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
 700         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 701                 more |= MSG_SENDPAGE_NOTLAST;
 702
 703         return file->f_op->sendpage(file, buf->page, buf->offset,
 704                                     sd->len, &pos, more);
 705 }
 706
 707 /*
 708  * This is a little more tricky than the file -> pipe splicing. There are
 709  * basically three cases:
 710  *
 711  *      - Destination page already exists in the address space and there
 712  *        are users of it. For that case we have no other option that
 713  *        copying the data. Tough luck.
 714  *      - Destination page already exists in the address space, but there
 715  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 716  *        through to last case.
 717  *      - Destination page does not exist, we can add the pipe page to
 718  *        the page cache and avoid the copy.
 719  *
 720  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 721  * sd->flags), we attempt to migrate pages from the pipe to the output
 722  * file address space page cache. This is possible if no one else has
 723  * the pipe page referenced outside of the pipe and page cache. If
 724  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 725  * a new page in the output file page cache and fill/dirty that.
 726  */
 727 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 728                  struct splice_desc *sd)
 729 {
 730         struct file *file = sd->u.file;
 731         struct address_space *mapping = file->f_mapping;
 732         unsigned int offset, this_len;
 733         struct page *page;
 734         void *fsdata;
 735         int ret;
 736
 737         offset = sd->pos & ~PAGE_CACHE_MASK;
 738
 739         this_len = sd->len;
 740         if (this_len + offset > PAGE_CACHE_SIZE)
 741                 this_len = PAGE_CACHE_SIZE - offset;
 742
 743         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 744                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 745         if (unlikely(ret))
 746                 goto out;
 747
 748         if (buf->page != page) {
 749                 char *src = buf->ops->map(pipe, buf, 1);
 750                 char *dst = kmap_atomic(page);
 751
 752                 memcpy(dst + offset, src + buf->offset, this_len);
 753                 flush_dcache_page(page);
 754                 kunmap_atomic(dst);
 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         sb_start_write(inode->i_sb);
1002
1003         pipe_lock(pipe);
1004
1005         splice_from_pipe_begin(&sd);
1006         do {
1007                 ret = splice_from_pipe_next(pipe, &sd);
1008                 if (ret <= 0)
1009                         break;
1010
1011                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1012                 ret = file_remove_suid(out);
1013                 if (!ret) {
1014                         ret = file_update_time(out);
1015                         if (!ret)
1016                                 ret = splice_from_pipe_feed(pipe, &sd,
1017                                                             pipe_to_file);
1018                 }
1019                 mutex_unlock(&inode->i_mutex);
1020         } while (ret > 0);
1021         splice_from_pipe_end(pipe, &sd);
1022
1023         pipe_unlock(pipe);
1024
1025         if (sd.num_spliced)
1026                 ret = sd.num_spliced;
1027
1028         if (ret > 0) {
1029                 int err;
1030
1031                 err = generic_write_sync(out, *ppos, ret);
1032                 if (err)
1033                         ret = err;
1034                 else
1035                         *ppos += ret;
1036                 balance_dirty_pages_ratelimited(mapping);
1037         }
1038         sb_end_write(inode->i_sb);
1039
1040         return ret;
1041 }
1042
1043 EXPORT_SYMBOL(generic_file_splice_write);
1044
1045 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1046                           struct splice_desc *sd)
1047 {
1048         int ret;
1049         void *data;
1050
1051         data = buf->ops->map(pipe, buf, 0);
1052         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1053         buf->ops->unmap(pipe, buf, data);
1054
1055         return ret;
1056 }
1057
1058 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1059                                          struct file *out, loff_t *ppos,
1060                                          size_t len, unsigned int flags)
1061 {
1062         ssize_t ret;
1063
1064         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1065         if (ret > 0)
1066                 *ppos += ret;
1067
1068         return ret;
1069 }
1070
1071 /**
1072  * generic_splice_sendpage - splice data from a pipe to a socket
1073  * @pipe:       pipe to splice from
1074  * @out:        socket to write to
1075  * @ppos:       position in @out
1076  * @len:        number of bytes to splice
1077  * @flags:      splice modifier flags
1078  *
1079  * Description:
1080  *    Will send @len bytes from the pipe to a network socket. No data copying
1081  *    is involved.
