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 #include <linux/gfp.h>
  34
  35 /*
  36  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  37  * a vm helper function, it's already simplified quite a bit by the
  38  * addition of remove_mapping(). If success is returned, the caller may
  39  * attempt to reuse this page for another destination.
  40  */
  41 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  42                                      struct pipe_buffer *buf)
  43 {
  44         struct page *page = buf->page;
  45         struct address_space *mapping;
  46
  47         lock_page(page);
  48
  49         mapping = page_mapping(page);
  50         if (mapping) {
  51                 WARN_ON(!PageUptodate(page));
  52
  53                 /*
  54                  * At least for ext2 with nobh option, we need to wait on
  55                  * writeback completing on this page, since we'll remove it
  56                  * from the pagecache.  Otherwise truncate wont wait on the
  57                  * page, allowing the disk blocks to be reused by someone else
  58                  * before we actually wrote our data to them. fs corruption
  59                  * ensues.
  60                  */
  61                 wait_on_page_writeback(page);
  62
  63                 if (page_has_private(page) &&
  64                     !try_to_release_page(page, GFP_KERNEL))
  65                         goto out_unlock;
  66
  67                 /*
  68                  * If we succeeded in removing the mapping, set LRU flag
  69                  * and return good.
  70                  */
  71                 if (remove_mapping(mapping, page)) {
  72                         buf->flags |= PIPE_BUF_FLAG_LRU;
  73                         return 0;
  74                 }
  75         }
  76
  77         /*
  78          * Raced with truncate or failed to remove page from current
  79          * address space, unlock and return failure.
  80          */
  81 out_unlock:
  82         unlock_page(page);
  83         return 1;
  84 }
  85
  86 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  87                                         struct pipe_buffer *buf)
  88 {
  89         page_cache_release(buf->page);
  90         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  91 }
  92
  93 /*
  94  * Check whether the contents of buf is OK to access. Since the content
  95  * is a page cache page, IO may be in flight.
  96  */
  97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  98                                        struct pipe_buffer *buf)
  99 {
 100         struct page *page = buf->page;
 101         int err;
 102
 103         if (!PageUptodate(page)) {
 104                 lock_page(page);
 105
 106                 /*
 107                  * Page got truncated/unhashed. This will cause a 0-byte
 108                  * splice, if this is the first page.
 109                  */
 110                 if (!page->mapping) {
 111                         err = -ENODATA;
 112                         goto error;
 113                 }
 114
 115                 /*
 116                  * Uh oh, read-error from disk.
 117                  */
 118                 if (!PageUptodate(page)) {
 119                         err = -EIO;
 120                         goto error;
 121                 }
 122
 123                 /*
 124                  * Page is ok afterall, we are done.
 125                  */
 126                 unlock_page(page);
 127         }
 128
 129         return 0;
 130 error:
 131         unlock_page(page);
 132         return err;
 133 }
 134
 135 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 136         .can_merge = 0,
 137         .map = generic_pipe_buf_map,
 138         .unmap = generic_pipe_buf_unmap,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .map = generic_pipe_buf_map,
 158         .unmap = generic_pipe_buf_unmap,
 159         .confirm = generic_pipe_buf_confirm,
 160         .release = page_cache_pipe_buf_release,
 161         .steal = user_page_pipe_buf_steal,
 162         .get = generic_pipe_buf_get,
 163 };
 164
 165 /**
 166  * splice_to_pipe - fill passed data into a pipe
 167  * @pipe:       pipe to fill
 168  * @spd:        data to fill
 169  *
 170  * Description:
 171  *    @spd contains a map of pages and len/offset tuples, along with
 172  *    the struct pipe_buf_operations associated with these pages. This
 173  *    function will link that data to the pipe.
 174  *
 175  */
 176 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 177                        struct splice_pipe_desc *spd)
 178 {
 179         unsigned int spd_pages = spd->nr_pages;
 180         int ret, do_wakeup, page_nr;
 181
 182         ret = 0;
 183         do_wakeup = 0;
 184         page_nr = 0;
 185
 186         pipe_lock(pipe);
 187
 188         for (;;) {
 189                 if (!pipe->readers) {
 190                         send_sig(SIGPIPE, current, 0);
 191                         if (!ret)
 192                                 ret = -EPIPE;
 193                         break;
 194                 }
 195
 196                 if (pipe->nrbufs < pipe->buffers) {
 197                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 198                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 199
 200                         buf->page = spd->pages[page_nr];
 201                         buf->offset = spd->partial[page_nr].offset;
 202                         buf->len = spd->partial[page_nr].len;
 203                         buf->private = spd->partial[page_nr].private;
 204                         buf->ops = spd->ops;
 205                         if (spd->flags & SPLICE_F_GIFT)
 206                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 207
 208                         pipe->nrbufs++;
 209                         page_nr++;
 210                         ret += buf->len;
 211
 212                         if (pipe->inode)
 213                                 do_wakeup = 1;
 214
 215                         if (!--spd->nr_pages)
 216                                 break;
 217                         if (pipe->nrbufs < pipe->buffers)
 218                                 continue;
 219
 220                         break;
 221                 }
 222
 223                 if (spd->flags & SPLICE_F_NONBLOCK) {
 224                         if (!ret)
 225                                 ret = -EAGAIN;
 226                         break;
 227                 }
 228
 229                 if (signal_pending(current)) {
 230                         if (!ret)
 231                                 ret = -ERESTARTSYS;
 232                         break;
 233                 }
 234
 235                 if (do_wakeup) {
 236                         smp_mb();
 237                         if (waitqueue_active(&pipe->wait))
 238                                 wake_up_interruptible_sync(&pipe->wait);
 239                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 240                         do_wakeup = 0;
 241                 }
 242
 243                 pipe->waiting_writers++;
 244                 pipe_wait(pipe);
 245                 pipe->waiting_writers--;
 246         }
 247
 248         pipe_unlock(pipe);
 249
 250         if (do_wakeup) {
 251                 smp_mb();
 252                 if (waitqueue_active(&pipe->wait))
 253                         wake_up_interruptible(&pipe->wait);
 254                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 255         }
 256
 257         while (page_nr < spd_pages)
 258                 spd->spd_release(spd, page_nr++);
 259
 260         return ret;
 261 }
 262
 263 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 264 {
 265         page_cache_release(spd->pages[i]);
 266 }
 267
 268 /*
 269  * Check if we need to grow the arrays holding pages and partial page
 270  * descriptions.
 271  */
 272 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 273 {
 274         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 275                 return 0;
 276
 277         spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
 278         spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
 279
 280         if (spd->pages && spd->partial)
 281                 return 0;
 282
 283         kfree(spd->pages);
 284         kfree(spd->partial);
 285         return -ENOMEM;
 286 }
 287
 288 void splice_shrink_spd(struct pipe_inode_info *pipe,
 289                        struct splice_pipe_desc *spd)
 290 {
 291         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 292                 return;
 293
 294         kfree(spd->pages);
 295         kfree(spd->partial);
 296 }
 297
 298 static int
 299 __generic_file_splice_read(struct file *in, loff_t *ppos,
 300                            struct pipe_inode_info *pipe, size_t len,
 301                            unsigned int flags)
 302 {
 303         struct address_space *mapping = in->f_mapping;
 304         unsigned int loff, nr_pages, req_pages;
 305         struct page *pages[PIPE_DEF_BUFFERS];
 306         struct partial_page partial[PIPE_DEF_BUFFERS];
 307         struct page *page;
 308         pgoff_t index, end_index;
 309         loff_t isize;
 310         int error, page_nr;
 311         struct splice_pipe_desc spd = {
 312                 .pages = pages,
 313                 .partial = partial,
 314                 .flags = flags,
 315                 .ops = &page_cache_pipe_buf_ops,
 316                 .spd_release = spd_release_page,
 317         };
 318
 319         if (splice_grow_spd(pipe, &spd))
 320                 return -ENOMEM;
 321
 322         index = *ppos >> PAGE_CACHE_SHIFT;
 323         loff = *ppos & ~PAGE_CACHE_MASK;
 324         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 325         nr_pages = min(req_pages, pipe->buffers);
 326
 327         /*
 328          * Lookup the (hopefully) full range of pages we need.
