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