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
 701         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 702                 more |= MSG_SENDPAGE_NOTLAST;
 703
 704         return file->f_op->sendpage(file, buf->page, buf->offset,
 705                                     sd->len, &pos, more);
 706 }
 707
 708 /*
 709  * This is a little more tricky than the file -> pipe splicing. There are
 710  * basically three cases:
 711  *
 712  *      - Destination page already exists in the address space and there
 713  *        are users of it. For that case we have no other option that
 714  *        copying the data. Tough luck.
 715  *      - Destination page already exists in the address space, but there
 716  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 717  *        through to last case.
 718  *      - Destination page does not exist, we can add the pipe page to
 719  *        the page cache and avoid the copy.
 720  *
 721  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 722  * sd->flags), we attempt to migrate pages from the pipe to the output
 723  * file address space page cache. This is possible if no one else has
 724  * the pipe page referenced outside of the pipe and page cache. If
 725  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 726  * a new page in the output file page cache and fill/dirty that.
 727  */
 728 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 729                  struct splice_desc *sd)
 730 {
 731         struct file *file = sd->u.file;
 732         struct address_space *mapping = file->f_mapping;
 733         unsigned int offset, this_len;
 734         struct page *page;
 735         void *fsdata;
 736         int ret;
 737
 738         offset = sd->pos & ~PAGE_CACHE_MASK;
 739
 740         this_len = sd->len;
 741         if (this_len + offset > PAGE_CACHE_SIZE)
 742                 this_len = PAGE_CACHE_SIZE - offset;
 743
 744         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 745                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 746         if (unlikely(ret))
 747                 goto out;
 748
 749         if (buf->page != page) {
 750                 /*
 751                  * Careful, ->map() uses KM_USER0!
 752                  */
 753                 char *src = buf->ops->map(pipe, buf, 1);
 754                 char *dst = kmap_atomic(page, KM_USER1);
 755
 756                 memcpy(dst + offset, src + buf->offset, this_len);
 757                 flush_dcache_page(page);
 758                 kunmap_atomic(dst, KM_USER1);
 759                 buf->ops->unmap(pipe, buf, src);
 760         }
 761         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 762                                 page, fsdata);
 763 out:
 764         return ret;
 765 }
 766 EXPORT_SYMBOL(pipe_to_file);
 767
 768 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 769 {
 770         smp_mb();
 771         if (waitqueue_active(&pipe->wait))
 772                 wake_up_interruptible(&pipe->wait);
 773         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 774 }
 775
 776 /**
 777  * splice_from_pipe_feed - feed available data from a pipe to a file
 778  * @pipe:       pipe to splice from
 779  * @sd:         information to @actor
 780  * @actor:      handler that splices the data
 781  *
 782  * Description:
 783  *    This function loops over the pipe and calls @actor to do the
 784  *    actual moving of a single struct pipe_buffer to the desired
 785  *    destination.  It returns when there's no more buffers left in
 786  *    the pipe or if the requested number of bytes (@sd->total_len)
 787  *    have been copied.  It returns a positive number (one) if the
 788  *    pipe needs to be filled with more data, zero if the required
 789  *    number of bytes have been copied and -errno on error.
 790  *
 791  *    This, together with splice_from_pipe_{begin,end,next}, may be
 792  *    used to implement the functionality of __splice_from_pipe() when
 793  *    locking is required around copying the pipe buffers to the
 794  *    destination.
