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