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
 642                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 643                                            sd->len, &pos, more);
 644         }
 645
 646         return ret;
 647 }
 648
 649 /*
 650  * This is a little more tricky than the file -> pipe splicing. There are
 651  * basically three cases:
 652  *
 653  *      - Destination page already exists in the address space and there
 654  *        are users of it. For that case we have no other option that
 655  *        copying the data. Tough luck.
 656  *      - Destination page already exists in the address space, but there
 657  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 658  *        through to last case.
 659  *      - Destination page does not exist, we can add the pipe page to
 660  *        the page cache and avoid the copy.
 661  *
 662  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 663  * sd->flags), we attempt to migrate pages from the pipe to the output
 664  * file address space page cache. This is possible if no one else has
 665  * the pipe page referenced outside of the pipe and page cache. If
 666  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 667  * a new page in the output file page cache and fill/dirty that.
 668  */
 669 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 670                  struct splice_desc *sd)
 671 {
 672         struct file *file = sd->u.file;
 673         struct address_space *mapping = file->f_mapping;
 674         unsigned int offset, this_len;
 675         struct page *page;
 676         void *fsdata;
 677         int ret;
 678
 679         /*
 680          * make sure the data in this buffer is uptodate
 681          */
 682         ret = buf->ops->confirm(pipe, buf);
 683         if (unlikely(ret))
 684                 return ret;
 685
 686         offset = sd->pos & ~PAGE_CACHE_MASK;
 687
 688         this_len = sd->len;
 689         if (this_len + offset > PAGE_CACHE_SIZE)
 690                 this_len = PAGE_CACHE_SIZE - offset;
 691
 692         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 693                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 694         if (unlikely(ret))
 695                 goto out;
 696
 697         if (buf->page != page) {
 698                 /*
 699                  * Careful, ->map() uses KM_USER0!
 700                  */
 701                 char *src = buf->ops->map(pipe, buf, 1);
 702                 char *dst = kmap_atomic(page, KM_USER1);
 703
 704                 memcpy(dst + offset, src + buf->offset, this_len);
 705                 flush_dcache_page(page);
 706                 kunmap_atomic(dst, KM_USER1);
 707                 buf->ops->unmap(pipe, buf, src);
 708         }
 709         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 710                                 page, fsdata);
 711 out:
 712         return ret;
 713 }
 714 EXPORT_SYMBOL(pipe_to_file);
 715
 716 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 717 {
 718         smp_mb();
 719         if (waitqueue_active(&pipe->wait))
 720                 wake_up_interruptible(&pipe->wait);
 721         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 722 }
 723
 724 /**
 725  * splice_from_pipe_feed - feed available data from a pipe to a file
 726  * @pipe:       pipe to splice from
 727  * @sd:         information to @actor
 728  * @actor:      handler that splices the data
 729  *
 730  * Description:
 731  *    This function loops over the pipe and calls @actor to do the
 732  *    actual moving of a single struct pipe_buffer to the desired
 733  *    destination.  It returns when there's no more buffers left in
 734  *    the pipe or if the requested number of bytes (@sd->total_len)
 735  *    have been copied.  It returns a positive number (one) if the
 736  *    pipe needs to be filled with more data, zero if the required
 737  *    number of bytes have been copied and -errno on error.
 738  *
 739  *    This, together with splice_from_pipe_{begin,end,next}, may be
 740  *    used to implement the functionality of __splice_from_pipe() when
 741  *    locking is required around copying the pipe buffers to the
 742  *    destination.
 743  */
 744 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 745                           splice_actor *actor)
 746 {
 747         int ret;
 748
 749         while (pipe->nrbufs) {
 750                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 751                 const struct pipe_buf_operations *ops = buf->ops;
 752
 753                 sd->len = buf->len;
 754                 if (sd->len > sd->total_len)
 755                         sd->len = sd->total_len;
 756
 757                 ret = actor(pipe, buf, sd);
 758                 if (ret <= 0) {
 759                         if (ret == -ENODATA)
 760                                 ret = 0;
 761                         return ret;
 762                 }
 763                 buf->offset += ret;
 764                 buf->len -= ret;
 765
 766                 sd->num_spliced += ret;
 767                 sd->len -= ret;
 768                 sd->pos += ret;
 769                 sd->total_len -= ret;
 770
 771                 if (!buf->len) {
 772                         buf->ops = NULL;
 773                         ops->release(pipe, buf);
 774                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 775                         pipe->nrbufs--;
 776                         if (pipe->inode)
 777                                 sd->need_wakeup = true;
 778                 }
 779
 780                 if (!sd->total_len)
 781                         return 0;
 782         }
 783
 784         return 1;
 785 }
 786 EXPORT_SYMBOL(splice_from_pipe_feed);
 787
 788 /**
 789  * splice_from_pipe_next - wait for some data to splice from
 790  * @pipe:       pipe to splice from
 791  * @sd:         information about the splice operation
 792  *
 793  * Description:
 794  *    This function will wait for some data and return a positive
 795  *    value (one) if pipe buffers are available.  It will return zero
 796  *    or -errno if no more data needs to be spliced.
