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/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34 #include <linux/compat.h>
  35 #include <linux/aio.h>
  36 #include "internal.h"
  37
  38 /*
  39  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  40  * a vm helper function, it's already simplified quite a bit by the
  41  * addition of remove_mapping(). If success is returned, the caller may
  42  * attempt to reuse this page for another destination.
  43  */
  44 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  45                                      struct pipe_buffer *buf)
  46 {
  47         struct page *page = buf->page;
  48         struct address_space *mapping;
  49
  50         lock_page(page);
  51
  52         mapping = page_mapping(page);
  53         if (mapping) {
  54                 WARN_ON(!PageUptodate(page));
  55
  56                 /*
  57                  * At least for ext2 with nobh option, we need to wait on
  58                  * writeback completing on this page, since we'll remove it
  59                  * from the pagecache.  Otherwise truncate wont wait on the
  60                  * page, allowing the disk blocks to be reused by someone else
  61                  * before we actually wrote our data to them. fs corruption
  62                  * ensues.
  63                  */
  64                 wait_on_page_writeback(page);
  65
  66                 if (page_has_private(page) &&
  67                     !try_to_release_page(page, GFP_KERNEL))
  68                         goto out_unlock;
  69
  70                 /*
  71                  * If we succeeded in removing the mapping, set LRU flag
  72                  * and return good.
  73                  */
  74                 if (remove_mapping(mapping, page)) {
  75                         buf->flags |= PIPE_BUF_FLAG_LRU;
  76                         return 0;
  77                 }
  78         }
  79
  80         /*
  81          * Raced with truncate or failed to remove page from current
  82          * address space, unlock and return failure.
  83          */
  84 out_unlock:
  85         unlock_page(page);
  86         return 1;
  87 }
  88
  89 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  90                                         struct pipe_buffer *buf)
  91 {
  92         page_cache_release(buf->page);
  93         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  94 }
  95
  96 /*
  97  * Check whether the contents of buf is OK to access. Since the content
  98  * is a page cache page, IO may be in flight.
  99  */
 100 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 101                                        struct pipe_buffer *buf)
 102 {
 103         struct page *page = buf->page;
 104         int err;
 105
 106         if (!PageUptodate(page)) {
 107                 lock_page(page);
 108
 109                 /*
 110                  * Page got truncated/unhashed. This will cause a 0-byte
 111                  * splice, if this is the first page.
 112                  */
 113                 if (!page->mapping) {
 114                         err = -ENODATA;
 115                         goto error;
 116                 }
 117
 118                 /*
 119                  * Uh oh, read-error from disk.
 120                  */
 121                 if (!PageUptodate(page)) {
 122                         err = -EIO;
 123                         goto error;
 124                 }
 125
 126                 /*
 127                  * Page is ok afterall, we are done.
 128                  */
 129                 unlock_page(page);
 130         }
 131
 132         return 0;
 133 error:
 134         unlock_page(page);
 135         return err;
 136 }
 137
 138 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 139         .can_merge = 0,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 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 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 165 {
 166         smp_mb();
 167         if (waitqueue_active(&pipe->wait))
 168                 wake_up_interruptible(&pipe->wait);
 169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 170 }
 171
 172 /**
 173  * splice_to_pipe - fill passed data into a pipe
 174  * @pipe:       pipe to fill
 175  * @spd:        data to fill
 176  *
 177  * Description:
 178  *    @spd contains a map of pages and len/offset tuples, along with
 179  *    the struct pipe_buf_operations associated with these pages. This
 180  *    function will link that data to the pipe.
 181  *
 182  */
 183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 184                        struct splice_pipe_desc *spd)
 185 {
 186         unsigned int spd_pages = spd->nr_pages;
 187         int ret, do_wakeup, page_nr;
 188
 189         ret = 0;
 190         do_wakeup = 0;
 191         page_nr = 0;
 192
 193         pipe_lock(pipe);
 194
 195         for (;;) {
 196                 if (!pipe->readers) {
 197                         send_sig(SIGPIPE, current, 0);
 198                         if (!ret)
 199                                 ret = -EPIPE;
 200                         break;
 201                 }
 202
 203                 if (pipe->nrbufs < pipe->buffers) {
 204                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 205                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 206
 207                         buf->page = spd->pages[page_nr];
 208                         buf->offset = spd->partial[page_nr].offset;
 209                         buf->len = spd->partial[page_nr].len;
 210                         buf->private = spd->partial[page_nr].private;
 211                         buf->ops = spd->ops;
 212                         if (spd->flags & SPLICE_F_GIFT)
 213                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 214
 215                         pipe->nrbufs++;
 216                         page_nr++;
 217                         ret += buf->len;
 218
 219                         if (pipe->files)
 220                                 do_wakeup = 1;
 221
 222                         if (!--spd->nr_pages)
 223                                 break;
 224                         if (pipe->nrbufs < pipe->buffers)
 225                                 continue;
 226
 227                         break;
 228                 }
 229
 230                 if (spd->flags & SPLICE_F_NONBLOCK) {
 231                         if (!ret)
 232                                 ret = -EAGAIN;
 233                         break;
 234                 }
 235
 236                 if (signal_pending(current)) {
 237                         if (!ret)
 238                                 ret = -ERESTARTSYS;
 239                         break;
 240                 }
 241
 242                 if (do_wakeup) {
 243                         smp_mb();
 244                         if (waitqueue_active(&pipe->wait))
 245                                 wake_up_interruptible_sync(&pipe->wait);
 246                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 247                         do_wakeup = 0;
 248                 }
 249
 250                 pipe->waiting_writers++;
 251                 pipe_wait(pipe);
 252                 pipe->waiting_writers--;
 253         }
 254
 255         pipe_unlock(pipe);
 256
 257         if (do_wakeup)
 258                 wakeup_pipe_readers(pipe);
 259
 260         while (page_nr < spd_pages)
 261                 spd->spd_release(spd, page_nr++);
 262
 263         return ret;
 264 }
 265
 266 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 267 {
 268         page_cache_release(spd->pages[i]);
 269 }
 270
 271 /*
 272  * Check if we need to grow the arrays holding pages and partial page
 273  * descriptions.
 274  */
 275 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 276 {
 277         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 278
 279         spd->nr_pages_max = buffers;
 280         if (buffers <= PIPE_DEF_BUFFERS)
 281                 return 0;
 282
 283         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 284         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 285
 286         if (spd->pages && spd->partial)
 287                 return 0;
 288
 289         kfree(spd->pages);
 290         kfree(spd->partial);
 291         return -ENOMEM;
 292 }
 293
 294 void splice_shrink_spd(struct splice_pipe_desc *spd)
 295 {
 296         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 297                 return;
 298
 299         kfree(spd->pages);
 300         kfree(spd->partial);
 301 }
 302
 303 static int
 304 __generic_file_splice_read(struct file *in, loff_t *ppos,
 305                            struct pipe_inode_info *pipe, size_t len,
 306                            unsigned int flags)
 307 {
 308         struct address_space *mapping = in->f_mapping;
 309         unsigned int loff, nr_pages, req_pages;
 310         struct page *pages[PIPE_DEF_BUFFERS];
 311         struct partial_page partial[PIPE_DEF_BUFFERS];
 312         struct page *page;
 313         pgoff_t index, end_index;
 314         loff_t isize;
 315         int error, page_nr;
 316         struct splice_pipe_desc spd = {
 317                 .pages = pages,
 318                 .partial = partial,
 319                 .nr_pages_max = PIPE_DEF_BUFFERS,
 320                 .flags = flags,
 321                 .ops = &page_cache_pipe_buf_ops,
 322                 .spd_release = spd_release_page,
 323         };
 324
 325         if (splice_grow_spd(pipe, &spd))
 326                 return -ENOMEM;
 327
 328         index = *ppos >> PAGE_CACHE_SHIFT;
 329         loff = *ppos & ~PAGE_CACHE_MASK;
 330         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 331         nr_pages = min(req_pages, spd.nr_pages_max);
 332
 333         /*
 334          * Lookup the (hopefully) full range of pages we need.
