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                         /*
 369                          * If in nonblock mode then dont block on waiting
 370                          * for an in-flight io page
 371                          */
 372                         if (flags & SPLICE_F_NONBLOCK) {
 373                                 if (!trylock_page(page)) {
 374                                         error = -EAGAIN;
 375                                         break;
 376                                 }
 377                         } else
 378                                 lock_page(page);
 379
 380                         /*
 381                          * Page was truncated, or invalidated by the
 382                          * filesystem.  Redo the find/create, but this time the
 383                          * page is kept locked, so there's no chance of another
 384                          * race with truncate/invalidate.
 385                          */
 386                         if (!page->mapping) {
 387                                 unlock_page(page);
 388                                 page = find_or_create_page(mapping, index,
 389                                                 mapping_gfp_mask(mapping));
 390
 391                                 if (!page) {
 392                                         error = -ENOMEM;
 393                                         break;
 394                                 }
 395                                 page_cache_release(pages[page_nr]);
 396                                 pages[page_nr] = page;
 397                         }
 398                         /*
 399                          * page was already under io and is now done, great
 400                          */
 401                         if (PageUptodate(page)) {
 402                                 unlock_page(page);
 403                                 goto fill_it;
 404                         }
 405
 406                         /*
 407                          * need to read in the page
 408                          */
 409                         error = mapping->a_ops->readpage(in, page);
 410                         if (unlikely(error)) {
 411                                 /*
 412                                  * We really should re-lookup the page here,
 413                                  * but it complicates things a lot. Instead
 414                                  * lets just do what we already stored, and
 415                                  * we'll get it the next time we are called.
 416                                  */
 417                                 if (error == AOP_TRUNCATED_PAGE)
 418                                         error = 0;
 419
 420                                 break;
 421                         }
 422                 }
 423 fill_it:
 424                 /*
 425                  * i_size must be checked after PageUptodate.
 426                  */
 427                 isize = i_size_read(mapping->host);
 428                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 429                 if (unlikely(!isize || index > end_index))
 430                         break;
 431
 432                 /*
 433                  * if this is the last page, see if we need to shrink
 434                  * the length and stop
 435                  */
 436                 if (end_index == index) {
 437                         unsigned int plen;
 438
 439                         /*
 440                          * max good bytes in this page
 441                          */
 442                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 443                         if (plen <= loff)
 444                                 break;
 445
 446                         /*
 447                          * force quit after adding this page
 448                          */
 449                         this_len = min(this_len, plen - loff);
 450                         len = this_len;
 451                 }
 452
 453                 partial[page_nr].offset = loff;
 454                 partial[page_nr].len = this_len;
 455                 len -= this_len;
 456                 loff = 0;
 457                 spd.nr_pages++;
 458                 index++;
 459         }
 460
 461         /*
 462          * Release any pages at the end, if we quit early. 'page_nr' is how far
 463          * we got, 'nr_pages' is how many pages are in the map.
 464          */
 465         while (page_nr < nr_pages)
 466                 page_cache_release(pages[page_nr++]);
 467         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 468
 469         if (spd.nr_pages)
 470                 return splice_to_pipe(pipe, &spd);
 471
 472         return error;
 473 }
 474
 475 /**
 476  * generic_file_splice_read - splice data from file to a pipe
 477  * @in:         file to splice from
 478  * @ppos:       position in @in
 479  * @pipe:       pipe to splice to
 480  * @len:        number of bytes to splice
 481  * @flags:      splice modifier flags
 482  *
 483  * Description:
 484  *    Will read pages from given file and fill them into a pipe. Can be
 485  *    used as long as the address_space operations for the source implements
 486  *    a readpage() hook.
 487  *
 488  */
 489 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 490                                  struct pipe_inode_info *pipe, size_t len,
 491                                  unsigned int flags)
 492 {
 493         loff_t isize, left;
 494         int ret;
 495
 496         isize = i_size_read(in->f_mapping->host);
 497         if (unlikely(*ppos >= isize))
 498                 return 0;
 499
 500         left = isize - *ppos;
 501         if (unlikely(left < len))
 502                 len = left;
 503
 504         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 505         if (ret > 0) {
 506                 *ppos += ret;
 507                 file_accessed(in);
 508         }
 509
 510         return ret;
 511 }
 512 EXPORT_SYMBOL(generic_file_splice_read);
 513
 514 static const struct pipe_buf_operations default_pipe_buf_ops = {
 515         .can_merge = 0,
 516         .map = generic_pipe_buf_map,
 517         .unmap = generic_pipe_buf_unmap,
 518         .confirm = generic_pipe_buf_confirm,
 519         .release = generic_pipe_buf_release,
 520         .steal = generic_pipe_buf_steal,
 521         .get = generic_pipe_buf_get,
 522 };
 523
 524 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 525                             unsigned long vlen, loff_t offset)
 526 {
 527         mm_segment_t old_fs;
 528         loff_t pos = offset;
 529         ssize_t res;
 530
 531         old_fs = get_fs();
 532         set_fs(get_ds());
 533         /* The cast to a user pointer is valid due to the set_fs() */
 534         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 535         set_fs(old_fs);
 536
 537         return res;
 538 }
 539
 540 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 541                             loff_t pos)
 542 {
 543         mm_segment_t old_fs;
 544         ssize_t res;
 545
 546         old_fs = get_fs();
 547         set_fs(get_ds());
 548         /* The cast to a user pointer is valid due to the set_fs() */
 549         res = vfs_write(file, (const char __user *)buf, count, &pos);
 550         set_fs(old_fs);
 551
 552         return res;
 553 }
 554
