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 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 EXPORT_SYMBOL(do_splice_from);
1120
1121 /*
1122  * Attempt to initiate a splice from a file to a pipe.
1123  */
1124 long do_splice_to(struct file *in, loff_t *ppos,
1125                   struct pipe_inode_info *pipe, size_t len,
1126                   unsigned int flags)
1127 {
1128         ssize_t (*splice_read)(struct file *, loff_t *,
1129                                struct pipe_inode_info *, size_t, unsigned int);
1130         int ret;
1131
1132         if (unlikely(!(in->f_mode & FMODE_READ)))
1133                 return -EBADF;
1134
1135         ret = rw_verify_area(READ, in, ppos, len);
1136         if (unlikely(ret < 0))
1137                 return ret;
1138
1139         if (in->f_op && in->f_op->splice_read)
1140                 splice_read = in->f_op->splice_read;
1141         else
1142                 splice_read = default_file_splice_read;
1143
1144         return splice_read(in, ppos, pipe, len, flags);
1145 }
1146 EXPORT_SYMBOL(do_splice_to);
1147
1148 /**
1149  * splice_direct_to_actor - splices data directly between two non-pipes
1150  * @in:         file to splice from
1151  * @sd:         actor information on where to splice to
1152  * @actor:      handles the data splicing
1153  *
1154  * Description:
1155  *    This is a special case helper to splice directly between two
1156  *    points, without requiring an explicit pipe. Internally an allocated
1157  *    pipe is cached in the process, and reused during the lifetime of
1158  *    that process.
1159  *
1160  */
1161 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1162                                splice_direct_actor *actor)
1163 {
1164         struct pipe_inode_info *pipe;
1165         long ret, bytes;
1166         umode_t i_mode;
1167         size_t len;
1168         int i, flags;
1169
1170         /*
1171          * We require the input being a regular file, as we don't want to
1172          * randomly drop data for eg socket -> socket splicing. Use the
1173          * piped splicing for that!
1174          */
1175         i_mode = in->f_path.dentry->d_inode->i_mode;
1176         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1177                 return -EINVAL;
1178
1179         /*
1180          * neither in nor out is a pipe, setup an internal pipe attached to
1181          * 'out' and transfer the wanted data from 'in' to 'out' through that
1182          */
1183         pipe = current->splice_pipe;
1184         if (unlikely(!pipe)) {
1185                 pipe = alloc_pipe_info(NULL);
1186                 if (!pipe)
1187                         return -ENOMEM;
1188
1189                 /*
1190                  * We don't have an immediate reader, but we'll read the stuff
1191                  * out of the pipe right after the splice_to_pipe(). So set
1192                  * PIPE_READERS appropriately.
1193                  */
1194                 pipe->readers = 1;
1195
1196                 current->splice_pipe = pipe;
1197         }
1198
1199         /*
1200          * Do the splice.
1201          */
1202         ret = 0;
1203         bytes = 0;
1204         len = sd->total_len;
1205         flags = sd->flags;
1206
1207         /*
1208          * Don't block on output, we have to drain the direct pipe.
