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
  34 /*
  35  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  36  * a vm helper function, it's already simplified quite a bit by the
  37  * addition of remove_mapping(). If success is returned, the caller may
  38  * attempt to reuse this page for another destination.
  39  */
  40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  41                                      struct pipe_buffer *buf)
  42 {
  43         struct page *page = buf->page;
  44         struct address_space *mapping;
  45
  46         lock_page(page);
  47
  48         mapping = page_mapping(page);
  49         if (mapping) {
  50                 WARN_ON(!PageUptodate(page));
  51
  52                 /*
  53                  * At least for ext2 with nobh option, we need to wait on
  54                  * writeback completing on this page, since we'll remove it
  55                  * from the pagecache.  Otherwise truncate wont wait on the
  56                  * page, allowing the disk blocks to be reused by someone else
  57                  * before we actually wrote our data to them. fs corruption
  58                  * ensues.
  59                  */
  60                 wait_on_page_writeback(page);
  61
  62                 if (page_has_private(page) &&
  63                     !try_to_release_page(page, GFP_KERNEL))
  64                         goto out_unlock;
  65
  66                 /*
  67                  * If we succeeded in removing the mapping, set LRU flag
  68                  * and return good.
  69                  */
  70                 if (remove_mapping(mapping, page)) {
  71                         buf->flags |= PIPE_BUF_FLAG_LRU;
  72                         return 0;
  73                 }
  74         }
  75
  76         /*
  77          * Raced with truncate or failed to remove page from current
  78          * address space, unlock and return failure.
  79          */
  80 out_unlock:
  81         unlock_page(page);
  82         return 1;
  83 }
  84
  85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  86                                         struct pipe_buffer *buf)
  87 {
  88         page_cache_release(buf->page);
  89         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  90 }
  91
  92 /*
  93  * Check whether the contents of buf is OK to access. Since the content
  94  * is a page cache page, IO may be in flight.
  95  */
  96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  97                                        struct pipe_buffer *buf)
  98 {
  99         struct page *page = buf->page;
 100         int err;
 101
 102         if (!PageUptodate(page)) {
 103                 lock_page(page);
 104
 105                 /*
 106                  * Page got truncated/unhashed. This will cause a 0-byte
 107                  * splice, if this is the first page.
 108                  */
 109                 if (!page->mapping) {
 110                         err = -ENODATA;
 111                         goto error;
 112                 }
 113
 114                 /*
 115                  * Uh oh, read-error from disk.
 116                  */
 117                 if (!PageUptodate(page)) {
 118                         err = -EIO;
 119                         goto error;
 120                 }
 121
 122                 /*
 123                  * Page is ok afterall, we are done.
 124                  */
 125                 unlock_page(page);
 126         }
 127
 128         return 0;
 129 error:
 130         unlock_page(page);
 131         return err;
 132 }
 133
 134 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 135         .can_merge = 0,
 136         .map = generic_pipe_buf_map,
 137         .unmap = generic_pipe_buf_unmap,
 138         .confirm = page_cache_pipe_buf_confirm,
 139         .release = page_cache_pipe_buf_release,
 140         .steal = page_cache_pipe_buf_steal,
 141         .get = generic_pipe_buf_get,
 142 };
 143
 144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 145                                     struct pipe_buffer *buf)
 146 {
 147         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 148                 return 1;
 149
 150         buf->flags |= PIPE_BUF_FLAG_LRU;
 151         return generic_pipe_buf_steal(pipe, buf);
 152 }
 153
 154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 155         .can_merge = 0,
 156         .map = generic_pipe_buf_map,
 157         .unmap = generic_pipe_buf_unmap,
 158         .confirm = generic_pipe_buf_confirm,
 159         .release = page_cache_pipe_buf_release,
 160         .steal = user_page_pipe_buf_steal,
 161         .get = generic_pipe_buf_get,
 162 };
 163
 164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 165 {
 166         smp_mb();
 167         if (waitqueue_active(&pipe->wait))
 168                 wake_up_interruptible(&pipe->wait);
 169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 170 }
 171
 172 /**
 173  * splice_to_pipe - fill passed data into a pipe
 174  * @pipe:       pipe to fill
 175  * @spd:        data to fill
 176  *
 177  * Description:
 178  *    @spd contains a map of pages and len/offset tuples, along with
 179  *    the struct pipe_buf_operations associated with these pages. This
 180  *    function will link that data to the pipe.
 181  *
 182  */
 183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 184                        struct splice_pipe_desc *spd)
 185 {
 186         unsigned int spd_pages = spd->nr_pages;
 187         int ret, do_wakeup, page_nr;
 188
 189         ret = 0;
 190         do_wakeup = 0;
 191         page_nr = 0;
 192
 193         pipe_lock(pipe);
 194
 195         for (;;) {
 196                 if (!pipe->readers) {
 197                         send_sig(SIGPIPE, current, 0);
 198                         if (!ret)
 199                                 ret = -EPIPE;
 200                         break;
 201                 }
 202
 203                 if (pipe->nrbufs < pipe->buffers) {
 204                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 205                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 206
 207                         buf->page = spd->pages[page_nr];
 208                         buf->offset = spd->partial[page_nr].offset;
 209                         buf->len = spd->partial[page_nr].len;
 210                         buf->private = spd->partial[page_nr].private;
 211                         buf->ops = spd->ops;
 212                         if (spd->flags & SPLICE_F_GIFT)
 213                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 214
 215                         pipe->nrbufs++;
 216                         page_nr++;
 217                         ret += buf->len;
 218
 219                         if (pipe->inode)
 220                                 do_wakeup = 1;
 221
 222                         if (!--spd->nr_pages)
 223                                 break;
 224                         if (pipe->nrbufs < pipe->buffers)
 225                                 continue;
 226
 227                         break;
 228                 }
 229
 230                 if (spd->flags & SPLICE_F_NONBLOCK) {
 231                         if (!ret)
 232                                 ret = -EAGAIN;
 233                         break;
 234                 }
 235
 236                 if (signal_pending(current)) {
 237                         if (!ret)
 238                                 ret = -ERESTARTSYS;
 239                         break;
 240                 }
 241
 242                 if (do_wakeup) {
 243                         smp_mb();
 244                         if (waitqueue_active(&pipe->wait))
 245                                 wake_up_interruptible_sync(&pipe->wait);
 246                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 247                         do_wakeup = 0;
 248                 }
 249
 250                 pipe->waiting_writers++;
 251                 pipe_wait(pipe);
 252                 pipe->waiting_writers--;
 253         }
 254
 255         pipe_unlock(pipe);
 256
 257         if (do_wakeup)
 258                 wakeup_pipe_readers(pipe);
 259
 260         while (page_nr < spd_pages)
 261                 spd->spd_release(spd, page_nr++);
 262
 263         return ret;
 264 }
 265
 266 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 267 {
 268         page_cache_release(spd->pages[i]);
 269 }
 270
 271 /*
 272  * Check if we need to grow the arrays holding pages and partial page
 273  * descriptions.
