fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/module.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33 #include <linux/gfp.h>
  34 #include <linux/socket.h>
  35
  36 /*
  37  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  38  * a vm helper function, it's already simplified quite a bit by the
  39  * addition of remove_mapping(). If success is returned, the caller may
  40  * attempt to reuse this page for another destination.
  41  */
  42 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  43                                      struct pipe_buffer *buf)
  44 {
  45         struct page *page = buf->page;
  46         struct address_space *mapping;
  47
  48         lock_page(page);
  49
  50         mapping = page_mapping(page);
  51         if (mapping) {
  52                 WARN_ON(!PageUptodate(page));
  53
  54                 /*
  55                  * At least for ext2 with nobh option, we need to wait on
  56                  * writeback completing on this page, since we'll remove it
  57                  * from the pagecache.  Otherwise truncate wont wait on the
  58                  * page, allowing the disk blocks to be reused by someone else
  59                  * before we actually wrote our data to them. fs corruption
  60                  * ensues.
  61                  */
  62                 wait_on_page_writeback(page);
  63
  64                 if (page_has_private(page) &&
  65                     !try_to_release_page(page, GFP_KERNEL))
  66                         goto out_unlock;
  67
  68                 /*
  69                  * If we succeeded in removing the mapping, set LRU flag
  70                  * and return good.
  71                  */
  72                 if (remove_mapping(mapping, page)) {
  73                         buf->flags |= PIPE_BUF_FLAG_LRU;
  74                         return 0;
  75                 }
  76         }
  77
  78         /*
  79          * Raced with truncate or failed to remove page from current
  80          * address space, unlock and return failure.
  81          */
  82 out_unlock:
  83         unlock_page(page);
  84         return 1;
  85 }
  86
  87 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  88                                         struct pipe_buffer *buf)
  89 {
  90         page_cache_release(buf->page);
  91         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  92 }
  93
  94 /*
  95  * Check whether the contents of buf is OK to access. Since the content
  96  * is a page cache page, IO may be in flight.
  97  */
  98 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  99                                        struct pipe_buffer *buf)
 100 {
 101         struct page *page = buf->page;
 102         int err;
 103
 104         if (!PageUptodate(page)) {
 105                 lock_page(page);
 106
 107                 /*
 108                  * Page got truncated/unhashed. This will cause a 0-byte
 109                  * splice, if this is the first page.
 110                  */
 111                 if (!page->mapping) {
 112                         err = -ENODATA;
 113                         goto error;
 114                 }
 115
 116                 /*
 117                  * Uh oh, read-error from disk.
 118                  */
 119                 if (!PageUptodate(page)) {
 120                         err = -EIO;
 121                         goto error;
 122                 }
 123
 124                 /*
 125                  * Page is ok afterall, we are done.
 126                  */
 127                 unlock_page(page);
 128         }
 129
 130         return 0;
 131 error:
 132         unlock_page(page);
 133         return err;
 134 }
 135
 136 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 137         .can_merge = 0,
 138         .map = generic_pipe_buf_map,
 139         .unmap = generic_pipe_buf_unmap,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 158         .map = generic_pipe_buf_map,
 159         .unmap = generic_pipe_buf_unmap,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 189         int ret, do_wakeup, page_nr;
 190
 191         ret = 0;
 192         do_wakeup = 0;
 193         page_nr = 0;
 194
 195         pipe_lock(pipe);
 196
 197         for (;;) {
 198                 if (!pipe->readers) {
 199                         send_sig(SIGPIPE, current, 0);
 200                         if (!ret)
 201                                 ret = -EPIPE;
 202                         break;
 203                 }
 204
 205                 if (pipe->nrbufs < pipe->buffers) {
 206                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 207                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 208
 209                         buf->page = spd->pages[page_nr];
 210                         buf->offset = spd->partial[page_nr].offset;
 211                         buf->len = spd->partial[page_nr].len;
 212                         buf->private = spd->partial[page_nr].private;
 213                         buf->ops = spd->ops;
 214                         if (spd->flags & SPLICE_F_GIFT)
 215                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 216
 217                         pipe->nrbufs++;
 218                         page_nr++;
 219                         ret += buf->len;
 220
 221                         if (pipe->inode)
 222                                 do_wakeup = 1;
 223
 224                         if (!--spd->nr_pages)
 225                                 break;
 226                         if (pipe->nrbufs < pipe->buffers)
 227                                 continue;
 228
 229                         break;
 230                 }
 231
 232                 if (spd->flags & SPLICE_F_NONBLOCK) {
 233                         if (!ret)
 234                                 ret = -EAGAIN;
 235                         break;
 236                 }
 237
 238                 if (signal_pending(current)) {
 239                         if (!ret)
 240                                 ret = -ERESTARTSYS;
 241                         break;
 242                 }
 243
 244                 if (do_wakeup) {
 245                         smp_mb();
 246                         if (waitqueue_active(&pipe->wait))
 247                                 wake_up_interruptible_sync(&pipe->wait);
 248                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 249                         do_wakeup = 0;
 250                 }
 251
 252                 pipe->waiting_writers++;
 253                 pipe_wait(pipe);
 254                 pipe->waiting_writers--;
 255         }
 256
 257         pipe_unlock(pipe);
 258
 259         if (do_wakeup)
 260                 wakeup_pipe_readers(pipe);
 261
 262         while (page_nr < spd_pages)
 263                 spd->spd_release(spd, page_nr++);
 264
 265         return ret;
 266 }
 267
 268 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 269 {
 270         page_cache_release(spd->pages[i]);
 271 }
 272
 273 /*
 274  * Check if we need to grow the arrays holding pages and partial page
 275  * descriptions.