1209          */
1210         sd->flags &= ~SPLICE_F_NONBLOCK;
1211
1212         while (len) {
1213                 size_t read_len;
1214                 loff_t pos = sd->pos, prev_pos = pos;
1215
1216                 ret = do_splice_to(in, &pos, pipe, len, flags);
1217                 if (unlikely(ret <= 0))
1218                         goto out_release;
1219
1220                 read_len = ret;
1221                 sd->total_len = read_len;
1222
1223                 /*
1224                  * NOTE: nonblocking mode only applies to the input. We
1225                  * must not do the output in nonblocking mode as then we
1226                  * could get stuck data in the internal pipe:
1227                  */
1228                 ret = actor(pipe, sd);
1229                 if (unlikely(ret <= 0)) {
1230                         sd->pos = prev_pos;
1231                         goto out_release;
1232                 }
1233
1234                 bytes += ret;
1235                 len -= ret;
1236                 sd->pos = pos;
1237
1238                 if (ret < read_len) {
1239                         sd->pos = prev_pos + ret;
1240                         goto out_release;
1241                 }
1242         }
1243
1244 done:
1245         pipe->nrbufs = pipe->curbuf = 0;
1246         file_accessed(in);
1247         return bytes;
1248
1249 out_release:
1250         /*
1251          * If we did an incomplete transfer we must release
1252          * the pipe buffers in question:
1253          */
1254         for (i = 0; i < pipe->buffers; i++) {
1255                 struct pipe_buffer *buf = pipe->bufs + i;
1256
1257                 if (buf->ops) {
1258                         buf->ops->release(pipe, buf);
1259                         buf->ops = NULL;
1260                 }
1261         }
1262
1263         if (!bytes)
1264                 bytes = ret;
1265
1266         goto done;
1267 }
1268 EXPORT_SYMBOL(splice_direct_to_actor);
1269
1270 static int direct_splice_actor(struct pipe_inode_info *pipe,
1271                                struct splice_desc *sd)
1272 {
1273         struct file *file = sd->u.file;
1274
1275         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1276                               sd->flags);
1277 }
1278
1279 /**
1280  * do_splice_direct - splices data directly between two files
1281  * @in:         file to splice from
1282  * @ppos:       input file offset
1283  * @out:        file to splice to
1284  * @len:        number of bytes to splice
1285  * @flags:      splice modifier flags
1286  *
1287  * Description:
1288  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1289  *    doing it in the application would incur an extra system call
1290  *    (splice in + splice out, as compared to just sendfile()). So this helper
1291  *    can splice directly through a process-private pipe.
1292  *
1293  */
1294 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1295                       size_t len, unsigned int flags)
1296 {
1297         struct splice_desc sd = {
1298                 .len            = len,
1299                 .total_len      = len,
1300                 .flags          = flags,
1301                 .pos            = *ppos,
1302                 .u.file         = out,
1303         };
1304         long ret;
1305
1306         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1307         if (ret > 0)
1308                 *ppos = sd.pos;
1309
1310         return ret;
1311 }
1312
1313 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1314                                struct pipe_inode_info *opipe,
1315                                size_t len, unsigned int flags);
1316
1317 /*
1318  * Determine where to splice to/from.
1319  */
1320 static long do_splice(struct file *in, loff_t __user *off_in,
1321                       struct file *out, loff_t __user *off_out,
1322                       size_t len, unsigned int flags)
1323 {
1324         struct pipe_inode_info *ipipe;
1325         struct pipe_inode_info *opipe;
1326         loff_t offset, *off;
1327         long ret;
1328
1329         ipipe = get_pipe_info(in);
1330         opipe = get_pipe_info(out);
1331
1332         if (ipipe && opipe) {
1333                 if (off_in || off_out)
1334                         return -ESPIPE;
1335
1336                 if (!(in->f_mode & FMODE_READ))
1337                         return -EBADF;
1338
1339                 if (!(out->f_mode & FMODE_WRITE))
1340                         return -EBADF;
1341
1342                 /* Splicing to self would be fun, but... */
1343                 if (ipipe == opipe)
1344                         return -EINVAL;
1345
1346                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1347         }
1348
1349         if (ipipe) {
1350                 if (off_in)
1351                         return -ESPIPE;
1352                 if (off_out) {
1353                         if (!(out->f_mode & FMODE_PWRITE))
1354                                 return -EINVAL;
1355                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1356                                 return -EFAULT;
1357                         off = &offset;
1358                 } else
1359                         off = &out->f_pos;
1360
1361                 ret = do_splice_from(ipipe, out, off, len, flags);
1362
1363                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1364                         ret = -EFAULT;
1365
1366                 return ret;
1367         }
1368
1369         if (opipe) {
1370                 if (off_out)
1371                         return -ESPIPE;
1372                 if (off_in) {
1373                         if (!(in->f_mode & FMODE_PREAD))
1374                                 return -EINVAL;
1375                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1376                                 return -EFAULT;
1377                         off = &offset;
1378                 } else
1379                         off = &in->f_pos;
1380
1381                 ret = do_splice_to(in, off, opipe, len, flags);
1382
1383                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1384                         ret = -EFAULT;
1385
1386                 return ret;
1387         }
1388
1389         return -EINVAL;
1390 }
1391
1392 /*
1393  * Map an iov into an array of pages and offset/length tupples. With the
1394  * partial_page structure, we can map several non-contiguous ranges into
1395  * our ones pages[] map instead of splitting that operation into pieces.
