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(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 278 {
 279         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 280
 281         spd->nr_pages_max = buffers;
 282         if (buffers <= PIPE_DEF_BUFFERS)
 283                 return 0;
 284
 285         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 286         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 287
 288         if (spd->pages && spd->partial)
 289                 return 0;
 290
 291         kfree(spd->pages);
 292         kfree(spd->partial);
 293         return -ENOMEM;
 294 }
 295
 296 void splice_shrink_spd(struct splice_pipe_desc *spd)
 297 {
 298         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 299                 return;
 300
 301         kfree(spd->pages);
 302         kfree(spd->partial);
 303 }
 304
 305 static int
 306 __generic_file_splice_read(struct file *in, loff_t *ppos,
 307                            struct pipe_inode_info *pipe, size_t len,
 308                            unsigned int flags)
 309 {
 310         struct address_space *mapping = in->f_mapping;
 311         unsigned int loff, nr_pages, req_pages;
 312         struct page *pages[PIPE_DEF_BUFFERS];
 313         struct partial_page partial[PIPE_DEF_BUFFERS];
 314         struct page *page;
 315         pgoff_t index, end_index;
 316         loff_t isize;
 317         int error, page_nr;
 318         struct splice_pipe_desc spd = {
 319                 .pages = pages,
 320                 .partial = partial,
 321                 .nr_pages_max = PIPE_DEF_BUFFERS,
 322                 .flags = flags,
 323                 .ops = &page_cache_pipe_buf_ops,
 324                 .spd_release = spd_release_page,
 325         };
 326
 327         if (splice_grow_spd(pipe, &spd))
 328                 return -ENOMEM;
 329
 330         index = *ppos >> PAGE_CACHE_SHIFT;
 331         loff = *ppos & ~PAGE_CACHE_MASK;
 332         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 333         nr_pages = min(req_pages, spd.nr_pages_max);
 334
 335         /*
 336          * Lookup the (hopefully) full range of pages we need.
 337          */
 338         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 339         index += spd.nr_pages;
 340
 341         /*
 342          * If find_get_pages_contig() returned fewer pages than we needed,
 343          * readahead/allocate the rest and fill in the holes.
 344          */
 345         if (spd.nr_pages < nr_pages)
 346                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 347                                 index, req_pages - spd.nr_pages);
 348
 349         error = 0;
 350         while (spd.nr_pages < nr_pages) {
 351                 /*
 352                  * Page could be there, find_get_pages_contig() breaks on
 353                  * the first hole.
 354                  */
 355                 page = find_get_page(mapping, index);
 356                 if (!page) {
 357                         /*
 358                          * page didn't exist, allocate one.
 359                          */
 360                         page = page_cache_alloc_cold(mapping);
 361                         if (!page)
 362                                 break;
 363
 364                         error = add_to_page_cache_lru(page, mapping, index,
 365                                                 GFP_KERNEL);
 366                         if (unlikely(error)) {
 367                                 page_cache_release(page);
 368                                 if (error == -EEXIST)
 369                                         continue;
 370                                 break;
 371                         }
 372                         /*
 373                          * add_to_page_cache() locks the page, unlock it
 374                          * to avoid convoluting the logic below even more.
 375                          */
 376                         unlock_page(page);
 377                 }
 378
 379                 spd.pages[spd.nr_pages++] = page;
 380                 index++;
 381         }
 382
 383         /*
 384          * Now loop over the map and see if we need to start IO on any
 385          * pages, fill in the partial map, etc.
 386          */
 387         index = *ppos >> PAGE_CACHE_SHIFT;
 388         nr_pages = spd.nr_pages;
 389         spd.nr_pages = 0;
 390         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 391                 unsigned int this_len;
 392
 393                 if (!len)
 394                         break;
 395
 396                 /*
 397                  * this_len is the max we'll use from this page
 398                  */
 399                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 400                 page = spd.pages[page_nr];
 401
 402                 if (PageReadahead(page))
 403                         page_cache_async_readahead(mapping, &in->f_ra, in,
 404                                         page, index, req_pages - page_nr);
 405
 406                 /*
 407                  * If the page isn't uptodate, we may need to start io on it
 408                  */
 409                 if (!PageUptodate(page)) {
 410                         lock_page(page);
 411
 412                         /*
 413                          * Page was truncated, or invalidated by the
 414                          * filesystem.  Redo the find/create, but this time the
 415                          * page is kept locked, so there's no chance of another
 416                          * race with truncate/invalidate.
 417                          */
 418                         if (!page->mapping) {
 419                                 unlock_page(page);
 420                                 page = find_or_create_page(mapping, index,
 421                                                 mapping_gfp_mask(mapping));
 422
 423                                 if (!page) {
 424                                         error = -ENOMEM;
 425                                         break;
 426                                 }
 427                                 page_cache_release(spd.pages[page_nr]);
 428                                 spd.pages[page_nr] = page;
 429                         }
 430                         /*
 431                          * page was already under io and is now done, great
 432                          */
 433                         if (PageUptodate(page)) {
 434                                 unlock_page(page);
 435                                 goto fill_it;
 436                         }
 437
 438                         /*
 439                          * need to read in the page
 440                          */
 441                         error = mapping->a_ops->readpage(in, page);
 442                         if (unlikely(error)) {
 443                                 /*
 444                                  * We really should re-lookup the page here,
 445                                  * but it complicates things a lot. Instead
 446                                  * lets just do what we already stored, and
 447                                  * we'll get it the next time we are called.
 448                                  */
 449                                 if (error == AOP_TRUNCATED_PAGE)
 450                                         error = 0;
 451
 452                                 break;
 453                         }
 454                 }
 455 fill_it:
 456                 /*
 457                  * i_size must be checked after PageUptodate.
 458                  */
 459                 isize = i_size_read(mapping->host);
 460                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 461                 if (unlikely(!isize || index > end_index))
 462                         break;
 463
 464                 /*
 465                  * if this is the last page, see if we need to shrink
 466                  * the length and stop
 467                  */
 468                 if (end_index == index) {
 469                         unsigned int plen;
 470
 471                         /*
 472                          * max good bytes in this page
 473                          */
 474                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 475                         if (plen <= loff)
 476                                 break;
 477
 478                         /*
 479                          * force quit after adding this page
 480                          */
 481                         this_len = min(this_len, plen - loff);
 482                         len = this_len;
 483                 }
 484
 485                 spd.partial[page_nr].offset = loff;
 486                 spd.partial[page_nr].len = this_len;
 487                 len -= this_len;
 488                 loff = 0;
 489                 spd.nr_pages++;
 490                 index++;
 491         }
 492
 493         /*
 494          * Release any pages at the end, if we quit early. 'page_nr' is how far
 495          * we got, 'nr_pages' is how many pages are in the map.
