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         ssize_t res;
 601         size_t this_len;
 602         int error;
 603         int i;
 604         struct splice_pipe_desc spd = {
 605                 .pages = pages,
 606                 .partial = partial,
 607                 .flags = flags,
 608                 .ops = &default_pipe_buf_ops,
 609                 .spd_release = spd_release_page,
 610         };
 611
 612         if (splice_grow_spd(pipe, &spd))
 613                 return -ENOMEM;
 614
 615         res = -ENOMEM;
 616         vec = __vec;
 617         if (pipe->buffers > PIPE_DEF_BUFFERS) {
 618                 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
 619                 if (!vec)
 620                         goto shrink_ret;
 621         }
 622
 623         offset = *ppos & ~PAGE_CACHE_MASK;
 624         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 625
 626         for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
 627                 struct page *page;
 628
 629                 page = alloc_page(GFP_USER);
 630                 error = -ENOMEM;
 631                 if (!page)
 632                         goto err;
 633
 634                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 635                 vec[i].iov_base = (void __user *) page_address(page);
 636                 vec[i].iov_len = this_len;
 637                 spd.pages[i] = page;
 638                 spd.nr_pages++;
 639                 len -= this_len;
 640                 offset = 0;
 641         }
 642
 643         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 644         if (res < 0) {
 645                 error = res;
 646                 goto err;
 647         }
 648
 649         error = 0;
 650         if (!res)
 651                 goto err;
 652
 653         nr_freed = 0;
 654         for (i = 0; i < spd.nr_pages; i++) {
 655                 this_len = min_t(size_t, vec[i].iov_len, res);
 656                 spd.partial[i].offset = 0;
 657                 spd.partial[i].len = this_len;
 658                 if (!this_len) {
 659                         __free_page(spd.pages[i]);
 660                         spd.pages[i] = NULL;
 661                         nr_freed++;
 662                 }
 663                 res -= this_len;
 664         }
 665         spd.nr_pages -= nr_freed;
 666
 667         res = splice_to_pipe(pipe, &spd);
 668         if (res > 0)
 669                 *ppos += res;
 670
 671 shrink_ret:
 672         if (vec != __vec)
 673                 kfree(vec);
 674         splice_shrink_spd(pipe, &spd);
 675         return res;
 676
 677 err:
 678         for (i = 0; i < spd.nr_pages; i++)
 679                 __free_page(spd.pages[i]);
 680
 681         res = error;
 682         goto shrink_ret;
 683 }
 684 EXPORT_SYMBOL(default_file_splice_read);
 685
 686 /*
 687  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 688  * using sendpage(). Return the number of bytes sent.
 689  */
 690 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 691                             struct pipe_buffer *buf, struct splice_desc *sd)
 692 {
 693         struct file *file = sd->u.file;
 694         loff_t pos = sd->pos;
 695         int ret, more;
 696
 697         ret = buf->ops->confirm(pipe, buf);
 698         if (!ret) {
 699                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 700                 if (file->f_op && file->f_op->sendpage)
 701                         ret = file->f_op->sendpage(file, buf->page, buf->offset,
 702                                                    sd->len, &pos, more);
 703                 else
 704                         ret = -EINVAL;
 705         }
 706
 707         return ret;
 708 }
 709
 710 /*
 711  * This is a little more tricky than the file -> pipe splicing. There are
 712  * basically three cases:
 713  *
 714  *      - Destination page already exists in the address space and there
 715  *        are users of it. For that case we have no other option that
 716  *        copying the data. Tough luck.
 717  *      - Destination page already exists in the address space, but there
 718  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 719  *        through to last case.
 720  *      - Destination page does not exist, we can add the pipe page to
 721  *        the page cache and avoid the copy.
 722  *
 723  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 724  * sd->flags), we attempt to migrate pages from the pipe to the output
 725  * file address space page cache. This is possible if no one else has
 726  * the pipe page referenced outside of the pipe and page cache. If
 727  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 728  * a new page in the output file page cache and fill/dirty that.
 729  */
 730 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 731                  struct splice_desc *sd)
 732 {
 733         struct file *file = sd->u.file;
 734         struct address_space *mapping = file->f_mapping;
 735         unsigned int offset, this_len;
 736         struct page *page;
 737         void *fsdata;
 738         int ret;
 739
 740         /*
 741          * make sure the data in this buffer is uptodate
 742          */
 743         ret = buf->ops->confirm(pipe, buf);
 744         if (unlikely(ret))
 745                 return ret;
 746
 747         offset = sd->pos & ~PAGE_CACHE_MASK;
 748
 749         this_len = sd->len;
 750         if (this_len + offset > PAGE_CACHE_SIZE)
 751                 this_len = PAGE_CACHE_SIZE - offset;
 752
 753         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 754                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 755         if (unlikely(ret))
 756                 goto out;
 757
 758         if (buf->page != page) {
 759                 /*
 760                  * Careful, ->map() uses KM_USER0!