 555 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 556                                  struct pipe_inode_info *pipe, size_t len,
 557                                  unsigned int flags)
 558 {
 559         unsigned int nr_pages;
 560         unsigned int nr_freed;
 561         size_t offset;
 562         struct page *pages[PIPE_BUFFERS];
 563         struct partial_page partial[PIPE_BUFFERS];
 564         struct iovec vec[PIPE_BUFFERS];
 565         pgoff_t index;
 566         ssize_t res;
 567         size_t this_len;
 568         int error;
 569         int i;
 570         struct splice_pipe_desc spd = {
 571                 .pages = pages,
 572                 .partial = partial,
 573                 .flags = flags,
 574                 .ops = &default_pipe_buf_ops,
 575                 .spd_release = spd_release_page,
 576         };
 577
 578         index = *ppos >> PAGE_CACHE_SHIFT;
 579         offset = *ppos & ~PAGE_CACHE_MASK;
 580         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 581
 582         for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
 583                 struct page *page;
 584
 585                 page = alloc_page(GFP_USER);
 586                 error = -ENOMEM;
 587                 if (!page)
 588                         goto err;
 589
 590                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 591                 vec[i].iov_base = (void __user *) page_address(page);
 592                 vec[i].iov_len = this_len;
 593                 pages[i] = page;
 594                 spd.nr_pages++;
 595                 len -= this_len;
 596                 offset = 0;
 597         }
 598
 599         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 600         if (res < 0) {
 601                 error = res;
 602                 goto err;
 603         }
 604
 605         error = 0;
 606         if (!res)
 607                 goto err;
 608
 609         nr_freed = 0;
 610         for (i = 0; i < spd.nr_pages; i++) {
 611                 this_len = min_t(size_t, vec[i].iov_len, res);
 612                 partial[i].offset = 0;
 613                 partial[i].len = this_len;
 614                 if (!this_len) {
 615                         __free_page(pages[i]);
 616                         pages[i] = NULL;
 617                         nr_freed++;
 618                 }
 619                 res -= this_len;
 620         }
 621         spd.nr_pages -= nr_freed;
 622
 623         res = splice_to_pipe(pipe, &spd);
 624         if (res > 0)
 625                 *ppos += res;
 626
 627         return res;
 628
 629 err:
 630         for (i = 0; i < spd.nr_pages; i++)
 631                 __free_page(pages[i]);
 632
 633         return error;
 634 }
 635 EXPORT_SYMBOL(default_file_splice_read);
 636
 637 /*
 638  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 639  * using sendpage(). Return the number of bytes sent.
 640  */
 641 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 642                             struct pipe_buffer *buf, struct splice_desc *sd)
 643 {
 644         struct file *file = sd->u.file;
 645         loff_t pos = sd->pos;
 646         int ret, more;
 647
 648         ret = buf->ops->confirm(pipe, buf);
 649         if (!ret) {
 650                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 651
 652                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
 653                                            sd->len, &pos, more);
 654         }
 655
 656         return ret;
 657 }
 658
 659 /*
 660  * This is a little more tricky than the file -> pipe splicing. There are
 661  * basically three cases:
 662  *
 663  *      - Destination page already exists in the address space and there
 664  *        are users of it. For that case we have no other option that
 665  *        copying the data. Tough luck.
 666  *      - Destination page already exists in the address space, but there
 667  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 668  *        through to last case.
 669  *      - Destination page does not exist, we can add the pipe page to
 670  *        the page cache and avoid the copy.
 671  *
 672  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 673  * sd->flags), we attempt to migrate pages from the pipe to the output
 674  * file address space page cache. This is possible if no one else has
 675  * the pipe page referenced outside of the pipe and page cache. If
 676  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 677  * a new page in the output file page cache and fill/dirty that.
 678  */
 679 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 680                  struct splice_desc *sd)
 681 {
 682         struct file *file = sd->u.file;
 683         struct address_space *mapping = file->f_mapping;
 684         unsigned int offset, this_len;
 685         struct page *page;
 686         void *fsdata;
 687         int ret;
 688
 689         /*
 690          * make sure the data in this buffer is uptodate
 691          */
 692         ret = buf->ops->confirm(pipe, buf);
 693         if (unlikely(ret))
 694                 return ret;
 695
 696         offset = sd->pos & ~PAGE_CACHE_MASK;
 697
 698         this_len = sd->len;
 699         if (this_len + offset > PAGE_CACHE_SIZE)
 700                 this_len = PAGE_CACHE_SIZE - offset;
 701
 702         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 703                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 704         if (unlikely(ret))
 705                 goto out;
 706
 707         if (buf->page != page) {
 708                 /*
 709                  * Careful, ->map() uses KM_USER0!
 710                  */
 711                 char *src = buf->ops->map(pipe, buf, 1);
 712                 char *dst = kmap_atomic(page, KM_USER1);
 713
 714                 memcpy(dst + offset, src + buf->offset, this_len);
 715                 flush_dcache_page(page);
 716                 kunmap_atomic(dst, KM_USER1);
 717                 buf->ops->unmap(pipe, buf, src);
 718         }
 719         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 720                                 page, fsdata);
 721 out:
 722         return ret;
 723 }
 724 EXPORT_SYMBOL(pipe_to_file);
 725
 726 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 727 {
 728         smp_mb();
 729         if (waitqueue_active(&pipe->wait))
 730                 wake_up_interruptible(&pipe->wait);
 731         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 732 }
 733
 734 /**
 735  * splice_from_pipe_feed - feed available data from a pipe to a file
 736  * @pipe:       pipe to splice from
 737  * @sd:         information to @actor
 738  * @actor:      handler that splices the data
 739  *
 740  * Description:
 741  *    This function loops over the pipe and calls @actor to do the
 742  *    actual moving of a single struct pipe_buffer to the desired
 743  *    destination.  It returns when there's no more buffers left in
 744  *    the pipe or if the requested number of bytes (@sd->total_len)
 745  *    have been copied.  It returns a positive number (one) if the
 746  *    pipe needs to be filled with more data, zero if the required
 747  *    number of bytes have been copied and -errno on error.
 748  *
 749  *    This, together with splice_from_pipe_{begin,end,next}, may be
 750  *    used to implement the functionality of __splice_from_pipe() when
 751  *    locking is required around copying the pipe buffers to the
 752  *    destination.