 274  */
 275 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 276 {
 277         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 278                 return 0;
 279
 280         spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
 281         spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
 282
 283         if (spd->pages && spd->partial)
 284                 return 0;
 285
 286         kfree(spd->pages);
 287         kfree(spd->partial);
 288         return -ENOMEM;
 289 }
 290
 291 void splice_shrink_spd(struct pipe_inode_info *pipe,
 292                        struct splice_pipe_desc *spd)
 293 {
 294         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 295                 return;
 296
 297         kfree(spd->pages);
 298         kfree(spd->partial);
 299 }
 300
 301 static int
 302 __generic_file_splice_read(struct file *in, loff_t *ppos,
 303                            struct pipe_inode_info *pipe, size_t len,
 304                            unsigned int flags)
 305 {
 306         struct address_space *mapping = in->f_mapping;
 307         unsigned int loff, nr_pages, req_pages;
 308         struct page *pages[PIPE_DEF_BUFFERS];
 309         struct partial_page partial[PIPE_DEF_BUFFERS];
 310         struct page *page;
 311         pgoff_t index, end_index;
 312         loff_t isize;
 313         int error, page_nr;
 314         struct splice_pipe_desc spd = {
 315                 .pages = pages,
 316                 .partial = partial,
 317                 .flags = flags,
 318                 .ops = &page_cache_pipe_buf_ops,
 319                 .spd_release = spd_release_page,
 320         };
 321
 322         if (splice_grow_spd(pipe, &spd))
 323                 return -ENOMEM;
 324
 325         index = *ppos >> PAGE_CACHE_SHIFT;
 326         loff = *ppos & ~PAGE_CACHE_MASK;
 327         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 328         nr_pages = min(req_pages, pipe->buffers);
 329
 330         /*
 331          * Lookup the (hopefully) full range of pages we need.
 332          */
 333         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 334         index += spd.nr_pages;
 335
 336         /*
 337          * If find_get_pages_contig() returned fewer pages than we needed,
 338          * readahead/allocate the rest and fill in the holes.
 339          */
 340         if (spd.nr_pages < nr_pages)
 341                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 342                                 index, req_pages - spd.nr_pages);
 343
 344         error = 0;
 345         while (spd.nr_pages < nr_pages) {
 346                 /*
 347                  * Page could be there, find_get_pages_contig() breaks on
 348                  * the first hole.
 349                  */
 350                 page = find_get_page(mapping, index);
 351                 if (!page) {
 352                         /*
 353                          * page didn't exist, allocate one.
 354                          */
 355                         page = page_cache_alloc_cold(mapping);
 356                         if (!page)
 357                                 break;
 358
 359                         error = add_to_page_cache_lru(page, mapping, index,
 360                                                 GFP_KERNEL);
 361                         if (unlikely(error)) {
 362                                 page_cache_release(page);
 363                                 if (error == -EEXIST)
 364                                         continue;
 365                                 break;
 366                         }
 367                         /*
 368                          * add_to_page_cache() locks the page, unlock it
 369                          * to avoid convoluting the logic below even more.
 370                          */
 371                         unlock_page(page);
 372                 }
 373
 374                 spd.pages[spd.nr_pages++] = page;
 375                 index++;
 376         }
 377
 378         /*
 379          * Now loop over the map and see if we need to start IO on any
 380          * pages, fill in the partial map, etc.
 381          */
 382         index = *ppos >> PAGE_CACHE_SHIFT;
 383         nr_pages = spd.nr_pages;
 384         spd.nr_pages = 0;
 385         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 386                 unsigned int this_len;
 387
 388                 if (!len)
 389                         break;
 390
 391                 /*
 392                  * this_len is the max we'll use from this page
 393                  */
 394                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 395                 page = spd.pages[page_nr];
 396
 397                 if (PageReadahead(page))
 398                         page_cache_async_readahead(mapping, &in->f_ra, in,
 399                                         page, index, req_pages - page_nr);
 400
 401                 /*
 402                  * If the page isn't uptodate, we may need to start io on it
 403                  */
 404                 if (!PageUptodate(page)) {
 405                         lock_page(page);
 406
 407                         /*
 408                          * Page was truncated, or invalidated by the
 409                          * filesystem.  Redo the find/create, but this time the
 410                          * page is kept locked, so there's no chance of another
 411                          * race with truncate/invalidate.
 412                          */
 413                         if (!page->mapping) {
 414                                 unlock_page(page);
 415                                 page = find_or_create_page(mapping, index,
 416                                                 mapping_gfp_mask(mapping));
 417
 418                                 if (!page) {
 419                                         error = -ENOMEM;
 420                                         break;
 421                                 }
 422                                 page_cache_release(spd.pages[page_nr]);
 423                                 spd.pages[page_nr] = page;
 424                         }
 425                         /*
 426                          * page was already under io and is now done, great
 427                          */
 428                         if (PageUptodate(page)) {
 429                                 unlock_page(page);
 430                                 goto fill_it;
 431                         }
 432
 433                         /*
 434                          * need to read in the page
 435                          */
 436                         error = mapping->a_ops->readpage(in, page);
 437                         if (unlikely(error)) {
 438                                 /*
 439                                  * We really should re-lookup the page here,
 440                                  * but it complicates things a lot. Instead
 441                                  * lets just do what we already stored, and
 442                                  * we'll get it the next time we are called.
 443                                  */
 444                                 if (error == AOP_TRUNCATED_PAGE)
 445                                         error = 0;
 446
 447                                 break;
 448                         }
 449                 }
 450 fill_it:
 451                 /*
 452                  * i_size must be checked after PageUptodate.
 453                  */
 454                 isize = i_size_read(mapping->host);
 455                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 456                 if (unlikely(!isize || index > end_index))
 457                         break;
 458
 459                 /*
 460                  * if this is the last page, see if we need to shrink
 461                  * the length and stop
 462                  */
 463                 if (end_index == index) {
 464                         unsigned int plen;
 465
 466                         /*
 467                          * max good bytes in this page
 468                          */
 469                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 470                         if (plen <= loff)
 471                                 break;
 472
 473                         /*
 474                          * force quit after adding this page
 475                          */
 476                         this_len = min(this_len, plen - loff);
 477                         len = this_len;
 478                 }
 479
 480                 spd.partial[page_nr].offset = loff;
 481                 spd.partial[page_nr].len = this_len;
 482                 len -= this_len;
 483                 loff = 0;
 484                 spd.nr_pages++;
 485                 index++;
 486         }
 487
 488         /*
 489          * Release any pages at the end, if we quit early. 'page_nr' is how far
 490          * we got, 'nr_pages' is how many pages are in the map.
 491          */
 492         while (page_nr < nr_pages)
 493                 page_cache_release(spd.pages[page_nr++]);
 494         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 495
 496         if (spd.nr_pages)
 497                 error = splice_to_pipe(pipe, &spd);
 498
 499         splice_shrink_spd(pipe, &spd);
 500         return error;
 501 }
 502
 503 /**
 504  * generic_file_splice_read - splice data from file to a pipe
 505  * @in:         file to splice from
 506  * @ppos:       position in @in
 507  * @pipe:       pipe to splice to
 508  * @len:        number of bytes to splice
 509  * @flags:      splice modifier flags
 510  *
 511  * Description:
 512  *    Will read pages from given file and fill them into a pipe. Can be
 513  *    used as long as the address_space operations for the source implements
 514  *    a readpage() hook.