 797  */
 798 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 799 {
 800         while (!pipe->nrbufs) {
 801                 if (!pipe->writers)
 802                         return 0;
 803
 804                 if (!pipe->waiting_writers && sd->num_spliced)
 805                         return 0;
 806
 807                 if (sd->flags & SPLICE_F_NONBLOCK)
 808                         return -EAGAIN;
 809
 810                 if (signal_pending(current))
 811                         return -ERESTARTSYS;
 812
 813                 if (sd->need_wakeup) {
 814                         wakeup_pipe_writers(pipe);
 815                         sd->need_wakeup = false;
 816                 }
 817
 818                 pipe_wait(pipe);
 819         }
 820
 821         return 1;
 822 }
 823 EXPORT_SYMBOL(splice_from_pipe_next);
 824
 825 /**
 826  * splice_from_pipe_begin - start splicing from pipe
 827  * @sd:         information about the splice operation
 828  *
 829  * Description:
 830  *    This function should be called before a loop containing
 831  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 832  *    initialize the necessary fields of @sd.
 833  */
 834 void splice_from_pipe_begin(struct splice_desc *sd)
 835 {
 836         sd->num_spliced = 0;
 837         sd->need_wakeup = false;
 838 }
 839 EXPORT_SYMBOL(splice_from_pipe_begin);
 840
 841 /**
 842  * splice_from_pipe_end - finish splicing from pipe
 843  * @pipe:       pipe to splice from
 844  * @sd:         information about the splice operation
 845  *
 846  * Description:
 847  *    This function will wake up pipe writers if necessary.  It should
 848  *    be called after a loop containing splice_from_pipe_next() and
 849  *    splice_from_pipe_feed().
 850  */
 851 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 852 {
 853         if (sd->need_wakeup)
 854                 wakeup_pipe_writers(pipe);
 855 }
 856 EXPORT_SYMBOL(splice_from_pipe_end);
 857
 858 /**
 859  * __splice_from_pipe - splice data from a pipe to given actor
 860  * @pipe:       pipe to splice from
 861  * @sd:         information to @actor
 862  * @actor:      handler that splices the data
 863  *
 864  * Description:
 865  *    This function does little more than loop over the pipe and call
 866  *    @actor to do the actual moving of a single struct pipe_buffer to
 867  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 868  *    pipe_to_user.
 869  *
 870  */
 871 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 872                            splice_actor *actor)
 873 {
 874         int ret;
 875
 876         splice_from_pipe_begin(sd);
 877         do {
 878                 ret = splice_from_pipe_next(pipe, sd);
 879                 if (ret > 0)
 880                         ret = splice_from_pipe_feed(pipe, sd, actor);
 881         } while (ret > 0);
 882         splice_from_pipe_end(pipe, sd);
 883
 884         return sd->num_spliced ? sd->num_spliced : ret;
 885 }
 886 EXPORT_SYMBOL(__splice_from_pipe);
 887
 888 /**
 889  * splice_from_pipe - splice data from a pipe to a file
 890  * @pipe:       pipe to splice from
 891  * @out:        file to splice to
 892  * @ppos:       position in @out
 893  * @len:        how many bytes to splice
 894  * @flags:      splice modifier flags
 895  * @actor:      handler that splices the data
 896  *
 897  * Description:
 898  *    See __splice_from_pipe. This function locks the pipe inode,
 899  *    otherwise it's identical to __splice_from_pipe().
 900  *
 901  */
 902 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 903                          loff_t *ppos, size_t len, unsigned int flags,
 904                          splice_actor *actor)
 905 {
 906         ssize_t ret;
 907         struct splice_desc sd = {
 908                 .total_len = len,
 909                 .flags = flags,
 910                 .pos = *ppos,
 911                 .u.file = out,
 912         };
 913
 914         pipe_lock(pipe);
 915         ret = __splice_from_pipe(pipe, &sd, actor);
 916         pipe_unlock(pipe);
 917
 918         return ret;
 919 }
 920
 921 /**
 922  * generic_file_splice_write - splice data from a pipe to a file
 923  * @pipe:       pipe info
 924  * @out:        file to write to
 925  * @ppos:       position in @out
 926  * @len:        number of bytes to splice
 927  * @flags:      splice modifier flags
 928  *
 929  * Description:
 930  *    Will either move or copy pages (determined by @flags options) from
 931  *    the given pipe inode to the given file.