1082  *
1083  */
1084 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1085                                 loff_t *ppos, size_t len, unsigned int flags)
1086 {
1087         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1088 }
1089
1090 EXPORT_SYMBOL(generic_splice_sendpage);
1091
1092 /*
1093  * Attempt to initiate a splice from pipe to file.
1094  */
1095 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1096                            loff_t *ppos, size_t len, unsigned int flags)
1097 {
1098         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1099                                 loff_t *, size_t, unsigned int);
1100         int ret;
1101
1102         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1103                 return -EBADF;
1104
1105         if (unlikely(out->f_flags & O_APPEND))
1106                 return -EINVAL;
1107
1108         ret = rw_verify_area(WRITE, out, ppos, len);
1109         if (unlikely(ret < 0))
1110                 return ret;
1111
1112         if (out->f_op && out->f_op->splice_write)
1113                 splice_write = out->f_op->splice_write;
1114         else
1115                 splice_write = default_file_splice_write;
1116
1117         return splice_write(pipe, out, ppos, len, flags);
1118 }
1119
1120 /*
1121  * Attempt to initiate a splice from a file to a pipe.
1122  */
1123 static long do_splice_to(struct file *in, loff_t *ppos,
1124                          struct pipe_inode_info *pipe, size_t len,
1125                          unsigned int flags)
1126 {
1127         ssize_t (*splice_read)(struct file *, loff_t *,
1128                                struct pipe_inode_info *, size_t, unsigned int);
1129         int ret;
1130
1131         if (unlikely(!(in->f_mode & FMODE_READ)))
1132                 return -EBADF;
1133
1134         ret = rw_verify_area(READ, in, ppos, len);
1135         if (unlikely(ret < 0))
1136                 return ret;
1137
1138         if (in->f_op && in->f_op->splice_read)
1139                 splice_read = in->f_op->splice_read;
1140         else
1141                 splice_read = default_file_splice_read;
1142
1143         return splice_read(in, ppos, pipe, len, flags);
1144 }
1145
1146 /**
1147  * splice_direct_to_actor - splices data directly between two non-pipes
1148  * @in:         file to splice from
1149  * @sd:         actor information on where to splice to
1150  * @actor:      handles the data splicing
1151  *
1152  * Description:
1153  *    This is a special case helper to splice directly between two
1154  *    points, without requiring an explicit pipe. Internally an allocated
1155  *    pipe is cached in the process, and reused during the lifetime of
1156  *    that process.
1157  *
1158  */
1159 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1160                                splice_direct_actor *actor)
1161 {
1162         struct pipe_inode_info *pipe;
1163         long ret, bytes;
1164         umode_t i_mode;
1165         size_t len;
1166         int i, flags;
1167
1168         /*
1169          * We require the input being a regular file, as we don't want to
1170          * randomly drop data for eg socket -> socket splicing. Use the
1171          * piped splicing for that!
1172          */
1173         i_mode = in->f_path.dentry->d_inode->i_mode;
1174         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1175                 return -EINVAL;
1176
1177         /*
1178          * neither in nor out is a pipe, setup an internal pipe attached to
1179          * 'out' and transfer the wanted data from 'in' to 'out' through that
1180          */
1181         pipe = current->splice_pipe;
1182         if (unlikely(!pipe)) {
1183                 pipe = alloc_pipe_info(NULL);
1184                 if (!pipe)
1185                         return -ENOMEM;
1186
1187                 /*
1188                  * We don't have an immediate reader, but we'll read the stuff
1189                  * out of the pipe right after the splice_to_pipe(). So set
1190                  * PIPE_READERS appropriately.
1191                  */
1192                 pipe->readers = 1;
1193
1194                 current->splice_pipe = pipe;
1195         }
1196
1197         /*
1198          * Do the splice.
1199          */
1200         ret = 0;
1201         bytes = 0;
1202         len = sd->total_len;
1203         flags = sd->flags;
1204
1205         /*
1206          * Don't block on output, we have to drain the direct pipe.