 496          */
 497         while (page_nr < nr_pages)
 498                 page_cache_release(spd.pages[page_nr++]);
 499         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 500
 501         if (spd.nr_pages)
 502                 error = splice_to_pipe(pipe, &spd);
 503
 504         splice_shrink_spd(&spd);
 505         return error;
 506 }
 507
 508 /**
 509  * generic_file_splice_read - splice data from file to a pipe
 510  * @in:         file to splice from
 511  * @ppos:       position in @in
 512  * @pipe:       pipe to splice to
 513  * @len:        number of bytes to splice
 514  * @flags:      splice modifier flags
 515  *
 516  * Description:
 517  *    Will read pages from given file and fill them into a pipe. Can be
 518  *    used as long as the address_space operations for the source implements
 519  *    a readpage() hook.
 520  *
 521  */
 522 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 523                                  struct pipe_inode_info *pipe, size_t len,
 524                                  unsigned int flags)
 525 {
 526         loff_t isize, left;
 527         int ret;
 528
 529         isize = i_size_read(in->f_mapping->host);
 530         if (unlikely(*ppos >= isize))
 531                 return 0;
 532
 533         left = isize - *ppos;
 534         if (unlikely(left < len))
 535                 len = left;
 536
 537         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 538         if (ret > 0) {
 539                 *ppos += ret;
 540                 file_accessed(in);
 541         }
 542
 543         return ret;
 544 }
 545 EXPORT_SYMBOL(generic_file_splice_read);
 546
 547 static const struct pipe_buf_operations default_pipe_buf_ops = {
 548         .can_merge = 0,
 549         .map = generic_pipe_buf_map,
 550         .unmap = generic_pipe_buf_unmap,
 551         .confirm = generic_pipe_buf_confirm,
 552         .release = generic_pipe_buf_release,
 553         .steal = generic_pipe_buf_steal,
 554         .get = generic_pipe_buf_get,
 555 };
 556
 557 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 558                             unsigned long vlen, loff_t offset)
 559 {
 560         mm_segment_t old_fs;
 561         loff_t pos = offset;
 562         ssize_t res;
 563
 564         old_fs = get_fs();
 565         set_fs(get_ds());
 566         /* The cast to a user pointer is valid due to the set_fs() */
 567         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 568         set_fs(old_fs);
 569
 570         return res;
 571 }
 572
 573 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 574                             loff_t pos)
 575 {
 576         mm_segment_t old_fs;
 577         ssize_t res;
 578
 579         old_fs = get_fs();
 580         set_fs(get_ds());
 581         /* The cast to a user pointer is valid due to the set_fs() */
 582         res = vfs_write(file, (const char __user *)buf, count, &pos);
 583         set_fs(old_fs);
 584
 585         return res;
 586 }
 587
 588 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 589                                  struct pipe_inode_info *pipe, size_t len,
 590                                  unsigned int flags)
 591 {
 592         unsigned int nr_pages;
 593         unsigned int nr_freed;
 594         size_t offset;
 595         struct page *pages[PIPE_DEF_BUFFERS];
 596         struct partial_page partial[PIPE_DEF_BUFFERS];
 597         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 598         ssize_t res;
 599         size_t this_len;
 600         int error;
 601         int i;
 602         struct splice_pipe_desc spd = {
 603                 .pages = pages,
 604                 .partial = partial,
 605                 .nr_pages_max = PIPE_DEF_BUFFERS,
 606                 .flags = flags,
 607                 .ops = &default_pipe_buf_ops,
 608                 .spd_release = spd_release_page,
 609         };
 610
 611         if (splice_grow_spd(pipe, &spd))
 612                 return -ENOMEM;
 613
 614         res = -ENOMEM;
 615         vec = __vec;
 616         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 617                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 618                 if (!vec)
 619                         goto shrink_ret;
 620         }
 621
 622         offset = *ppos & ~PAGE_CACHE_MASK;
 623         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 624
 625         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 626                 struct page *page;
 627
 628                 page = alloc_page(GFP_USER);
 629                 error = -ENOMEM;
 630                 if (!page)
 631                         goto err;
 632
 633                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 634                 vec[i].iov_base = (void __user *) page_address(page);
 635                 vec[i].iov_len = this_len;
 636                 spd.pages[i] = page;
 637                 spd.nr_pages++;
 638                 len -= this_len;
 639                 offset = 0;
 640         }
 641
 642         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 643         if (res < 0) {
 644                 error = res;
 645                 goto err;
 646         }
 647
 648         error = 0;
 649         if (!res)
 650                 goto err;
 651
 652         nr_freed = 0;
 653         for (i = 0; i < spd.nr_pages; i++) {
 654                 this_len = min_t(size_t, vec[i].iov_len, res);
 655                 spd.partial[i].offset = 0;
 656                 spd.partial[i].len = this_len;
 657                 if (!this_len) {
 658                         __free_page(spd.pages[i]);
 659                         spd.pages[i] = NULL;
 660                         nr_freed++;
 661                 }
 662                 res -= this_len;
 663         }
 664         spd.nr_pages -= nr_freed;
 665
 666         res = splice_to_pipe(pipe, &spd);
 667         if (res > 0)
 668                 *ppos += res;
 669
 670 shrink_ret:
 671         if (vec != __vec)
 672                 kfree(vec);
 673         splice_shrink_spd(&spd);
 674         return res;
 675
 676 err:
 677         for (i = 0; i < spd.nr_pages; i++)
 678                 __free_page(spd.pages[i]);
 679
 680         res = error;
 681         goto shrink_ret;
 682 }
 683 EXPORT_SYMBOL(default_file_splice_read);
 684
 685 /*
 686  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 687  * using sendpage(). Return the number of bytes sent.
 688  */
 689 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 690                             struct pipe_buffer *buf, struct splice_desc *sd)
 691 {
 692         struct file *file = sd->u.file;
 693         loff_t pos = sd->pos;
 694         int more;
 695
 696         if (!likely(file->f_op && file->f_op->sendpage))
 697                 return -EINVAL;
 698
 699         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 700         if (sd->len < sd->total_len)
 701                 more |= MSG_SENDPAGE_NOTLAST;
 702         return file->f_op->sendpage(file, buf->page, buf->offset,
 703                                     sd->len, &pos, more);
 704 }
 705
 706 /*
 707  * This is a little more tricky than the file -> pipe splicing. There are
 708  * basically three cases:
 709  *
 710  *      - Destination page already exists in the address space and there
 711  *        are users of it. For that case we have no other option that
 712  *        copying the data. Tough luck.
 713  *      - Destination page already exists in the address space, but there
 714  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 715  *        through to last case.