 329          */
 330         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 331         index += spd.nr_pages;
 332
 333         /*
 334          * If find_get_pages_contig() returned fewer pages than we needed,
 335          * readahead/allocate the rest and fill in the holes.
 336          */
 337         if (spd.nr_pages < nr_pages)
 338                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 339                                 index, req_pages - spd.nr_pages);
 340
 341         error = 0;
 342         while (spd.nr_pages < nr_pages) {
 343                 /*
 344                  * Page could be there, find_get_pages_contig() breaks on
 345                  * the first hole.
 346                  */
 347                 page = find_get_page(mapping, index);
 348                 if (!page) {
 349                         /*
 350                          * page didn't exist, allocate one.
 351                          */
 352                         page = page_cache_alloc_cold(mapping);
 353                         if (!page)
 354                                 break;
 355
 356                         error = add_to_page_cache_lru(page, mapping, index,
 357                                                 GFP_KERNEL);
 358                         if (unlikely(error)) {
 359                                 page_cache_release(page);
 360                                 if (error == -EEXIST)
 361                                         continue;
 362                                 break;
 363                         }
 364                         /*
 365                          * add_to_page_cache() locks the page, unlock it
 366                          * to avoid convoluting the logic below even more.
 367                          */
 368                         unlock_page(page);
 369                 }
 370
 371                 spd.pages[spd.nr_pages++] = page;
 372                 index++;
 373         }
 374
 375         /*
 376          * Now loop over the map and see if we need to start IO on any
 377          * pages, fill in the partial map, etc.
 378          */
 379         index = *ppos >> PAGE_CACHE_SHIFT;
 380         nr_pages = spd.nr_pages;
 381         spd.nr_pages = 0;
 382         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 383                 unsigned int this_len;
 384
 385                 if (!len)
 386                         break;
 387
 388                 /*
 389                  * this_len is the max we'll use from this page
 390                  */
 391                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 392                 page = spd.pages[page_nr];
 393
 394                 if (PageReadahead(page))
 395                         page_cache_async_readahead(mapping, &in->f_ra, in,
 396                                         page, index, req_pages - page_nr);
 397
 398                 /*
 399                  * If the page isn't uptodate, we may need to start io on it
 400                  */
 401                 if (!PageUptodate(page)) {
 402                         /*
 403                          * If in nonblock mode then dont block on waiting
 404                          * for an in-flight io page
 405                          */
 406                         if (flags & SPLICE_F_NONBLOCK) {
 407                                 if (!trylock_page(page)) {
 408                                         error = -EAGAIN;
 409                                         break;
 410                                 }
 411                         } else
 412                                 lock_page(page);
 413
 414                         /*
 415                          * Page was truncated, or invalidated by the
 416                          * filesystem.  Redo the find/create, but this time the
 417                          * page is kept locked, so there's no chance of another
 418                          * race with truncate/invalidate.
 419                          */
 420                         if (!page->mapping) {
 421                                 unlock_page(page);
 422                                 page = find_or_create_page(mapping, index,
 423                                                 mapping_gfp_mask(mapping));
 424
 425                                 if (!page) {
 426                                         error = -ENOMEM;
 427                                         break;
 428                                 }
 429                                 page_cache_release(spd.pages[page_nr]);
 430                                 spd.pages[page_nr] = page;
 431                         }
 432                         /*
 433                          * page was already under io and is now done, great
 434                          */
 435                         if (PageUptodate(page)) {
 436                                 unlock_page(page);
 437                                 goto fill_it;
 438                         }
 439
 440                         /*
 441                          * need to read in the page
 442                          */
 443                         error = mapping->a_ops->readpage(in, page);
 444                         if (unlikely(error)) {
 445                                 /*
 446                                  * We really should re-lookup the page here,
 447                                  * but it complicates things a lot. Instead
 448                                  * lets just do what we already stored, and
 449                                  * we'll get it the next time we are called.
 450                                  */
 451                                 if (error == AOP_TRUNCATED_PAGE)
 452                                         error = 0;
 453
 454                                 break;
 455                         }
 456                 }
 457 fill_it:
 458                 /*
 459                  * i_size must be checked after PageUptodate.
 460                  */
 461                 isize = i_size_read(mapping->host);
 462                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 463                 if (unlikely(!isize || index > end_index))
 464                         break;
 465
 466                 /*
 467                  * if this is the last page, see if we need to shrink
 468                  * the length and stop
 469                  */
 470                 if (end_index == index) {
 471                         unsigned int plen;
 472
 473                         /*
 474                          * max good bytes in this page
 475                          */
 476                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 477                         if (plen <= loff)
 478                                 break;
 479
 480                         /*
 481                          * force quit after adding this page
 482                          */
 483                         this_len = min(this_len, plen - loff);
 484                         len = this_len;
 485                 }
 486
 487                 spd.partial[page_nr].offset = loff;
 488                 spd.partial[page_nr].len = this_len;
 489                 len -= this_len;
 490                 loff = 0;
 491                 spd.nr_pages++;
 492                 index++;
 493         }
 494
 495         /*
 496          * Release any pages at the end, if we quit early. 'page_nr' is how far
 497          * we got, 'nr_pages' is how many pages are in the map.
 498          */
 499         while (page_nr < nr_pages)
 500                 page_cache_release(spd.pages[page_nr++]);
 501         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 502
 503         if (spd.nr_pages)
 504                 error = splice_to_pipe(pipe, &spd);
 505
 506         splice_shrink_spd(pipe, &spd);
 507         return error;
 508 }
 509
 510 /**
 511  * generic_file_splice_read - splice data from file to a pipe
 512  * @in:         file to splice from
 513  * @ppos:       position in @in
 514  * @pipe:       pipe to splice to
 515  * @len:        number of bytes to splice
 516  * @flags:      splice modifier flags
 517  *
 518  * Description:
 519  *    Will read pages from given file and fill them into a pipe. Can be
 520  *    used as long as the address_space operations for the source implements
 521  *    a readpage() hook.