 276  */
 277 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 278 {
 279         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 280                 return 0;
 281
 282         spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
 283         spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
 284
 285         if (spd->pages && spd->partial)
 286                 return 0;
 287
 288         kfree(spd->pages);
 289         kfree(spd->partial);
 290         return -ENOMEM;
 291 }
 292
 293 void splice_shrink_spd(struct pipe_inode_info *pipe,
 294                        struct splice_pipe_desc *spd)
 295 {
 296         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 297                 return;
 298
 299         kfree(spd->pages);
 300         kfree(spd->partial);
 301 }
 302
 303 static int
 304 __generic_file_splice_read(struct file *in, loff_t *ppos,
 305                            struct pipe_inode_info *pipe, size_t len,
 306                            unsigned int flags)
 307 {
 308         struct address_space *mapping = in->f_mapping;
 309         unsigned int loff, nr_pages, req_pages;
 310         struct page *pages[PIPE_DEF_BUFFERS];
 311         struct partial_page partial[PIPE_DEF_BUFFERS];
 312         struct page *page;
 313         pgoff_t index, end_index;
 314         loff_t isize;
 315         int error, page_nr;
 316         struct splice_pipe_desc spd = {
 317                 .pages = pages,
 318                 .partial = partial,
 319                 .flags = flags,
 320                 .ops = &page_cache_pipe_buf_ops,
 321                 .spd_release = spd_release_page,
 322         };
 323
 324         if (splice_grow_spd(pipe, &spd))
 325                 return -ENOMEM;
 326
 327         index = *ppos >> PAGE_CACHE_SHIFT;
 328         loff = *ppos & ~PAGE_CACHE_MASK;
 329         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 330         nr_pages = min(req_pages, pipe->buffers);
 331
 332         /*
 333          * Lookup the (hopefully) full range of pages we need.
 334          */
 335         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 336         index += spd.nr_pages;
 337
 338         /*
 339          * If find_get_pages_contig() returned fewer pages than we needed,
 340          * readahead/allocate the rest and fill in the holes.
 341          */
 342         if (spd.nr_pages < nr_pages)
 343                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 344                                 index, req_pages - spd.nr_pages);
 345
 346         error = 0;
 347         while (spd.nr_pages < nr_pages) {
 348                 /*
 349                  * Page could be there, find_get_pages_contig() breaks on
 350                  * the first hole.
 351                  */
 352                 page = find_get_page(mapping, index);
 353                 if (!page) {
 354                         /*
 355                          * page didn't exist, allocate one.
 356                          */
 357                         page = page_cache_alloc_cold(mapping);
 358                         if (!page)
 359                                 break;
 360
 361                         error = add_to_page_cache_lru(page, mapping, index,
 362                                                 GFP_KERNEL);
 363                         if (unlikely(error)) {
 364                                 page_cache_release(page);
 365                                 if (error == -EEXIST)
 366                                         continue;
 367                                 break;
 368                         }
 369                         /*
 370                          * add_to_page_cache() locks the page, unlock it
 371                          * to avoid convoluting the logic below even more.
 372                          */
 373                         unlock_page(page);
 374                 }
 375
 376                 spd.pages[spd.nr_pages++] = page;
 377                 index++;
 378         }
 379
 380         /*
 381          * Now loop over the map and see if we need to start IO on any
 382          * pages, fill in the partial map, etc.
 383          */
 384         index = *ppos >> PAGE_CACHE_SHIFT;
 385         nr_pages = spd.nr_pages;
 386         spd.nr_pages = 0;
 387         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 388                 unsigned int this_len;
 389
 390                 if (!len)
 391                         break;
 392
 393                 /*
 394                  * this_len is the max we'll use from this page
 395                  */
 396                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 397                 page = spd.pages[page_nr];
 398
 399                 if (PageReadahead(page))
 400                         page_cache_async_readahead(mapping, &in->f_ra, in,
 401                                         page, index, req_pages - page_nr);
 402
 403                 /*
 404                  * If the page isn't uptodate, we may need to start io on it
 405                  */
 406                 if (!PageUptodate(page)) {
 407                         lock_page(page);
 408
 409                         /*
 410                          * Page was truncated, or invalidated by the
 411                          * filesystem.  Redo the find/create, but this time the
 412                          * page is kept locked, so there's no chance of another
 413                          * race with truncate/invalidate.
 414                          */
 415                         if (!page->mapping) {
 416                                 unlock_page(page);
 417                                 page = find_or_create_page(mapping, index,
 418                                                 mapping_gfp_mask(mapping));
 419
 420                                 if (!page) {
 421                                         error = -ENOMEM;
 422                                         break;
 423                                 }
 424                                 page_cache_release(spd.pages[page_nr]);
 425                                 spd.pages[page_nr] = page;
 426                         }
 427                         /*
 428                          * page was already under io and is now done, great
 429                          */
 430                         if (PageUptodate(page)) {
 431                                 unlock_page(page);
 432                                 goto fill_it;
 433                         }
 434
 435                         /*
 436                          * need to read in the page
 437                          */
 438                         error = mapping->a_ops->readpage(in, page);
 439                         if (unlikely(error)) {
 440                                 /*
 441                                  * We really should re-lookup the page here,
 442                                  * but it complicates things a lot. Instead
 443                                  * lets just do what we already stored, and
 444                                  * we'll get it the next time we are called.
 445                                  */
 446                                 if (error == AOP_TRUNCATED_PAGE)
 447                                         error = 0;
 448
 449                                 break;
 450                         }
 451                 }
 452 fill_it:
 453                 /*
 454                  * i_size must be checked after PageUptodate.
 455                  */
 456                 isize = i_size_read(mapping->host);
 457                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 458                 if (unlikely(!isize || index > end_index))
 459                         break;
 460
 461                 /*
 462                  * if this is the last page, see if we need to shrink
 463                  * the length and stop
 464                  */
 465                 if (end_index == index) {
 466                         unsigned int plen;
 467
 468                         /*
 469                          * max good bytes in this page
 470                          */
 471                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 472                         if (plen <= loff)
 473                                 break;
 474
 475                         /*
 476                          * force quit after adding this page
 477                          */
 478                         this_len = min(this_len, plen - loff);
 479                         len = this_len;
 480                 }
 481
 482                 spd.partial[page_nr].offset = loff;
 483                 spd.partial[page_nr].len = this_len;
 484                 len -= this_len;
 485                 loff = 0;
 486                 spd.nr_pages++;
 487                 index++;
 488         }
 489
 490         /*
 491          * Release any pages at the end, if we quit early. 'page_nr' is how far
 492          * we got, 'nr_pages' is how many pages are in the map.
 493          */
 494         while (page_nr < nr_pages)
 495                 page_cache_release(spd.pages[page_nr++]);
 496         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 497
 498         if (spd.nr_pages)
 499                 error = splice_to_pipe(pipe, &spd);
 500
 501         splice_shrink_spd(pipe, &spd);
 502         return error;
 503 }
 504
 505 /**
 506  * generic_file_splice_read - splice data from file to a pipe
 507  * @in:         file to splice from
 508  * @ppos:       position in @in
 509  * @pipe:       pipe to splice to
 510  * @len:        number of bytes to splice
 511  * @flags:      splice modifier flags
 512  *
 513  * Description:
 514  *    Will read pages from given file and fill them into a pipe. Can be
 515  *    used as long as the address_space operations for the source implements
 516  *    a readpage() hook.