1396  * Could easily be exported as a generic helper for other users, in which
1397  * case one would probably want to add a 'max_nr_pages' parameter as well.
1398  */
1399 static int get_iovec_page_array(const struct iovec __user *iov,
1400                                 unsigned int nr_vecs, struct page **pages,
1401                                 struct partial_page *partial, int aligned,
1402                                 unsigned int pipe_buffers)
1403 {
1404         int buffers = 0, error = 0;
1405
1406         while (nr_vecs) {
1407                 unsigned long off, npages;
1408                 struct iovec entry;
1409                 void __user *base;
1410                 size_t len;
1411                 int i;
1412
1413                 error = -EFAULT;
1414                 if (copy_from_user(&entry, iov, sizeof(entry)))
1415                         break;
1416
1417                 base = entry.iov_base;
1418                 len = entry.iov_len;
1419
1420                 /*
1421                  * Sanity check this iovec. 0 read succeeds.
1422                  */
1423                 error = 0;
1424                 if (unlikely(!len))
1425                         break;
1426                 error = -EFAULT;
1427                 if (!access_ok(VERIFY_READ, base, len))
1428                         break;
1429
1430                 /*
1431                  * Get this base offset and number of pages, then map
1432                  * in the user pages.
1433                  */
1434                 off = (unsigned long) base & ~PAGE_MASK;
1435
1436                 /*
1437                  * If asked for alignment, the offset must be zero and the
1438                  * length a multiple of the PAGE_SIZE.
1439                  */
1440                 error = -EINVAL;
1441                 if (aligned && (off || len & ~PAGE_MASK))
1442                         break;
1443
1444                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1445                 if (npages > pipe_buffers - buffers)
1446                         npages = pipe_buffers - buffers;
1447
1448                 error = get_user_pages_fast((unsigned long)base, npages,
1449                                         0, &pages[buffers]);
1450
1451                 if (unlikely(error <= 0))
1452                         break;
1453
1454                 /*
1455                  * Fill this contiguous range into the partial page map.
1456                  */
1457                 for (i = 0; i < error; i++) {
1458                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1459
1460                         partial[buffers].offset = off;
1461                         partial[buffers].len = plen;
1462
1463                         off = 0;
1464                         len -= plen;
1465                         buffers++;
1466                 }
1467
1468                 /*
1469                  * We didn't complete this iov, stop here since it probably
1470                  * means we have to move some of this into a pipe to
1471                  * be able to continue.
1472                  */
1473                 if (len)
1474                         break;
1475
1476                 /*
1477                  * Don't continue if we mapped fewer pages than we asked for,
1478                  * or if we mapped the max number of pages that we have
1479                  * room for.
1480                  */
1481                 if (error < npages || buffers == pipe_buffers)
1482                         break;
1483
1484                 nr_vecs--;
1485                 iov++;
1486         }
1487
1488         if (buffers)
1489                 return buffers;
1490
1491         return error;
1492 }
1493
1494 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1495                         struct splice_desc *sd)
1496 {
1497         char *src;
1498         int ret;
1499
1500         /*
1501          * See if we can use the atomic maps, by prefaulting in the
1502          * pages and doing an atomic copy
1503          */
1504         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1505                 src = buf->ops->map(pipe, buf, 1);
1506                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1507                                                         sd->len);
1508                 buf->ops->unmap(pipe, buf, src);
1509                 if (!ret) {
1510                         ret = sd->len;
1511                         goto out;
1512                 }
1513         }
1514
1515         /*
1516          * No dice, use slow non-atomic map and copy
1517          */
1518         src = buf->ops->map(pipe, buf, 0);
1519
1520         ret = sd->len;
1521         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1522                 ret = -EFAULT;
1523
1524         buf->ops->unmap(pipe, buf, src);
1525 out:
1526         if (ret > 0)
1527                 sd->u.userptr += ret;
1528         return ret;
1529 }
1530
1531 /*
1532  * For lack of a better implementation, implement vmsplice() to userspace
1533  * as a simple copy of the pipes pages to the user iov.