 335          */
 336         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 337         index += spd.nr_pages;
 338
 339         /*
 340          * If find_get_pages_contig() returned fewer pages than we needed,
 341          * readahead/allocate the rest and fill in the holes.
 342          */
 343         if (spd.nr_pages < nr_pages)
 344                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 345                                 index, req_pages - spd.nr_pages);
 346
 347         error = 0;
 348         while (spd.nr_pages < nr_pages) {
 349                 /*
 350                  * Page could be there, find_get_pages_contig() breaks on
 351                  * the first hole.
 352                  */
 353                 page = find_get_page(mapping, index);
 354                 if (!page) {
 355                         /*
 356                          * page didn't exist, allocate one.
 357                          */
 358                         page = page_cache_alloc_cold(mapping);
 359                         if (!page)
 360                                 break;
 361
 362                         error = add_to_page_cache_lru(page, mapping, index,
 363                                                 GFP_KERNEL);
 364                         if (unlikely(error)) {
 365                                 page_cache_release(page);
 366                                 if (error == -EEXIST)
 367                                         continue;
 368                                 break;
 369                         }
 370                         /*
 371                          * add_to_page_cache() locks the page, unlock it
 372                          * to avoid convoluting the logic below even more.
 373                          */
 374                         unlock_page(page);
 375                 }
 376
 377                 spd.pages[spd.nr_pages++] = page;
 378                 index++;
 379         }
 380
 381         /*
 382          * Now loop over the map and see if we need to start IO on any
 383          * pages, fill in the partial map, etc.
 384          */
 385         index = *ppos >> PAGE_CACHE_SHIFT;
 386         nr_pages = spd.nr_pages;
 387         spd.nr_pages = 0;
 388         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 389                 unsigned int this_len;
 390
 391                 if (!len)
 392                         break;
 393
 394                 /*
 395                  * this_len is the max we'll use from this page
 396                  */
 397                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 398                 page = spd.pages[page_nr];
 399
 400                 if (PageReadahead(page))
 401                         page_cache_async_readahead(mapping, &in->f_ra, in,
 402                                         page, index, req_pages - page_nr);
 403
 404                 /*
 405                  * If the page isn't uptodate, we may need to start io on it
 406                  */
 407                 if (!PageUptodate(page)) {
 408                         lock_page(page);
 409
 410                         /*
 411                          * Page was truncated, or invalidated by the
 412                          * filesystem.  Redo the find/create, but this time the
 413                          * page is kept locked, so there's no chance of another
 414                          * race with truncate/invalidate.
 415                          */
 416                         if (!page->mapping) {
 417                                 unlock_page(page);
 418                                 page = find_or_create_page(mapping, index,
 419                                                 mapping_gfp_mask(mapping));
 420
 421                                 if (!page) {
 422                                         error = -ENOMEM;
 423                                         break;
 424                                 }
 425                                 page_cache_release(spd.pages[page_nr]);
 426                                 spd.pages[page_nr] = page;
 427                         }
 428                         /*
 429                          * page was already under io and is now done, great
 430                          */
 431                         if (PageUptodate(page)) {
 432                                 unlock_page(page);
 433                                 goto fill_it;
 434                         }
 435
 436                         /*
 437                          * need to read in the page
 438                          */
 439                         error = mapping->a_ops->readpage(in, page);
 440                         if (unlikely(error)) {
 441                                 /*
 442                                  * We really should re-lookup the page here,
 443                                  * but it complicates things a lot. Instead
 444                                  * lets just do what we already stored, and
 445                                  * we'll get it the next time we are called.
 446                                  */
 447                                 if (error == AOP_TRUNCATED_PAGE)
 448                                         error = 0;
 449
 450                                 break;
 451                         }
 452                 }
 453 fill_it:
 454                 /*
 455                  * i_size must be checked after PageUptodate.
 456                  */
 457                 isize = i_size_read(mapping->host);
 458                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 459                 if (unlikely(!isize || index > end_index))
 460                         break;
 461
 462                 /*
 463                  * if this is the last page, see if we need to shrink
 464                  * the length and stop
 465                  */
 466                 if (end_index == index) {
 467                         unsigned int plen;
 468
 469                         /*
 470                          * max good bytes in this page
 471                          */
 472                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 473                         if (plen <= loff)
 474                                 break;
 475
 476                         /*
 477                          * force quit after adding this page
 478                          */
 479                         this_len = min(this_len, plen - loff);
 480                         len = this_len;
 481                 }
 482
 483                 spd.partial[page_nr].offset = loff;
 484                 spd.partial[page_nr].len = this_len;
 485                 len -= this_len;
 486                 loff = 0;
 487                 spd.nr_pages++;
 488                 index++;
 489         }
 490
 491         /*
 492          * Release any pages at the end, if we quit early. 'page_nr' is how far
 493          * we got, 'nr_pages' is how many pages are in the map.
 494          */
 495         while (page_nr < nr_pages)
 496                 page_cache_release(spd.pages[page_nr++]);
 497         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 498
 499         if (spd.nr_pages)
 500                 error = splice_to_pipe(pipe, &spd);
 501
 502         splice_shrink_spd(&spd);
 503         return error;
 504 }
 505
 506 /**
 507  * generic_file_splice_read - splice data from file to a pipe
 508  * @in:         file to splice from
 509  * @ppos:       position in @in
 510  * @pipe:       pipe to splice to
 511  * @len:        number of bytes to splice
 512  * @flags:      splice modifier flags
 513  *
 514  * Description:
 515  *    Will read pages from given file and fill them into a pipe. Can be
 516  *    used as long as the address_space operations for the source implements
 517  *    a readpage() hook.