 517  *
 518  */
 519 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 520                                  struct pipe_inode_info *pipe, size_t len,
 521                                  unsigned int flags)
 522 {
 523         loff_t isize, left;
 524         int ret;
 525
 526         isize = i_size_read(in->f_mapping->host);
 527         if (unlikely(*ppos >= isize))
 528                 return 0;
 529
 530         left = isize - *ppos;
 531         if (unlikely(left < len))
 532                 len = left;
 533
 534         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 535         if (ret > 0) {
 536                 *ppos += ret;
 537                 file_accessed(in);
 538         }
 539
 540         return ret;
 541 }
 542 EXPORT_SYMBOL(generic_file_splice_read);
 543
 544 static const struct pipe_buf_operations default_pipe_buf_ops = {
 545         .can_merge = 0,
 546         .map = generic_pipe_buf_map,
 547         .unmap = generic_pipe_buf_unmap,
 548         .confirm = generic_pipe_buf_confirm,
 549         .release = generic_pipe_buf_release,
 550         .steal = generic_pipe_buf_steal,
 551         .get = generic_pipe_buf_get,
 552 };
 553
 554 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 555                             unsigned long vlen, loff_t offset)
 556 {
 557         mm_segment_t old_fs;
 558         loff_t pos = offset;
 559         ssize_t res;
 560
 561         old_fs = get_fs();
 562         set_fs(get_ds());
 563         /* The cast to a user pointer is valid due to the set_fs() */
 564         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 565         set_fs(old_fs);
 566
 567         return res;
 568 }
 569
 570 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 571                             loff_t pos)
 572 {
 573         mm_segment_t old_fs;
 574         ssize_t res;
 575
 576         old_fs = get_fs();
 577         set_fs(get_ds());
 578         /* The cast to a user pointer is valid due to the set_fs() */
 579         res = vfs_write(file, (const char __user *)buf, count, &pos);
 580         set_fs(old_fs);
 581
 582         return res;
 583 }
 584
 585 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 586                                  struct pipe_inode_info *pipe, size_t len,
 587                                  unsigned int flags)
 588 {
 589         unsigned int nr_pages;
 590         unsigned int nr_freed;
 591         size_t offset;
 592         struct page *pages[PIPE_DEF_BUFFERS];
 593         struct partial_page partial[PIPE_DEF_BUFFERS];
 594         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 595         ssize_t res;
 596         size_t this_len;
 597         int error;
 598         int i;
 599         struct splice_pipe_desc spd = {
 600                 .pages = pages,
 601                 .partial = partial,
 602                 .flags = flags,
 603                 .ops = &default_pipe_buf_ops,
 604                 .spd_release = spd_release_page,
 605         };
 606
 607         if (splice_grow_spd(pipe, &spd))
 608                 return -ENOMEM;
 609
 610         res = -ENOMEM;
 611         vec = __vec;
 612         if (pipe->buffers > PIPE_DEF_BUFFERS) {
 613                 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
 614                 if (!vec)
 615                         goto shrink_ret;
 616         }
 617
 618         offset = *ppos & ~PAGE_CACHE_MASK;
 619         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 620
 621         for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
 622                 struct page *page;
 623
 624                 page = alloc_page(GFP_USER);
 625                 error = -ENOMEM;
 626                 if (!page)
 627                         goto err;
 628
 629                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 630                 vec[i].iov_base = (void __user *) page_address(page);
 631                 vec[i].iov_len = this_len;
 632                 spd.pages[i] = page;
 633                 spd.nr_pages++;
 634                 len -= this_len;
 635                 offset = 0;
 636         }
 637
 638         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 639         if (res < 0) {
 640                 error = res;
 641                 goto err;
 642         }
 643
 644         error = 0;
 645         if (!res)
 646                 goto err;
 647
 648         nr_freed = 0;
 649         for (i = 0; i < spd.nr_pages; i++) {
 650                 this_len = min_t(size_t, vec[i].iov_len, res);
 651                 spd.partial[i].offset = 0;
 652                 spd.partial[i].len = this_len;
 653                 if (!this_len) {
 654                         __free_page(spd.pages[i]);
 655                         spd.pages[i] = NULL;
 656                         nr_freed++;
 657                 }
 658                 res -= this_len;
 659         }
 660         spd.nr_pages -= nr_freed;
 661
 662         res = splice_to_pipe(pipe, &spd);
 663         if (res > 0)
 664                 *ppos += res;
 665
 666 shrink_ret:
 667         if (vec != __vec)
 668                 kfree(vec);
 669         splice_shrink_spd(pipe, &spd);
 670         return res;
 671
 672 err:
 673         for (i = 0; i < spd.nr_pages; i++)
 674                 __free_page(spd.pages[i]);
 675
 676         res = error;
 677         goto shrink_ret;
 678 }
 679 EXPORT_SYMBOL(default_file_splice_read);
 680
 681 /*
 682  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 683  * using sendpage(). Return the number of bytes sent.
 684  */
 685 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 686                             struct pipe_buffer *buf, struct splice_desc *sd)
 687 {
 688         struct file *file = sd->u.file;
 689         loff_t pos = sd->pos;
 690         int more;
 691
 692         if (!likely(file->f_op && file->f_op->sendpage))
 693                 return -EINVAL;
 694
 695         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 696         if (sd->len < sd->total_len)
 697                 more |= MSG_SENDPAGE_NOTLAST;
 698         return file->f_op->sendpage(file, buf->page, buf->offset,
 699                                     sd->len, &pos, more);
 700 }
 701
 702 /*
 703  * This is a little more tricky than the file -> pipe splicing. There are
 704  * basically three cases:
 705  *
 706  *      - Destination page already exists in the address space and there
 707  *        are users of it. For that case we have no other option that
 708  *        copying the data. Tough luck.
 709  *      - Destination page already exists in the address space, but there
 710  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 711  *        through to last case.
 712  *      - Destination page does not exist, we can add the pipe page to
 713  *        the page cache and avoid the copy.