 761                  */
 762                 char *src = buf->ops->map(pipe, buf, 1);
 763                 char *dst = kmap_atomic(page, KM_USER1);
 764
 765                 memcpy(dst + offset, src + buf->offset, this_len);
 766                 flush_dcache_page(page);
 767                 kunmap_atomic(dst, KM_USER1);
 768                 buf->ops->unmap(pipe, buf, src);
 769         }
 770         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 771                                 page, fsdata);
 772 out:
 773         return ret;
 774 }
 775 EXPORT_SYMBOL(pipe_to_file);
 776
 777 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 778 {
 779         smp_mb();
 780         if (waitqueue_active(&pipe->wait))
 781                 wake_up_interruptible(&pipe->wait);
 782         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 783 }
 784
 785 /**
 786  * splice_from_pipe_feed - feed available data from a pipe to a file
 787  * @pipe:       pipe to splice from
 788  * @sd:         information to @actor
 789  * @actor:      handler that splices the data
 790  *
 791  * Description:
 792  *    This function loops over the pipe and calls @actor to do the
 793  *    actual moving of a single struct pipe_buffer to the desired
 794  *    destination.  It returns when there's no more buffers left in
 795  *    the pipe or if the requested number of bytes (@sd->total_len)
 796  *    have been copied.  It returns a positive number (one) if the
 797  *    pipe needs to be filled with more data, zero if the required
 798  *    number of bytes have been copied and -errno on error.
 799  *
 800  *    This, together with splice_from_pipe_{begin,end,next}, may be
 801  *    used to implement the functionality of __splice_from_pipe() when
 802  *    locking is required around copying the pipe buffers to the
 803  *    destination.
 804  */
 805 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 806                           splice_actor *actor)
 807 {
 808         int ret;
 809
 810         while (pipe->nrbufs) {
 811                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 812                 const struct pipe_buf_operations *ops = buf->ops;
 813
 814                 sd->len = buf->len;
 815                 if (sd->len > sd->total_len)
 816                         sd->len = sd->total_len;
 817
 818                 ret = actor(pipe, buf, sd);
 819                 if (ret <= 0) {
 820                         if (ret == -ENODATA)
 821                                 ret = 0;
 822                         return ret;
 823                 }
 824                 buf->offset += ret;
 825                 buf->len -= ret;
 826
 827                 sd->num_spliced += ret;
 828                 sd->len -= ret;
 829                 sd->pos += ret;
 830                 sd->total_len -= ret;
 831
 832                 if (!buf->len) {
 833                         buf->ops = NULL;
 834                         ops->release(pipe, buf);
 835                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 836                         pipe->nrbufs--;
 837                         if (pipe->inode)
 838                                 sd->need_wakeup = true;
 839                 }
 840
 841                 if (!sd->total_len)
 842                         return 0;
 843         }
 844
 845         return 1;
 846 }
 847 EXPORT_SYMBOL(splice_from_pipe_feed);
 848
 849 /**
 850  * splice_from_pipe_next - wait for some data to splice from
 851  * @pipe:       pipe to splice from
 852  * @sd:         information about the splice operation
 853  *
 854  * Description:
 855  *    This function will wait for some data and return a positive
 856  *    value (one) if pipe buffers are available.  It will return zero
 857  *    or -errno if no more data needs to be spliced.
 858  */
 859 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 860 {
 861         while (!pipe->nrbufs) {
 862                 if (!pipe->writers)
 863                         return 0;
 864
 865                 if (!pipe->waiting_writers && sd->num_spliced)
 866                         return 0;
 867
 868                 if (sd->flags & SPLICE_F_NONBLOCK)
 869                         return -EAGAIN;
 870
 871                 if (signal_pending(current))
 872                         return -ERESTARTSYS;
 873
 874                 if (sd->need_wakeup) {
 875                         wakeup_pipe_writers(pipe);
 876                         sd->need_wakeup = false;
 877                 }
 878
 879                 pipe_wait(pipe);
 880         }
 881
 882         return 1;
 883 }
 884 EXPORT_SYMBOL(splice_from_pipe_next);
 885
 886 /**
 887  * splice_from_pipe_begin - start splicing from pipe
 888  * @sd:         information about the splice operation
 889  *
 890  * Description:
 891  *    This function should be called before a loop containing
 892  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 893  *    initialize the necessary fields of @sd.
 894  */
 895 void splice_from_pipe_begin(struct splice_desc *sd)
 896 {
 897         sd->num_spliced = 0;
 898         sd->need_wakeup = false;
 899 }
 900 EXPORT_SYMBOL(splice_from_pipe_begin);
 901
 902 /**
 903  * splice_from_pipe_end - finish splicing from pipe
 904  * @pipe:       pipe to splice from
 905  * @sd:         information about the splice operation
 906  *
 907  * Description:
 908  *    This function will wake up pipe writers if necessary.  It should
 909  *    be called after a loop containing splice_from_pipe_next() and
 910  *    splice_from_pipe_feed().
 911  */
 912 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 913 {
 914         if (sd->need_wakeup)
 915                 wakeup_pipe_writers(pipe);
 916 }
 917 EXPORT_SYMBOL(splice_from_pipe_end);
 918
 919 /**
 920  * __splice_from_pipe - splice data from a pipe to given actor
 921  * @pipe:       pipe to splice from
 922  * @sd:         information to @actor
 923  * @actor:      handler that splices the data
 924  *
 925  * Description:
 926  *    This function does little more than loop over the pipe and call
 927  *    @actor to do the actual moving of a single struct pipe_buffer to
 928  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 929  *    pipe_to_user.
 930  *
 931  */
 932 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 933                            splice_actor *actor)
 934 {
 935         int ret;
 936
 937         splice_from_pipe_begin(sd);
 938         do {
 939                 ret = splice_from_pipe_next(pipe, sd);
 940                 if (ret > 0)
 941                         ret = splice_from_pipe_feed(pipe, sd, actor);
 942         } while (ret > 0);
 943         splice_from_pipe_end(pipe, sd);
 944
 945         return sd->num_spliced ? sd->num_spliced : ret;
 946 }
 947 EXPORT_SYMBOL(__splice_from_pipe);
 948
 949 /**
 950  * splice_from_pipe - splice data from a pipe to a file
 951  * @pipe:       pipe to splice from
 952  * @out:        file to splice to
 953  * @ppos:       position in @out
 954  * @len:        how many bytes to splice
 955  * @flags:      splice modifier flags
 956  * @actor:      handler that splices the data
 957  *
 958  * Description:
 959  *    See __splice_from_pipe. This function locks the pipe inode,
 960  *    otherwise it's identical to __splice_from_pipe().