 753  */
 754 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 755                           splice_actor *actor)
 756 {
 757         int ret;
 758
 759         while (pipe->nrbufs) {
 760                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 761                 const struct pipe_buf_operations *ops = buf->ops;
 762
 763                 sd->len = buf->len;
 764                 if (sd->len > sd->total_len)
 765                         sd->len = sd->total_len;
 766
 767                 ret = actor(pipe, buf, sd);
 768                 if (ret <= 0) {
 769                         if (ret == -ENODATA)
 770                                 ret = 0;
 771                         return ret;
 772                 }
 773                 buf->offset += ret;
 774                 buf->len -= ret;
 775
 776                 sd->num_spliced += ret;
 777                 sd->len -= ret;
 778                 sd->pos += ret;
 779                 sd->total_len -= ret;
 780
 781                 if (!buf->len) {
 782                         buf->ops = NULL;
 783                         ops->release(pipe, buf);
 784                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
 785                         pipe->nrbufs--;
 786                         if (pipe->inode)
 787                                 sd->need_wakeup = true;
 788                 }
 789
 790                 if (!sd->total_len)
 791                         return 0;
 792         }
 793
 794         return 1;
 795 }
 796 EXPORT_SYMBOL(splice_from_pipe_feed);
 797
 798 /**
 799  * splice_from_pipe_next - wait for some data to splice from
 800  * @pipe:       pipe to splice from
 801  * @sd:         information about the splice operation
 802  *
 803  * Description:
 804  *    This function will wait for some data and return a positive
 805  *    value (one) if pipe buffers are available.  It will return zero
 806  *    or -errno if no more data needs to be spliced.
 807  */
 808 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 809 {
 810         while (!pipe->nrbufs) {
 811                 if (!pipe->writers)
 812                         return 0;
 813
 814                 if (!pipe->waiting_writers && sd->num_spliced)
 815                         return 0;
 816
 817                 if (sd->flags & SPLICE_F_NONBLOCK)
 818                         return -EAGAIN;
 819
 820                 if (signal_pending(current))
 821                         return -ERESTARTSYS;
 822
 823                 if (sd->need_wakeup) {
 824                         wakeup_pipe_writers(pipe);
 825                         sd->need_wakeup = false;
 826                 }
 827
 828                 pipe_wait(pipe);
 829         }
 830
 831         return 1;
 832 }
 833 EXPORT_SYMBOL(splice_from_pipe_next);
 834
 835 /**
 836  * splice_from_pipe_begin - start splicing from pipe
 837  * @sd:         information about the splice operation
 838  *
 839  * Description:
 840  *    This function should be called before a loop containing
 841  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 842  *    initialize the necessary fields of @sd.
 843  */
 844 void splice_from_pipe_begin(struct splice_desc *sd)
 845 {
 846         sd->num_spliced = 0;
 847         sd->need_wakeup = false;
 848 }
 849 EXPORT_SYMBOL(splice_from_pipe_begin);
 850
 851 /**
 852  * splice_from_pipe_end - finish splicing from pipe
 853  * @pipe:       pipe to splice from
 854  * @sd:         information about the splice operation
 855  *
 856  * Description:
 857  *    This function will wake up pipe writers if necessary.  It should
 858  *    be called after a loop containing splice_from_pipe_next() and
 859  *    splice_from_pipe_feed().
 860  */
 861 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 862 {
 863         if (sd->need_wakeup)
 864                 wakeup_pipe_writers(pipe);
 865 }
 866 EXPORT_SYMBOL(splice_from_pipe_end);
 867
 868 /**
 869  * __splice_from_pipe - splice data from a pipe to given actor
 870  * @pipe:       pipe to splice from
 871  * @sd:         information to @actor
 872  * @actor:      handler that splices the data
 873  *
 874  * Description:
 875  *    This function does little more than loop over the pipe and call
 876  *    @actor to do the actual moving of a single struct pipe_buffer to
 877  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 878  *    pipe_to_user.
 879  *
 880  */
 881 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 882                            splice_actor *actor)
 883 {
 884         int ret;
 885
 886         splice_from_pipe_begin(sd);
 887         do {
 888                 ret = splice_from_pipe_next(pipe, sd);
 889                 if (ret > 0)
 890                         ret = splice_from_pipe_feed(pipe, sd, actor);
 891         } while (ret > 0);
 892         splice_from_pipe_end(pipe, sd);
 893
 894         return sd->num_spliced ? sd->num_spliced : ret;
 895 }
 896 EXPORT_SYMBOL(__splice_from_pipe);
 897
 898 /**
 899  * splice_from_pipe - splice data from a pipe to a file
 900  * @pipe:       pipe to splice from
 901  * @out:        file to splice to
 902  * @ppos:       position in @out
 903  * @len:        how many bytes to splice
 904  * @flags:      splice modifier flags
 905  * @actor:      handler that splices the data
 906  *
 907  * Description:
 908  *    See __splice_from_pipe. This function locks the pipe inode,
 909  *    otherwise it's identical to __splice_from_pipe().
 910  *
 911  */
 912 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 913                          loff_t *ppos, size_t len, unsigned int flags,
 914                          splice_actor *actor)
 915 {
 916         ssize_t ret;
 917         struct splice_desc sd = {
 918                 .total_len = len,
 919                 .flags = flags,
 920                 .pos = *ppos,
 921                 .u.file = out,
 922         };
 923
 924         pipe_lock(pipe);
 925         ret = __splice_from_pipe(pipe, &sd, actor);
 926         pipe_unlock(pipe);
 927
 928         return ret;
 929 }
 930
 931 /**
 932  * generic_file_splice_write - splice data from a pipe to a file
 933  * @pipe:       pipe info
 934  * @out:        file to write to
 935  * @ppos:       position in @out
 936  * @len:        number of bytes to splice
 937  * @flags:      splice modifier flags
 938  *
 939  * Description:
 940  *    Will either move or copy pages (determined by @flags options) from
 941  *    the given pipe inode to the given file.