 795  */
 796 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 797                           splice_actor *actor)
 798 {
 799         int ret;
 800
 801         while (pipe->nrbufs) {
 802                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 803                 const struct pipe_buf_operations *ops = buf->ops;
 804
 805                 sd->len = buf->len;
 806                 if (sd->len > sd->total_len)
 807                         sd->len = sd->total_len;
 808
 809                 ret = buf->ops->confirm(pipe, buf);
 810                 if (unlikely(ret)) {
 811                         if (ret == -ENODATA)
 812                                 ret = 0;
 813                         return ret;
 814                 }
 815
 816                 ret = actor(pipe, buf, sd);
 817                 if (ret <= 0)
 818                         return ret;
 819
 820                 buf->offset += ret;
 821                 buf->len -= ret;
 822
 823                 sd->num_spliced += ret;
 824                 sd->len -= ret;
 825                 sd->pos += ret;
 826                 sd->total_len -= ret;
 827
 828                 if (!buf->len) {
 829                         buf->ops = NULL;
 830                         ops->release(pipe, buf);
 831                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 832                         pipe->nrbufs--;
 833                         if (pipe->inode)
 834                                 sd->need_wakeup = true;
 835                 }
 836
 837                 if (!sd->total_len)
 838                         return 0;
 839         }
 840
 841         return 1;
 842 }
 843 EXPORT_SYMBOL(splice_from_pipe_feed);
 844
 845 /**
 846  * splice_from_pipe_next - wait for some data to splice from
 847  * @pipe:       pipe to splice from
 848  * @sd:         information about the splice operation
 849  *
 850  * Description:
 851  *    This function will wait for some data and return a positive
 852  *    value (one) if pipe buffers are available.  It will return zero
 853  *    or -errno if no more data needs to be spliced.
 854  */
 855 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 856 {
 857         while (!pipe->nrbufs) {
 858                 if (!pipe->writers)
 859                         return 0;
 860
 861                 if (!pipe->waiting_writers && sd->num_spliced)
 862                         return 0;
 863
 864                 if (sd->flags & SPLICE_F_NONBLOCK)
 865                         return -EAGAIN;
 866
 867                 if (signal_pending(current))
 868                         return -ERESTARTSYS;
 869
 870                 if (sd->need_wakeup) {
 871                         wakeup_pipe_writers(pipe);
 872                         sd->need_wakeup = false;
 873                 }
 874
 875                 pipe_wait(pipe);
 876         }
 877
 878         return 1;
 879 }
 880 EXPORT_SYMBOL(splice_from_pipe_next);
 881
 882 /**
 883  * splice_from_pipe_begin - start splicing from pipe
 884  * @sd:         information about the splice operation
 885  *
 886  * Description:
 887  *    This function should be called before a loop containing
 888  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 889  *    initialize the necessary fields of @sd.
 890  */
 891 void splice_from_pipe_begin(struct splice_desc *sd)
 892 {
 893         sd->num_spliced = 0;
 894         sd->need_wakeup = false;
 895 }
 896 EXPORT_SYMBOL(splice_from_pipe_begin);
 897
 898 /**
 899  * splice_from_pipe_end - finish splicing from pipe
 900  * @pipe:       pipe to splice from
 901  * @sd:         information about the splice operation
 902  *
 903  * Description:
 904  *    This function will wake up pipe writers if necessary.  It should
 905  *    be called after a loop containing splice_from_pipe_next() and
 906  *    splice_from_pipe_feed().
 907  */
 908 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 909 {
 910         if (sd->need_wakeup)
 911                 wakeup_pipe_writers(pipe);
 912 }
 913 EXPORT_SYMBOL(splice_from_pipe_end);
 914
 915 /**
 916  * __splice_from_pipe - splice data from a pipe to given actor
 917  * @pipe:       pipe to splice from
 918  * @sd:         information to @actor
 919  * @actor:      handler that splices the data
 920  *
 921  * Description:
 922  *    This function does little more than loop over the pipe and call
 923  *    @actor to do the actual moving of a single struct pipe_buffer to
 924  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 925  *    pipe_to_user.
 926  *
 927  */
 928 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 929                            splice_actor *actor)
 930 {
 931         int ret;
 932
 933         splice_from_pipe_begin(sd);
 934         do {
 935                 ret = splice_from_pipe_next(pipe, sd);
 936                 if (ret > 0)
 937                         ret = splice_from_pipe_feed(pipe, sd, actor);
 938         } while (ret > 0);
 939         splice_from_pipe_end(pipe, sd);
 940
 941         return sd->num_spliced ? sd->num_spliced : ret;
 942 }
 943 EXPORT_SYMBOL(__splice_from_pipe);
 944
 945 /**
 946  * splice_from_pipe - splice data from a pipe to a file
 947  * @pipe:       pipe to splice from
 948  * @out:        file to splice to
 949  * @ppos:       position in @out
 950  * @len:        how many bytes to splice
 951  * @flags:      splice modifier flags
 952  * @actor:      handler that splices the data
 953  *
 954  * Description:
 955  *    See __splice_from_pipe. This function locks the pipe inode,
 956  *    otherwise it's identical to __splice_from_pipe().