 515  *
 516  */
 517 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 518                                  struct pipe_inode_info *pipe, size_t len,
 519                                  unsigned int flags)
 520 {
 521         loff_t isize, left;
 522         int ret;
 523
 524         isize = i_size_read(in->f_mapping->host);
 525         if (unlikely(*ppos >= isize))
 526                 return 0;
 527
 528         left = isize - *ppos;
 529         if (unlikely(left < len))
 530                 len = left;
 531
 532         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 533         if (ret > 0) {
 534                 *ppos += ret;
 535                 file_accessed(in);
 536         }
 537
 538         return ret;
 539 }
 540 EXPORT_SYMBOL(generic_file_splice_read);
 541
 542 static const struct pipe_buf_operations default_pipe_buf_ops = {
 543         .can_merge = 0,
 544         .map = generic_pipe_buf_map,
 545         .unmap = generic_pipe_buf_unmap,
 546         .confirm = generic_pipe_buf_confirm,
 547         .release = generic_pipe_buf_release,
 548         .steal = generic_pipe_buf_steal,
 549         .get = generic_pipe_buf_get,
 550 };
 551
 552 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 553                             unsigned long vlen, loff_t offset)
 554 {
 555         mm_segment_t old_fs;
 556         loff_t pos = offset;
 557         ssize_t res;
 558
 559         old_fs = get_fs();
 560         set_fs(get_ds());
 561         /* The cast to a user pointer is valid due to the set_fs() */
 562         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 563         set_fs(old_fs);
 564
 565         return res;
 566 }
 567
 568 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 569                             loff_t pos)
 570 {
 571         mm_segment_t old_fs;
 572         ssize_t res;
 573
 574         old_fs = get_fs();
 575         set_fs(get_ds());
 576         /* The cast to a user pointer is valid due to the set_fs() */
 577         res = vfs_write(file, (const char __user *)buf, count, &pos);
 578         set_fs(old_fs);
 579
 580         return res;
 581 }
 582
 583 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 584                                  struct pipe_inode_info *pipe, size_t len,
 585                                  unsigned int flags)
 586 {
 587         unsigned int nr_pages;
 588         unsigned int nr_freed;
 589         size_t offset;
 590         struct page *pages[PIPE_DEF_BUFFERS];
 591         struct partial_page partial[PIPE_DEF_BUFFERS];
 592         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 593         ssize_t res;
 594         size_t this_len;
 595         int error;
 596         int i;
 597         struct splice_pipe_desc spd = {
 598                 .pages = pages,
 599                 .partial = partial,
 600                 .flags = flags,
 601                 .ops = &default_pipe_buf_ops,
 602                 .spd_release = spd_release_page,
 603         };
 604
 605         if (splice_grow_spd(pipe, &spd))
 606                 return -ENOMEM;
 607
 608         res = -ENOMEM;
 609         vec = __vec;
 610         if (pipe->buffers > PIPE_DEF_BUFFERS) {
 611                 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
 612                 if (!vec)
 613                         goto shrink_ret;
 614         }
 615
 616         offset = *ppos & ~PAGE_CACHE_MASK;
 617         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 618
 619         for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
 620                 struct page *page;
 621
 622                 page = alloc_page(GFP_USER);
 623                 error = -ENOMEM;
 624                 if (!page)
 625                         goto err;
 626
 627                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 628                 vec[i].iov_base = (void __user *) page_address(page);
 629                 vec[i].iov_len = this_len;
 630                 spd.pages[i] = page;
 631                 spd.nr_pages++;
 632                 len -= this_len;
 633                 offset = 0;
 634         }
 635
 636         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 637         if (res < 0) {
 638                 error = res;
 639                 goto err;
 640         }
 641
 642         error = 0;
 643         if (!res)
 644                 goto err;
 645
 646         nr_freed = 0;
 647         for (i = 0; i < spd.nr_pages; i++) {
 648                 this_len = min_t(size_t, vec[i].iov_len, res);
 649                 spd.partial[i].offset = 0;
 650                 spd.partial[i].len = this_len;
 651                 if (!this_len) {
 652                         __free_page(spd.pages[i]);
 653                         spd.pages[i] = NULL;
 654                         nr_freed++;
 655                 }
 656                 res -= this_len;
 657         }
 658         spd.nr_pages -= nr_freed;
 659
 660         res = splice_to_pipe(pipe, &spd);
 661         if (res > 0)
 662                 *ppos += res;
 663
 664 shrink_ret:
 665         if (vec != __vec)
 666                 kfree(vec);
 667         splice_shrink_spd(pipe, &spd);
 668         return res;
 669
 670 err:
 671         for (i = 0; i < spd.nr_pages; i++)
 672                 __free_page(spd.pages[i]);
 673
 674         res = error;
 675         goto shrink_ret;
 676 }
 677 EXPORT_SYMBOL(default_file_splice_read);
 678
 679 /*
 680  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 681  * using sendpage(). Return the number of bytes sent.
 682  */
 683 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 684                             struct pipe_buffer *buf, struct splice_desc *sd)
 685 {
 686         struct file *file = sd->u.file;
 687         loff_t pos = sd->pos;
 688         int more;
 689
 690         if (!likely(file->f_op && file->f_op->sendpage))
 691                 return -EINVAL;
 692
 693         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 694         return file->f_op->sendpage(file, buf->page, buf->offset,
 695                                     sd->len, &pos, more);
 696 }
 697
 698 /*
 699  * This is a little more tricky than the file -> pipe splicing. There are
 700  * basically three cases:
 701  *
 702  *      - Destination page already exists in the address space and there
 703  *        are users of it. For that case we have no other option that
 704  *        copying the data. Tough luck.
 705  *      - Destination page already exists in the address space, but there
 706  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 707  *        through to last case.
 708  *      - Destination page does not exist, we can add the pipe page to
 709  *        the page cache and avoid the copy.
 710  *
 711  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 712  * sd->flags), we attempt to migrate pages from the pipe to the output
 713  * file address space page cache. This is possible if no one else has
 714  * the pipe page referenced outside of the pipe and page cache. If
 715  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 716  * a new page in the output file page cache and fill/dirty that.
 717  */
 718 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 719                  struct splice_desc *sd)
 720 {
 721         struct file *file = sd->u.file;
 722         struct address_space *mapping = file->f_mapping;
 723         unsigned int offset, this_len;
 724         struct page *page;
 725         void *fsdata;
 726         int ret;
 727
 728         offset = sd->pos & ~PAGE_CACHE_MASK;
 729
 730         this_len = sd->len;
 731         if (this_len + offset > PAGE_CACHE_SIZE)
 732                 this_len = PAGE_CACHE_SIZE - offset;
 733
 734         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 735                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 736         if (unlikely(ret))
 737                 goto out;
 738
 739         if (buf->page != page) {
 740                 char *src = buf->ops->map(pipe, buf, 1);
 741                 char *dst = kmap_atomic(page);
 742
 743                 memcpy(dst + offset, src + buf->offset, this_len);
 744                 flush_dcache_page(page);
 745                 kunmap_atomic(dst);
 746                 buf->ops->unmap(pipe, buf, src);
 747         }
 748         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 749                                 page, fsdata);
 750 out:
 751         return ret;
 752 }
 753 EXPORT_SYMBOL(pipe_to_file);
 754
 755 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 756 {
 757         smp_mb();
 758         if (waitqueue_active(&pipe->wait))
 759                 wake_up_interruptible(&pipe->wait);
 760         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 761 }
 762
 763 /**
 764  * splice_from_pipe_feed - feed available data from a pipe to a file
 765  * @pipe:       pipe to splice from
 766  * @sd:         information to @actor
 767  * @actor:      handler that splices the data
 768  *
 769  * Description:
 770  *    This function loops over the pipe and calls @actor to do the
 771  *    actual moving of a single struct pipe_buffer to the desired
 772  *    destination.  It returns when there's no more buffers left in
 773  *    the pipe or if the requested number of bytes (@sd->total_len)
 774  *    have been copied.  It returns a positive number (one) if the
 775  *    pipe needs to be filled with more data, zero if the required
 776  *    number of bytes have been copied and -errno on error.