 961  *
 962  */
 963 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 964                          loff_t *ppos, size_t len, unsigned int flags,
 965                          splice_actor *actor)
 966 {
 967         ssize_t ret;
 968         struct splice_desc sd = {
 969                 .total_len = len,
 970                 .flags = flags,
 971                 .pos = *ppos,
 972                 .u.file = out,
 973         };
 974
 975         pipe_lock(pipe);
 976         ret = __splice_from_pipe(pipe, &sd, actor);
 977         pipe_unlock(pipe);
 978
 979         return ret;
 980 }
 981
 982 /**
 983  * generic_file_splice_write - splice data from a pipe to a file
 984  * @pipe:       pipe info
 985  * @out:        file to write to
 986  * @ppos:       position in @out
 987  * @len:        number of bytes to splice
 988  * @flags:      splice modifier flags
 989  *
 990  * Description:
 991  *    Will either move or copy pages (determined by @flags options) from
 992  *    the given pipe inode to the given file.
 993  *
 994  */
 995 ssize_t
 996 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 997                           loff_t *ppos, size_t len, unsigned int flags)
 998 {
 999         struct address_space *mapping = out->f_mapping;
1000         struct inode *inode = mapping->host;
1001         struct splice_desc sd = {
1002                 .total_len = len,
1003                 .flags = flags,
1004                 .pos = *ppos,
1005                 .u.file = out,
1006         };
1007         ssize_t ret;
1008
1009         pipe_lock(pipe);
1010
1011         splice_from_pipe_begin(&sd);
1012         do {
1013                 ret = splice_from_pipe_next(pipe, &sd);
1014                 if (ret <= 0)
1015                         break;
1016
1017                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1018                 ret = file_remove_suid(out);
1019                 if (!ret) {
1020                         file_update_time(out);
1021                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1022                 }
1023                 mutex_unlock(&inode->i_mutex);
1024         } while (ret > 0);
1025         splice_from_pipe_end(pipe, &sd);
1026
1027         pipe_unlock(pipe);
1028
1029         if (sd.num_spliced)
1030                 ret = sd.num_spliced;
1031
1032         if (ret > 0) {
1033                 unsigned long nr_pages;
1034                 int err;
1035
1036                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1037
1038                 err = generic_write_sync(out, *ppos, ret);
1039                 if (err)
1040                         ret = err;
1041                 else
1042                         *ppos += ret;
1043                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1044         }
1045
1046         return ret;
1047 }
1048
1049 EXPORT_SYMBOL(generic_file_splice_write);
1050
1051 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1052                           struct splice_desc *sd)
1053 {
1054         int ret;
1055         void *data;
1056
1057         ret = buf->ops->confirm(pipe, buf);
1058         if (ret)
1059                 return ret;
1060
1061         data = buf->ops->map(pipe, buf, 0);
1062         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1063         buf->ops->unmap(pipe, buf, data);
1064
1065         return ret;
1066 }
1067
1068 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1069                                          struct file *out, loff_t *ppos,
1070                                          size_t len, unsigned int flags)
1071 {
1072         ssize_t ret;
1073
1074         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1075         if (ret > 0)
1076                 *ppos += ret;
1077
1078         return ret;
1079 }
1080
1081 /**
1082  * generic_splice_sendpage - splice data from a pipe to a socket
1083  * @pipe:       pipe to splice from
1084  * @out:        socket to write to
1085  * @ppos:       position in @out
1086  * @len:        number of bytes to splice
1087  * @flags:      splice modifier flags
1088  *
1089  * Description:
1090  *    Will send @len bytes from the pipe to a network socket. No data copying
1091  *    is involved.
1092  *
1093  */
1094 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1095                                 loff_t *ppos, size_t len, unsigned int flags)
1096 {
1097         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1098 }
1099
1100 EXPORT_SYMBOL(generic_splice_sendpage);
1101
1102 /*
1103  * Attempt to initiate a splice from pipe to file.
1104  */
1105 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1106                            loff_t *ppos, size_t len, unsigned int flags)
1107 {
1108         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1109                                 loff_t *, size_t, unsigned int);
1110         int ret;
1111
1112         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1113                 return -EBADF;
1114
1115         if (unlikely(out->f_flags & O_APPEND))
1116                 return -EINVAL;
1117
1118         ret = rw_verify_area(WRITE, out, ppos, len);
1119         if (unlikely(ret < 0))
1120                 return ret;
1121
1122         if (out->f_op && out->f_op->splice_write)
1123                 splice_write = out->f_op->splice_write;
1124         else
1125                 splice_write = default_file_splice_write;
1126
1127         return splice_write(pipe, out, ppos, len, flags);
1128 }
1129
1130 /*
1131  * Attempt to initiate a splice from a file to a pipe.