 518  *
 519  */
 520 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 521                                  struct pipe_inode_info *pipe, size_t len,
 522                                  unsigned int flags)
 523 {
 524         loff_t isize, left;
 525         int ret;
 526
 527         isize = i_size_read(in->f_mapping->host);
 528         if (unlikely(*ppos >= isize))
 529                 return 0;
 530
 531         left = isize - *ppos;
 532         if (unlikely(left < len))
 533                 len = left;
 534
 535         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 536         if (ret > 0) {
 537                 *ppos += ret;
 538                 file_accessed(in);
 539         }
 540
 541         return ret;
 542 }
 543 EXPORT_SYMBOL(generic_file_splice_read);
 544
 545 static const struct pipe_buf_operations default_pipe_buf_ops = {
 546         .can_merge = 0,
 547         .confirm = generic_pipe_buf_confirm,
 548         .release = generic_pipe_buf_release,
 549         .steal = generic_pipe_buf_steal,
 550         .get = generic_pipe_buf_get,
 551 };
 552
 553 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 554                                     struct pipe_buffer *buf)
 555 {
 556         return 1;
 557 }
 558
 559 /* Pipe buffer operations for a socket and similar. */
 560 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 561         .can_merge = 0,
 562         .confirm = generic_pipe_buf_confirm,
 563         .release = generic_pipe_buf_release,
 564         .steal = generic_pipe_buf_nosteal,
 565         .get = generic_pipe_buf_get,
 566 };
 567 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 568
 569 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 570                             unsigned long vlen, loff_t offset)
 571 {
 572         mm_segment_t old_fs;
 573         loff_t pos = offset;
 574         ssize_t res;
 575
 576         old_fs = get_fs();
 577         set_fs(get_ds());
 578         /* The cast to a user pointer is valid due to the set_fs() */
 579         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 580         set_fs(old_fs);
 581
 582         return res;
 583 }
 584
 585 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 586                             loff_t pos)
 587 {
 588         mm_segment_t old_fs;
 589         ssize_t res;
 590
 591         old_fs = get_fs();
 592         set_fs(get_ds());
 593         /* The cast to a user pointer is valid due to the set_fs() */
 594         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 595         set_fs(old_fs);
 596
 597         return res;
 598 }
 599 EXPORT_SYMBOL(kernel_write);
 600
 601 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 602                                  struct pipe_inode_info *pipe, size_t len,
 603                                  unsigned int flags)
 604 {
 605         unsigned int nr_pages;
 606         unsigned int nr_freed;
 607         size_t offset;
 608         struct page *pages[PIPE_DEF_BUFFERS];
 609         struct partial_page partial[PIPE_DEF_BUFFERS];
 610         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 611         ssize_t res;
 612         size_t this_len;
 613         int error;
 614         int i;
 615         struct splice_pipe_desc spd = {
 616                 .pages = pages,
 617                 .partial = partial,
 618                 .nr_pages_max = PIPE_DEF_BUFFERS,
 619                 .flags = flags,
 620                 .ops = &default_pipe_buf_ops,
 621                 .spd_release = spd_release_page,
 622         };
 623
 624         if (splice_grow_spd(pipe, &spd))
 625                 return -ENOMEM;
 626
 627         res = -ENOMEM;
 628         vec = __vec;
 629         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 630                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 631                 if (!vec)
 632                         goto shrink_ret;
 633         }
 634
 635         offset = *ppos & ~PAGE_CACHE_MASK;
 636         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 637
 638         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 639                 struct page *page;
 640
 641                 page = alloc_page(GFP_USER);
 642                 error = -ENOMEM;
 643                 if (!page)
 644                         goto err;
 645
 646                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 647                 vec[i].iov_base = (void __user *) page_address(page);
 648                 vec[i].iov_len = this_len;
 649                 spd.pages[i] = page;
 650                 spd.nr_pages++;
 651                 len -= this_len;
 652                 offset = 0;
 653         }
 654
 655         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 656         if (res < 0) {
 657                 error = res;
 658                 goto err;
 659         }
 660
 661         error = 0;
 662         if (!res)
 663                 goto err;
 664
 665         nr_freed = 0;
 666         for (i = 0; i < spd.nr_pages; i++) {
 667                 this_len = min_t(size_t, vec[i].iov_len, res);
 668                 spd.partial[i].offset = 0;
 669                 spd.partial[i].len = this_len;
 670                 if (!this_len) {
 671                         __free_page(spd.pages[i]);
 672                         spd.pages[i] = NULL;
 673                         nr_freed++;
 674                 }
 675                 res -= this_len;
 676         }
 677         spd.nr_pages -= nr_freed;
 678
 679         res = splice_to_pipe(pipe, &spd);
 680         if (res > 0)
 681                 *ppos += res;
 682
 683 shrink_ret:
 684         if (vec != __vec)
 685                 kfree(vec);
 686         splice_shrink_spd(&spd);
 687         return res;
 688
 689 err:
 690         for (i = 0; i < spd.nr_pages; i++)
 691                 __free_page(spd.pages[i]);
 692
 693         res = error;
 694         goto shrink_ret;
 695 }
 696 EXPORT_SYMBOL(default_file_splice_read);
 697
 698 /*
 699  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 700  * using sendpage(). Return the number of bytes sent.
 701  */
 702 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 703                             struct pipe_buffer *buf, struct splice_desc *sd)
 704 {
 705         struct file *file = sd->u.file;
 706         loff_t pos = sd->pos;
 707         int more;
 708
 709         if (!likely(file->f_op->sendpage))
 710                 return -EINVAL;
 711
 712         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 713
 714         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 715                 more |= MSG_SENDPAGE_NOTLAST;
 716
 717         return file->f_op->sendpage(file, buf->page, buf->offset,
 718                                     sd->len, &pos, more);
 719 }
 720
 721 /*
 722  * This is a little more tricky than the file -> pipe splicing. There are
 723  * basically three cases:
 724  *
 725  *      - Destination page already exists in the address space and there
 726  *        are users of it. For that case we have no other option that
 727  *        copying the data. Tough luck.
 728  *      - Destination page already exists in the address space, but there
 729  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 730  *        through to last case.
 731  *      - Destination page does not exist, we can add the pipe page to
 732  *        the page cache and avoid the copy.
 733  *
 734  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 735  * sd->flags), we attempt to migrate pages from the pipe to the output
 736  * file address space page cache. This is possible if no one else has
 737  * the pipe page referenced outside of the pipe and page cache. If
 738  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 739  * a new page in the output file page cache and fill/dirty that.
 740  */
 741 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 742                  struct splice_desc *sd)
 743 {
 744         struct file *file = sd->u.file;
 745         struct address_space *mapping = file->f_mapping;
 746         unsigned int offset, this_len;
 747         struct page *page;
 748         void *fsdata;
 749         int ret;
 750
 751         offset = sd->pos & ~PAGE_CACHE_MASK;
 752
 753         this_len = sd->len;
 754         if (this_len + offset > PAGE_CACHE_SIZE)
 755                 this_len = PAGE_CACHE_SIZE - offset;
 756
 757         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 758                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 759         if (unlikely(ret))
 760                 goto out;
 761
 762         if (buf->page != page) {
 763                 char *src = kmap_atomic(buf->page);
 764                 char *dst = kmap_atomic(page);
 765
 766                 memcpy(dst + offset, src + buf->offset, this_len);
 767                 flush_dcache_page(page);
 768                 kunmap_atomic(dst);
 769                 kunmap_atomic(src);
 770         }
 771         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 772                                 page, fsdata);
 773 out:
 774         return ret;
 775 }
 776 EXPORT_SYMBOL(pipe_to_file);
 777
 778 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 779 {
 780         smp_mb();
 781         if (waitqueue_active(&pipe->wait))
 782                 wake_up_interruptible(&pipe->wait);
 783         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 784 }
 785
 786 /**
 787  * splice_from_pipe_feed - feed available data from a pipe to a file
 788  * @pipe:       pipe to splice from
 789  * @sd:         information to @actor
 790  * @actor:      handler that splices the data
 791  *
 792  * Description:
 793  *    This function loops over the pipe and calls @actor to do the
 794  *    actual moving of a single struct pipe_buffer to the desired
 795  *    destination.  It returns when there's no more buffers left in
 796  *    the pipe or if the requested number of bytes (@sd->total_len)
 797  *    have been copied.  It returns a positive number (one) if the
 798  *    pipe needs to be filled with more data, zero if the required
 799  *    number of bytes have been copied and -errno on error.
 800  *
 801  *    This, together with splice_from_pipe_{begin,end,next}, may be
 802  *    used to implement the functionality of __splice_from_pipe() when
 803  *    locking is required around copying the pipe buffers to the
 804  *    destination.
 805  */
 806 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 807                           splice_actor *actor)
 808 {
 809         int ret;
 810
 811         while (pipe->nrbufs) {
 812                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 813                 const struct pipe_buf_operations *ops = buf->ops;
 814
 815                 sd->len = buf->len;
 816                 if (sd->len > sd->total_len)
 817                         sd->len = sd->total_len;
 818
 819                 ret = buf->ops->confirm(pipe, buf);
 820                 if (unlikely(ret)) {
 821                         if (ret == -ENODATA)
 822                                 ret = 0;
 823                         return ret;
 824                 }
 825
 826                 ret = actor(pipe, buf, sd);
 827                 if (ret <= 0)
 828                         return ret;
 829
 830                 buf->offset += ret;
 831                 buf->len -= ret;
 832
 833                 sd->num_spliced += ret;
 834                 sd->len -= ret;
 835                 sd->pos += ret;
 836                 sd->total_len -= ret;
 837
 838                 if (!buf->len) {
 839                         buf->ops = NULL;
 840                         ops->release(pipe, buf);
 841                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 842                         pipe->nrbufs--;
 843                         if (pipe->files)
 844                                 sd->need_wakeup = true;
 845                 }
 846
 847                 if (!sd->total_len)
 848                         return 0;
 849         }
 850
 851         return 1;
 852 }
 853 EXPORT_SYMBOL(splice_from_pipe_feed);
 854
 855 /**
 856  * splice_from_pipe_next - wait for some data to splice from
 857  * @pipe:       pipe to splice from
 858  * @sd:         information about the splice operation
 859  *
 860  * Description:
 861  *    This function will wait for some data and return a positive
 862  *    value (one) if pipe buffers are available.  It will return zero
 863  *    or -errno if no more data needs to be spliced.