1534  */
1535 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1536                              unsigned long nr_segs, unsigned int flags)
1537 {
1538         struct pipe_inode_info *pipe;
1539         struct splice_desc sd;
1540         ssize_t size;
1541         int error;
1542         long ret;
1543
1544         pipe = get_pipe_info(file);
1545         if (!pipe)
1546                 return -EBADF;
1547
1548         pipe_lock(pipe);
1549
1550         error = ret = 0;
1551         while (nr_segs) {
1552                 void __user *base;
1553                 size_t len;
1554
1555                 /*
1556                  * Get user address base and length for this iovec.
1557                  */
1558                 error = get_user(base, &iov->iov_base);
1559                 if (unlikely(error))
1560                         break;
1561                 error = get_user(len, &iov->iov_len);
1562                 if (unlikely(error))
1563                         break;
1564
1565                 /*
1566                  * Sanity check this iovec. 0 read succeeds.
1567                  */
1568                 if (unlikely(!len))
1569                         break;
1570                 if (unlikely(!base)) {
1571                         error = -EFAULT;
1572                         break;
1573                 }
1574
1575                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1576                         error = -EFAULT;
1577                         break;
1578                 }
1579
1580                 sd.len = 0;
1581                 sd.total_len = len;
1582                 sd.flags = flags;
1583                 sd.u.userptr = base;
1584                 sd.pos = 0;
1585
1586                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1587                 if (size < 0) {
1588                         if (!ret)
1589                                 ret = size;
1590
1591                         break;
1592                 }
1593
1594                 ret += size;
1595
1596                 if (size < len)
1597                         break;
1598
1599                 nr_segs--;
1600                 iov++;
1601         }
1602
1603         pipe_unlock(pipe);
1604
1605         if (!ret)
1606                 ret = error;
1607
1608         return ret;
1609 }
1610
1611 /*
1612  * vmsplice splices a user address range into a pipe. It can be thought of
1613  * as splice-from-memory, where the regular splice is splice-from-file (or
1614  * to file). In both cases the output is a pipe, naturally.
1615  */
1616 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1617                              unsigned long nr_segs, unsigned int flags)
1618 {
1619         struct pipe_inode_info *pipe;
1620         struct page *pages[PIPE_DEF_BUFFERS];
1621         struct partial_page partial[PIPE_DEF_BUFFERS];
1622         struct splice_pipe_desc spd = {
1623                 .pages = pages,
1624                 .partial = partial,
1625                 .nr_pages_max = PIPE_DEF_BUFFERS,
1626                 .flags = flags,
1627                 .ops = &user_page_pipe_buf_ops,
1628                 .spd_release = spd_release_page,
1629         };
1630         long ret;
1631
1632         pipe = get_pipe_info(file);
1633         if (!pipe)
1634                 return -EBADF;
1635
1636         if (splice_grow_spd(pipe, &spd))
1637                 return -ENOMEM;
1638
1639         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1640                                             spd.partial, flags & SPLICE_F_GIFT,
1641                                             spd.nr_pages_max);
1642         if (spd.nr_pages <= 0)
1643                 ret = spd.nr_pages;
1644         else
1645                 ret = splice_to_pipe(pipe, &spd);
1646
1647         splice_shrink_spd(&spd);
1648         return ret;
1649 }
1650
1651 /*
1652  * Note that vmsplice only really supports true splicing _from_ user memory
1653  * to a pipe, not the other way around. Splicing from user memory is a simple
1654  * operation that can be supported without any funky alignment restrictions
1655  * or nasty vm tricks. We simply map in the user memory and fill them into
1656  * a pipe. The reverse isn't quite as easy, though. There are two possible
1657  * solutions for that:
1658  *
1659  *      - memcpy() the data internally, at which point we might as well just
1660  *        do a regular read() on the buffer anyway.
