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 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 EXPORT_SYMBOL(do_splice_from);
1156
1157 /*
1158  * Attempt to initiate a splice from a file to a pipe.
1159  */
1160 long do_splice_to(struct file *in, loff_t *ppos,
1161                   struct pipe_inode_info *pipe, size_t len,
1162                   unsigned int flags)
1163 {
1164         ssize_t (*splice_read)(struct file *, loff_t *,
1165                                struct pipe_inode_info *, size_t, unsigned int);
1166         int ret;
1167
1168         if (unlikely(!(in->f_mode & FMODE_READ)))
1169                 return -EBADF;
1170
1171         ret = rw_verify_area(READ, in, ppos, len);
1172         if (unlikely(ret < 0))
1173                 return ret;
1174
1175         if (in->f_op && in->f_op->splice_read)
1176                 splice_read = in->f_op->splice_read;
1177         else
1178                 splice_read = default_file_splice_read;
1179
1180         return splice_read(in, ppos, pipe, len, flags);
1181 }
1182 EXPORT_SYMBOL(do_splice_to);
1183
1184 /**
1185  * splice_direct_to_actor - splices data directly between two non-pipes
1186  * @in:         file to splice from
1187  * @sd:         actor information on where to splice to
1188  * @actor:      handles the data splicing
1189  *
1190  * Description:
1191  *    This is a special case helper to splice directly between two
1192  *    points, without requiring an explicit pipe. Internally an allocated
1193  *    pipe is cached in the process, and reused during the lifetime of
1194  *    that process.
1195  *
1196  */
1197 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1198                                splice_direct_actor *actor)
1199 {
1200         struct pipe_inode_info *pipe;
1201         long ret, bytes;
1202         umode_t i_mode;
1203         size_t len;
1204         int i, flags, more;
1205
1206         /*
1207          * We require the input being a regular file, as we don't want to
1208          * randomly drop data for eg socket -> socket splicing. Use the
1209          * piped splicing for that!
1210          */
1211         i_mode = in->f_path.dentry->d_inode->i_mode;
1212         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1213                 return -EINVAL;
1214
1215         /*
1216          * neither in nor out is a pipe, setup an internal pipe attached to
1217          * 'out' and transfer the wanted data from 'in' to 'out' through that
1218          */
1219         pipe = current->splice_pipe;
1220         if (unlikely(!pipe)) {
1221                 pipe = alloc_pipe_info(NULL);
1222                 if (!pipe)
1223                         return -ENOMEM;
1224
1225                 /*
1226                  * We don't have an immediate reader, but we'll read the stuff
1227                  * out of the pipe right after the splice_to_pipe(). So set
1228                  * PIPE_READERS appropriately.
1229                  */
1230                 pipe->readers = 1;
1231
1232                 current->splice_pipe = pipe;
1233         }
1234
1235         /*
1236          * Do the splice.
1237          */
1238         ret = 0;
1239         bytes = 0;
1240         len = sd->total_len;
1241         flags = sd->flags;
1242
1243         /*
1244          * Don't block on output, we have to drain the direct pipe.
1245          */
1246         sd->flags &= ~SPLICE_F_NONBLOCK;
1247         more = sd->flags & SPLICE_F_MORE;
1248
1249         while (len) {
1250                 size_t read_len;
1251                 loff_t pos = sd->pos, prev_pos = pos;
1252
1253                 ret = do_splice_to(in, &pos, pipe, len, flags);
1254                 if (unlikely(ret <= 0))
1255                         goto out_release;
1256
1257                 read_len = ret;
1258                 sd->total_len = read_len;
1259
1260                 /*
1261                  * If more data is pending, set SPLICE_F_MORE
1262                  * If this is the last data and SPLICE_F_MORE was not set
1263                  * initially, clears it.