 864  */
 865 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 866 {
 867         while (!pipe->nrbufs) {
 868                 if (!pipe->writers)
 869                         return 0;
 870
 871                 if (!pipe->waiting_writers && sd->num_spliced)
 872                         return 0;
 873
 874                 if (sd->flags & SPLICE_F_NONBLOCK)
 875                         return -EAGAIN;
 876
 877                 if (signal_pending(current))
 878                         return -ERESTARTSYS;
 879
 880                 if (sd->need_wakeup) {
 881                         wakeup_pipe_writers(pipe);
 882                         sd->need_wakeup = false;
 883                 }
 884
 885                 pipe_wait(pipe);
 886         }
 887
 888         return 1;
 889 }
 890 EXPORT_SYMBOL(splice_from_pipe_next);
 891
 892 /**
 893  * splice_from_pipe_begin - start splicing from pipe
 894  * @sd:         information about the splice operation
 895  *
 896  * Description:
 897  *    This function should be called before a loop containing
 898  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 899  *    initialize the necessary fields of @sd.
 900  */
 901 void splice_from_pipe_begin(struct splice_desc *sd)
 902 {
 903         sd->num_spliced = 0;
 904         sd->need_wakeup = false;
 905 }
 906 EXPORT_SYMBOL(splice_from_pipe_begin);
 907
 908 /**
 909  * splice_from_pipe_end - finish splicing from pipe
 910  * @pipe:       pipe to splice from
 911  * @sd:         information about the splice operation
 912  *
 913  * Description:
 914  *    This function will wake up pipe writers if necessary.  It should
 915  *    be called after a loop containing splice_from_pipe_next() and
 916  *    splice_from_pipe_feed().
 917  */
 918 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 919 {
 920         if (sd->need_wakeup)
 921                 wakeup_pipe_writers(pipe);
 922 }
 923 EXPORT_SYMBOL(splice_from_pipe_end);
 924
 925 /**
 926  * __splice_from_pipe - splice data from a pipe to given actor
 927  * @pipe:       pipe to splice from
 928  * @sd:         information to @actor
 929  * @actor:      handler that splices the data
 930  *
 931  * Description:
 932  *    This function does little more than loop over the pipe and call
 933  *    @actor to do the actual moving of a single struct pipe_buffer to
 934  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 935  *    pipe_to_user.
 936  *
 937  */
 938 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 939                            splice_actor *actor)
 940 {
 941         int ret;
 942
 943         splice_from_pipe_begin(sd);
 944         do {
 945                 ret = splice_from_pipe_next(pipe, sd);
 946                 if (ret > 0)
 947                         ret = splice_from_pipe_feed(pipe, sd, actor);
 948         } while (ret > 0);
 949         splice_from_pipe_end(pipe, sd);
 950
 951         return sd->num_spliced ? sd->num_spliced : ret;
 952 }
 953 EXPORT_SYMBOL(__splice_from_pipe);
 954
 955 /**
 956  * splice_from_pipe - splice data from a pipe to a file
 957  * @pipe:       pipe to splice from
 958  * @out:        file to splice to
 959  * @ppos:       position in @out
 960  * @len:        how many bytes to splice
 961  * @flags:      splice modifier flags
 962  * @actor:      handler that splices the data
 963  *
 964  * Description:
 965  *    See __splice_from_pipe. This function locks the pipe inode,
 966  *    otherwise it's identical to __splice_from_pipe().
 967  *
 968  */
 969 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 970                          loff_t *ppos, size_t len, unsigned int flags,
 971                          splice_actor *actor)
 972 {
 973         ssize_t ret;
 974         struct splice_desc sd = {
 975                 .total_len = len,
 976                 .flags = flags,
 977                 .pos = *ppos,
 978                 .u.file = out,
 979         };
 980
 981         pipe_lock(pipe);
 982         ret = __splice_from_pipe(pipe, &sd, actor);
 983         pipe_unlock(pipe);
 984
 985         return ret;
 986 }
 987
 988 /**
 989  * generic_file_splice_write - splice data from a pipe to a file
 990  * @pipe:       pipe info
 991  * @out:        file to write to
 992  * @ppos:       position in @out
 993  * @len:        number of bytes to splice
 994  * @flags:      splice modifier flags
 995  *
 996  * Description:
 997  *    Will either move or copy pages (determined by @flags options) from
 998  *    the given pipe inode to the given file.
 999  *
1000  */
1001 ssize_t
1002 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1003                           loff_t *ppos, size_t len, unsigned int flags)
1004 {
1005         struct address_space *mapping = out->f_mapping;
1006         struct inode *inode = mapping->host;
1007         struct splice_desc sd = {
1008                 .total_len = len,
1009                 .flags = flags,
1010                 .pos = *ppos,
1011                 .u.file = out,
1012         };
1013         ssize_t ret;
1014
1015         pipe_lock(pipe);
1016
1017         splice_from_pipe_begin(&sd);
1018         do {
1019                 ret = splice_from_pipe_next(pipe, &sd);
1020                 if (ret <= 0)
1021                         break;
1022
1023                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1024                 ret = file_remove_suid(out);
1025                 if (!ret) {
1026                         ret = file_update_time(out);
1027                         if (!ret)
1028                                 ret = splice_from_pipe_feed(pipe, &sd,
1029                                                             pipe_to_file);
1030                 }
1031                 mutex_unlock(&inode->i_mutex);
1032         } while (ret > 0);
1033         splice_from_pipe_end(pipe, &sd);
1034
1035         pipe_unlock(pipe);
1036
1037         if (sd.num_spliced)
1038                 ret = sd.num_spliced;
1039
1040         if (ret > 0) {
1041                 int err;
1042
1043                 err = generic_write_sync(out, *ppos, ret);
1044                 if (err)
1045                         ret = err;
1046                 else
1047                         *ppos += ret;
1048                 balance_dirty_pages_ratelimited(mapping);
1049         }
1050
1051         return ret;
1052 }
1053
1054 EXPORT_SYMBOL(generic_file_splice_write);
1055
1056 /**
1057  * iter_file_splice_write - splice data from a pipe to a file
1058  * @pipe:       pipe info
1059  * @out:        file to write to
1060  * @ppos:       position in @out
1061  * @len:        number of bytes to splice
1062  * @flags:      splice modifier flags
1063  *
1064  * Description:
1065  *    Will either move or copy pages (determined by @flags options) from
1066  *    the given pipe inode to the given file.
1067  *    This one is ->write_iter-based.