 957  *
 958  */
 959 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 960                          loff_t *ppos, size_t len, unsigned int flags,
 961                          splice_actor *actor)
 962 {
 963         ssize_t ret;
 964         struct splice_desc sd = {
 965                 .total_len = len,
 966                 .flags = flags,
 967                 .pos = *ppos,
 968                 .u.file = out,
 969         };
 970
 971         pipe_lock(pipe);
 972         ret = __splice_from_pipe(pipe, &sd, actor);
 973         pipe_unlock(pipe);
 974
 975         return ret;
 976 }
 977
 978 /**
 979  * generic_file_splice_write - splice data from a pipe to a file
 980  * @pipe:       pipe info
 981  * @out:        file to write to
 982  * @ppos:       position in @out
 983  * @len:        number of bytes to splice
 984  * @flags:      splice modifier flags
 985  *
 986  * Description:
 987  *    Will either move or copy pages (determined by @flags options) from
 988  *    the given pipe inode to the given file.
 989  *
 990  */
 991 ssize_t
 992 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 993                           loff_t *ppos, size_t len, unsigned int flags)
 994 {
 995         struct address_space *mapping = out->f_mapping;
 996         struct inode *inode = mapping->host;
 997         struct splice_desc sd = {
 998                 .total_len = len,
 999                 .flags = flags,
1000                 .pos = *ppos,
1001                 .u.file = out,
1002         };
1003         ssize_t ret;
1004
1005         pipe_lock(pipe);
1006
1007         splice_from_pipe_begin(&sd);
1008         do {
1009                 ret = splice_from_pipe_next(pipe, &sd);
1010                 if (ret <= 0)
1011                         break;
1012
1013                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1014                 ret = file_remove_suid(out);
1015                 if (!ret) {
1016                         file_update_time(out);
1017                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1018                 }
1019                 mutex_unlock(&inode->i_mutex);
1020         } while (ret > 0);
1021         splice_from_pipe_end(pipe, &sd);
1022
1023         pipe_unlock(pipe);
1024
1025         if (sd.num_spliced)
1026                 ret = sd.num_spliced;
1027
1028         if (ret > 0) {
1029                 unsigned long nr_pages;
1030                 int err;
1031
1032                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1033
1034                 err = generic_write_sync(out, *ppos, ret);
1035                 if (err)
1036                         ret = err;
1037                 else
1038                         *ppos += ret;
1039                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1040         }
1041
1042         return ret;
1043 }
1044
1045 EXPORT_SYMBOL(generic_file_splice_write);
1046
1047 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1048                           struct splice_desc *sd)
1049 {
1050         int ret;
1051         void *data;
1052
1053         data = buf->ops->map(pipe, buf, 0);
1054         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1055         buf->ops->unmap(pipe, buf, data);
1056
1057         return ret;
1058 }
1059
1060 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1061                                          struct file *out, loff_t *ppos,
1062                                          size_t len, unsigned int flags)
1063 {
1064         ssize_t ret;
1065
1066         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1067         if (ret > 0)
1068                 *ppos += ret;
1069
1070         return ret;
1071 }
1072
1073 /**
1074  * generic_splice_sendpage - splice data from a pipe to a socket
1075  * @pipe:       pipe to splice from
1076  * @out:        socket to write to
1077  * @ppos:       position in @out
1078  * @len:        number of bytes to splice
1079  * @flags:      splice modifier flags
1080  *
1081  * Description:
1082  *    Will send @len bytes from the pipe to a network socket. No data copying
1083  *    is involved.
1084  *
1085  */
1086 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1087                                 loff_t *ppos, size_t len, unsigned int flags)
1088 {
1089         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1090 }
1091
1092 EXPORT_SYMBOL(generic_splice_sendpage);
1093
1094 /*
1095  * Attempt to initiate a splice from pipe to file.