1264                  */
1265                 if (read_len < len)
1266                         sd->flags |= SPLICE_F_MORE;
1267                 else if (!more)
1268                         sd->flags &= ~SPLICE_F_MORE;
1269                 /*
1270                  * NOTE: nonblocking mode only applies to the input. We
1271                  * must not do the output in nonblocking mode as then we
1272                  * could get stuck data in the internal pipe:
1273                  */
1274                 ret = actor(pipe, sd);
1275                 if (unlikely(ret <= 0)) {
1276                         sd->pos = prev_pos;
1277                         goto out_release;
1278                 }
1279
1280                 bytes += ret;
1281                 len -= ret;
1282                 sd->pos = pos;
1283
1284                 if (ret < read_len) {
1285                         sd->pos = prev_pos + ret;
1286                         goto out_release;
1287                 }
1288         }
1289
1290 done:
1291         pipe->nrbufs = pipe->curbuf = 0;
1292         file_accessed(in);
1293         return bytes;
1294
1295 out_release:
1296         /*
1297          * If we did an incomplete transfer we must release
1298          * the pipe buffers in question:
1299          */
1300         for (i = 0; i < pipe->buffers; i++) {
1301                 struct pipe_buffer *buf = pipe->bufs + i;
1302
1303                 if (buf->ops) {
1304                         buf->ops->release(pipe, buf);
1305                         buf->ops = NULL;
1306                 }
1307         }
1308
1309         if (!bytes)
1310                 bytes = ret;
1311
1312         goto done;
1313 }
1314 EXPORT_SYMBOL(splice_direct_to_actor);
1315
1316 static int direct_splice_actor(struct pipe_inode_info *pipe,
1317                                struct splice_desc *sd)
1318 {
1319         struct file *file = sd->u.file;
1320
1321         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1322                               sd->flags);
1323 }
1324
1325 /**
1326  * do_splice_direct - splices data directly between two files
1327  * @in:         file to splice from
1328  * @ppos:       input file offset
1329  * @out:        file to splice to
1330  * @len:        number of bytes to splice
1331  * @flags:      splice modifier flags
1332  *
1333  * Description:
1334  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1335  *    doing it in the application would incur an extra system call
1336  *    (splice in + splice out, as compared to just sendfile()). So this helper
1337  *    can splice directly through a process-private pipe.
1338  *
1339  */
1340 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1341                       size_t len, unsigned int flags)
1342 {
1343         struct splice_desc sd = {
1344                 .len            = len,
1345                 .total_len      = len,
1346                 .flags          = flags,
1347                 .pos            = *ppos,
1348                 .u.file         = out,
1349         };
1350         long ret;
1351
1352         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1353         if (ret > 0)
1354                 *ppos = sd.pos;
1355
1356         return ret;
1357 }
1358
1359 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1360                                struct pipe_inode_info *opipe,
1361                                size_t len, unsigned int flags);
1362
1363 /*
1364  * Determine where to splice to/from.
1365  */
1366 static long do_splice(struct file *in, loff_t __user *off_in,
1367                       struct file *out, loff_t __user *off_out,
1368                       size_t len, unsigned int flags)
1369 {
1370         struct pipe_inode_info *ipipe;
1371         struct pipe_inode_info *opipe;
1372         loff_t offset, *off;
1373         long ret;
1374
1375         ipipe = get_pipe_info(in);
1376         opipe = get_pipe_info(out);
1377
1378         if (ipipe && opipe) {
1379                 if (off_in || off_out)
1380                         return -ESPIPE;
1381
1382                 if (!(in->f_mode & FMODE_READ))
1383                         return -EBADF;
1384
1385                 if (!(out->f_mode & FMODE_WRITE))
1386                         return -EBADF;
1387
1388                 /* Splicing to self would be fun, but... */
1389                 if (ipipe == opipe)
1390                         return -EINVAL;
1391
1392                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1393         }
1394
1395         if (ipipe) {
1396                 if (off_in)
1397                         return -ESPIPE;
1398                 if (off_out) {
1399                         if (!(out->f_mode & FMODE_PWRITE))
1400                                 return -EINVAL;
1401                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1402                                 return -EFAULT;
1403                         off = &offset;
1404                 } else
1405                         off = &out->f_pos;
1406
1407                 ret = do_splice_from(ipipe, out, off, len, flags);
1408
1409                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1410                         ret = -EFAULT;
1411
1412                 return ret;
1413         }
1414
1415         if (opipe) {
1416                 if (off_out)
1417                         return -ESPIPE;
1418                 if (off_in) {
1419                         if (!(in->f_mode & FMODE_PREAD))
1420                                 return -EINVAL;
1421                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1422                                 return -EFAULT;
1423                         off = &offset;
1424                 } else
1425                         off = &in->f_pos;
1426
1427                 ret = do_splice_to(in, off, opipe, len, flags);
1428
1429                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1430                         ret = -EFAULT;
1431
1432                 return ret;
1433         }
1434
1435         return -EINVAL;
1436 }
1437
1438 /*
1439  * Map an iov into an array of pages and offset/length tupples. With the
1440  * partial_page structure, we can map several non-contiguous ranges into
1441  * our ones pages[] map instead of splitting that operation into pieces.