 932  *
 933  */
 934 ssize_t
 935 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 936                           loff_t *ppos, size_t len, unsigned int flags)
 937 {
 938         struct address_space *mapping = out->f_mapping;
 939         struct inode *inode = mapping->host;
 940         struct splice_desc sd = {
 941                 .total_len = len,
 942                 .flags = flags,
 943                 .pos = *ppos,
 944                 .u.file = out,
 945         };
 946         ssize_t ret;
 947
 948         pipe_lock(pipe);
 949
 950         splice_from_pipe_begin(&sd);
 951         do {
 952                 ret = splice_from_pipe_next(pipe, &sd);
 953                 if (ret <= 0)
 954                         break;
 955
 956                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
 957                 ret = file_remove_suid(out);
 958                 if (!ret) {
 959                         file_update_time(out);
 960                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
 961                 }
 962                 mutex_unlock(&inode->i_mutex);
 963         } while (ret > 0);
 964         splice_from_pipe_end(pipe, &sd);
 965
 966         pipe_unlock(pipe);
 967
 968         if (sd.num_spliced)
 969                 ret = sd.num_spliced;
 970
 971         if (ret > 0) {
 972                 unsigned long nr_pages;
 973                 int err;
 974
 975                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 976
 977                 err = generic_write_sync(out, *ppos, ret);
 978                 if (err)
 979                         ret = err;
 980                 else
 981                         *ppos += ret;
 982                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 983         }
 984
 985         return ret;
 986 }
 987
 988 EXPORT_SYMBOL(generic_file_splice_write);
 989
 990 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 991                           struct splice_desc *sd)
 992 {
 993         int ret;
 994         void *data;
 995
 996         ret = buf->ops->confirm(pipe, buf);
 997         if (ret)
 998                 return ret;
 999
1000         data = buf->ops->map(pipe, buf, 0);
1001         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1002         buf->ops->unmap(pipe, buf, data);
1003
1004         return ret;
1005 }
1006
1007 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1008                                          struct file *out, loff_t *ppos,
1009                                          size_t len, unsigned int flags)
1010 {
1011         ssize_t ret;
1012
1013         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1014         if (ret > 0)
1015                 *ppos += ret;
1016
1017         return ret;
1018 }
1019
1020 /**
1021  * generic_splice_sendpage - splice data from a pipe to a socket
1022  * @pipe:       pipe to splice from
1023  * @out:        socket to write to
1024  * @ppos:       position in @out
1025  * @len:        number of bytes to splice
1026  * @flags:      splice modifier flags
1027  *
1028  * Description:
1029  *    Will send @len bytes from the pipe to a network socket. No data copying
1030  *    is involved.
1031  *
1032  */
1033 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1034                                 loff_t *ppos, size_t len, unsigned int flags)
1035 {
1036         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1037 }
1038
1039 EXPORT_SYMBOL(generic_splice_sendpage);
1040
1041 /*
1042  * Attempt to initiate a splice from pipe to file.
1043  */
1044 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1045                            loff_t *ppos, size_t len, unsigned int flags)
1046 {
1047         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1048                                 loff_t *, size_t, unsigned int);
1049         int ret;
1050
1051         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1052                 return -EBADF;
1053
1054         if (unlikely(out->f_flags & O_APPEND))
1055                 return -EINVAL;
1056
1057         ret = rw_verify_area(WRITE, out, ppos, len);
1058         if (unlikely(ret < 0))
1059                 return ret;
1060
1061         splice_write = out->f_op->splice_write;
1062         if (!splice_write)
1063                 splice_write = default_file_splice_write;
1064
1065         return splice_write(pipe, out, ppos, len, flags);
1066 }
1067
1068 /*
1069  * Attempt to initiate a splice from a file to a pipe.
1070  */
1071 static long do_splice_to(struct file *in, loff_t *ppos,
1072                          struct pipe_inode_info *pipe, size_t len,
1073                          unsigned int flags)
1074 {
1075         ssize_t (*splice_read)(struct file *, loff_t *,
1076                                struct pipe_inode_info *, size_t, unsigned int);
1077         int ret;
1078
1079         if (unlikely(!(in->f_mode & FMODE_READ)))
1080                 return -EBADF;
1081
1082         ret = rw_verify_area(READ, in, ppos, len);
1083         if (unlikely(ret < 0))
1084                 return ret;
1085
1086         splice_read = in->f_op->splice_read;
1087         if (!splice_read)
1088                 splice_read = default_file_splice_read;
1089
1090         return splice_read(in, ppos, pipe, len, flags);
1091 }
1092
1093 /**
1094  * splice_direct_to_actor - splices data directly between two non-pipes
1095  * @in:         file to splice from
1096  * @sd:         actor information on where to splice to
1097  * @actor:      handles the data splicing
1098  *
1099  * Description:
1100  *    This is a special case helper to splice directly between two
1101  *    points, without requiring an explicit pipe. Internally an allocated
1102  *    pipe is cached in the process, and reused during the lifetime of
1103  *    that process.
1104  *
1105  */
1106 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1107                                splice_direct_actor *actor)
1108 {
1109         struct pipe_inode_info *pipe;
1110         long ret, bytes;
1111         umode_t i_mode;
1112         size_t len;
1113         int i, flags;
1114
1115         /*
1116          * We require the input being a regular file, as we don't want to
1117          * randomly drop data for eg socket -> socket splicing. Use the
1118          * piped splicing for that!
1119          */
1120         i_mode = in->f_path.dentry->d_inode->i_mode;
1121         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1122                 return -EINVAL;
1123
1124         /*
1125          * neither in nor out is a pipe, setup an internal pipe attached to
1126          * 'out' and transfer the wanted data from 'in' to 'out' through that
1127          */
1128         pipe = current->splice_pipe;
1129         if (unlikely(!pipe)) {
1130                 pipe = alloc_pipe_info(NULL);
1131                 if (!pipe)
1132                         return -ENOMEM;
1133
1134                 /*
1135                  * We don't have an immediate reader, but we'll read the stuff
1136                  * out of the pipe right after the splice_to_pipe(). So set
1137                  * PIPE_READERS appropriately.