 716  *      - Destination page does not exist, we can add the pipe page to
 717  *        the page cache and avoid the copy.
 718  *
 719  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 720  * sd->flags), we attempt to migrate pages from the pipe to the output
 721  * file address space page cache. This is possible if no one else has
 722  * the pipe page referenced outside of the pipe and page cache. If
 723  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 724  * a new page in the output file page cache and fill/dirty that.
 725  */
 726 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 727                  struct splice_desc *sd)
 728 {
 729         struct file *file = sd->u.file;
 730         struct address_space *mapping = file->f_mapping;
 731         unsigned int offset, this_len;
 732         struct page *page;
 733         void *fsdata;
 734         int ret;
 735
 736         offset = sd->pos & ~PAGE_CACHE_MASK;
 737
 738         this_len = sd->len;
 739         if (this_len + offset > PAGE_CACHE_SIZE)
 740                 this_len = PAGE_CACHE_SIZE - offset;
 741
 742         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 743                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 744         if (unlikely(ret))
 745                 goto out;
 746
 747         if (buf->page != page) {
 748                 /*
 749                  * Careful, ->map() uses KM_USER0!
 750                  */
 751                 char *src = buf->ops->map(pipe, buf, 1);
 752                 char *dst = kmap_atomic(page, KM_USER1);
 753
 754                 memcpy(dst + offset, src + buf->offset, this_len);
 755                 flush_dcache_page(page);
 756                 kunmap_atomic(dst, KM_USER1);
 757                 buf->ops->unmap(pipe, buf, src);
 758         }
 759         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 760                                 page, fsdata);
 761 out:
 762         return ret;
 763 }
 764 EXPORT_SYMBOL(pipe_to_file);
 765
 766 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 767 {
 768         smp_mb();
 769         if (waitqueue_active(&pipe->wait))
 770                 wake_up_interruptible(&pipe->wait);
 771         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 772 }
 773
 774 /**
 775  * splice_from_pipe_feed - feed available data from a pipe to a file
 776  * @pipe:       pipe to splice from
 777  * @sd:         information to @actor
 778  * @actor:      handler that splices the data
 779  *
 780  * Description:
 781  *    This function loops over the pipe and calls @actor to do the
 782  *    actual moving of a single struct pipe_buffer to the desired
 783  *    destination.  It returns when there's no more buffers left in
 784  *    the pipe or if the requested number of bytes (@sd->total_len)
 785  *    have been copied.  It returns a positive number (one) if the
 786  *    pipe needs to be filled with more data, zero if the required
 787  *    number of bytes have been copied and -errno on error.
 788  *
 789  *    This, together with splice_from_pipe_{begin,end,next}, may be
 790  *    used to implement the functionality of __splice_from_pipe() when
 791  *    locking is required around copying the pipe buffers to the
 792  *    destination.
 793  */
 794 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 795                           splice_actor *actor)
 796 {
 797         int ret;
 798
 799         while (pipe->nrbufs) {
 800                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 801                 const struct pipe_buf_operations *ops = buf->ops;
 802
 803                 sd->len = buf->len;
 804                 if (sd->len > sd->total_len)
 805                         sd->len = sd->total_len;
 806
 807                 ret = buf->ops->confirm(pipe, buf);
 808                 if (unlikely(ret)) {
 809                         if (ret == -ENODATA)
 810                                 ret = 0;
 811                         return ret;
 812                 }
 813
 814                 ret = actor(pipe, buf, sd);
 815                 if (ret <= 0)
 816                         return ret;
 817
 818                 buf->offset += ret;
 819                 buf->len -= ret;
 820
 821                 sd->num_spliced += ret;
 822                 sd->len -= ret;
 823                 sd->pos += ret;
 824                 sd->total_len -= ret;
 825
 826                 if (!buf->len) {
 827                         buf->ops = NULL;
 828                         ops->release(pipe, buf);
 829                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 830                         pipe->nrbufs--;
 831                         if (pipe->inode)
 832                                 sd->need_wakeup = true;
 833                 }
 834
 835                 if (!sd->total_len)
 836                         return 0;
 837         }
 838
 839         return 1;
 840 }
 841 EXPORT_SYMBOL(splice_from_pipe_feed);
 842
 843 /**
 844  * splice_from_pipe_next - wait for some data to splice from
 845  * @pipe:       pipe to splice from
 846  * @sd:         information about the splice operation
 847  *
 848  * Description:
 849  *    This function will wait for some data and return a positive
 850  *    value (one) if pipe buffers are available.  It will return zero
 851  *    or -errno if no more data needs to be spliced.
 852  */
 853 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 854 {
 855         while (!pipe->nrbufs) {
 856                 if (!pipe->writers)
 857                         return 0;
 858
 859                 if (!pipe->waiting_writers && sd->num_spliced)
 860                         return 0;
 861
 862                 if (sd->flags & SPLICE_F_NONBLOCK)
 863                         return -EAGAIN;
 864
 865                 if (signal_pending(current))
 866                         return -ERESTARTSYS;
 867
 868                 if (sd->need_wakeup) {
 869                         wakeup_pipe_writers(pipe);
 870                         sd->need_wakeup = false;
 871                 }
 872
 873                 pipe_wait(pipe);
 874         }
 875
 876         return 1;
 877 }
 878 EXPORT_SYMBOL(splice_from_pipe_next);
 879
 880 /**
 881  * splice_from_pipe_begin - start splicing from pipe
 882  * @sd:         information about the splice operation
 883  *
 884  * Description:
 885  *    This function should be called before a loop containing
 886  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 887  *    initialize the necessary fields of @sd.
 888  */
 889 void splice_from_pipe_begin(struct splice_desc *sd)
 890 {
 891         sd->num_spliced = 0;
 892         sd->need_wakeup = false;
 893 }
 894 EXPORT_SYMBOL(splice_from_pipe_begin);
 895
 896 /**
 897  * splice_from_pipe_end - finish splicing from pipe
 898  * @pipe:       pipe to splice from
 899  * @sd:         information about the splice operation
 900  *
 901  * Description:
 902  *    This function will wake up pipe writers if necessary.  It should
 903  *    be called after a loop containing splice_from_pipe_next() and
 904  *    splice_from_pipe_feed().
 905  */
 906 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 907 {
 908         if (sd->need_wakeup)
 909                 wakeup_pipe_writers(pipe);
 910 }
 911 EXPORT_SYMBOL(splice_from_pipe_end);
 912
 913 /**
 914  * __splice_from_pipe - splice data from a pipe to given actor
 915  * @pipe:       pipe to splice from
 916  * @sd:         information to @actor
 917  * @actor:      handler that splices the data
 918  *
 919  * Description:
 920  *    This function does little more than loop over the pipe and call
 921  *    @actor to do the actual moving of a single struct pipe_buffer to
 922  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 923  *    pipe_to_user.