1207          */
1208         sd->flags &= ~SPLICE_F_NONBLOCK;
1209
1210         while (len) {
1211                 size_t read_len;
1212                 loff_t pos = sd->pos, prev_pos = pos;
1213
1214                 ret = do_splice_to(in, &pos, pipe, len, flags);
1215                 if (unlikely(ret <= 0))
1216                         goto out_release;
1217
1218                 read_len = ret;
1219                 sd->total_len = read_len;
1220
1221                 /*
1222                  * NOTE: nonblocking mode only applies to the input. We
1223                  * must not do the output in nonblocking mode as then we
1224                  * could get stuck data in the internal pipe:
1225                  */
1226                 ret = actor(pipe, sd);
1227                 if (unlikely(ret <= 0)) {
1228                         sd->pos = prev_pos;
1229                         goto out_release;
1230                 }
1231
1232                 bytes += ret;
1233                 len -= ret;
1234                 sd->pos = pos;
1235
1236                 if (ret < read_len) {
1237                         sd->pos = prev_pos + ret;
1238                         goto out_release;
1239                 }
1240         }
1241
1242 done:
1243         pipe->nrbufs = pipe->curbuf = 0;
1244         file_accessed(in);
1245         return bytes;
1246
1247 out_release:
1248         /*
1249          * If we did an incomplete transfer we must release
1250          * the pipe buffers in question:
1251          */
1252         for (i = 0; i < pipe->buffers; i++) {
1253                 struct pipe_buffer *buf = pipe->bufs + i;
1254
1255                 if (buf->ops) {
1256                         buf->ops->release(pipe, buf);
1257                         buf->ops = NULL;
1258                 }
1259         }
1260
1261         if (!bytes)
1262                 bytes = ret;
1263
1264         goto done;
1265 }
1266 EXPORT_SYMBOL(splice_direct_to_actor);
1267
1268 static int direct_splice_actor(struct pipe_inode_info *pipe,
1269                                struct splice_desc *sd)
1270 {
1271         struct file *file = sd->u.file;
1272
1273         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1274                               sd->flags);
1275 }
1276
1277 /**
1278  * do_splice_direct - splices data directly between two files
1279  * @in:         file to splice from
1280  * @ppos:       input file offset
1281  * @out:        file to splice to
1282  * @len:        number of bytes to splice
1283  * @flags:      splice modifier flags
1284  *
1285  * Description:
1286  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1287  *    doing it in the application would incur an extra system call
1288  *    (splice in + splice out, as compared to just sendfile()). So this helper
1289  *    can splice directly through a process-private pipe.
1290  *
1291  */
1292 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1293                       size_t len, unsigned int flags)
1294 {
1295         struct splice_desc sd = {
1296                 .len            = len,
1297                 .total_len      = len,
1298                 .flags          = flags,
1299                 .pos            = *ppos,
1300                 .u.file         = out,
1301         };
1302         long ret;
1303
1304         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1305         if (ret > 0)
1306                 *ppos = sd.pos;
1307
1308         return ret;
1309 }
1310
1311 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1312                                struct pipe_inode_info *opipe,
1313                                size_t len, unsigned int flags);
1314
1315 /*
1316  * Determine where to splice to/from.