1132  */
1133 static long do_splice_to(struct file *in, loff_t *ppos,
1134                          struct pipe_inode_info *pipe, size_t len,
1135                          unsigned int flags)
1136 {
1137         ssize_t (*splice_read)(struct file *, loff_t *,
1138                                struct pipe_inode_info *, size_t, unsigned int);
1139         int ret;
1140
1141         if (unlikely(!(in->f_mode & FMODE_READ)))
1142                 return -EBADF;
1143
1144         ret = rw_verify_area(READ, in, ppos, len);
1145         if (unlikely(ret < 0))
1146                 return ret;
1147
1148         if (in->f_op && in->f_op->splice_read)
1149                 splice_read = in->f_op->splice_read;
1150         else
1151                 splice_read = default_file_splice_read;
1152
1153         return splice_read(in, ppos, pipe, len, flags);
1154 }
1155
1156 /**
1157  * splice_direct_to_actor - splices data directly between two non-pipes
1158  * @in:         file to splice from
1159  * @sd:         actor information on where to splice to
1160  * @actor:      handles the data splicing
1161  *
1162  * Description:
1163  *    This is a special case helper to splice directly between two
1164  *    points, without requiring an explicit pipe. Internally an allocated
1165  *    pipe is cached in the process, and reused during the lifetime of
1166  *    that process.
1167  *
1168  */
1169 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1170                                splice_direct_actor *actor)
1171 {
1172         struct pipe_inode_info *pipe;
1173         long ret, bytes;
1174         umode_t i_mode;
1175         size_t len;
1176         int i, flags;
1177
1178         /*
1179          * We require the input being a regular file, as we don't want to
1180          * randomly drop data for eg socket -> socket splicing. Use the
1181          * piped splicing for that!
1182          */
1183         i_mode = in->f_path.dentry->d_inode->i_mode;
1184         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1185                 return -EINVAL;
1186
1187         /*
1188          * neither in nor out is a pipe, setup an internal pipe attached to
1189          * 'out' and transfer the wanted data from 'in' to 'out' through that
1190          */
1191         pipe = current->splice_pipe;
1192         if (unlikely(!pipe)) {
1193                 pipe = alloc_pipe_info(NULL);
1194                 if (!pipe)
1195                         return -ENOMEM;
1196
1197                 /*
1198                  * We don't have an immediate reader, but we'll read the stuff
1199                  * out of the pipe right after the splice_to_pipe(). So set
1200                  * PIPE_READERS appropriately.
1201                  */
1202                 pipe->readers = 1;
1203
1204                 current->splice_pipe = pipe;
1205         }
1206
1207         /*
1208          * Do the splice.
1209          */
1210         ret = 0;
1211         bytes = 0;
1212         len = sd->total_len;
1213         flags = sd->flags;
1214
1215         /*
1216          * Don't block on output, we have to drain the direct pipe.
1217          */
1218         sd->flags &= ~SPLICE_F_NONBLOCK;
1219
1220         while (len) {
1221                 size_t read_len;
1222                 loff_t pos = sd->pos, prev_pos = pos;
1223
1224                 ret = do_splice_to(in, &pos, pipe, len, flags);
1225                 if (unlikely(ret <= 0))
1226                         goto out_release;
1227
1228                 read_len = ret;
1229                 sd->total_len = read_len;
1230
1231                 /*
1232                  * NOTE: nonblocking mode only applies to the input. We
1233                  * must not do the output in nonblocking mode as then we
1234                  * could get stuck data in the internal pipe:
1235                  */
1236                 ret = actor(pipe, sd);
1237                 if (unlikely(ret <= 0)) {
1238                         sd->pos = prev_pos;
1239                         goto out_release;
1240                 }
1241
1242                 bytes += ret;
1243                 len -= ret;
1244                 sd->pos = pos;
1245
1246                 if (ret < read_len) {
1247                         sd->pos = prev_pos + ret;
1248                         goto out_release;
1249                 }
1250         }
1251
1252 done:
1253         pipe->nrbufs = pipe->curbuf = 0;
1254         file_accessed(in);
1255         return bytes;
1256
1257 out_release:
1258         /*
1259          * If we did an incomplete transfer we must release
1260          * the pipe buffers in question:
1261          */
1262         for (i = 0; i < pipe->buffers; i++) {
1263                 struct pipe_buffer *buf = pipe->bufs + i;
1264
1265                 if (buf->ops) {
1266                         buf->ops->release(pipe, buf);
1267                         buf->ops = NULL;
1268                 }
1269         }
1270
1271         if (!bytes)
1272                 bytes = ret;
1273
1274         goto done;
1275 }
1276 EXPORT_SYMBOL(splice_direct_to_actor);
1277
1278 static int direct_splice_actor(struct pipe_inode_info *pipe,
1279                                struct splice_desc *sd)
1280 {
1281         struct file *file = sd->u.file;
1282
1283         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1284                               sd->flags);
1285 }
1286
1287 /**
1288  * do_splice_direct - splices data directly between two files
1289  * @in:         file to splice from
1290  * @ppos:       input file offset
1291  * @out:        file to splice to
1292  * @len:        number of bytes to splice
1293  * @flags:      splice modifier flags
1294  *
1295  * Description:
1296  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1297  *    doing it in the application would incur an extra system call
1298  *    (splice in + splice out, as compared to just sendfile()). So this helper
1299  *    can splice directly through a process-private pipe.
1300  *
1301  */
1302 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1303                       size_t len, unsigned int flags)
1304 {
1305         struct splice_desc sd = {
1306                 .len            = len,
1307                 .total_len      = len,
1308                 .flags          = flags,
1309                 .pos            = *ppos,
1310                 .u.file         = out,
1311         };
1312         long ret;
1313
1314         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1315         if (ret > 0)
1316                 *ppos = sd.pos;
1317
1318         return ret;
1319 }
1320
1321 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1322                                struct pipe_inode_info *opipe,
1323                                size_t len, unsigned int flags);
1324 /*
1325  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1326  * location, so checking ->i_pipe is not enough to verify that this is a
1327  * pipe.