 924  *
 925  */
 926 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 927                            splice_actor *actor)
 928 {
 929         int ret;
 930
 931         splice_from_pipe_begin(sd);
 932         do {
 933                 ret = splice_from_pipe_next(pipe, sd);
 934                 if (ret > 0)
 935                         ret = splice_from_pipe_feed(pipe, sd, actor);
 936         } while (ret > 0);
 937         splice_from_pipe_end(pipe, sd);
 938
 939         return sd->num_spliced ? sd->num_spliced : ret;
 940 }
 941 EXPORT_SYMBOL(__splice_from_pipe);
 942
 943 /**
 944  * splice_from_pipe - splice data from a pipe to a file
 945  * @pipe:       pipe to splice from
 946  * @out:        file to splice to
 947  * @ppos:       position in @out
 948  * @len:        how many bytes to splice
 949  * @flags:      splice modifier flags
 950  * @actor:      handler that splices the data
 951  *
 952  * Description:
 953  *    See __splice_from_pipe. This function locks the pipe inode,
 954  *    otherwise it's identical to __splice_from_pipe().
 955  *
 956  */
 957 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 958                          loff_t *ppos, size_t len, unsigned int flags,
 959                          splice_actor *actor)
 960 {
 961         ssize_t ret;
 962         struct splice_desc sd = {
 963                 .total_len = len,
 964                 .flags = flags,
 965                 .pos = *ppos,
 966                 .u.file = out,
 967         };
 968
 969         pipe_lock(pipe);
 970         ret = __splice_from_pipe(pipe, &sd, actor);
 971         pipe_unlock(pipe);
 972
 973         return ret;
 974 }
 975
 976 /**
 977  * generic_file_splice_write - splice data from a pipe to a file
 978  * @pipe:       pipe info
 979  * @out:        file to write to
 980  * @ppos:       position in @out
 981  * @len:        number of bytes to splice
 982  * @flags:      splice modifier flags
 983  *
 984  * Description:
 985  *    Will either move or copy pages (determined by @flags options) from
 986  *    the given pipe inode to the given file.
 987  *
 988  */
 989 ssize_t
 990 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 991                           loff_t *ppos, size_t len, unsigned int flags)
 992 {
 993         struct address_space *mapping = out->f_mapping;
 994         struct inode *inode = mapping->host;
 995         struct splice_desc sd = {
 996                 .total_len = len,
 997                 .flags = flags,
 998                 .pos = *ppos,
 999                 .u.file = out,
1000         };
1001         ssize_t ret;
1002
1003         pipe_lock(pipe);
1004
1005         splice_from_pipe_begin(&sd);
1006         do {
1007                 ret = splice_from_pipe_next(pipe, &sd);
1008                 if (ret <= 0)
1009                         break;
1010
1011                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1012                 ret = file_remove_suid(out);
1013                 if (!ret) {
1014                         file_update_time(out);
1015                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1016                 }
1017                 mutex_unlock(&inode->i_mutex);
1018         } while (ret > 0);
1019         splice_from_pipe_end(pipe, &sd);
1020
1021         pipe_unlock(pipe);
1022
1023         if (sd.num_spliced)
1024                 ret = sd.num_spliced;
1025
1026         if (ret > 0) {
1027                 unsigned long nr_pages;
1028                 int err;
1029
1030                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1031
1032                 err = generic_write_sync(out, *ppos, ret);
1033                 if (err)
1034                         ret = err;
1035                 else
1036                         *ppos += ret;
1037                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1038         }
1039
1040         return ret;
1041 }
1042
1043 EXPORT_SYMBOL(generic_file_splice_write);
1044
1045 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1046                           struct splice_desc *sd)
1047 {
1048         int ret;
1049         void *data;
1050
1051         data = buf->ops->map(pipe, buf, 0);
1052         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1053         buf->ops->unmap(pipe, buf, data);
1054
1055         return ret;
1056 }
1057
1058 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1059                                          struct file *out, loff_t *ppos,
1060                                          size_t len, unsigned int flags)
1061 {
1062         ssize_t ret;
1063
1064         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1065         if (ret > 0)
1066                 *ppos += ret;
1067
1068         return ret;
1069 }
1070
1071 /**
1072  * generic_splice_sendpage - splice data from a pipe to a socket
1073  * @pipe:       pipe to splice from
1074  * @out:        socket to write to
1075  * @ppos:       position in @out
1076  * @len:        number of bytes to splice
1077  * @flags:      splice modifier flags
1078  *
1079  * Description:
1080  *    Will send @len bytes from the pipe to a network socket. No data copying
1081  *    is involved.
1082  *
1083  */
1084 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1085                                 loff_t *ppos, size_t len, unsigned int flags)
1086 {
1087         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1088 }
1089
1090 EXPORT_SYMBOL(generic_splice_sendpage);
1091
1092 /*
1093  * Attempt to initiate a splice from pipe to file.
1094  */
1095 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1096                            loff_t *ppos, size_t len, unsigned int flags)
1097 {
1098         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1099                                 loff_t *, size_t, unsigned int);
1100         int ret;
1101
1102         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1103                 return -EBADF;
1104
1105         if (unlikely(out->f_flags & O_APPEND))
1106                 return -EINVAL;
1107
1108         ret = rw_verify_area(WRITE, out, ppos, len);
1109         if (unlikely(ret < 0))
1110                 return ret;
1111
1112         if (out->f_op && out->f_op->splice_write)
1113                 splice_write = out->f_op->splice_write;
1114         else
1115                 splice_write = default_file_splice_write;
1116
1117         return splice_write(pipe, out, ppos, len, flags);
1118 }
1119
1120 /*
1121  * Attempt to initiate a splice from a file to a pipe.
1122  */
1123 static long do_splice_to(struct file *in, loff_t *ppos,
1124                          struct pipe_inode_info *pipe, size_t len,
1125                          unsigned int flags)
1126 {
1127         ssize_t (*splice_read)(struct file *, loff_t *,
1128                                struct pipe_inode_info *, size_t, unsigned int);
1129         int ret;
1130
1131         if (unlikely(!(in->f_mode & FMODE_READ)))
1132                 return -EBADF;
1133
1134         ret = rw_verify_area(READ, in, ppos, len);
1135         if (unlikely(ret < 0))
1136                 return ret;
1137
1138         if (in->f_op && in->f_op->splice_read)
1139                 splice_read = in->f_op->splice_read;
1140         else
1141                 splice_read = default_file_splice_read;
1142
1143         return splice_read(in, ppos, pipe, len, flags);
1144 }
1145
1146 /**
1147  * splice_direct_to_actor - splices data directly between two non-pipes
1148  * @in:         file to splice from
1149  * @sd:         actor information on where to splice to
1150  * @actor:      handles the data splicing
1151  *
1152  * Description:
1153  *    This is a special case helper to splice directly between two
1154  *    points, without requiring an explicit pipe. Internally an allocated
1155  *    pipe is cached in the process, and reused during the lifetime of
1156  *    that process.