1138                  */
1139                 pipe->readers = 1;
1140
1141                 current->splice_pipe = pipe;
1142         }
1143
1144         /*
1145          * Do the splice.
1146          */
1147         ret = 0;
1148         bytes = 0;
1149         len = sd->total_len;
1150         flags = sd->flags;
1151
1152         /*
1153          * Don't block on output, we have to drain the direct pipe.
1154          */
1155         sd->flags &= ~SPLICE_F_NONBLOCK;
1156
1157         while (len) {
1158                 size_t read_len;
1159                 loff_t pos = sd->pos, prev_pos = pos;
1160
1161                 ret = do_splice_to(in, &pos, pipe, len, flags);
1162                 if (unlikely(ret <= 0))
1163                         goto out_release;
1164
1165                 read_len = ret;
1166                 sd->total_len = read_len;
1167
1168                 /*
1169                  * NOTE: nonblocking mode only applies to the input. We
1170                  * must not do the output in nonblocking mode as then we
1171                  * could get stuck data in the internal pipe:
1172                  */
1173                 ret = actor(pipe, sd);
1174                 if (unlikely(ret <= 0)) {
1175                         sd->pos = prev_pos;
1176                         goto out_release;
1177                 }
1178
1179                 bytes += ret;
1180                 len -= ret;
1181                 sd->pos = pos;
1182
1183                 if (ret < read_len) {
1184                         sd->pos = prev_pos + ret;
1185                         goto out_release;
1186                 }
1187         }
1188
1189 done:
1190         pipe->nrbufs = pipe->curbuf = 0;
1191         file_accessed(in);
1192         return bytes;
1193
1194 out_release:
1195         /*
1196          * If we did an incomplete transfer we must release
1197          * the pipe buffers in question:
1198          */
1199         for (i = 0; i < PIPE_BUFFERS; i++) {
1200                 struct pipe_buffer *buf = pipe->bufs + i;
1201
1202                 if (buf->ops) {
1203                         buf->ops->release(pipe, buf);
1204                         buf->ops = NULL;
1205                 }
1206         }
1207
1208         if (!bytes)
1209                 bytes = ret;
1210
1211         goto done;
1212 }
1213 EXPORT_SYMBOL(splice_direct_to_actor);
1214
1215 static int direct_splice_actor(struct pipe_inode_info *pipe,
1216                                struct splice_desc *sd)
1217 {
1218         struct file *file = sd->u.file;
1219
1220         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1221 }
1222
1223 /**
1224  * do_splice_direct - splices data directly between two files
1225  * @in:         file to splice from
1226  * @ppos:       input file offset
1227  * @out:        file to splice to
1228  * @len:        number of bytes to splice
1229  * @flags:      splice modifier flags
1230  *
1231  * Description:
1232  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1233  *    doing it in the application would incur an extra system call
1234  *    (splice in + splice out, as compared to just sendfile()). So this helper
1235  *    can splice directly through a process-private pipe.
1236  *
1237  */
1238 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1239                       size_t len, unsigned int flags)
1240 {
1241         struct splice_desc sd = {
1242                 .len            = len,
1243                 .total_len      = len,
1244                 .flags          = flags,
1245                 .pos            = *ppos,
1246                 .u.file         = out,
1247         };
1248         long ret;
1249
1250         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1251         if (ret > 0)
1252                 *ppos = sd.pos;
1253
1254         return ret;
1255 }
1256
1257 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1258                                struct pipe_inode_info *opipe,
1259                                size_t len, unsigned int flags);
1260 /*
1261  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1262  * location, so checking ->i_pipe is not enough to verify that this is a
1263  * pipe.
1264  */
1265 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1266 {
1267         if (S_ISFIFO(inode->i_mode))
1268                 return inode->i_pipe;
1269
1270         return NULL;
1271 }
1272
1273 /*
1274  * Determine where to splice to/from.