1068  *
1069  */
1070 ssize_t
1071 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1072                           loff_t *ppos, size_t len, unsigned int flags)
1073 {
1074         struct splice_desc sd = {
1075                 .total_len = len,
1076                 .flags = flags,
1077                 .pos = *ppos,
1078                 .u.file = out,
1079         };
1080         int nbufs = pipe->buffers;
1081         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
1082                                         GFP_KERNEL);
1083         ssize_t ret;
1084
1085         if (unlikely(!array))
1086                 return -ENOMEM;
1087
1088         pipe_lock(pipe);
1089
1090         splice_from_pipe_begin(&sd);
1091         while (sd.total_len) {
1092                 struct iov_iter from;
1093                 struct kiocb kiocb;
1094                 size_t left;
1095                 int n, idx;
1096
1097                 ret = splice_from_pipe_next(pipe, &sd);
1098                 if (ret <= 0)
1099                         break;
1100
1101                 if (unlikely(nbufs < pipe->buffers)) {
1102                         kfree(array);
1103                         nbufs = pipe->buffers;
1104                         array = kcalloc(nbufs, sizeof(struct bio_vec),
1105                                         GFP_KERNEL);
1106                         if (!array) {
1107                                 ret = -ENOMEM;
1108                                 break;
1109                         }
1110                 }
1111
1112                 /* build the vector */
1113                 left = sd.total_len;
1114                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
1115                         struct pipe_buffer *buf = pipe->bufs + idx;
1116                         size_t this_len = buf->len;
1117
1118                         if (this_len > left)
1119                                 this_len = left;
1120
1121                         if (idx == pipe->buffers - 1)
1122                                 idx = -1;
1123
1124                         ret = buf->ops->confirm(pipe, buf);
1125                         if (unlikely(ret)) {
1126                                 if (ret == -ENODATA)
1127                                         ret = 0;
1128                                 goto done;
1129                         }
1130
1131                         array[n].bv_page = buf->page;
1132                         array[n].bv_len = this_len;
1133                         array[n].bv_offset = buf->offset;
1134                         left -= this_len;
1135                 }
1136
1137                 /* ... iov_iter */
1138                 from.type = ITER_BVEC | WRITE;
1139                 from.bvec = array;
1140                 from.nr_segs = n;
1141                 from.count = sd.total_len - left;
1142                 from.iov_offset = 0;
1143
1144                 /* ... and iocb */
1145                 init_sync_kiocb(&kiocb, out);
1146                 kiocb.ki_pos = sd.pos;
1147                 kiocb.ki_nbytes = sd.total_len - left;
1148
1149                 /* now, send it */
1150                 ret = out->f_op->write_iter(&kiocb, &from);
1151                 if (-EIOCBQUEUED == ret)
1152                         ret = wait_on_sync_kiocb(&kiocb);
1153
1154                 if (ret <= 0)
1155                         break;
1156
1157                 sd.num_spliced += ret;
1158                 sd.total_len -= ret;
1159                 *ppos = sd.pos = kiocb.ki_pos;
1160
1161                 /* dismiss the fully eaten buffers, adjust the partial one */
1162                 while (ret) {
1163                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1164                         if (ret >= buf->len) {
1165                                 const struct pipe_buf_operations *ops = buf->ops;
1166                                 ret -= buf->len;
1167                                 buf->len = 0;
1168                                 buf->ops = NULL;
1169                                 ops->release(pipe, buf);
1170                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1171                                 pipe->nrbufs--;
1172                                 if (pipe->files)
1173                                         sd.need_wakeup = true;
1174                         } else {
1175                                 buf->offset += ret;
1176                                 buf->len -= ret;
1177                                 ret = 0;
1178                         }
1179                 }
1180         }
1181 done:
1182         kfree(array);
1183         splice_from_pipe_end(pipe, &sd);
1184
1185         pipe_unlock(pipe);
1186
1187         if (sd.num_spliced)
1188                 ret = sd.num_spliced;
1189
1190         return ret;
1191 }
1192
1193 EXPORT_SYMBOL(iter_file_splice_write);
1194
1195 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1196                           struct splice_desc *sd)
1197 {
1198         int ret;
1199         void *data;
1200         loff_t tmp = sd->pos;
1201
1202         data = kmap(buf->page);
1203         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1204         kunmap(buf->page);
1205
1206         return ret;
1207 }
1208
1209 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1210                                          struct file *out, loff_t *ppos,
1211                                          size_t len, unsigned int flags)
1212 {
1213         ssize_t ret;
1214
1215         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1216         if (ret > 0)
1217                 *ppos += ret;
1218
1219         return ret;
1220 }
1221
1222 /**
1223  * generic_splice_sendpage - splice data from a pipe to a socket
1224  * @pipe:       pipe to splice from
1225  * @out:        socket to write to
1226  * @ppos:       position in @out
1227  * @len:        number of bytes to splice
1228  * @flags:      splice modifier flags
1229  *
1230  * Description:
1231  *    Will send @len bytes from the pipe to a network socket. No data copying
1232  *    is involved.
1233  *
1234  */
1235 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1236                                 loff_t *ppos, size_t len, unsigned int flags)
1237 {
1238         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1239 }
1240
1241 EXPORT_SYMBOL(generic_splice_sendpage);
1242
1243 /*
1244  * Attempt to initiate a splice from pipe to file.
1245  */
1246 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1247                            loff_t *ppos, size_t len, unsigned int flags)
1248 {
1249         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1250                                 loff_t *, size_t, unsigned int);
1251
1252         if (out->f_op->splice_write)
1253                 splice_write = out->f_op->splice_write;
1254         else
1255                 splice_write = default_file_splice_write;
1256
1257         return splice_write(pipe, out, ppos, len, flags);
1258 }
1259
1260 /*
1261  * Attempt to initiate a splice from a file to a pipe.
1262  */
1263 static long do_splice_to(struct file *in, loff_t *ppos,
1264                          struct pipe_inode_info *pipe, size_t len,
1265                          unsigned int flags)
1266 {
1267         ssize_t (*splice_read)(struct file *, loff_t *,
1268                                struct pipe_inode_info *, size_t, unsigned int);
1269         int ret;
1270
1271         if (unlikely(!(in->f_mode & FMODE_READ)))
1272                 return -EBADF;
1273
1274         ret = rw_verify_area(READ, in, ppos, len);
1275         if (unlikely(ret < 0))
1276                 return ret;
1277
1278         if (in->f_op->splice_read)
1279                 splice_read = in->f_op->splice_read;
1280         else
1281                 splice_read = default_file_splice_read;
1282
1283         return splice_read(in, ppos, pipe, len, flags);
1284 }
1285
1286 /**
1287  * splice_direct_to_actor - splices data directly between two non-pipes
1288  * @in:         file to splice from
1289  * @sd:         actor information on where to splice to
1290  * @actor:      handles the data splicing
1291  *
1292  * Description:
1293  *    This is a special case helper to splice directly between two
1294  *    points, without requiring an explicit pipe. Internally an allocated
1295  *    pipe is cached in the process, and reused during the lifetime of
1296  *    that process.
1297  *
1298  */
1299 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1300                                splice_direct_actor *actor)
1301 {
1302         struct pipe_inode_info *pipe;
1303         long ret, bytes;
1304         umode_t i_mode;
1305         size_t len;
1306         int i, flags;
1307
1308         /*
1309          * We require the input being a regular file, as we don't want to
1310          * randomly drop data for eg socket -> socket splicing. Use the
1311          * piped splicing for that!
1312          */
1313         i_mode = file_inode(in)->i_mode;
1314         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1315                 return -EINVAL;
1316
1317         /*
1318          * neither in nor out is a pipe, setup an internal pipe attached to
1319          * 'out' and transfer the wanted data from 'in' to 'out' through that
1320          */
1321         pipe = current->splice_pipe;
1322         if (unlikely(!pipe)) {
1323                 pipe = alloc_pipe_info();
1324                 if (!pipe)
1325                         return -ENOMEM;
1326
1327                 /*
1328                  * We don't have an immediate reader, but we'll read the stuff
1329                  * out of the pipe right after the splice_to_pipe(). So set
1330                  * PIPE_READERS appropriately.
1331                  */
1332                 pipe->readers = 1;
1333
1334                 current->splice_pipe = pipe;
1335         }
1336
1337         /*
1338          * Do the splice.
1339          */
1340         ret = 0;
1341         bytes = 0;
1342         len = sd->total_len;
1343         flags = sd->flags;
1344
1345         /*
1346          * Don't block on output, we have to drain the direct pipe.
