;
 193                 }
 194
 195                 if (pipe->nrbufs < PIPE_BUFFERS) {
 196                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
 197                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 198
 199                         buf->page = spd->pages[page_nr];
 200                         buf->offset = spd->partial[page_nr].offset;
 201                         buf->len = spd->partial[page_nr].len;
 202                         buf->private = spd->partial[page_nr].private;
 203                         buf->ops = spd->ops;
 204                         if (spd->flags & SPLICE_F_GIFT)
 205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 206
 207                         pipe->nrbufs++;
 208                         page_nr++;
 209                         ret += buf->len;
 210
 211                         if (pipe->inode)
 212                                 do_wakeup = 1;
 213
 214                         if (!--spd->nr_pages)
 215                                 break;
 216                         if (pipe->nrbufs < PIPE_BUFFERS)
 217                                 continue;
 218
 219                         break;
 220                 }
 221
 222                 if (spd->flags & SPLICE_F_NONBLOCK) {
 223                         if (!ret)
 224                                 ret = -EAGAIN;
 225                         break;
 226                 }
 227
 228                 if (signal_pending(current)) {
 229                         if (!ret)
 230                                 ret = -ERESTARTSYS;
 231                         break;
 232                 }
 233
 234                 if (do_wakeup) {
 235                         smp_mb();
 236                         if (waitqueue_active(&pipe->wait))
 237                                 wake_up_interruptible_sync(&pipe->wait);
 238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 239                         do_wakeup = 0;
 240                 }
 241
 242                 pipe->waiting_writers++;
 243                 pipe_wait(pipe);
 244                 pipe->waiting_writers--;
 245         }
 246
 247         pipe_unlock(pipe);
 248
 249         if (do_wakeup) {
 250                 smp_mb();
 251                 if (waitqueue_active(&pipe->wait))
 252                         wake_up_interruptible(&pipe->wait);
 253                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 254         }
 255
 256         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                                                 mapping_gfp_mask(mapping));
 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                         lock_page(page);
 369
 370                         /*
 371                          * Page was truncated, or invalidated by the
 372                          * filesystem.  Redo the find/create, but this time the
 373                          * page is kept locked, so there's no chance of another
 374                          * race with truncate/invalidate.
 375                          */
 376                         if (!page->mapping) {
 377                                 unlock_page(page);
 378                                 page = find_or_create_page(mapping, index,
 379                                                 mapping_gfp_mask(mapping));
 380
 381                                 if (!page) {
 382                                         error = -ENOMEM;
 383                                         break;
 384                                 }
 385                                 page_cache_release(pages[page_nr]);
 386                                 pages[page_nr] = page;
 387                         }
 388                         /*
 389                          * page was already under io and is now done, great
 390                          */
 391                         if (PageUptodate(page)) {
 392                                 unlock_page(page);
 393                                 goto fill_it;
 394                         }
 395
 396                         /*
 397                          * need to read in the page
 398                          */
 399                         error = mapping->a_ops->readpage(in, page);
 400                         if (unlikely(error)) {
 401                                 /*
 402                                  * We really should re-lookup the page here,
 403                                  * but it complicates things a lot. Instead
 404                                  * lets just do what we already stored, and
 405                                  * we'll get it the next time we are called.
 406                                  */
 407                                 if (error == AOP_TRUNCATED_PAGE)
 408                                         error = 0;
 409
 410                                 break;
 411                         }
 412                 }
 413 fill_it:
 414                 /*
 415                  * i_size must be checked after PageUptodate.
 416                  */
 417                 isize = i_size_read(mapping->host);
 418                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 419                 if (unlikely(!isize || index > end_index))
 420                         break;
 421
 422                 /*
 423                  * if this is the last page, see if we need to shrink
 424                  * the length and stop
 425                  */
 426                 if (end_index == index) {
 427                         unsigned int plen;
 428
 429                         /*
 430                          * max good bytes in this page
 431                          */
 432                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 433                         if (plen <= loff)
 434                                 break;
 435
 436                         /*
 437                          * force quit after adding this page
 438                          */
 439                         this_len = min(this_len, plen - loff);
 440                         len = this_len;
 441                 }
 442
 443                 partial[page_nr].offset = loff;
 444                 partial[page_nr].len = this_len;
 445                 len -= this_len;
 446                 loff = 0;
 447                 spd.nr_pages++;
 448                 index++;
 449         }
 450
 451         /*
 452          * Release any pages at the end, if we quit early. 'page_nr' is how far
 453          * we got, 'nr_pages' is how many pages are in the map.
 454          */
 455         while (page_nr < nr_pages)
 456                 page_cache_release(pages[page_nr++]);
 457         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 458
 459         if (spd.nr_pages)
 460                 return splice_to_pipe(pipe, &spd);
 461
 462         return error;
 463 }
 464
 465 /**
 466  * generic_file_splice_read - splice data from file to a pipe
 467  * @in:         file to splice from
 468  * @ppos:       position in @in
 469  * @pipe:       pipe to splice to
 470  * @len:        number of bytes to splice
 471  * @flags:      splice modifier flags
 472  *
 473  * Description:
 474  *    Will read pages from given file and fill them into a pipe. Can be
 475  *    used as long as the address_space operations for the source implements
 476  *    a readpage() hook.
