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 #include <linux/socket.h>
  35
  36 /*
  37  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  38  * a vm helper function, it's already simplified quite a bit by the
  39  * addition of remove_mapping(). If success is returned, the caller may
  40  * attempt to reuse this page for another destination.
  41  */
  42 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  43                                      struct pipe_buffer *buf)
  44 {
  45         struct page *page = buf->page;
  46         struct address_space *mapping;
  47
  48         lock_page(page);
  49
  50         mapping = page_mapping(page);
  51         if (mapping) {
  52                 WARN_ON(!PageUptodate(page));
  53
  54                 /*
  55                  * At least for ext2 with nobh option, we need to wait on
  56                  * writeback completing on this page, since we'll remove it
  57                  * from the pagecache.  Otherwise truncate wont wait on the
  58                  * page, allowing the disk blocks to be reused by someone else
  59                  * before we actually wrote our data to them. fs corruption
  60                  * ensues.
  61                  */
  62                 wait_on_page_writeback(page);
  63
  64                 if (page_has_private(page) &&
  65                     !try_to_release_page(page, GFP_KERNEL))
  66                         goto out_unlock;
  67
  68                 /*
  69                  * If we succeeded in removing the mapping, set LRU flag
  70                  * and return good.
  71                  */
  72                 if (remove_mapping(mapping, page)) {
  73                         buf->flags |= PIPE_BUF_FLAG_LRU;
  74                         return 0;
  75                 }
  76         }
  77
  78         /*
  79          * Raced with truncate or failed to remove page from current
  80          * address space, unlock and return failure.
  81          */
  82 out_unlock:
  83         unlock_page(page);
  84         return 1;
  85 }
  86
  87 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  88                                         struct pipe_buffer *buf)
  89 {
  90         page_cache_release(buf->page);
  91         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  92 }
  93
  94 /*
  95  * Check whether the contents of buf is OK to access. Since the content
  96  * is a page cache page, IO may be in flight.
  97  */
  98 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  99                                        struct pipe_buffer *buf)
 100 {
 101         struct page *page = buf->page;
 102         int err;
 103
 104         if (!PageUptodate(page)) {
 105                 lock_page(page);
 106
 107                 /*
 108                  * Page got truncated/unhashed. This will cause a 0-byte
 109                  * splice, if this is the first page.
 110                  */
 111                 if (!page->mapping) {
 112                         err = -ENODATA;
 113                         goto error;
 114                 }
 115
 116                 /*
 117                  * Uh oh, read-error from disk.
 118                  */
 119                 if (!PageUptodate(page)) {
 120                         err = -EIO;
 121                         goto error;
 122                 }
 123
 124                 /*
 125                  * Page is ok afterall, we are done.
 126                  */
 127                 unlock_page(page);
 128         }
 129
 130         return 0;
 131 error:
 132         unlock_page(page);
 133         return err;
 134 }
 135
 136 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 137         .can_merge = 0,
 138         .map = generic_pipe_buf_map,
 139         .unmap = generic_pipe_buf_unmap,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 158         .map = generic_pipe_buf_map,
 159         .unmap = generic_pipe_buf_unmap,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 189         int ret, do_wakeup, page_nr;
 190
 191         if (!spd_pages)
 192                 return 0;
 193
 194         ret = 0;
 195         do_wakeup = 0;
 196         page_nr = 0;
 197
 198         pipe_lock(pipe);
 199
 200         for (;;) {
 201                 if (!pipe->readers) {
 202                         send_sig(SIGPIPE, current, 0);
 203                         if (!ret)
 204                                 ret = -EPIPE;
 205                         break;
 206                 }
 207
 208                 if (pipe->nrbufs < pipe->buffers) {
 209                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 210                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 211
 212                         buf->page = spd->pages[page_nr];
 213                         buf->offset = spd->partial[page_nr].offset;
 214                         buf->len = spd->partial[page_nr].len;
 215                         buf->private = spd->partial[page_nr].private;
 216                         buf->ops = spd->ops;
 217                         buf->flags = 0;
 218                         if (spd->flags & SPLICE_F_GIFT)
 219                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 220
 221                         pipe->nrbufs++;
 222                         page_nr++;
 223                         ret += buf->len;
 224
 225                         if (pipe->inode)
 226                                 do_wakeup = 1;
 227
 228                         if (!--spd->nr_pages)
 229                                 break;
 230                         if (pipe->nrbufs < pipe->buffers)
 231                                 continue;
 232
 233                         break;
 234                 }
 235
 236                 if (spd->flags & SPLICE_F_NONBLOCK) {
 237                         if (!ret)
 238                                 ret = -EAGAIN;
 239                         break;
 240                 }
 241
 242                 if (signal_pending(current)) {
 243                         if (!ret)
 244                                 ret = -ERESTARTSYS;
 245                         break;
 246                 }
 247
 248                 if (do_wakeup) {
 249                         smp_mb();
 250                         if (waitqueue_active(&pipe->wait))
 251                                 wake_up_interruptible_sync(&pipe->wait);
 252                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 253                         do_wakeup = 0;
 254                 }
 255
 256                 pipe->waiting_writers++;
 257                 pipe_wait(pipe);
 258                 pipe->waiting_writers--;
 259         }
 260
 261         pipe_unlock(pipe);
 262
 263         if (do_wakeup)
 264                 wakeup_pipe_readers(pipe);
 265
 266         while (page_nr < spd_pages)
 267                 spd->spd_release(spd, page_nr++);
 268
 269         return ret;
 270 }
 271
 272 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 273 {
 274         page_cache_release(spd->pages[i]);
 275 }
 276
 277 /*
 278  * Check if we need to grow the arrays holding pages and partial page
 279  * descriptions.
 280  */
 281 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 282 {
 283         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 284
 285         spd->nr_pages_max = buffers;
 286         if (buffers <= PIPE_DEF_BUFFERS)
 287                 return 0;
 288
 289         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 290         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 291
 292         if (spd->pages && spd->partial)
 293                 return 0;
 294
 295         kfree(spd->pages);
 296         kfree(spd->partial);
 297         return -ENOMEM;
 298 }
 299
 300 void splice_shrink_spd(struct splice_pipe_desc *spd)
 301 {
 302         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 303                 return;
 304
 305         kfree(spd->pages);
 306         kfree(spd->partial);
 307 }
 308
 309 static int
 310 __generic_file_splice_read(struct file *in, loff_t *ppos,
 311                            struct pipe_inode_info *pipe, size_t len,
 312                            unsigned int flags)
 313 {
 314         struct address_space *mapping = in->f_mapping;
 315         unsigned int loff, nr_pages, req_pages;
 316         struct page *pages[PIPE_DEF_BUFFERS];
 317         struct partial_page partial[PIPE_DEF_BUFFERS];
 318         struct page *page;
 319         pgoff_t index, end_index;
 320         loff_t isize;
 321         int error, page_nr;
 322         struct splice_pipe_desc spd = {
 323                 .pages = pages,
 324                 .partial = partial,
 325                 .nr_pages_max = PIPE_DEF_BUFFERS,
 326                 .flags = flags,
 327                 .ops = &page_cache_pipe_buf_ops,
 328                 .spd_release = spd_release_page,
 329         };
 330
 331         if (splice_grow_spd(pipe, &spd))
 332                 return -ENOMEM;
 333
 334         index = *ppos >> PAGE_CACHE_SHIFT;
 335         loff = *ppos & ~PAGE_CACHE_MASK;
 336         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 337         nr_pages = min(req_pages, spd.nr_pages_max);
 338
 339         /*
 340          * Lookup the (hopefully) full range of pages we need.