 942  *
 943  */
 944 ssize_t
 945 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 946                           loff_t *ppos, size_t len, unsigned int flags)
 947 {
 948         struct address_space *mapping = out->f_mapping;
 949         struct inode *inode = mapping->host;
 950         struct splice_desc sd = {
 951                 .total_len = len,
 952                 .flags = flags,
 953                 .pos = *ppos,
 954                 .u.file = out,
 955         };
 956         ssize_t ret;
 957
 958         pipe_lock(pipe);
 959
 960         splice_from_pipe_begin(&sd);
 961         do {
 962                 ret = splice_from_pipe_next(pipe, &sd);
 963                 if (ret <= 0)
 964                         break;
 965
 966                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
 967                 ret = file_remove_suid(out);
 968                 if (!ret) {
 969                         file_update_time(out);
 970                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
 971                 }
 972                 mutex_unlock(&inode->i_mutex);
 973         } while (ret > 0);
 974         splice_from_pipe_end(pipe, &sd);
 975
 976         pipe_unlock(pipe);
 977
 978         if (sd.num_spliced)
 979                 ret = sd.num_spliced;
 980
 981         if (ret > 0) {
 982                 unsigned long nr_pages;
 983                 int err;
 984
 985                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 986
 987                 err = generic_write_sync(out, *ppos, ret);
 988                 if (err)
 989                         ret = err;
 990                 else
 991                         *ppos += ret;
 992                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
 993         }
 994
 995         return ret;
 996 }
 997
 998 EXPORT_SYMBOL(generic_file_splice_write);
 999
1000 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1001                           struct splice_desc *sd)
1002 {
1003         int ret;
1004         void *data;
1005
1006         ret = buf->ops->confirm(pipe, buf);
1007         if (ret)
1008                 return ret;
1009
1010         data = buf->ops->map(pipe, buf, 0);
1011         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1012         buf->ops->unmap(pipe, buf, data);
1013
1014         return ret;
1015 }
1016
1017 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1018                                          struct file *out, loff_t *ppos,
1019                                          size_t len, unsigned int flags)
1020 {
1021         ssize_t ret;
1022
1023         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1024         if (ret > 0)
1025                 *ppos += ret;
1026
1027         return ret;
1028 }
1029
1030 /**
1031  * generic_splice_sendpage - splice data from a pipe to a socket
1032  * @pipe:       pipe to splice from
1033  * @out:        socket to write to
1034  * @ppos:       position in @out
1035  * @len:        number of bytes to splice
1036  * @flags:      splice modifier flags
1037  *
1038  * Description:
1039  *    Will send @len bytes from the pipe to a network socket. No data copying
1040  *    is involved.
1041  *
1042  */
1043 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1044                                 loff_t *ppos, size_t len, unsigned int flags)
1045 {
1046         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1047 }
1048
1049 EXPORT_SYMBOL(generic_splice_sendpage);
1050
1051 /*
1052  * Attempt to initiate a splice from pipe to file.
1053  */
1054 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1055                            loff_t *ppos, size_t len, unsigned int flags)
1056 {
1057         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1058                                 loff_t *, size_t, unsigned int);
1059         int ret;
1060
1061         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1062                 return -EBADF;
1063
1064         if (unlikely(out->f_flags & O_APPEND))
1065                 return -EINVAL;
1066
1067         ret = rw_verify_area(WRITE, out, ppos, len);
1068         if (unlikely(ret < 0))
1069                 return ret;
1070
1071         splice_write = out->f_op->splice_write;
1072         if (!splice_write)
1073                 splice_write = default_file_splice_write;
1074
1075         return splice_write(pipe, out, ppos, len, flags);
1076 }
1077
1078 /*
1079  * Attempt to initiate a splice from a file to a pipe.
1080  */
1081 static long do_splice_to(struct file *in, loff_t *ppos,
1082                          struct pipe_inode_info *pipe, size_t len,
1083                          unsigned int flags)
1084 {
1085         ssize_t (*splice_read)(struct file *, loff_t *,
1086                                struct pipe_inode_info *, size_t, unsigned int);
1087         int ret;
1088
1089         if (unlikely(!(in->f_mode & FMODE_READ)))
1090                 return -EBADF;
1091
1092         ret = rw_verify_area(READ, in, ppos, len);
1093         if (unlikely(ret < 0))
1094                 return ret;
1095
1096         splice_read = in->f_op->splice_read;
1097         if (!splice_read)
1098                 splice_read = default_file_splice_read;
1099
1100         return splice_read(in, ppos, pipe, len, flags);
1101 }
1102
1103 /**
1104  * splice_direct_to_actor - splices data directly between two non-pipes
1105  * @in:         file to splice from
1106  * @sd:         actor information on where to splice to
1107  * @actor:      handles the data splicing
1108  *
1109  * Description:
1110  *    This is a special case helper to splice directly between two
1111  *    points, without requiring an explicit pipe. Internally an allocated
1112  *    pipe is cached in the process, and reused during the lifetime of
1113  *    that process.
1114  *
1115  */
1116 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1117                                splice_direct_actor *actor)
1118 {
1119         struct pipe_inode_info *pipe;
1120         long ret, bytes;
1121         umode_t i_mode;
1122         size_t len;
1123         int i, flags;
1124
1125         /*
1126          * We require the input being a regular file, as we don't want to
1127          * randomly drop data for eg socket -> socket splicing. Use the
1128          * piped splicing for that!
1129          */
1130         i_mode = in->f_path.dentry->d_inode->i_mode;
1131         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1132                 return -EINVAL;
1133
1134         /*
1135          * neither in nor out is a pipe, setup an internal pipe attached to
1136          * 'out' and transfer the wanted data from 'in' to 'out' through that
1137          */
1138         pipe = current->splice_pipe;
1139         if (unlikely(!pipe)) {
1140                 pipe = alloc_pipe_info(NULL);
1141                 if (!pipe)
1142                         return -ENOMEM;
1143
1144                 /*
1145                  * We don't have an immediate reader, but we'll read the stuff
1146                  * out of the pipe right after the splice_to_pipe(). So set
1147                  * PIPE_READERS appropriately.
1148                  */
1149                 pipe->readers = 1;
1150
1151                 current->splice_pipe = pipe;
1152         }
1153
1154         /*
1155          * Do the splice.
1156          */
1157         ret = 0;
1158         bytes = 0;
1159         len = sd->total_len;
1160         flags = sd->flags;
1161
1162         /*
1163          * Don't block on output, we have to drain the direct pipe.