 477  *
 478  */
 479 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 480                                  struct pipe_inode_info *pipe, size_t len,
 481                                  unsigned int flags)
 482 {
 483         loff_t isize, left;
 484         int ret;
 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 = __generic_file_splice_read(in, ppos, pipe, len, flags);
 495         if (ret > 0) {
 496                 *ppos += ret;
 497                 file_accessed(in);
 498         }
 499
 500         return ret;
 501 }
 502 EXPORT_SYMBOL(generic_file_splice_read);
 503
 504 static const struct pipe_buf_operations default_pipe_buf_ops = {
 505         .can_merge = 0,
 506         .map = generic_pipe_buf_map,
 507         .unmap = generic_pipe_buf_unmap,
 508         .confirm = generic_pipe_buf_confirm,
 509         .release = generic_pipe_buf_release,
 510         .steal = generic_pipe_buf_steal,
 511         .get = generic_pipe_buf_get,
 512 };
 513
 514 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 515                             unsigned long vlen, loff_t offset)
 516 {
 517         mm_segment_t old_fs;
 518         loff_t pos = offset;
 519         ssize_t res;
 520
 521         old_fs = get_fs();
 522         set_fs(get_ds());
 523         /* The cast to a user pointer is valid due to the set_fs() */
 524         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 525         set_fs(old_fs);
 526
 527         return res;
 528 }
 529
 530 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 531                             loff_t pos)
 532 {
 533         mm_segment_t old_fs;
 534         ssize_t res;
 535
 536         old_fs = get_fs();
 537         set_fs(get_ds());
 538         /* The cast to a user pointer is valid due to the set_fs() */
 539         res = vfs_write(file, (const char __user *)buf, count, &pos);
 540         set_fs(old_fs);
 541
 542         return res;
 543 }
 544
 545 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 546                                  struct pipe_inode_info *pipe, size_t len,
 547                                  unsigned int flags)
 548 {
 549         unsigned int nr_pages;
 550         unsigned int nr_freed;
 551         size_t offset;
 552         struct page *pages[PIPE_BUFFERS];
 553         struct partial_page partial[PIPE_BUFFERS];
 554         struct iovec vec[PIPE_BUFFERS];
 555         pgoff_t index;
 556         ssize_t res;
 557         size_t this_len;
 558         int error;
 559         int i;
 560         struct splice_pipe_desc spd = {
 561                 .pages = pages,
 562                 .partial = partial,
 563                 .flags = flags,
 564                 .ops = &default_pipe_buf_ops,
 565                 .spd_release = spd_release_page,
 566         };
 567
 568         index = *ppos >> PAGE_CACHE_SHIFT;
 569         offset = *ppos & ~PAGE_CACHE_MASK;
 570         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 571
 572         for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
 573                 struct page *page;
 574
 575                 page = alloc_page(GFP_USER);
 576                 error = -ENOMEM;
 577                 if (!page)
 578                         goto err;
 579
 580                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 581                 vec[i].iov_base = (void __user *) page_address(page);
 582                 vec[i].iov_len = this_len;
 583                 pages[i] = page;
 584                 spd.nr_pages++;
 585                 len -= this_len;
 586                 offset = 0;
 587         }
 588
 589         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 590         if (res < 0) {
 591                 error = res;
 592                 goto err;
 593         }
 594
 595         error = 0;
 596         if (!res)
 597                 goto err;
 598
 599         nr_freed = 0;
 600         for (i = 0; i < spd.nr_pages; i++) {
 601                 this_len = min_t(size_t, vec[i].iov_len, res);
 602                 partial[i].offset = 0;
 603                 partial[i].len = this_len;
 604                 if (!this_len) {
 605                         __free_page(pages[i]);
 606                         pages[i] = NULL;
 607                         nr_freed++;
 608                 }
 609                 res -= this_len;
 610         }
 611         spd.nr_pages -= nr_freed;
 612
 613         res = splice_to_pipe(pipe, &spd);
 614         if (res > 0)
 615                 *ppos += res;
 616
 617         return res;
 618
 619 err:
 620         for (i = 0; i < spd.nr_pages; i++)
 621                 __free_page(pages[i]);
 622
 623         return error;
 624 }
 625 EXPORT_SYMBOL(default_file_splice_read);
 626
 627 /*
 628  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 629  * using sendpage(). Return the number of bytes sent.
 630  */
 631 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 632                             struct pipe_buffer *buf, struct splice_desc *sd)
 633 {
 634         struct file *file = sd->u.file;
 635         loff_t pos = sd->pos;
 636         int ret, more;
 637
 638         ret = buf->ops->confirm(pipe, buf);
 639         if (!ret) {
 640                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 641                 if (file->f_op && file->f_op->sendpage)
 642                         ret = file->f_op->sendpage(file, buf->page, buf->offset,
 643                                                    sd->len, &pos, more);
 644                 else
 645                         ret = -EINVAL;
 646         }
 647
 648         return ret;
 649 }
 650
 651 /*
 652  * This is a little more tricky than the file -> pipe splicing. There are
 653  * basically three cases:
 654  *
 655  *      - Destination page already exists in the address space and there
 656  *        are users of it. For that case we have no other option that
 657  *        copying the data. Tough luck.
 658  *      - Destination page already exists in the address space, but there
 659  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 660  *        through to last case.
 661  *      - Destination page does not exist, we can add the pipe page to
 662  *        the page cache and avoid the copy.
 663  *
 664  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 665  * sd->flags), we attempt to migrate pages from the pipe to the output
 666  * file address space page cache. This is possible if no one else has
 667  * the pipe page referenced outside of the pipe and page cache. If
 668  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 669  * a new page in the output file page cache and fill/dirty that.
 670  */
 671 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 672                  struct splice_desc *sd)
 673 {
 674         struct file *file = sd->u.file;
 675         struct address_space *mapping = file->f_mapping;
 676         unsigned int offset, this_len;
 677         struct page *page;
 678         void *fsdata;
 679         int ret;
 680
 681         /*
 682          * make sure the data in this buffer is uptodate
 683          */
 684         ret = buf->ops->confirm(pipe, buf);
 685         if (unlikely(ret))
 686                 return ret;
 687
 688         offset = sd->pos & ~PAGE_CACHE_MASK;
 689
 690         this_len = sd->len;
 691         if (this_len + offset > PAGE_CACHE_SIZE)
 692                 this_len = PAGE_CACHE_SIZE - offset;
 693
 694         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 695                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 696         if (unlikely(ret))
 697                 goto out;
 698
 699         if (buf->page != page) {
 700                 /*
 701                  * Careful, ->map() uses KM_USER0!
 702                  */
 703                 char *src = buf->ops->map(pipe, buf, 1);
 704                 char *dst = kmap_atomic(page, KM_USER1);
 705
 706                 memcpy(dst + offset, src + buf->offset, this_len);
 707                 flush_dcache_page(page);
 708                 kunmap_atomic(dst, KM_USER1);
 709                 buf->ops->unmap(pipe, buf, src);
 710         }
 711         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 712                                 page, fsdata);
 713 out:
 714         return ret;
 715 }
 716 EXPORT_SYMBOL(pipe_to_file);
 717
 718 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 719 {
 720         smp_mb();
 721         if (waitqueue_active(&pipe->wait))
 722                 wake_up_interruptible(&pipe->wait);
 723         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 724 }
 725
 726 /**
 727  * splice_from_pipe_feed - feed available data from a pipe to a file
 728  * @pipe:       pipe to splice from
 729  * @sd:         information to @actor
 730  * @actor:      handler that splices the data
 731  *
 732  * Description:
 733  *    This function loops over the pipe and calls @actor to do the
 734  *    actual moving of a single struct pipe_buffer to the desired
 735  *    destination.  It returns when there's no more buffers left in
 736  *    the pipe or if the requested number of bytes (@sd->total_len)
 737  *    have been copied.  It returns a positive number (one) if the
 738  *    pipe needs to be filled with more data, zero if the required
 739  *    number of bytes have been copied and -errno on error.