1317  */
1318 static long do_splice(struct file *in, loff_t __user *off_in,
1319                       struct file *out, loff_t __user *off_out,
1320                       size_t len, unsigned int flags)
1321 {
1322         struct pipe_inode_info *ipipe;
1323         struct pipe_inode_info *opipe;
1324         loff_t offset, *off;
1325         long ret;
1326
1327         ipipe = get_pipe_info(in);
1328         opipe = get_pipe_info(out);
1329
1330         if (ipipe && opipe) {
1331                 if (off_in || off_out)
1332                         return -ESPIPE;
1333
1334                 if (!(in->f_mode & FMODE_READ))
1335                         return -EBADF;
1336
1337                 if (!(out->f_mode & FMODE_WRITE))
1338                         return -EBADF;
1339
1340                 /* Splicing to self would be fun, but... */
1341                 if (ipipe == opipe)
1342                         return -EINVAL;
1343
1344                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1345         }
1346
1347         if (ipipe) {
1348                 if (off_in)
1349                         return -ESPIPE;
1350                 if (off_out) {
1351                         if (!(out->f_mode & FMODE_PWRITE))
1352                                 return -EINVAL;
1353                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1354                                 return -EFAULT;
1355                         off = &offset;
1356                 } else
1357                         off = &out->f_pos;
1358
1359                 ret = do_splice_from(ipipe, out, off, len, flags);
1360
1361                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1362                         ret = -EFAULT;
1363
1364                 return ret;
1365         }
1366
1367         if (opipe) {
1368                 if (off_out)
1369                         return -ESPIPE;
1370                 if (off_in) {
1371                         if (!(in->f_mode & FMODE_PREAD))
1372                                 return -EINVAL;
1373                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1374                                 return -EFAULT;
1375                         off = &offset;
1376                 } else
1377                         off = &in->f_pos;
1378
1379                 ret = do_splice_to(in, off, opipe, len, flags);
1380
1381                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1382                         ret = -EFAULT;
1383
1384                 return ret;
1385         }
1386
1387         return -EINVAL;
1388 }
1389
1390 /*
1391  * Map an iov into an array of pages and offset/length tupples. With the
1392  * partial_page structure, we can map several non-contiguous ranges into
1393  * our ones pages[] map instead of splitting that operation into pieces.
1394  * Could easily be exported as a generic helper for other users, in which
1395  * case one would probably want to add a 'max_nr_pages' parameter as well.
1396  */
1397 static int get_iovec_page_array(const struct iovec __user *iov,
1398                                 unsigned int nr_vecs, struct page **pages,
1399                                 struct partial_page *partial, bool aligned,
1400                                 unsigned int pipe_buffers)
1401 {
1402         int buffers = 0, error = 0;
1403
1404         while (nr_vecs) {
1405                 unsigned long off, npages;
1406                 struct iovec entry;
1407                 void __user *base;
1408                 size_t len;
1409                 int i;
1410
1411                 error = -EFAULT;
1412                 if (copy_from_user(&entry, iov, sizeof(entry)))
1413                         break;
1414
1415                 base = entry.iov_base;
1416                 len = entry.iov_len;
1417
1418                 /*
1419                  * Sanity check this iovec. 0 read succeeds.
1420                  */
1421                 error = 0;
1422                 if (unlikely(!len))
1423                         break;
1424                 error = -EFAULT;
1425                 if (!access_ok(VERIFY_READ, base, len))
1426                         break;
1427
1428                 /*
1429                  * Get this base offset and number of pages, then map
1430                  * in the user pages.
1431                  */
1432                 off = (unsigned long) base & ~PAGE_MASK;
1433
1434                 /*
1435                  * If asked for alignment, the offset must be zero and the
1436                  * length a multiple of the PAGE_SIZE.
1437                  */
1438                 error = -EINVAL;
1439                 if (aligned && (off || len & ~PAGE_MASK))
1440                         break;
1441
1442                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1443                 if (npages > pipe_buffers - buffers)
1444                         npages = pipe_buffers - buffers;
1445
1446                 error = get_user_pages_fast((unsigned long)base, npages,
1447                                         0, &pages[buffers]);
1448
1449                 if (unlikely(error <= 0))
1450                         break;
1451
1452                 /*
1453                  * Fill this contiguous range into the partial page map.
1454                  */
1455                 for (i = 0; i < error; i++) {
1456                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1457
1458                         partial[buffers].offset = off;
1459                         partial[buffers].len = plen;
1460
1461                         off = 0;
1462                         len -= plen;
1463                         buffers++;
1464                 }
1465
1466                 /*
1467                  * We didn't complete this iov, stop here since it probably
1468                  * means we have to move some of this into a pipe to
1469                  * be able to continue.
1470                  */
1471                 if (len)
1472                         break;
1473
1474                 /*
1475                  * Don't continue if we mapped fewer pages than we asked for,
1476                  * or if we mapped the max number of pages that we have
1477                  * room for.