1328  */
1329 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1330 {
1331         if (S_ISFIFO(inode->i_mode))
1332                 return inode->i_pipe;
1333
1334         return NULL;
1335 }
1336
1337 /*
1338  * Determine where to splice to/from.
1339  */
1340 static long do_splice(struct file *in, loff_t __user *off_in,
1341                       struct file *out, loff_t __user *off_out,
1342                       size_t len, unsigned int flags)
1343 {
1344         struct pipe_inode_info *ipipe;
1345         struct pipe_inode_info *opipe;
1346         loff_t offset, *off;
1347         long ret;
1348
1349         ipipe = pipe_info(in->f_path.dentry->d_inode);
1350         opipe = pipe_info(out->f_path.dentry->d_inode);
1351
1352         if (ipipe && opipe) {
1353                 if (off_in || off_out)
1354                         return -ESPIPE;
1355
1356                 if (!(in->f_mode & FMODE_READ))
1357                         return -EBADF;
1358
1359                 if (!(out->f_mode & FMODE_WRITE))
1360                         return -EBADF;
1361
1362                 /* Splicing to self would be fun, but... */
1363                 if (ipipe == opipe)
1364                         return -EINVAL;
1365
1366                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1367         }
1368
1369         if (ipipe) {
1370                 if (off_in)
1371                         return -ESPIPE;
1372                 if (off_out) {
1373                         if (!(out->f_mode & FMODE_PWRITE))
1374                                 return -EINVAL;
1375                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1376                                 return -EFAULT;
1377                         off = &offset;
1378                 } else
1379                         off = &out->f_pos;
1380
1381                 ret = do_splice_from(ipipe, out, off, len, flags);
1382
1383                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1384                         ret = -EFAULT;
1385
1386                 return ret;
1387         }
1388
1389         if (opipe) {
1390                 if (off_out)
1391                         return -ESPIPE;
1392                 if (off_in) {
1393                         if (!(in->f_mode & FMODE_PREAD))
1394                                 return -EINVAL;
1395                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1396                                 return -EFAULT;
1397                         off = &offset;
1398                 } else
1399                         off = &in->f_pos;
1400
1401                 ret = do_splice_to(in, off, opipe, len, flags);
1402
1403                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1404                         ret = -EFAULT;
1405
1406                 return ret;
1407         }
1408
1409         return -EINVAL;
1410 }
1411
1412 /*
1413  * Map an iov into an array of pages and offset/length tupples. With the
1414  * partial_page structure, we can map several non-contiguous ranges into
1415  * our ones pages[] map instead of splitting that operation into pieces.
1416  * Could easily be exported as a generic helper for other users, in which
1417  * case one would probably want to add a 'max_nr_pages' parameter as well.
1418  */
1419 static int get_iovec_page_array(const struct iovec __user *iov,
1420                                 unsigned int nr_vecs, struct page **pages,
1421                                 struct partial_page *partial, int aligned,
1422                                 unsigned int pipe_buffers)
1423 {
1424         int buffers = 0, error = 0;
1425
1426         while (nr_vecs) {
1427                 unsigned long off, npages;
1428                 struct iovec entry;
1429                 void __user *base;
1430                 size_t len;
1431                 int i;
1432
1433                 error = -EFAULT;
1434                 if (copy_from_user(&entry, iov, sizeof(entry)))
1435                         break;
1436
1437                 base = entry.iov_base;
1438                 len = entry.iov_len;
1439
1440                 /*
1441                  * Sanity check this iovec. 0 read succeeds.
1442                  */
1443                 error = 0;
1444                 if (unlikely(!len))
1445                         break;
1446                 error = -EFAULT;
1447                 if (!access_ok(VERIFY_READ, base, len))
1448                         break;
1449
1450                 /*
1451                  * Get this base offset and number of pages, then map
1452                  * in the user pages.
1453                  */
1454                 off = (unsigned long) base & ~PAGE_MASK;
1455
1456                 /*
1457                  * If asked for alignment, the offset must be zero and the
1458                  * length a multiple of the PAGE_SIZE.
1459                  */
1460                 error = -EINVAL;
1461                 if (aligned && (off || len & ~PAGE_MASK))
1462                         break;
1463
1464                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1465                 if (npages > pipe_buffers - buffers)
1466                         npages = pipe_buffers - buffers;
1467
1468                 error = get_user_pages_fast((unsigned long)base, npages,
1469                                         0, &pages[buffers]);
1470
1471                 if (unlikely(error <= 0))
1472                         break;
1473
1474                 /*
1475                  * Fill this contiguous range into the partial page map.
1476                  */
1477                 for (i = 0; i < error; i++) {
1478                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1479
1480                         partial[buffers].offset = off;
1481                         partial[buffers].len = plen;
1482
1483                         off = 0;
1484                         len -= plen;
1485                         buffers++;
1486                 }
1487
1488                 /*
1489                  * We didn't complete this iov, stop here since it probably
1490                  * means we have to move some of this into a pipe to
1491                  * be able to continue.
1492                  */
1493                 if (len)
1494                         break;
1495
1496                 /*
1497                  * Don't continue if we mapped fewer pages than we asked for,
1498                  * or if we mapped the max number of pages that we have
1499                  * room for.