 714  *
 715  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 716  * sd->flags), we attempt to migrate pages from the pipe to the output
 717  * file address space page cache. This is possible if no one else has
 718  * the pipe page referenced outside of the pipe and page cache. If
 719  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 720  * a new page in the output file page cache and fill/dirty that.
 721  */
 722 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 723                  struct splice_desc *sd)
 724 {
 725         struct file *file = sd->u.file;
 726         struct address_space *mapping = file->f_mapping;
 727         unsigned int offset, this_len;
 728         struct page *page;
 729         void *fsdata;
 730         int ret;
 731
 732         offset = sd->pos & ~PAGE_CACHE_MASK;
 733
 734         this_len = sd->len;
 735         if (this_len + offset > PAGE_CACHE_SIZE)
 736                 this_len = PAGE_CACHE_SIZE - offset;
 737
 738         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 739                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 740         if (unlikely(ret))
 741                 goto out;
 742
 743         if (buf->page != page) {
 744                 /*
 745                  * Careful, ->map() uses KM_USER0!
 746                  */
 747                 char *src = buf->ops->map(pipe, buf, 1);
 748                 char *dst = kmap_atomic(page, KM_USER1);
 749
 750                 memcpy(dst + offset, src + buf->offset, this_len);
 751                 flush_dcache_page(page);
 752                 kunmap_atomic(dst, KM_USER1);
 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                         file_update_time(out);
1011                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1012                 }
1013                 mutex_unlock(&inode->i_mutex);
1014         } while (ret > 0);
1015         splice_from_pipe_end(pipe, &sd);
1016
1017         pipe_unlock(pipe);
1018
1019         if (sd.num_spliced)
1020                 ret = sd.num_spliced;
1021
1022         if (ret > 0) {
1023                 unsigned long nr_pages;
1024                 int err;
1025
1026                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1027
1028                 err = generic_write_sync(out, *ppos, ret);
1029                 if (err)
1030                         ret = err;
1031                 else
1032                         *ppos += ret;
1033                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1034         }
1035
1036         return ret;
1037 }
1038
1039 EXPORT_SYMBOL(generic_file_splice_write);
1040
1041 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1042                           struct splice_desc *sd)
1043 {
1044         int ret;
1045         void *data;
1046
1047         data = buf->ops->map(pipe, buf, 0);
1048         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1049         buf->ops->unmap(pipe, buf, data);
1050
1051         return ret;
1052 }
1053
1054 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1055                                          struct file *out, loff_t *ppos,
1056                                          size_t len, unsigned int flags)
1057 {
1058         ssize_t ret;
1059
1060         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1061         if (ret > 0)
1062                 *ppos += ret;
1063
1064         return ret;
1065 }
1066
1067 /**
1068  * generic_splice_sendpage - splice data from a pipe to a socket
1069  * @pipe:       pipe to splice from
1070  * @out:        socket to write to
1071  * @ppos:       position in @out
1072  * @len:        number of bytes to splice
1073  * @flags:      splice modifier flags
1074  *
1075  * Description:
1076  *    Will send @len bytes from the pipe to a network socket. No data copying
1077  *    is involved.
1078  *
1079  */
1080 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1081                                 loff_t *ppos, size_t len, unsigned int flags)
1082 {
1083         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1084 }
1085
1086 EXPORT_SYMBOL(generic_splice_sendpage);
1087
1088 /*
1089  * Attempt to initiate a splice from pipe to file.
1090  */
1091 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1092                            loff_t *ppos, size_t len, unsigned int flags)
1093 {
1094         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1095                                 loff_t *, size_t, unsigned int);
1096         int ret;
1097
1098         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1099                 return -EBADF;
1100
1101         if (unlikely(out->f_flags & O_APPEND))
1102                 return -EINVAL;
1103
1104         ret = rw_verify_area(WRITE, out, ppos, len);
1105         if (unlikely(ret < 0))
1106                 return ret;
1107
1108         if (out->f_op && out->f_op->splice_write)
1109                 splice_write = out->f_op->splice_write;
1110         else
1111                 splice_write = default_file_splice_write;
1112
1113         return splice_write(pipe, out, ppos, len, flags);
1114 }
1115
1116 /*
1117  * Attempt to initiate a splice from a file to a pipe.
1118  */
1119 static long do_splice_to(struct file *in, loff_t *ppos,
1120                          struct pipe_inode_info *pipe, size_t len,
1121                          unsigned int flags)
1122 {
1123         ssize_t (*splice_read)(struct file *, loff_t *,
1124                                struct pipe_inode_info *, size_t, unsigned int);
1125         int ret;
1126
1127         if (unlikely(!(in->f_mode & FMODE_READ)))
1128                 return -EBADF;
1129
1130         ret = rw_verify_area(READ, in, ppos, len);
1131         if (unlikely(ret < 0))
1132                 return ret;
1133
1134         if (in->f_op && in->f_op->splice_read)
1135                 splice_read = in->f_op->splice_read;
1136         else
1137                 splice_read = default_file_splice_read;
1138
1139         return splice_read(in, ppos, pipe, len, flags);
1140 }
1141
1142 /**
1143  * splice_direct_to_actor - splices data directly between two non-pipes
1144  * @in:         file to splice from
1145  * @sd:         actor information on where to splice to
1146  * @actor:      handles the data splicing
1147  *
1148  * Description:
1149  *    This is a special case helper to splice directly between two
1150  *    points, without requiring an explicit pipe. Internally an allocated
1151  *    pipe is cached in the process, and reused during the lifetime of
1152  *    that process.
1153  *
1154  */
1155 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1156                                splice_direct_actor *actor)
1157 {
1158         struct pipe_inode_info *pipe;
1159         long ret, bytes;
1160         umode_t i_mode;
1161         size_t len;
1162         int i, flags;
1163
1164         /*
1165          * We require the input being a regular file, as we don't want to
1166          * randomly drop data for eg socket -> socket splicing. Use the
1167          * piped splicing for that!
1168          */
1169         i_mode = in->f_path.dentry->d_inode->i_mode;
1170         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1171                 return -EINVAL;
1172
1173         /*
1174          * neither in nor out is a pipe, setup an internal pipe attached to
1175          * 'out' and transfer the wanted data from 'in' to 'out' through that
1176          */
1177         pipe = current->splice_pipe;
1178         if (unlikely(!pipe)) {
1179                 pipe = alloc_pipe_info(NULL);
1180                 if (!pipe)
1181                         return -ENOMEM;
1182
1183                 /*
1184                  * We don't have an immediate reader, but we'll read the stuff
1185                  * out of the pipe right after the splice_to_pipe(). So set
1186                  * PIPE_READERS appropriately.