1478                  */
1479                 if (error < npages || buffers == pipe_buffers)
1480                         break;
1481
1482                 nr_vecs--;
1483                 iov++;
1484         }
1485
1486         if (buffers)
1487                 return buffers;
1488
1489         return error;
1490 }
1491
1492 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1493                         struct splice_desc *sd)
1494 {
1495         char *src;
1496         int ret;
1497
1498         /*
1499          * See if we can use the atomic maps, by prefaulting in the
1500          * pages and doing an atomic copy
1501          */
1502         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1503                 src = buf->ops->map(pipe, buf, 1);
1504                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1505                                                         sd->len);
1506                 buf->ops->unmap(pipe, buf, src);
1507                 if (!ret) {
1508                         ret = sd->len;
1509                         goto out;
1510                 }
1511         }
1512
1513         /*
1514          * No dice, use slow non-atomic map and copy
1515          */
1516         src = buf->ops->map(pipe, buf, 0);
1517
1518         ret = sd->len;
1519         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1520                 ret = -EFAULT;
1521
1522         buf->ops->unmap(pipe, buf, src);
1523 out:
1524         if (ret > 0)
1525                 sd->u.userptr += ret;
1526         return ret;
1527 }
1528
1529 /*
1530  * For lack of a better implementation, implement vmsplice() to userspace
1531  * as a simple copy of the pipes pages to the user iov.
1532  */
1533 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1534                              unsigned long nr_segs, unsigned int flags)
1535 {
1536         struct pipe_inode_info *pipe;
1537         struct splice_desc sd;
1538         ssize_t size;
1539         int error;
1540         long ret;
1541
1542         pipe = get_pipe_info(file);
1543         if (!pipe)
1544                 return -EBADF;
1545
1546         pipe_lock(pipe);
1547
1548         error = ret = 0;
1549         while (nr_segs) {
1550                 void __user *base;
1551                 size_t len;
1552
1553                 /*
1554                  * Get user address base and length for this iovec.
1555                  */
1556                 error = get_user(base, &iov->iov_base);
1557                 if (unlikely(error))
1558                         break;
1559                 error = get_user(len, &iov->iov_len);
1560                 if (unlikely(error))
1561                         break;
1562
1563                 /*
1564                  * Sanity check this iovec. 0 read succeeds.
1565                  */
1566                 if (unlikely(!len))
1567                         break;
1568                 if (unlikely(!base)) {
1569                         error = -EFAULT;
1570                         break;
1571                 }
1572
1573                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1574                         error = -EFAULT;
1575                         break;
1576                 }
1577
1578                 sd.len = 0;
1579                 sd.total_len = len;
1580                 sd.flags = flags;
1581                 sd.u.userptr = base;
1582                 sd.pos = 0;
1583
1584                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1585                 if (size < 0) {
1586                         if (!ret)
1587                                 ret = size;
1588
1589                         break;
1590                 }
1591
1592                 ret += size;
1593
1594                 if (size < len)
1595                         break;
1596
1597                 nr_segs--;
1598                 iov++;
1599         }
1600
1601         pipe_unlock(pipe);
1602
1603         if (!ret)
1604                 ret = error;
1605
1606         return ret;
1607 }
1608
1609 /*
1610  * vmsplice splices a user address range into a pipe. It can be thought of
1611  * as splice-from-memory, where the regular splice is splice-from-file (or
1612  * to file). In both cases the output is a pipe, naturally.