 341          */
 342         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 343         index += spd.nr_pages;
 344
 345         /*
 346          * If find_get_pages_contig() returned fewer pages than we needed,
 347          * readahead/allocate the rest and fill in the holes.
 348          */
 349         if (spd.nr_pages < nr_pages)
 350                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 351                                 index, req_pages - spd.nr_pages);
 352
 353         error = 0;
 354         while (spd.nr_pages < nr_pages) {
 355                 /*
 356                  * Page could be there, find_get_pages_contig() breaks on
 357                  * the first hole.
 358                  */
 359                 page = find_get_page(mapping, index);
 360                 if (!page) {
 361                         /*
 362                          * page didn't exist, allocate one.
 363                          */
 364                         page = page_cache_alloc_cold(mapping);
 365                         if (!page)
 366                                 break;
 367
 368                         error = add_to_page_cache_lru(page, mapping, index,
 369                                                 GFP_KERNEL);
 370                         if (unlikely(error)) {
 371                                 page_cache_release(page);
 372                                 if (error == -EEXIST)
 373                                         continue;
 374                                 break;
 375                         }
 376                         /*
 377                          * add_to_page_cache() locks the page, unlock it
 378                          * to avoid convoluting the logic below even more.
 379                          */
 380                         unlock_page(page);
 381                 }
 382
 383                 spd.pages[spd.nr_pages++] = page;
 384                 index++;
 385         }
 386
 387         /*
 388          * Now loop over the map and see if we need to start IO on any
 389          * pages, fill in the partial map, etc.
 390          */
 391         index = *ppos >> PAGE_CACHE_SHIFT;
 392         nr_pages = spd.nr_pages;
 393         spd.nr_pages = 0;
 394         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 395                 unsigned int this_len;
 396
 397                 if (!len)
 398                         break;
 399
 400                 /*
 401                  * this_len is the max we'll use from this page
 402                  */
 403                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 404                 page = spd.pages[page_nr];
 405
 406                 if (PageReadahead(page))
 407                         page_cache_async_readahead(mapping, &in->f_ra, in,
 408                                         page, index, req_pages - page_nr);
 409
 410                 /*
 411                  * If the page isn't uptodate, we may need to start io on it
 412                  */
 413                 if (!PageUptodate(page)) {
 414                         lock_page(page);
 415
 416                         /*
 417                          * Page was truncated, or invalidated by the
 418                          * filesystem.  Redo the find/create, but this time the
 419                          * page is kept locked, so there's no chance of another
 420                          * race with truncate/invalidate.
 421                          */
 422                         if (!page->mapping) {
 423                                 unlock_page(page);
 424                                 page = find_or_create_page(mapping, index,
 425                                                 mapping_gfp_mask(mapping));
 426
 427                                 if (!page) {
 428                                         error = -ENOMEM;
 429                                         break;
 430                                 }
 431                                 page_cache_release(spd.pages[page_nr]);
 432                                 spd.pages[page_nr] = page;
 433                         }
 434                         /*
 435                          * page was already under io and is now done, great
 436                          */
 437                         if (PageUptodate(page)) {
 438                                 unlock_page(page);
 439                                 goto fill_it;
 440                         }
 441
 442                         /*
 443                          * need to read in the page
 444                          */
 445                         error = mapping->a_ops->readpage(in, page);
 446                         if (unlikely(error)) {
 447                                 /*
 448                                  * We really should re-lookup the page here,
 449                                  * but it complicates things a lot. Instead
 450                                  * lets just do what we already stored, and
 451                                  * we'll get it the next time we are called.
 452                                  */
 453                                 if (error == AOP_TRUNCATED_PAGE)
 454                                         error = 0;
 455
 456                                 break;
 457                         }
 458                 }
 459 fill_it:
 460                 /*
 461                  * i_size must be checked after PageUptodate.
 462                  */
 463                 isize = i_size_read(mapping->host);
 464                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 465                 if (unlikely(!isize || index > end_index))
 466                         break;
 467
 468                 /*
 469                  * if this is the last page, see if we need to shrink
 470                  * the length and stop
 471                  */
 472                 if (end_index == index) {
 473                         unsigned int plen;
 474
 475                         /*
 476                          * max good bytes in this page
 477                          */
 478                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 479                         if (plen <= loff)
 480                                 break;
 481
 482                         /*
 483                          * force quit after adding this page
 484                          */
 485                         this_len = min(this_len, plen - loff);
 486                         len = this_len;
 487                 }
 488
 489                 spd.partial[page_nr].offset = loff;
 490                 spd.partial[page_nr].len = this_len;
 491                 len -= this_len;
 492                 loff = 0;
 493                 spd.nr_pages++;
 494                 index++;
 495         }
 496
 497         /*
 498          * Release any pages at the end, if we quit early. 'page_nr' is how far
 499          * we got, 'nr_pages' is how many pages are in the map.
 500          */
 501         while (page_nr < nr_pages)
 502                 page_cache_release(spd.pages[page_nr++]);
 503         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 504
 505         if (spd.nr_pages)
 506                 error = splice_to_pipe(pipe, &spd);
 507
 508         splice_shrink_spd(&spd);
 509         return error;
 510 }
 511
 512 /**
 513  * generic_file_splice_read - splice data from file to a pipe
 514  * @in:         file to splice from
 515  * @ppos:       position in @in
 516  * @pipe:       pipe to splice to
 517  * @len:        number of bytes to splice
 518  * @flags:      splice modifier flags
 519  *
 520  * Description:
 521  *    Will read pages from given file and fill them into a pipe. Can be
 522  *    used as long as the address_space operations for the source implements
 523  *    a readpage() hook.