1661  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1662  *        has restriction limitations on both ends of the pipe).
1663  *
1664  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1665  *
1666  */
1667 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1668                 unsigned long, nr_segs, unsigned int, flags)
1669 {
1670         struct file *file;
1671         long error;
1672         int fput;
1673
1674         if (unlikely(nr_segs > UIO_MAXIOV))
1675                 return -EINVAL;
1676         else if (unlikely(!nr_segs))
1677                 return 0;
1678
1679         error = -EBADF;
1680         file = fget_light(fd, &fput);
1681         if (file) {
1682                 if (file->f_mode & FMODE_WRITE)
1683                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1684                 else if (file->f_mode & FMODE_READ)
1685                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1686
1687                 fput_light(file, fput);
1688         }
1689
1690         return error;
1691 }
1692
1693 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1694                 int, fd_out, loff_t __user *, off_out,
1695                 size_t, len, unsigned int, flags)
1696 {
1697         long error;
1698         struct file *in, *out;
1699         int fput_in, fput_out;
1700
1701         if (unlikely(!len))
1702                 return 0;
1703
1704         error = -EBADF;
1705         in = fget_light(fd_in, &fput_in);
1706         if (in) {
1707                 if (in->f_mode & FMODE_READ) {
1708                         out = fget_light(fd_out, &fput_out);
1709                         if (out) {
1710                                 if (out->f_mode & FMODE_WRITE)
1711                                         error = do_splice(in, off_in,
1712                                                           out, off_out,
1713                                                           len, flags);
1714                                 fput_light(out, fput_out);
1715                         }
1716                 }
1717
1718                 fput_light(in, fput_in);
1719         }
1720
1721         return error;
1722 }
1723
1724 /*
1725  * Make sure there's data to read. Wait for input if we can, otherwise
1726  * return an appropriate error.
1727  */
1728 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1729 {
1730         int ret;
1731
1732         /*
1733          * Check ->nrbufs without the inode lock first. This function
1734          * is speculative anyways, so missing one is ok.
1735          */
1736         if (pipe->nrbufs)
1737                 return 0;
1738
1739         ret = 0;
1740         pipe_lock(pipe);
1741
1742         while (!pipe->nrbufs) {
1743                 if (signal_pending(current)) {
1744                         ret = -ERESTARTSYS;
1745                         break;
1746                 }
1747                 if (!pipe->writers)
1748                         break;
1749                 if (!pipe->waiting_writers) {
1750                         if (flags & SPLICE_F_NONBLOCK) {
1751                                 ret = -EAGAIN;
1752                                 break;
1753                         }
1754                 }
1755                 pipe_wait(pipe);
1756         }
1757
1758         pipe_unlock(pipe);
1759         return ret;
1760 }
1761
1762 /*
1763  * Make sure there's writeable room. Wait for room if we can, otherwise
1764  * return an appropriate error.
1765  */
1766 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1767 {
1768         int ret;
1769
1770         /*
1771          * Check ->nrbufs without the inode lock first. This function
1772          * is speculative anyways, so missing one is ok.
1773          */
1774         if (pipe->nrbufs < pipe->buffers)
1775                 return 0;
1776
1777         ret = 0;
1778         pipe_lock(pipe);
1779
1780         while (pipe->nrbufs >= pipe->buffers) {
1781                 if (!pipe->readers) {
1782                         send_sig(SIGPIPE, current, 0);
1783                         ret = -EPIPE;
1784                         break;
1785                 }
1786                 if (flags & SPLICE_F_NONBLOCK) {
1787                         ret = -EAGAIN;
1788                         break;
1789                 }
1790                 if (signal_pending(current)) {
1791                         ret = -ERESTARTSYS;
1792                         break;
1793                 }
1794                 pipe->waiting_writers++;
1795                 pipe_wait(pipe);
1796                 pipe->waiting_writers--;
1797         }
1798
1799         pipe_unlock(pipe);
1800         return ret;
1801 }
1802
1803 /*
1804  * Splice contents of ipipe to opipe.