1613  */
1614 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1615                              unsigned long nr_segs, unsigned int flags)
1616 {
1617         struct pipe_inode_info *pipe;
1618         struct page *pages[PIPE_DEF_BUFFERS];
1619         struct partial_page partial[PIPE_DEF_BUFFERS];
1620         struct splice_pipe_desc spd = {
1621                 .pages = pages,
1622                 .partial = partial,
1623                 .nr_pages_max = PIPE_DEF_BUFFERS,
1624                 .flags = flags,
1625                 .ops = &user_page_pipe_buf_ops,
1626                 .spd_release = spd_release_page,
1627         };
1628         long ret;
1629
1630         pipe = get_pipe_info(file);
1631         if (!pipe)
1632                 return -EBADF;
1633
1634         if (splice_grow_spd(pipe, &spd))
1635                 return -ENOMEM;
1636
1637         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1638                                             spd.partial, false,
1639                                             spd.nr_pages_max);
1640         if (spd.nr_pages <= 0)
1641                 ret = spd.nr_pages;
1642         else
1643                 ret = splice_to_pipe(pipe, &spd);
1644
1645         splice_shrink_spd(&spd);
1646         return ret;
1647 }
1648
1649 /*
1650  * Note that vmsplice only really supports true splicing _from_ user memory
1651  * to a pipe, not the other way around. Splicing from user memory is a simple
1652  * operation that can be supported without any funky alignment restrictions
1653  * or nasty vm tricks. We simply map in the user memory and fill them into
1654  * a pipe. The reverse isn't quite as easy, though. There are two possible
1655  * solutions for that:
1656  *
1657  *      - memcpy() the data internally, at which point we might as well just
1658  *        do a regular read() on the buffer anyway.
1659  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1660  *        has restriction limitations on both ends of the pipe).
1661  *
1662  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1663  *
1664  */
1665 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1666                 unsigned long, nr_segs, unsigned int, flags)
1667 {
1668         struct fd f;
1669         long error;
1670
1671         if (unlikely(nr_segs > UIO_MAXIOV))
1672                 return -EINVAL;
1673         else if (unlikely(!nr_segs))
1674                 return 0;
1675
1676         error = -EBADF;
1677         f = fdget(fd);
1678         if (f.file) {
1679                 if (f.file->f_mode & FMODE_WRITE)
1680                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1681                 else if (f.file->f_mode & FMODE_READ)
1682                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1683
1684                 fdput(f);
1685         }
1686
1687         return error;
1688 }
1689
1690 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1691                 int, fd_out, loff_t __user *, off_out,
1692                 size_t, len, unsigned int, flags)
1693 {
1694         struct fd in, out;
1695         long error;
1696
1697         if (unlikely(!len))
1698                 return 0;
1699
1700         error = -EBADF;
1701         in = fdget(fd_in);
1702         if (in.file) {
1703                 if (in.file->f_mode & FMODE_READ) {
1704                         out = fdget(fd_out);
1705                         if (out.file) {
1706                                 if (out.file->f_mode & FMODE_WRITE)
1707                                         error = do_splice(in.file, off_in,
1708                                                           out.file, off_out,
1709                                                           len, flags);
1710                                 fdput(out);
1711                         }
1712                 }
1713                 fdput(in);
1714         }
1715         return error;
1716 }
1717
1718 /*
1719  * Make sure there's data to read. Wait for input if we can, otherwise
1720  * return an appropriate error.
1721  */
1722 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1723 {
1724         int ret;
1725
1726         /*
1727          * Check ->nrbufs without the inode lock first. This function
1728          * is speculative anyways, so missing one is ok.
1729          */
1730         if (pipe->nrbufs)
1731                 return 0;
1732
1733         ret = 0;
1734         pipe_lock(pipe);
1735
1736         while (!pipe->nrbufs) {
1737                 if (signal_pending(current)) {
1738                         ret = -ERESTARTSYS;
1739                         break;
1740                 }
1741                 if (!pipe->writers)
1742                         break;
1743                 if (!pipe->waiting_writers) {
1744                         if (flags & SPLICE_F_NONBLOCK) {
1745                                 ret = -EAGAIN;
1746                                 break;
1747                         }
1748                 }
1749                 pipe_wait(pipe);
1750         }
1751
1752         pipe_unlock(pipe);
1753         return ret;
1754 }
1755
1756 /*
1757  * Make sure there's writeable room. Wait for room if we can, otherwise
1758  * return an appropriate error.
1759  */
1760 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1761 {
1762         int ret;
1763
1764         /*
1765          * Check ->nrbufs without the inode lock first. This function
1766          * is speculative anyways, so missing one is ok.