1187                  */
1188                 pipe->readers = 1;
1189
1190                 current->splice_pipe = pipe;
1191         }
1192
1193         /*
1194          * Do the splice.
1195          */
1196         ret = 0;
1197         bytes = 0;
1198         len = sd->total_len;
1199         flags = sd->flags;
1200
1201         /*
1202          * Don't block on output, we have to drain the direct pipe.
1203          */
1204         sd->flags &= ~SPLICE_F_NONBLOCK;
1205
1206         while (len) {
1207                 size_t read_len;
1208                 loff_t pos = sd->pos, prev_pos = pos;
1209
1210                 ret = do_splice_to(in, &pos, pipe, len, flags);
1211                 if (unlikely(ret <= 0))
1212                         goto out_release;
1213
1214                 read_len = ret;
1215                 sd->total_len = read_len;
1216
1217                 /*
1218                  * NOTE: nonblocking mode only applies to the input. We
1219                  * must not do the output in nonblocking mode as then we
1220                  * could get stuck data in the internal pipe:
1221                  */
1222                 ret = actor(pipe, sd);
1223                 if (unlikely(ret <= 0)) {
1224                         sd->pos = prev_pos;
1225                         goto out_release;
1226                 }
1227
1228                 bytes += ret;
1229                 len -= ret;
1230                 sd->pos = pos;
1231
1232                 if (ret < read_len) {
1233                         sd->pos = prev_pos + ret;
1234                         goto out_release;
1235                 }
1236         }
1237
1238 done:
1239         pipe->nrbufs = pipe->curbuf = 0;
1240         file_accessed(in);
1241         return bytes;
1242
1243 out_release:
1244         /*
1245          * If we did an incomplete transfer we must release
1246          * the pipe buffers in question:
1247          */
1248         for (i = 0; i < pipe->buffers; i++) {
1249                 struct pipe_buffer *buf = pipe->bufs + i;
1250
1251                 if (buf->ops) {
1252                         buf->ops->release(pipe, buf);
1253                         buf->ops = NULL;
1254                 }
1255         }
1256
1257         if (!bytes)
1258                 bytes = ret;
1259
1260         goto done;
1261 }
1262 EXPORT_SYMBOL(splice_direct_to_actor);
1263
1264 static int direct_splice_actor(struct pipe_inode_info *pipe,
1265                                struct splice_desc *sd)
1266 {
1267         struct file *file = sd->u.file;
1268
1269         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1270                               sd->flags);
1271 }
1272
1273 /**
1274  * do_splice_direct - splices data directly between two files
1275  * @in:         file to splice from
1276  * @ppos:       input file offset
1277  * @out:        file to splice to
1278  * @len:        number of bytes to splice
1279  * @flags:      splice modifier flags
1280  *
1281  * Description:
1282  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1283  *    doing it in the application would incur an extra system call
1284  *    (splice in + splice out, as compared to just sendfile()). So this helper
1285  *    can splice directly through a process-private pipe.
1286  *
1287  */
1288 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1289                       size_t len, unsigned int flags)
1290 {
1291         struct splice_desc sd = {
1292                 .len            = len,
1293                 .total_len      = len,
1294                 .flags          = flags,
1295                 .pos            = *ppos,
1296                 .u.file         = out,
1297         };
1298         long ret;
1299
1300         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1301         if (ret > 0)
1302                 *ppos = sd.pos;
1303
1304         return ret;
1305 }
1306
1307 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1308                                struct pipe_inode_info *opipe,
1309                                size_t len, unsigned int flags);
1310
1311 /*
1312  * Determine where to splice to/from.
1313  */
1314 static long do_splice(struct file *in, loff_t __user *off_in,
1315                       struct file *out, loff_t __user *off_out,
1316                       size_t len, unsigned int flags)
1317 {
1318         struct pipe_inode_info *ipipe;
1319         struct pipe_inode_info *opipe;
1320         loff_t offset, *off;
1321         long ret;
1322
1323         ipipe = get_pipe_info(in);
1324         opipe = get_pipe_info(out);
1325
1326         if (ipipe && opipe) {
1327                 if (off_in || off_out)
1328                         return -ESPIPE;
1329
1330                 if (!(in->f_mode & FMODE_READ))
1331                         return -EBADF;
1332
1333                 if (!(out->f_mode & FMODE_WRITE))
1334                         return -EBADF;
1335
1336                 /* Splicing to self would be fun, but... */
1337                 if (ipipe == opipe)
1338                         return -EINVAL;
1339
1340                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1341         }
1342
1343         if (ipipe) {
1344                 if (off_in)
1345                         return -ESPIPE;
1346                 if (off_out) {
1347                         if (!(out->f_mode & FMODE_PWRITE))
1348                                 return -EINVAL;
1349                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1350                                 return -EFAULT;
1351                         off = &offset;
1352                 } else
1353                         off = &out->f_pos;
1354
1355                 ret = do_splice_from(ipipe, out, off, len, flags);
1356
1357                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1358                         ret = -EFAULT;
1359
1360                 return ret;
1361         }
1362
1363         if (opipe) {
1364                 if (off_out)
1365                         return -ESPIPE;
1366                 if (off_in) {
1367                         if (!(in->f_mode & FMODE_PREAD))
1368                                 return -EINVAL;
1369                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1370                                 return -EFAULT;
1371                         off = &offset;
1372                 } else
1373                         off = &in->f_pos;
1374
1375                 ret = do_splice_to(in, off, opipe, len, flags);
1376
1377                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1378                         ret = -EFAULT;
1379
1380                 return ret;
1381         }
1382
1383         return -EINVAL;
1384 }
1385
1386 /*
1387  * Map an iov into an array of pages and offset/length tupples. With the
1388  * partial_page structure, we can map several non-contiguous ranges into
1389  * our ones pages[] map instead of splitting that operation into pieces.
1390  * Could easily be exported as a generic helper for other users, in which
1391  * case one would probably want to add a 'max_nr_pages' parameter as well.