1096  */
1097 long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1098                     loff_t *ppos, size_t len, unsigned int flags)
1099 {
1100         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1101                                 loff_t *, size_t, unsigned int);
1102         int ret;
1103
1104         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1105                 return -EBADF;
1106
1107         if (unlikely(out->f_flags & O_APPEND))
1108                 return -EINVAL;
1109
1110         ret = rw_verify_area(WRITE, out, ppos, len);
1111         if (unlikely(ret < 0))
1112                 return ret;
1113
1114         if (out->f_op && out->f_op->splice_write)
1115                 splice_write = out->f_op->splice_write;
1116         else
1117                 splice_write = default_file_splice_write;
1118
1119         return splice_write(pipe, out, ppos, len, flags);
1120 }
1121 EXPORT_SYMBOL(do_splice_from);
1122
1123 /*
1124  * Attempt to initiate a splice from a file to a pipe.
1125  */
1126 long do_splice_to(struct file *in, loff_t *ppos,
1127                   struct pipe_inode_info *pipe, size_t len,
1128                   unsigned int flags)
1129 {
1130         ssize_t (*splice_read)(struct file *, loff_t *,
1131                                struct pipe_inode_info *, size_t, unsigned int);
1132         int ret;
1133
1134         if (unlikely(!(in->f_mode & FMODE_READ)))
1135                 return -EBADF;
1136
1137         ret = rw_verify_area(READ, in, ppos, len);
1138         if (unlikely(ret < 0))
1139                 return ret;
1140
1141         if (in->f_op && in->f_op->splice_read)
1142                 splice_read = in->f_op->splice_read;
1143         else
1144                 splice_read = default_file_splice_read;
1145
1146         return splice_read(in, ppos, pipe, len, flags);
1147 }
1148 EXPORT_SYMBOL(do_splice_to);
1149
1150 /**
1151  * splice_direct_to_actor - splices data directly between two non-pipes
1152  * @in:         file to splice from
1153  * @sd:         actor information on where to splice to
1154  * @actor:      handles the data splicing
1155  *
1156  * Description:
1157  *    This is a special case helper to splice directly between two
1158  *    points, without requiring an explicit pipe. Internally an allocated
1159  *    pipe is cached in the process, and reused during the lifetime of
1160  *    that process.
1161  *
1162  */
1163 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1164                                splice_direct_actor *actor)
1165 {
1166         struct pipe_inode_info *pipe;
1167         long ret, bytes;
1168         umode_t i_mode;
1169         size_t len;
1170         int i, flags;
1171
1172         /*
1173          * We require the input being a regular file, as we don't want to
1174          * randomly drop data for eg socket -> socket splicing. Use the
1175          * piped splicing for that!
1176          */
1177         i_mode = in->f_path.dentry->d_inode->i_mode;
1178         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1179                 return -EINVAL;
1180
1181         /*
1182          * neither in nor out is a pipe, setup an internal pipe attached to
1183          * 'out' and transfer the wanted data from 'in' to 'out' through that
1184          */
1185         pipe = current->splice_pipe;
1186         if (unlikely(!pipe)) {
1187                 pipe = alloc_pipe_info(NULL);
1188                 if (!pipe)
1189                         return -ENOMEM;
1190
1191                 /*
1192                  * We don't have an immediate reader, but we'll read the stuff
1193                  * out of the pipe right after the splice_to_pipe(). So set
1194                  * PIPE_READERS appropriately.
1195                  */
1196                 pipe->readers = 1;
1197
1198                 current->splice_pipe = pipe;
1199         }
1200
1201         /*
1202          * Do the splice.
1203          */
1204         ret = 0;
1205         bytes = 0;
1206         len = sd->total_len;
1207         flags = sd->flags;
1208
1209         /*
1210          * Don't block on output, we have to drain the direct pipe.