1805  */
1806 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1807                                struct pipe_inode_info *opipe,
1808                                size_t len, unsigned int flags)
1809 {
1810         struct pipe_buffer *ibuf, *obuf;
1811         int ret = 0, nbuf;
1812         bool input_wakeup = false;
1813
1814
1815 retry:
1816         ret = ipipe_prep(ipipe, flags);
1817         if (ret)
1818                 return ret;
1819
1820         ret = opipe_prep(opipe, flags);
1821         if (ret)
1822                 return ret;
1823
1824         /*
1825          * Potential ABBA deadlock, work around it by ordering lock
1826          * grabbing by pipe info address. Otherwise two different processes
1827          * could deadlock (one doing tee from A -> B, the other from B -> A).
1828          */
1829         pipe_double_lock(ipipe, opipe);
1830
1831         do {
1832                 if (!opipe->readers) {
1833                         send_sig(SIGPIPE, current, 0);
1834                         if (!ret)
1835                                 ret = -EPIPE;
1836                         break;
1837                 }
1838
1839                 if (!ipipe->nrbufs && !ipipe->writers)
1840                         break;
1841
1842                 /*
1843                  * Cannot make any progress, because either the input
1844                  * pipe is empty or the output pipe is full.
1845                  */
1846                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1847                         /* Already processed some buffers, break */
1848                         if (ret)
1849                                 break;
1850
1851                         if (flags & SPLICE_F_NONBLOCK) {
1852                                 ret = -EAGAIN;
1853                                 break;
1854                         }
1855
1856                         /*
1857                          * We raced with another reader/writer and haven't
1858                          * managed to process any buffers.  A zero return
1859                          * value means EOF, so retry instead.
1860                          */
1861                         pipe_unlock(ipipe);
1862                         pipe_unlock(opipe);
1863                         goto retry;
1864                 }
1865
1866                 ibuf = ipipe->bufs + ipipe->curbuf;
1867                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1868                 obuf = opipe->bufs + nbuf;
1869
1870                 if (len >= ibuf->len) {
1871                         /*
1872                          * Simply move the whole buffer from ipipe to opipe
1873                          */
1874                         *obuf = *ibuf;
1875                         ibuf->ops = NULL;
1876                         opipe->nrbufs++;
1877                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1878                         ipipe->nrbufs--;
1879                         input_wakeup = true;
1880                 } else {
1881                         /*
1882                          * Get a reference to this pipe buffer,
1883                          * so we can copy the contents over.
1884                          */
1885                         ibuf->ops->get(ipipe, ibuf);
1886                         *obuf = *ibuf;
1887
1888                         /*
1889                          * Don't inherit the gift flag, we need to
1890                          * prevent multiple steals of this page.
1891                          */
1892                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1893
1894                         obuf->len = len;
1895                         opipe->nrbufs++;
1896                         ibuf->offset += obuf->len;
1897                         ibuf->len -= obuf->len;
1898                 }
1899                 ret += obuf->len;
1900                 len -= obuf->len;
1901         } while (len);
1902
1903         pipe_unlock(ipipe);
1904         pipe_unlock(opipe);
1905
1906         /*
1907          * If we put data in the output pipe, wakeup any potential readers.
1908          */
1909         if (ret > 0)
1910                 wakeup_pipe_readers(opipe);
1911
1912         if (input_wakeup)
1913                 wakeup_pipe_writers(ipipe);
1914
1915         return ret;
1916 }
1917
1918 /*
1919  * Link contents of ipipe to opipe.