1392  */
1393 static int get_iovec_page_array(const struct iovec __user *iov,
1394                                 unsigned int nr_vecs, struct page **pages,
1395                                 struct partial_page *partial, int aligned,
1396                                 unsigned int pipe_buffers)
1397 {
1398         int buffers = 0, error = 0;
1399
1400         while (nr_vecs) {
1401                 unsigned long off, npages;
1402                 struct iovec entry;
1403                 void __user *base;
1404                 size_t len;
1405                 int i;
1406
1407                 error = -EFAULT;
1408                 if (copy_from_user(&entry, iov, sizeof(entry)))
1409                         break;
1410
1411                 base = entry.iov_base;
1412                 len = entry.iov_len;
1413
1414                 /*
1415                  * Sanity check this iovec. 0 read succeeds.
1416                  */
1417                 error = 0;
1418                 if (unlikely(!len))
1419                         break;
1420                 error = -EFAULT;
1421                 if (!access_ok(VERIFY_READ, base, len))
1422                         break;
1423
1424                 /*
1425                  * Get this base offset and number of pages, then map
1426                  * in the user pages.
1427                  */
1428                 off = (unsigned long) base & ~PAGE_MASK;
1429
1430                 /*
1431                  * If asked for alignment, the offset must be zero and the
1432                  * length a multiple of the PAGE_SIZE.
1433                  */
1434                 error = -EINVAL;
1435                 if (aligned && (off || len & ~PAGE_MASK))
1436                         break;
1437
1438                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1439                 if (npages > pipe_buffers - buffers)
1440                         npages = pipe_buffers - buffers;
1441
1442                 error = get_user_pages_fast((unsigned long)base, npages,
1443                                         0, &pages[buffers]);
1444
1445                 if (unlikely(error <= 0))
1446                         break;
1447
1448                 /*
1449                  * Fill this contiguous range into the partial page map.
1450                  */
1451                 for (i = 0; i < error; i++) {
1452                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1453
1454                         partial[buffers].offset = off;
1455                         partial[buffers].len = plen;
1456
1457                         off = 0;
1458                         len -= plen;
1459                         buffers++;
1460                 }
1461
1462                 /*
1463                  * We didn't complete this iov, stop here since it probably
1464                  * means we have to move some of this into a pipe to
1465                  * be able to continue.
1466                  */
1467                 if (len)
1468                         break;
1469
1470                 /*
1471                  * Don't continue if we mapped fewer pages than we asked for,
1472                  * or if we mapped the max number of pages that we have
1473                  * room for.
1474                  */
1475                 if (error < npages || buffers == pipe_buffers)
1476                         break;
1477
1478                 nr_vecs--;
1479                 iov++;
1480         }
1481
1482         if (buffers)
1483                 return buffers;
1484
1485         return error;
1486 }
1487
1488 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1489                         struct splice_desc *sd)
1490 {
1491         char *src;
1492         int ret;
1493
1494         /*
1495          * See if we can use the atomic maps, by prefaulting in the
1496          * pages and doing an atomic copy
1497          */
1498         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1499                 src = buf->ops->map(pipe, buf, 1);
1500                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1501                                                         sd->len);
1502                 buf->ops->unmap(pipe, buf, src);
1503                 if (!ret) {
1504                         ret = sd->len;
1505                         goto out;
1506                 }
1507         }
1508
1509         /*
1510          * No dice, use slow non-atomic map and copy
1511          */
1512         src = buf->ops->map(pipe, buf, 0);
1513
1514         ret = sd->len;
1515         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1516                 ret = -EFAULT;
1517
1518         buf->ops->unmap(pipe, buf, src);
1519 out:
1520         if (ret > 0)
1521                 sd->u.userptr += ret;
1522         return ret;
1523 }
1524
1525 /*
1526  * For lack of a better implementation, implement vmsplice() to userspace
1527  * as a simple copy of the pipes pages to the user iov.
1528  */
1529 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1530                              unsigned long nr_segs, unsigned int flags)
1531 {
1532         struct pipe_inode_info *pipe;
1533         struct splice_desc sd;
1534         ssize_t size;
1535         int error;
1536         long ret;
1537
1538         pipe = get_pipe_info(file);
1539         if (!pipe)
1540                 return -EBADF;
1541
1542         pipe_lock(pipe);
1543
1544         error = ret = 0;
1545         while (nr_segs) {
1546                 void __user *base;
1547                 size_t len;
1548
1549                 /*
1550                  * Get user address base and length for this iovec.
1551                  */
1552                 error = get_user(base, &iov->iov_base);
1553                 if (unlikely(error))
1554                         break;
1555                 error = get_user(len, &iov->iov_len);
1556                 if (unlikely(error))
1557                         break;
1558
1559                 /*
1560                  * Sanity check this iovec. 0 read succeeds.
1561                  */
1562                 if (unlikely(!len))
1563                         break;
1564                 if (unlikely(!base)) {
1565                         error = -EFAULT;
1566                         break;
1567                 }
1568
1569                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1570                         error = -EFAULT;
1571                         break;
1572                 }
1573
1574                 sd.len = 0;
1575                 sd.total_len = len;
1576                 sd.flags = flags;
1577                 sd.u.userptr = base;
1578                 sd.pos = 0;
1579
1580                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1581                 if (size < 0) {
1582                         if (!ret)
1583                                 ret = size;
1584
1585                         break;
1586                 }
1587
1588                 ret += size;
1589
1590                 if (size < len)
1591                         break;
1592
1593                 nr_segs--;
1594                 iov++;
1595         }
1596
1597         pipe_unlock(pipe);
1598
1599         if (!ret)
1600                 ret = error;
1601
1602         return ret;
1603 }
1604
1605 /*
1606  * vmsplice splices a user address range into a pipe. It can be thought of
1607  * as splice-from-memory, where the regular splice is splice-from-file (or
1608  * to file). In both cases the output is a pipe, naturally.