 777  *
 778  *    This, together with splice_from_pipe_{begin,end,next}, may be
 779  *    used to implement the functionality of __splice_from_pipe() when
 780  *    locking is required around copying the pipe buffers to the
 781  *    destination.
 782  */
 783 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 784                           splice_actor *actor)
 785 {
 786         int ret;
 787
 788         while (pipe->nrbufs) {
 789                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 790                 const struct pipe_buf_operations *ops = buf->ops;
 791
 792                 sd->len = buf->len;
 793                 if (sd->len > sd->total_len)
 794                         sd->len = sd->total_len;
 795
 796                 ret = buf->ops->confirm(pipe, buf);
 797                 if (unlikely(ret)) {
 798                         if (ret == -ENODATA)
 799                                 ret = 0;
 800                         return ret;
 801                 }
 802
 803                 ret = actor(pipe, buf, sd);
 804                 if (ret <= 0)
 805                         return ret;
 806
 807                 buf->offset += ret;
 808                 buf->len -= ret;
 809
 810                 sd->num_spliced += ret;
 811                 sd->len -= ret;
 812                 sd->pos += ret;
 813                 sd->total_len -= ret;
 814
 815                 if (!buf->len) {
 816                         buf->ops = NULL;
 817                         ops->release(pipe, buf);
 818                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 819                         pipe->nrbufs--;
 820                         if (pipe->inode)
 821                                 sd->need_wakeup = true;
 822                 }
 823
 824                 if (!sd->total_len)
 825                         return 0;
 826         }
 827
 828         return 1;
 829 }
 830 EXPORT_SYMBOL(splice_from_pipe_feed);
 831
 832 /**
 833  * splice_from_pipe_next - wait for some data to splice from
 834  * @pipe:       pipe to splice from
 835  * @sd:         information about the splice operation
 836  *
 837  * Description:
 838  *    This function will wait for some data and return a positive
 839  *    value (one) if pipe buffers are available.  It will return zero
 840  *    or -errno if no more data needs to be spliced.
 841  */
 842 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 843 {
 844         while (!pipe->nrbufs) {
 845                 if (!pipe->writers)
 846                         return 0;
 847
 848                 if (!pipe->waiting_writers && sd->num_spliced)
 849                         return 0;
 850
 851                 if (sd->flags & SPLICE_F_NONBLOCK)
 852                         return -EAGAIN;
 853
 854                 if (signal_pending(current))
 855                         return -ERESTARTSYS;
 856
 857                 if (sd->need_wakeup) {
 858                         wakeup_pipe_writers(pipe);
 859                         sd->need_wakeup = false;
 860                 }
 861
 862                 pipe_wait(pipe);
 863         }
 864
 865         return 1;
 866 }
 867 EXPORT_SYMBOL(splice_from_pipe_next);
 868
 869 /**
 870  * splice_from_pipe_begin - start splicing from pipe
 871  * @sd:         information about the splice operation
 872  *
 873  * Description:
 874  *    This function should be called before a loop containing
 875  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 876  *    initialize the necessary fields of @sd.
 877  */
 878 void splice_from_pipe_begin(struct splice_desc *sd)
 879 {
 880         sd->num_spliced = 0;
 881         sd->need_wakeup = false;
 882 }
 883 EXPORT_SYMBOL(splice_from_pipe_begin);
 884
 885 /**
 886  * splice_from_pipe_end - finish splicing from pipe
 887  * @pipe:       pipe to splice from
 888  * @sd:         information about the splice operation
 889  *
 890  * Description:
 891  *    This function will wake up pipe writers if necessary.  It should
 892  *    be called after a loop containing splice_from_pipe_next() and
 893  *    splice_from_pipe_feed().
 894  */
 895 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 896 {
 897         if (sd->need_wakeup)
 898                 wakeup_pipe_writers(pipe);
 899 }
 900 EXPORT_SYMBOL(splice_from_pipe_end);
 901
 902 /**
 903  * __splice_from_pipe - splice data from a pipe to given actor
 904  * @pipe:       pipe to splice from
 905  * @sd:         information to @actor
 906  * @actor:      handler that splices the data
 907  *
 908  * Description:
 909  *    This function does little more than loop over the pipe and call
 910  *    @actor to do the actual moving of a single struct pipe_buffer to
 911  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 912  *    pipe_to_user.
 913  *
 914  */
 915 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 916                            splice_actor *actor)
 917 {
 918         int ret;
 919
 920         splice_from_pipe_begin(sd);
 921         do {
 922                 ret = splice_from_pipe_next(pipe, sd);
 923                 if (ret > 0)
 924                         ret = splice_from_pipe_feed(pipe, sd, actor);
 925         } while (ret > 0);
 926         splice_from_pipe_end(pipe, sd);
 927
 928         return sd->num_spliced ? sd->num_spliced : ret;
 929 }
 930 EXPORT_SYMBOL(__splice_from_pipe);
 931
 932 /**
 933  * splice_from_pipe - splice data from a pipe to a file
 934  * @pipe:       pipe to splice from
 935  * @out:        file to splice to
 936  * @ppos:       position in @out
 937  * @len:        how many bytes to splice
 938  * @flags:      splice modifier flags
 939  * @actor:      handler that splices the data
 940  *
 941  * Description:
 942  *    See __splice_from_pipe. This function locks the pipe inode,
 943  *    otherwise it's identical to __splice_from_pipe().
 944  *
 945  */
 946 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 947                          loff_t *ppos, size_t len, unsigned int flags,
 948                          splice_actor *actor)
 949 {
 950         ssize_t ret;
 951         struct splice_desc sd = {
 952                 .total_len = len,
 953                 .flags = flags,
 954                 .pos = *ppos,
 955                 .u.file = out,
 956         };
 957
 958         pipe_lock(pipe);
 959         ret = __splice_from_pipe(pipe, &sd, actor);
 960         pipe_unlock(pipe);
 961
 962         return ret;
 963 }
 964
 965 /**
 966  * generic_file_splice_write - splice data from a pipe to a file
 967  * @pipe:       pipe info
 968  * @out:        file to write to
 969  * @ppos:       position in @out
 970  * @len:        number of bytes to splice
 971  * @flags:      splice modifier flags
 972  *
 973  * Description:
 974  *    Will either move or copy pages (determined by @flags options) from
 975  *    the given pipe inode to the given file.