 522  *
 523  */
 524 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 525                                  struct pipe_inode_info *pipe, size_t len,
 526                                  unsigned int flags)
 527 {
 528         loff_t isize, left;
 529         int ret;
 530
 531         isize = i_size_read(in->f_mapping->host);
 532         if (unlikely(*ppos >= isize))
 533                 return 0;
 534
 535         left = isize - *ppos;
 536         if (unlikely(left < len))
 537                 len = left;
 538
 539         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 540         if (ret > 0) {
 541                 *ppos += ret;
 542                 file_accessed(in);
 543         }
 544
 545         return ret;
 546 }
 547 EXPORT_SYMBOL(generic_file_splice_read);
 548
 549 static const struct pipe_buf_operations default_pipe_buf_ops = {
 550         .can_merge = 0,
 551         .map = generic_pipe_buf_map,
 552         .unmap = generic_pipe_buf_unmap,
 553         .confirm = generic_pipe_buf_confirm,
 554         .release = generic_pipe_buf_release,
 555         .steal = generic_pipe_buf_steal,
 556         .get = generic_pipe_buf_get,
 557 };
 558
 559 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 560                             unsigned long vlen, loff_t offset)
 561 {
 562         mm_segment_t old_fs;
 563         loff_t pos = offset;
 564         ssize_t res;
 565
 566         old_fs = get_fs();
 567         set_fs(get_ds());
 568         /* The cast to a user pointer is valid due to the set_fs() */
 569         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 570         set_fs(old_fs);
 571
 572         return res;
 573 }
 574
 575 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 576                             loff_t pos)
 577 {
 578         mm_segment_t old_fs;
 579         ssize_t res;
 580
 581         old_fs = get_fs();
 582         set_fs(get_ds());
 583         /* The cast to a user pointer is valid due to the set_fs() */
 584         res = vfs_write(file, (const char __user *)buf, count, &pos);
 585         set_fs(old_fs);
 586
 587         return res;
 588 }
 589
 590 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 591                                  struct pipe_inode_info *pipe, size_t len,
 592                                  unsigned int flags)
 593 {
 594         unsigned int nr_pages;
 595         unsigned int nr_freed;
 596         size_t offset;
 597         struct page *pages[PIPE_DEF_BUFFERS];
 598         struct partial_page partial[PIPE_DEF_BUFFERS];
 599         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 600         pgoff_t index;
 601         ssize_t res;
 602         size_t this_len;
 603         int error;
 604         int i;
 605         struct splice_pipe_desc spd = {
 606                 .pages = pages,
 607                 .partial = partial,
 608                 .flags = flags,
 609                 .ops = &default_pipe_buf_ops,
 610                 .spd_release = spd_release_page,
 611         };
 612
 613         if (splice_grow_spd(pipe, &spd))
 614                 return -ENOMEM;
 615
 616         res = -ENOMEM;
 617         vec = __vec;
 618         if (pipe->buffers > PIPE_DEF_BUFFERS) {
 619                 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
 620                 if (!vec)
 621                         goto shrink_ret;
 622         }
 623
 624         index = *ppos >> PAGE_CACHE_SHIFT;
 625         offset = *ppos & ~PAGE_CACHE_MASK;
 626         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 627
 628         for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
 629                 struct page *page;
 630
 631                 page = alloc_page(GFP_USER);
 632                 error = -ENOMEM;
 633                 if (!page)
 634                         goto err;
 635
 636                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 637                 vec[i].iov_base = (void __user *) page_address(page);
 638                 vec[i].iov_len = this_len;
 639                 spd.pages[i] = page;
 640                 spd.nr_pages++;
 641                 len -= this_len;
 642                 offset = 0;
 643         }
 644
 645         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 646         if (res < 0) {
 647                 error = res;
 648                 goto err;
 649         }
 650
 651         error = 0;
 652         if (!res)
 653                 goto err;
 654
 655         nr_freed = 0;
 656         for (i = 0; i < spd.nr_pages; i++) {
 657                 this_len = min_t(size_t, vec[i].iov_len, res);
 658                 spd.partial[i].offset = 0;
 659                 spd.partial[i].len = this_len;
 660                 if (!this_len) {
 661                         __free_page(spd.pages[i]);
 662                         spd.pages[i] = NULL;
 663                         nr_freed++;
 664                 }
 665                 res -= this_len;
 666         }
 667         spd.nr_pages -= nr_freed;
 668
 669         res = splice_to_pipe(pipe, &spd);
 670         if (res > 0)
 671                 *ppos += res;
 672
 673 shrink_ret:
 674         if (vec != __vec)
 675                 kfree(vec);
 676         splice_shrink_spd(pipe, &spd);
 677         return res;
 678
 679 err:
 680         for (i = 0; i < spd.nr_pages; i++)
 681                 __free_page(spd.pages[i]);
 682
 683         res = error;
 684         goto shrink_ret;
 685 }
 686 EXPORT_SYMBOL(default_file_splice_read);
 687
 688 /*
 689  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 690  * using sendpage(). Return the number of bytes sent.
 691  */
 692 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 693                             struct pipe_buffer *buf, struct splice_desc *sd)
 694 {
 695         struct file *file = sd->u.file;
 696         loff_t pos = sd->pos;
 697         int ret, more;
 698
 699         ret = buf->ops->confirm(pipe, buf);
 700         if (!ret) {
 701                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 702                 if (file->f_op && file->f_op->sendpage)
 703                         ret = file->f_op->sendpage(file, buf->page, buf->offset,
 704                                                    sd->len, &pos, more);
 705                 else
 706                         ret = -EINVAL;
 707         }
 708
 709         return ret;
 710 }
 711
 712 /*
 713  * This is a little more tricky than the file -> pipe splicing. There are
 714  * basically three cases:
 715  *
 716  *      - Destination page already exists in the address space and there
 717  *        are users of it. For that case we have no other option that
 718  *        copying the data. Tough luck.
 719  *      - Destination page already exists in the address space, but there
 720  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 721  *        through to last case.
 722  *      - Destination page does not exist, we can add the pipe page to
 723  *        the page cache and avoid the copy.
 724  *
 725  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 726  * sd->flags), we attempt to migrate pages from the pipe to the output
 727  * file address space page cache. This is possible if no one else has
 728  * the pipe page referenced outside of the pipe and page cache. If
 729  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 730  * a new page in the output file page cache and fill/dirty that.
 731  */
 732 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 733                  struct splice_desc *sd)
 734 {
 735         struct file *file = sd->u.file;
 736         struct address_space *mapping = file->f_mapping;
 737         unsigned int offset, this_len;
 738         struct page *page;
 739         void *fsdata;
 740         int ret;
 741
 742         /*
 743          * make sure the data in this buffer is uptodate
 744          */
 745         ret = buf->ops->confirm(pipe, buf);
 746         if (unlikely(ret))
 747                 return ret;
 748
 749         offset = sd->pos & ~PAGE_CACHE_MASK;
 750
 751         this_len = sd->len;
 752         if (this_len + offset > PAGE_CACHE_SIZE)
 753                 this_len = PAGE_CACHE_SIZE - offset;
 754
 755         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 756                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 757         if (unlikely(ret))
 758                 goto out;
 759
 760         if (buf->page != page) {
 761                 /*
 762                  * Careful, ->map() uses KM_USER0!
 763                  */
 764                 char *src = buf->ops->map(pipe, buf, 1);
 765                 char *dst = kmap_atomic(page, KM_USER1);
 766
 767                 memcpy(dst + offset, src + buf->offset, this_len);
 768                 flush_dcache_page(page);
 769                 kunmap_atomic(dst, KM_USER1);
 770                 buf->ops->unmap(pipe, buf, src);
 771         }
 772         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 773                                 page, fsdata);
 774 out:
 775         return ret;
 776 }
 777 EXPORT_SYMBOL(pipe_to_file);
 778
 779 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 780 {
 781         smp_mb();
 782         if (waitqueue_active(&pipe->wait))
 783                 wake_up_interruptible(&pipe->wait);
 784         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 785 }
 786
 787 /**
 788  * splice_from_pipe_feed - feed available data from a pipe to a file
 789  * @pipe:       pipe to splice from
 790  * @sd:         information to @actor
 791  * @actor:      handler that splices the data
 792  *
 793  * Description:
 794  *    This function loops over the pipe and calls @actor to do the
 795  *    actual moving of a single struct pipe_buffer to the desired
 796  *    destination.  It returns when there's no more buffers left in
 797  *    the pipe or if the requested number of bytes (@sd->total_len)
 798  *    have been copied.  It returns a positive number (one) if the
 799  *    pipe needs to be filled with more data, zero if the required
 800  *    number of bytes have been copied and -errno on error.