 740  *
 741  *    This, together with splice_from_pipe_{begin,end,next}, may be
 742  *    used to implement the functionality of __splice_from_pipe() when
 743  *    locking is required around copying the pipe buffers to the
 744  *    destination.
 745  */
 746 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 747                           splice_actor *actor)
 748 {
 749         int ret;
 750
 751         while (pipe->nrbufs) {
 752                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 753                 const struct pipe_buf_operations *ops = buf->ops;
 754
 755                 sd->len = buf->len;
 756                 if (sd->len > sd->total_len)
 757                         sd->len = sd->total_len;
 758
 759                 ret = actor(pipe, buf, sd);
 760                 if (ret <= 0) {
 761                         if (ret == -ENODATA)
 762                                 ret = 0;
 763                         return ret;
 764                 }
 765                 buf->offset += ret;
 766                 buf->len -= ret;
 767
 768                 sd->num_spliced += ret;
 769                 sd->len -= ret;
 770                 sd->pos += ret;
 771                 sd->total_len -= ret;
 772
 773                 if (!buf->len) {
 774                         buf->ops = NULL;
 775                         ops->release(pipe, buf);
 776                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 777                         pipe->nrbufs--;
 778                         if (pipe->inode)
 779                                 sd->need_wakeup = true;
 780                 }
 781
 782                 if (!sd->total_len)
 783                         return 0;
 784         }
 785
 786         return 1;
 787 }
 788 EXPORT_SYMBOL(splice_from_pipe_feed);
 789
 790 /**
 791  * splice_from_pipe_next - wait for some data to splice from
 792  * @pipe:       pipe to splice from
 793  * @sd:         information about the splice operation
 794  *
 795  * Description:
 796  *    This function will wait for some data and return a positive
 797  *    value (one) if pipe buffers are available.  It will return zero
 798  *    or -errno if no more data needs to be spliced.
 799  */
 800 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 801 {
 802         while (!pipe->nrbufs) {
 803                 if (!pipe->writers)
 804                         return 0;
 805
 806                 if (!pipe->waiting_writers && sd->num_spliced)
 807                         return 0;
 808
 809                 if (sd->flags & SPLICE_F_NONBLOCK)
 810                         return -EAGAIN;
 811
 812                 if (signal_pending(current))
 813                         return -ERESTARTSYS;
 814
 815                 if (sd->need_wakeup) {
 816                         wakeup_pipe_writers(pipe);
 817                         sd->need_wakeup = false;
 818                 }
 819
 820                 pipe_wait(pipe);
 821         }
 822
 823         return 1;
 824 }
 825 EXPORT_SYMBOL(splice_from_pipe_next);
 826
 827 /**
 828  * splice_from_pipe_begin - start splicing from pipe
 829  * @sd:         information about the splice operation
 830  *
 831  * Description:
 832  *    This function should be called before a loop containing
 833  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 834  *    initialize the necessary fields of @sd.
 835  */
 836 void splice_from_pipe_begin(struct splice_desc *sd)
 837 {
 838         sd->num_spliced = 0;
 839         sd->need_wakeup = false;
 840 }
 841 EXPORT_SYMBOL(splice_from_pipe_begin);
 842
 843 /**
 844  * splice_from_pipe_end - finish splicing from pipe
 845  * @pipe:       pipe to splice from
 846  * @sd:         information about the splice operation
 847  *
 848  * Description:
 849  *    This function will wake up pipe writers if necessary.  It should
 850  *    be called after a loop containing splice_from_pipe_next() and
 851  *    splice_from_pipe_feed().
 852  */
 853 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 854 {
 855         if (sd->need_wakeup)
 856                 wakeup_pipe_writers(pipe);
 857 }
 858 EXPORT_SYMBOL(splice_from_pipe_end);
 859
 860 /**
 861  * __splice_from_pipe - splice data from a pipe to given actor
 862  * @pipe:       pipe to splice from
 863  * @sd:         information to @actor
 864  * @actor:      handler that splices the data
 865  *
 866  * Description:
 867  *    This function does little more than loop over the pipe and call
 868  *    @actor to do the actual moving of a single struct pipe_buffer to
 869  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 870  *    pipe_to_user.
 871  *
 872  */
 873 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 874                            splice_actor *actor)
 875 {
 876         int ret;
 877
 878         splice_from_pipe_begin(sd);
 879         do {
 880                 ret = splice_from_pipe_next(pipe, sd);
 881                 if (ret > 0)
 882                         ret = splice_from_pipe_feed(pipe, sd, actor);
 883         } while (ret > 0);
 884         splice_from_pipe_end(pipe, sd);
 885
 886         return sd->num_spliced ? sd->num_spliced : ret;
 887 }
 888 EXPORT_SYMBOL(__splice_from_pipe);
 889
 890 /**
 891  * splice_from_pipe - splice data from a pipe to a file
 892  * @pipe:       pipe to splice from
 893  * @out:        file to splice to
 894  * @ppos:       position in @out
 895  * @len:        how many bytes to splice
 896  * @flags:      splice modifier flags
 897  * @actor:      handler that splices the data
 898  *
 899  * Description:
 900  *    See __splice_from_pipe. This function locks the pipe inode,
 901  *    otherwise it's identical to __splice_from_pipe().
 902  *
 903  */
 904 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 905                          loff_t *ppos, size_t len, unsigned int flags,
 906                          splice_actor *actor)
 907 {
 908         ssize_t ret;
 909         struct splice_desc sd = {
 910                 .total_len = len,
 911                 .flags = flags,
 912                 .pos = *ppos,
 913                 .u.file = out,
 914         };
 915
 916         pipe_lock(pipe);
 917         ret = __splice_from_pipe(pipe, &sd, actor);
 918         pipe_unlock(pipe);
 919
 920         return ret;
 921 }
 922
 923 /**
 924  * generic_file_splice_write - splice data from a pipe to a file
 925  * @pipe:       pipe info
 926  * @out:        file to write to
 927  * @ppos:       position in @out
 928  * @len:        number of bytes to splice
 929  * @flags:      splice modifier flags
 930  *
 931  * Description:
 932  *    Will either move or copy pages (determined by @flags options) from
 933  *    the given pipe inode to the given file.
