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