1767          */
1768         if (pipe->nrbufs < pipe->buffers)
1769                 return 0;
1770
1771         ret = 0;
1772         pipe_lock(pipe);
1773
1774         while (pipe->nrbufs >= pipe->buffers) {
1775                 if (!pipe->readers) {
1776                         send_sig(SIGPIPE, current, 0);
1777                         ret = -EPIPE;
1778                         break;
1779                 }
1780                 if (flags & SPLICE_F_NONBLOCK) {
1781                         ret = -EAGAIN;
1782                         break;
1783                 }
1784                 if (signal_pending(current)) {
1785                         ret = -ERESTARTSYS;
1786                         break;
1787                 }
1788                 pipe->waiting_writers++;
1789                 pipe_wait(pipe);
1790                 pipe->waiting_writers--;
1791         }
1792
1793         pipe_unlock(pipe);
1794         return ret;
1795 }
1796
1797 /*
1798  * Splice contents of ipipe to opipe.
1799  */
1800 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1801                                struct pipe_inode_info *opipe,
1802                                size_t len, unsigned int flags)
1803 {
1804         struct pipe_buffer *ibuf, *obuf;
1805         int ret = 0, nbuf;
1806         bool input_wakeup = false;
1807
1808
1809 retry:
1810         ret = ipipe_prep(ipipe, flags);
1811         if (ret)
1812                 return ret;
1813
1814         ret = opipe_prep(opipe, flags);
1815         if (ret)
1816                 return ret;
1817
1818         /*
1819          * Potential ABBA deadlock, work around it by ordering lock
1820          * grabbing by pipe info address. Otherwise two different processes
1821          * could deadlock (one doing tee from A -> B, the other from B -> A).
1822          */
1823         pipe_double_lock(ipipe, opipe);
1824
1825         do {
1826                 if (!opipe->readers) {
1827                         send_sig(SIGPIPE, current, 0);
1828                         if (!ret)
1829                                 ret = -EPIPE;
1830                         break;
1831                 }
1832
1833                 if (!ipipe->nrbufs && !ipipe->writers)
1834                         break;
1835
1836                 /*
1837                  * Cannot make any progress, because either the input
1838                  * pipe is empty or the output pipe is full.
1839                  */
1840                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1841                         /* Already processed some buffers, break */
1842                         if (ret)
1843                                 break;
1844
1845                         if (flags & SPLICE_F_NONBLOCK) {
1846                                 ret = -EAGAIN;
1847                                 break;
1848                         }
1849
1850                         /*
1851                          * We raced with another reader/writer and haven't
1852                          * managed to process any buffers.  A zero return
1853                          * value means EOF, so retry instead.
1854                          */
1855                         pipe_unlock(ipipe);
1856                         pipe_unlock(opipe);
1857                         goto retry;
1858                 }
1859
1860                 ibuf = ipipe->bufs + ipipe->curbuf;
1861                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1862                 obuf = opipe->bufs + nbuf;
1863
1864                 if (len >= ibuf->len) {
1865                         /*
1866                          * Simply move the whole buffer from ipipe to opipe
1867                          */
1868                         *obuf = *ibuf;
1869                         ibuf->ops = NULL;
1870                         opipe->nrbufs++;
1871                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1872                         ipipe->nrbufs--;
1873                         input_wakeup = true;
1874                 } else {
1875                         /*
1876                          * Get a reference to this pipe buffer,
1877                          * so we can copy the contents over.
1878                          */
1879                         ibuf->ops->get(ipipe, ibuf);
1880                         *obuf = *ibuf;
1881
1882                         /*
1883                          * Don't inherit the gift flag, we need to
1884                          * prevent multiple steals of this page.
1885                          */
1886                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1887
1888                         obuf->len = len;
1889                         opipe->nrbufs++;
1890                         ibuf->offset += obuf->len;
1891                         ibuf->len -= obuf->len;
1892                 }
1893                 ret += obuf->len;
1894                 len -= obuf->len;
1895         } while (len);
1896
1897         pipe_unlock(ipipe);
1898         pipe_unlock(opipe);
1899
1900         /*
1901          * If we put data in the output pipe, wakeup any potential readers.
1902          */
1903         if (ret > 0)
1904                 wakeup_pipe_readers(opipe);
1905
1906         if (input_wakeup)
1907                 wakeup_pipe_writers(ipipe);
1908
1909         return ret;
1910 }
1911
1912 /*
1913  * Link contents of ipipe to opipe.