1157  *
1158  */
1159 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1160                                splice_direct_actor *actor)
1161 {
1162         struct pipe_inode_info *pipe;
1163         long ret, bytes;
1164         umode_t i_mode;
1165         size_t len;
1166         int i, flags;
1167
1168         /*
1169          * We require the input being a regular file, as we don't want to
1170          * randomly drop data for eg socket -> socket splicing. Use the
1171          * piped splicing for that!
1172          */
1173         i_mode = in->f_path.dentry->d_inode->i_mode;
1174         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1175                 return -EINVAL;
1176
1177         /*
1178          * neither in nor out is a pipe, setup an internal pipe attached to
1179          * 'out' and transfer the wanted data from 'in' to 'out' through that
1180          */
1181         pipe = current->splice_pipe;
1182         if (unlikely(!pipe)) {
1183                 pipe = alloc_pipe_info(NULL);
1184                 if (!pipe)
1185                         return -ENOMEM;
1186
1187                 /*
1188                  * We don't have an immediate reader, but we'll read the stuff
1189                  * out of the pipe right after the splice_to_pipe(). So set
1190                  * PIPE_READERS appropriately.
1191                  */
1192                 pipe->readers = 1;
1193
1194                 current->splice_pipe = pipe;
1195         }
1196
1197         /*
1198          * Do the splice.
1199          */
1200         ret = 0;
1201         bytes = 0;
1202         len = sd->total_len;
1203         flags = sd->flags;
1204
1205         /*
1206          * Don't block on output, we have to drain the direct pipe.
1207          */
1208         sd->flags &= ~SPLICE_F_NONBLOCK;
1209
1210         while (len) {
1211                 size_t read_len;
1212                 loff_t pos = sd->pos, prev_pos = pos;
1213
1214                 ret = do_splice_to(in, &pos, pipe, len, flags);
1215                 if (unlikely(ret <= 0))
1216                         goto out_release;
1217
1218                 read_len = ret;
1219                 sd->total_len = read_len;
1220
1221                 /*
1222                  * NOTE: nonblocking mode only applies to the input. We
1223                  * must not do the output in nonblocking mode as then we
1224                  * could get stuck data in the internal pipe:
1225                  */
1226                 ret = actor(pipe, sd);
1227                 if (unlikely(ret <= 0)) {
1228                         sd->pos = prev_pos;
1229                         goto out_release;
1230                 }
1231
1232                 bytes += ret;
1233                 len -= ret;
1234                 sd->pos = pos;
1235
1236                 if (ret < read_len) {
1237                         sd->pos = prev_pos + ret;
1238                         goto out_release;
1239                 }
1240         }
1241
1242 done:
1243         pipe->nrbufs = pipe->curbuf = 0;
1244         file_accessed(in);
1245         return bytes;
1246
1247 out_release:
1248         /*
1249          * If we did an incomplete transfer we must release
1250          * the pipe buffers in question:
1251          */
1252         for (i = 0; i < pipe->buffers; i++) {
1253                 struct pipe_buffer *buf = pipe->bufs + i;
1254
1255                 if (buf->ops) {
1256                         buf->ops->release(pipe, buf);
1257                         buf->ops = NULL;
1258                 }
1259         }
1260
1261         if (!bytes)
1262                 bytes = ret;
1263
1264         goto done;
1265 }
1266 EXPORT_SYMBOL(splice_direct_to_actor);
1267
1268 static int direct_splice_actor(struct pipe_inode_info *pipe,
1269                                struct splice_desc *sd)
1270 {
1271         struct file *file = sd->u.file;
1272
1273         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1274                               sd->flags);
1275 }
1276
1277 /**
1278  * do_splice_direct - splices data directly between two files
1279  * @in:         file to splice from
1280  * @ppos:       input file offset
1281  * @out:        file to splice to
1282  * @len:        number of bytes to splice
1283  * @flags:      splice modifier flags
1284  *
1285  * Description:
1286  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1287  *    doing it in the application would incur an extra system call
1288  *    (splice in + splice out, as compared to just sendfile()). So this helper
1289  *    can splice directly through a process-private pipe.
1290  *
1291  */
1292 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1293                       size_t len, unsigned int flags)
1294 {
1295         struct splice_desc sd = {
1296                 .len            = len,
1297                 .total_len      = len,
1298                 .flags          = flags,
1299                 .pos            = *ppos,
1300                 .u.file         = out,
1301         };
1302         long ret;
1303
1304         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1305         if (ret > 0)
1306                 *ppos = sd.pos;
1307
1308         return ret;
1309 }
1310
1311 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1312                                struct pipe_inode_info *opipe,
1313                                size_t len, unsigned int flags);
1314
1315 /*
1316  * Determine where to splice to/from.
1317  */
1318 static long do_splice(struct file *in, loff_t __user *off_in,
1319                       struct file *out, loff_t __user *off_out,
1320                       size_t len, unsigned int flags)
1321 {
1322         struct pipe_inode_info *ipipe;
1323         struct pipe_inode_info *opipe;
1324         loff_t offset, *off;
1325         long ret;
1326
1327         ipipe = get_pipe_info(in);
1328         opipe = get_pipe_info(out);
1329
1330         if (ipipe && opipe) {
1331                 if (off_in || off_out)
1332                         return -ESPIPE;
1333
1334                 if (!(in->f_mode & FMODE_READ))
1335                         return -EBADF;
1336
1337                 if (!(out->f_mode & FMODE_WRITE))
1338                         return -EBADF;
1339
1340                 /* Splicing to self would be fun, but... */
1341                 if (ipipe == opipe)
1342                         return -EINVAL;
1343
1344                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1345         }
1346
1347         if (ipipe) {
1348                 if (off_in)
1349                         return -ESPIPE;
1350                 if (off_out) {
1351                         if (!(out->f_mode & FMODE_PWRITE))
1352                                 return -EINVAL;
1353                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1354                                 return -EFAULT;
1355                         off = &offset;
1356                 } else
1357                         off = &out->f_pos;
1358
1359                 ret = do_splice_from(ipipe, out, off, len, flags);
1360
1361                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1362                         ret = -EFAULT;
1363
1364                 return ret;
1365         }
1366
1367         if (opipe) {
1368                 if (off_out)
1369                         return -ESPIPE;
1370                 if (off_in) {
1371                         if (!(in->f_mode & FMODE_PREAD))
1372                                 return -EINVAL;
1373                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1374                                 return -EFAULT;
1375                         off = &offset;
1376                 } else
1377                         off = &in->f_pos;
1378
1379                 ret = do_splice_to(in, off, opipe, len, flags);
1380
1381                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1382                         ret = -EFAULT;
1383
1384                 return ret;
1385         }
1386
1387         return -EINVAL;
1388 }
1389
1390 /*
1391  * Map an iov into an array of pages and offset/length tupples. With the
1392  * partial_page structure, we can map several non-contiguous ranges into
1393  * our ones pages[] map instead of splitting that operation into pieces.