1275  */
1276 static long do_splice(struct file *in, loff_t __user *off_in,
1277                       struct file *out, loff_t __user *off_out,
1278                       size_t len, unsigned int flags)
1279 {
1280         struct pipe_inode_info *ipipe;
1281         struct pipe_inode_info *opipe;
1282         loff_t offset, *off;
1283         long ret;
1284
1285         ipipe = pipe_info(in->f_path.dentry->d_inode);
1286         opipe = pipe_info(out->f_path.dentry->d_inode);
1287
1288         if (ipipe && opipe) {
1289                 if (off_in || off_out)
1290                         return -ESPIPE;
1291
1292                 if (!(in->f_mode & FMODE_READ))
1293                         return -EBADF;
1294
1295                 if (!(out->f_mode & FMODE_WRITE))
1296                         return -EBADF;
1297
1298                 /* Splicing to self would be fun, but... */
1299                 if (ipipe == opipe)
1300                         return -EINVAL;
1301
1302                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1303         }
1304
1305         if (ipipe) {
1306                 if (off_in)
1307                         return -ESPIPE;
1308                 if (off_out) {
1309                         if (out->f_op->llseek == no_llseek)
1310                                 return -EINVAL;
1311                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1312                                 return -EFAULT;
1313                         off = &offset;
1314                 } else
1315                         off = &out->f_pos;
1316
1317                 ret = do_splice_from(ipipe, out, off, len, flags);
1318
1319                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1320                         ret = -EFAULT;
1321
1322                 return ret;
1323         }
1324
1325         if (opipe) {
1326                 if (off_out)
1327                         return -ESPIPE;
1328                 if (off_in) {
1329                         if (in->f_op->llseek == no_llseek)
1330                                 return -EINVAL;
1331                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1332                                 return -EFAULT;
1333                         off = &offset;
1334                 } else
1335                         off = &in->f_pos;
1336
1337                 ret = do_splice_to(in, off, opipe, len, flags);
1338
1339                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1340                         ret = -EFAULT;
1341
1342                 return ret;
1343         }
1344
1345         return -EINVAL;
1346 }
1347
1348 /*
1349  * Map an iov into an array of pages and offset/length tupples. With the
1350  * partial_page structure, we can map several non-contiguous ranges into
1351  * our ones pages[] map instead of splitting that operation into pieces.
1352  * Could easily be exported as a generic helper for other users, in which
1353  * case one would probably want to add a 'max_nr_pages' parameter as well.
1354  */
1355 static int get_iovec_page_array(const struct iovec __user *iov,
1356                                 unsigned int nr_vecs, struct page **pages,
1357                                 struct partial_page *partial, int aligned)
1358 {
1359         int buffers = 0, error = 0;
1360
1361         while (nr_vecs) {
1362                 unsigned long off, npages;
1363                 struct iovec entry;
1364                 void __user *base;
1365                 size_t len;
1366                 int i;
1367
1368                 error = -EFAULT;
1369                 if (copy_from_user(&entry, iov, sizeof(entry)))
1370                         break;
1371
1372                 base = entry.iov_base;
1373                 len = entry.iov_len;
1374
1375                 /*
1376                  * Sanity check this iovec. 0 read succeeds.
1377                  */
1378                 error = 0;
1379                 if (unlikely(!len))
1380                         break;
1381                 error = -EFAULT;
1382                 if (!access_ok(VERIFY_READ, base, len))
1383                         break;
1384
1385                 /*
1386                  * Get this base offset and number of pages, then map
1387                  * in the user pages.
1388                  */
1389                 off = (unsigned long) base & ~PAGE_MASK;
1390
1391                 /*
1392                  * If asked for alignment, the offset must be zero and the
1393                  * length a multiple of the PAGE_SIZE.
1394                  */
1395                 error = -EINVAL;
1396                 if (aligned && (off || len & ~PAGE_MASK))
1397                         break;
1398
1399                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1400                 if (npages > PIPE_BUFFERS - buffers)
1401                         npages = PIPE_BUFFERS - buffers;
1402
1403                 error = get_user_pages_fast((unsigned long)base, npages,
1404                                         0, &pages[buffers]);
1405
1406                 if (unlikely(error <= 0))
1407                         break;
1408
1409                 /*
1410                  * Fill this contiguous range into the partial page map.
1411                  */
1412                 for (i = 0; i < error; i++) {
1413                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1414
1415                         partial[buffers].offset = off;
1416                         partial[buffers].len = plen;
1417
1418                         off = 0;
1419                         len -= plen;
1420                         buffers++;
1421                 }
1422
1423                 /*
1424                  * We didn't complete this iov, stop here since it probably
1425                  * means we have to move some of this into a pipe to
1426                  * be able to continue.
1427                  */
1428                 if (len)
1429                         break;
1430
1431                 /*
1432                  * Don't continue if we mapped fewer pages than we asked for,
1433                  * or if we mapped the max number of pages that we have
1434                  * room for.
1435                  */
1436                 if (error < npages || buffers == PIPE_BUFFERS)
1437                         break;
1438
1439                 nr_vecs--;
1440                 iov++;
1441         }
1442
1443         if (buffers)
1444                 return buffers;
1445
1446         return error;
1447 }
1448
1449 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1450                         struct splice_desc *sd)
1451 {
1452         char *src;
1453         int ret;
1454
1455         ret = buf->ops->confirm(pipe, buf);
1456         if (unlikely(ret))
1457                 return ret;
1458
1459         /*
1460          * See if we can use the atomic maps, by prefaulting in the
1461          * pages and doing an atomic copy
1462          */
1463         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1464                 src = buf->ops->map(pipe, buf, 1);
1465                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1466                                                         sd->len);
1467                 buf->ops->unmap(pipe, buf, src);
1468                 if (!ret) {
1469                         ret = sd->len;
1470                         goto out;
1471                 }
1472         }
1473
1474         /*
1475          * No dice, use slow non-atomic map and copy
1476          */
1477         src = buf->ops->map(pipe, buf, 0);
1478
1479         ret = sd->len;
1480         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1481                 ret = -EFAULT;
1482
1483         buf->ops->unmap(pipe, buf, src);
1484 out:
1485         if (ret > 0)
1486                 sd->u.userptr += ret;
1487         return ret;
1488 }
1489
1490 /*
1491  * For lack of a better implementation, implement vmsplice() to userspace
1492  * as a simple copy of the pipes pages to the user iov.