 524  *
 525  */
 526 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 527                                  struct pipe_inode_info *pipe, size_t len,
 528                                  unsigned int flags)
 529 {
 530         loff_t isize, left;
 531         int ret;
 532
 533         isize = i_size_read(in->f_mapping->host);
 534         if (unlikely(*ppos >= isize))
 535                 return 0;
 536
 537         left = isize - *ppos;
 538         if (unlikely(left < len))
 539                 len = left;
 540
 541         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 542         if (ret > 0) {
 543                 *ppos += ret;
 544                 file_accessed(in);
 545         }
 546
 547         return ret;
 548 }
 549 EXPORT_SYMBOL(generic_file_splice_read);
 550
 551 static const struct pipe_buf_operations default_pipe_buf_ops = {
 552         .can_merge = 0,
 553         .map = generic_pipe_buf_map,
 554         .unmap = generic_pipe_buf_unmap,
 555         .confirm = generic_pipe_buf_confirm,
 556         .release = generic_pipe_buf_release,
 557         .steal = generic_pipe_buf_steal,
 558         .get = generic_pipe_buf_get,
 559 };
 560
 561 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 562                                     struct pipe_buffer *buf)
 563 {
 564         return 1;
 565 }
 566
 567 /* Pipe buffer operations for a socket and similar. */
 568 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 569         .can_merge = 0,
 570         .map = generic_pipe_buf_map,
 571         .unmap = generic_pipe_buf_unmap,
 572         .confirm = generic_pipe_buf_confirm,
 573         .release = generic_pipe_buf_release,
 574         .steal = generic_pipe_buf_nosteal,
 575         .get = generic_pipe_buf_get,
 576 };
 577 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 578
 579 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 580                             unsigned long vlen, loff_t offset)
 581 {
 582         mm_segment_t old_fs;
 583         loff_t pos = offset;
 584         ssize_t res;
 585
 586         old_fs = get_fs();
 587         set_fs(get_ds());
 588         /* The cast to a user pointer is valid due to the set_fs() */
 589         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 590         set_fs(old_fs);
 591
 592         return res;
 593 }
 594
 595 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 596                             loff_t pos)
 597 {
 598         mm_segment_t old_fs;
 599         ssize_t res;
 600
 601         old_fs = get_fs();
 602         set_fs(get_ds());
 603         /* The cast to a user pointer is valid due to the set_fs() */
 604         res = vfs_write(file, (const char __user *)buf, count, &pos);
 605         set_fs(old_fs);
 606
 607         return res;
 608 }
 609
 610 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 611                                  struct pipe_inode_info *pipe, size_t len,
 612                                  unsigned int flags)
 613 {
 614         unsigned int nr_pages;
 615         unsigned int nr_freed;
 616         size_t offset;
 617         struct page *pages[PIPE_DEF_BUFFERS];
 618         struct partial_page partial[PIPE_DEF_BUFFERS];
 619         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 620         ssize_t res;
 621         size_t this_len;
 622         int error;
 623         int i;
 624         struct splice_pipe_desc spd = {
 625                 .pages = pages,
 626                 .partial = partial,
 627                 .nr_pages_max = PIPE_DEF_BUFFERS,
 628                 .flags = flags,
 629                 .ops = &default_pipe_buf_ops,
 630                 .spd_release = spd_release_page,
 631         };
 632
 633         if (splice_grow_spd(pipe, &spd))
 634                 return -ENOMEM;
 635
 636         res = -ENOMEM;
 637         vec = __vec;
 638         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 639                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 640                 if (!vec)
 641                         goto shrink_ret;
 642         }
 643
 644         offset = *ppos & ~PAGE_CACHE_MASK;
 645         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 646
 647         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 648                 struct page *page;
 649
 650                 page = alloc_page(GFP_USER);
 651                 error = -ENOMEM;
 652                 if (!page)
 653                         goto err;
 654
 655                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 656                 vec[i].iov_base = (void __user *) page_address(page);
 657                 vec[i].iov_len = this_len;
 658                 spd.pages[i] = page;
 659                 spd.nr_pages++;
 660                 len -= this_len;
 661                 offset = 0;
 662         }
 663
 664         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 665         if (res < 0) {
 666                 error = res;
 667                 goto err;
 668         }
 669
 670         error = 0;
 671         if (!res)
 672                 goto err;
 673
 674         nr_freed = 0;
 675         for (i = 0; i < spd.nr_pages; i++) {
 676                 this_len = min_t(size_t, vec[i].iov_len, res);
 677                 spd.partial[i].offset = 0;
 678                 spd.partial[i].len = this_len;
 679                 if (!this_len) {
 680                         __free_page(spd.pages[i]);
 681                         spd.pages[i] = NULL;
 682                         nr_freed++;
 683                 }
 684                 res -= this_len;
 685         }
 686         spd.nr_pages -= nr_freed;
 687
 688         res = splice_to_pipe(pipe, &spd);
 689         if (res > 0)
 690                 *ppos += res;
 691
 692 shrink_ret:
 693         if (vec != __vec)
 694                 kfree(vec);
 695         splice_shrink_spd(&spd);
 696         return res;
 697
 698 err:
 699         for (i = 0; i < spd.nr_pages; i++)
 700                 __free_page(spd.pages[i]);
 701
 702         res = error;
 703         goto shrink_ret;
 704 }
 705 EXPORT_SYMBOL(default_file_splice_read);
 706
 707 /*
 708  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 709  * using sendpage(). Return the number of bytes sent.
 710  */
 711 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 712                             struct pipe_buffer *buf, struct splice_desc *sd)
 713 {
 714         struct file *file = sd->u.file;
 715         loff_t pos = sd->pos;
 716         int more;
 717
 718         if (!likely(file->f_op && file->f_op->sendpage))
 719                 return -EINVAL;
 720
 721         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 722
 723         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 724                 more |= MSG_SENDPAGE_NOTLAST;
 725
 726         return file->f_op->sendpage(file, buf->page, buf->offset,
 727                                     sd->len, &pos, more);
 728 }
 729
 730 /*
 731  * This is a little more tricky than the file -> pipe splicing. There are
 732  * basically three cases:
 733  *
 734  *      - Destination page already exists in the address space and there
 735  *        are users of it. For that case we have no other option that
 736  *        copying the data. Tough luck.
 737  *      - Destination page already exists in the address space, but there
 738  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 739  *        through to last case.
 740  *      - Destination page does not exist, we can add the pipe page to
 741  *        the page cache and avoid the copy.
 742  *
 743  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 744  * sd->flags), we attempt to migrate pages from the pipe to the output
 745  * file address space page cache. This is possible if no one else has
 746  * the pipe page referenced outside of the pipe and page cache. If
 747  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 748  * a new page in the output file page cache and fill/dirty that.
 749  */
 750 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 751                  struct splice_desc *sd)
 752 {
 753         struct file *file = sd->u.file;
 754         struct address_space *mapping = file->f_mapping;
 755         unsigned int offset, this_len;
 756         struct page *page;
 757         void *fsdata;
 758         int ret;
 759
 760         offset = sd->pos & ~PAGE_CACHE_MASK;
 761
 762         this_len = sd->len;
 763         if (this_len + offset > PAGE_CACHE_SIZE)
 764                 this_len = PAGE_CACHE_SIZE - offset;
 765
 766         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 767                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 768         if (unlikely(ret))
 769                 goto out;
 770
 771         if (buf->page != page) {
 772                 /*
 773                  * Careful, ->map() uses KM_USER0!
 774                  */
 775                 char *src = buf->ops->map(pipe, buf, 1);
 776                 char *dst = kmap_atomic(page, KM_USER1);
 777
 778                 memcpy(dst + offset, src + buf->offset, this_len);
 779                 flush_dcache_page(page);
 780                 kunmap_atomic(dst, KM_USER1);
 781                 buf->ops->unmap(pipe, buf, src);
 782         }
 783         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 784                                 page, fsdata);
 785 out:
 786         return ret;
 787 }
 788 EXPORT_SYMBOL(pipe_to_file);
 789
 790 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 791 {
 792         smp_mb();
 793         if (waitqueue_active(&pipe->wait))
 794                 wake_up_interruptible(&pipe->wait);
 795         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 796 }
 797
 798 /**
 799  * splice_from_pipe_feed - feed available data from a pipe to a file
 800  * @pipe:       pipe to splice from
 801  * @sd:         information to @actor
 802  * @actor:      handler that splices the data
 803  *
 804  * Description:
 805  *    This function loops over the pipe and calls @actor to do the
 806  *    actual moving of a single struct pipe_buffer to the desired
 807  *    destination.  It returns when there's no more buffers left in
 808  *    the pipe or if the requested number of bytes (@sd->total_len)
 809  *    have been copied.  It returns a positive number (one) if the
 810  *    pipe needs to be filled with more data, zero if the required
 811  *    number of bytes have been copied and -errno on error.
 812  *
 813  *    This, together with splice_from_pipe_{begin,end,next}, may be
 814  *    used to implement the functionality of __splice_from_pipe() when
 815  *    locking is required around copying the pipe buffers to the
 816  *    destination.