 934  *
 935  */
 936 ssize_t
 937 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 938                           loff_t *ppos, size_t len, unsigned int flags)
 939 {
 940         struct address_space *mapping = out->f_mapping;
 941         struct inode *inode = mapping->host;
 942         struct splice_desc sd = {
 943                 .total_len = len,
 944                 .flags = flags,
 945                 .pos = *ppos,
 946                 .u.file = out,
 947         };
 948         ssize_t ret;
 949
 950         pipe_lock(pipe);
 951
 952         splice_from_pipe_begin(&sd);
 953         do {
 954                 ret = splice_from_pipe_next(pipe, &sd);
 955                 if (ret <= 0)
 956                         break;
 957
 958                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
 959                 ret = file_remove_suid(out);
 960                 if (!ret) {
 961                         file_update_time(out);
 962                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
 963                 }
 964                 mutex_unlock(&inode->i_mutex);
 965         } while (ret > 0);
 966         splice_from_pipe_end(pipe, &sd);
 967
 968         pipe_unlock(pipe);
 969
 970         if (sd.num_spliced)
 971                 ret = sd.num_spliced;
 972
 973         if (ret > 0) {
 974                 unsigned long nr_pages;
 975                 int err;
 976
 977                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 978
 979                 err = generic_write_sync(out, *ppos, ret);
 980                 if (err)
 981                         ret = err;
 982                 else
 983                         *ppos += ret;
 984                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 985         }
 986
 987         return ret;
 988 }
 989
 990 EXPORT_SYMBOL(generic_file_splice_write);
 991
 992 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 993                           struct splice_desc *sd)
 994 {
 995         int ret;
 996         void *data;
 997
 998         ret = buf->ops->confirm(pipe, buf);
 999         if (ret)
1000                 return ret;
1001
1002         data = buf->ops->map(pipe, buf, 0);
1003         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1004         buf->ops->unmap(pipe, buf, data);
1005
1006         return ret;
1007 }
1008
1009 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1010                                          struct file *out, loff_t *ppos,
1011                                          size_t len, unsigned int flags)
1012 {
1013         ssize_t ret;
1014
1015         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1016         if (ret > 0)
1017                 *ppos += ret;
1018
1019         return ret;
1020 }
1021
1022 /**
1023  * generic_splice_sendpage - splice data from a pipe to a socket
1024  * @pipe:       pipe to splice from
1025  * @out:        socket to write to
1026  * @ppos:       position in @out
1027  * @len:        number of bytes to splice
1028  * @flags:      splice modifier flags
1029  *
1030  * Description:
1031  *    Will send @len bytes from the pipe to a network socket. No data copying
1032  *    is involved.
1033  *
1034  */
1035 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1036                                 loff_t *ppos, size_t len, unsigned int flags)
1037 {
1038         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1039 }
1040
1041 EXPORT_SYMBOL(generic_splice_sendpage);
1042
1043 /*
1044  * Attempt to initiate a splice from pipe to file.
1045  */
1046 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1047                            loff_t *ppos, size_t len, unsigned int flags)
1048 {
1049         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1050                                 loff_t *, size_t, unsigned int);
1051         int ret;
1052
1053         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1054                 return -EBADF;
1055
1056         if (unlikely(out->f_flags & O_APPEND))
1057                 return -EINVAL;
1058
1059         ret = rw_verify_area(WRITE, out, ppos, len);
1060         if (unlikely(ret < 0))
1061                 return ret;
1062
1063         if (out->f_op && out->f_op->splice_write)
1064                 splice_write = out->f_op->splice_write;
1065         else
1066                 splice_write = default_file_splice_write;
1067
1068         return splice_write(pipe, out, ppos, len, flags);
1069 }
1070
1071 /*
1072  * Attempt to initiate a splice from a file to a pipe.
1073  */
1074 static long do_splice_to(struct file *in, loff_t *ppos,
1075                          struct pipe_inode_info *pipe, size_t len,
1076                          unsigned int flags)
1077 {
1078         ssize_t (*splice_read)(struct file *, loff_t *,
1079                                struct pipe_inode_info *, size_t, unsigned int);
1080         int ret;
1081
1082         if (unlikely(!(in->f_mode & FMODE_READ)))
1083                 return -EBADF;
1084
1085         ret = rw_verify_area(READ, in, ppos, len);
1086         if (unlikely(ret < 0))
1087                 return ret;
1088
1089         if (in->f_op && in->f_op->splice_read)
1090                 splice_read = in->f_op->splice_read;
1091         else
1092                 splice_read = default_file_splice_read;
1093
1094         return splice_read(in, ppos, pipe, len, flags);
1095 }
1096
1097 /**
1098  * splice_direct_to_actor - splices data directly between two non-pipes
1099  * @in:         file to splice from
1100  * @sd:         actor information on where to splice to
1101  * @actor:      handles the data splicing
1102  *
1103  * Description:
1104  *    This is a special case helper to splice directly between two
1105  *    points, without requiring an explicit pipe. Internally an allocated
1106  *    pipe is cached in the process, and reused during the lifetime of
1107  *    that process.
1108  *
1109  */
1110 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1111                                splice_direct_actor *actor)
1112 {
1113         struct pipe_inode_info *pipe;
1114         long ret, bytes;
1115         umode_t i_mode;
1116         size_t len;
1117         int i, flags;
1118
1119         /*
1120          * We require the input being a regular file, as we don't want to
1121          * randomly drop data for eg socket -> socket splicing. Use the
1122          * piped splicing for that!
1123          */
1124         i_mode = in->f_path.dentry->d_inode->i_mode;
1125         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1126                 return -EINVAL;
1127
1128         /*
1129          * neither in nor out is a pipe, setup an internal pipe attached to
1130          * 'out' and transfer the wanted data from 'in' to 'out' through that
1131          */
1132         pipe = current->splice_pipe;
1133         if (unlikely(!pipe)) {
1134                 pipe = alloc_pipe_info(NULL);
1135                 if (!pipe)
1136                         return -ENOMEM;
1137
1138                 /*
1139                  * We don't have an immediate reader, but we'll read the stuff
1140                  * out of the pipe right after the splice_to_pipe(). So set
1141                  * PIPE_READERS appropriately.