1211          */
1212         sd->flags &= ~SPLICE_F_NONBLOCK;
1213
1214         while (len) {
1215                 size_t read_len;
1216                 loff_t pos = sd->pos, prev_pos = pos;
1217
1218                 ret = do_splice_to(in, &pos, pipe, len, flags);
1219                 if (unlikely(ret <= 0))
1220                         goto out_release;
1221
1222                 read_len = ret;
1223                 sd->total_len = read_len;
1224
1225                 /*
1226                  * NOTE: nonblocking mode only applies to the input. We
1227                  * must not do the output in nonblocking mode as then we
1228                  * could get stuck data in the internal pipe:
1229                  */
1230                 ret = actor(pipe, sd);
1231                 if (unlikely(ret <= 0)) {
1232                         sd->pos = prev_pos;
1233                         goto out_release;
1234                 }
1235
1236                 bytes += ret;
1237                 len -= ret;
1238                 sd->pos = pos;
1239
1240                 if (ret < read_len) {
1241                         sd->pos = prev_pos + ret;
1242                         goto out_release;
1243                 }
1244         }
1245
1246 done:
1247         pipe->nrbufs = pipe->curbuf = 0;
1248         file_accessed(in);
1249         return bytes;
1250
1251 out_release:
1252         /*
1253          * If we did an incomplete transfer we must release
1254          * the pipe buffers in question:
1255          */
1256         for (i = 0; i < pipe->buffers; i++) {
1257                 struct pipe_buffer *buf = pipe->bufs + i;
1258
1259                 if (buf->ops) {
1260                         buf->ops->release(pipe, buf);
1261                         buf->ops = NULL;
1262                 }
1263         }
1264
1265         if (!bytes)
1266                 bytes = ret;
1267
1268         goto done;
1269 }
1270 EXPORT_SYMBOL(splice_direct_to_actor);
1271
1272 static int direct_splice_actor(struct pipe_inode_info *pipe,
1273                                struct splice_desc *sd)
1274 {
1275         struct file *file = sd->u.file;
1276
1277         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1278                               sd->flags);
1279 }
1280
1281 /**
1282  * do_splice_direct - splices data directly between two files
1283  * @in:         file to splice from
1284  * @ppos:       input file offset
1285  * @out:        file to splice to
1286  * @len:        number of bytes to splice
1287  * @flags:      splice modifier flags
1288  *
1289  * Description:
1290  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1291  *    doing it in the application would incur an extra system call
1292  *    (splice in + splice out, as compared to just sendfile()). So this helper
1293  *    can splice directly through a process-private pipe.
1294  *
1295  */
1296 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1297                       size_t len, unsigned int flags)
1298 {
1299         struct splice_desc sd = {
1300                 .len            = len,
1301                 .total_len      = len,
1302                 .flags          = flags,
1303                 .pos            = *ppos,
1304                 .u.file         = out,
1305         };
1306         long ret;
1307
1308         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1309         if (ret > 0)
1310                 *ppos = sd.pos;
1311
1312         return ret;
1313 }
1314
1315 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1316                                struct pipe_inode_info *opipe,
1317                                size_t len, unsigned int flags);
1318
1319 /*
1320  * Determine where to splice to/from.
1321  */
1322 static long do_splice(struct file *in, loff_t __user *off_in,
1323                       struct file *out, loff_t __user *off_out,
1324                       size_t len, unsigned int flags)
1325 {
1326         struct pipe_inode_info *ipipe;
1327         struct pipe_inode_info *opipe;
1328         loff_t offset, *off;
1329         long ret;
1330
1331         ipipe = get_pipe_info(in);
1332         opipe = get_pipe_info(out);
1333
1334         if (ipipe && opipe) {
1335                 if (off_in || off_out)
1336                         return -ESPIPE;
1337
1338                 if (!(in->f_mode & FMODE_READ))
1339                         return -EBADF;
1340
1341                 if (!(out->f_mode & FMODE_WRITE))
1342                         return -EBADF;
1343
1344                 /* Splicing to self would be fun, but... */
1345                 if (ipipe == opipe)
1346                         return -EINVAL;
1347
1348                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1349         }
1350
1351         if (ipipe) {
1352                 if (off_in)
1353                         return -ESPIPE;
1354                 if (off_out) {
1355                         if (!(out->f_mode & FMODE_PWRITE))
1356                                 return -EINVAL;
1357                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1358                                 return -EFAULT;
1359                         off = &offset;
1360                 } else
1361                         off = &out->f_pos;
1362
1363                 ret = do_splice_from(ipipe, out, off, len, flags);
1364
1365                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1366                         ret = -EFAULT;
1367
1368                 return ret;
1369         }
1370
1371         if (opipe) {
1372                 if (off_out)
1373                         return -ESPIPE;
1374                 if (off_in) {
1375                         if (!(in->f_mode & FMODE_PREAD))
1376                                 return -EINVAL;
1377                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1378                                 return -EFAULT;
1379                         off = &offset;
1380                 } else
1381                         off = &in->f_pos;
1382
1383                 ret = do_splice_to(in, off, opipe, len, flags);
1384
1385                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1386                         ret = -EFAULT;
1387
1388                 return ret;
1389         }
1390
1391         return -EINVAL;
1392 }
1393
1394 /*
1395  * Map an iov into an array of pages and offset/length tupples. With the
1396  * partial_page structure, we can map several non-contiguous ranges into
1397  * our ones pages[] map instead of splitting that operation into pieces.