1164          */
1165         sd->flags &= ~SPLICE_F_NONBLOCK;
1166
1167         while (len) {
1168                 size_t read_len;
1169                 loff_t pos = sd->pos, prev_pos = pos;
1170
1171                 ret = do_splice_to(in, &pos, pipe, len, flags);
1172                 if (unlikely(ret <= 0))
1173                         goto out_release;
1174
1175                 read_len = ret;
1176                 sd->total_len = read_len;
1177
1178                 /*
1179                  * NOTE: nonblocking mode only applies to the input. We
1180                  * must not do the output in nonblocking mode as then we
1181                  * could get stuck data in the internal pipe:
1182                  */
1183                 ret = actor(pipe, sd);
1184                 if (unlikely(ret <= 0)) {
1185                         sd->pos = prev_pos;
1186                         goto out_release;
1187                 }
1188
1189                 bytes += ret;
1190                 len -= ret;
1191                 sd->pos = pos;
1192
1193                 if (ret < read_len) {
1194                         sd->pos = prev_pos + ret;
1195                         goto out_release;
1196                 }
1197         }
1198
1199 done:
1200         pipe->nrbufs = pipe->curbuf = 0;
1201         file_accessed(in);
1202         return bytes;
1203
1204 out_release:
1205         /*
1206          * If we did an incomplete transfer we must release
1207          * the pipe buffers in question:
1208          */
1209         for (i = 0; i < PIPE_BUFFERS; i++) {
1210                 struct pipe_buffer *buf = pipe->bufs + i;
1211
1212                 if (buf->ops) {
1213                         buf->ops->release(pipe, buf);
1214                         buf->ops = NULL;
1215                 }
1216         }
1217
1218         if (!bytes)
1219                 bytes = ret;
1220
1221         goto done;
1222 }
1223 EXPORT_SYMBOL(splice_direct_to_actor);
1224
1225 static int direct_splice_actor(struct pipe_inode_info *pipe,
1226                                struct splice_desc *sd)
1227 {
1228         struct file *file = sd->u.file;
1229
1230         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1231 }
1232
1233 /**
1234  * do_splice_direct - splices data directly between two files
1235  * @in:         file to splice from
1236  * @ppos:       input file offset
1237  * @out:        file to splice to
1238  * @len:        number of bytes to splice
1239  * @flags:      splice modifier flags
1240  *
1241  * Description:
1242  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1243  *    doing it in the application would incur an extra system call
1244  *    (splice in + splice out, as compared to just sendfile()). So this helper
1245  *    can splice directly through a process-private pipe.
1246  *
1247  */
1248 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1249                       size_t len, unsigned int flags)
1250 {
1251         struct splice_desc sd = {
1252                 .len            = len,
1253                 .total_len      = len,
1254                 .flags          = flags,
1255                 .pos            = *ppos,
1256                 .u.file         = out,
1257         };
1258         long ret;
1259
1260         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1261         if (ret > 0)
1262                 *ppos = sd.pos;
1263
1264         return ret;
1265 }
1266
1267 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1268                                struct pipe_inode_info *opipe,
1269                                size_t len, unsigned int flags);
1270 /*
1271  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1272  * location, so checking ->i_pipe is not enough to verify that this is a
1273  * pipe.
1274  */
1275 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1276 {
1277         if (S_ISFIFO(inode->i_mode))
1278                 return inode->i_pipe;
1279
1280         return NULL;
1281 }
1282
1283 /*
1284  * Determine where to splice to/from.
1285  */
1286 static long do_splice(struct file *in, loff_t __user *off_in,
1287                       struct file *out, loff_t __user *off_out,
1288                       size_t len, unsigned int flags)
1289 {
1290         struct pipe_inode_info *ipipe;
1291         struct pipe_inode_info *opipe;
1292         loff_t offset, *off;
1293         long ret;
1294
1295         ipipe = pipe_info(in->f_path.dentry->d_inode);
1296         opipe = pipe_info(out->f_path.dentry->d_inode);
1297
1298         if (ipipe && opipe) {
1299                 if (off_in || off_out)
1300                         return -ESPIPE;
1301
1302                 if (!(in->f_mode & FMODE_READ))
1303                         return -EBADF;
1304
1305                 if (!(out->f_mode & FMODE_WRITE))
1306                         return -EBADF;
1307
1308                 /* Splicing to self would be fun, but... */
1309                 if (ipipe == opipe)
1310                         return -EINVAL;
1311
1312                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1313         }
1314
1315         if (ipipe) {
1316                 if (off_in)
1317                         return -ESPIPE;
1318                 if (off_out) {
1319                         if (out->f_op->llseek == no_llseek)
1320                                 return -EINVAL;
1321                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1322                                 return -EFAULT;
1323                         off = &offset;
1324                 } else
1325                         off = &out->f_pos;
1326
1327                 ret = do_splice_from(ipipe, out, off, len, flags);
1328
1329                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1330                         ret = -EFAULT;
1331
1332                 return ret;
1333         }
1334
1335         if (opipe) {
1336                 if (off_out)
1337                         return -ESPIPE;
1338                 if (off_in) {
1339                         if (in->f_op->llseek == no_llseek)
1340                                 return -EINVAL;
1341                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1342                                 return -EFAULT;
1343                         off = &offset;
1344                 } else
1345                         off = &in->f_pos;
1346
1347                 ret = do_splice_to(in, off, opipe, len, flags);
1348
1349                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1350                         ret = -EFAULT;
1351
1352                 return ret;
1353         }
1354
1355         return -EINVAL;
1356 }
1357
1358 /*
1359  * Map an iov into an array of pages and offset/length tupples. With the
1360  * partial_page structure, we can map several non-contiguous ranges into
1361  * our ones pages[] map instead of splitting that operation into pieces.
1362  * Could easily be exported as a generic helper for other users, in which
1363  * case one would probably want to add a 'max_nr_pages' parameter as well.
1364  */
1365 static int get_iovec_page_array(const struct iovec __user *iov,
1366                                 unsigned int nr_vecs, struct page **pages,
1367                                 struct partial_page *partial, int aligned)
1368 {
1369         int buffers = 0, error = 0;
1370
1371         while (nr_vecs) {
1372                 unsigned long off, npages;
1373                 struct iovec entry;
1374                 void __user *base;
1375                 size_t len;
1376                 int i;
1377
1378                 error = -EFAULT;
1379                 if (copy_from_user(&entry, iov, sizeof(entry)))
1380                         break;
1381
1382                 base = entry.iov_base;
1383                 len = entry.iov_len;
1384
1385                 /*
1386                  * Sanity check this iovec. 0 read succeeds.