1347          */
1348         sd->flags &= ~SPLICE_F_NONBLOCK;
1349
1350         while (len) {
1351                 size_t read_len;
1352                 loff_t pos = sd->pos, prev_pos = pos;
1353
1354                 ret = do_splice_to(in, &pos, pipe, len, flags);
1355                 if (unlikely(ret <= 0))
1356                         goto out_release;
1357
1358                 read_len = ret;
1359                 sd->total_len = read_len;
1360
1361                 /*
1362                  * NOTE: nonblocking mode only applies to the input. We
1363                  * must not do the output in nonblocking mode as then we
1364                  * could get stuck data in the internal pipe:
1365                  */
1366                 ret = actor(pipe, sd);
1367                 if (unlikely(ret <= 0)) {
1368                         sd->pos = prev_pos;
1369                         goto out_release;
1370                 }
1371
1372                 bytes += ret;
1373                 len -= ret;
1374                 sd->pos = pos;
1375
1376                 if (ret < read_len) {
1377                         sd->pos = prev_pos + ret;
1378                         goto out_release;
1379                 }
1380         }
1381
1382 done:
1383         pipe->nrbufs = pipe->curbuf = 0;
1384         file_accessed(in);
1385         return bytes;
1386
1387 out_release:
1388         /*
1389          * If we did an incomplete transfer we must release
1390          * the pipe buffers in question:
1391          */
1392         for (i = 0; i < pipe->buffers; i++) {
1393                 struct pipe_buffer *buf = pipe->bufs + i;
1394
1395                 if (buf->ops) {
1396                         buf->ops->release(pipe, buf);
1397                         buf->ops = NULL;
1398                 }
1399         }
1400
1401         if (!bytes)
1402                 bytes = ret;
1403
1404         goto done;
1405 }
1406 EXPORT_SYMBOL(splice_direct_to_actor);
1407
1408 static int direct_splice_actor(struct pipe_inode_info *pipe,
1409                                struct splice_desc *sd)
1410 {
1411         struct file *file = sd->u.file;
1412
1413         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1414                               sd->flags);
1415 }
1416
1417 /**
1418  * do_splice_direct - splices data directly between two files
1419  * @in:         file to splice from
1420  * @ppos:       input file offset
1421  * @out:        file to splice to
1422  * @opos:       output file offset
1423  * @len:        number of bytes to splice
1424  * @flags:      splice modifier flags
1425  *
1426  * Description:
1427  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1428  *    doing it in the application would incur an extra system call
1429  *    (splice in + splice out, as compared to just sendfile()). So this helper
1430  *    can splice directly through a process-private pipe.
1431  *
1432  */
1433 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1434                       loff_t *opos, size_t len, unsigned int flags)
1435 {
1436         struct splice_desc sd = {
1437                 .len            = len,
1438                 .total_len      = len,
1439                 .flags          = flags,
1440                 .pos            = *ppos,
1441                 .u.file         = out,
1442                 .opos           = opos,
1443         };
1444         long ret;
1445
1446         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1447                 return -EBADF;
1448
1449         if (unlikely(out->f_flags & O_APPEND))
1450                 return -EINVAL;
1451
1452         ret = rw_verify_area(WRITE, out, opos, len);
1453         if (unlikely(ret < 0))
1454                 return ret;
1455
1456         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1457         if (ret > 0)
1458                 *ppos = sd.pos;
1459
1460         return ret;
1461 }
1462
1463 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1464                                struct pipe_inode_info *opipe,
1465                                size_t len, unsigned int flags);
1466
1467 /*
1468  * Determine where to splice to/from.
1469  */
1470 static long do_splice(struct file *in, loff_t __user *off_in,
1471                       struct file *out, loff_t __user *off_out,
1472                       size_t len, unsigned int flags)
1473 {
1474         struct pipe_inode_info *ipipe;
1475         struct pipe_inode_info *opipe;
1476         loff_t offset;
1477         long ret;
1478
1479         ipipe = get_pipe_info(in);
1480         opipe = get_pipe_info(out);
1481
1482         if (ipipe && opipe) {
1483                 if (off_in || off_out)
1484                         return -ESPIPE;
1485
1486                 if (!(in->f_mode & FMODE_READ))
1487                         return -EBADF;
1488
1489                 if (!(out->f_mode & FMODE_WRITE))
1490                         return -EBADF;
1491
1492                 /* Splicing to self would be fun, but... */
1493                 if (ipipe == opipe)
1494                         return -EINVAL;
1495
1496                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1497         }
1498
1499         if (ipipe) {
1500                 if (off_in)
1501                         return -ESPIPE;
1502                 if (off_out) {
1503                         if (!(out->f_mode & FMODE_PWRITE))
1504                                 return -EINVAL;
1505                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1506                                 return -EFAULT;
1507                 } else {
1508                         offset = out->f_pos;
1509                 }
1510
1511                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1512                         return -EBADF;
1513
1514                 if (unlikely(out->f_flags & O_APPEND))
1515                         return -EINVAL;
1516
1517                 ret = rw_verify_area(WRITE, out, &offset, len);
1518                 if (unlikely(ret < 0))
1519                         return ret;
1520
1521                 file_start_write(out);
1522                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1523                 file_end_write(out);
1524
1525                 if (!off_out)
1526                         out->f_pos = offset;
1527                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1528                         ret = -EFAULT;
1529
1530                 return ret;
1531         }
1532
1533         if (opipe) {
1534                 if (off_out)
1535                         return -ESPIPE;
1536                 if (off_in) {
1537                         if (!(in->f_mode & FMODE_PREAD))
1538                                 return -EINVAL;
1539                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1540                                 return -EFAULT;
1541                 } else {
1542                         offset = in->f_pos;
1543                 }
1544
1545                 ret = do_splice_to(in, &offset, opipe, len, flags);
1546
1547                 if (!off_in)
1548                         in->f_pos = offset;
1549                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1550                         ret = -EFAULT;
1551
1552                 return ret;
1553         }
1554
1555         return -EINVAL;
1556 }
1557
1558 /*
1559  * Map an iov into an array of pages and offset/length tupples. With the
1560  * partial_page structure, we can map several non-contiguous ranges into
1561  * our ones pages[] map instead of splitting that operation into pieces.
1562  * Could easily be exported as a generic helper for other users, in which
1563  * case one would probably want to add a 'max_nr_pages' parameter as well.
1564  */
1565 static int get_iovec_page_array(const struct iovec __user *iov,
1566                                 unsigned int nr_vecs, struct page **pages,
1567                                 struct partial_page *partial, bool aligned,
1568                                 unsigned int pipe_buffers)
1569 {
1570         int buffers = 0, error = 0;
1571
1572         while (nr_vecs) {
1573                 unsigned long off, npages;
1574                 struct iovec entry;
1575                 void __user *base;
1576                 size_t len;
1577                 int i;
1578
1579                 error = -EFAULT;
1580                 if (copy_from_user(&entry, iov, sizeof(entry)))
1581                         break;
1582
1583                 base = entry.iov_base;
1584                 len = entry.iov_len;
1585
1586                 /*
1587                  * Sanity check this iovec. 0 read succeeds.
1588                  */
1589                 error = 0;
1590                 if (unlikely(!len))
1591                         break;
1592                 error = -EFAULT;
1593                 if (!access_ok(VERIFY_READ, base, len))
1594                         break;
1595
1596                 /*
1597                  * Get this base offset and number of pages, then map
1598                  * in the user pages.
1599                  */
1600                 off = (unsigned long) base & ~PAGE_MASK;
1601
1602                 /*
1603                  * If asked for alignment, the offset must be zero and the
1604                  * length a multiple of the PAGE_SIZE.
1605                  */
1606                 error = -EINVAL;
1607                 if (aligned && (off || len & ~PAGE_MASK))
1608                         break;
1609
1610                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1611                 if (npages > pipe_buffers - buffers)
1612                         npages = pipe_buffers - buffers;
1613
1614                 error = get_user_pages_fast((unsigned long)base, npages,
1615                                         0, &pages[buffers]);
1616
1617                 if (unlikely(error <= 0))
1618                         break;
1619
1620                 /*
1621                  * Fill this contiguous range into the partial page map.
1622                  */
1623                 for (i = 0; i < error; i++) {
1624                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1625
1626                         partial[buffers].offset = off;
1627                         partial[buffers].len = plen;
1628
1629                         off = 0;
1630                         len -= plen;
1631                         buffers++;
1632                 }
1633
1634                 /*
1635                  * We didn't complete this iov, stop here since it probably
1636                  * means we have to move some of this into a pipe to
1637                  * be able to continue.