1394  * Could easily be exported as a generic helper for other users, in which
1395  * case one would probably want to add a 'max_nr_pages' parameter as well.
1396  */
1397 static int get_iovec_page_array(const struct iovec __user *iov,
1398                                 unsigned int nr_vecs, struct page **pages,
1399                                 struct partial_page *partial, int aligned,
1400                                 unsigned int pipe_buffers)
1401 {
1402         int buffers = 0, error = 0;
1403
1404         while (nr_vecs) {
1405                 unsigned long off, npages;
1406                 struct iovec entry;
1407                 void __user *base;
1408                 size_t len;
1409                 int i;
1410
1411                 error = -EFAULT;
1412                 if (copy_from_user(&entry, iov, sizeof(entry)))
1413                         break;
1414
1415                 base = entry.iov_base;
1416                 len = entry.iov_len;
1417
1418                 /*
1419                  * Sanity check this iovec. 0 read succeeds.
1420                  */
1421                 error = 0;
1422                 if (unlikely(!len))
1423                         break;
1424                 error = -EFAULT;
1425                 if (!access_ok(VERIFY_READ, base, len))
1426                         break;
1427
1428                 /*
1429                  * Get this base offset and number of pages, then map
1430                  * in the user pages.
1431                  */
1432                 off = (unsigned long) base & ~PAGE_MASK;
1433
1434                 /*
1435                  * If asked for alignment, the offset must be zero and the
1436                  * length a multiple of the PAGE_SIZE.
1437                  */
1438                 error = -EINVAL;
1439                 if (aligned && (off || len & ~PAGE_MASK))
1440                         break;
1441
1442                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1443                 if (npages > pipe_buffers - buffers)
1444                         npages = pipe_buffers - buffers;
1445
1446                 error = get_user_pages_fast((unsigned long)base, npages,
1447                                         0, &pages[buffers]);
1448
1449                 if (unlikely(error <= 0))
1450                         break;
1451
1452                 /*
1453                  * Fill this contiguous range into the partial page map.
1454                  */
1455                 for (i = 0; i < error; i++) {
1456                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1457
1458                         partial[buffers].offset = off;
1459                         partial[buffers].len = plen;
1460
1461                         off = 0;
1462                         len -= plen;
1463                         buffers++;
1464                 }
1465
1466                 /*
1467                  * We didn't complete this iov, stop here since it probably
1468                  * means we have to move some of this into a pipe to
1469                  * be able to continue.
1470                  */
1471                 if (len)
1472                         break;
1473
1474                 /*
1475                  * Don't continue if we mapped fewer pages than we asked for,
1476                  * or if we mapped the max number of pages that we have
1477                  * room for.
1478                  */
1479                 if (error < npages || buffers == pipe_buffers)
1480                         break;
1481
1482                 nr_vecs--;
1483                 iov++;
1484         }
1485
1486         if (buffers)
1487                 return buffers;
1488
1489         return error;
1490 }
1491
1492 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1493                         struct splice_desc *sd)
1494 {
1495         char *src;
1496         int ret;
1497
1498         /*
1499          * See if we can use the atomic maps, by prefaulting in the
1500          * pages and doing an atomic copy
1501          */
1502         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1503                 src = buf->ops->map(pipe, buf, 1);
1504                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1505                                                         sd->len);
1506                 buf->ops->unmap(pipe, buf, src);
1507                 if (!ret) {
1508                         ret = sd->len;
1509                         goto out;
1510                 }
1511         }
1512
1513         /*
1514          * No dice, use slow non-atomic map and copy
1515          */
1516         src = buf->ops->map(pipe, buf, 0);
1517
1518         ret = sd->len;
1519         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1520                 ret = -EFAULT;
1521
1522         buf->ops->unmap(pipe, buf, src);
1523 out:
1524         if (ret > 0)
1525                 sd->u.userptr += ret;
1526         return ret;
1527 }
1528
1529 /*
1530  * For lack of a better implementation, implement vmsplice() to userspace
1531  * as a simple copy of the pipes pages to the user iov.
1532  */
1533 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1534                              unsigned long nr_segs, unsigned int flags)
1535 {
1536         struct pipe_inode_info *pipe;
1537         struct splice_desc sd;
1538         ssize_t size;
1539         int error;
1540         long ret;
1541
1542         pipe = get_pipe_info(file);
1543         if (!pipe)
1544                 return -EBADF;
1545
1546         pipe_lock(pipe);
1547
1548         error = ret = 0;
1549         while (nr_segs) {
1550                 void __user *base;
1551                 size_t len;
1552
1553                 /*
1554                  * Get user address base and length for this iovec.
1555                  */
1556                 error = get_user(base, &iov->iov_base);
1557                 if (unlikely(error))
1558                         break;
1559                 error = get_user(len, &iov->iov_len);
1560                 if (unlikely(error))
1561                         break;
1562
1563                 /*
1564                  * Sanity check this iovec. 0 read succeeds.
1565                  */
1566                 if (unlikely(!len))
1567                         break;
1568                 if (unlikely(!base)) {
1569                         error = -EFAULT;
1570                         break;
1571                 }
1572
1573                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1574                         error = -EFAULT;
1575                         break;
1576                 }
1577
1578                 sd.len = 0;
1579                 sd.total_len = len;
1580                 sd.flags = flags;
1581                 sd.u.userptr = base;
1582                 sd.pos = 0;
1583
1584                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1585                 if (size < 0) {
1586                         if (!ret)
1587                                 ret = size;
1588
1589                         break;
1590                 }
1591
1592                 ret += size;
1593
1594                 if (size < len)
1595                         break;
1596
1597                 nr_segs--;
1598                 iov++;
1599         }
1600
1601         pipe_unlock(pipe);
1602
1603         if (!ret)
1604                 ret = error;
1605
1606         return ret;
1607 }
1608
1609 /*
1610  * vmsplice splices a user address range into a pipe. It can be thought of
1611  * as splice-from-memory, where the regular splice is splice-from-file (or
1612  * to file). In both cases the output is a pipe, naturally.