1387                  */
1388                 error = 0;
1389                 if (unlikely(!len))
1390                         break;
1391                 error = -EFAULT;
1392                 if (!access_ok(VERIFY_READ, base, len))
1393                         break;
1394
1395                 /*
1396                  * Get this base offset and number of pages, then map
1397                  * in the user pages.
1398                  */
1399                 off = (unsigned long) base & ~PAGE_MASK;
1400
1401                 /*
1402                  * If asked for alignment, the offset must be zero and the
1403                  * length a multiple of the PAGE_SIZE.
1404                  */
1405                 error = -EINVAL;
1406                 if (aligned && (off || len & ~PAGE_MASK))
1407                         break;
1408
1409                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1410                 if (npages > PIPE_BUFFERS - buffers)
1411                         npages = PIPE_BUFFERS - buffers;
1412
1413                 error = get_user_pages_fast((unsigned long)base, npages,
1414                                         0, &pages[buffers]);
1415
1416                 if (unlikely(error <= 0))
1417                         break;
1418
1419                 /*
1420                  * Fill this contiguous range into the partial page map.
1421                  */
1422                 for (i = 0; i < error; i++) {
1423                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1424
1425                         partial[buffers].offset = off;
1426                         partial[buffers].len = plen;
1427
1428                         off = 0;
1429                         len -= plen;
1430                         buffers++;
1431                 }
1432
1433                 /*
1434                  * We didn't complete this iov, stop here since it probably
1435                  * means we have to move some of this into a pipe to
1436                  * be able to continue.
1437                  */
1438                 if (len)
1439                         break;
1440
1441                 /*
1442                  * Don't continue if we mapped fewer pages than we asked for,
1443                  * or if we mapped the max number of pages that we have
1444                  * room for.
1445                  */
1446                 if (error < npages || buffers == PIPE_BUFFERS)
1447                         break;
1448
1449                 nr_vecs--;
1450                 iov++;
1451         }
1452
1453         if (buffers)
1454                 return buffers;
1455
1456         return error;
1457 }
1458
1459 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1460                         struct splice_desc *sd)
1461 {
1462         char *src;
1463         int ret;
1464
1465         ret = buf->ops->confirm(pipe, buf);
1466         if (unlikely(ret))
1467                 return ret;
1468
1469         /*
1470          * See if we can use the atomic maps, by prefaulting in the
1471          * pages and doing an atomic copy
1472          */
1473         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1474                 src = buf->ops->map(pipe, buf, 1);
1475                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1476                                                         sd->len);
1477                 buf->ops->unmap(pipe, buf, src);
1478                 if (!ret) {
1479                         ret = sd->len;
1480                         goto out;
1481                 }
1482         }
1483
1484         /*
1485          * No dice, use slow non-atomic map and copy
1486          */
1487         src = buf->ops->map(pipe, buf, 0);
1488
1489         ret = sd->len;
1490         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1491                 ret = -EFAULT;
1492
1493         buf->ops->unmap(pipe, buf, src);
1494 out:
1495         if (ret > 0)
1496                 sd->u.userptr += ret;
1497         return ret;
1498 }
1499
1500 /*
1501  * For lack of a better implementation, implement vmsplice() to userspace
1502  * as a simple copy of the pipes pages to the user iov.
1503  */
1504 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1505                              unsigned long nr_segs, unsigned int flags)
1506 {
1507         struct pipe_inode_info *pipe;
1508         struct splice_desc sd;
1509         ssize_t size;
1510         int error;
1511         long ret;
1512
1513         pipe = pipe_info(file->f_path.dentry->d_inode);
1514         if (!pipe)
1515                 return -EBADF;
1516
1517         pipe_lock(pipe);
1518
1519         error = ret = 0;
1520         while (nr_segs) {
1521                 void __user *base;
1522                 size_t len;
1523
1524                 /*
1525                  * Get user address base and length for this iovec.
1526                  */
1527                 error = get_user(base, &iov->iov_base);
1528                 if (unlikely(error))
1529                         break;
1530                 error = get_user(len, &iov->iov_len);
1531                 if (unlikely(error))
1532                         break;
1533
1534                 /*
1535                  * Sanity check this iovec. 0 read succeeds.
1536                  */
1537                 if (unlikely(!len))
1538                         break;
1539                 if (unlikely(!base)) {
1540                         error = -EFAULT;
1541                         break;
1542                 }
1543
1544                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1545                         error = -EFAULT;
1546                         break;
1547                 }
1548
1549                 sd.len = 0;
1550                 sd.total_len = len;
1551                 sd.flags = flags;
1552                 sd.u.userptr = base;
1553                 sd.pos = 0;
1554
1555                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1556                 if (size < 0) {
1557                         if (!ret)
1558                                 ret = size;
1559
1560                         break;
1561                 }
1562
1563                 ret += size;
1564
1565                 if (size < len)
1566                         break;
1567
1568                 nr_segs--;
1569                 iov++;
1570         }
1571
1572         pipe_unlock(pipe);
1573
1574         if (!ret)
1575                 ret = error;
1576
1577         return ret;
1578 }
1579
1580 /*
1581  * vmsplice splices a user address range into a pipe. It can be thought of
1582  * as splice-from-memory, where the regular splice is splice-from-file (or
1583  * to file). In both cases the output is a pipe, naturally.
1584  */
1585 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1586                              unsigned long nr_segs, unsigned int flags)
1587 {
1588         struct pipe_inode_info *pipe;
1589         struct page *pages[PIPE_BUFFERS];
1590         struct partial_page partial[PIPE_BUFFERS];
1591         struct splice_pipe_desc spd = {
1592                 .pages = pages,
1593                 .partial = partial,
1594                 .flags = flags,
1595                 .ops = &user_page_pipe_buf_ops,
1596                 .spd_release = spd_release_page,
1597         };
1598
1599         pipe = pipe_info(file->f_path.dentry->d_inode);
1600         if (!pipe)
1601                 return -EBADF;
1602
1603         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1604                                             flags & SPLICE_F_GIFT);
1605         if (spd.nr_pages <= 0)
1606                 return spd.nr_pages;
1607
1608         return splice_to_pipe(pipe, &spd);
1609 }
1610
1611 /*
1612  * Note that vmsplice only really supports true splicing _from_ user memory
1613  * to a pipe, not the other way around. Splicing from user memory is a simple
1614  * operation that can be supported without any funky alignment restrictions
1615  * or nasty vm tricks. We simply map in the user memory and fill them into
1616  * a pipe. The reverse isn't quite as easy, though. There are two possible
1617  * solutions for that:
1618  *
1619  *      - memcpy() the data internally, at which point we might as well just
1620  *        do a regular read() on the buffer anyway.