1493  */
1494 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1495                              unsigned long nr_segs, unsigned int flags)
1496 {
1497         struct pipe_inode_info *pipe;
1498         struct splice_desc sd;
1499         ssize_t size;
1500         int error;
1501         long ret;
1502
1503         pipe = pipe_info(file->f_path.dentry->d_inode);
1504         if (!pipe)
1505                 return -EBADF;
1506
1507         pipe_lock(pipe);
1508
1509         error = ret = 0;
1510         while (nr_segs) {
1511                 void __user *base;
1512                 size_t len;
1513
1514                 /*
1515                  * Get user address base and length for this iovec.
1516                  */
1517                 error = get_user(base, &iov->iov_base);
1518                 if (unlikely(error))
1519                         break;
1520                 error = get_user(len, &iov->iov_len);
1521                 if (unlikely(error))
1522                         break;
1523
1524                 /*
1525                  * Sanity check this iovec. 0 read succeeds.
1526                  */
1527                 if (unlikely(!len))
1528                         break;
1529                 if (unlikely(!base)) {
1530                         error = -EFAULT;
1531                         break;
1532                 }
1533
1534                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1535                         error = -EFAULT;
1536                         break;
1537                 }
1538
1539                 sd.len = 0;
1540                 sd.total_len = len;
1541                 sd.flags = flags;
1542                 sd.u.userptr = base;
1543                 sd.pos = 0;
1544
1545                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1546                 if (size < 0) {
1547                         if (!ret)
1548                                 ret = size;
1549
1550                         break;
1551                 }
1552
1553                 ret += size;
1554
1555                 if (size < len)
1556                         break;
1557
1558                 nr_segs--;
1559                 iov++;
1560         }
1561
1562         pipe_unlock(pipe);
1563
1564         if (!ret)
1565                 ret = error;
1566
1567         return ret;
1568 }
1569
1570 /*
1571  * vmsplice splices a user address range into a pipe. It can be thought of
1572  * as splice-from-memory, where the regular splice is splice-from-file (or
1573  * to file). In both cases the output is a pipe, naturally.
1574  */
1575 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1576                              unsigned long nr_segs, unsigned int flags)
1577 {
1578         struct pipe_inode_info *pipe;
1579         struct page *pages[PIPE_BUFFERS];
1580         struct partial_page partial[PIPE_BUFFERS];
1581         struct splice_pipe_desc spd = {
1582                 .pages = pages,
1583                 .partial = partial,
1584                 .flags = flags,
1585                 .ops = &user_page_pipe_buf_ops,
1586                 .spd_release = spd_release_page,
1587         };
1588
1589         pipe = pipe_info(file->f_path.dentry->d_inode);
1590         if (!pipe)
1591                 return -EBADF;
1592
1593         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1594                                             flags & SPLICE_F_GIFT);
1595         if (spd.nr_pages <= 0)
1596                 return spd.nr_pages;
1597
1598         return splice_to_pipe(pipe, &spd);
1599 }
1600
1601 /*
1602  * Note that vmsplice only really supports true splicing _from_ user memory
1603  * to a pipe, not the other way around. Splicing from user memory is a simple
1604  * operation that can be supported without any funky alignment restrictions
1605  * or nasty vm tricks. We simply map in the user memory and fill them into
1606  * a pipe. The reverse isn't quite as easy, though. There are two possible
1607  * solutions for that:
1608  *
1609  *      - memcpy() the data internally, at which point we might as well just
1610  *        do a regular read() on the buffer anyway.
1611  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1612  *        has restriction limitations on both ends of the pipe).
1613  *
1614  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1615  *
1616  */
1617 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1618                 unsigned long, nr_segs, unsigned int, flags)
1619 {
1620         struct file *file;
1621         long error;
1622         int fput;
1623
1624         if (unlikely(nr_segs > UIO_MAXIOV))
1625                 return -EINVAL;
1626         else if (unlikely(!nr_segs))
1627                 return 0;
1628
1629         error = -EBADF;
1630         file = fget_light(fd, &fput);
1631         if (file) {
1632                 if (file->f_mode & FMODE_WRITE)
1633                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1634                 else if (file->f_mode & FMODE_READ)
1635                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1636
1637                 fput_light(file, fput);
1638         }
1639
1640         return error;
1641 }
1642
1643 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1644                 int, fd_out, loff_t __user *, off_out,
1645                 size_t, len, unsigned int, flags)
1646 {
1647         long error;
1648         struct file *in, *out;
1649         int fput_in, fput_out;
1650
1651         if (unlikely(!len))
1652                 return 0;
1653
1654         error = -EBADF;
1655         in = fget_light(fd_in, &fput_in);
1656         if (in) {
1657                 if (in->f_mode & FMODE_READ) {
1658                         out = fget_light(fd_out, &fput_out);
1659                         if (out) {
1660                                 if (out->f_mode & FMODE_WRITE)
1661                                         error = do_splice(in, off_in,
1662                                                           out, off_out,
1663                                                           len, flags);
1664                                 fput_light(out, fput_out);
1665                         }
1666                 }
1667
1668                 fput_light(in, fput_in);
1669         }
1670
1671         return error;
1672 }
1673
1674 /*
1675  * Make sure there's data to read. Wait for input if we can, otherwise
1676  * return an appropriate error.