 976  *
 977  */
 978 ssize_t
 979 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 980                           loff_t *ppos, size_t len, unsigned int flags)
 981 {
 982         struct address_space *mapping = out->f_mapping;
 983         struct inode *inode = mapping->host;
 984         struct splice_desc sd = {
 985                 .total_len = len,
 986                 .flags = flags,
 987                 .pos = *ppos,
 988                 .u.file = out,
 989         };
 990         ssize_t ret;
 991
 992         pipe_lock(pipe);
 993
 994         splice_from_pipe_begin(&sd);
 995         do {
 996                 ret = splice_from_pipe_next(pipe, &sd);
 997                 if (ret <= 0)
 998                         break;
 999
1000                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1001                 ret = file_remove_suid(out);
1002                 if (!ret) {
1003                         file_update_time(out);
1004                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1005                 }
1006                 mutex_unlock(&inode->i_mutex);
1007         } while (ret > 0);
1008         splice_from_pipe_end(pipe, &sd);
1009
1010         pipe_unlock(pipe);
1011
1012         if (sd.num_spliced)
1013                 ret = sd.num_spliced;
1014
1015         if (ret > 0) {
1016                 unsigned long nr_pages;
1017                 int err;
1018
1019                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1020
1021                 err = generic_write_sync(out, *ppos, ret);
1022                 if (err)
1023                         ret = err;
1024                 else
1025                         *ppos += ret;
1026                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1027         }
1028
1029         return ret;
1030 }
1031
1032 EXPORT_SYMBOL(generic_file_splice_write);
1033
1034 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1035                           struct splice_desc *sd)
1036 {
1037         int ret;
1038         void *data;
1039
1040         data = buf->ops->map(pipe, buf, 0);
1041         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1042         buf->ops->unmap(pipe, buf, data);
1043
1044         return ret;
1045 }
1046
1047 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1048                                          struct file *out, loff_t *ppos,
1049                                          size_t len, unsigned int flags)
1050 {
1051         ssize_t ret;
1052
1053         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1054         if (ret > 0)
1055                 *ppos += ret;
1056
1057         return ret;
1058 }
1059
1060 /**
1061  * generic_splice_sendpage - splice data from a pipe to a socket
1062  * @pipe:       pipe to splice from
1063  * @out:        socket to write to
1064  * @ppos:       position in @out
1065  * @len:        number of bytes to splice
1066  * @flags:      splice modifier flags
1067  *
1068  * Description:
1069  *    Will send @len bytes from the pipe to a network socket. No data copying
1070  *    is involved.
1071  *
1072  */
1073 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1074                                 loff_t *ppos, size_t len, unsigned int flags)
1075 {
1076         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1077 }
1078
1079 EXPORT_SYMBOL(generic_splice_sendpage);
1080
1081 /*
1082  * Attempt to initiate a splice from pipe to file.
1083  */
1084 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1085                            loff_t *ppos, size_t len, unsigned int flags)
1086 {
1087         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1088                                 loff_t *, size_t, unsigned int);
1089         int ret;
1090
1091         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1092                 return -EBADF;
1093
1094         if (unlikely(out->f_flags & O_APPEND))
1095                 return -EINVAL;
1096
1097         ret = rw_verify_area(WRITE, out, ppos, len);
1098         if (unlikely(ret < 0))
1099                 return ret;
1100
1101         if (out->f_op && out->f_op->splice_write)
1102                 splice_write = out->f_op->splice_write;
1103         else
1104                 splice_write = default_file_splice_write;
1105
1106         return splice_write(pipe, out, ppos, len, flags);
1107 }
1108
1109 /*
1110  * Attempt to initiate a splice from a file to a pipe.
1111  */
1112 static long do_splice_to(struct file *in, loff_t *ppos,
1113                          struct pipe_inode_info *pipe, size_t len,
1114                          unsigned int flags)
1115 {
1116         ssize_t (*splice_read)(struct file *, loff_t *,
1117                                struct pipe_inode_info *, size_t, unsigned int);
1118         int ret;
1119
1120         if (unlikely(!(in->f_mode & FMODE_READ)))
1121                 return -EBADF;
1122
1123         ret = rw_verify_area(READ, in, ppos, len);
1124         if (unlikely(ret < 0))
1125                 return ret;
1126
1127         if (in->f_op && in->f_op->splice_read)
1128                 splice_read = in->f_op->splice_read;
1129         else
1130                 splice_read = default_file_splice_read;
1131
1132         return splice_read(in, ppos, pipe, len, flags);
1133 }
1134
1135 /**
1136  * splice_direct_to_actor - splices data directly between two non-pipes
1137  * @in:         file to splice from
1138  * @sd:         actor information on where to splice to
1139  * @actor:      handles the data splicing
1140  *
1141  * Description:
1142  *    This is a special case helper to splice directly between two
1143  *    points, without requiring an explicit pipe. Internally an allocated
1144  *    pipe is cached in the process, and reused during the lifetime of
1145  *    that process.
1146  *
1147  */
1148 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1149                                splice_direct_actor *actor)
1150 {
1151         struct pipe_inode_info *pipe;
1152         long ret, bytes;
1153         umode_t i_mode;
1154         size_t len;
1155         int i, flags;
1156
1157         /*
1158          * We require the input being a regular file, as we don't want to
1159          * randomly drop data for eg socket -> socket splicing. Use the
1160          * piped splicing for that!
1161          */
1162         i_mode = in->f_path.dentry->d_inode->i_mode;
1163         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1164                 return -EINVAL;
1165
1166         /*
1167          * neither in nor out is a pipe, setup an internal pipe attached to
1168          * 'out' and transfer the wanted data from 'in' to 'out' through that
1169          */
1170         pipe = current->splice_pipe;
1171         if (unlikely(!pipe)) {
1172                 pipe = alloc_pipe_info(NULL);
1173                 if (!pipe)
1174                         return -ENOMEM;
1175
1176                 /*
1177                  * We don't have an immediate reader, but we'll read the stuff
1178                  * out of the pipe right after the splice_to_pipe(). So set
1179                  * PIPE_READERS appropriately.
1180                  */
1181                 pipe->readers = 1;
1182
1183                 current->splice_pipe = pipe;
1184         }
1185
1186         /*
1187          * Do the splice.
1188          */
1189         ret = 0;
1190         bytes = 0;
1191         len = sd->total_len;
1192         flags = sd->flags;
1193
1194         /*
1195          * Don't block on output, we have to drain the direct pipe.
1196          */
1197         sd->flags &= ~SPLICE_F_NONBLOCK;
1198
1199         while (len) {
1200                 size_t read_len;
1201                 loff_t pos = sd->pos, prev_pos = pos;
1202
1203                 ret = do_splice_to(in, &pos, pipe, len, flags);
1204                 if (unlikely(ret <= 0))
1205                         goto out_release;
1206
1207                 read_len = ret;
1208                 sd->total_len = read_len;
1209
1210                 /*
1211                  * NOTE: nonblocking mode only applies to the input. We
1212                  * must not do the output in nonblocking mode as then we
1213                  * could get stuck data in the internal pipe:
1214                  */
1215                 ret = actor(pipe, sd);
1216                 if (unlikely(ret <= 0)) {
1217                         sd->pos = prev_pos;
1218                         goto out_release;
1219                 }
1220
1221                 bytes += ret;
1222                 len -= ret;
1223                 sd->pos = pos;
1224
1225                 if (ret < read_len) {
1226                         sd->pos = prev_pos + ret;
1227                         goto out_release;
1228                 }
1229         }
1230
1231 done:
1232         pipe->nrbufs = pipe->curbuf = 0;
1233         file_accessed(in);
1234         return bytes;
1235
1236 out_release:
1237         /*
1238          * If we did an incomplete transfer we must release
1239          * the pipe buffers in question:
1240          */
1241         for (i = 0; i < pipe->buffers; i++) {
1242                 struct pipe_buffer *buf = pipe->bufs + i;
1243
1244                 if (buf->ops) {
1245                         buf->ops->release(pipe, buf);
1246                         buf->ops = NULL;
1247                 }
1248         }
1249
1250         if (!bytes)
1251                 bytes = ret;
1252
1253         goto done;
1254 }
1255 EXPORT_SYMBOL(splice_direct_to_actor);
1256
1257 static int direct_splice_actor(struct pipe_inode_info *pipe,
1258                                struct splice_desc *sd)
1259 {
1260         struct file *file = sd->u.file;
1261
1262         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1263                               sd->flags);
1264 }
1265
1266 /**
1267  * do_splice_direct - splices data directly between two files
1268  * @in:         file to splice from
1269  * @ppos:       input file offset
1270  * @out:        file to splice to
1271  * @len:        number of bytes to splice
1272  * @flags:      splice modifier flags
1273  *
1274  * Description:
1275  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1276  *    doing it in the application would incur an extra system call
1277  *    (splice in + splice out, as compared to just sendfile()). So this helper
1278  *    can splice directly through a process-private pipe.