 817  */
 818 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 819                           splice_actor *actor)
 820 {
 821         int ret;
 822
 823         while (pipe->nrbufs) {
 824                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 825                 const struct pipe_buf_operations *ops = buf->ops;
 826
 827                 sd->len = buf->len;
 828                 if (sd->len > sd->total_len)
 829                         sd->len = sd->total_len;
 830
 831                 ret = buf->ops->confirm(pipe, buf);
 832                 if (unlikely(ret)) {
 833                         if (ret == -ENODATA)
 834                                 ret = 0;
 835                         return ret;
 836                 }
 837
 838                 ret = actor(pipe, buf, sd);
 839                 if (ret <= 0)
 840                         return ret;
 841
 842                 buf->offset += ret;
 843                 buf->len -= ret;
 844
 845                 sd->num_spliced += ret;
 846                 sd->len -= ret;
 847                 sd->pos += ret;
 848                 sd->total_len -= ret;
 849
 850                 if (!buf->len) {
 851                         buf->ops = NULL;
 852                         ops->release(pipe, buf);
 853                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 854                         pipe->nrbufs--;
 855                         if (pipe->inode)
 856                                 sd->need_wakeup = true;
 857                 }
 858
 859                 if (!sd->total_len)
 860                         return 0;
 861         }
 862
 863         return 1;
 864 }
 865 EXPORT_SYMBOL(splice_from_pipe_feed);
 866
 867 /**
 868  * splice_from_pipe_next - wait for some data to splice from
 869  * @pipe:       pipe to splice from
 870  * @sd:         information about the splice operation
 871  *
 872  * Description:
 873  *    This function will wait for some data and return a positive
 874  *    value (one) if pipe buffers are available.  It will return zero
 875  *    or -errno if no more data needs to be spliced.
 876  */
 877 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 878 {
 879         /*
 880          * Check for signal early to make process killable when there are
 881          * always buffers available
 882          */
 883         if (signal_pending(current))
 884                 return -ERESTARTSYS;
 885
 886         while (!pipe->nrbufs) {
 887                 if (!pipe->writers)
 888                         return 0;
 889
 890                 if (!pipe->waiting_writers && sd->num_spliced)
 891                         return 0;
 892
 893                 if (sd->flags & SPLICE_F_NONBLOCK)
 894                         return -EAGAIN;
 895
 896                 if (signal_pending(current))
 897                         return -ERESTARTSYS;
 898
 899                 if (sd->need_wakeup) {
 900                         wakeup_pipe_writers(pipe);
 901                         sd->need_wakeup = false;
 902                 }
 903
 904                 pipe_wait(pipe);
 905         }
 906
 907         return 1;
 908 }
 909 EXPORT_SYMBOL(splice_from_pipe_next);
 910
 911 /**
 912  * splice_from_pipe_begin - start splicing from pipe
 913  * @sd:         information about the splice operation
 914  *
 915  * Description:
 916  *    This function should be called before a loop containing
 917  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 918  *    initialize the necessary fields of @sd.
 919  */
 920 void splice_from_pipe_begin(struct splice_desc *sd)
 921 {
 922         sd->num_spliced = 0;
 923         sd->need_wakeup = false;
 924 }
 925 EXPORT_SYMBOL(splice_from_pipe_begin);
 926
 927 /**
 928  * splice_from_pipe_end - finish splicing from pipe
 929  * @pipe:       pipe to splice from
 930  * @sd:         information about the splice operation
 931  *
 932  * Description:
 933  *    This function will wake up pipe writers if necessary.  It should
 934  *    be called after a loop containing splice_from_pipe_next() and
 935  *    splice_from_pipe_feed().
 936  */
 937 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 938 {
 939         if (sd->need_wakeup)
 940                 wakeup_pipe_writers(pipe);
 941 }
 942 EXPORT_SYMBOL(splice_from_pipe_end);
 943
 944 /**
 945  * __splice_from_pipe - splice data from a pipe to given actor
 946  * @pipe:       pipe to splice from
 947  * @sd:         information to @actor
 948  * @actor:      handler that splices the data
 949  *
 950  * Description:
 951  *    This function does little more than loop over the pipe and call
 952  *    @actor to do the actual moving of a single struct pipe_buffer to
 953  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 954  *    pipe_to_user.
 955  *
 956  */
 957 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 958                            splice_actor *actor)
 959 {
 960         int ret;
 961
 962         splice_from_pipe_begin(sd);
 963         do {
 964                 cond_resched();
 965                 ret = splice_from_pipe_next(pipe, sd);
 966                 if (ret > 0)
 967                         ret = splice_from_pipe_feed(pipe, sd, actor);
 968         } while (ret > 0);
 969         splice_from_pipe_end(pipe, sd);
 970
 971         return sd->num_spliced ? sd->num_spliced : ret;
 972 }
 973 EXPORT_SYMBOL(__splice_from_pipe);
 974
 975 /**
 976  * splice_from_pipe - splice data from a pipe to a file
 977  * @pipe:       pipe to splice from
 978  * @out:        file to splice to
 979  * @ppos:       position in @out
 980  * @len:        how many bytes to splice
 981  * @flags:      splice modifier flags
 982  * @actor:      handler that splices the data
 983  *
 984  * Description:
 985  *    See __splice_from_pipe. This function locks the pipe inode,
 986  *    otherwise it's identical to __splice_from_pipe().
 987  *
 988  */
 989 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 990                          loff_t *ppos, size_t len, unsigned int flags,
 991                          splice_actor *actor)
 992 {
 993         ssize_t ret;
 994         struct splice_desc sd = {
 995                 .total_len = len,
 996                 .flags = flags,
 997                 .pos = *ppos,
 998                 .u.file = out,
 999         };
1000
1001         pipe_lock(pipe);
1002         ret = __splice_from_pipe(pipe, &sd, actor);
1003         pipe_unlock(pipe);
1004
1005         return ret;
1006 }
1007
1008 /**
1009  * generic_file_splice_write - splice data from a pipe to a file
1010  * @pipe:       pipe info
1011  * @out:        file to write to
1012  * @ppos:       position in @out
1013  * @len:        number of bytes to splice
1014  * @flags:      splice modifier flags
1015  *
1016  * Description:
1017  *    Will either move or copy pages (determined by @flags options) from
1018  *    the given pipe inode to the given file.
1019  *
1020  */
1021 ssize_t
1022 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1023                           loff_t *ppos, size_t len, unsigned int flags)
1024 {
1025         struct address_space *mapping = out->f_mapping;
1026         struct inode *inode = mapping->host;
1027         struct splice_desc sd = {
1028                 .flags = flags,
1029                 .u.file = out,
1030         };
1031         ssize_t ret;
1032
1033         ret = generic_write_checks(out, ppos, &len, S_ISBLK(inode->i_mode));
1034         if (ret)
1035                 return ret;
1036         sd.total_len = len;
1037         sd.pos = *ppos;
1038
1039         pipe_lock(pipe);
1040
1041         splice_from_pipe_begin(&sd);
1042         do {
1043                 ret = splice_from_pipe_next(pipe, &sd);
1044                 if (ret <= 0)
1045                         break;
1046
1047                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1048                 ret = file_remove_suid(out);
1049                 if (!ret) {
1050                         file_update_time(out);
1051                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1052                 }
1053                 mutex_unlock(&inode->i_mutex);
1054         } while (ret > 0);
1055         splice_from_pipe_end(pipe, &sd);
1056
1057         pipe_unlock(pipe);
1058
1059         if (sd.num_spliced)
1060                 ret = sd.num_spliced;
1061
1062         if (ret > 0) {
1063                 unsigned long nr_pages;
1064                 int err;
1065
1066                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1067
1068                 err = generic_write_sync(out, *ppos, ret);
1069                 if (err)
1070                         ret = err;
1071                 else
1072                         *ppos += ret;
1073                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1074         }
1075
1076         return ret;
1077 }
1078
1079 EXPORT_SYMBOL(generic_file_splice_write);
1080
1081 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1082                           struct splice_desc *sd)
1083 {
1084         int ret;
1085         void *data;
1086
1087         data = buf->ops->map(pipe, buf, 0);
1088         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1089         buf->ops->unmap(pipe, buf, data);
1090
1091         return ret;
1092 }
1093
1094 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1095                                          struct file *out, loff_t *ppos,
1096                                          size_t len, unsigned int flags)
1097 {
1098         ssize_t ret;
1099
1100         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1101         if (ret > 0)
1102                 *ppos += ret;
1103
1104         return ret;
1105 }
1106
1107 /**
1108  * generic_splice_sendpage - splice data from a pipe to a socket
1109  * @pipe:       pipe to splice from
1110  * @out:        socket to write to
1111  * @ppos:       position in @out
1112  * @len:        number of bytes to splice
1113  * @flags:      splice modifier flags
1114  *
1115  * Description:
1116  *    Will send @len bytes from the pipe to a network socket. No data copying
1117  *    is involved.