1398  * Could easily be exported as a generic helper for other users, in which
1399  * case one would probably want to add a 'max_nr_pages' parameter as well.
1400  */
1401 static int get_iovec_page_array(const struct iovec __user *iov,
1402                                 unsigned int nr_vecs, struct page **pages,
1403                                 struct partial_page *partial, int aligned,
1404                                 unsigned int pipe_buffers)
1405 {
1406         int buffers = 0, error = 0;
1407
1408         while (nr_vecs) {
1409                 unsigned long off, npages;
1410                 struct iovec entry;
1411                 void __user *base;
1412                 size_t len;
1413                 int i;
1414
1415                 error = -EFAULT;
1416                 if (copy_from_user(&entry, iov, sizeof(entry)))
1417                         break;
1418
1419                 base = entry.iov_base;
1420                 len = entry.iov_len;
1421
1422                 /*
1423                  * Sanity check this iovec. 0 read succeeds.
1424                  */
1425                 error = 0;
1426                 if (unlikely(!len))
1427                         break;
1428                 error = -EFAULT;
1429                 if (!access_ok(VERIFY_READ, base, len))
1430                         break;
1431
1432                 /*
1433                  * Get this base offset and number of pages, then map
1434                  * in the user pages.
1435                  */
1436                 off = (unsigned long) base & ~PAGE_MASK;
1437
1438                 /*
1439                  * If asked for alignment, the offset must be zero and the
1440                  * length a multiple of the PAGE_SIZE.
1441                  */
1442                 error = -EINVAL;
1443                 if (aligned && (off || len & ~PAGE_MASK))
1444                         break;
1445
1446                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1447                 if (npages > pipe_buffers - buffers)
1448                         npages = pipe_buffers - buffers;
1449
1450                 error = get_user_pages_fast((unsigned long)base, npages,
1451                                         0, &pages[buffers]);
1452
1453                 if (unlikely(error <= 0))
1454                         break;
1455
1456                 /*
1457                  * Fill this contiguous range into the partial page map.
1458                  */
1459                 for (i = 0; i < error; i++) {
1460                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1461
1462                         partial[buffers].offset = off;
1463                         partial[buffers].len = plen;
1464
1465                         off = 0;
1466                         len -= plen;
1467                         buffers++;
1468                 }
1469
1470                 /*
1471                  * We didn't complete this iov, stop here since it probably
1472                  * means we have to move some of this into a pipe to
1473                  * be able to continue.
1474                  */
1475                 if (len)
1476                         break;
1477
1478                 /*
1479                  * Don't continue if we mapped fewer pages than we asked for,
1480                  * or if we mapped the max number of pages that we have
1481                  * room for.