1500                  */
1501                 if (error < npages || buffers == pipe_buffers)
1502                         break;
1503
1504                 nr_vecs--;
1505                 iov++;
1506         }
1507
1508         if (buffers)
1509                 return buffers;
1510
1511         return error;
1512 }
1513
1514 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1515                         struct splice_desc *sd)
1516 {
1517         char *src;
1518         int ret;
1519
1520         ret = buf->ops->confirm(pipe, buf);
1521         if (unlikely(ret))
1522                 return ret;
1523
1524         /*
1525          * See if we can use the atomic maps, by prefaulting in the
1526          * pages and doing an atomic copy
1527          */
1528         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1529                 src = buf->ops->map(pipe, buf, 1);
1530                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1531                                                         sd->len);
1532                 buf->ops->unmap(pipe, buf, src);
1533                 if (!ret) {
1534                         ret = sd->len;
1535                         goto out;
1536                 }
1537         }
1538
1539         /*
1540          * No dice, use slow non-atomic map and copy
1541          */
1542         src = buf->ops->map(pipe, buf, 0);
1543
1544         ret = sd->len;
1545         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1546                 ret = -EFAULT;
1547
1548         buf->ops->unmap(pipe, buf, src);
1549 out:
1550         if (ret > 0)
1551                 sd->u.userptr += ret;
1552         return ret;
1553 }
1554
1555 /*
1556  * For lack of a better implementation, implement vmsplice() to userspace
1557  * as a simple copy of the pipes pages to the user iov.
1558  */
1559 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1560                              unsigned long nr_segs, unsigned int flags)
1561 {
1562         struct pipe_inode_info *pipe;
1563         struct splice_desc sd;
1564         ssize_t size;
1565         int error;
1566         long ret;
1567
1568         pipe = pipe_info(file->f_path.dentry->d_inode);
1569         if (!pipe)
1570                 return -EBADF;
1571
1572         pipe_lock(pipe);
1573
1574         error = ret = 0;
1575         while (nr_segs) {
1576                 void __user *base;
1577                 size_t len;
1578
1579                 /*
1580                  * Get user address base and length for this iovec.
1581                  */
1582                 error = get_user(base, &iov->iov_base);
1583                 if (unlikely(error))
1584                         break;
1585                 error = get_user(len, &iov->iov_len);
1586                 if (unlikely(error))
1587                         break;
1588
1589                 /*
1590                  * Sanity check this iovec. 0 read succeeds.
1591                  */
1592                 if (unlikely(!len))
1593                         break;
1594                 if (unlikely(!base)) {
1595                         error = -EFAULT;
1596                         break;
1597                 }
1598
1599                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1600                         error = -EFAULT;
1601                         break;
1602                 }
1603
1604                 sd.len = 0;
1605                 sd.total_len = len;
1606                 sd.flags = flags;
1607                 sd.u.userptr = base;
1608                 sd.pos = 0;
1609
1610                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1611                 if (size < 0) {
1612                         if (!ret)
1613                                 ret = size;
1614
1615                         break;
1616                 }
1617
1618                 ret += size;
1619
1620                 if (size < len)
1621                         break;
1622
1623                 nr_segs--;
1624                 iov++;
1625         }
1626
1627         pipe_unlock(pipe);
1628
1629         if (!ret)
1630                 ret = error;
1631
1632         return ret;
1633 }
1634
1635 /*
1636  * vmsplice splices a user address range into a pipe. It can be thought of
1637  * as splice-from-memory, where the regular splice is splice-from-file (or
1638  * to file). In both cases the output is a pipe, naturally.
1639  */
1640 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1641                              unsigned long nr_segs, unsigned int flags)
1642 {
1643         struct pipe_inode_info *pipe;
1644         struct page *pages[PIPE_DEF_BUFFERS];
1645         struct partial_page partial[PIPE_DEF_BUFFERS];
1646         struct splice_pipe_desc spd = {
1647                 .pages = pages,
1648                 .partial = partial,
1649                 .flags = flags,
1650                 .ops = &user_page_pipe_buf_ops,
1651                 .spd_release = spd_release_page,
1652         };
1653         long ret;
1654
1655         pipe = pipe_info(file->f_path.dentry->d_inode);
1656         if (!pipe)
1657                 return -EBADF;
1658
1659         if (splice_grow_spd(pipe, &spd))
1660                 return -ENOMEM;
1661
1662         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1663                                             spd.partial, flags & SPLICE_F_GIFT,
1664                                             pipe->buffers);
1665         if (spd.nr_pages <= 0)
1666                 ret = spd.nr_pages;
1667         else
1668                 ret = splice_to_pipe(pipe, &spd);
1669
1670         splice_shrink_spd(pipe, &spd);
1671         return ret;
1672 }
1673
1674 /*
1675  * Note that vmsplice only really supports true splicing _from_ user memory
1676  * to a pipe, not the other way around. Splicing from user memory is a simple
1677  * operation that can be supported without any funky alignment restrictions
1678  * or nasty vm tricks. We simply map in the user memory and fill them into
1679  * a pipe. The reverse isn't quite as easy, though. There are two possible
1680  * solutions for that:
1681  *
1682  *      - memcpy() the data internally, at which point we might as well just
1683  *        do a regular read() on the buffer anyway.
1684  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1685  *        has restriction limitations on both ends of the pipe).