1442  * Could easily be exported as a generic helper for other users, in which
1443  * case one would probably want to add a 'max_nr_pages' parameter as well.
1444  */
1445 static int get_iovec_page_array(const struct iovec __user *iov,
1446                                 unsigned int nr_vecs, struct page **pages,
1447                                 struct partial_page *partial, int aligned,
1448                                 unsigned int pipe_buffers)
1449 {
1450         int buffers = 0, error = 0;
1451
1452         while (nr_vecs) {
1453                 unsigned long off, npages;
1454                 struct iovec entry;
1455                 void __user *base;
1456                 size_t len;
1457                 int i;
1458
1459                 error = -EFAULT;
1460                 if (copy_from_user(&entry, iov, sizeof(entry)))
1461                         break;
1462
1463                 base = entry.iov_base;
1464                 len = entry.iov_len;
1465
1466                 /*
1467                  * Sanity check this iovec. 0 read succeeds.
1468                  */
1469                 error = 0;
1470                 if (unlikely(!len))
1471                         break;
1472                 error = -EFAULT;
1473                 if (!access_ok(VERIFY_READ, base, len))
1474                         break;
1475
1476                 /*
1477                  * Get this base offset and number of pages, then map
1478                  * in the user pages.
1479                  */
1480                 off = (unsigned long) base & ~PAGE_MASK;
1481
1482                 /*
1483                  * If asked for alignment, the offset must be zero and the
1484                  * length a multiple of the PAGE_SIZE.
1485                  */
1486                 error = -EINVAL;
1487                 if (aligned && (off || len & ~PAGE_MASK))
1488                         break;
1489
1490                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1491                 if (npages > pipe_buffers - buffers)
1492                         npages = pipe_buffers - buffers;
1493
1494                 error = get_user_pages_fast((unsigned long)base, npages,
1495                                         0, &pages[buffers]);
1496
1497                 if (unlikely(error <= 0))
1498                         break;
1499
1500                 /*
1501                  * Fill this contiguous range into the partial page map.
1502                  */
1503                 for (i = 0; i < error; i++) {
1504                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1505
1506                         partial[buffers].offset = off;
1507                         partial[buffers].len = plen;
1508
1509                         off = 0;
1510                         len -= plen;
1511                         buffers++;
1512                 }
1513
1514                 /*
1515                  * We didn't complete this iov, stop here since it probably
1516                  * means we have to move some of this into a pipe to
1517                  * be able to continue.
1518                  */
1519                 if (len)
1520                         break;
1521
1522                 /*
1523                  * Don't continue if we mapped fewer pages than we asked for,
1524                  * or if we mapped the max number of pages that we have
1525                  * room for.