 801  *
 802  *    This, together with splice_from_pipe_{begin,end,next}, may be
 803  *    used to implement the functionality of __splice_from_pipe() when
 804  *    locking is required around copying the pipe buffers to the
 805  *    destination.
 806  */
 807 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 808                           splice_actor *actor)
 809 {
 810         int ret;
 811
 812         while (pipe->nrbufs) {
 813                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 814                 const struct pipe_buf_operations *ops = buf->ops;
 815
 816                 sd->len = buf->len;
 817                 if (sd->len > sd->total_len)
 818                         sd->len = sd->total_len;
 819
 820                 ret = actor(pipe, buf, sd);
 821                 if (ret <= 0) {
 822                         if (ret == -ENODATA)
 823                                 ret = 0;
 824                         return ret;
 825                 }
 826                 buf->offset += ret;
 827                 buf->len -= ret;
 828
 829                 sd->num_spliced += ret;
 830                 sd->len -= ret;
 831                 sd->pos += ret;
 832                 sd->total_len -= ret;
 833
 834                 if (!buf->len) {
 835                         buf->ops = NULL;
 836                         ops->release(pipe, buf);
 837                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 838                         pipe->nrbufs--;
 839                         if (pipe->inode)
 840                                 sd->need_wakeup = true;
 841                 }
 842
 843                 if (!sd->total_len)
 844                         return 0;
 845         }
 846
 847         return 1;
 848 }
 849 EXPORT_SYMBOL(splice_from_pipe_feed);
 850
 851 /**
 852  * splice_from_pipe_next - wait for some data to splice from
 853  * @pipe:       pipe to splice from
 854  * @sd:         information about the splice operation
 855  *
 856  * Description:
 857  *    This function will wait for some data and return a positive
 858  *    value (one) if pipe buffers are available.  It will return zero
 859  *    or -errno if no more data needs to be spliced.
 860  */
 861 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 862 {
 863         while (!pipe->nrbufs) {
 864                 if (!pipe->writers)
 865                         return 0;
 866
 867                 if (!pipe->waiting_writers && sd->num_spliced)
 868                         return 0;
 869
 870                 if (sd->flags & SPLICE_F_NONBLOCK)
 871                         return -EAGAIN;
 872
 873                 if (signal_pending(current))
 874                         return -ERESTARTSYS;
 875
 876                 if (sd->need_wakeup) {
 877                         wakeup_pipe_writers(pipe);
 878                         sd->need_wakeup = false;
 879                 }
 880
 881                 pipe_wait(pipe);
 882         }
 883
 884         return 1;
 885 }
 886 EXPORT_SYMBOL(splice_from_pipe_next);
 887
 888 /**
 889  * splice_from_pipe_begin - start splicing from pipe
 890  * @sd:         information about the splice operation
 891  *
 892  * Description:
 893  *    This function should be called before a loop containing
 894  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 895  *    initialize the necessary fields of @sd.
 896  */
 897 void splice_from_pipe_begin(struct splice_desc *sd)
 898 {
 899         sd->num_spliced = 0;
 900         sd->need_wakeup = false;
 901 }
 902 EXPORT_SYMBOL(splice_from_pipe_begin);
 903
 904 /**
 905  * splice_from_pipe_end - finish splicing from pipe
 906  * @pipe:       pipe to splice from
 907  * @sd:         information about the splice operation
 908  *
 909  * Description:
 910  *    This function will wake up pipe writers if necessary.  It should
 911  *    be called after a loop containing splice_from_pipe_next() and
 912  *    splice_from_pipe_feed().
 913  */
 914 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 915 {
 916         if (sd->need_wakeup)
 917                 wakeup_pipe_writers(pipe);
 918 }
 919 EXPORT_SYMBOL(splice_from_pipe_end);
 920
 921 /**
 922  * __splice_from_pipe - splice data from a pipe to given actor
 923  * @pipe:       pipe to splice from
 924  * @sd:         information to @actor
 925  * @actor:      handler that splices the data
 926  *
 927  * Description:
 928  *    This function does little more than loop over the pipe and call
 929  *    @actor to do the actual moving of a single struct pipe_buffer to
 930  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 931  *    pipe_to_user.
 932  *
 933  */
 934 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 935                            splice_actor *actor)
 936 {
 937         int ret;
 938
 939         splice_from_pipe_begin(sd);
 940         do {
 941                 ret = splice_from_pipe_next(pipe, sd);
 942                 if (ret > 0)
 943                         ret = splice_from_pipe_feed(pipe, sd, actor);
 944         } while (ret > 0);
 945         splice_from_pipe_end(pipe, sd);
 946
 947         return sd->num_spliced ? sd->num_spliced : ret;
 948 }
 949 EXPORT_SYMBOL(__splice_from_pipe);
 950
 951 /**
 952  * splice_from_pipe - splice data from a pipe to a file
 953  * @pipe:       pipe to splice from
 954  * @out:        file to splice to
 955  * @ppos:       position in @out
 956  * @len:        how many bytes to splice
 957  * @flags:      splice modifier flags
 958  * @actor:      handler that splices the data
 959  *
 960  * Description:
 961  *    See __splice_from_pipe. This function locks the pipe inode,
 962  *    otherwise it's identical to __splice_from_pipe().
 963  *
 964  */
 965 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 966                          loff_t *ppos, size_t len, unsigned int flags,
 967                          splice_actor *actor)
 968 {
 969         ssize_t ret;
 970         struct splice_desc sd = {
 971                 .total_len = len,
 972                 .flags = flags,
 973                 .pos = *ppos,
 974                 .u.file = out,
 975         };
 976
 977         pipe_lock(pipe);
 978         ret = __splice_from_pipe(pipe, &sd, actor);
 979         pipe_unlock(pipe);
 980
 981         return ret;
 982 }
 983
 984 /**
 985  * generic_file_splice_write - splice data from a pipe to a file
 986  * @pipe:       pipe info
 987  * @out:        file to write to
 988  * @ppos:       position in @out
 989  * @len:        number of bytes to splice
 990  * @flags:      splice modifier flags
 991  *
 992  * Description:
 993  *    Will either move or copy pages (determined by @flags options) from
 994  *    the given pipe inode to the given file.
 995  *
 996  */
 997 ssize_t
 998 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 999                           loff_t *ppos, size_t len, unsigned int flags)
1000 {
1001         struct address_space *mapping = out->f_mapping;
1002         struct inode *inode = mapping->host;
1003         struct splice_desc sd = {
1004                 .total_len = len,
1005                 .flags = flags,
1006                 .pos = *ppos,
1007                 .u.file = out,
1008         };
1009         ssize_t ret;
1010
1011         pipe_lock(pipe);
1012
1013         splice_from_pipe_begin(&sd);
1014         do {
1015                 ret = splice_from_pipe_next(pipe, &sd);
1016                 if (ret <= 0)
1017                         break;
1018
1019                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1020                 ret = file_remove_suid(out);
1021                 if (!ret) {
1022                         file_update_time(out);
1023                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1024                 }
1025                 mutex_unlock(&inode->i_mutex);
1026         } while (ret > 0);
1027         splice_from_pipe_end(pipe, &sd);
1028
1029         pipe_unlock(pipe);
1030
1031         if (sd.num_spliced)
1032                 ret = sd.num_spliced;
1033
1034         if (ret > 0) {
1035                 unsigned long nr_pages;
1036                 int err;
1037
1038                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1039
1040                 err = generic_write_sync(out, *ppos, ret);
1041                 if (err)
1042                         ret = err;
1043                 else
1044                         *ppos += ret;
1045                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1046         }
1047
1048         return ret;
1049 }
1050
1051 EXPORT_SYMBOL(generic_file_splice_write);
1052
1053 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1054                           struct splice_desc *sd)
1055 {
1056         int ret;
1057         void *data;
1058
1059         ret = buf->ops->confirm(pipe, buf);
1060         if (ret)
1061                 return ret;
1062
1063         data = buf->ops->map(pipe, buf, 0);
1064         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1065         buf->ops->unmap(pipe, buf, data);
1066
1067         return ret;
1068 }
1069
1070 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1071                                          struct file *out, loff_t *ppos,
1072                                          size_t len, unsigned int flags)
1073 {
1074         ssize_t ret;
1075
1076         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1077         if (ret > 0)
1078                 *ppos += ret;
1079
1080         return ret;
1081 }
1082
1083 /**
1084  * generic_splice_sendpage - splice data from a pipe to a socket
1085  * @pipe:       pipe to splice from
1086  * @out:        socket to write to
1087  * @ppos:       position in @out
1088  * @len:        number of bytes to splice
1089  * @flags:      splice modifier flags
1090  *
1091  * Description:
1092  *    Will send @len bytes from the pipe to a network socket. No data copying
1093  *    is involved.