1621  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1622  *        has restriction limitations on both ends of the pipe).
1623  *
1624  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1625  *
1626  */
1627 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1628                 unsigned long, nr_segs, unsigned int, flags)
1629 {
1630         struct file *file;
1631         long error;
1632         int fput;
1633
1634         if (unlikely(nr_segs > UIO_MAXIOV))
1635                 return -EINVAL;
1636         else if (unlikely(!nr_segs))
1637                 return 0;
1638
1639         error = -EBADF;
1640         file = fget_light(fd, &fput);
1641         if (file) {
1642                 if (file->f_mode & FMODE_WRITE)
1643                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1644                 else if (file->f_mode & FMODE_READ)
1645                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1646
1647                 fput_light(file, fput);
1648         }
1649
1650         return error;
1651 }
1652
1653 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1654                 int, fd_out, loff_t __user *, off_out,
1655                 size_t, len, unsigned int, flags)
1656 {
1657         long error;
1658         struct file *in, *out;
1659         int fput_in, fput_out;
1660
1661         if (unlikely(!len))
1662                 return 0;
1663
1664         error = -EBADF;
1665         in = fget_light(fd_in, &fput_in);
1666         if (in) {
1667                 if (in->f_mode & FMODE_READ) {
1668                         out = fget_light(fd_out, &fput_out);
1669                         if (out) {
1670                                 if (out->f_mode & FMODE_WRITE)
1671                                         error = do_splice(in, off_in,
1672                                                           out, off_out,
1673                                                           len, flags);
1674                                 fput_light(out, fput_out);
1675                         }
1676                 }
1677
1678                 fput_light(in, fput_in);
1679         }
1680
1681         return error;
1682 }
1683
1684 /*
1685  * Make sure there's data to read. Wait for input if we can, otherwise
1686  * return an appropriate error.
1687  */
1688 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1689 {
1690         int ret;
1691
1692         /*
1693          * Check ->nrbufs without the inode lock first. This function
1694          * is speculative anyways, so missing one is ok.
1695          */
1696         if (pipe->nrbufs)
1697                 return 0;
1698
1699         ret = 0;
1700         pipe_lock(pipe);
1701
1702         while (!pipe->nrbufs) {
1703                 if (signal_pending(current)) {
1704                         ret = -ERESTARTSYS;
1705                         break;
1706                 }
1707                 if (!pipe->writers)
1708                         break;
1709                 if (!pipe->waiting_writers) {
1710                         if (flags & SPLICE_F_NONBLOCK) {
1711                                 ret = -EAGAIN;
1712                                 break;
1713                         }
1714                 }
1715                 pipe_wait(pipe);
1716         }
1717
1718         pipe_unlock(pipe);
1719         return ret;
1720 }
1721
1722 /*
1723  * Make sure there's writeable room. Wait for room if we can, otherwise
1724  * return an appropriate error.
1725  */
1726 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1727 {
1728         int ret;
1729
1730         /*
1731          * Check ->nrbufs without the inode lock first. This function
1732          * is speculative anyways, so missing one is ok.
1733          */
1734         if (pipe->nrbufs < PIPE_BUFFERS)
1735                 return 0;
1736
1737         ret = 0;
1738         pipe_lock(pipe);
1739
1740         while (pipe->nrbufs >= PIPE_BUFFERS) {
1741                 if (!pipe->readers) {
1742                         send_sig(SIGPIPE, current, 0);
1743                         ret = -EPIPE;
1744                         break;
1745                 }
1746                 if (flags & SPLICE_F_NONBLOCK) {
1747                         ret = -EAGAIN;
1748                         break;
1749                 }
1750                 if (signal_pending(current)) {
1751                         ret = -ERESTARTSYS;
1752                         break;
1753                 }
1754                 pipe->waiting_writers++;
1755                 pipe_wait(pipe);
1756                 pipe->waiting_writers--;
1757         }
1758
1759         pipe_unlock(pipe);
1760         return ret;
1761 }
1762
1763 /*
1764  * Splice contents of ipipe to opipe.
1765  */
1766 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1767                                struct pipe_inode_info *opipe,
1768                                size_t len, unsigned int flags)
1769 {
1770         struct pipe_buffer *ibuf, *obuf;
1771         int ret = 0, nbuf;
1772         bool input_wakeup = false;
1773
1774
1775 retry:
1776         ret = ipipe_prep(ipipe, flags);
1777         if (ret)
1778                 return ret;
1779
1780         ret = opipe_prep(opipe, flags);
1781         if (ret)
1782                 return ret;
1783
1784         /*
1785          * Potential ABBA deadlock, work around it by ordering lock
1786          * grabbing by pipe info address. Otherwise two different processes
1787          * could deadlock (one doing tee from A -> B, the other from B -> A).
1788          */
1789         pipe_double_lock(ipipe, opipe);
1790
1791         do {
1792                 if (!opipe->readers) {
1793                         send_sig(SIGPIPE, current, 0);
1794                         if (!ret)
1795                                 ret = -EPIPE;
1796                         break;
1797                 }
1798
1799                 if (!ipipe->nrbufs && !ipipe->writers)
1800                         break;
1801
1802                 /*
1803                  * Cannot make any progress, because either the input
1804                  * pipe is empty or the output pipe is full.
1805                  */
1806                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1807                         /* Already processed some buffers, break */
1808                         if (ret)
1809                                 break;
1810
1811                         if (flags & SPLICE_F_NONBLOCK) {
1812                                 ret = -EAGAIN;
1813                                 break;
1814                         }
1815
1816                         /*
1817                          * We raced with another reader/writer and haven't
1818                          * managed to process any buffers.  A zero return
1819                          * value means EOF, so retry instead.