1914  */
1915 static int link_pipe(struct pipe_inode_info *ipipe,
1916                      struct pipe_inode_info *opipe,
1917                      size_t len, unsigned int flags)
1918 {
1919         struct pipe_buffer *ibuf, *obuf;
1920         int ret = 0, i = 0, nbuf;
1921
1922         /*
1923          * Potential ABBA deadlock, work around it by ordering lock
1924          * grabbing by pipe info address. Otherwise two different processes
1925          * could deadlock (one doing tee from A -> B, the other from B -> A).
1926          */
1927         pipe_double_lock(ipipe, opipe);
1928
1929         do {
1930                 if (!opipe->readers) {
1931                         send_sig(SIGPIPE, current, 0);
1932                         if (!ret)
1933                                 ret = -EPIPE;
1934                         break;
1935                 }
1936
1937                 /*
1938                  * If we have iterated all input buffers or ran out of
1939                  * output room, break.
1940                  */
1941                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1942                         break;
1943
1944                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1945                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1946
1947                 /*
1948                  * Get a reference to this pipe buffer,
1949                  * so we can copy the contents over.
1950                  */
1951                 ibuf->ops->get(ipipe, ibuf);
1952
1953                 obuf = opipe->bufs + nbuf;
1954                 *obuf = *ibuf;
1955
1956                 /*
1957                  * Don't inherit the gift flag, we need to
1958                  * prevent multiple steals of this page.
1959                  */
1960                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1961
1962                 if (obuf->len > len)
1963                         obuf->len = len;
1964
1965                 opipe->nrbufs++;
1966                 ret += obuf->len;
1967                 len -= obuf->len;
1968                 i++;
1969         } while (len);
1970
1971         /*
1972          * return EAGAIN if we have the potential of some data in the
1973          * future, otherwise just return 0
1974          */
1975         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1976                 ret = -EAGAIN;
1977
1978         pipe_unlock(ipipe);
1979         pipe_unlock(opipe);
1980
1981         /*
1982          * If we put data in the output pipe, wakeup any potential readers.
1983          */
1984         if (ret > 0)
1985                 wakeup_pipe_readers(opipe);
1986
1987         return ret;
1988 }
1989
1990 /*
1991  * This is a tee(1) implementation that works on pipes. It doesn't copy
1992  * any data, it simply references the 'in' pages on the 'out' pipe.
1993  * The 'flags' used are the SPLICE_F_* variants, currently the only
1994  * applicable one is SPLICE_F_NONBLOCK.
1995  */
1996 static long do_tee(struct file *in, struct file *out, size_t len,
1997                    unsigned int flags)
1998 {
1999         struct pipe_inode_info *ipipe = get_pipe_info(in);
2000         struct pipe_inode_info *opipe = get_pipe_info(out);
2001         int ret = -EINVAL;
2002
2003         /*
2004          * Duplicate the contents of ipipe to opipe without actually
2005          * copying the data.
2006          */
2007         if (ipipe && opipe && ipipe != opipe) {
2008                 /*
2009                  * Keep going, unless we encounter an error. The ipipe/opipe
2010                  * ordering doesn't really matter.
2011                  */
2012                 ret = ipipe_prep(ipipe, flags);
2013                 if (!ret) {
2014                         ret = opipe_prep(opipe, flags);
2015                         if (!ret)
2016                                 ret = link_pipe(ipipe, opipe, len, flags);
2017                 }
2018         }
2019
2020         return ret;
2021 }
2022
2023 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2024 {
2025         struct fd in;
2026         int error;
2027
2028         if (unlikely(!len))
2029                 return 0;
2030
2031         error = -EBADF;
2032         in = fdget(fdin);
2033         if (in.file) {
2034                 if (in.file->f_mode & FMODE_READ) {
2035                         struct fd out = fdget(fdout);
2036                         if (out.file) {
2037                                 if (out.file->f_mode & FMODE_WRITE)
2038                                         error = do_tee(in.file, out.file,
2039                                                         len, flags);
2040                                 fdput(out);
2041                         }
2042                 }
2043                 fdput(in);
2044         }
2045
2046         return error;
2047 }