1526                  */
1527                 if (error < npages || buffers == pipe_buffers)
1528                         break;
1529
1530                 nr_vecs--;
1531                 iov++;
1532         }
1533
1534         if (buffers)
1535                 return buffers;
1536
1537         return error;
1538 }
1539
1540 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1541                         struct splice_desc *sd)
1542 {
1543         char *src;
1544         int ret;
1545
1546         /*
1547          * See if we can use the atomic maps, by prefaulting in the
1548          * pages and doing an atomic copy
1549          */
1550         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1551                 src = buf->ops->map(pipe, buf, 1);
1552                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1553                                                         sd->len);
1554                 buf->ops->unmap(pipe, buf, src);
1555                 if (!ret) {
1556                         ret = sd->len;
1557                         goto out;
1558                 }
1559         }
1560
1561         /*
1562          * No dice, use slow non-atomic map and copy
1563          */
1564         src = buf->ops->map(pipe, buf, 0);
1565
1566         ret = sd->len;
1567         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1568                 ret = -EFAULT;
1569
1570         buf->ops->unmap(pipe, buf, src);
1571 out:
1572         if (ret > 0)
1573                 sd->u.userptr += ret;
1574         return ret;
1575 }
1576
1577 /*
1578  * For lack of a better implementation, implement vmsplice() to userspace
1579  * as a simple copy of the pipes pages to the user iov.
1580  */
1581 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1582                              unsigned long nr_segs, unsigned int flags)
1583 {
1584         struct pipe_inode_info *pipe;
1585         struct splice_desc sd;
1586         ssize_t size;
1587         int error;
1588         long ret;
1589
1590         pipe = get_pipe_info(file);
1591         if (!pipe)
1592                 return -EBADF;
1593
1594         pipe_lock(pipe);
1595
1596         error = ret = 0;
1597         while (nr_segs) {
1598                 void __user *base;
1599                 size_t len;
1600
1601                 /*
1602                  * Get user address base and length for this iovec.
1603                  */
1604                 error = get_user(base, &iov->iov_base);
1605                 if (unlikely(error))
1606                         break;
1607                 error = get_user(len, &iov->iov_len);
1608                 if (unlikely(error))
1609                         break;
1610
1611                 /*
1612                  * Sanity check this iovec. 0 read succeeds.
1613                  */
1614                 if (unlikely(!len))
1615                         break;
1616                 if (unlikely(!base)) {
1617                         error = -EFAULT;
1618                         break;
1619                 }
1620
1621                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1622                         error = -EFAULT;
1623                         break;
1624                 }
1625
1626                 sd.len = 0;
1627                 sd.total_len = len;
1628                 sd.flags = flags;
1629                 sd.u.userptr = base;
1630                 sd.pos = 0;
1631
1632                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1633                 if (size < 0) {
1634                         if (!ret)
1635                                 ret = size;
1636
1637                         break;
1638                 }
1639
1640                 ret += size;
1641
1642                 if (size < len)
1643                         break;
1644
1645                 nr_segs--;
1646                 iov++;
1647         }
1648
1649         pipe_unlock(pipe);
1650
1651         if (!ret)
1652                 ret = error;
1653
1654         return ret;
1655 }
1656
1657 /*
1658  * vmsplice splices a user address range into a pipe. It can be thought of
1659  * as splice-from-memory, where the regular splice is splice-from-file (or
1660  * to file). In both cases the output is a pipe, naturally.
1661  */
1662 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1663                              unsigned long nr_segs, unsigned int flags)
1664 {
1665         struct pipe_inode_info *pipe;
1666         struct page *pages[PIPE_DEF_BUFFERS];
1667         struct partial_page partial[PIPE_DEF_BUFFERS];
1668         struct splice_pipe_desc spd = {
1669                 .pages = pages,
1670                 .partial = partial,
1671                 .nr_pages_max = PIPE_DEF_BUFFERS,
1672                 .flags = flags,
1673                 .ops = &user_page_pipe_buf_ops,
1674                 .spd_release = spd_release_page,
1675         };
1676         long ret;
1677
1678         pipe = get_pipe_info(file);
1679         if (!pipe)
1680                 return -EBADF;
1681
1682         if (splice_grow_spd(pipe, &spd))
1683                 return -ENOMEM;
1684
1685         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1686                                             spd.partial, flags & SPLICE_F_GIFT,
1687                                             spd.nr_pages_max);
1688         if (spd.nr_pages <= 0)
1689                 ret = spd.nr_pages;
1690         else
1691                 ret = splice_to_pipe(pipe, &spd);
1692
1693         splice_shrink_spd(&spd);
1694         return ret;
1695 }
1696
1697 /*
1698  * Note that vmsplice only really supports true splicing _from_ user memory
1699  * to a pipe, not the other way around. Splicing from user memory is a simple
1700  * operation that can be supported without any funky alignment restrictions
1701  * or nasty vm tricks. We simply map in the user memory and fill them into
1702  * a pipe. The reverse isn't quite as easy, though. There are two possible
1703  * solutions for that:
1704  *
1705  *      - memcpy() the data internally, at which point we might as well just
1706  *        do a regular read() on the buffer anyway.
