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/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34
  35 /*
  36  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  37  * a vm helper function, it's already simplified quite a bit by the
  38  * addition of remove_mapping(). If success is returned, the caller may
  39  * attempt to reuse this page for another destination.
  40  */
  41 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  42                                      struct pipe_buffer *buf)
  43 {
  44         struct page *page = buf->page;
  45         struct address_space *mapping;
  46
  47         lock_page(page);
  48
  49         mapping = page_mapping(page);
  50         if (mapping) {
  51                 WARN_ON(!PageUptodate(page));
  52
  53                 /*
  54                  * At least for ext2 with nobh option, we need to wait on
  55                  * writeback completing on this page, since we'll remove it
  56                  * from the pagecache.  Otherwise truncate wont wait on the
  57                  * page, allowing the disk blocks to be reused by someone else
  58                  * before we actually wrote our data to them. fs corruption
  59                  * ensues.
  60                  */
  61                 wait_on_page_writeback(page);
  62
  63                 if (page_has_private(page) &&
  64                     !try_to_release_page(page, GFP_KERNEL))
  65                         goto out_unlock;
  66
  67                 /*
  68                  * If we succeeded in removing the mapping, set LRU flag
  69                  * and return good.
  70                  */
  71                 if (remove_mapping(mapping, page)) {
  72                         buf->flags |= PIPE_BUF_FLAG_LRU;
  73                         return 0;
  74                 }
  75         }
  76
  77         /*
  78          * Raced with truncate or failed to remove page from current
  79          * address space, unlock and return failure.
  80          */
  81 out_unlock:
  82         unlock_page(page);
  83         return 1;
  84 }
  85
  86 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  87                                         struct pipe_buffer *buf)
  88 {
  89         page_cache_release(buf->page);
  90         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  91 }
  92
  93 /*
  94  * Check whether the contents of buf is OK to access. Since the content
  95  * is a page cache page, IO may be in flight.
  96  */
  97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  98                                        struct pipe_buffer *buf)
  99 {
 100         struct page *page = buf->page;
 101         int err;
 102
 103         if (!PageUptodate(page)) {
 104                 lock_page(page);
 105
 106                 /*
 107                  * Page got truncated/unhashed. This will cause a 0-byte
 108                  * splice, if this is the first page.
 109                  */
 110                 if (!page->mapping) {
 111                         err = -ENODATA;
 112                         goto error;
 113                 }
 114
 115                 /*
 116                  * Uh oh, read-error from disk.
 117                  */
 118                 if (!PageUptodate(page)) {
 119                         err = -EIO;
 120                         goto error;
 121                 }
 122
 123                 /*
 124                  * Page is ok afterall, we are done.
 125                  */
 126                 unlock_page(page);
 127         }
 128
 129         return 0;
 130 error:
 131         unlock_page(page);
 132         return err;
 133 }
 134
 135 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 136         .can_merge = 0,
 137         .map = generic_pipe_buf_map,
 138         .unmap = generic_pipe_buf_unmap,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .map = generic_pipe_buf_map,
 158         .unmap = generic_pipe_buf_unmap,
 159         .confirm = generic_pipe_buf_confirm,
 160         .release = page_cache_pipe_buf_release,
 161         .steal = user_page_pipe_buf_steal,
 162         .get = generic_pipe_buf_get,
 163 };
 164
 165 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 166 {
 167         smp_mb();
 168         if (waitqueue_active(&pipe->wait))
 169                 wake_up_interruptible(&pipe->wait);
 170         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 171 }
 172
 173 /**
 174  * splice_to_pipe - fill passed data into a pipe
 175  * @pipe:       pipe to fill
 176  * @spd:        data to fill
 177  *
 178  * Description:
 179  *    @spd contains a map of pages and len/offset tuples, along with
 180  *    the struct pipe_buf_operations associated with these pages. This
 181  *    function will link that data to the pipe.
 182  *
 183  */
 184 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 185                        struct splice_pipe_desc *spd)
 186 {
 187         unsigned int spd_pages = spd->nr_pages;
 188         int ret, do_wakeup, page_nr;
 189
 190         ret = 0;
 191         do_wakeup = 0;
 192         page_nr = 0;
 193
 194         pipe_lock(pipe);
 195
 196         for (;;) {
 197                 if (!pipe->readers) {
 198                         send_sig(SIGPIPE, current, 0);
 199                         if (!ret)
 200                                 ret = -EPIPE;
 201                         break;
 202                 }
 203
 204                 if (pipe->nrbufs < pipe->buffers) {
 205                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 206                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 207
 208                         buf->page = spd->pages[page_nr];
 209                         buf->offset = spd->partial[page_nr].offset;
 210                         buf->len = spd->partial[page_nr].len;
 211                         buf->private = spd->partial[page_nr].private;
 212                         buf->ops = spd->ops;
 213                         if (spd->flags & SPLICE_F_GIFT)
 214                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 215
 216                         pipe->nrbufs++;
 217                         page_nr++;
 218                         ret += buf->len;
 219
 220                         if (pipe->inode)
 221                                 do_wakeup = 1;
 222
 223                         if (!--spd->nr_pages)
 224                                 break;
 225                         if (pipe->nrbufs < pipe->buffers)
 226                                 continue;
 227
 228                         break;
 229                 }
 230
 231                 if (spd->flags & SPLICE_F_NONBLOCK) {
 232                         if (!ret)
 233                                 ret = -EAGAIN;
 234                         break;
 235                 }
 236
 237                 if (signal_pending(current)) {
 238                         if (!ret)
 239                                 ret = -ERESTARTSYS;
 240                         break;
 241                 }
 242
 243                 if (do_wakeup) {
 244                         smp_mb();
 245                         if (waitqueue_active(&pipe->wait))
 246                                 wake_up_interruptible_sync(&pipe->wait);
 247                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 248                         do_wakeup = 0;
 249                 }
 250
 251                 pipe->waiting_writers++;
 252                 pipe_wait(pipe);
 253                 pipe->waiting_writers--;
 254         }
 255
 256         pipe_unlock(pipe);
 257
 258         if (do_wakeup)
 259                 wakeup_pipe_readers(pipe);
 260
 261         while (page_nr < spd_pages)
 262                 spd->spd_release(spd, page_nr++);
 263
 264         return ret;
 265 }
 266
 267 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 268 {
 269         page_cache_release(spd->pages[i]);
 270 }
 271
 272 /*
 273  * Check if we need to grow the arrays holding pages and partial page
 274  * descriptions.