1118  *
1119  */
1120 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1121                                 loff_t *ppos, size_t len, unsigned int flags)
1122 {
1123         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1124 }
1125
1126 EXPORT_SYMBOL(generic_splice_sendpage);
1127
1128 /*
1129  * Attempt to initiate a splice from pipe to file.
1130  */
1131 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1132                            loff_t *ppos, size_t len, unsigned int flags)
1133 {
1134         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1135                                 loff_t *, size_t, unsigned int);
1136         int ret;
1137
1138         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1139                 return -EBADF;
1140
1141         if (unlikely(out->f_flags & O_APPEND))
1142                 return -EINVAL;
1143
1144         ret = rw_verify_area(WRITE, out, ppos, len);
1145         if (unlikely(ret < 0))
1146                 return ret;
1147
1148         if (out->f_op && out->f_op->splice_write)
1149                 splice_write = out->f_op->splice_write;
1150         else
1151                 splice_write = default_file_splice_write;
1152
1153         return splice_write(pipe, out, ppos, len, flags);
1154 }
1155
1156 /*
1157  * Attempt to initiate a splice from a file to a pipe.
1158  */
1159 static long do_splice_to(struct file *in, loff_t *ppos,
1160                          struct pipe_inode_info *pipe, size_t len,
1161                          unsigned int flags)
1162 {
1163         ssize_t (*splice_read)(struct file *, loff_t *,
1164                                struct pipe_inode_info *, size_t, unsigned int);
1165         int ret;
1166
1167         if (unlikely(!(in->f_mode & FMODE_READ)))
1168                 return -EBADF;
1169
1170         ret = rw_verify_area(READ, in, ppos, len);
1171         if (unlikely(ret < 0))
1172                 return ret;
1173
1174         if (in->f_op && in->f_op->splice_read)
1175                 splice_read = in->f_op->splice_read;
1176         else
1177                 splice_read = default_file_splice_read;
1178
1179         return splice_read(in, ppos, pipe, len, flags);
1180 }
1181
1182 /**
1183  * splice_direct_to_actor - splices data directly between two non-pipes
1184  * @in:         file to splice from
1185  * @sd:         actor information on where to splice to
1186  * @actor:      handles the data splicing
1187  *
1188  * Description:
1189  *    This is a special case helper to splice directly between two
1190  *    points, without requiring an explicit pipe. Internally an allocated
1191  *    pipe is cached in the process, and reused during the lifetime of
1192  *    that process.
1193  *
1194  */
1195 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1196                                splice_direct_actor *actor)
1197 {
1198         struct pipe_inode_info *pipe;
1199         long ret, bytes;
1200         umode_t i_mode;
1201         size_t len;
1202         int i, flags, more;
1203
1204         /*
1205          * We require the input being a regular file, as we don't want to
1206          * randomly drop data for eg socket -> socket splicing. Use the
1207          * piped splicing for that!
1208          */
1209         i_mode = in->f_path.dentry->d_inode->i_mode;
1210         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1211                 return -EINVAL;
1212
1213         /*
1214          * neither in nor out is a pipe, setup an internal pipe attached to
1215          * 'out' and transfer the wanted data from 'in' to 'out' through that
1216          */
1217         pipe = current->splice_pipe;
1218         if (unlikely(!pipe)) {
1219                 pipe = alloc_pipe_info(NULL);
1220                 if (!pipe)
1221                         return -ENOMEM;
1222
1223                 /*
1224                  * We don't have an immediate reader, but we'll read the stuff
1225                  * out of the pipe right after the splice_to_pipe(). So set
1226                  * PIPE_READERS appropriately.
1227                  */
1228                 pipe->readers = 1;
1229
1230                 current->splice_pipe = pipe;
1231         }
1232
1233         /*
1234          * Do the splice.
1235          */
1236         ret = 0;
1237         bytes = 0;
1238         len = sd->total_len;
1239         flags = sd->flags;
1240
1241         /*
1242          * Don't block on output, we have to drain the direct pipe.
1243          */
1244         sd->flags &= ~SPLICE_F_NONBLOCK;
1245         more = sd->flags & SPLICE_F_MORE;
1246
1247         while (len) {
1248                 size_t read_len;
1249                 loff_t pos = sd->pos, prev_pos = pos;
1250
1251                 ret = do_splice_to(in, &pos, pipe, len, flags);
1252                 if (unlikely(ret <= 0))
1253                         goto out_release;
1254
1255                 read_len = ret;
1256                 sd->total_len = read_len;
1257
1258                 /*
1259                  * If more data is pending, set SPLICE_F_MORE
1260                  * If this is the last data and SPLICE_F_MORE was not set
1261                  * initially, clears it.
1262                  */
1263                 if (read_len < len)
1264                         sd->flags |= SPLICE_F_MORE;
1265                 else if (!more)
1266                         sd->flags &= ~SPLICE_F_MORE;
1267                 /*
1268                  * NOTE: nonblocking mode only applies to the input. We
1269                  * must not do the output in nonblocking mode as then we
1270                  * could get stuck data in the internal pipe:
1271                  */
1272                 ret = actor(pipe, sd);
1273                 if (unlikely(ret <= 0)) {
1274                         sd->pos = prev_pos;
1275                         goto out_release;
1276                 }
1277
1278                 bytes += ret;
1279                 len -= ret;
1280                 sd->pos = pos;
1281
1282                 if (ret < read_len) {
1283                         sd->pos = prev_pos + ret;
1284                         goto out_release;
1285                 }
1286         }
1287
1288 done:
1289         pipe->nrbufs = pipe->curbuf = 0;
1290         file_accessed(in);
1291         return bytes;
1292
1293 out_release:
1294         /*
1295          * If we did an incomplete transfer we must release
1296          * the pipe buffers in question:
1297          */
1298         for (i = 0; i < pipe->buffers; i++) {
1299                 struct pipe_buffer *buf = pipe->bufs + i;
1300
1301                 if (buf->ops) {
1302                         buf->ops->release(pipe, buf);
1303                         buf->ops = NULL;
1304                 }
1305         }
1306
1307         if (!bytes)
1308                 bytes = ret;
1309
1310         goto done;
1311 }
1312 EXPORT_SYMBOL(splice_direct_to_actor);
1313
1314 static int direct_splice_actor(struct pipe_inode_info *pipe,
1315                                struct splice_desc *sd)
1316 {
1317         struct file *file = sd->u.file;
1318
1319         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1320                               sd->flags);
1321 }
1322
1323 /**
1324  * do_splice_direct - splices data directly between two files
1325  * @in:         file to splice from
1326  * @ppos:       input file offset
1327  * @out:        file to splice to
1328  * @len:        number of bytes to splice
1329  * @flags:      splice modifier flags
1330  *
1331  * Description:
1332  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1333  *    doing it in the application would incur an extra system call
1334  *    (splice in + splice out, as compared to just sendfile()). So this helper
1335  *    can splice directly through a process-private pipe.