1686  *
1687  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1688  *
1689  */
1690 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1691                 unsigned long, nr_segs, unsigned int, flags)
1692 {
1693         struct file *file;
1694         long error;
1695         int fput;
1696
1697         if (unlikely(nr_segs > UIO_MAXIOV))
1698                 return -EINVAL;
1699         else if (unlikely(!nr_segs))
1700                 return 0;
1701
1702         error = -EBADF;
1703         file = fget_light(fd, &fput);
1704         if (file) {
1705                 if (file->f_mode & FMODE_WRITE)
1706                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1707                 else if (file->f_mode & FMODE_READ)
1708                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1709
1710                 fput_light(file, fput);
1711         }
1712
1713         return error;
1714 }
1715
1716 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1717                 int, fd_out, loff_t __user *, off_out,
1718                 size_t, len, unsigned int, flags)
1719 {
1720         long error;
1721         struct file *in, *out;
1722         int fput_in, fput_out;
1723
1724         if (unlikely(!len))
1725                 return 0;
1726
1727         error = -EBADF;
1728         in = fget_light(fd_in, &fput_in);
1729         if (in) {
1730                 if (in->f_mode & FMODE_READ) {
1731                         out = fget_light(fd_out, &fput_out);
1732                         if (out) {
1733                                 if (out->f_mode & FMODE_WRITE)
1734                                         error = do_splice(in, off_in,
1735                                                           out, off_out,
1736                                                           len, flags);
1737                                 fput_light(out, fput_out);
1738                         }
1739                 }
1740
1741                 fput_light(in, fput_in);
1742         }
1743
1744         return error;
1745 }
1746
1747 /*
1748  * Make sure there's data to read. Wait for input if we can, otherwise
1749  * return an appropriate error.
1750  */
1751 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1752 {
1753         int ret;
1754
1755         /*
1756          * Check ->nrbufs without the inode lock first. This function
1757          * is speculative anyways, so missing one is ok.
1758          */
1759         if (pipe->nrbufs)
1760                 return 0;
1761
1762         ret = 0;
1763         pipe_lock(pipe);
1764
1765         while (!pipe->nrbufs) {
1766                 if (signal_pending(current)) {
1767                         ret = -ERESTARTSYS;
1768                         break;
1769                 }
1770                 if (!pipe->writers)
1771                         break;
1772                 if (!pipe->waiting_writers) {
1773                         if (flags & SPLICE_F_NONBLOCK) {
1774                                 ret = -EAGAIN;
1775                                 break;
1776                         }
1777                 }
1778                 pipe_wait(pipe);
1779         }
1780
1781         pipe_unlock(pipe);
1782         return ret;
1783 }
1784
1785 /*
1786  * Make sure there's writeable room. Wait for room if we can, otherwise
1787  * return an appropriate error.
1788  */
1789 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1790 {
1791         int ret;
1792
1793         /*
1794          * Check ->nrbufs without the inode lock first. This function
1795          * is speculative anyways, so missing one is ok.
1796          */
1797         if (pipe->nrbufs < pipe->buffers)
1798                 return 0;
1799
1800         ret = 0;
1801         pipe_lock(pipe);
1802
1803         while (pipe->nrbufs >= pipe->buffers) {
1804                 if (!pipe->readers) {
1805                         send_sig(SIGPIPE, current, 0);
1806                         ret = -EPIPE;
1807                         break;
1808                 }
1809                 if (flags & SPLICE_F_NONBLOCK) {
1810                         ret = -EAGAIN;
1811                         break;
1812                 }
1813                 if (signal_pending(current)) {
1814                         ret = -ERESTARTSYS;
1815                         break;
1816                 }
1817                 pipe->waiting_writers++;
1818                 pipe_wait(pipe);
1819                 pipe->waiting_writers--;
1820         }
1821
1822         pipe_unlock(pipe);
1823         return ret;
1824 }
1825
1826 /*
1827  * Splice contents of ipipe to opipe.
1828  */
1829 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1830                                struct pipe_inode_info *opipe,
1831                                size_t len, unsigned int flags)
1832 {
1833         struct pipe_buffer *ibuf, *obuf;
1834         int ret = 0, nbuf;
1835         bool input_wakeup = false;
1836
1837
1838 retry:
1839         ret = ipipe_prep(ipipe, flags);
1840         if (ret)
1841                 return ret;
1842
1843         ret = opipe_prep(opipe, flags);
1844         if (ret)
1845                 return ret;
1846
1847         /*
1848          * Potential ABBA deadlock, work around it by ordering lock
1849          * grabbing by pipe info address. Otherwise two different processes
1850          * could deadlock (one doing tee from A -> B, the other from B -> A).
1851          */
1852         pipe_double_lock(ipipe, opipe);
1853
1854         do {
1855                 if (!opipe->readers) {
1856                         send_sig(SIGPIPE, current, 0);
1857                         if (!ret)
1858                                 ret = -EPIPE;
1859                         break;
1860                 }
1861
1862                 if (!ipipe->nrbufs && !ipipe->writers)
1863                         break;
1864
1865                 /*
1866                  * Cannot make any progress, because either the input
1867                  * pipe is empty or the output pipe is full.
1868                  */
1869                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1870                         /* Already processed some buffers, break */
1871                         if (ret)
1872                                 break;
1873
1874                         if (flags & SPLICE_F_NONBLOCK) {
1875                                 ret = -EAGAIN;
1876                                 break;
1877                         }
1878
1879                         /*
1880                          * We raced with another reader/writer and haven't
1881                          * managed to process any buffers.  A zero return
1882                          * value means EOF, so retry instead.
1883                          */
1884                         pipe_unlock(ipipe);
1885                         pipe_unlock(opipe);
1886                         goto retry;
1887                 }
1888
1889                 ibuf = ipipe->bufs + ipipe->curbuf;
1890                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1891                 obuf = opipe->bufs + nbuf;
1892
1893                 if (len >= ibuf->len) {
1894                         /*
1895                          * Simply move the whole buffer from ipipe to opipe
1896                          */
1897                         *obuf = *ibuf;
1898                         ibuf->ops = NULL;
1899                         opipe->nrbufs++;
1900                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1901                         ipipe->nrbufs--;
1902                         input_wakeup = true;
1903                 } else {
1904                         /*
1905                          * Get a reference to this pipe buffer,
1906                          * so we can copy the contents over.