 275  */
 276 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 277 {
 278         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 279
 280         spd->nr_pages_max = buffers;
 281         if (buffers <= PIPE_DEF_BUFFERS)
 282                 return 0;
 283
 284         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 285         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 286
 287         if (spd->pages && spd->partial)
 288                 return 0;
 289
 290         kfree(spd->pages);
 291         kfree(spd->partial);
 292         return -ENOMEM;
 293 }
 294
 295 void splice_shrink_spd(struct splice_pipe_desc *spd)
 296 {
 297         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 298                 return;
 299
 300         kfree(spd->pages);
 301         kfree(spd->partial);
 302 }
 303
 304 static int
 305 __generic_file_splice_read(struct file *in, loff_t *ppos,
 306                            struct pipe_inode_info *pipe, size_t len,
 307                            unsigned int flags)
 308 {
 309         struct address_space *mapping = in->f_mapping;
 310         unsigned int loff, nr_pages, req_pages;
 311         struct page *pages[PIPE_DEF_BUFFERS];
 312         struct partial_page partial[PIPE_DEF_BUFFERS];
 313         struct page *page;
 314         pgoff_t index, end_index;
 315         loff_t isize;
 316         int error, page_nr;
 317         struct splice_pipe_desc spd = {
 318                 .pages = pages,
 319                 .partial = partial,
 320                 .nr_pages_max = PIPE_DEF_BUFFERS,
 321                 .flags = flags,
 322                 .ops = &page_cache_pipe_buf_ops,
 323                 .spd_release = spd_release_page,
 324         };
 325
 326         if (splice_grow_spd(pipe, &spd))
 327                 return -ENOMEM;
 328
 329         index = *ppos >> PAGE_CACHE_SHIFT;
 330         loff = *ppos & ~PAGE_CACHE_MASK;
 331         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 332         nr_pages = min(req_pages, spd.nr_pages_max);
 333
 334         /*
 335          * Lookup the (hopefully) full range of pages we need.
 336          */
 337         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 338         index += spd.nr_pages;
 339
 340         /*
 341          * If find_get_pages_contig() returned fewer pages than we needed,
 342          * readahead/allocate the rest and fill in the holes.
 343          */
 344         if (spd.nr_pages < nr_pages)
 345                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 346                                 index, req_pages - spd.nr_pages);
 347
 348         error = 0;
 349         while (spd.nr_pages < nr_pages) {
 350                 /*
 351                  * Page could be there, find_get_pages_contig() breaks on
 352                  * the first hole.
 353                  */
 354                 page = find_get_page(mapping, index);
 355                 if (!page) {
 356                         /*
 357                          * page didn't exist, allocate one.
 358                          */
 359                         page = page_cache_alloc_cold(mapping);
 360                         if (!page)
 361                                 break;
 362
 363                         error = add_to_page_cache_lru(page, mapping, index,
 364                                                 GFP_KERNEL);
 365                         if (unlikely(error)) {
 366                                 page_cache_release(page);
 367                                 if (error == -EEXIST)
 368                                         continue;
 369                                 break;
 370                         }
 371                         /*
 372                          * add_to_page_cache() locks the page, unlock it
 373                          * to avoid convoluting the logic below even more.
 374                          */
 375                         unlock_page(page);
 376                 }
 377
 378                 spd.pages[spd.nr_pages++] = page;
 379                 index++;
 380         }
 381
 382         /*
 383          * Now loop over the map and see if we need to start IO on any
 384          * pages, fill in the partial map, etc.
 385          */
 386         index = *ppos >> PAGE_CACHE_SHIFT;
 387         nr_pages = spd.nr_pages;
 388         spd.nr_pages = 0;
 389         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 390                 unsigned int this_len;
 391
 392                 if (!len)
 393                         break;
 394
 395                 /*
 396                  * this_len is the max we'll use from this page
 397                  */
 398                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 399                 page = spd.pages[page_nr];
 400
 401                 if (PageReadahead(page))
 402                         page_cache_async_readahead(mapping, &in->f_ra, in,
 403                                         page, index, req_pages - page_nr);
 404
 405                 /*
 406                  * If the page isn't uptodate, we may need to start io on it
 407                  */
 408                 if (!PageUptodate(page)) {
 409                         lock_page(page);
 410
 411                         /*
 412                          * Page was truncated, or invalidated by the
 413                          * filesystem.  Redo the find/create, but this time the
 414                          * page is kept locked, so there's no chance of another
 415                          * race with truncate/invalidate.