1336  *
1337  */
1338 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1339                       size_t len, unsigned int flags)
1340 {
1341         struct splice_desc sd = {
1342                 .len            = len,
1343                 .total_len      = len,
1344                 .flags          = flags,
1345                 .pos            = *ppos,
1346                 .u.file         = out,
1347         };
1348         long ret;
1349
1350         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1351         if (ret > 0)
1352                 *ppos = sd.pos;
1353
1354         return ret;
1355 }
1356
1357 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1358                                struct pipe_inode_info *opipe,
1359                                size_t len, unsigned int flags);
1360
1361 /*
1362  * Determine where to splice to/from.
1363  */
1364 static long do_splice(struct file *in, loff_t __user *off_in,
1365                       struct file *out, loff_t __user *off_out,
1366                       size_t len, unsigned int flags)
1367 {
1368         struct pipe_inode_info *ipipe;
1369         struct pipe_inode_info *opipe;
1370         loff_t offset, *off;
1371         long ret;
1372
1373         ipipe = get_pipe_info(in);
1374         opipe = get_pipe_info(out);
1375
1376         if (ipipe && opipe) {
1377                 if (off_in || off_out)
1378                         return -ESPIPE;
1379
1380                 if (!(in->f_mode & FMODE_READ))
1381                         return -EBADF;
1382
1383                 if (!(out->f_mode & FMODE_WRITE))
1384                         return -EBADF;
1385
1386                 /* Splicing to self would be fun, but... */
1387                 if (ipipe == opipe)
1388                         return -EINVAL;
1389
1390                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1391         }
1392
1393         if (ipipe) {
1394                 if (off_in)
1395                         return -ESPIPE;
1396                 if (off_out) {
1397                         if (!(out->f_mode & FMODE_PWRITE))
1398                                 return -EINVAL;
1399                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1400                                 return -EFAULT;
1401                         off = &offset;
1402                 } else
1403                         off = &out->f_pos;
1404
1405                 ret = do_splice_from(ipipe, out, off, len, flags);
1406
1407                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1408                         ret = -EFAULT;
1409
1410                 return ret;
1411         }
1412
1413         if (opipe) {
1414                 if (off_out)
1415                         return -ESPIPE;
1416                 if (off_in) {
1417                         if (!(in->f_mode & FMODE_PREAD))
1418                                 return -EINVAL;
1419                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1420                                 return -EFAULT;
1421                         off = &offset;
1422                 } else
1423                         off = &in->f_pos;
1424
1425                 ret = do_splice_to(in, off, opipe, len, flags);
1426
1427                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1428                         ret = -EFAULT;
1429
1430                 return ret;
1431         }
1432
1433         return -EINVAL;
1434 }
1435
1436 /*
1437  * Map an iov into an array of pages and offset/length tupples. With the
1438  * partial_page structure, we can map several non-contiguous ranges into
1439  * our ones pages[] map instead of splitting that operation into pieces.
1440  * Could easily be exported as a generic helper for other users, in which
1441  * case one would probably want to add a 'max_nr_pages' parameter as well.
1442  */
1443 static int get_iovec_page_array(const struct iovec __user *iov,
1444                                 unsigned int nr_vecs, struct page **pages,
1445                                 struct partial_page *partial, int aligned,
1446                                 unsigned int pipe_buffers)
1447 {
1448         int buffers = 0, error = 0;
1449
1450         while (nr_vecs) {
1451                 unsigned long off, npages;
1452                 struct iovec entry;
1453                 void __user *base;
1454                 size_t len;
1455                 int i;
1456
1457                 error = -EFAULT;
1458                 if (copy_from_user(&entry, iov, sizeof(entry)))
1459                         break;
1460
1461                 base = entry.iov_base;
1462                 len = entry.iov_len;
1463
1464                 /*
1465                  * Sanity check this iovec. 0 read succeeds.
1466                  */
1467                 error = 0;
1468                 if (unlikely(!len))
1469                         break;
1470                 error = -EFAULT;
1471                 if (!access_ok(VERIFY_READ, base, len))
1472                         break;
1473
1474                 /*
1475                  * Get this base offset and number of pages, then map
1476                  * in the user pages.
1477                  */
1478                 off = (unsigned long) base & ~PAGE_MASK;
1479
1480                 /*
1481                  * If asked for alignment, the offset must be zero and the
1482                  * length a multiple of the PAGE_SIZE.
1483                  */
1484                 error = -EINVAL;
1485                 if (aligned && (off || len & ~PAGE_MASK))
1486                         break;
1487
1488                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1489                 if (npages > pipe_buffers - buffers)
1490                         npages = pipe_buffers - buffers;
1491
1492                 error = get_user_pages_fast((unsigned long)base, npages,
1493                                         0, &pages[buffers]);
1494
1495                 if (unlikely(error <= 0))
1496                         break;
1497
1498                 /*
1499                  * Fill this contiguous range into the partial page map.
1500                  */
1501                 for (i = 0; i < error; i++) {
1502                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1503
1504                         partial[buffers].offset = off;
1505                         partial[buffers].len = plen;
1506
1507                         off = 0;
1508                         len -= plen;
1509                         buffers++;
1510                 }
1511
1512                 /*
1513                  * We didn't complete this iov, stop here since it probably
1514                  * means we have to move some of this into a pipe to
1515                  * be able to continue.
1516                  */
1517                 if (len)
1518                         break;
1519
1520                 /*
1521                  * Don't continue if we mapped fewer pages than we asked for,
1522                  * or if we mapped the max number of pages that we have
1523                  * room for.
1524                  */
1525                 if (error < npages || buffers == pipe_buffers)
1526                         break;
1527
1528                 nr_vecs--;
1529                 iov++;
1530         }
1531
1532         if (buffers)
1533                 return buffers;
1534
1535         return error;
1536 }
1537
1538 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1539                         struct splice_desc *sd)
1540 {
1541         char *src;
1542         int ret;
1543
1544         /*
1545          * See if we can use the atomic maps, by prefaulting in the
1546          * pages and doing an atomic copy
1547          */
1548         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1549                 src = buf->ops->map(pipe, buf, 1);
1550                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1551                                                         sd->len);
1552                 buf->ops->unmap(pipe, buf, src);
1553                 if (!ret) {
1554                         ret = sd->len;
1555                         goto out;
1556                 }
1557         }
1558
1559         /*
1560          * No dice, use slow non-atomic map and copy
1561          */
1562         src = buf->ops->map(pipe, buf, 0);
1563
1564         ret = sd->len;
1565         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1566                 ret = -EFAULT;
1567
1568         buf->ops->unmap(pipe, buf, src);
1569 out:
1570         if (ret > 0)
1571                 sd->u.userptr += ret;
1572         return ret;
1573 }
1574
1575 /*
1576  * For lack of a better implementation, implement vmsplice() to userspace
1577  * as a simple copy of the pipes pages to the user iov.