1482                  */
1483                 if (error < npages || buffers == pipe_buffers)
1484                         break;
1485
1486                 nr_vecs--;
1487                 iov++;
1488         }
1489
1490         if (buffers)
1491                 return buffers;
1492
1493         return error;
1494 }
1495
1496 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1497                         struct splice_desc *sd)
1498 {
1499         char *src;
1500         int ret;
1501
1502         /*
1503          * See if we can use the atomic maps, by prefaulting in the
1504          * pages and doing an atomic copy
1505          */
1506         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1507                 src = buf->ops->map(pipe, buf, 1);
1508                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1509                                                         sd->len);
1510                 buf->ops->unmap(pipe, buf, src);
1511                 if (!ret) {
1512                         ret = sd->len;
1513                         goto out;
1514                 }
1515         }
1516
1517         /*
1518          * No dice, use slow non-atomic map and copy
1519          */
1520         src = buf->ops->map(pipe, buf, 0);
1521
1522         ret = sd->len;
1523         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1524                 ret = -EFAULT;
1525
1526         buf->ops->unmap(pipe, buf, src);
1527 out:
1528         if (ret > 0)
1529                 sd->u.userptr += ret;
1530         return ret;
1531 }
1532
1533 /*
1534  * For lack of a better implementation, implement vmsplice() to userspace
1535  * as a simple copy of the pipes pages to the user iov.
1536  */
1537 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1538                              unsigned long nr_segs, unsigned int flags)
1539 {
1540         struct pipe_inode_info *pipe;
1541         struct splice_desc sd;
1542         ssize_t size;
1543         int error;
1544         long ret;
1545
1546         pipe = get_pipe_info(file);
1547         if (!pipe)
1548                 return -EBADF;
1549
1550         pipe_lock(pipe);
1551
1552         error = ret = 0;
1553         while (nr_segs) {
1554                 void __user *base;
1555                 size_t len;
1556
1557                 /*
1558                  * Get user address base and length for this iovec.
1559                  */
1560                 error = get_user(base, &iov->iov_base);
1561                 if (unlikely(error))
1562                         break;
1563                 error = get_user(len, &iov->iov_len);
1564                 if (unlikely(error))
1565                         break;
1566
1567                 /*
1568                  * Sanity check this iovec. 0 read succeeds.
1569                  */
1570                 if (unlikely(!len))
1571                         break;
1572                 if (unlikely(!base)) {
1573                         error = -EFAULT;
1574                         break;
1575                 }
1576
1577                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1578                         error = -EFAULT;
1579                         break;
1580                 }
1581
1582                 sd.len = 0;
1583                 sd.total_len = len;
1584                 sd.flags = flags;
1585                 sd.u.userptr = base;
1586                 sd.pos = 0;
1587
1588                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1589                 if (size < 0) {
1590                         if (!ret)
1591                                 ret = size;
1592
1593                         break;
1594                 }
1595
1596                 ret += size;
1597
1598                 if (size < len)
1599                         break;
1600
1601                 nr_segs--;
1602                 iov++;
1603         }
1604
1605         pipe_unlock(pipe);
1606
1607         if (!ret)
1608                 ret = error;
1609
1610         return ret;
1611 }
1612
1613 /*
1614  * vmsplice splices a user address range into a pipe. It can be thought of
1615  * as splice-from-memory, where the regular splice is splice-from-file (or
1616  * to file). In both cases the output is a pipe, naturally.
1617  */
1618 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1619                              unsigned long nr_segs, unsigned int flags)
1620 {
1621         struct pipe_inode_info *pipe;
1622         struct page *pages[PIPE_DEF_BUFFERS];
1623         struct partial_page partial[PIPE_DEF_BUFFERS];
1624         struct splice_pipe_desc spd = {
1625                 .pages = pages,
1626                 .partial = partial,
1627                 .nr_pages_max = PIPE_DEF_BUFFERS,
1628                 .flags = flags,
1629                 .ops = &user_page_pipe_buf_ops,
1630                 .spd_release = spd_release_page,
1631         };
1632         long ret;
1633
1634         pipe = get_pipe_info(file);
1635         if (!pipe)
1636                 return -EBADF;
1637
1638         if (splice_grow_spd(pipe, &spd))
1639                 return -ENOMEM;
1640
1641         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1642                                             spd.partial, flags & SPLICE_F_GIFT,
1643                                             spd.nr_pages_max);
1644         if (spd.nr_pages <= 0)
1645                 ret = spd.nr_pages;
1646         else
1647                 ret = splice_to_pipe(pipe, &spd);
1648
1649         splice_shrink_spd(&spd);
1650         return ret;
1651 }
1652
1653 /*
1654  * Note that vmsplice only really supports true splicing _from_ user memory
1655  * to a pipe, not the other way around. Splicing from user memory is a simple
1656  * operation that can be supported without any funky alignment restrictions
1657  * or nasty vm tricks. We simply map in the user memory and fill them into
1658  * a pipe. The reverse isn't quite as easy, though. There are two possible
1659  * solutions for that:
1660  *
1661  *      - memcpy() the data internally, at which point we might as well just
1662  *        do a regular read() on the buffer anyway.