 416                          */
 417                         if (!page->mapping) {
 418                                 unlock_page(page);
 419                                 page = find_or_create_page(mapping, index,
 420                                                 mapping_gfp_mask(mapping));
 421
 422                                 if (!page) {
 423                                         error = -ENOMEM;
 424                                         break;
 425                                 }
 426                                 page_cache_release(spd.pages[page_nr]);
 427                                 spd.pages[page_nr] = page;
 428                         }
 429                         /*
 430                          * page was already under io and is now done, great
 431                          */
 432                         if (PageUptodate(page)) {
 433                                 unlock_page(page);
 434                                 goto fill_it;
 435                         }
 436
 437                         /*
 438                          * need to read in the page
 439                          */
 440                         error = mapping->a_ops->readpage(in, page);
 441                         if (unlikely(error)) {
 442                                 /*
 443                                  * We really should re-lookup the page here,
 444                                  * but it complicates things a lot. Instead
 445                                  * lets just do what we already stored, and
 446                                  * we'll get it the next time we are called.
 447                                  */
 448                                 if (error == AOP_TRUNCATED_PAGE)
 449                                         error = 0;
 450
 451                                 break;
 452                         }
 453                 }
 454 fill_it:
 455                 /*
 456                  * i_size must be checked after PageUptodate.
 457                  */
 458                 isize = i_size_read(mapping->host);
 459                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 460                 if (unlikely(!isize || index > end_index))
 461                         break;
 462
 463                 /*
 464                  * if this is the last page, see if we need to shrink
 465                  * the length and stop
 466                  */
 467                 if (end_index == index) {
 468                         unsigned int plen;
 469
 470                         /*
 471                          * max good bytes in this page
 472                          */
 473                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 474                         if (plen <= loff)
 475                                 break;
 476
 477                         /*
 478                          * force quit after adding this page
 479                          */
 480                         this_len = min(this_len, plen - loff);
 481                         len = this_len;
 482                 }
 483
 484                 spd.partial[page_nr].offset = loff;
 485                 spd.partial[page_nr].len = this_len;
 486                 len -= this_len;
 487                 loff = 0;
 488                 spd.nr_pages++;
 489                 index++;
 490         }
 491
 492         /*
 493          * Release any pages at the end, if we quit early. 'page_nr' is how far
 494          * we got, 'nr_pages' is how many pages are in the map.
 495          */
 496         while (page_nr < nr_pages)
 497                 page_cache_release(spd.pages[page_nr++]);
 498         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 499
 500         if (spd.nr_pages)
 501                 error = splice_to_pipe(pipe, &spd);
 502
 503         splice_shrink_spd(&spd);
 504         return error;
 505 }
 506
 507 /**
 508  * generic_file_splice_read - splice data from file to a pipe
 509  * @in:         file to splice from
 510  * @ppos:       position in @in
 511  * @pipe:       pipe to splice to
 512  * @len:        number of bytes to splice
 513  * @flags:      splice modifier flags
 514  *
 515  * Description:
 516  *    Will read pages from given file and fill them into a pipe. Can be
 517  *    used as long as the address_space operations for the source implements
 518  *    a readpage() hook.
 519  *
 520  */
 521 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 522                                  struct pipe_inode_info *pipe, size_t len,
 523                                  unsigned int flags)
 524 {
 525         loff_t isize, left;
 526         int ret;
 527
 528         isize = i_size_read(in->f_mapping->host);
 529         if (unlikely(*ppos >= isize))
 530                 return 0;
 531
 532         left = isize - *ppos;
 533         if (unlikely(left < len))
 534                 len = left;
 535
 536         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 537         if (ret > 0) {
 538                 *ppos += ret;
 539                 file_accessed(in);
 540         }
 541
 542         return ret;
 543 }
 544 EXPORT_SYMBOL(generic_file_splice_read);
 545
 546 static const struct pipe_buf_operations default_pipe_buf_ops = {
 547         .can_merge = 0,
 548         .map = generic_pipe_buf_map,
 549         .unmap = generic_pipe_buf_unmap,
 550         .confirm = generic_pipe_buf_confirm,
 551         .release = generic_pipe_buf_release,
 552         .steal = generic_pipe_buf_steal,
 553         .get = generic_pipe_buf_get,
 554 };
 555
 556 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 557                             unsigned long vlen, loff_t offset)
 558 {
 559         mm_segment_t old_fs;
 560         loff_t pos = offset;
 561         ssize_t res;
 562
 563         old_fs = get_fs();
 564         set_fs(get_ds());