1677  */
1678 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1679 {
1680         int ret;
1681
1682         /*
1683          * Check ->nrbufs without the inode lock first. This function
1684          * is speculative anyways, so missing one is ok.
1685          */
1686         if (pipe->nrbufs)
1687                 return 0;
1688
1689         ret = 0;
1690         pipe_lock(pipe);
1691
1692         while (!pipe->nrbufs) {
1693                 if (signal_pending(current)) {
1694                         ret = -ERESTARTSYS;
1695                         break;
1696                 }
1697                 if (!pipe->writers)
1698                         break;
1699                 if (!pipe->waiting_writers) {
1700                         if (flags & SPLICE_F_NONBLOCK) {
1701                                 ret = -EAGAIN;
1702                                 break;
1703                         }
1704                 }
1705                 pipe_wait(pipe);
1706         }
1707
1708         pipe_unlock(pipe);
1709         return ret;
1710 }
1711
1712 /*
1713  * Make sure there's writeable room. Wait for room if we can, otherwise
1714  * return an appropriate error.
1715  */
1716 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1717 {
1718         int ret;
1719
1720         /*
1721          * Check ->nrbufs without the inode lock first. This function
1722          * is speculative anyways, so missing one is ok.
1723          */
1724         if (pipe->nrbufs < PIPE_BUFFERS)
1725                 return 0;
1726
1727         ret = 0;
1728         pipe_lock(pipe);
1729
1730         while (pipe->nrbufs >= PIPE_BUFFERS) {
1731                 if (!pipe->readers) {
1732                         send_sig(SIGPIPE, current, 0);
1733                         ret = -EPIPE;
1734                         break;
1735                 }
1736                 if (flags & SPLICE_F_NONBLOCK) {
1737                         ret = -EAGAIN;
1738                         break;
1739                 }
1740                 if (signal_pending(current)) {
1741                         ret = -ERESTARTSYS;
1742                         break;
1743                 }
1744                 pipe->waiting_writers++;
1745                 pipe_wait(pipe);
1746                 pipe->waiting_writers--;
1747         }
1748
1749         pipe_unlock(pipe);
1750         return ret;
1751 }
1752
1753 /*
1754  * Splice contents of ipipe to opipe.
1755  */
1756 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1757                                struct pipe_inode_info *opipe,
1758                                size_t len, unsigned int flags)
1759 {
1760         struct pipe_buffer *ibuf, *obuf;
1761         int ret = 0, nbuf;
1762         bool input_wakeup = false;
1763
1764
1765 retry:
1766         ret = ipipe_prep(ipipe, flags);
1767         if (ret)
1768                 return ret;
1769
1770         ret = opipe_prep(opipe, flags);
1771         if (ret)
1772                 return ret;
1773
1774         /*
1775          * Potential ABBA deadlock, work around it by ordering lock
1776          * grabbing by pipe info address. Otherwise two different processes
1777          * could deadlock (one doing tee from A -> B, the other from B -> A).
1778          */
1779         pipe_double_lock(ipipe, opipe);
1780
1781         do {
1782                 if (!opipe->readers) {
1783                         send_sig(SIGPIPE, current, 0);
1784                         if (!ret)
1785                                 ret = -EPIPE;
1786                         break;
1787                 }
1788
1789                 if (!ipipe->nrbufs && !ipipe->writers)
1790                         break;
1791
1792                 /*
1793                  * Cannot make any progress, because either the input
1794                  * pipe is empty or the output pipe is full.
1795                  */
1796                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1797                         /* Already processed some buffers, break */
1798                         if (ret)
1799                                 break;
1800
1801                         if (flags & SPLICE_F_NONBLOCK) {
1802                                 ret = -EAGAIN;
1803                                 break;
1804                         }
1805
1806                         /*
1807                          * We raced with another reader/writer and haven't
1808                          * managed to process any buffers.  A zero return
1809                          * value means EOF, so retry instead.
1810                          */
1811                         pipe_unlock(ipipe);
1812                         pipe_unlock(opipe);
1813                         goto retry;
1814                 }
1815
1816                 ibuf = ipipe->bufs + ipipe->curbuf;
1817                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1818                 obuf = opipe->bufs + nbuf;
1819
1820                 if (len >= ibuf->len) {
1821                         /*
1822                          * Simply move the whole buffer from ipipe to opipe
1823                          */
1824                         *obuf = *ibuf;
1825                         ibuf->ops = NULL;
1826                         opipe->nrbufs++;
1827                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1828                         ipipe->nrbufs--;
1829                         input_wakeup = true;
1830                 } else {
1831                         /*
1832                          * Get a reference to this pipe buffer,
1833                          * so we can copy the contents over.
1834                          */
1835                         ibuf->ops->get(ipipe, ibuf);
1836                         *obuf = *ibuf;
1837
1838                         /*
1839                          * Don't inherit the gift flag, we need to
1840                          * prevent multiple steals of this page.