1613  */
1614 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1615                              unsigned long nr_segs, unsigned int flags)
1616 {
1617         struct pipe_inode_info *pipe;
1618         struct page *pages[PIPE_DEF_BUFFERS];
1619         struct partial_page partial[PIPE_DEF_BUFFERS];
1620         struct splice_pipe_desc spd = {
1621                 .pages = pages,
1622                 .partial = partial,
1623                 .nr_pages_max = PIPE_DEF_BUFFERS,
1624                 .flags = flags,
1625                 .ops = &user_page_pipe_buf_ops,
1626                 .spd_release = spd_release_page,
1627         };
1628         long ret;
1629
1630         pipe = get_pipe_info(file);
1631         if (!pipe)
1632                 return -EBADF;
1633
1634         if (splice_grow_spd(pipe, &spd))
1635                 return -ENOMEM;
1636
1637         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1638                                             spd.partial, flags & SPLICE_F_GIFT,
1639                                             spd.nr_pages_max);
1640         if (spd.nr_pages <= 0)
1641                 ret = spd.nr_pages;
1642         else
1643                 ret = splice_to_pipe(pipe, &spd);
1644
1645         splice_shrink_spd(&spd);
1646         return ret;
1647 }
1648
1649 /*
1650  * Note that vmsplice only really supports true splicing _from_ user memory
1651  * to a pipe, not the other way around. Splicing from user memory is a simple
1652  * operation that can be supported without any funky alignment restrictions
1653  * or nasty vm tricks. We simply map in the user memory and fill them into
1654  * a pipe. The reverse isn't quite as easy, though. There are two possible
1655  * solutions for that:
1656  *
1657  *      - memcpy() the data internally, at which point we might as well just
1658  *        do a regular read() on the buffer anyway.
1659  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1660  *        has restriction limitations on both ends of the pipe).
1661  *
1662  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1663  *
1664  */
1665 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1666                 unsigned long, nr_segs, unsigned int, flags)
1667 {
1668         struct file *file;
1669         long error;
1670         int fput;
1671
1672         if (unlikely(nr_segs > UIO_MAXIOV))
1673                 return -EINVAL;
1674         else if (unlikely(!nr_segs))
1675                 return 0;
1676
1677         error = -EBADF;
1678         file = fget_light(fd, &fput);
1679         if (file) {
1680                 if (file->f_mode & FMODE_WRITE)
1681                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1682                 else if (file->f_mode & FMODE_READ)
1683                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1684
1685                 fput_light(file, fput);
1686         }
1687
1688         return error;
1689 }
1690
1691 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1692                 int, fd_out, loff_t __user *, off_out,
1693                 size_t, len, unsigned int, flags)
1694 {
1695         long error;
1696         struct file *in, *out;
1697         int fput_in, fput_out;
1698
1699         if (unlikely(!len))
1700                 return 0;
1701
1702         error = -EBADF;
1703         in = fget_light(fd_in, &fput_in);
1704         if (in) {
1705                 if (in->f_mode & FMODE_READ) {
1706                         out = fget_light(fd_out, &fput_out);
1707                         if (out) {
1708                                 if (out->f_mode & FMODE_WRITE)
1709                                         error = do_splice(in, off_in,
1710                                                           out, off_out,
1711                                                           len, flags);
1712                                 fput_light(out, fput_out);
1713                         }
1714                 }
1715
1716                 fput_light(in, fput_in);
1717         }
1718
1719         return error;
1720 }
1721
1722 /*
1723  * Make sure there's data to read. Wait for input if we can, otherwise
1724  * return an appropriate error.
1725  */
1726 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1727 {
1728         int ret;
1729
1730         /*
1731          * Check ->nrbufs without the inode lock first. This function
1732          * is speculative anyways, so missing one is ok.
1733          */
1734         if (pipe->nrbufs)
1735                 return 0;
1736
1737         ret = 0;
1738         pipe_lock(pipe);
1739
1740         while (!pipe->nrbufs) {
1741                 if (signal_pending(current)) {
1742                         ret = -ERESTARTSYS;
1743                         break;
1744                 }
1745                 if (!pipe->writers)
1746                         break;
1747                 if (!pipe->waiting_writers) {
1748                         if (flags & SPLICE_F_NONBLOCK) {
1749                                 ret = -EAGAIN;
1750                                 break;
1751                         }
1752                 }
1753                 pipe_wait(pipe);
1754         }
1755
1756         pipe_unlock(pipe);
1757         return ret;
1758 }
1759
1760 /*
1761  * Make sure there's writeable room. Wait for room if we can, otherwise
1762  * return an appropriate error.
1763  */
1764 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1765 {
1766         int ret;
1767
1768         /*
1769          * Check ->nrbufs without the inode lock first. This function
1770          * is speculative anyways, so missing one is ok.
1771          */
1772         if (pipe->nrbufs < pipe->buffers)
1773                 return 0;
1774
1775         ret = 0;
1776         pipe_lock(pipe);
1777
1778         while (pipe->nrbufs >= pipe->buffers) {
1779                 if (!pipe->readers) {
1780                         send_sig(SIGPIPE, current, 0);
1781                         ret = -EPIPE;
1782                         break;
1783                 }
1784                 if (flags & SPLICE_F_NONBLOCK) {
1785                         ret = -EAGAIN;
1786                         break;
1787                 }
1788                 if (signal_pending(current)) {
1789                         ret = -ERESTARTSYS;
1790                         break;
1791                 }
1792                 pipe->waiting_writers++;
1793                 pipe_wait(pipe);
1794                 pipe->waiting_writers--;
1795         }
1796
1797         pipe_unlock(pipe);
1798         return ret;
1799 }
1800
1801 /*
1802  * Splice contents of ipipe to opipe.
1803  */
1804 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1805                                struct pipe_inode_info *opipe,
1806                                size_t len, unsigned int flags)
1807 {
1808         struct pipe_buffer *ibuf, *obuf;
1809         int ret = 0, nbuf;
1810         bool input_wakeup = false;
1811
1812
1813 retry:
1814         ret = ipipe_prep(ipipe, flags);
1815         if (ret)
1816                 return ret;
1817
1818         ret = opipe_prep(opipe, flags);
1819         if (ret)
1820                 return ret;
1821
1822         /*
1823          * Potential ABBA deadlock, work around it by ordering lock
1824          * grabbing by pipe info address. Otherwise two different processes
1825          * could deadlock (one doing tee from A -> B, the other from B -> A).