 565         /* The cast to a user pointer is valid due to the set_fs() */
 566         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 567         set_fs(old_fs);
 568
 569         return res;
 570 }
 571
 572 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 573                             loff_t pos)
 574 {
 575         mm_segment_t old_fs;
 576         ssize_t res;
 577
 578         old_fs = get_fs();
 579         set_fs(get_ds());
 580         /* The cast to a user pointer is valid due to the set_fs() */
 581         res = vfs_write(file, (const char __user *)buf, count, &pos);
 582         set_fs(old_fs);
 583
 584         return res;
 585 }
 586
 587 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 588                                  struct pipe_inode_info *pipe, size_t len,
 589                                  unsigned int flags)
 590 {
 591         unsigned int nr_pages;
 592         unsigned int nr_freed;
 593         size_t offset;
 594         struct page *pages[PIPE_DEF_BUFFERS];
 595         struct partial_page partial[PIPE_DEF_BUFFERS];
 596         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 597         ssize_t res;
 598         size_t this_len;
 599         int error;
 600         int i;
 601         struct splice_pipe_desc spd = {
 602                 .pages = pages,
 603                 .partial = partial,
 604                 .nr_pages_max = PIPE_DEF_BUFFERS,
 605                 .flags = flags,
 606                 .ops = &default_pipe_buf_ops,
 607                 .spd_release = spd_release_page,
 608         };
 609
 610         if (splice_grow_spd(pipe, &spd))
 611                 return -ENOMEM;
 612
 613         res = -ENOMEM;
 614         vec = __vec;
 615         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 616                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 617                 if (!vec)
 618                         goto shrink_ret;
 619         }
 620
 621         offset = *ppos & ~PAGE_CACHE_MASK;
 622         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 623
 624         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 625                 struct page *page;
 626
 627                 page = alloc_page(GFP_USER);
 628                 error = -ENOMEM;
 629                 if (!page)
 630                         goto err;
 631
 632                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 633                 vec[i].iov_base = (void __user *) page_address(page);
 634                 vec[i].iov_len = this_len;
 635                 spd.pages[i] = page;
 636                 spd.nr_pages++;
 637                 len -= this_len;
 638                 offset = 0;
 639         }
 640
 641         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 642         if (res < 0) {
 643                 error = res;
 644                 goto err;
 645         }
 646
 647         error = 0;
 648         if (!res)
 649                 goto err;
 650
 651         nr_freed = 0;
 652         for (i = 0; i < spd.nr_pages; i++) {
 653                 this_len = min_t(size_t, vec[i].iov_len, res);
 654                 spd.partial[i].offset = 0;
 655                 spd.partial[i].len = this_len;
 656                 if (!this_len) {
 657                         __free_page(spd.pages[i]);
 658                         spd.pages[i] = NULL;
 659                         nr_freed++;
 660                 }
 661                 res -= this_len;
 662         }
 663         spd.nr_pages -= nr_freed;
 664
 665         res = splice_to_pipe(pipe, &spd);
 666         if (res > 0)
 667                 *ppos += res;
 668
 669 shrink_ret:
 670         if (vec != __vec)
 671                 kfree(vec);
 672         splice_shrink_spd(&spd);
 673         return res;
 674
 675 err:
 676         for (i = 0; i < spd.nr_pages; i++)
 677                 __free_page(spd.pages[i]);
 678
 679         res = error;
 680         goto shrink_ret;
 681 }
 682 EXPORT_SYMBOL(default_file_splice_read);
 683
 684 /*
 685  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 686  * using sendpage(). Return the number of bytes sent.
 687  */
 688 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 689                             struct pipe_buffer *buf, struct splice_desc *sd)
 690 {
 691         struct file *file = sd->u.file;
 692         loff_t pos = sd->pos;
 693         int more;
 694
 695         if (!likely(file->f_op && file->f_op->sendpage))
 696                 return -EINVAL;
 697
 698         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 699         if (sd->len < sd->total_len)
 700                 more |= MSG_SENDPAGE_NOTLAST;
 701         return file->f_op->sendpage(file, buf->page, buf->offset,
 702                                     sd->len, &pos, more);
 703 }
 704
 705 /*
 706  * This is a little more tricky than the file -> pipe splicing. There are
 707  * basically three cases:
 708  *
 709  *      - Destination page already exists in the address space and there
 710  *        are users of it. For that case we have no other option that
 711  *        copying the data. Tough luck.
 712  *      - Destination page already exists in the address space, but there
 713  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 714  *        through to last case.
 715  *      - Destination page does not exist, we can add the pipe page to
 716  *        the page cache and avoid the copy.
 717  *
 718  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 719  * sd->flags), we attempt to migrate pages from the pipe to the output
 720  * file address space page cache. This is possible if no one else has
 721  * the pipe page referenced outside of the pipe and page cache. If
 722  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 723  * a new page in the output file page cache and fill/dirty that.