1638                  */
1639                 if (len)
1640                         break;
1641
1642                 /*
1643                  * Don't continue if we mapped fewer pages than we asked for,
1644                  * or if we mapped the max number of pages that we have
1645                  * room for.
1646                  */
1647                 if (error < npages || buffers == pipe_buffers)
1648                         break;
1649
1650                 nr_vecs--;
1651                 iov++;
1652         }
1653
1654         if (buffers)
1655                 return buffers;
1656
1657         return error;
1658 }
1659
1660 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1661                         struct splice_desc *sd)
1662 {
1663         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1664         return n == sd->len ? n : -EFAULT;
1665 }
1666
1667 /*
1668  * For lack of a better implementation, implement vmsplice() to userspace
1669  * as a simple copy of the pipes pages to the user iov.
1670  */
1671 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1672                              unsigned long nr_segs, unsigned int flags)
1673 {
1674         struct pipe_inode_info *pipe;
1675         struct splice_desc sd;
1676         long ret;
1677         struct iovec iovstack[UIO_FASTIOV];
1678         struct iovec *iov = iovstack;
1679         struct iov_iter iter;
1680         ssize_t count = 0;
1681
1682         pipe = get_pipe_info(file);
1683         if (!pipe)
1684                 return -EBADF;
1685
1686         ret = rw_copy_check_uvector(READ, uiov, nr_segs,
1687                                     ARRAY_SIZE(iovstack), iovstack, &iov);
1688         if (ret <= 0)
1689                 return ret;
1690
1691         iov_iter_init(&iter, READ, iov, nr_segs, count);
1692
1693         sd.len = 0;
1694         sd.total_len = count;
1695         sd.flags = flags;
1696         sd.u.data = &iter;
1697         sd.pos = 0;
1698
1699         pipe_lock(pipe);
1700         ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1701         pipe_unlock(pipe);
1702
1703         if (iov != iovstack)
1704                 kfree(iov);
1705
1706         return ret;
1707 }
1708
1709 /*
1710  * vmsplice splices a user address range into a pipe. It can be thought of
1711  * as splice-from-memory, where the regular splice is splice-from-file (or
1712  * to file). In both cases the output is a pipe, naturally.
1713  */
1714 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1715                              unsigned long nr_segs, unsigned int flags)
1716 {
1717         struct pipe_inode_info *pipe;
1718         struct page *pages[PIPE_DEF_BUFFERS];
1719         struct partial_page partial[PIPE_DEF_BUFFERS];
1720         struct splice_pipe_desc spd = {
1721                 .pages = pages,
1722                 .partial = partial,
1723                 .nr_pages_max = PIPE_DEF_BUFFERS,
1724                 .flags = flags,
1725                 .ops = &user_page_pipe_buf_ops,
1726                 .spd_release = spd_release_page,
1727         };
1728         long ret;
1729
1730         pipe = get_pipe_info(file);
1731         if (!pipe)
1732                 return -EBADF;
1733
1734         if (splice_grow_spd(pipe, &spd))
1735                 return -ENOMEM;
1736
1737         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1738                                             spd.partial, false,
1739                                             spd.nr_pages_max);
1740         if (spd.nr_pages <= 0)
1741                 ret = spd.nr_pages;
1742         else
1743                 ret = splice_to_pipe(pipe, &spd);
1744
1745         splice_shrink_spd(&spd);
1746         return ret;
1747 }
1748
1749 /*
1750  * Note that vmsplice only really supports true splicing _from_ user memory
1751  * to a pipe, not the other way around. Splicing from user memory is a simple
1752  * operation that can be supported without any funky alignment restrictions
1753  * or nasty vm tricks. We simply map in the user memory and fill them into
1754  * a pipe. The reverse isn't quite as easy, though. There are two possible
1755  * solutions for that:
1756  *
1757  *      - memcpy() the data internally, at which point we might as well just
1758  *        do a regular read() on the buffer anyway.
1759  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1760  *        has restriction limitations on both ends of the pipe).
1761  *
1762  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1763  *
1764  */
1765 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1766                 unsigned long, nr_segs, unsigned int, flags)
1767 {
1768         struct fd f;
1769         long error;
1770
1771         if (unlikely(nr_segs > UIO_MAXIOV))
1772                 return -EINVAL;
1773         else if (unlikely(!nr_segs))
1774                 return 0;
1775
1776         error = -EBADF;
1777         f = fdget(fd);
1778         if (f.file) {
1779                 if (f.file->f_mode & FMODE_WRITE)
1780                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1781                 else if (f.file->f_mode & FMODE_READ)
1782                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1783
1784                 fdput(f);
1785         }
1786
1787         return error;
1788 }
1789
1790 #ifdef CONFIG_COMPAT
1791 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1792                     unsigned int, nr_segs, unsigned int, flags)
1793 {
1794         unsigned i;
1795         struct iovec __user *iov;
1796         if (nr_segs > UIO_MAXIOV)
1797                 return -EINVAL;
1798         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1799         for (i = 0; i < nr_segs; i++) {
1800                 struct compat_iovec v;
1801                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1802                     get_user(v.iov_len, &iov32[i].iov_len) ||
1803                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1804                     put_user(v.iov_len, &iov[i].iov_len))
1805                         return -EFAULT;
1806         }
1807         return sys_vmsplice(fd, iov, nr_segs, flags);
1808 }
1809 #endif
1810
1811 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1812                 int, fd_out, loff_t __user *, off_out,
1813                 size_t, len, unsigned int, flags)
1814 {
1815         struct fd in, out;
1816         long error;
1817
1818         if (unlikely(!len))
1819                 return 0;
1820
1821         error = -EBADF;
1822         in = fdget(fd_in);
1823         if (in.file) {
1824                 if (in.file->f_mode & FMODE_READ) {
1825                         out = fdget(fd_out);
1826                         if (out.file) {
1827                                 if (out.file->f_mode & FMODE_WRITE)
1828                                         error = do_splice(in.file, off_in,
1829                                                           out.file, off_out,
1830                                                           len, flags);
1831                                 fdput(out);
1832                         }
1833                 }
1834                 fdput(in);
1835         }
1836         return error;
1837 }
1838
1839 /*
1840  * Make sure there's data to read. Wait for input if we can, otherwise
1841  * return an appropriate error.
1842  */
1843 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1844 {
1845         int ret;
1846
1847         /*
1848          * Check ->nrbufs without the inode lock first. This function
1849          * is speculative anyways, so missing one is ok.
1850          */
1851         if (pipe->nrbufs)
1852                 return 0;
1853
1854         ret = 0;
1855         pipe_lock(pipe);
1856
1857         while (!pipe->nrbufs) {
1858                 if (signal_pending(current)) {
1859                         ret = -ERESTARTSYS;
1860                         break;
1861                 }
1862                 if (!pipe->writers)
1863                         break;
1864                 if (!pipe->waiting_writers) {
1865                         if (flags & SPLICE_F_NONBLOCK) {
1866                                 ret = -EAGAIN;
1867                                 break;
1868                         }
1869                 }
1870                 pipe_wait(pipe);
1871         }
1872
1873         pipe_unlock(pipe);
1874         return ret;
1875 }
1876
1877 /*
1878  * Make sure there's writeable room. Wait for room if we can, otherwise
1879  * return an appropriate error.
1880  */
1881 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1882 {
1883         int ret;
1884
1885         /*
1886          * Check ->nrbufs without the inode lock first. This function
1887          * is speculative anyways, so missing one is ok.
1888          */
1889         if (pipe->nrbufs < pipe->buffers)
1890                 return 0;
1891
1892         ret = 0;
1893         pipe_lock(pipe);
1894
1895         while (pipe->nrbufs >= pipe->buffers) {
1896                 if (!pipe->readers) {
1897                         send_sig(SIGPIPE, current, 0);
1898                         ret = -EPIPE;
1899                         break;
1900                 }
1901                 if (flags & SPLICE_F_NONBLOCK) {
1902                         ret = -EAGAIN;
1903                         break;
1904                 }
1905                 if (signal_pending(current)) {
1906                         ret = -ERESTARTSYS;
1907                         break;
1908                 }
1909                 pipe->waiting_writers++;
1910                 pipe_wait(pipe);
1911                 pipe->waiting_writers--;
1912         }
1913
1914         pipe_unlock(pipe);
1915         return ret;
1916 }
1917
1918 /*
1919  * Splice contents of ipipe to opipe.