1820                          */
1821                         pipe_unlock(ipipe);
1822                         pipe_unlock(opipe);
1823                         goto retry;
1824                 }
1825
1826                 ibuf = ipipe->bufs + ipipe->curbuf;
1827                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1828                 obuf = opipe->bufs + nbuf;
1829
1830                 if (len >= ibuf->len) {
1831                         /*
1832                          * Simply move the whole buffer from ipipe to opipe
1833                          */
1834                         *obuf = *ibuf;
1835                         ibuf->ops = NULL;
1836                         opipe->nrbufs++;
1837                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1838                         ipipe->nrbufs--;
1839                         input_wakeup = true;
1840                 } else {
1841                         /*
1842                          * Get a reference to this pipe buffer,
1843                          * so we can copy the contents over.
1844                          */
1845                         ibuf->ops->get(ipipe, ibuf);
1846                         *obuf = *ibuf;
1847
1848                         /*
1849                          * Don't inherit the gift flag, we need to
1850                          * prevent multiple steals of this page.
1851                          */
1852                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1853
1854                         obuf->len = len;
1855                         opipe->nrbufs++;
1856                         ibuf->offset += obuf->len;
1857                         ibuf->len -= obuf->len;
1858                 }
1859                 ret += obuf->len;
1860                 len -= obuf->len;
1861         } while (len);
1862
1863         pipe_unlock(ipipe);
1864         pipe_unlock(opipe);
1865
1866         /*
1867          * If we put data in the output pipe, wakeup any potential readers.
1868          */
1869         if (ret > 0) {
1870                 smp_mb();
1871                 if (waitqueue_active(&opipe->wait))
1872                         wake_up_interruptible(&opipe->wait);
1873                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1874         }
1875         if (input_wakeup)
1876                 wakeup_pipe_writers(ipipe);
1877
1878         return ret;
1879 }
1880
1881 /*
1882  * Link contents of ipipe to opipe.
1883  */
1884 static int link_pipe(struct pipe_inode_info *ipipe,
1885                      struct pipe_inode_info *opipe,
1886                      size_t len, unsigned int flags)
1887 {
1888         struct pipe_buffer *ibuf, *obuf;
1889         int ret = 0, i = 0, nbuf;
1890
1891         /*
1892          * Potential ABBA deadlock, work around it by ordering lock
1893          * grabbing by pipe info address. Otherwise two different processes
1894          * could deadlock (one doing tee from A -> B, the other from B -> A).
1895          */
1896         pipe_double_lock(ipipe, opipe);
1897
1898         do {
1899                 if (!opipe->readers) {
1900                         send_sig(SIGPIPE, current, 0);
1901                         if (!ret)
1902                                 ret = -EPIPE;
1903                         break;
1904                 }
1905
1906                 /*
1907                  * If we have iterated all input buffers or ran out of
1908                  * output room, break.
1909                  */
1910                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1911                         break;
1912
1913                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1914                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1915
1916                 /*
1917                  * Get a reference to this pipe buffer,
1918                  * so we can copy the contents over.
1919                  */
1920                 ibuf->ops->get(ipipe, ibuf);
1921
1922                 obuf = opipe->bufs + nbuf;
1923                 *obuf = *ibuf;
1924
1925                 /*
1926                  * Don't inherit the gift flag, we need to
1927                  * prevent multiple steals of this page.
1928                  */
1929                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1930
1931                 if (obuf->len > len)
1932                         obuf->len = len;
1933
1934                 opipe->nrbufs++;
1935                 ret += obuf->len;
1936                 len -= obuf->len;
1937                 i++;
1938         } while (len);
1939
1940         /*
1941          * return EAGAIN if we have the potential of some data in the
1942          * future, otherwise just return 0
1943          */
1944         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1945                 ret = -EAGAIN;
1946
1947         pipe_unlock(ipipe);
1948         pipe_unlock(opipe);
1949
1950         /*
1951          * If we put data in the output pipe, wakeup any potential readers.
1952          */
1953         if (ret > 0) {
1954                 smp_mb();
1955                 if (waitqueue_active(&opipe->wait))
1956                         wake_up_interruptible(&opipe->wait);
1957                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1958         }
1959
1960         return ret;
1961 }
1962
1963 /*
1964  * This is a tee(1) implementation that works on pipes. It doesn't copy
1965  * any data, it simply references the 'in' pages on the 'out' pipe.
1966  * The 'flags' used are the SPLICE_F_* variants, currently the only
1967  * applicable one is SPLICE_F_NONBLOCK.
1968  */
1969 static long do_tee(struct file *in, struct file *out, size_t len,
1970                    unsigned int flags)
1971 {
1972         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1973         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1974         int ret = -EINVAL;
1975
1976         /*
1977          * Duplicate the contents of ipipe to opipe without actually
1978          * copying the data.
1979          */
1980         if (ipipe && opipe && ipipe != opipe) {
1981                 /*
1982                  * Keep going, unless we encounter an error. The ipipe/opipe
1983                  * ordering doesn't really matter.
1984                  */
1985                 ret = ipipe_prep(ipipe, flags);
1986                 if (!ret) {
1987                         ret = opipe_prep(opipe, flags);
1988                         if (!ret)
1989                                 ret = link_pipe(ipipe, opipe, len, flags);
1990                 }
1991         }
1992
1993         return ret;
1994 }
1995
1996 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1997 {
1998         struct file *in;
1999         int error, fput_in;
2000
2001         if (unlikely(!len))
2002                 return 0;
2003
2004         error = -EBADF;
2005         in = fget_light(fdin, &fput_in);
2006         if (in) {
2007                 if (in->f_mode & FMODE_READ) {
2008                         int fput_out;
2009                         struct file *out = fget_light(fdout, &fput_out);
2010
2011                         if (out) {
2012                                 if (out->f_mode & FMODE_WRITE)
2013                                         error = do_tee(in, out, len, flags);
2014                                 fput_light(out, fput_out);
2015                         }
2016                 }
2017                 fput_light(in, fput_in);
2018         }
2019
2020         return error;
2021 }