 724  */
 725 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 726                  struct splice_desc *sd)
 727 {
 728         struct file *file = sd->u.file;
 729         struct address_space *mapping = file->f_mapping;
 730         unsigned int offset, this_len;
 731         struct page *page;
 732         void *fsdata;
 733         int ret;
 734
 735         offset = sd->pos & ~PAGE_CACHE_MASK;
 736
 737         this_len = sd->len;
 738         if (this_len + offset > PAGE_CACHE_SIZE)
 739                 this_len = PAGE_CACHE_SIZE - offset;
 740
 741         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 742                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 743         if (unlikely(ret))
 744                 goto out;
 745
 746         if (buf->page != page) {
 747                 char *src = buf->ops->map(pipe, buf, 1);
 748                 char *dst = kmap_atomic(page);
 749
 750                 memcpy(dst + offset, src + buf->offset, this_len);
 751                 flush_dcache_page(page);
 752                 kunmap_atomic(dst);
 753                 buf->ops->unmap(pipe, buf, src);
 754         }
 755         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 756                                 page, fsdata);
 757 out:
 758         return ret;
 759 }
 760 EXPORT_SYMBOL(pipe_to_file);
 761
 762 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 763 {
 764         smp_mb();
 765         if (waitqueue_active(&pipe->wait))
 766                 wake_up_interruptible(&pipe->wait);
 767         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 768 }
 769
 770 /**
 771  * splice_from_pipe_feed - feed available data from a pipe to a file
 772  * @pipe:       pipe to splice from
 773  * @sd:         information to @actor
 774  * @actor:      handler that splices the data
 775  *
 776  * Description:
 777  *    This function loops over the pipe and calls @actor to do the
 778  *    actual moving of a single struct pipe_buffer to the desired
 779  *    destination.  It returns when there's no more buffers left in
 780  *    the pipe or if the requested number of bytes (@sd->total_len)
 781  *    have been copied.  It returns a positive number (one) if the
 782  *    pipe needs to be filled with more data, zero if the required
 783  *    number of bytes have been copied and -errno on error.
 784  *
 785  *    This, together with splice_from_pipe_{begin,end,next}, may be
 786  *    used to implement the functionality of __splice_from_pipe() when
 787  *    locking is required around copying the pipe buffers to the
 788  *    destination.
 789  */
 790 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 791                           splice_actor *actor)
 792 {
 793         int ret;
 794
 795         while (pipe->nrbufs) {
 796                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 797                 const struct pipe_buf_operations *ops = buf->ops;
 798
 799                 sd->len = buf->len;
 800                 if (sd->len > sd->total_len)
 801                         sd->len = sd->total_len;
 802
 803                 ret = buf->ops->confirm(pipe, buf);
 804                 if (unlikely(ret)) {
 805                         if (ret == -ENODATA)
 806                                 ret = 0;
 807                         return ret;
 808                 }
 809
 810                 ret = actor(pipe, buf, sd);
 811                 if (ret <= 0)
 812                         return ret;
 813
 814                 buf->offset += ret;
 815                 buf->len -= ret;
 816
 817                 sd->num_spliced += ret;
 818                 sd->len -= ret;
 819                 sd->pos += ret;
 820                 sd->total_len -= ret;
 821
 822                 if (!buf->len) {
 823                         buf->ops = NULL;
 824                         ops->release(pipe, buf);
 825                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 826                         pipe->nrbufs--;
 827                         if (pipe->inode)
 828                                 sd->need_wakeup = true;
 829                 }
 830
 831                 if (!sd->total_len)
 832                         return 0;
 833         }
 834
 835         return 1;
 836 }
 837 EXPORT_SYMBOL(splice_from_pipe_feed);
 838
 839 /**
 840  * splice_from_pipe_next - wait for some data to splice from
 841  * @pipe:       pipe to splice from
 842  * @sd:         information about the splice operation
 843  *
 844  * Description:
 845  *    This function will wait for some data and return a positive
 846  *    value (one) if pipe buffers are available.  It will return zero
 847  *    or -errno if no more data needs to be spliced.
 848  */
 849 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 850 {
 851         while (!pipe->nrbufs) {
 852                 if (!pipe->writers)
 853                         return 0;
 854
 855                 if (!pipe->waiting_writers && sd->num_spliced)
 856                         return 0;
 857
 858                 if (sd->flags & SPLICE_F_NONBLOCK)
 859                         return -EAGAIN;
 860
 861                 if (signal_pending(current))
 862                         return -ERESTARTSYS;
 863
 864                 if (sd->need_wakeup) {
 865                         wakeup_pipe_writers(pipe);
 866                         sd->need_wakeup = false;
 867                 }
 868
 869                 pipe_wait(pipe);
 870         }
 871
 872         return 1;
 873 }
 874 EXPORT_SYMBOL(splice_from_pipe_next);
 875
 876 /**
 877  * splice_from_pipe_begin - start splicing from pipe
 878  * @sd:         information about the splice operation
 879  *
 880  * Description:
 881  *    This function should be called before a loop containing
 882  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 883  *    initialize the necessary fields of @sd.
 884  */
 885 void splice_from_pipe_begin(struct splice_desc *sd)
 886 {
 887         sd->num_spliced = 0;
 888         sd->need_wakeup = false;
 889 }
 890 EXPORT_SYMBOL(splice_from_pipe_begin);
 891
 892 /**
 893  * splice_from_pipe_end - finish splicing from pipe
 894  * @pipe:       pipe to splice from
 895  * @sd:         information about the splice operation
 896  *
 897  * Description:
 898  *    This function will wake up pipe writers if necessary.  It should
 899  *    be called after a loop containing splice_from_pipe_next() and
 900  *    splice_from_pipe_feed().
 901  */
 902 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 903 {
 904         if (sd->need_wakeup)
 905                 wakeup_pipe_writers(pipe);
 906 }
 907 EXPORT_SYMBOL(splice_from_pipe_end);
 908
 909 /**
 910  * __splice_from_pipe - splice data from a pipe to given actor
 911  * @pipe:       pipe to splice from
 912  * @sd:         information to @actor
 913  * @actor:      handler that splices the data
 914  *
 915  * Description:
 916  *    This function does little more than loop over the pipe and call
 917  *    @actor to do the actual moving of a single struct pipe_buffer to
 918  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 919  *    pipe_to_user.