1578  */
1579 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1580                              unsigned long nr_segs, unsigned int flags)
1581 {
1582         struct pipe_inode_info *pipe;
1583         struct splice_desc sd;
1584         ssize_t size;
1585         int error;
1586         long ret;
1587
1588         pipe = get_pipe_info(file);
1589         if (!pipe)
1590                 return -EBADF;
1591
1592         pipe_lock(pipe);
1593
1594         error = ret = 0;
1595         while (nr_segs) {
1596                 void __user *base;
1597                 size_t len;
1598
1599                 /*
1600                  * Get user address base and length for this iovec.
1601                  */
1602                 error = get_user(base, &iov->iov_base);
1603                 if (unlikely(error))
1604                         break;
1605                 error = get_user(len, &iov->iov_len);
1606                 if (unlikely(error))
1607                         break;
1608
1609                 /*
1610                  * Sanity check this iovec. 0 read succeeds.
1611                  */
1612                 if (unlikely(!len))
1613                         break;
1614                 if (unlikely(!base)) {
1615                         error = -EFAULT;
1616                         break;
1617                 }
1618
1619                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1620                         error = -EFAULT;
1621                         break;
1622                 }
1623
1624                 sd.len = 0;
1625                 sd.total_len = len;
1626                 sd.flags = flags;
1627                 sd.u.userptr = base;
1628                 sd.pos = 0;
1629
1630                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1631                 if (size < 0) {
1632                         if (!ret)
1633                                 ret = size;
1634
1635                         break;
1636                 }
1637
1638                 ret += size;
1639
1640                 if (size < len)
1641                         break;
1642
1643                 nr_segs--;
1644                 iov++;
1645         }
1646
1647         pipe_unlock(pipe);
1648
1649         if (!ret)
1650                 ret = error;
1651
1652         return ret;
1653 }
1654
1655 /*
1656  * vmsplice splices a user address range into a pipe. It can be thought of
1657  * as splice-from-memory, where the regular splice is splice-from-file (or
1658  * to file). In both cases the output is a pipe, naturally.
1659  */
1660 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1661                              unsigned long nr_segs, unsigned int flags)
1662 {
1663         struct pipe_inode_info *pipe;
1664         struct page *pages[PIPE_DEF_BUFFERS];
1665         struct partial_page partial[PIPE_DEF_BUFFERS];
1666         struct splice_pipe_desc spd = {
1667                 .pages = pages,
1668                 .partial = partial,
1669                 .nr_pages_max = PIPE_DEF_BUFFERS,
1670                 .flags = flags,
1671                 .ops = &user_page_pipe_buf_ops,
1672                 .spd_release = spd_release_page,
1673         };
1674         long ret;
1675
1676         pipe = get_pipe_info(file);
1677         if (!pipe)
1678                 return -EBADF;
1679
1680         if (splice_grow_spd(pipe, &spd))
1681                 return -ENOMEM;
1682
1683         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1684                                             spd.partial, flags & SPLICE_F_GIFT,
1685                                             spd.nr_pages_max);
1686         if (spd.nr_pages <= 0)
1687                 ret = spd.nr_pages;
1688         else
1689                 ret = splice_to_pipe(pipe, &spd);
1690
1691         splice_shrink_spd(&spd);
1692         return ret;
1693 }
1694
1695 /*
1696  * Note that vmsplice only really supports true splicing _from_ user memory
1697  * to a pipe, not the other way around. Splicing from user memory is a simple
1698  * operation that can be supported without any funky alignment restrictions
1699  * or nasty vm tricks. We simply map in the user memory and fill them into
1700  * a pipe. The reverse isn't quite as easy, though. There are two possible
1701  * solutions for that:
1702  *
1703  *      - memcpy() the data internally, at which point we might as well just
1704  *        do a regular read() on the buffer anyway.
1705  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1706  *        has restriction limitations on both ends of the pipe).
1707  *
1708  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1709  *
1710  */
1711 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1712                 unsigned long, nr_segs, unsigned int, flags)
1713 {
1714         struct file *file;
1715         long error;
1716         int fput;
1717
1718         if (unlikely(nr_segs > UIO_MAXIOV))
1719                 return -EINVAL;
1720         else if (unlikely(!nr_segs))
1721                 return 0;
1722
1723         error = -EBADF;
1724         file = fget_light(fd, &fput);
1725         if (file) {
1726                 if (file->f_mode & FMODE_WRITE)
1727                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1728                 else if (file->f_mode & FMODE_READ)
1729                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1730
1731                 fput_light(file, fput);
1732         }
1733
1734         return error;
1735 }
1736
1737 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1738                 int, fd_out, loff_t __user *, off_out,
1739                 size_t, len, unsigned int, flags)
1740 {
1741         long error;
1742         struct file *in, *out;
1743         int fput_in, fput_out;
1744
1745         if (unlikely(!len))
1746                 return 0;
1747
1748         error = -EBADF;
1749         in = fget_light(fd_in, &fput_in);
1750         if (in) {
1751                 if (in->f_mode & FMODE_READ) {
1752                         out = fget_light(fd_out, &fput_out);
1753                         if (out) {
1754                                 if (out->f_mode & FMODE_WRITE)
1755                                         error = do_splice(in, off_in,
1756                                                           out, off_out,
1757                                                           len, flags);
1758                                 fput_light(out, fput_out);
1759                         }
1760                 }
1761
1762                 fput_light(in, fput_in);
1763         }
1764
1765         return error;
1766 }
1767
1768 /*
1769  * Make sure there's data to read. Wait for input if we can, otherwise
1770  * return an appropriate error.
1771  */
1772 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1773 {
1774         int ret;
1775
1776         /*
1777          * Check ->nrbufs without the inode lock first. This function
1778          * is speculative anyways, so missing one is ok.
1779          */
1780         if (pipe->nrbufs)
1781                 return 0;
1782
1783         ret = 0;
1784         pipe_lock(pipe);
1785
1786         while (!pipe->nrbufs) {
1787                 if (signal_pending(current)) {
1788                         ret = -ERESTARTSYS;
1789                         break;
1790                 }
1791                 if (!pipe->writers)
1792                         break;
1793                 if (!pipe->waiting_writers) {
1794                         if (flags & SPLICE_F_NONBLOCK) {
1795                                 ret = -EAGAIN;
1796                                 break;
1797                         }
1798                 }
1799                 pipe_wait(pipe);
1800         }
1801
1802         pipe_unlock(pipe);
1803         return ret;
1804 }
1805
1806 /*
1807  * Make sure there's writeable room. Wait for room if we can, otherwise
1808  * return an appropriate error.
1809  */
1810 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1811 {
1812         int ret;
1813
1814         /*
1815          * Check ->nrbufs without the inode lock first. This function
1816          * is speculative anyways, so missing one is ok.
1817          */
1818         if (pipe->nrbufs < pipe->buffers)
1819                 return 0;
1820
1821         ret = 0;
1822         pipe_lock(pipe);
1823
1824         while (pipe->nrbufs >= pipe->buffers) {
1825                 if (!pipe->readers) {
1826                         send_sig(SIGPIPE, current, 0);
1827                         ret = -EPIPE;
1828                         break;
1829                 }
1830                 if (flags & SPLICE_F_NONBLOCK) {
1831                         ret = -EAGAIN;
1832                         break;
1833                 }
1834                 if (signal_pending(current)) {
1835                         ret = -ERESTARTSYS;
1836                         break;
1837                 }
1838                 pipe->waiting_writers++;
1839                 pipe_wait(pipe);
1840                 pipe->waiting_writers--;
1841         }
1842
1843         pipe_unlock(pipe);
1844         return ret;
1845 }
1846
1847 /*
1848  * Splice contents of ipipe to opipe.