1279  *
1280  */
1281 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1282                       size_t len, unsigned int flags)
1283 {
1284         struct splice_desc sd = {
1285                 .len            = len,
1286                 .total_len      = len,
1287                 .flags          = flags,
1288                 .pos            = *ppos,
1289                 .u.file         = out,
1290         };
1291         long ret;
1292
1293         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1294         if (ret > 0)
1295                 *ppos = sd.pos;
1296
1297         return ret;
1298 }
1299
1300 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1301                                struct pipe_inode_info *opipe,
1302                                size_t len, unsigned int flags);
1303
1304 /*
1305  * Determine where to splice to/from.
1306  */
1307 static long do_splice(struct file *in, loff_t __user *off_in,
1308                       struct file *out, loff_t __user *off_out,
1309                       size_t len, unsigned int flags)
1310 {
1311         struct pipe_inode_info *ipipe;
1312         struct pipe_inode_info *opipe;
1313         loff_t offset, *off;
1314         long ret;
1315
1316         ipipe = get_pipe_info(in);
1317         opipe = get_pipe_info(out);
1318
1319         if (ipipe && opipe) {
1320                 if (off_in || off_out)
1321                         return -ESPIPE;
1322
1323                 if (!(in->f_mode & FMODE_READ))
1324                         return -EBADF;
1325
1326                 if (!(out->f_mode & FMODE_WRITE))
1327                         return -EBADF;
1328
1329                 /* Splicing to self would be fun, but... */
1330                 if (ipipe == opipe)
1331                         return -EINVAL;
1332
1333                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1334         }
1335
1336         if (ipipe) {
1337                 if (off_in)
1338                         return -ESPIPE;
1339                 if (off_out) {
1340                         if (!(out->f_mode & FMODE_PWRITE))
1341                                 return -EINVAL;
1342                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1343                                 return -EFAULT;
1344                         off = &offset;
1345                 } else
1346                         off = &out->f_pos;
1347
1348                 ret = do_splice_from(ipipe, out, off, len, flags);
1349
1350                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1351                         ret = -EFAULT;
1352
1353                 return ret;
1354         }
1355
1356         if (opipe) {
1357                 if (off_out)
1358                         return -ESPIPE;
1359                 if (off_in) {
1360                         if (!(in->f_mode & FMODE_PREAD))
1361                                 return -EINVAL;
1362                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1363                                 return -EFAULT;
1364                         off = &offset;
1365                 } else
1366                         off = &in->f_pos;
1367
1368                 ret = do_splice_to(in, off, opipe, len, flags);
1369
1370                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1371                         ret = -EFAULT;
1372
1373                 return ret;
1374         }
1375
1376         return -EINVAL;
1377 }
1378
1379 /*
1380  * Map an iov into an array of pages and offset/length tupples. With the
1381  * partial_page structure, we can map several non-contiguous ranges into
1382  * our ones pages[] map instead of splitting that operation into pieces.
1383  * Could easily be exported as a generic helper for other users, in which
1384  * case one would probably want to add a 'max_nr_pages' parameter as well.
1385  */
1386 static int get_iovec_page_array(const struct iovec __user *iov,
1387                                 unsigned int nr_vecs, struct page **pages,
1388                                 struct partial_page *partial, int aligned,
1389                                 unsigned int pipe_buffers)
1390 {
1391         int buffers = 0, error = 0;
1392
1393         while (nr_vecs) {
1394                 unsigned long off, npages;
1395                 struct iovec entry;
1396                 void __user *base;
1397                 size_t len;
1398                 int i;
1399
1400                 error = -EFAULT;
1401                 if (copy_from_user(&entry, iov, sizeof(entry)))
1402                         break;
1403
1404                 base = entry.iov_base;
1405                 len = entry.iov_len;
1406
1407                 /*
1408                  * Sanity check this iovec. 0 read succeeds.
1409                  */
1410                 error = 0;
1411                 if (unlikely(!len))
1412                         break;
1413                 error = -EFAULT;
1414                 if (!access_ok(VERIFY_READ, base, len))
1415                         break;
1416
1417                 /*
1418                  * Get this base offset and number of pages, then map
1419                  * in the user pages.
1420                  */
1421                 off = (unsigned long) base & ~PAGE_MASK;
1422
1423                 /*
1424                  * If asked for alignment, the offset must be zero and the
1425                  * length a multiple of the PAGE_SIZE.
1426                  */
1427                 error = -EINVAL;
1428                 if (aligned && (off || len & ~PAGE_MASK))
1429                         break;
1430
1431                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1432                 if (npages > pipe_buffers - buffers)
1433                         npages = pipe_buffers - buffers;
1434
1435                 error = get_user_pages_fast((unsigned long)base, npages,
1436                                         0, &pages[buffers]);
1437
1438                 if (unlikely(error <= 0))
1439                         break;
1440
1441                 /*
1442                  * Fill this contiguous range into the partial page map.
1443                  */
1444                 for (i = 0; i < error; i++) {
1445                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1446
1447                         partial[buffers].offset = off;
1448                         partial[buffers].len = plen;
1449
1450                         off = 0;
1451                         len -= plen;
1452                         buffers++;
1453                 }
1454
1455                 /*
1456                  * We didn't complete this iov, stop here since it probably
1457                  * means we have to move some of this into a pipe to
1458                  * be able to continue.
1459                  */
1460                 if (len)
1461                         break;
1462
1463                 /*
1464                  * Don't continue if we mapped fewer pages than we asked for,
1465                  * or if we mapped the max number of pages that we have
1466                  * room for.
1467                  */
1468                 if (error < npages || buffers == pipe_buffers)
1469                         break;
1470
1471                 nr_vecs--;
1472                 iov++;
1473         }
1474
1475         if (buffers)
1476                 return buffers;
1477
1478         return error;
1479 }
1480
1481 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1482                         struct splice_desc *sd)
1483 {
1484         char *src;
1485         int ret;
1486
1487         /*
1488          * See if we can use the atomic maps, by prefaulting in the
1489          * pages and doing an atomic copy
1490          */
1491         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1492                 src = buf->ops->map(pipe, buf, 1);
1493                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1494                                                         sd->len);
1495                 buf->ops->unmap(pipe, buf, src);
1496                 if (!ret) {
1497                         ret = sd->len;
1498                         goto out;
1499                 }
1500         }
1501
1502         /*
1503          * No dice, use slow non-atomic map and copy
1504          */
1505         src = buf->ops->map(pipe, buf, 0);
1506
1507         ret = sd->len;
1508         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1509                 ret = -EFAULT;
1510
1511         buf->ops->unmap(pipe, buf, src);
1512 out:
1513         if (ret > 0)
1514                 sd->u.userptr += ret;
1515         return ret;
1516 }
1517
1518 /*
1519  * For lack of a better implementation, implement vmsplice() to userspace
1520  * as a simple copy of the pipes pages to the user iov.