1907                          */
1908                         ibuf->ops->get(ipipe, ibuf);
1909                         *obuf = *ibuf;
1910
1911                         /*
1912                          * Don't inherit the gift flag, we need to
1913                          * prevent multiple steals of this page.
1914                          */
1915                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1916
1917                         obuf->len = len;
1918                         opipe->nrbufs++;
1919                         ibuf->offset += obuf->len;
1920                         ibuf->len -= obuf->len;
1921                 }
1922                 ret += obuf->len;
1923                 len -= obuf->len;
1924         } while (len);
1925
1926         pipe_unlock(ipipe);
1927         pipe_unlock(opipe);
1928
1929         /*
1930          * If we put data in the output pipe, wakeup any potential readers.
1931          */
1932         if (ret > 0) {
1933                 smp_mb();
1934                 if (waitqueue_active(&opipe->wait))
1935                         wake_up_interruptible(&opipe->wait);
1936                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1937         }
1938         if (input_wakeup)
1939                 wakeup_pipe_writers(ipipe);
1940
1941         return ret;
1942 }
1943
1944 /*
1945  * Link contents of ipipe to opipe.
1946  */
1947 static int link_pipe(struct pipe_inode_info *ipipe,
1948                      struct pipe_inode_info *opipe,
1949                      size_t len, unsigned int flags)
1950 {
1951         struct pipe_buffer *ibuf, *obuf;
1952         int ret = 0, i = 0, nbuf;
1953
1954         /*
1955          * Potential ABBA deadlock, work around it by ordering lock
1956          * grabbing by pipe info address. Otherwise two different processes
1957          * could deadlock (one doing tee from A -> B, the other from B -> A).
1958          */
1959         pipe_double_lock(ipipe, opipe);
1960
1961         do {
1962                 if (!opipe->readers) {
1963                         send_sig(SIGPIPE, current, 0);
1964                         if (!ret)
1965                                 ret = -EPIPE;
1966                         break;
1967                 }
1968
1969                 /*
1970                  * If we have iterated all input buffers or ran out of
1971                  * output room, break.
1972                  */
1973                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1974                         break;
1975
1976                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1977                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1978
1979                 /*
1980                  * Get a reference to this pipe buffer,
1981                  * so we can copy the contents over.
1982                  */
1983                 ibuf->ops->get(ipipe, ibuf);
1984
1985                 obuf = opipe->bufs + nbuf;
1986                 *obuf = *ibuf;
1987
1988                 /*
1989                  * Don't inherit the gift flag, we need to
1990                  * prevent multiple steals of this page.
1991                  */
1992                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1993
1994                 if (obuf->len > len)
1995                         obuf->len = len;
1996
1997                 opipe->nrbufs++;
1998                 ret += obuf->len;
1999                 len -= obuf->len;
2000                 i++;
2001         } while (len);
2002
2003         /*
2004          * return EAGAIN if we have the potential of some data in the
2005          * future, otherwise just return 0
2006          */
2007         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2008                 ret = -EAGAIN;
2009
2010         pipe_unlock(ipipe);
2011         pipe_unlock(opipe);
2012
2013         /*
2014          * If we put data in the output pipe, wakeup any potential readers.
2015          */
2016         if (ret > 0) {
2017                 smp_mb();
2018                 if (waitqueue_active(&opipe->wait))
2019                         wake_up_interruptible(&opipe->wait);
2020                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2021         }
2022
2023         return ret;
2024 }
2025
2026 /*
2027  * This is a tee(1) implementation that works on pipes. It doesn't copy
2028  * any data, it simply references the 'in' pages on the 'out' pipe.
2029  * The 'flags' used are the SPLICE_F_* variants, currently the only
2030  * applicable one is SPLICE_F_NONBLOCK.
2031  */
2032 static long do_tee(struct file *in, struct file *out, size_t len,
2033                    unsigned int flags)
2034 {
2035         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2036         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2037         int ret = -EINVAL;
2038
2039         /*
2040          * Duplicate the contents of ipipe to opipe without actually
2041          * copying the data.
2042          */
2043         if (ipipe && opipe && ipipe != opipe) {
2044                 /*
2045                  * Keep going, unless we encounter an error. The ipipe/opipe
2046                  * ordering doesn't really matter.
2047                  */
2048                 ret = ipipe_prep(ipipe, flags);
2049                 if (!ret) {
2050                         ret = opipe_prep(opipe, flags);
2051                         if (!ret)
2052                                 ret = link_pipe(ipipe, opipe, len, flags);
2053                 }
2054         }
2055
2056         return ret;
2057 }
2058
2059 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2060 {
2061         struct file *in;
2062         int error, fput_in;
2063
2064         if (unlikely(!len))
2065                 return 0;
2066
2067         error = -EBADF;
2068         in = fget_light(fdin, &fput_in);
2069         if (in) {
2070                 if (in->f_mode & FMODE_READ) {
2071                         int fput_out;
2072                         struct file *out = fget_light(fdout, &fput_out);
2073
2074                         if (out) {
2075                                 if (out->f_mode & FMODE_WRITE)
2076                                         error = do_tee(in, out, len, flags);
2077                                 fput_light(out, fput_out);
2078                         }
2079                 }
2080                 fput_light(in, fput_in);
2081         }
2082
2083         return error;
2084 }