 920  *
 921  */
 922 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 923                            splice_actor *actor)
 924 {
 925         int ret;
 926
 927         splice_from_pipe_begin(sd);
 928         do {
 929                 ret = splice_from_pipe_next(pipe, sd);
 930                 if (ret > 0)
 931                         ret = splice_from_pipe_feed(pipe, sd, actor);
 932         } while (ret > 0);
 933         splice_from_pipe_end(pipe, sd);
 934
 935         return sd->num_spliced ? sd->num_spliced : ret;
 936 }
 937 EXPORT_SYMBOL(__splice_from_pipe);
 938
 939 /**
 940  * splice_from_pipe - splice data from a pipe to a file
 941  * @pipe:       pipe to splice from
 942  * @out:        file to splice to
 943  * @ppos:       position in @out
 944  * @len:        how many bytes to splice
 945  * @flags:      splice modifier flags
 946  * @actor:      handler that splices the data
 947  *
 948  * Description:
 949  *    See __splice_from_pipe. This function locks the pipe inode,
 950  *    otherwise it's identical to __splice_from_pipe().
 951  *
 952  */
 953 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 954                          loff_t *ppos, size_t len, unsigned int flags,
 955                          splice_actor *actor)
 956 {
 957         ssize_t ret;
 958         struct splice_desc sd = {
 959                 .total_len = len,
 960                 .flags = flags,
 961                 .pos = *ppos,
 962                 .u.file = out,
 963         };
 964
 965         pipe_lock(pipe);
 966         ret = __splice_from_pipe(pipe, &sd, actor);
 967         pipe_unlock(pipe);
 968
 969         return ret;
 970 }
 971
 972 /**
 973  * generic_file_splice_write - splice data from a pipe to a file
 974  * @pipe:       pipe info
 975  * @out:        file to write to
 976  * @ppos:       position in @out
 977  * @len:        number of bytes to splice
 978  * @flags:      splice modifier flags
 979  *
 980  * Description:
 981  *    Will either move or copy pages (determined by @flags options) from
 982  *    the given pipe inode to the given file.
 983  *
 984  */
 985 ssize_t
 986 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 987                           loff_t *ppos, size_t len, unsigned int flags)
 988 {
 989         struct address_space *mapping = out->f_mapping;
 990         struct inode *inode = mapping->host;
 991         struct splice_desc sd = {
 992                 .total_len = len,
 993                 .flags = flags,
 994                 .pos = *ppos,
 995                 .u.file = out,
 996         };
 997         ssize_t ret;
 998
 999         pipe_lock(pipe);
1000
1001         splice_from_pipe_begin(&sd);
1002         do {
1003                 ret = splice_from_pipe_next(pipe, &sd);
1004                 if (ret <= 0)
1005                         break;
1006
1007                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1008                 ret = file_remove_suid(out);
1009                 if (!ret) {
1010                         ret = file_update_time(out);
1011                         if (!ret)
1012                                 ret = splice_from_pipe_feed(pipe, &sd,
1013                                                             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, bool 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                 .nr_pages_max = PIPE_DEF_BUFFERS,
1622                 .flags = flags,
1623                 .ops = &user_page_pipe_buf_ops,
1624                 .spd_release = spd_release_page,
1625         };
1626         long ret;
1627
1628         pipe = get_pipe_info(file);
1629         if (!pipe)
1630                 return -EBADF;
1631
1632         if (splice_grow_spd(pipe, &spd))
1633                 return -ENOMEM;
1634
1635         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1636                                             spd.partial, false,
1637                                             spd.nr_pages_max);
1638         if (spd.nr_pages <= 0)
1639                 ret = spd.nr_pages;
1640         else
1641                 ret = splice_to_pipe(pipe, &spd);
1642
1643         splice_shrink_spd(&spd);
1644         return ret;
1645 }
1646
1647 /*
1648  * Note that vmsplice only really supports true splicing _from_ user memory
1649  * to a pipe, not the other way around. Splicing from user memory is a simple
1650  * operation that can be supported without any funky alignment restrictions
1651  * or nasty vm tricks. We simply map in the user memory and fill them into
1652  * a pipe. The reverse isn't quite as easy, though. There are two possible
1653  * solutions for that:
1654  *
1655  *      - memcpy() the data internally, at which point we might as well just
1656  *        do a regular read() on the buffer anyway.
1657  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1658  *        has restriction limitations on both ends of the pipe).
1659  *
1660  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1661  *
1662  */
1663 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1664                 unsigned long, nr_segs, unsigned int, flags)
1665 {
1666         struct file *file;
1667         long error;
1668         int fput;
1669
1670         if (unlikely(nr_segs > UIO_MAXIOV))
1671                 return -EINVAL;
1672         else if (unlikely(!nr_segs))
1673                 return 0;
1674
1675         error = -EBADF;
1676         file = fget_light(fd, &fput);
1677         if (file) {
1678                 if (file->f_mode & FMODE_WRITE)
1679                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1680                 else if (file->f_mode & FMODE_READ)
1681                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1682
1683                 fput_light(file, fput);
1684         }
1685
1686         return error;
1687 }
1688
1689 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1690                 int, fd_out, loff_t __user *, off_out,
1691                 size_t, len, unsigned int, flags)
1692 {
1693         long error;
1694         struct file *in, *out;
1695         int fput_in, fput_out;
1696
1697         if (unlikely(!len))
1698                 return 0;
1699
1700         error = -EBADF;
1701         in = fget_light(fd_in, &fput_in);
1702         if (in) {
1703                 if (in->f_mode & FMODE_READ) {
1704                         out = fget_light(fd_out, &fput_out);
1705                         if (out) {
1706                                 if (out->f_mode & FMODE_WRITE)
1707                                         error = do_splice(in, off_in,
1708                                                           out, off_out,
1709                                                           len, flags);
1710                                 fput_light(out, fput_out);
1711                         }
1712                 }
1713
1714                 fput_light(in, fput_in);
1715         }
1716
1717         return error;
1718 }
1719
1720 /*
1721  * Make sure there's data to read. Wait for input if we can, otherwise
1722  * return an appropriate error.