1142                  */
1143                 pipe->readers = 1;
1144
1145                 current->splice_pipe = pipe;
1146         }
1147
1148         /*
1149          * Do the splice.
1150          */
1151         ret = 0;
1152         bytes = 0;
1153         len = sd->total_len;
1154         flags = sd->flags;
1155
1156         /*
1157          * Don't block on output, we have to drain the direct pipe.
1158          */
1159         sd->flags &= ~SPLICE_F_NONBLOCK;
1160
1161         while (len) {
1162                 size_t read_len;
1163                 loff_t pos = sd->pos, prev_pos = pos;
1164
1165                 ret = do_splice_to(in, &pos, pipe, len, flags);
1166                 if (unlikely(ret <= 0))
1167                         goto out_release;
1168
1169                 read_len = ret;
1170                 sd->total_len = read_len;
1171
1172                 /*
1173                  * NOTE: nonblocking mode only applies to the input. We
1174                  * must not do the output in nonblocking mode as then we
1175                  * could get stuck data in the internal pipe:
1176                  */
1177                 ret = actor(pipe, sd);
1178                 if (unlikely(ret <= 0)) {
1179                         sd->pos = prev_pos;
1180                         goto out_release;
1181                 }
1182
1183                 bytes += ret;
1184                 len -= ret;
1185                 sd->pos = pos;
1186
1187                 if (ret < read_len) {
1188                         sd->pos = prev_pos + ret;
1189                         goto out_release;
1190                 }
1191         }
1192
1193 done:
1194         pipe->nrbufs = pipe->curbuf = 0;
1195         file_accessed(in);
1196         return bytes;
1197
1198 out_release:
1199         /*
1200          * If we did an incomplete transfer we must release
1201          * the pipe buffers in question:
1202          */
1203         for (i = 0; i < PIPE_BUFFERS; i++) {
1204                 struct pipe_buffer *buf = pipe->bufs + i;
1205
1206                 if (buf->ops) {
1207                         buf->ops->release(pipe, buf);
1208                         buf->ops = NULL;
1209                 }
1210         }
1211
1212         if (!bytes)
1213                 bytes = ret;
1214
1215         goto done;
1216 }
1217 EXPORT_SYMBOL(splice_direct_to_actor);
1218
1219 static int direct_splice_actor(struct pipe_inode_info *pipe,
1220                                struct splice_desc *sd)
1221 {
1222         struct file *file = sd->u.file;
1223
1224         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1225                               sd->flags);
1226 }
1227
1228 /**
1229  * do_splice_direct - splices data directly between two files
1230  * @in:         file to splice from
1231  * @ppos:       input file offset
1232  * @out:        file to splice to
1233  * @len:        number of bytes to splice
1234  * @flags:      splice modifier flags
1235  *
1236  * Description:
1237  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1238  *    doing it in the application would incur an extra system call
1239  *    (splice in + splice out, as compared to just sendfile()). So this helper
1240  *    can splice directly through a process-private pipe.
1241  *
1242  */
1243 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1244                       size_t len, unsigned int flags)
1245 {
1246         struct splice_desc sd = {
1247                 .len            = len,
1248                 .total_len      = len,
1249                 .flags          = flags,
1250                 .pos            = *ppos,
1251                 .u.file         = out,
1252         };
1253         long ret;
1254
1255         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1256         if (ret > 0)
1257                 *ppos = sd.pos;
1258
1259         return ret;
1260 }
1261
1262 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1263                                struct pipe_inode_info *opipe,
1264                                size_t len, unsigned int flags);
1265 /*
1266  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1267  * location, so checking ->i_pipe is not enough to verify that this is a
1268  * pipe.
1269  */
1270 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1271 {
1272         if (S_ISFIFO(inode->i_mode))
1273                 return inode->i_pipe;
1274
1275         return NULL;
1276 }
1277
1278 /*
1279  * Determine where to splice to/from.
1280  */
1281 static long do_splice(struct file *in, loff_t __user *off_in,
1282                       struct file *out, loff_t __user *off_out,
1283                       size_t len, unsigned int flags)
1284 {
1285         struct pipe_inode_info *ipipe;
1286         struct pipe_inode_info *opipe;
1287         loff_t offset, *off;
1288         long ret;
1289
1290         ipipe = pipe_info(in->f_path.dentry->d_inode);
1291         opipe = pipe_info(out->f_path.dentry->d_inode);
1292
1293         if (ipipe && opipe) {
1294                 if (off_in || off_out)
1295                         return -ESPIPE;
1296
1297                 if (!(in->f_mode & FMODE_READ))
1298                         return -EBADF;
1299
1300                 if (!(out->f_mode & FMODE_WRITE))
1301                         return -EBADF;
1302
1303                 /* Splicing to self would be fun, but... */
1304                 if (ipipe == opipe)
1305                         return -EINVAL;
1306
1307                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1308         }
1309
1310         if (ipipe) {
1311                 if (off_in)
1312                         return -ESPIPE;
1313                 if (off_out) {
1314                         if (!out->f_op || !out->f_op->llseek ||
1315                             out->f_op->llseek == no_llseek)
1316                                 return -EINVAL;
1317                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1318                                 return -EFAULT;
1319                         off = &offset;
1320                 } else
1321                         off = &out->f_pos;
1322
1323                 ret = do_splice_from(ipipe, out, off, len, flags);
1324
1325                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1326                         ret = -EFAULT;
1327
1328                 return ret;
1329         }
1330
1331         if (opipe) {
1332                 if (off_out)
1333                         return -ESPIPE;
1334                 if (off_in) {
1335                         if (!in->f_op || !in->f_op->llseek ||
1336                             in->f_op->llseek == no_llseek)
1337                                 return -EINVAL;
1338                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1339                                 return -EFAULT;
1340                         off = &offset;
1341                 } else
1342                         off = &in->f_pos;
1343
1344                 ret = do_splice_to(in, off, opipe, len, flags);
1345
1346                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1347                         ret = -EFAULT;
1348
1349                 return ret;
1350         }
1351
1352         return -EINVAL;
1353 }
1354
1355 /*
1356  * Map an iov into an array of pages and offset/length tupples. With the
1357  * partial_page structure, we can map several non-contiguous ranges into
1358  * our ones pages[] map instead of splitting that operation into pieces.
1359  * Could easily be exported as a generic helper for other users, in which
1360  * case one would probably want to add a 'max_nr_pages' parameter as well.