1841                          */
1842                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1843
1844                         obuf->len = len;
1845                         opipe->nrbufs++;
1846                         ibuf->offset += obuf->len;
1847                         ibuf->len -= obuf->len;
1848                 }
1849                 ret += obuf->len;
1850                 len -= obuf->len;
1851         } while (len);
1852
1853         pipe_unlock(ipipe);
1854         pipe_unlock(opipe);
1855
1856         /*
1857          * If we put data in the output pipe, wakeup any potential readers.
1858          */
1859         if (ret > 0) {
1860                 smp_mb();
1861                 if (waitqueue_active(&opipe->wait))
1862                         wake_up_interruptible(&opipe->wait);
1863                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1864         }
1865         if (input_wakeup)
1866                 wakeup_pipe_writers(ipipe);
1867
1868         return ret;
1869 }
1870
1871 /*
1872  * Link contents of ipipe to opipe.
1873  */
1874 static int link_pipe(struct pipe_inode_info *ipipe,
1875                      struct pipe_inode_info *opipe,
1876                      size_t len, unsigned int flags)
1877 {
1878         struct pipe_buffer *ibuf, *obuf;
1879         int ret = 0, i = 0, nbuf;
1880
1881         /*
1882          * Potential ABBA deadlock, work around it by ordering lock
1883          * grabbing by pipe info address. Otherwise two different processes
1884          * could deadlock (one doing tee from A -> B, the other from B -> A).
1885          */
1886         pipe_double_lock(ipipe, opipe);
1887
1888         do {
1889                 if (!opipe->readers) {
1890                         send_sig(SIGPIPE, current, 0);
1891                         if (!ret)
1892                                 ret = -EPIPE;
1893                         break;
1894                 }
1895
1896                 /*
1897                  * If we have iterated all input buffers or ran out of
1898                  * output room, break.
1899                  */
1900                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1901                         break;
1902
1903                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1904                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1905
1906                 /*
1907                  * Get a reference to this pipe buffer,
1908                  * so we can copy the contents over.
1909                  */
1910                 ibuf->ops->get(ipipe, ibuf);
1911
1912                 obuf = opipe->bufs + nbuf;
1913                 *obuf = *ibuf;
1914
1915                 /*
1916                  * Don't inherit the gift flag, we need to
1917                  * prevent multiple steals of this page.
1918                  */
1919                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1920
1921                 if (obuf->len > len)
1922                         obuf->len = len;
1923
1924                 opipe->nrbufs++;
1925                 ret += obuf->len;
1926                 len -= obuf->len;
1927                 i++;
1928         } while (len);
1929
1930         /*
1931          * return EAGAIN if we have the potential of some data in the
1932          * future, otherwise just return 0
1933          */
1934         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1935                 ret = -EAGAIN;
1936
1937         pipe_unlock(ipipe);
1938         pipe_unlock(opipe);
1939
1940         /*
1941          * If we put data in the output pipe, wakeup any potential readers.
1942          */
1943         if (ret > 0) {
1944                 smp_mb();
1945                 if (waitqueue_active(&opipe->wait))
1946                         wake_up_interruptible(&opipe->wait);
1947                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1948         }
1949
1950         return ret;
1951 }
1952
1953 /*
1954  * This is a tee(1) implementation that works on pipes. It doesn't copy
1955  * any data, it simply references the 'in' pages on the 'out' pipe.
1956  * The 'flags' used are the SPLICE_F_* variants, currently the only
1957  * applicable one is SPLICE_F_NONBLOCK.
1958  */
1959 static long do_tee(struct file *in, struct file *out, size_t len,
1960                    unsigned int flags)
1961 {
1962         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1963         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1964         int ret = -EINVAL;
1965
1966         /*
1967          * Duplicate the contents of ipipe to opipe without actually
1968          * copying the data.
1969          */
1970         if (ipipe && opipe && ipipe != opipe) {
1971                 /*
1972                  * Keep going, unless we encounter an error. The ipipe/opipe
1973                  * ordering doesn't really matter.
1974                  */
1975                 ret = ipipe_prep(ipipe, flags);
1976                 if (!ret) {
1977                         ret = opipe_prep(opipe, flags);
1978                         if (!ret)
1979                                 ret = link_pipe(ipipe, opipe, len, flags);
1980                 }
1981         }
1982
1983         return ret;
1984 }
1985
1986 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1987 {
1988         struct file *in;
1989         int error, fput_in;
1990
1991         if (unlikely(!len))
1992                 return 0;
1993
1994         error = -EBADF;
1995         in = fget_light(fdin, &fput_in);
1996         if (in) {
1997                 if (in->f_mode & FMODE_READ) {
1998                         int fput_out;
1999                         struct file *out = fget_light(fdout, &fput_out);
2000
2001                         if (out) {
2002                                 if (out->f_mode & FMODE_WRITE)
2003                                         error = do_tee(in, out, len, flags);
2004                                 fput_light(out, fput_out);
2005                         }
2006                 }
2007                 fput_light(in, fput_in);
2008         }
2009
2010         return error;
2011 }