1826          */
1827         pipe_double_lock(ipipe, opipe);
1828
1829         do {
1830                 if (!opipe->readers) {
1831                         send_sig(SIGPIPE, current, 0);
1832                         if (!ret)
1833                                 ret = -EPIPE;
1834                         break;
1835                 }
1836
1837                 if (!ipipe->nrbufs && !ipipe->writers)
1838                         break;
1839
1840                 /*
1841                  * Cannot make any progress, because either the input
1842                  * pipe is empty or the output pipe is full.
1843                  */
1844                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1845                         /* Already processed some buffers, break */
1846                         if (ret)
1847                                 break;
1848
1849                         if (flags & SPLICE_F_NONBLOCK) {
1850                                 ret = -EAGAIN;
1851                                 break;
1852                         }
1853
1854                         /*
1855                          * We raced with another reader/writer and haven't
1856                          * managed to process any buffers.  A zero return
1857                          * value means EOF, so retry instead.
1858                          */
1859                         pipe_unlock(ipipe);
1860                         pipe_unlock(opipe);
1861                         goto retry;
1862                 }
1863
1864                 ibuf = ipipe->bufs + ipipe->curbuf;
1865                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1866                 obuf = opipe->bufs + nbuf;
1867
1868                 if (len >= ibuf->len) {
1869                         /*
1870                          * Simply move the whole buffer from ipipe to opipe
1871                          */
1872                         *obuf = *ibuf;
1873                         ibuf->ops = NULL;
1874                         opipe->nrbufs++;
1875                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1876                         ipipe->nrbufs--;
1877                         input_wakeup = true;
1878                 } else {
1879                         /*
1880                          * Get a reference to this pipe buffer,
1881                          * so we can copy the contents over.
1882                          */
1883                         ibuf->ops->get(ipipe, ibuf);
1884                         *obuf = *ibuf;
1885
1886                         /*
1887                          * Don't inherit the gift flag, we need to
1888                          * prevent multiple steals of this page.
1889                          */
1890                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1891
1892                         obuf->len = len;
1893                         opipe->nrbufs++;
1894                         ibuf->offset += obuf->len;
1895                         ibuf->len -= obuf->len;
1896                 }
1897                 ret += obuf->len;
1898                 len -= obuf->len;
1899         } while (len);
1900
1901         pipe_unlock(ipipe);
1902         pipe_unlock(opipe);
1903
1904         /*
1905          * If we put data in the output pipe, wakeup any potential readers.
1906          */
1907         if (ret > 0)
1908                 wakeup_pipe_readers(opipe);
1909
1910         if (input_wakeup)
1911                 wakeup_pipe_writers(ipipe);
1912
1913         return ret;
1914 }
1915
1916 /*
1917  * Link contents of ipipe to opipe.
1918  */
1919 static int link_pipe(struct pipe_inode_info *ipipe,
1920                      struct pipe_inode_info *opipe,
1921                      size_t len, unsigned int flags)
1922 {
1923         struct pipe_buffer *ibuf, *obuf;
1924         int ret = 0, i = 0, nbuf;
1925
1926         /*
1927          * Potential ABBA deadlock, work around it by ordering lock
1928          * grabbing by pipe info address. Otherwise two different processes
1929          * could deadlock (one doing tee from A -> B, the other from B -> A).
1930          */
1931         pipe_double_lock(ipipe, opipe);
1932
1933         do {
1934                 if (!opipe->readers) {
1935                         send_sig(SIGPIPE, current, 0);
1936                         if (!ret)
1937                                 ret = -EPIPE;
1938                         break;
1939                 }
1940
1941                 /*
1942                  * If we have iterated all input buffers or ran out of
1943                  * output room, break.
1944                  */
1945                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1946                         break;
1947
1948                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1949                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1950
1951                 /*
1952                  * Get a reference to this pipe buffer,
1953                  * so we can copy the contents over.
1954                  */
1955                 ibuf->ops->get(ipipe, ibuf);
1956
1957                 obuf = opipe->bufs + nbuf;
1958                 *obuf = *ibuf;
1959
1960                 /*
1961                  * Don't inherit the gift flag, we need to
1962                  * prevent multiple steals of this page.
1963                  */
1964                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1965
1966                 if (obuf->len > len)
1967                         obuf->len = len;
1968
1969                 opipe->nrbufs++;
1970                 ret += obuf->len;
1971                 len -= obuf->len;
1972                 i++;
1973         } while (len);
1974
1975         /*
1976          * return EAGAIN if we have the potential of some data in the
1977          * future, otherwise just return 0
1978          */
1979         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1980                 ret = -EAGAIN;
1981
1982         pipe_unlock(ipipe);
1983         pipe_unlock(opipe);
1984
1985         /*
1986          * If we put data in the output pipe, wakeup any potential readers.
1987          */
1988         if (ret > 0)
1989                 wakeup_pipe_readers(opipe);
1990
1991         return ret;
1992 }
1993
1994 /*
1995  * This is a tee(1) implementation that works on pipes. It doesn't copy
1996  * any data, it simply references the 'in' pages on the 'out' pipe.
1997  * The 'flags' used are the SPLICE_F_* variants, currently the only
1998  * applicable one is SPLICE_F_NONBLOCK.
1999  */
2000 static long do_tee(struct file *in, struct file *out, size_t len,
2001                    unsigned int flags)
2002 {
2003         struct pipe_inode_info *ipipe = get_pipe_info(in);
2004         struct pipe_inode_info *opipe = get_pipe_info(out);
2005         int ret = -EINVAL;
2006
2007         /*
2008          * Duplicate the contents of ipipe to opipe without actually
2009          * copying the data.
2010          */
2011         if (ipipe && opipe && ipipe != opipe) {
2012                 /*
2013                  * Keep going, unless we encounter an error. The ipipe/opipe
2014                  * ordering doesn't really matter.
2015                  */
2016                 ret = ipipe_prep(ipipe, flags);
2017                 if (!ret) {
2018                         ret = opipe_prep(opipe, flags);
2019                         if (!ret)
2020                                 ret = link_pipe(ipipe, opipe, len, flags);
2021                 }
2022         }
2023
2024         return ret;
2025 }
2026
2027 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2028 {
2029         struct file *in;
2030         int error, fput_in;
2031
2032         if (unlikely(!len))
2033                 return 0;
2034
2035         error = -EBADF;
2036         in = fget_light(fdin, &fput_in);
2037         if (in) {
2038                 if (in->f_mode & FMODE_READ) {
2039                         int fput_out;
2040                         struct file *out = fget_light(fdout, &fput_out);
2041
2042                         if (out) {
2043                                 if (out->f_mode & FMODE_WRITE)
2044                                         error = do_tee(in, out, len, flags);
2045                                 fput_light(out, fput_out);
2046                         }
2047                 }
2048                 fput_light(in, fput_in);
2049         }
2050
2051         return error;
2052 }