1609  */
1610 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1611                              unsigned long nr_segs, unsigned int flags)
1612 {
1613         struct pipe_inode_info *pipe;
1614         struct page *pages[PIPE_DEF_BUFFERS];
1615         struct partial_page partial[PIPE_DEF_BUFFERS];
1616         struct splice_pipe_desc spd = {
1617                 .pages = pages,
1618                 .partial = partial,
1619                 .flags = flags,
1620                 .ops = &user_page_pipe_buf_ops,
1621                 .spd_release = spd_release_page,
1622         };
1623         long ret;
1624
1625         pipe = get_pipe_info(file);
1626         if (!pipe)
1627                 return -EBADF;
1628
1629         if (splice_grow_spd(pipe, &spd))
1630                 return -ENOMEM;
1631
1632         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1633                                             spd.partial, flags & SPLICE_F_GIFT,
1634                                             pipe->buffers);
1635         if (spd.nr_pages <= 0)
1636                 ret = spd.nr_pages;
1637         else
1638                 ret = splice_to_pipe(pipe, &spd);
1639
1640         splice_shrink_spd(pipe, &spd);
1641         return ret;
1642 }
1643
1644 /*
1645  * Note that vmsplice only really supports true splicing _from_ user memory
1646  * to a pipe, not the other way around. Splicing from user memory is a simple
1647  * operation that can be supported without any funky alignment restrictions
1648  * or nasty vm tricks. We simply map in the user memory and fill them into
1649  * a pipe. The reverse isn't quite as easy, though. There are two possible
1650  * solutions for that:
1651  *
1652  *      - memcpy() the data internally, at which point we might as well just
1653  *        do a regular read() on the buffer anyway.
1654  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1655  *        has restriction limitations on both ends of the pipe).
1656  *
1657  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1658  *
1659  */
1660 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1661                 unsigned long, nr_segs, unsigned int, flags)
1662 {
1663         struct file *file;
1664         long error;
1665         int fput;
1666
1667         if (unlikely(nr_segs > UIO_MAXIOV))
1668                 return -EINVAL;
1669         else if (unlikely(!nr_segs))
1670                 return 0;
1671
1672         error = -EBADF;
1673         file = fget_light(fd, &fput);
1674         if (file) {
1675                 if (file->f_mode & FMODE_WRITE)
1676                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1677                 else if (file->f_mode & FMODE_READ)
1678                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1679
1680                 fput_light(file, fput);
1681         }
1682
1683         return error;
1684 }
1685
1686 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1687                 int, fd_out, loff_t __user *, off_out,
1688                 size_t, len, unsigned int, flags)
1689 {
1690         long error;
1691         struct file *in, *out;
1692         int fput_in, fput_out;
1693
1694         if (unlikely(!len))
1695                 return 0;
1696
1697         error = -EBADF;
1698         in = fget_light(fd_in, &fput_in);
1699         if (in) {
1700                 if (in->f_mode & FMODE_READ) {
1701                         out = fget_light(fd_out, &fput_out);
1702                         if (out) {
1703                                 if (out->f_mode & FMODE_WRITE)
1704                                         error = do_splice(in, off_in,
1705                                                           out, off_out,
1706                                                           len, flags);
1707                                 fput_light(out, fput_out);
1708                         }
1709                 }
1710
1711                 fput_light(in, fput_in);
1712         }
1713
1714         return error;
1715 }
1716
1717 /*
1718  * Make sure there's data to read. Wait for input if we can, otherwise
1719  * return an appropriate error.
1720  */
1721 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1722 {
1723         int ret;
1724
1725         /*
1726          * Check ->nrbufs without the inode lock first. This function
1727          * is speculative anyways, so missing one is ok.
1728          */
1729         if (pipe->nrbufs)
1730                 return 0;
1731
1732         ret = 0;
1733         pipe_lock(pipe);
1734
1735         while (!pipe->nrbufs) {
1736                 if (signal_pending(current)) {
1737                         ret = -ERESTARTSYS;
1738                         break;
1739                 }
1740                 if (!pipe->writers)
1741                         break;
1742                 if (!pipe->waiting_writers) {
1743                         if (flags & SPLICE_F_NONBLOCK) {
1744                                 ret = -EAGAIN;
1745                                 break;
1746                         }
1747                 }
1748                 pipe_wait(pipe);
1749         }
1750
1751         pipe_unlock(pipe);
1752         return ret;
1753 }
1754
1755 /*
1756  * Make sure there's writeable room. Wait for room if we can, otherwise
1757  * return an appropriate error.
1758  */
1759 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1760 {
1761         int ret;
1762
1763         /*
1764          * Check ->nrbufs without the inode lock first. This function
1765          * is speculative anyways, so missing one is ok.
1766          */
1767         if (pipe->nrbufs < pipe->buffers)
1768                 return 0;
1769
1770         ret = 0;
1771         pipe_lock(pipe);
1772
1773         while (pipe->nrbufs >= pipe->buffers) {
1774                 if (!pipe->readers) {
1775                         send_sig(SIGPIPE, current, 0);
1776                         ret = -EPIPE;
1777                         break;
1778                 }
1779                 if (flags & SPLICE_F_NONBLOCK) {
1780                         ret = -EAGAIN;
1781                         break;
1782                 }
1783                 if (signal_pending(current)) {
1784                         ret = -ERESTARTSYS;
1785                         break;
1786                 }
1787                 pipe->waiting_writers++;
1788                 pipe_wait(pipe);
1789                 pipe->waiting_writers--;
1790         }
1791
1792         pipe_unlock(pipe);
1793         return ret;
1794 }
1795
1796 /*
1797  * Splice contents of ipipe to opipe.
1798  */
1799 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1800                                struct pipe_inode_info *opipe,
1801                                size_t len, unsigned int flags)
1802 {
1803         struct pipe_buffer *ibuf, *obuf;
1804         int ret = 0, nbuf;
1805         bool input_wakeup = false;
1806
1807
1808 retry:
1809         ret = ipipe_prep(ipipe, flags);
1810         if (ret)
1811                 return ret;
1812
1813         ret = opipe_prep(opipe, flags);
1814         if (ret)
1815                 return ret;
1816
1817         /*
1818          * Potential ABBA deadlock, work around it by ordering lock
1819          * grabbing by pipe info address. Otherwise two different processes
1820          * could deadlock (one doing tee from A -> B, the other from B -> A).
1821          */
1822         pipe_double_lock(ipipe, opipe);
1823
1824         do {
1825                 if (!opipe->readers) {
1826                         send_sig(SIGPIPE, current, 0);
1827                         if (!ret)
1828                                 ret = -EPIPE;
1829                         break;
1830                 }
1831
1832                 if (!ipipe->nrbufs && !ipipe->writers)
1833                         break;
1834
1835                 /*
1836                  * Cannot make any progress, because either the input
1837                  * pipe is empty or the output pipe is full.