1707  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1708  *        has restriction limitations on both ends of the pipe).
1709  *
1710  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1711  *
1712  */
1713 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1714                 unsigned long, nr_segs, unsigned int, flags)
1715 {
1716         struct file *file;
1717         long error;
1718         int fput;
1719
1720         if (unlikely(nr_segs > UIO_MAXIOV))
1721                 return -EINVAL;
1722         else if (unlikely(!nr_segs))
1723                 return 0;
1724
1725         error = -EBADF;
1726         file = fget_light(fd, &fput);
1727         if (file) {
1728                 if (file->f_mode & FMODE_WRITE)
1729                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1730                 else if (file->f_mode & FMODE_READ)
1731                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1732
1733                 fput_light(file, fput);
1734         }
1735
1736         return error;
1737 }
1738
1739 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1740                 int, fd_out, loff_t __user *, off_out,
1741                 size_t, len, unsigned int, flags)
1742 {
1743         long error;
1744         struct file *in, *out;
1745         int fput_in, fput_out;
1746
1747         if (unlikely(!len))
1748                 return 0;
1749
1750         error = -EBADF;
1751         in = fget_light(fd_in, &fput_in);
1752         if (in) {
1753                 if (in->f_mode & FMODE_READ) {
1754                         out = fget_light(fd_out, &fput_out);
1755                         if (out) {
1756                                 if (out->f_mode & FMODE_WRITE)
1757                                         error = do_splice(in, off_in,
1758                                                           out, off_out,
1759                                                           len, flags);
1760                                 fput_light(out, fput_out);
1761                         }
1762                 }
1763
1764                 fput_light(in, fput_in);
1765         }
1766
1767         return error;
1768 }
1769
1770 /*
1771  * Make sure there's data to read. Wait for input if we can, otherwise
1772  * return an appropriate error.
1773  */
1774 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1775 {
1776         int ret;
1777
1778         /*
1779          * Check ->nrbufs without the inode lock first. This function
1780          * is speculative anyways, so missing one is ok.
1781          */
1782         if (pipe->nrbufs)
1783                 return 0;
1784
1785         ret = 0;
1786         pipe_lock(pipe);
1787
1788         while (!pipe->nrbufs) {
1789                 if (signal_pending(current)) {
1790                         ret = -ERESTARTSYS;
1791                         break;
1792                 }
1793                 if (!pipe->writers)
1794                         break;
1795                 if (!pipe->waiting_writers) {
1796                         if (flags & SPLICE_F_NONBLOCK) {
1797                                 ret = -EAGAIN;
1798                                 break;
1799                         }
1800                 }
1801                 pipe_wait(pipe);
1802         }
1803
1804         pipe_unlock(pipe);
1805         return ret;
1806 }
1807
1808 /*
1809  * Make sure there's writeable room. Wait for room if we can, otherwise
1810  * return an appropriate error.
1811  */
1812 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1813 {
1814         int ret;
1815
1816         /*
1817          * Check ->nrbufs without the inode lock first. This function
1818          * is speculative anyways, so missing one is ok.
1819          */
1820         if (pipe->nrbufs < pipe->buffers)
1821                 return 0;
1822
1823         ret = 0;
1824         pipe_lock(pipe);
1825
1826         while (pipe->nrbufs >= pipe->buffers) {
1827                 if (!pipe->readers) {
1828                         send_sig(SIGPIPE, current, 0);
1829                         ret = -EPIPE;
1830                         break;
1831                 }
1832                 if (flags & SPLICE_F_NONBLOCK) {
1833                         ret = -EAGAIN;
1834                         break;
1835                 }
1836                 if (signal_pending(current)) {
1837                         ret = -ERESTARTSYS;
1838                         break;
1839                 }
1840                 pipe->waiting_writers++;
1841                 pipe_wait(pipe);
1842                 pipe->waiting_writers--;
1843         }
1844
1845         pipe_unlock(pipe);
1846         return ret;
1847 }
1848
1849 /*
1850  * Splice contents of ipipe to opipe.