1094  *
1095  */
1096 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1097                                 loff_t *ppos, size_t len, unsigned int flags)
1098 {
1099         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1100 }
1101
1102 EXPORT_SYMBOL(generic_splice_sendpage);
1103
1104 /*
1105  * Attempt to initiate a splice from pipe to file.
1106  */
1107 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1108                            loff_t *ppos, size_t len, unsigned int flags)
1109 {
1110         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1111                                 loff_t *, size_t, unsigned int);
1112         int ret;
1113
1114         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1115                 return -EBADF;
1116
1117         if (unlikely(out->f_flags & O_APPEND))
1118                 return -EINVAL;
1119
1120         ret = rw_verify_area(WRITE, out, ppos, len);
1121         if (unlikely(ret < 0))
1122                 return ret;
1123
1124         if (out->f_op && out->f_op->splice_write)
1125                 splice_write = out->f_op->splice_write;
1126         else
1127                 splice_write = default_file_splice_write;
1128
1129         return splice_write(pipe, out, ppos, len, flags);
1130 }
1131
1132 /*
1133  * Attempt to initiate a splice from a file to a pipe.
1134  */
1135 static long do_splice_to(struct file *in, loff_t *ppos,
1136                          struct pipe_inode_info *pipe, size_t len,
1137                          unsigned int flags)
1138 {
1139         ssize_t (*splice_read)(struct file *, loff_t *,
1140                                struct pipe_inode_info *, size_t, unsigned int);
1141         int ret;
1142
1143         if (unlikely(!(in->f_mode & FMODE_READ)))
1144                 return -EBADF;
1145
1146         ret = rw_verify_area(READ, in, ppos, len);
1147         if (unlikely(ret < 0))
1148                 return ret;
1149
1150         if (in->f_op && in->f_op->splice_read)
1151                 splice_read = in->f_op->splice_read;
1152         else
1153                 splice_read = default_file_splice_read;
1154
1155         return splice_read(in, ppos, pipe, len, flags);
1156 }
1157
1158 /**
1159  * splice_direct_to_actor - splices data directly between two non-pipes
1160  * @in:         file to splice from
1161  * @sd:         actor information on where to splice to
1162  * @actor:      handles the data splicing
1163  *
1164  * Description:
1165  *    This is a special case helper to splice directly between two
1166  *    points, without requiring an explicit pipe. Internally an allocated
1167  *    pipe is cached in the process, and reused during the lifetime of
1168  *    that process.
1169  *
1170  */
1171 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1172                                splice_direct_actor *actor)
1173 {
1174         struct pipe_inode_info *pipe;
1175         long ret, bytes;
1176         umode_t i_mode;
1177         size_t len;
1178         int i, flags;
1179
1180         /*
1181          * We require the input being a regular file, as we don't want to
1182          * randomly drop data for eg socket -> socket splicing. Use the
1183          * piped splicing for that!
1184          */
1185         i_mode = in->f_path.dentry->d_inode->i_mode;
1186         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1187                 return -EINVAL;
1188
1189         /*
1190          * neither in nor out is a pipe, setup an internal pipe attached to
1191          * 'out' and transfer the wanted data from 'in' to 'out' through that
1192          */
1193         pipe = current->splice_pipe;
1194         if (unlikely(!pipe)) {
1195                 pipe = alloc_pipe_info(NULL);
1196                 if (!pipe)
1197                         return -ENOMEM;
1198
1199                 /*
1200                  * We don't have an immediate reader, but we'll read the stuff
1201                  * out of the pipe right after the splice_to_pipe(). So set
1202                  * PIPE_READERS appropriately.
1203                  */
1204                 pipe->readers = 1;
1205
1206                 current->splice_pipe = pipe;
1207         }
1208
1209         /*
1210          * Do the splice.
1211          */
1212         ret = 0;
1213         bytes = 0;
1214         len = sd->total_len;
1215         flags = sd->flags;
1216
1217         /*
1218          * Don't block on output, we have to drain the direct pipe.
1219          */
1220         sd->flags &= ~SPLICE_F_NONBLOCK;
1221
1222         while (len) {
1223                 size_t read_len;
1224                 loff_t pos = sd->pos, prev_pos = pos;
1225
1226                 ret = do_splice_to(in, &pos, pipe, len, flags);
1227                 if (unlikely(ret <= 0))
1228                         goto out_release;
1229
1230                 read_len = ret;
1231                 sd->total_len = read_len;
1232
1233                 /*
1234                  * NOTE: nonblocking mode only applies to the input. We
1235                  * must not do the output in nonblocking mode as then we
1236                  * could get stuck data in the internal pipe:
1237                  */
1238                 ret = actor(pipe, sd);
1239                 if (unlikely(ret <= 0)) {
1240                         sd->pos = prev_pos;
1241                         goto out_release;
1242                 }
1243
1244                 bytes += ret;
1245                 len -= ret;
1246                 sd->pos = pos;
1247
1248                 if (ret < read_len) {
1249                         sd->pos = prev_pos + ret;
1250                         goto out_release;
1251                 }
1252         }
1253
1254 done:
1255         pipe->nrbufs = pipe->curbuf = 0;
1256         file_accessed(in);
1257         return bytes;
1258
1259 out_release:
1260         /*
1261          * If we did an incomplete transfer we must release
1262          * the pipe buffers in question:
1263          */
1264         for (i = 0; i < pipe->buffers; i++) {
1265                 struct pipe_buffer *buf = pipe->bufs + i;
1266
1267                 if (buf->ops) {
1268                         buf->ops->release(pipe, buf);
1269                         buf->ops = NULL;
1270                 }
1271         }
1272
1273         if (!bytes)
1274                 bytes = ret;
1275
1276         goto done;
1277 }
1278 EXPORT_SYMBOL(splice_direct_to_actor);
1279
1280 static int direct_splice_actor(struct pipe_inode_info *pipe,
1281                                struct splice_desc *sd)
1282 {
1283         struct file *file = sd->u.file;
1284
1285         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1286 }
1287
1288 /**
1289  * do_splice_direct - splices data directly between two files
1290  * @in:         file to splice from
1291  * @ppos:       input file offset
1292  * @out:        file to splice to
1293  * @len:        number of bytes to splice
1294  * @flags:      splice modifier flags
1295  *
1296  * Description:
1297  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1298  *    doing it in the application would incur an extra system call
1299  *    (splice in + splice out, as compared to just sendfile()). So this helper
1300  *    can splice directly through a process-private pipe.