1920  */
1921 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1922                                struct pipe_inode_info *opipe,
1923                                size_t len, unsigned int flags)
1924 {
1925         struct pipe_buffer *ibuf, *obuf;
1926         int ret = 0, nbuf;
1927         bool input_wakeup = false;
1928
1929
1930 retry:
1931         ret = ipipe_prep(ipipe, flags);
1932         if (ret)
1933                 return ret;
1934
1935         ret = opipe_prep(opipe, flags);
1936         if (ret)
1937                 return ret;
1938
1939         /*
1940          * Potential ABBA deadlock, work around it by ordering lock
1941          * grabbing by pipe info address. Otherwise two different processes
1942          * could deadlock (one doing tee from A -> B, the other from B -> A).
1943          */
1944         pipe_double_lock(ipipe, opipe);
1945
1946         do {
1947                 if (!opipe->readers) {
1948                         send_sig(SIGPIPE, current, 0);
1949                         if (!ret)
1950                                 ret = -EPIPE;
1951                         break;
1952                 }
1953
1954                 if (!ipipe->nrbufs && !ipipe->writers)
1955                         break;
1956
1957                 /*
1958                  * Cannot make any progress, because either the input
1959                  * pipe is empty or the output pipe is full.
1960                  */
1961                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1962                         /* Already processed some buffers, break */
1963                         if (ret)
1964                                 break;
1965
1966                         if (flags & SPLICE_F_NONBLOCK) {
1967                                 ret = -EAGAIN;
1968                                 break;
1969                         }
1970
1971                         /*
1972                          * We raced with another reader/writer and haven't
1973                          * managed to process any buffers.  A zero return
1974                          * value means EOF, so retry instead.
1975                          */
1976                         pipe_unlock(ipipe);
1977                         pipe_unlock(opipe);
1978                         goto retry;
1979                 }
1980
1981                 ibuf = ipipe->bufs + ipipe->curbuf;
1982                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1983                 obuf = opipe->bufs + nbuf;
1984
1985                 if (len >= ibuf->len) {
1986                         /*
1987                          * Simply move the whole buffer from ipipe to opipe
1988                          */
1989                         *obuf = *ibuf;
1990                         ibuf->ops = NULL;
1991                         opipe->nrbufs++;
1992                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1993                         ipipe->nrbufs--;
1994                         input_wakeup = true;
1995                 } else {
1996                         /*
1997                          * Get a reference to this pipe buffer,
1998                          * so we can copy the contents over.
1999                          */
2000                         ibuf->ops->get(ipipe, ibuf);
2001                         *obuf = *ibuf;
2002
2003                         /*
2004                          * Don't inherit the gift flag, we need to
2005                          * prevent multiple steals of this page.
2006                          */
2007                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2008
2009                         obuf->len = len;
2010                         opipe->nrbufs++;
2011                         ibuf->offset += obuf->len;
2012                         ibuf->len -= obuf->len;
2013                 }
2014                 ret += obuf->len;
2015                 len -= obuf->len;
2016         } while (len);
2017
2018         pipe_unlock(ipipe);
2019         pipe_unlock(opipe);
2020
2021         /*
2022          * If we put data in the output pipe, wakeup any potential readers.
2023          */
2024         if (ret > 0)
2025                 wakeup_pipe_readers(opipe);
2026
2027         if (input_wakeup)
2028                 wakeup_pipe_writers(ipipe);
2029
2030         return ret;
2031 }
2032
2033 /*
2034  * Link contents of ipipe to opipe.
2035  */
2036 static int link_pipe(struct pipe_inode_info *ipipe,
2037                      struct pipe_inode_info *opipe,
2038                      size_t len, unsigned int flags)
2039 {
2040         struct pipe_buffer *ibuf, *obuf;
2041         int ret = 0, i = 0, nbuf;
2042
2043         /*
2044          * Potential ABBA deadlock, work around it by ordering lock
2045          * grabbing by pipe info address. Otherwise two different processes
2046          * could deadlock (one doing tee from A -> B, the other from B -> A).
2047          */
2048         pipe_double_lock(ipipe, opipe);
2049
2050         do {
2051                 if (!opipe->readers) {
2052                         send_sig(SIGPIPE, current, 0);
2053                         if (!ret)
2054                                 ret = -EPIPE;
2055                         break;
2056                 }
2057
2058                 /*
2059                  * If we have iterated all input buffers or ran out of
2060                  * output room, break.
2061                  */
2062                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
2063                         break;
2064
2065                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
2066                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2067
2068                 /*
2069                  * Get a reference to this pipe buffer,
2070                  * so we can copy the contents over.
2071                  */
2072                 ibuf->ops->get(ipipe, ibuf);
2073
2074                 obuf = opipe->bufs + nbuf;
2075                 *obuf = *ibuf;
2076
2077                 /*
2078                  * Don't inherit the gift flag, we need to
2079                  * prevent multiple steals of this page.
2080                  */
2081                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2082
2083                 if (obuf->len > len)
2084                         obuf->len = len;
2085
2086                 opipe->nrbufs++;
2087                 ret += obuf->len;
2088                 len -= obuf->len;
2089                 i++;
2090         } while (len);
2091
2092         /*
2093          * return EAGAIN if we have the potential of some data in the
2094          * future, otherwise just return 0
2095          */
2096         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2097                 ret = -EAGAIN;
2098
2099         pipe_unlock(ipipe);
2100         pipe_unlock(opipe);
2101
2102         /*
2103          * If we put data in the output pipe, wakeup any potential readers.
2104          */
2105         if (ret > 0)
2106                 wakeup_pipe_readers(opipe);
2107
2108         return ret;
2109 }
2110
2111 /*
2112  * This is a tee(1) implementation that works on pipes. It doesn't copy
2113  * any data, it simply references the 'in' pages on the 'out' pipe.
2114  * The 'flags' used are the SPLICE_F_* variants, currently the only
2115  * applicable one is SPLICE_F_NONBLOCK.
2116  */
2117 static long do_tee(struct file *in, struct file *out, size_t len,
2118                    unsigned int flags)
2119 {
2120         struct pipe_inode_info *ipipe = get_pipe_info(in);
2121         struct pipe_inode_info *opipe = get_pipe_info(out);
2122         int ret = -EINVAL;
2123
2124         /*
2125          * Duplicate the contents of ipipe to opipe without actually
2126          * copying the data.
2127          */
2128         if (ipipe && opipe && ipipe != opipe) {
2129                 /*
2130                  * Keep going, unless we encounter an error. The ipipe/opipe
2131                  * ordering doesn't really matter.
2132                  */
2133                 ret = ipipe_prep(ipipe, flags);
2134                 if (!ret) {
2135                         ret = opipe_prep(opipe, flags);
2136                         if (!ret)
2137                                 ret = link_pipe(ipipe, opipe, len, flags);
2138                 }
2139         }
2140
2141         return ret;
2142 }
2143
2144 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2145 {
2146         struct fd in;
2147         int error;
2148
2149         if (unlikely(!len))
2150                 return 0;
2151
2152         error = -EBADF;
2153         in = fdget(fdin);
2154         if (in.file) {
2155                 if (in.file->f_mode & FMODE_READ) {
2156                         struct fd out = fdget(fdout);
2157                         if (out.file) {
2158                                 if (out.file->f_mode & FMODE_WRITE)
2159                                         error = do_tee(in.file, out.file,
2160                                                         len, flags);
2161                                 fdput(out);
2162                         }
2163                 }
2164                 fdput(in);
2165         }
2166
2167         return error;
2168 }