1851  */
1852 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1853                                struct pipe_inode_info *opipe,
1854                                size_t len, unsigned int flags)
1855 {
1856         struct pipe_buffer *ibuf, *obuf;
1857         int ret = 0, nbuf;
1858         bool input_wakeup = false;
1859
1860
1861 retry:
1862         ret = ipipe_prep(ipipe, flags);
1863         if (ret)
1864                 return ret;
1865
1866         ret = opipe_prep(opipe, flags);
1867         if (ret)
1868                 return ret;
1869
1870         /*
1871          * Potential ABBA deadlock, work around it by ordering lock
1872          * grabbing by pipe info address. Otherwise two different processes
1873          * could deadlock (one doing tee from A -> B, the other from B -> A).
1874          */
1875         pipe_double_lock(ipipe, opipe);
1876
1877         do {
1878                 if (!opipe->readers) {
1879                         send_sig(SIGPIPE, current, 0);
1880                         if (!ret)
1881                                 ret = -EPIPE;
1882                         break;
1883                 }
1884
1885                 if (!ipipe->nrbufs && !ipipe->writers)
1886                         break;
1887
1888                 /*
1889                  * Cannot make any progress, because either the input
1890                  * pipe is empty or the output pipe is full.
1891                  */
1892                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1893                         /* Already processed some buffers, break */
1894                         if (ret)
1895                                 break;
1896
1897                         if (flags & SPLICE_F_NONBLOCK) {
1898                                 ret = -EAGAIN;
1899                                 break;
1900                         }
1901
1902                         /*
1903                          * We raced with another reader/writer and haven't
1904                          * managed to process any buffers.  A zero return
1905                          * value means EOF, so retry instead.
1906                          */
1907                         pipe_unlock(ipipe);
1908                         pipe_unlock(opipe);
1909                         goto retry;
1910                 }
1911
1912                 ibuf = ipipe->bufs + ipipe->curbuf;
1913                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1914                 obuf = opipe->bufs + nbuf;
1915
1916                 if (len >= ibuf->len) {
1917                         /*
1918                          * Simply move the whole buffer from ipipe to opipe
1919                          */
1920                         *obuf = *ibuf;
1921                         ibuf->ops = NULL;
1922                         opipe->nrbufs++;
1923                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1924                         ipipe->nrbufs--;
1925                         input_wakeup = true;
1926                 } else {
1927                         /*
1928                          * Get a reference to this pipe buffer,
1929                          * so we can copy the contents over.
1930                          */
1931                         ibuf->ops->get(ipipe, ibuf);
1932                         *obuf = *ibuf;
1933
1934                         /*
1935                          * Don't inherit the gift flag, we need to
1936                          * prevent multiple steals of this page.
1937                          */
1938                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1939
1940                         obuf->len = len;
1941                         opipe->nrbufs++;
1942                         ibuf->offset += obuf->len;
1943                         ibuf->len -= obuf->len;
1944                 }
1945                 ret += obuf->len;
1946                 len -= obuf->len;
1947         } while (len);
1948
1949         pipe_unlock(ipipe);
1950         pipe_unlock(opipe);
1951
1952         /*
1953          * If we put data in the output pipe, wakeup any potential readers.
1954          */
1955         if (ret > 0)
1956                 wakeup_pipe_readers(opipe);
1957
1958         if (input_wakeup)
1959                 wakeup_pipe_writers(ipipe);
1960
1961         return ret;
1962 }
1963
1964 /*
1965  * Link contents of ipipe to opipe.