1301  *
1302  */
1303 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1304                       size_t len, unsigned int flags)
1305 {
1306         struct splice_desc sd = {
1307                 .len            = len,
1308                 .total_len      = len,
1309                 .flags          = flags,
1310                 .pos            = *ppos,
1311                 .u.file         = out,
1312         };
1313         long ret;
1314
1315         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1316         if (ret > 0)
1317                 *ppos = sd.pos;
1318
1319         return ret;
1320 }
1321
1322 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1323                                struct pipe_inode_info *opipe,
1324                                size_t len, unsigned int flags);
1325 /*
1326  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1327  * location, so checking ->i_pipe is not enough to verify that this is a
1328  * pipe.
1329  */
1330 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1331 {
1332         if (S_ISFIFO(inode->i_mode))
1333                 return inode->i_pipe;
1334
1335         return NULL;
1336 }
1337
1338 /*
1339  * Determine where to splice to/from.
1340  */
1341 static long do_splice(struct file *in, loff_t __user *off_in,
1342                       struct file *out, loff_t __user *off_out,
1343                       size_t len, unsigned int flags)
1344 {
1345         struct pipe_inode_info *ipipe;
1346         struct pipe_inode_info *opipe;
1347         loff_t offset, *off;
1348         long ret;
1349
1350         ipipe = pipe_info(in->f_path.dentry->d_inode);
1351         opipe = pipe_info(out->f_path.dentry->d_inode);
1352
1353         if (ipipe && opipe) {
1354                 if (off_in || off_out)
1355                         return -ESPIPE;
1356
1357                 if (!(in->f_mode & FMODE_READ))
1358                         return -EBADF;
1359
1360                 if (!(out->f_mode & FMODE_WRITE))
1361                         return -EBADF;
1362
1363                 /* Splicing to self would be fun, but... */
1364                 if (ipipe == opipe)
1365                         return -EINVAL;
1366
1367                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1368         }
1369
1370         if (ipipe) {
1371                 if (off_in)
1372                         return -ESPIPE;
1373                 if (off_out) {
1374                         if (!out->f_op || !out->f_op->llseek ||
1375                             out->f_op->llseek == no_llseek)
1376                                 return -EINVAL;
1377                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1378                                 return -EFAULT;
1379                         off = &offset;
1380                 } else
1381                         off = &out->f_pos;
1382
1383                 ret = do_splice_from(ipipe, out, off, len, flags);
1384
1385                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1386                         ret = -EFAULT;
1387
1388                 return ret;
1389         }
1390
1391         if (opipe) {
1392                 if (off_out)
1393                         return -ESPIPE;
1394                 if (off_in) {
1395                         if (!in->f_op || !in->f_op->llseek ||
1396                             in->f_op->llseek == no_llseek)
1397                                 return -EINVAL;
1398                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1399                                 return -EFAULT;
1400                         off = &offset;
1401                 } else
1402                         off = &in->f_pos;
1403
1404                 ret = do_splice_to(in, off, opipe, len, flags);
1405
1406                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1407                         ret = -EFAULT;
1408
1409                 return ret;
1410         }
1411
1412         return -EINVAL;
1413 }
1414
1415 /*
1416  * Map an iov into an array of pages and offset/length tupples. With the
1417  * partial_page structure, we can map several non-contiguous ranges into
1418  * our ones pages[] map instead of splitting that operation into pieces.
1419  * Could easily be exported as a generic helper for other users, in which
1420  * case one would probably want to add a 'max_nr_pages' parameter as well.
1421  */
1422 static int get_iovec_page_array(const struct iovec __user *iov,
1423                                 unsigned int nr_vecs, struct page **pages,
1424                                 struct partial_page *partial, int aligned,
1425                                 unsigned int pipe_buffers)
1426 {
1427         int buffers = 0, error = 0;
1428
1429         while (nr_vecs) {
1430                 unsigned long off, npages;
1431                 struct iovec entry;
1432                 void __user *base;
1433                 size_t len;
1434                 int i;
1435
1436                 error = -EFAULT;
1437                 if (copy_from_user(&entry, iov, sizeof(entry)))
1438                         break;
1439
1440                 base = entry.iov_base;
1441                 len = entry.iov_len;
1442
1443                 /*
1444                  * Sanity check this iovec. 0 read succeeds.
1445                  */
1446                 error = 0;
1447                 if (unlikely(!len))
1448                         break;
1449                 error = -EFAULT;
1450                 if (!access_ok(VERIFY_READ, base, len))
1451                         break;
1452
1453                 /*
1454                  * Get this base offset and number of pages, then map
1455                  * in the user pages.
1456                  */
1457                 off = (unsigned long) base & ~PAGE_MASK;
1458
1459                 /*
1460                  * If asked for alignment, the offset must be zero and the
1461                  * length a multiple of the PAGE_SIZE.
1462                  */
1463                 error = -EINVAL;
1464                 if (aligned && (off || len & ~PAGE_MASK))
1465                         break;
1466
1467                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1468                 if (npages > pipe_buffers - buffers)
1469                         npages = pipe_buffers - buffers;
1470
1471                 error = get_user_pages_fast((unsigned long)base, npages,
1472                                         0, &pages[buffers]);
1473
1474                 if (unlikely(error <= 0))
1475                         break;
1476
1477                 /*
1478                  * Fill this contiguous range into the partial page map.
1479                  */
1480                 for (i = 0; i < error; i++) {
1481                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1482
1483                         partial[buffers].offset = off;
1484                         partial[buffers].len = plen;
1485
1486                         off = 0;
1487                         len -= plen;
1488                         buffers++;
1489                 }
1490
1491                 /*
1492                  * We didn't complete this iov, stop here since it probably
1493                  * means we have to move some of this into a pipe to
1494                  * be able to continue.
1495                  */
1496                 if (len)
1497                         break;
1498
1499                 /*
1500                  * Don't continue if we mapped fewer pages than we asked for,
1501                  * or if we mapped the max number of pages that we have
1502                  * room for.
1503                  */
1504                 if (error < npages || buffers == pipe_buffers)
1505                         break;
1506
1507                 nr_vecs--;
1508                 iov++;
1509         }
1510
1511         if (buffers)
1512                 return buffers;
1513
1514         return error;
1515 }
1516
1517 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1518                         struct splice_desc *sd)
1519 {
1520         char *src;
1521         int ret;
1522
1523         ret = buf->ops->confirm(pipe, buf);
1524         if (unlikely(ret))
1525                 return ret;
1526
1527         /*
1528          * See if we can use the atomic maps, by prefaulting in the
1529          * pages and doing an atomic copy
1530          */
1531         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1532                 src = buf->ops->map(pipe, buf, 1);
1533                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1534                                                         sd->len);
1535                 buf->ops->unmap(pipe, buf, src);
1536                 if (!ret) {
1537                         ret = sd->len;
1538                         goto out;
1539                 }
1540         }
1541
1542         /*
1543          * No dice, use slow non-atomic map and copy
1544          */
1545         src = buf->ops->map(pipe, buf, 0);
1546
1547         ret = sd->len;
1548         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1549                 ret = -EFAULT;
1550
1551         buf->ops->unmap(pipe, buf, src);
1552 out:
1553         if (ret > 0)
1554                 sd->u.userptr += ret;
1555         return ret;
1556 }
1557
1558 /*
1559  * For lack of a better implementation, implement vmsplice() to userspace
1560  * as a simple copy of the pipes pages to the user iov.