1521  */
1522 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1523                              unsigned long nr_segs, unsigned int flags)
1524 {
1525         struct pipe_inode_info *pipe;
1526         struct splice_desc sd;
1527         ssize_t size;
1528         int error;
1529         long ret;
1530
1531         pipe = get_pipe_info(file);
1532         if (!pipe)
1533                 return -EBADF;
1534
1535         pipe_lock(pipe);
1536
1537         error = ret = 0;
1538         while (nr_segs) {
1539                 void __user *base;
1540                 size_t len;
1541
1542                 /*
1543                  * Get user address base and length for this iovec.
1544                  */
1545                 error = get_user(base, &iov->iov_base);
1546                 if (unlikely(error))
1547                         break;
1548                 error = get_user(len, &iov->iov_len);
1549                 if (unlikely(error))
1550                         break;
1551
1552                 /*
1553                  * Sanity check this iovec. 0 read succeeds.
1554                  */
1555                 if (unlikely(!len))
1556                         break;
1557                 if (unlikely(!base)) {
1558                         error = -EFAULT;
1559                         break;
1560                 }
1561
1562                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1563                         error = -EFAULT;
1564                         break;
1565                 }
1566
1567                 sd.len = 0;
1568                 sd.total_len = len;
1569                 sd.flags = flags;
1570                 sd.u.userptr = base;
1571                 sd.pos = 0;
1572
1573                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1574                 if (size < 0) {
1575                         if (!ret)
1576                                 ret = size;
1577
1578                         break;
1579                 }
1580
1581                 ret += size;
1582
1583                 if (size < len)
1584                         break;
1585
1586                 nr_segs--;
1587                 iov++;
1588         }
1589
1590         pipe_unlock(pipe);
1591
1592         if (!ret)
1593                 ret = error;
1594
1595         return ret;
1596 }
1597
1598 /*
1599  * vmsplice splices a user address range into a pipe. It can be thought of
1600  * as splice-from-memory, where the regular splice is splice-from-file (or
1601  * to file). In both cases the output is a pipe, naturally.
1602  */
1603 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1604                              unsigned long nr_segs, unsigned int flags)
1605 {
1606         struct pipe_inode_info *pipe;
1607         struct page *pages[PIPE_DEF_BUFFERS];
1608         struct partial_page partial[PIPE_DEF_BUFFERS];
1609         struct splice_pipe_desc spd = {
1610                 .pages = pages,
1611                 .partial = partial,
1612                 .flags = flags,
1613                 .ops = &user_page_pipe_buf_ops,
1614                 .spd_release = spd_release_page,
1615         };
1616         long ret;
1617
1618         pipe = get_pipe_info(file);
1619         if (!pipe)
1620                 return -EBADF;
1621
1622         if (splice_grow_spd(pipe, &spd))
1623                 return -ENOMEM;
1624
1625         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1626                                             spd.partial, flags & SPLICE_F_GIFT,
1627                                             pipe->buffers);
1628         if (spd.nr_pages <= 0)
1629                 ret = spd.nr_pages;
1630         else
1631                 ret = splice_to_pipe(pipe, &spd);
1632
1633         splice_shrink_spd(pipe, &spd);
1634         return ret;
1635 }
1636
1637 /*
1638  * Note that vmsplice only really supports true splicing _from_ user memory
1639  * to a pipe, not the other way around. Splicing from user memory is a simple
1640  * operation that can be supported without any funky alignment restrictions
1641  * or nasty vm tricks. We simply map in the user memory and fill them into
1642  * a pipe. The reverse isn't quite as easy, though. There are two possible
1643  * solutions for that:
1644  *
1645  *      - memcpy() the data internally, at which point we might as well just
1646  *        do a regular read() on the buffer anyway.
1647  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1648  *        has restriction limitations on both ends of the pipe).
1649  *
1650  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1651  *
1652  */
1653 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1654                 unsigned long, nr_segs, unsigned int, flags)
1655 {
1656         struct file *file;
1657         long error;
1658         int fput;
1659
1660         if (unlikely(nr_segs > UIO_MAXIOV))
1661                 return -EINVAL;
1662         else if (unlikely(!nr_segs))
1663                 return 0;
1664
1665         error = -EBADF;
1666         file = fget_light(fd, &fput);
1667         if (file) {
1668                 if (file->f_mode & FMODE_WRITE)
1669                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1670                 else if (file->f_mode & FMODE_READ)
1671                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1672
1673                 fput_light(file, fput);
1674         }
1675
1676         return error;
1677 }
1678
1679 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1680                 int, fd_out, loff_t __user *, off_out,
1681                 size_t, len, unsigned int, flags)
1682 {
1683         long error;
1684         struct file *in, *out;
1685         int fput_in, fput_out;
1686
1687         if (unlikely(!len))
1688                 return 0;
1689
1690         error = -EBADF;
1691         in = fget_light(fd_in, &fput_in);
1692         if (in) {
1693                 if (in->f_mode & FMODE_READ) {
1694                         out = fget_light(fd_out, &fput_out);
1695                         if (out) {
1696                                 if (out->f_mode & FMODE_WRITE)
1697                                         error = do_splice(in, off_in,
1698                                                           out, off_out,
1699                                                           len, flags);
1700                                 fput_light(out, fput_out);
1701                         }
1702                 }
1703
1704                 fput_light(in, fput_in);
1705         }
1706
1707         return error;
1708 }
1709
1710 /*
1711  * Make sure there's data to read. Wait for input if we can, otherwise
1712  * return an appropriate error.
1713  */
1714 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1715 {
1716         int ret;
1717
1718         /*
1719          * Check ->nrbufs without the inode lock first. This function
1720          * is speculative anyways, so missing one is ok.
1721          */
1722         if (pipe->nrbufs)
1723                 return 0;
1724
1725         ret = 0;
1726         pipe_lock(pipe);
1727
1728         while (!pipe->nrbufs) {
1729                 if (signal_pending(current)) {
1730                         ret = -ERESTARTSYS;
1731                         break;
1732                 }
1733                 if (!pipe->writers)
1734                         break;
1735                 if (!pipe->waiting_writers) {
1736                         if (flags & SPLICE_F_NONBLOCK) {
1737                                 ret = -EAGAIN;
1738                                 break;
1739                         }
1740                 }
1741                 pipe_wait(pipe);
1742         }
1743
1744         pipe_unlock(pipe);
1745         return ret;
1746 }
1747
1748 /*
1749  * Make sure there's writeable room. Wait for room if we can, otherwise
1750  * return an appropriate error.
1751  */
1752 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1753 {
1754         int ret;
1755
1756         /*
1757          * Check ->nrbufs without the inode lock first. This function
1758          * is speculative anyways, so missing one is ok.
1759          */
1760         if (pipe->nrbufs < pipe->buffers)
1761                 return 0;
1762
1763         ret = 0;
1764         pipe_lock(pipe);
1765
1766         while (pipe->nrbufs >= pipe->buffers) {
1767                 if (!pipe->readers) {
1768                         send_sig(SIGPIPE, current, 0);
1769                         ret = -EPIPE;
1770                         break;
1771                 }
1772                 if (flags & SPLICE_F_NONBLOCK) {
1773                         ret = -EAGAIN;
1774                         break;
1775                 }
1776                 if (signal_pending(current)) {
1777                         ret = -ERESTARTSYS;
1778                         break;
1779                 }
1780                 pipe->waiting_writers++;
1781                 pipe_wait(pipe);
1782                 pipe->waiting_writers--;
1783         }
1784
1785         pipe_unlock(pipe);
1786         return ret;
1787 }
1788
1789 /*
1790  * Splice contents of ipipe to opipe.