1361  */
1362 static int get_iovec_page_array(const struct iovec __user *iov,
1363                                 unsigned int nr_vecs, struct page **pages,
1364                                 struct partial_page *partial, int aligned)
1365 {
1366         int buffers = 0, error = 0;
1367
1368         while (nr_vecs) {
1369                 unsigned long off, npages;
1370                 struct iovec entry;
1371                 void __user *base;
1372                 size_t len;
1373                 int i;
1374
1375                 error = -EFAULT;
1376                 if (copy_from_user(&entry, iov, sizeof(entry)))
1377                         break;
1378
1379                 base = entry.iov_base;
1380                 len = entry.iov_len;
1381
1382                 /*
1383                  * Sanity check this iovec. 0 read succeeds.
1384                  */
1385                 error = 0;
1386                 if (unlikely(!len))
1387                         break;
1388                 error = -EFAULT;
1389                 if (!access_ok(VERIFY_READ, base, len))
1390                         break;
1391
1392                 /*
1393                  * Get this base offset and number of pages, then map
1394                  * in the user pages.
1395                  */
1396                 off = (unsigned long) base & ~PAGE_MASK;
1397
1398                 /*
1399                  * If asked for alignment, the offset must be zero and the
1400                  * length a multiple of the PAGE_SIZE.
1401                  */
1402                 error = -EINVAL;
1403                 if (aligned && (off || len & ~PAGE_MASK))
1404                         break;
1405
1406                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1407                 if (npages > PIPE_BUFFERS - buffers)
1408                         npages = PIPE_BUFFERS - buffers;
1409
1410                 error = get_user_pages_fast((unsigned long)base, npages,
1411                                         0, &pages[buffers]);
1412
1413                 if (unlikely(error <= 0))
1414                         break;
1415
1416                 /*
1417                  * Fill this contiguous range into the partial page map.
1418                  */
1419                 for (i = 0; i < error; i++) {
1420                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1421
1422                         partial[buffers].offset = off;
1423                         partial[buffers].len = plen;
1424
1425                         off = 0;
1426                         len -= plen;
1427                         buffers++;
1428                 }
1429
1430                 /*
1431                  * We didn't complete this iov, stop here since it probably
1432                  * means we have to move some of this into a pipe to
1433                  * be able to continue.
1434                  */
1435                 if (len)
1436                         break;
1437
1438                 /*
1439                  * Don't continue if we mapped fewer pages than we asked for,
1440                  * or if we mapped the max number of pages that we have
1441                  * room for.
1442                  */
1443                 if (error < npages || buffers == PIPE_BUFFERS)
1444                         break;
1445
1446                 nr_vecs--;
1447                 iov++;
1448         }
1449
1450         if (buffers)
1451                 return buffers;
1452
1453         return error;
1454 }
1455
1456 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1457                         struct splice_desc *sd)
1458 {
1459         char *src;
1460         int ret;
1461
1462         ret = buf->ops->confirm(pipe, buf);
1463         if (unlikely(ret))
1464                 return ret;
1465
1466         /*
1467          * See if we can use the atomic maps, by prefaulting in the
1468          * pages and doing an atomic copy
1469          */
1470         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1471                 src = buf->ops->map(pipe, buf, 1);
1472                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1473                                                         sd->len);
1474                 buf->ops->unmap(pipe, buf, src);
1475                 if (!ret) {
1476                         ret = sd->len;
1477                         goto out;
1478                 }
1479         }
1480
1481         /*
1482          * No dice, use slow non-atomic map and copy
1483          */
1484         src = buf->ops->map(pipe, buf, 0);
1485
1486         ret = sd->len;
1487         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1488                 ret = -EFAULT;
1489
1490         buf->ops->unmap(pipe, buf, src);
1491 out:
1492         if (ret > 0)
1493                 sd->u.userptr += ret;
1494         return ret;
1495 }
1496
1497 /*
1498  * For lack of a better implementation, implement vmsplice() to userspace
1499  * as a simple copy of the pipes pages to the user iov.
1500  */
1501 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1502                              unsigned long nr_segs, unsigned int flags)
1503 {
1504         struct pipe_inode_info *pipe;
1505         struct splice_desc sd;
1506         ssize_t size;
1507         int error;
1508         long ret;
1509
1510         pipe = pipe_info(file->f_path.dentry->d_inode);
1511         if (!pipe)
1512                 return -EBADF;
1513
1514         pipe_lock(pipe);
1515
1516         error = ret = 0;
1517         while (nr_segs) {
1518                 void __user *base;
1519                 size_t len;
1520
1521                 /*
1522                  * Get user address base and length for this iovec.
1523                  */
1524                 error = get_user(base, &iov->iov_base);
1525                 if (unlikely(error))
1526                         break;
1527                 error = get_user(len, &iov->iov_len);
1528                 if (unlikely(error))
1529                         break;
1530
1531                 /*
1532                  * Sanity check this iovec. 0 read succeeds.
1533                  */
1534                 if (unlikely(!len))
1535                         break;
1536                 if (unlikely(!base)) {
1537                         error = -EFAULT;
1538                         break;
1539                 }
1540
1541                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1542                         error = -EFAULT;
1543                         break;
1544                 }
1545
1546                 sd.len = 0;
1547                 sd.total_len = len;
1548                 sd.flags = flags;
1549                 sd.u.userptr = base;
1550                 sd.pos = 0;
1551
1552                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1553                 if (size < 0) {
1554                         if (!ret)
1555                                 ret = size;
1556
1557                         break;
1558                 }
1559
1560                 ret += size;
1561
1562                 if (size < len)
1563                         break;
1564
1565                 nr_segs--;
1566                 iov++;
1567         }
1568
1569         pipe_unlock(pipe);
1570
1571         if (!ret)
1572                 ret = error;
1573
1574         return ret;
1575 }
1576
1577 /*
1578  * vmsplice splices a user address range into a pipe. It can be thought of
1579  * as splice-from-memory, where the regular splice is splice-from-file (or
1580  * to file). In both cases the output is a pipe, naturally.