1723  */
1724 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1725 {
1726         int ret;
1727
1728         /*
1729          * Check ->nrbufs without the inode lock first. This function
1730          * is speculative anyways, so missing one is ok.
1731          */
1732         if (pipe->nrbufs)
1733                 return 0;
1734
1735         ret = 0;
1736         pipe_lock(pipe);
1737
1738         while (!pipe->nrbufs) {
1739                 if (signal_pending(current)) {
1740                         ret = -ERESTARTSYS;
1741                         break;
1742                 }
1743                 if (!pipe->writers)
1744                         break;
1745                 if (!pipe->waiting_writers) {
1746                         if (flags & SPLICE_F_NONBLOCK) {
1747                                 ret = -EAGAIN;
1748                                 break;
1749                         }
1750                 }
1751                 pipe_wait(pipe);
1752         }
1753
1754         pipe_unlock(pipe);
1755         return ret;
1756 }
1757
1758 /*
1759  * Make sure there's writeable room. Wait for room if we can, otherwise
1760  * return an appropriate error.
1761  */
1762 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1763 {
1764         int ret;
1765
1766         /*
1767          * Check ->nrbufs without the inode lock first. This function
1768          * is speculative anyways, so missing one is ok.
1769          */
1770         if (pipe->nrbufs < pipe->buffers)
1771                 return 0;
1772
1773         ret = 0;
1774         pipe_lock(pipe);
1775
1776         while (pipe->nrbufs >= pipe->buffers) {
1777                 if (!pipe->readers) {
1778                         send_sig(SIGPIPE, current, 0);
1779                         ret = -EPIPE;
1780                         break;
1781                 }
1782                 if (flags & SPLICE_F_NONBLOCK) {
1783                         ret = -EAGAIN;
1784                         break;
1785                 }
1786                 if (signal_pending(current)) {
1787                         ret = -ERESTARTSYS;
1788                         break;
1789                 }
1790                 pipe->waiting_writers++;
1791                 pipe_wait(pipe);
1792                 pipe->waiting_writers--;
1793         }
1794
1795         pipe_unlock(pipe);
1796         return ret;
1797 }
1798
1799 /*
1800  * Splice contents of ipipe to opipe.
1801  */
1802 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1803                                struct pipe_inode_info *opipe,
1804                                size_t len, unsigned int flags)
1805 {
1806         struct pipe_buffer *ibuf, *obuf;
1807         int ret = 0, nbuf;
1808         bool input_wakeup = false;
1809
1810
1811 retry:
1812         ret = ipipe_prep(ipipe, flags);
1813         if (ret)
1814                 return ret;
1815
1816         ret = opipe_prep(opipe, flags);
1817         if (ret)
1818                 return ret;
1819
1820         /*
1821          * Potential ABBA deadlock, work around it by ordering lock
1822          * grabbing by pipe info address. Otherwise two different processes
1823          * could deadlock (one doing tee from A -> B, the other from B -> A).
1824          */
1825         pipe_double_lock(ipipe, opipe);
1826
1827         do {
1828                 if (!opipe->readers) {
1829                         send_sig(SIGPIPE, current, 0);
1830                         if (!ret)
1831                                 ret = -EPIPE;
1832                         break;
1833                 }
1834
1835                 if (!ipipe->nrbufs && !ipipe->writers)
1836                         break;
1837
1838                 /*
1839                  * Cannot make any progress, because either the input
1840                  * pipe is empty or the output pipe is full.
1841                  */
1842                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1843                         /* Already processed some buffers, break */
1844                         if (ret)
1845                                 break;
1846
1847                         if (flags & SPLICE_F_NONBLOCK) {
1848                                 ret = -EAGAIN;
1849                                 break;
1850                         }
1851
1852                         /*
1853                          * We raced with another reader/writer and haven't
1854                          * managed to process any buffers.  A zero return
1855                          * value means EOF, so retry instead.
1856                          */
1857                         pipe_unlock(ipipe);
1858                         pipe_unlock(opipe);
1859                         goto retry;
1860                 }
1861
1862                 ibuf = ipipe->bufs + ipipe->curbuf;
1863                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1864                 obuf = opipe->bufs + nbuf;
1865
1866                 if (len >= ibuf->len) {
1867                         /*
1868                          * Simply move the whole buffer from ipipe to opipe
1869                          */
1870                         *obuf = *ibuf;
1871                         ibuf->ops = NULL;
1872                         opipe->nrbufs++;
1873                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1874                         ipipe->nrbufs--;
1875                         input_wakeup = true;
1876                 } else {
1877                         /*
1878                          * Get a reference to this pipe buffer,
1879                          * so we can copy the contents over.