1791  */
1792 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1793                                struct pipe_inode_info *opipe,
1794                                size_t len, unsigned int flags)
1795 {
1796         struct pipe_buffer *ibuf, *obuf;
1797         int ret = 0, nbuf;
1798         bool input_wakeup = false;
1799
1800
1801 retry:
1802         ret = ipipe_prep(ipipe, flags);
1803         if (ret)
1804                 return ret;
1805
1806         ret = opipe_prep(opipe, flags);
1807         if (ret)
1808                 return ret;
1809
1810         /*
1811          * Potential ABBA deadlock, work around it by ordering lock
1812          * grabbing by pipe info address. Otherwise two different processes
1813          * could deadlock (one doing tee from A -> B, the other from B -> A).
1814          */
1815         pipe_double_lock(ipipe, opipe);
1816
1817         do {
1818                 if (!opipe->readers) {
1819                         send_sig(SIGPIPE, current, 0);
1820                         if (!ret)
1821                                 ret = -EPIPE;
1822                         break;
1823                 }
1824
1825                 if (!ipipe->nrbufs && !ipipe->writers)
1826                         break;
1827
1828                 /*
1829                  * Cannot make any progress, because either the input
1830                  * pipe is empty or the output pipe is full.
1831                  */
1832                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1833                         /* Already processed some buffers, break */
1834                         if (ret)
1835                                 break;
1836
1837                         if (flags & SPLICE_F_NONBLOCK) {
1838                                 ret = -EAGAIN;
1839                                 break;
1840                         }
1841
1842                         /*
1843                          * We raced with another reader/writer and haven't
1844                          * managed to process any buffers.  A zero return
1845                          * value means EOF, so retry instead.
1846                          */
1847                         pipe_unlock(ipipe);
1848                         pipe_unlock(opipe);
1849                         goto retry;
1850                 }
1851
1852                 ibuf = ipipe->bufs + ipipe->curbuf;
1853                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1854                 obuf = opipe->bufs + nbuf;
1855
1856                 if (len >= ibuf->len) {
1857                         /*
1858                          * Simply move the whole buffer from ipipe to opipe
1859                          */
1860                         *obuf = *ibuf;
1861                         ibuf->ops = NULL;
1862                         opipe->nrbufs++;
1863                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1864                         ipipe->nrbufs--;
1865                         input_wakeup = true;
1866                 } else {
1867                         /*
1868                          * Get a reference to this pipe buffer,
1869                          * so we can copy the contents over.
1870                          */
1871                         ibuf->ops->get(ipipe, ibuf);
1872                         *obuf = *ibuf;
1873
1874                         /*
1875                          * Don't inherit the gift flag, we need to
1876                          * prevent multiple steals of this page.
1877                          */
1878                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1879
1880                         obuf->len = len;
1881                         opipe->nrbufs++;
1882                         ibuf->offset += obuf->len;
1883                         ibuf->len -= obuf->len;
1884                 }
1885                 ret += obuf->len;
1886                 len -= obuf->len;
1887         } while (len);
1888
1889         pipe_unlock(ipipe);
1890         pipe_unlock(opipe);
1891
1892         /*
1893          * If we put data in the output pipe, wakeup any potential readers.
1894          */
1895         if (ret > 0)
1896                 wakeup_pipe_readers(opipe);
1897
1898         if (input_wakeup)
1899                 wakeup_pipe_writers(ipipe);
1900
1901         return ret;
1902 }
1903
1904 /*
1905  * Link contents of ipipe to opipe.
1906  */
1907 static int link_pipe(struct pipe_inode_info *ipipe,
1908                      struct pipe_inode_info *opipe,
1909                      size_t len, unsigned int flags)
1910 {
1911         struct pipe_buffer *ibuf, *obuf;
1912         int ret = 0, i = 0, nbuf;
1913
1914         /*
1915          * Potential ABBA deadlock, work around it by ordering lock
1916          * grabbing by pipe info address. Otherwise two different processes
1917          * could deadlock (one doing tee from A -> B, the other from B -> A).
1918          */
1919         pipe_double_lock(ipipe, opipe);
1920
1921         do {
1922                 if (!opipe->readers) {
1923                         send_sig(SIGPIPE, current, 0);
1924                         if (!ret)
1925                                 ret = -EPIPE;
1926                         break;
1927                 }
1928
1929                 /*
1930                  * If we have iterated all input buffers or ran out of
1931                  * output room, break.
1932                  */
1933                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1934                         break;
1935
1936                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1937                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1938
1939                 /*
1940                  * Get a reference to this pipe buffer,
1941                  * so we can copy the contents over.
1942                  */
1943                 ibuf->ops->get(ipipe, ibuf);
1944
1945                 obuf = opipe->bufs + nbuf;
1946                 *obuf = *ibuf;
1947
1948                 /*
1949                  * Don't inherit the gift flag, we need to
1950                  * prevent multiple steals of this page.
1951                  */
1952                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1953
1954                 if (obuf->len > len)
1955                         obuf->len = len;
1956
1957                 opipe->nrbufs++;
1958                 ret += obuf->len;
1959                 len -= obuf->len;
1960                 i++;
1961         } while (len);
1962
1963         /*
1964          * return EAGAIN if we have the potential of some data in the
1965          * future, otherwise just return 0
1966          */
1967         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1968                 ret = -EAGAIN;
1969
1970         pipe_unlock(ipipe);
1971         pipe_unlock(opipe);
1972
1973         /*
1974          * If we put data in the output pipe, wakeup any potential readers.
1975          */
1976         if (ret > 0)
1977                 wakeup_pipe_readers(opipe);
1978
1979         return ret;
1980 }
1981
1982 /*
1983  * This is a tee(1) implementation that works on pipes. It doesn't copy
1984  * any data, it simply references the 'in' pages on the 'out' pipe.
1985  * The 'flags' used are the SPLICE_F_* variants, currently the only
1986  * applicable one is SPLICE_F_NONBLOCK.
1987  */
1988 static long do_tee(struct file *in, struct file *out, size_t len,
1989                    unsigned int flags)
1990 {
1991         struct pipe_inode_info *ipipe = get_pipe_info(in);
1992         struct pipe_inode_info *opipe = get_pipe_info(out);
1993         int ret = -EINVAL;
1994
1995         /*
1996          * Duplicate the contents of ipipe to opipe without actually
1997          * copying the data.
1998          */
1999         if (ipipe && opipe && ipipe != opipe) {
2000                 /*
2001                  * Keep going, unless we encounter an error. The ipipe/opipe
2002                  * ordering doesn't really matter.
2003                  */
2004                 ret = ipipe_prep(ipipe, flags);
2005                 if (!ret) {
2006                         ret = opipe_prep(opipe, flags);
2007                         if (!ret)
2008                                 ret = link_pipe(ipipe, opipe, len, flags);
2009                 }
2010         }
2011
2012         return ret;
2013 }
2014
2015 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2016 {
2017         struct file *in;
2018         int error, fput_in;
2019
2020         if (unlikely(!len))
2021                 return 0;
2022
2023         error = -EBADF;
2024         in = fget_light(fdin, &fput_in);
2025         if (in) {
2026                 if (in->f_mode & FMODE_READ) {
2027                         int fput_out;
2028                         struct file *out = fget_light(fdout, &fput_out);
2029
2030                         if (out) {
2031                                 if (out->f_mode & FMODE_WRITE)
2032                                         error = do_tee(in, out, len, flags);
2033                                 fput_light(out, fput_out);
2034                         }
2035                 }
2036                 fput_light(in, fput_in);
2037         }
2038
2039         return error;
2040 }