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