1581  */
1582 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1583                              unsigned long nr_segs, unsigned int flags)
1584 {
1585         struct pipe_inode_info *pipe;
1586         struct page *pages[PIPE_BUFFERS];
1587         struct partial_page partial[PIPE_BUFFERS];
1588         struct splice_pipe_desc spd = {
1589                 .pages = pages,
1590                 .partial = partial,
1591                 .flags = flags,
1592                 .ops = &user_page_pipe_buf_ops,
1593                 .spd_release = spd_release_page,
1594         };
1595
1596         pipe = pipe_info(file->f_path.dentry->d_inode);
1597         if (!pipe)
1598                 return -EBADF;
1599
1600         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1601                                             flags & SPLICE_F_GIFT);
1602         if (spd.nr_pages <= 0)
1603                 return spd.nr_pages;
1604
1605         return splice_to_pipe(pipe, &spd);
1606 }
1607
1608 /*
1609  * Note that vmsplice only really supports true splicing _from_ user memory
1610  * to a pipe, not the other way around. Splicing from user memory is a simple
1611  * operation that can be supported without any funky alignment restrictions
1612  * or nasty vm tricks. We simply map in the user memory and fill them into
1613  * a pipe. The reverse isn't quite as easy, though. There are two possible
1614  * solutions for that:
1615  *
1616  *      - memcpy() the data internally, at which point we might as well just
1617  *        do a regular read() on the buffer anyway.
1618  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1619  *        has restriction limitations on both ends of the pipe).
1620  *
1621  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1622  *
1623  */
1624 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1625                 unsigned long, nr_segs, unsigned int, flags)
1626 {
1627         struct file *file;
1628         long error;
1629         int fput;
1630
1631         if (unlikely(nr_segs > UIO_MAXIOV))
1632                 return -EINVAL;
1633         else if (unlikely(!nr_segs))
1634                 return 0;
1635
1636         error = -EBADF;
1637         file = fget_light(fd, &fput);
1638         if (file) {
1639                 if (file->f_mode & FMODE_WRITE)
1640                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1641                 else if (file->f_mode & FMODE_READ)
1642                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1643
1644                 fput_light(file, fput);
1645         }
1646
1647         return error;
1648 }
1649
1650 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1651                 int, fd_out, loff_t __user *, off_out,
1652                 size_t, len, unsigned int, flags)
1653 {
1654         long error;
1655         struct file *in, *out;
1656         int fput_in, fput_out;
1657
1658         if (unlikely(!len))
1659                 return 0;
1660
1661         error = -EBADF;
1662         in = fget_light(fd_in, &fput_in);
1663         if (in) {
1664                 if (in->f_mode & FMODE_READ) {
1665                         out = fget_light(fd_out, &fput_out);
1666                         if (out) {
1667                                 if (out->f_mode & FMODE_WRITE)
1668                                         error = do_splice(in, off_in,
1669                                                           out, off_out,
1670                                                           len, flags);
1671                                 fput_light(out, fput_out);
1672                         }
1673                 }
1674
1675                 fput_light(in, fput_in);
1676         }
1677
1678         return error;
1679 }
1680
1681 /*
1682  * Make sure there's data to read. Wait for input if we can, otherwise
1683  * return an appropriate error.
1684  */
1685 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1686 {
1687         int ret;
1688
1689         /*
1690          * Check ->nrbufs without the inode lock first. This function
1691          * is speculative anyways, so missing one is ok.
1692          */
1693         if (pipe->nrbufs)
1694                 return 0;
1695
1696         ret = 0;
1697         pipe_lock(pipe);
1698
1699         while (!pipe->nrbufs) {
1700                 if (signal_pending(current)) {
1701                         ret = -ERESTARTSYS;
1702                         break;
1703                 }
1704                 if (!pipe->writers)
1705                         break;
1706                 if (!pipe->waiting_writers) {
1707                         if (flags & SPLICE_F_NONBLOCK) {
1708                                 ret = -EAGAIN;
1709                                 break;
1710                         }
1711                 }
1712                 pipe_wait(pipe);
1713         }
1714
1715         pipe_unlock(pipe);
1716         return ret;
1717 }
1718
1719 /*
1720  * Make sure there's writeable room. Wait for room if we can, otherwise
1721  * return an appropriate error.
1722  */
1723 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1724 {
1725         int ret;
1726
1727         /*
1728          * Check ->nrbufs without the inode lock first. This function
1729          * is speculative anyways, so missing one is ok.
1730          */
1731         if (pipe->nrbufs < PIPE_BUFFERS)
1732                 return 0;
1733
1734         ret = 0;
1735         pipe_lock(pipe);
1736
1737         while (pipe->nrbufs >= PIPE_BUFFERS) {
1738                 if (!pipe->readers) {
1739                         send_sig(SIGPIPE, current, 0);
1740                         ret = -EPIPE;
1741                         break;
1742                 }
1743                 if (flags & SPLICE_F_NONBLOCK) {
1744                         ret = -EAGAIN;
1745                         break;
1746                 }
1747                 if (signal_pending(current)) {
1748                         ret = -ERESTARTSYS;
1749                         break;
1750                 }
1751                 pipe->waiting_writers++;
1752                 pipe_wait(pipe);
1753                 pipe->waiting_writers--;
1754         }
1755
1756         pipe_unlock(pipe);
1757         return ret;
1758 }
1759
1760 /*
1761  * Splice contents of ipipe to opipe.
1762  */
1763 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1764                                struct pipe_inode_info *opipe,
1765                                size_t len, unsigned int flags)
1766 {
1767         struct pipe_buffer *ibuf, *obuf;
1768         int ret = 0, nbuf;
1769         bool input_wakeup = false;
1770
1771
1772 retry:
1773         ret = ipipe_prep(ipipe, flags);
1774         if (ret)
1775                 return ret;
1776
1777         ret = opipe_prep(opipe, flags);
1778         if (ret)
1779                 return ret;
1780
1781         /*
1782          * Potential ABBA deadlock, work around it by ordering lock
1783          * grabbing by pipe info address. Otherwise two different processes
1784          * could deadlock (one doing tee from A -> B, the other from B -> A).
1785          */
1786         pipe_double_lock(ipipe, opipe);
1787
1788         do {
1789                 if (!opipe->readers) {
1790                         send_sig(SIGPIPE, current, 0);
1791                         if (!ret)
1792                                 ret = -EPIPE;
1793                         break;
1794                 }
1795
1796                 if (!ipipe->nrbufs && !ipipe->writers)
1797                         break;
1798
1799                 /*
1800                  * Cannot make any progress, because either the input
1801                  * pipe is empty or the output pipe is full.