1838                  */
1839                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1840                         /* Already processed some buffers, break */
1841                         if (ret)
1842                                 break;
1843
1844                         if (flags & SPLICE_F_NONBLOCK) {
1845                                 ret = -EAGAIN;
1846                                 break;
1847                         }
1848
1849                         /*
1850                          * We raced with another reader/writer and haven't
1851                          * managed to process any buffers.  A zero return
1852                          * value means EOF, so retry instead.
1853                          */
1854                         pipe_unlock(ipipe);
1855                         pipe_unlock(opipe);
1856                         goto retry;
1857                 }
1858
1859                 ibuf = ipipe->bufs + ipipe->curbuf;
1860                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1861                 obuf = opipe->bufs + nbuf;
1862
1863                 if (len >= ibuf->len) {
1864                         /*
1865                          * Simply move the whole buffer from ipipe to opipe
1866                          */
1867                         *obuf = *ibuf;
1868                         ibuf->ops = NULL;
1869                         opipe->nrbufs++;
1870                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1871                         ipipe->nrbufs--;
1872                         input_wakeup = true;
1873                 } else {
1874                         /*
1875                          * Get a reference to this pipe buffer,
1876                          * so we can copy the contents over.
1877                          */
1878                         ibuf->ops->get(ipipe, ibuf);
1879                         *obuf = *ibuf;
1880
1881                         /*
1882                          * Don't inherit the gift flag, we need to
1883                          * prevent multiple steals of this page.
1884                          */
1885                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1886
1887                         obuf->len = len;
1888                         opipe->nrbufs++;
1889                         ibuf->offset += obuf->len;
1890                         ibuf->len -= obuf->len;
1891                 }
1892                 ret += obuf->len;
1893                 len -= obuf->len;
1894         } while (len);
1895
1896         pipe_unlock(ipipe);
1897         pipe_unlock(opipe);
1898
1899         /*
1900          * If we put data in the output pipe, wakeup any potential readers.
1901          */
1902         if (ret > 0)
1903                 wakeup_pipe_readers(opipe);
1904
1905         if (input_wakeup)
1906                 wakeup_pipe_writers(ipipe);
1907
1908         return ret;
1909 }
1910
1911 /*
1912  * Link contents of ipipe to opipe.
1913  */
1914 static int link_pipe(struct pipe_inode_info *ipipe,
1915                      struct pipe_inode_info *opipe,
1916                      size_t len, unsigned int flags)
1917 {
1918         struct pipe_buffer *ibuf, *obuf;
1919         int ret = 0, i = 0, nbuf;
1920
1921         /*
1922          * Potential ABBA deadlock, work around it by ordering lock
1923          * grabbing by pipe info address. Otherwise two different processes
1924          * could deadlock (one doing tee from A -> B, the other from B -> A).
1925          */
1926         pipe_double_lock(ipipe, opipe);
1927
1928         do {
1929                 if (!opipe->readers) {
1930                         send_sig(SIGPIPE, current, 0);
1931                         if (!ret)
1932                                 ret = -EPIPE;
1933                         break;
1934                 }
1935
1936                 /*
1937                  * If we have iterated all input buffers or ran out of
1938                  * output room, break.
1939                  */
1940                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1941                         break;
1942
1943                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1944                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1945
1946                 /*
1947                  * Get a reference to this pipe buffer,
1948                  * so we can copy the contents over.
1949                  */
1950                 ibuf->ops->get(ipipe, ibuf);
1951
1952                 obuf = opipe->bufs + nbuf;
1953                 *obuf = *ibuf;
1954
1955                 /*
1956                  * Don't inherit the gift flag, we need to
1957                  * prevent multiple steals of this page.
1958                  */
1959                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1960
1961                 if (obuf->len > len)
1962                         obuf->len = len;
1963
1964                 opipe->nrbufs++;
1965                 ret += obuf->len;
1966                 len -= obuf->len;
1967                 i++;
1968         } while (len);
1969
1970         /*
1971          * return EAGAIN if we have the potential of some data in the
1972          * future, otherwise just return 0
1973          */
1974         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1975                 ret = -EAGAIN;
1976
1977         pipe_unlock(ipipe);
1978         pipe_unlock(opipe);
1979
1980         /*
1981          * If we put data in the output pipe, wakeup any potential readers.
1982          */
1983         if (ret > 0)
1984                 wakeup_pipe_readers(opipe);
1985
1986         return ret;
1987 }
1988
1989 /*
1990  * This is a tee(1) implementation that works on pipes. It doesn't copy
1991  * any data, it simply references the 'in' pages on the 'out' pipe.
1992  * The 'flags' used are the SPLICE_F_* variants, currently the only
1993  * applicable one is SPLICE_F_NONBLOCK.
1994  */
1995 static long do_tee(struct file *in, struct file *out, size_t len,
1996                    unsigned int flags)
1997 {
1998         struct pipe_inode_info *ipipe = get_pipe_info(in);
1999         struct pipe_inode_info *opipe = get_pipe_info(out);
2000         int ret = -EINVAL;
2001
2002         /*
2003          * Duplicate the contents of ipipe to opipe without actually
2004          * copying the data.
2005          */
2006         if (ipipe && opipe && ipipe != opipe) {
2007                 /*
2008                  * Keep going, unless we encounter an error. The ipipe/opipe
2009                  * ordering doesn't really matter.
2010                  */
2011                 ret = ipipe_prep(ipipe, flags);
2012                 if (!ret) {
2013                         ret = opipe_prep(opipe, flags);
2014                         if (!ret)
2015                                 ret = link_pipe(ipipe, opipe, len, flags);
2016                 }
2017         }
2018
2019         return ret;
2020 }
2021
2022 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2023 {
2024         struct file *in;
2025         int error, fput_in;
2026
2027         if (unlikely(!len))
2028                 return 0;
2029
2030         error = -EBADF;
2031         in = fget_light(fdin, &fput_in);
2032         if (in) {
2033                 if (in->f_mode & FMODE_READ) {
2034                         int fput_out;
2035                         struct file *out = fget_light(fdout, &fput_out);
2036
2037                         if (out) {
2038                                 if (out->f_mode & FMODE_WRITE)
2039                                         error = do_tee(in, out, len, flags);
2040                                 fput_light(out, fput_out);
2041                         }
2042                 }
2043                 fput_light(in, fput_in);
2044         }
2045
2046         return error;
2047 }