1561  */
1562 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1563                              unsigned long nr_segs, unsigned int flags)
1564 {
1565         struct pipe_inode_info *pipe;
1566         struct splice_desc sd;
1567         ssize_t size;
1568         int error;
1569         long ret;
1570
1571         pipe = pipe_info(file->f_path.dentry->d_inode);
1572         if (!pipe)
1573                 return -EBADF;
1574
1575         pipe_lock(pipe);
1576
1577         error = ret = 0;
1578         while (nr_segs) {
1579                 void __user *base;
1580                 size_t len;
1581
1582                 /*
1583                  * Get user address base and length for this iovec.
1584                  */
1585                 error = get_user(base, &iov->iov_base);
1586                 if (unlikely(error))
1587                         break;
1588                 error = get_user(len, &iov->iov_len);
1589                 if (unlikely(error))
1590                         break;
1591
1592                 /*
1593                  * Sanity check this iovec. 0 read succeeds.
1594                  */
1595                 if (unlikely(!len))
1596                         break;
1597                 if (unlikely(!base)) {
1598                         error = -EFAULT;
1599                         break;
1600                 }
1601
1602                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1603                         error = -EFAULT;
1604                         break;
1605                 }
1606
1607                 sd.len = 0;
1608                 sd.total_len = len;
1609                 sd.flags = flags;
1610                 sd.u.userptr = base;
1611                 sd.pos = 0;
1612
1613                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1614                 if (size < 0) {
1615                         if (!ret)
1616                                 ret = size;
1617
1618                         break;
1619                 }
1620
1621                 ret += size;
1622
1623                 if (size < len)
1624                         break;
1625
1626                 nr_segs--;
1627                 iov++;
1628         }
1629
1630         pipe_unlock(pipe);
1631
1632         if (!ret)
1633                 ret = error;
1634
1635         return ret;
1636 }
1637
1638 /*
1639  * vmsplice splices a user address range into a pipe. It can be thought of
1640  * as splice-from-memory, where the regular splice is splice-from-file (or
1641  * to file). In both cases the output is a pipe, naturally.
1642  */
1643 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1644                              unsigned long nr_segs, unsigned int flags)
1645 {
1646         struct pipe_inode_info *pipe;
1647         struct page *pages[PIPE_DEF_BUFFERS];
1648         struct partial_page partial[PIPE_DEF_BUFFERS];
1649         struct splice_pipe_desc spd = {
1650                 .pages = pages,
1651                 .partial = partial,
1652                 .flags = flags,
1653                 .ops = &user_page_pipe_buf_ops,
1654                 .spd_release = spd_release_page,
1655         };
1656         long ret;
1657
1658         pipe = pipe_info(file->f_path.dentry->d_inode);
1659         if (!pipe)
1660                 return -EBADF;
1661
1662         if (splice_grow_spd(pipe, &spd))
1663                 return -ENOMEM;
1664
1665         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1666                                             spd.partial, flags & SPLICE_F_GIFT,
1667                                             pipe->buffers);
1668         if (spd.nr_pages <= 0)
1669                 ret = spd.nr_pages;
1670         else
1671                 ret = splice_to_pipe(pipe, &spd);
1672
1673         splice_shrink_spd(pipe, &spd);
1674         return ret;
1675 }
1676
1677 /*
1678  * Note that vmsplice only really supports true splicing _from_ user memory
1679  * to a pipe, not the other way around. Splicing from user memory is a simple
1680  * operation that can be supported without any funky alignment restrictions
1681  * or nasty vm tricks. We simply map in the user memory and fill them into
1682  * a pipe. The reverse isn't quite as easy, though. There are two possible
1683  * solutions for that:
1684  *
1685  *      - memcpy() the data internally, at which point we might as well just
1686  *        do a regular read() on the buffer anyway.
1687  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1688  *        has restriction limitations on both ends of the pipe).
1689  *
1690  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1691  *
1692  */
1693 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1694                 unsigned long, nr_segs, unsigned int, flags)
1695 {
1696         struct file *file;
1697         long error;
1698         int fput;
1699
1700         if (unlikely(nr_segs > UIO_MAXIOV))
1701                 return -EINVAL;
1702         else if (unlikely(!nr_segs))
1703                 return 0;
1704
1705         error = -EBADF;
1706         file = fget_light(fd, &fput);
1707         if (file) {
1708                 if (file->f_mode & FMODE_WRITE)
1709                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1710                 else if (file->f_mode & FMODE_READ)
1711                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1712
1713                 fput_light(file, fput);
1714         }
1715
1716         return error;
1717 }
1718
1719 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1720                 int, fd_out, loff_t __user *, off_out,
1721                 size_t, len, unsigned int, flags)
1722 {
1723         long error;
1724         struct file *in, *out;
1725         int fput_in, fput_out;
1726
1727         if (unlikely(!len))
1728                 return 0;
1729
1730         error = -EBADF;
1731         in = fget_light(fd_in, &fput_in);
1732         if (in) {
1733                 if (in->f_mode & FMODE_READ) {
1734                         out = fget_light(fd_out, &fput_out);
1735                         if (out) {
1736                                 if (out->f_mode & FMODE_WRITE)
1737                                         error = do_splice(in, off_in,
1738                                                           out, off_out,
1739                                                           len, flags);
1740                                 fput_light(out, fput_out);
1741                         }
1742                 }
1743
1744                 fput_light(in, fput_in);
1745         }
1746
1747         return error;
1748 }
1749
1750 /*
1751  * Make sure there's data to read. Wait for input if we can, otherwise
1752  * return an appropriate error.
1753  */
1754 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1755 {
1756         int ret;
1757
1758         /*
1759          * Check ->nrbufs without the inode lock first. This function
1760          * is speculative anyways, so missing one is ok.
1761          */
1762         if (pipe->nrbufs)
1763                 return 0;
1764
1765         ret = 0;
1766         pipe_lock(pipe);
1767
1768         while (!pipe->nrbufs) {
1769                 if (signal_pending(current)) {
1770                         ret = -ERESTARTSYS;
1771                         break;
1772                 }
1773                 if (!pipe->writers)
1774                         break;
1775                 if (!pipe->waiting_writers) {
1776                         if (flags & SPLICE_F_NONBLOCK) {
1777                                 ret = -EAGAIN;
1778                                 break;
1779                         }
1780                 }
1781                 pipe_wait(pipe);
1782         }
1783
1784         pipe_unlock(pipe);
1785         return ret;
1786 }
1787
1788 /*
1789  * Make sure there's writeable room. Wait for room if we can, otherwise
1790  * return an appropriate error.
1791  */
1792 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1793 {
1794         int ret;
1795
1796         /*
1797          * Check ->nrbufs without the inode lock first. This function
1798          * is speculative anyways, so missing one is ok.
1799          */
1800         if (pipe->nrbufs < pipe->buffers)
1801                 return 0;
1802
1803         ret = 0;
1804         pipe_lock(pipe);
1805
1806         while (pipe->nrbufs >= pipe->buffers) {
1807                 if (!pipe->readers) {
1808                         send_sig(SIGPIPE, current, 0);
1809                         ret = -EPIPE;
1810                         break;
1811                 }
1812                 if (flags & SPLICE_F_NONBLOCK) {
1813                         ret = -EAGAIN;
1814                         break;
1815                 }
1816                 if (signal_pending(current)) {
1817                         ret = -ERESTARTSYS;
1818                         break;
1819                 }
1820                 pipe->waiting_writers++;
1821                 pipe_wait(pipe);
1822                 pipe->waiting_writers--;
1823         }
1824
1825         pipe_unlock(pipe);
1826         return ret;
1827 }
1828
1829 /*
1830  * Splice contents of ipipe to opipe.