1802                  */
1803                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1804                         /* Already processed some buffers, break */
1805                         if (ret)
1806                                 break;
1807
1808                         if (flags & SPLICE_F_NONBLOCK) {
1809                                 ret = -EAGAIN;
1810                                 break;
1811                         }
1812
1813                         /*
1814                          * We raced with another reader/writer and haven't
1815                          * managed to process any buffers.  A zero return
1816                          * value means EOF, so retry instead.
1817                          */
1818                         pipe_unlock(ipipe);
1819                         pipe_unlock(opipe);
1820                         goto retry;
1821                 }
1822
1823                 ibuf = ipipe->bufs + ipipe->curbuf;
1824                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1825                 obuf = opipe->bufs + nbuf;
1826
1827                 if (len >= ibuf->len) {
1828                         /*
1829                          * Simply move the whole buffer from ipipe to opipe
1830                          */
1831                         *obuf = *ibuf;
1832                         ibuf->ops = NULL;
1833                         opipe->nrbufs++;
1834                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1835                         ipipe->nrbufs--;
1836                         input_wakeup = true;
1837                 } else {
1838                         /*
1839                          * Get a reference to this pipe buffer,
1840                          * so we can copy the contents over.
1841                          */
1842                         ibuf->ops->get(ipipe, ibuf);
1843                         *obuf = *ibuf;
1844
1845                         /*
1846                          * Don't inherit the gift flag, we need to
1847                          * prevent multiple steals of this page.
1848                          */
1849                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1850
1851                         obuf->len = len;
1852                         opipe->nrbufs++;
1853                         ibuf->offset += obuf->len;
1854                         ibuf->len -= obuf->len;
1855                 }
1856                 ret += obuf->len;
1857                 len -= obuf->len;
1858         } while (len);
1859
1860         pipe_unlock(ipipe);
1861         pipe_unlock(opipe);
1862
1863         /*
1864          * If we put data in the output pipe, wakeup any potential readers.
1865          */
1866         if (ret > 0) {
1867                 smp_mb();
1868                 if (waitqueue_active(&opipe->wait))
1869                         wake_up_interruptible(&opipe->wait);
1870                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1871         }
1872         if (input_wakeup)
1873                 wakeup_pipe_writers(ipipe);
1874
1875         return ret;
1876 }
1877
1878 /*
1879  * Link contents of ipipe to opipe.
1880  */
1881 static int link_pipe(struct pipe_inode_info *ipipe,
1882                      struct pipe_inode_info *opipe,
1883                      size_t len, unsigned int flags)
1884 {
1885         struct pipe_buffer *ibuf, *obuf;
1886         int ret = 0, i = 0, nbuf;
1887
1888         /*
1889          * Potential ABBA deadlock, work around it by ordering lock
1890          * grabbing by pipe info address. Otherwise two different processes
1891          * could deadlock (one doing tee from A -> B, the other from B -> A).
1892          */
1893         pipe_double_lock(ipipe, opipe);
1894
1895         do {
1896                 if (!opipe->readers) {
1897                         send_sig(SIGPIPE, current, 0);
1898                         if (!ret)
1899                                 ret = -EPIPE;
1900                         break;
1901                 }
1902
1903                 /*
1904                  * If we have iterated all input buffers or ran out of
1905                  * output room, break.
1906                  */
1907                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1908                         break;
1909
1910                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1911                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1912
1913                 /*
1914                  * Get a reference to this pipe buffer,
1915                  * so we can copy the contents over.
1916                  */
1917                 ibuf->ops->get(ipipe, ibuf);
1918
1919                 obuf = opipe->bufs + nbuf;
1920                 *obuf = *ibuf;
1921
1922                 /*
1923                  * Don't inherit the gift flag, we need to
1924                  * prevent multiple steals of this page.
1925                  */
1926                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1927
1928                 if (obuf->len > len)
1929                         obuf->len = len;
1930
1931                 opipe->nrbufs++;
1932                 ret += obuf->len;
1933                 len -= obuf->len;
1934                 i++;
1935         } while (len);
1936
1937         /*
1938          * return EAGAIN if we have the potential of some data in the
1939          * future, otherwise just return 0
1940          */
1941         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1942                 ret = -EAGAIN;
1943
1944         pipe_unlock(ipipe);
1945         pipe_unlock(opipe);
1946
1947         /*
1948          * If we put data in the output pipe, wakeup any potential readers.
1949          */
1950         if (ret > 0) {
1951                 smp_mb();
1952                 if (waitqueue_active(&opipe->wait))
1953                         wake_up_interruptible(&opipe->wait);
1954                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1955         }
1956
1957         return ret;
1958 }
1959
1960 /*
1961  * This is a tee(1) implementation that works on pipes. It doesn't copy
1962  * any data, it simply references the 'in' pages on the 'out' pipe.
1963  * The 'flags' used are the SPLICE_F_* variants, currently the only
1964  * applicable one is SPLICE_F_NONBLOCK.
1965  */
1966 static long do_tee(struct file *in, struct file *out, size_t len,
1967                    unsigned int flags)
1968 {
1969         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1970         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1971         int ret = -EINVAL;
1972
1973         /*
1974          * Duplicate the contents of ipipe to opipe without actually
1975          * copying the data.
1976          */
1977         if (ipipe && opipe && ipipe != opipe) {
1978                 /*
1979                  * Keep going, unless we encounter an error. The ipipe/opipe
1980                  * ordering doesn't really matter.
1981                  */
1982                 ret = ipipe_prep(ipipe, flags);
1983                 if (!ret) {
1984                         ret = opipe_prep(opipe, flags);
1985                         if (!ret)
1986                                 ret = link_pipe(ipipe, opipe, len, flags);
1987                 }
1988         }
1989
1990         return ret;
1991 }
1992
1993 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1994 {
1995         struct file *in;
1996         int error, fput_in;
1997
1998         if (unlikely(!len))
1999                 return 0;
2000
2001         error = -EBADF;
2002         in = fget_light(fdin, &fput_in);
2003         if (in) {
2004                 if (in->f_mode & FMODE_READ) {
2005                         int fput_out;
2006                         struct file *out = fget_light(fdout, &fput_out);
2007
2008                         if (out) {
2009                                 if (out->f_mode & FMODE_WRITE)
2010                                         error = do_tee(in, out, len, flags);
2011                                 fput_light(out, fput_out);
2012                         }
2013                 }
2014                 fput_light(in, fput_in);
2015         }
2016
2017         return error;
2018 }