1920  */
1921 static int link_pipe(struct pipe_inode_info *ipipe,
1922                      struct pipe_inode_info *opipe,
1923                      size_t len, unsigned int flags)
1924 {
1925         struct pipe_buffer *ibuf, *obuf;
1926         int ret = 0, i = 0, nbuf;
1927
1928         /*
1929          * Potential ABBA deadlock, work around it by ordering lock
1930          * grabbing by pipe info address. Otherwise two different processes
1931          * could deadlock (one doing tee from A -> B, the other from B -> A).
1932          */
1933         pipe_double_lock(ipipe, opipe);
1934
1935         do {
1936                 if (!opipe->readers) {
1937                         send_sig(SIGPIPE, current, 0);
1938                         if (!ret)
1939                                 ret = -EPIPE;
1940                         break;
1941                 }
1942
1943                 /*
1944                  * If we have iterated all input buffers or ran out of
1945                  * output room, break.
1946                  */
1947                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1948                         break;
1949
1950                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1951                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1952
1953                 /*
1954                  * Get a reference to this pipe buffer,
1955                  * so we can copy the contents over.
1956                  */
1957                 ibuf->ops->get(ipipe, ibuf);
1958
1959                 obuf = opipe->bufs + nbuf;
1960                 *obuf = *ibuf;
1961
1962                 /*
1963                  * Don't inherit the gift flag, we need to
1964                  * prevent multiple steals of this page.
1965                  */
1966                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1967
1968                 if (obuf->len > len)
1969                         obuf->len = len;
1970
1971                 opipe->nrbufs++;
1972                 ret += obuf->len;
1973                 len -= obuf->len;
1974                 i++;
1975         } while (len);
1976
1977         /*
1978          * return EAGAIN if we have the potential of some data in the
1979          * future, otherwise just return 0
1980          */
1981         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1982                 ret = -EAGAIN;
1983
1984         pipe_unlock(ipipe);
1985         pipe_unlock(opipe);
1986
1987         /*
1988          * If we put data in the output pipe, wakeup any potential readers.
1989          */
1990         if (ret > 0)
1991                 wakeup_pipe_readers(opipe);
1992
1993         return ret;
1994 }
1995
1996 /*
1997  * This is a tee(1) implementation that works on pipes. It doesn't copy
1998  * any data, it simply references the 'in' pages on the 'out' pipe.
1999  * The 'flags' used are the SPLICE_F_* variants, currently the only
2000  * applicable one is SPLICE_F_NONBLOCK.
2001  */
2002 static long do_tee(struct file *in, struct file *out, size_t len,
2003                    unsigned int flags)
2004 {
2005         struct pipe_inode_info *ipipe = get_pipe_info(in);
2006         struct pipe_inode_info *opipe = get_pipe_info(out);
2007         int ret = -EINVAL;
2008
2009         /*
2010          * Duplicate the contents of ipipe to opipe without actually
2011          * copying the data.
2012          */
2013         if (ipipe && opipe && ipipe != opipe) {
2014                 /*
2015                  * Keep going, unless we encounter an error. The ipipe/opipe
2016                  * ordering doesn't really matter.
2017                  */
2018                 ret = ipipe_prep(ipipe, flags);
2019                 if (!ret) {
2020                         ret = opipe_prep(opipe, flags);
2021                         if (!ret)
2022                                 ret = link_pipe(ipipe, opipe, len, flags);
2023                 }
2024         }
2025
2026         return ret;
2027 }
2028
2029 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2030 {
2031         struct file *in;
2032         int error, fput_in;
2033
2034         if (unlikely(!len))
2035                 return 0;
2036
2037         error = -EBADF;
2038         in = fget_light(fdin, &fput_in);
2039         if (in) {
2040                 if (in->f_mode & FMODE_READ) {
2041                         int fput_out;
2042                         struct file *out = fget_light(fdout, &fput_out);
2043
2044                         if (out) {
2045                                 if (out->f_mode & FMODE_WRITE)
2046                                         error = do_tee(in, out, len, flags);
2047                                 fput_light(out, fput_out);
2048                         }
2049                 }
2050                 fput_light(in, fput_in);
2051         }
2052
2053         return error;
2054 }