1831  */
1832 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1833                                struct pipe_inode_info *opipe,
1834                                size_t len, unsigned int flags)
1835 {
1836         struct pipe_buffer *ibuf, *obuf;
1837         int ret = 0, nbuf;
1838         bool input_wakeup = false;
1839
1840
1841 retry:
1842         ret = ipipe_prep(ipipe, flags);
1843         if (ret)
1844                 return ret;
1845
1846         ret = opipe_prep(opipe, flags);
1847         if (ret)
1848                 return ret;
1849
1850         /*
1851          * Potential ABBA deadlock, work around it by ordering lock
1852          * grabbing by pipe info address. Otherwise two different processes
1853          * could deadlock (one doing tee from A -> B, the other from B -> A).
1854          */
1855         pipe_double_lock(ipipe, opipe);
1856
1857         do {
1858                 if (!opipe->readers) {
1859                         send_sig(SIGPIPE, current, 0);
1860                         if (!ret)
1861                                 ret = -EPIPE;
1862                         break;
1863                 }
1864
1865                 if (!ipipe->nrbufs && !ipipe->writers)
1866                         break;
1867
1868                 /*
1869                  * Cannot make any progress, because either the input
1870                  * pipe is empty or the output pipe is full.
1871                  */
1872                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1873                         /* Already processed some buffers, break */
1874                         if (ret)
1875                                 break;
1876
1877                         if (flags & SPLICE_F_NONBLOCK) {
1878                                 ret = -EAGAIN;
1879                                 break;
1880                         }
1881
1882                         /*
1883                          * We raced with another reader/writer and haven't
1884                          * managed to process any buffers.  A zero return
1885                          * value means EOF, so retry instead.
1886                          */
1887                         pipe_unlock(ipipe);
1888                         pipe_unlock(opipe);
1889                         goto retry;
1890                 }
1891
1892                 ibuf = ipipe->bufs + ipipe->curbuf;
1893                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1894                 obuf = opipe->bufs + nbuf;
1895
1896                 if (len >= ibuf->len) {
1897                         /*
1898                          * Simply move the whole buffer from ipipe to opipe
1899                          */
1900                         *obuf = *ibuf;
1901                         ibuf->ops = NULL;
1902                         opipe->nrbufs++;
1903                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1904                         ipipe->nrbufs--;
1905                         input_wakeup = true;
1906                 } else {
1907                         /*
1908                          * Get a reference to this pipe buffer,
1909                          * so we can copy the contents over.
1910                          */
1911                         ibuf->ops->get(ipipe, ibuf);
1912                         *obuf = *ibuf;
1913
1914                         /*
1915                          * Don't inherit the gift flag, we need to
1916                          * prevent multiple steals of this page.
1917                          */
1918                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1919
1920                         obuf->len = len;
1921                         opipe->nrbufs++;
1922                         ibuf->offset += obuf->len;
1923                         ibuf->len -= obuf->len;
1924                 }
1925                 ret += obuf->len;
1926                 len -= obuf->len;
1927         } while (len);
1928
1929         pipe_unlock(ipipe);
1930         pipe_unlock(opipe);
1931
1932         /*
1933          * If we put data in the output pipe, wakeup any potential readers.
1934          */
1935         if (ret > 0) {
1936                 smp_mb();
1937                 if (waitqueue_active(&opipe->wait))
1938                         wake_up_interruptible(&opipe->wait);
1939                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1940         }
1941         if (input_wakeup)
1942                 wakeup_pipe_writers(ipipe);
1943
1944         return ret;
1945 }
1946
1947 /*
1948  * Link contents of ipipe to opipe.
1949  */
1950 static int link_pipe(struct pipe_inode_info *ipipe,
1951                      struct pipe_inode_info *opipe,
1952                      size_t len, unsigned int flags)
1953 {
1954         struct pipe_buffer *ibuf, *obuf;
1955         int ret = 0, i = 0, nbuf;
1956
1957         /*
1958          * Potential ABBA deadlock, work around it by ordering lock
1959          * grabbing by pipe info address. Otherwise two different processes
1960          * could deadlock (one doing tee from A -> B, the other from B -> A).
1961          */
1962         pipe_double_lock(ipipe, opipe);
1963
1964         do {
1965                 if (!opipe->readers) {
1966                         send_sig(SIGPIPE, current, 0);
1967                         if (!ret)
1968                                 ret = -EPIPE;
1969                         break;
1970                 }
1971
1972                 /*
1973                  * If we have iterated all input buffers or ran out of
1974                  * output room, break.
1975                  */
1976                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1977                         break;
1978
1979                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1980                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1981
1982                 /*
1983                  * Get a reference to this pipe buffer,
1984                  * so we can copy the contents over.
1985                  */
1986                 ibuf->ops->get(ipipe, ibuf);
1987
1988                 obuf = opipe->bufs + nbuf;
1989                 *obuf = *ibuf;
1990
1991                 /*
1992                  * Don't inherit the gift flag, we need to
1993                  * prevent multiple steals of this page.
1994                  */
1995                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1996
1997                 if (obuf->len > len)
1998                         obuf->len = len;
1999
2000                 opipe->nrbufs++;
2001                 ret += obuf->len;
2002                 len -= obuf->len;
2003                 i++;
2004         } while (len);
2005
2006         /*
2007          * return EAGAIN if we have the potential of some data in the
2008          * future, otherwise just return 0
2009          */
2010         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2011                 ret = -EAGAIN;
2012
2013         pipe_unlock(ipipe);
2014         pipe_unlock(opipe);
2015
2016         /*
2017          * If we put data in the output pipe, wakeup any potential readers.
2018          */
2019         if (ret > 0) {
2020                 smp_mb();
2021                 if (waitqueue_active(&opipe->wait))
2022                         wake_up_interruptible(&opipe->wait);
2023                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2024         }
2025
2026         return ret;
2027 }
2028
2029 /*
2030  * This is a tee(1) implementation that works on pipes. It doesn't copy
2031  * any data, it simply references the 'in' pages on the 'out' pipe.
2032  * The 'flags' used are the SPLICE_F_* variants, currently the only
2033  * applicable one is SPLICE_F_NONBLOCK.
2034  */
2035 static long do_tee(struct file *in, struct file *out, size_t len,
2036                    unsigned int flags)
2037 {
2038         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2039         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2040         int ret = -EINVAL;
2041
2042         /*
2043          * Duplicate the contents of ipipe to opipe without actually
2044          * copying the data.
2045          */
2046         if (ipipe && opipe && ipipe != opipe) {
2047                 /*
2048                  * Keep going, unless we encounter an error. The ipipe/opipe
2049                  * ordering doesn't really matter.
2050                  */
2051                 ret = ipipe_prep(ipipe, flags);
2052                 if (!ret) {
2053                         ret = opipe_prep(opipe, flags);
2054                         if (!ret)
2055                                 ret = link_pipe(ipipe, opipe, len, flags);
2056                 }
2057         }
2058
2059         return ret;
2060 }
2061
2062 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2063 {
2064         struct file *in;
2065         int error, fput_in;
2066
2067         if (unlikely(!len))
2068                 return 0;
2069
2070         error = -EBADF;
2071         in = fget_light(fdin, &fput_in);
2072         if (in) {
2073                 if (in->f_mode & FMODE_READ) {
2074                         int fput_out;
2075                         struct file *out = fget_light(fdout, &fput_out);
2076
2077                         if (out) {
2078                                 if (out->f_mode & FMODE_WRITE)
2079                                         error = do_tee(in, out, len, flags);
2080                                 fput_light(out, fput_out);
2081                         }
2082                 }
2083                 fput_light(in, fput_in);
2084         }
2085
2086         return error;
2087 }