1880                          */
1881                         ibuf->ops->get(ipipe, ibuf);
1882                         *obuf = *ibuf;
1883
1884                         /*
1885                          * Don't inherit the gift flag, we need to
1886                          * prevent multiple steals of this page.
1887                          */
1888                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1889
1890                         obuf->len = len;
1891                         opipe->nrbufs++;
1892                         ibuf->offset += obuf->len;
1893                         ibuf->len -= obuf->len;
1894                 }
1895                 ret += obuf->len;
1896                 len -= obuf->len;
1897         } while (len);
1898
1899         pipe_unlock(ipipe);
1900         pipe_unlock(opipe);
1901
1902         /*
1903          * If we put data in the output pipe, wakeup any potential readers.
1904          */
1905         if (ret > 0)
1906                 wakeup_pipe_readers(opipe);
1907
1908         if (input_wakeup)
1909                 wakeup_pipe_writers(ipipe);
1910
1911         return ret;
1912 }
1913
1914 /*
1915  * Link contents of ipipe to opipe.
1916  */
1917 static int link_pipe(struct pipe_inode_info *ipipe,
1918                      struct pipe_inode_info *opipe,
1919                      size_t len, unsigned int flags)
1920 {
1921         struct pipe_buffer *ibuf, *obuf;
1922         int ret = 0, i = 0, nbuf;
1923
1924         /*
1925          * Potential ABBA deadlock, work around it by ordering lock
1926          * grabbing by pipe info address. Otherwise two different processes
1927          * could deadlock (one doing tee from A -> B, the other from B -> A).
1928          */
1929         pipe_double_lock(ipipe, opipe);
1930
1931         do {
1932                 if (!opipe->readers) {
1933                         send_sig(SIGPIPE, current, 0);
1934                         if (!ret)
1935                                 ret = -EPIPE;
1936                         break;
1937                 }
1938
1939                 /*
1940                  * If we have iterated all input buffers or ran out of
1941                  * output room, break.
1942                  */
1943                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1944                         break;
1945
1946                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1947                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1948
1949                 /*
1950                  * Get a reference to this pipe buffer,
1951                  * so we can copy the contents over.
1952                  */
1953                 ibuf->ops->get(ipipe, ibuf);
1954
1955                 obuf = opipe->bufs + nbuf;
1956                 *obuf = *ibuf;
1957
1958                 /*
1959                  * Don't inherit the gift flag, we need to
1960                  * prevent multiple steals of this page.
1961                  */
1962                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1963
1964                 if (obuf->len > len)
1965                         obuf->len = len;
1966
1967                 opipe->nrbufs++;
1968                 ret += obuf->len;
1969                 len -= obuf->len;
1970                 i++;
1971         } while (len);
1972
1973         /*
1974          * return EAGAIN if we have the potential of some data in the
1975          * future, otherwise just return 0
1976          */
1977         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1978                 ret = -EAGAIN;
1979
1980         pipe_unlock(ipipe);
1981         pipe_unlock(opipe);
1982
1983         /*
1984          * If we put data in the output pipe, wakeup any potential readers.
1985          */
1986         if (ret > 0)
1987                 wakeup_pipe_readers(opipe);
1988
1989         return ret;
1990 }
1991
1992 /*
1993  * This is a tee(1) implementation that works on pipes. It doesn't copy
1994  * any data, it simply references the 'in' pages on the 'out' pipe.
1995  * The 'flags' used are the SPLICE_F_* variants, currently the only
1996  * applicable one is SPLICE_F_NONBLOCK.
1997  */
1998 static long do_tee(struct file *in, struct file *out, size_t len,
1999                    unsigned int flags)
2000 {
2001         struct pipe_inode_info *ipipe = get_pipe_info(in);
2002         struct pipe_inode_info *opipe = get_pipe_info(out);
2003         int ret = -EINVAL;
2004
2005         /*
2006          * Duplicate the contents of ipipe to opipe without actually
2007          * copying the data.
2008          */
2009         if (ipipe && opipe && ipipe != opipe) {
2010                 /*
2011                  * Keep going, unless we encounter an error. The ipipe/opipe
2012                  * ordering doesn't really matter.
2013                  */
2014                 ret = ipipe_prep(ipipe, flags);
2015                 if (!ret) {
2016                         ret = opipe_prep(opipe, flags);
2017                         if (!ret)
2018                                 ret = link_pipe(ipipe, opipe, len, flags);
2019                 }
2020         }
2021
2022         return ret;
2023 }
2024
2025 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2026 {
2027         struct file *in;
2028         int error, fput_in;
2029
2030         if (unlikely(!len))
2031                 return 0;
2032
2033         error = -EBADF;
2034         in = fget_light(fdin, &fput_in);
2035         if (in) {
2036                 if (in->f_mode & FMODE_READ) {
2037                         int fput_out;
2038                         struct file *out = fget_light(fdout, &fput_out);
2039
2040                         if (out) {
2041                                 if (out->f_mode & FMODE_WRITE)
2042                                         error = do_tee(in, out, len, flags);
2043                                 fput_light(out, fput_out);
2044                         }
2045                 }
2046                 fput_light(in, fput_in);
2047         }
2048
2049         return error;
2050 }