1966  */
1967 static int link_pipe(struct pipe_inode_info *ipipe,
1968                      struct pipe_inode_info *opipe,
1969                      size_t len, unsigned int flags)
1970 {
1971         struct pipe_buffer *ibuf, *obuf;
1972         int ret = 0, i = 0, nbuf;
1973
1974         /*
1975          * Potential ABBA deadlock, work around it by ordering lock
1976          * grabbing by pipe info address. Otherwise two different processes
1977          * could deadlock (one doing tee from A -> B, the other from B -> A).
1978          */
1979         pipe_double_lock(ipipe, opipe);
1980
1981         do {
1982                 if (!opipe->readers) {
1983                         send_sig(SIGPIPE, current, 0);
1984                         if (!ret)
1985                                 ret = -EPIPE;
1986                         break;
1987                 }
1988
1989                 /*
1990                  * If we have iterated all input buffers or ran out of
1991                  * output room, break.
1992                  */
1993                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1994                         break;
1995
1996                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1997                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1998
1999                 /*
2000                  * Get a reference to this pipe buffer,
2001                  * so we can copy the contents over.
2002                  */
2003                 ibuf->ops->get(ipipe, ibuf);
2004
2005                 obuf = opipe->bufs + nbuf;
2006                 *obuf = *ibuf;
2007
2008                 /*
2009                  * Don't inherit the gift flag, we need to
2010                  * prevent multiple steals of this page.
2011                  */
2012                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2013
2014                 if (obuf->len > len)
2015                         obuf->len = len;
2016
2017                 opipe->nrbufs++;
2018                 ret += obuf->len;
2019                 len -= obuf->len;
2020                 i++;
2021         } while (len);
2022
2023         /*
2024          * return EAGAIN if we have the potential of some data in the
2025          * future, otherwise just return 0
2026          */
2027         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2028                 ret = -EAGAIN;
2029
2030         pipe_unlock(ipipe);
2031         pipe_unlock(opipe);
2032
2033         /*
2034          * If we put data in the output pipe, wakeup any potential readers.
2035          */
2036         if (ret > 0)
2037                 wakeup_pipe_readers(opipe);
2038
2039         return ret;
2040 }
2041
2042 /*
2043  * This is a tee(1) implementation that works on pipes. It doesn't copy
2044  * any data, it simply references the 'in' pages on the 'out' pipe.
2045  * The 'flags' used are the SPLICE_F_* variants, currently the only
2046  * applicable one is SPLICE_F_NONBLOCK.
2047  */
2048 static long do_tee(struct file *in, struct file *out, size_t len,
2049                    unsigned int flags)
2050 {
2051         struct pipe_inode_info *ipipe = get_pipe_info(in);
2052         struct pipe_inode_info *opipe = get_pipe_info(out);
2053         int ret = -EINVAL;
2054
2055         /*
2056          * Duplicate the contents of ipipe to opipe without actually
2057          * copying the data.
2058          */
2059         if (ipipe && opipe && ipipe != opipe) {
2060                 /*
2061                  * Keep going, unless we encounter an error. The ipipe/opipe
2062                  * ordering doesn't really matter.
2063                  */
2064                 ret = ipipe_prep(ipipe, flags);
2065                 if (!ret) {
2066                         ret = opipe_prep(opipe, flags);
2067                         if (!ret)
2068                                 ret = link_pipe(ipipe, opipe, len, flags);
2069                 }
2070         }
2071
2072         return ret;
2073 }
2074
2075 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2076 {
2077         struct file *in;
2078         int error, fput_in;
2079
2080         if (unlikely(!len))
2081                 return 0;
2082
2083         error = -EBADF;
2084         in = fget_light(fdin, &fput_in);
2085         if (in) {
2086                 if (in->f_mode & FMODE_READ) {
2087                         int fput_out;
2088                         struct file *out = fget_light(fdout, &fput_out);
2089
2090                         if (out) {
2091                                 if (out->f_mode & FMODE_WRITE)
2092                                         error = do_tee(in, out, len, flags);
2093                                 fput_light(out, fput_out);
2094                         }
2095                 }
2096                 fput_light(in, fput_in);
2097         }
2098
2099         return error;
2100 }