1849  */
1850 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1851                                struct pipe_inode_info *opipe,
1852                                size_t len, unsigned int flags)
1853 {
1854         struct pipe_buffer *ibuf, *obuf;
1855         int ret = 0, nbuf;
1856         bool input_wakeup = false;
1857
1858
1859 retry:
1860         ret = ipipe_prep(ipipe, flags);
1861         if (ret)
1862                 return ret;
1863
1864         ret = opipe_prep(opipe, flags);
1865         if (ret)
1866                 return ret;
1867
1868         /*
1869          * Potential ABBA deadlock, work around it by ordering lock
1870          * grabbing by pipe info address. Otherwise two different processes
1871          * could deadlock (one doing tee from A -> B, the other from B -> A).
1872          */
1873         pipe_double_lock(ipipe, opipe);
1874
1875         do {
1876                 if (!opipe->readers) {
1877                         send_sig(SIGPIPE, current, 0);
1878                         if (!ret)
1879                                 ret = -EPIPE;
1880                         break;
1881                 }
1882
1883                 if (!ipipe->nrbufs && !ipipe->writers)
1884                         break;
1885
1886                 /*
1887                  * Cannot make any progress, because either the input
1888                  * pipe is empty or the output pipe is full.
1889                  */
1890                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1891                         /* Already processed some buffers, break */
1892                         if (ret)
1893                                 break;
1894
1895                         if (flags & SPLICE_F_NONBLOCK) {
1896                                 ret = -EAGAIN;
1897                                 break;
1898                         }
1899
1900                         /*
1901                          * We raced with another reader/writer and haven't
1902                          * managed to process any buffers.  A zero return
1903                          * value means EOF, so retry instead.
1904                          */
1905                         pipe_unlock(ipipe);
1906                         pipe_unlock(opipe);
1907                         goto retry;
1908                 }
1909
1910                 ibuf = ipipe->bufs + ipipe->curbuf;
1911                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1912                 obuf = opipe->bufs + nbuf;
1913
1914                 if (len >= ibuf->len) {
1915                         /*
1916                          * Simply move the whole buffer from ipipe to opipe
1917                          */
1918                         *obuf = *ibuf;
1919                         ibuf->ops = NULL;
1920                         opipe->nrbufs++;
1921                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1922                         ipipe->nrbufs--;
1923                         input_wakeup = true;
1924                 } else {
1925                         /*
1926                          * Get a reference to this pipe buffer,
1927                          * so we can copy the contents over.
1928                          */
1929                         ibuf->ops->get(ipipe, ibuf);
1930                         *obuf = *ibuf;
1931
1932                         /*
1933                          * Don't inherit the gift flag, we need to
1934                          * prevent multiple steals of this page.
1935                          */
1936                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1937
1938                         obuf->len = len;
1939                         opipe->nrbufs++;
1940                         ibuf->offset += obuf->len;
1941                         ibuf->len -= obuf->len;
1942                 }
1943                 ret += obuf->len;
1944                 len -= obuf->len;
1945         } while (len);
1946
1947         pipe_unlock(ipipe);
1948         pipe_unlock(opipe);
1949
1950         /*
1951          * If we put data in the output pipe, wakeup any potential readers.
1952          */
1953         if (ret > 0)
1954                 wakeup_pipe_readers(opipe);
1955
1956         if (input_wakeup)
1957                 wakeup_pipe_writers(ipipe);
1958
1959         return ret;
1960 }
1961
1962 /*
1963  * Link contents of ipipe to opipe.
1964  */
1965 static int link_pipe(struct pipe_inode_info *ipipe,
1966                      struct pipe_inode_info *opipe,
1967                      size_t len, unsigned int flags)
1968 {
1969         struct pipe_buffer *ibuf, *obuf;
1970         int ret = 0, i = 0, nbuf;
1971
1972         /*
1973          * Potential ABBA deadlock, work around it by ordering lock
1974          * grabbing by pipe info address. Otherwise two different processes
1975          * could deadlock (one doing tee from A -> B, the other from B -> A).
1976          */
1977         pipe_double_lock(ipipe, opipe);
1978
1979         do {
1980                 if (!opipe->readers) {
1981                         send_sig(SIGPIPE, current, 0);
1982                         if (!ret)
1983                                 ret = -EPIPE;
1984                         break;
1985                 }
1986
1987                 /*
1988                  * If we have iterated all input buffers or ran out of
1989                  * output room, break.
1990                  */
1991                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1992                         break;
1993
1994                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1995                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1996
1997                 /*
1998                  * Get a reference to this pipe buffer,
1999                  * so we can copy the contents over.
2000                  */
2001                 ibuf->ops->get(ipipe, ibuf);
2002
2003                 obuf = opipe->bufs + nbuf;
2004                 *obuf = *ibuf;
2005
2006                 /*
2007                  * Don't inherit the gift flag, we need to
2008                  * prevent multiple steals of this page.
2009                  */
2010                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2011
2012                 if (obuf->len > len)
2013                         obuf->len = len;
2014
2015                 opipe->nrbufs++;
2016                 ret += obuf->len;
2017                 len -= obuf->len;
2018                 i++;
2019         } while (len);
2020
2021         /*
2022          * return EAGAIN if we have the potential of some data in the
2023          * future, otherwise just return 0
2024          */
2025         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2026                 ret = -EAGAIN;
2027
2028         pipe_unlock(ipipe);
2029         pipe_unlock(opipe);
2030
2031         /*
2032          * If we put data in the output pipe, wakeup any potential readers.
2033          */
2034         if (ret > 0)
2035                 wakeup_pipe_readers(opipe);
2036
2037         return ret;
2038 }
2039
2040 /*
2041  * This is a tee(1) implementation that works on pipes. It doesn't copy
2042  * any data, it simply references the 'in' pages on the 'out' pipe.
2043  * The 'flags' used are the SPLICE_F_* variants, currently the only
2044  * applicable one is SPLICE_F_NONBLOCK.
2045  */
2046 static long do_tee(struct file *in, struct file *out, size_t len,
2047                    unsigned int flags)
2048 {
2049         struct pipe_inode_info *ipipe = get_pipe_info(in);
2050         struct pipe_inode_info *opipe = get_pipe_info(out);
2051         int ret = -EINVAL;
2052
2053         /*
2054          * Duplicate the contents of ipipe to opipe without actually
2055          * copying the data.
2056          */
2057         if (ipipe && opipe && ipipe != opipe) {
2058                 /*
2059                  * Keep going, unless we encounter an error. The ipipe/opipe
2060                  * ordering doesn't really matter.
2061                  */
2062                 ret = ipipe_prep(ipipe, flags);
2063                 if (!ret) {
2064                         ret = opipe_prep(opipe, flags);
2065                         if (!ret)
2066                                 ret = link_pipe(ipipe, opipe, len, flags);
2067                 }
2068         }
2069
2070         return ret;
2071 }
2072
2073 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2074 {
2075         struct file *in;
2076         int error, fput_in;
2077
2078         if (unlikely(!len))
2079                 return 0;
2080
2081         error = -EBADF;
2082         in = fget_light(fdin, &fput_in);
2083         if (in) {
2084                 if (in->f_mode & FMODE_READ) {
2085                         int fput_out;
2086                         struct file *out = fget_light(fdout, &fput_out);
2087
2088                         if (out) {
2089                                 if (out->f_mode & FMODE_WRITE)
2090                                         error = do_tee(in, out, len, flags);
2091                                 fput_light(out, fput_out);
2092                         }
2093                 }
2094                 fput_light(in, fput_in);
2095         }
2096
2097         return error;
2098 }