1663  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1664  *        has restriction limitations on both ends of the pipe).
1665  *
1666  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1667  *
1668  */
1669 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1670                 unsigned long, nr_segs, unsigned int, flags)
1671 {
1672         struct file *file;
1673         long error;
1674         int fput;
1675
1676         if (unlikely(nr_segs > UIO_MAXIOV))
1677                 return -EINVAL;
1678         else if (unlikely(!nr_segs))
1679                 return 0;
1680
1681         error = -EBADF;
1682         file = fget_light(fd, &fput);
1683         if (file) {
1684                 if (file->f_mode & FMODE_WRITE)
1685                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1686                 else if (file->f_mode & FMODE_READ)
1687                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1688
1689                 fput_light(file, fput);
1690         }
1691
1692         return error;
1693 }
1694
1695 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1696                 int, fd_out, loff_t __user *, off_out,
1697                 size_t, len, unsigned int, flags)
1698 {
1699         long error;
1700         struct file *in, *out;
1701         int fput_in, fput_out;
1702
1703         if (unlikely(!len))
1704                 return 0;
1705
1706         error = -EBADF;
1707         in = fget_light(fd_in, &fput_in);
1708         if (in) {
1709                 if (in->f_mode & FMODE_READ) {
1710                         out = fget_light(fd_out, &fput_out);
1711                         if (out) {
1712                                 if (out->f_mode & FMODE_WRITE)
1713                                         error = do_splice(in, off_in,
1714                                                           out, off_out,
1715                                                           len, flags);
1716                                 fput_light(out, fput_out);
1717                         }
1718                 }
1719
1720                 fput_light(in, fput_in);
1721         }
1722
1723         return error;
1724 }
1725
1726 /*
1727  * Make sure there's data to read. Wait for input if we can, otherwise
1728  * return an appropriate error.
1729  */
1730 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1731 {
1732         int ret;
1733
1734         /*
1735          * Check ->nrbufs without the inode lock first. This function
1736          * is speculative anyways, so missing one is ok.
1737          */
1738         if (pipe->nrbufs)
1739                 return 0;
1740
1741         ret = 0;
1742         pipe_lock(pipe);
1743
1744         while (!pipe->nrbufs) {
1745                 if (signal_pending(current)) {
1746                         ret = -ERESTARTSYS;
1747                         break;
1748                 }
1749                 if (!pipe->writers)
1750                         break;
1751                 if (!pipe->waiting_writers) {
1752                         if (flags & SPLICE_F_NONBLOCK) {
1753                                 ret = -EAGAIN;
1754                                 break;
1755                         }
1756                 }
1757                 pipe_wait(pipe);
1758         }
1759
1760         pipe_unlock(pipe);
1761         return ret;
1762 }
1763
1764 /*
1765  * Make sure there's writeable room. Wait for room if we can, otherwise
1766  * return an appropriate error.
1767  */
1768 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1769 {
1770         int ret;
1771
1772         /*
1773          * Check ->nrbufs without the inode lock first. This function
1774          * is speculative anyways, so missing one is ok.
1775          */
1776         if (pipe->nrbufs < pipe->buffers)
1777                 return 0;
1778
1779         ret = 0;
1780         pipe_lock(pipe);
1781
1782         while (pipe->nrbufs >= pipe->buffers) {
1783                 if (!pipe->readers) {
1784                         send_sig(SIGPIPE, current, 0);
1785                         ret = -EPIPE;
1786                         break;
1787                 }
1788                 if (flags & SPLICE_F_NONBLOCK) {
1789                         ret = -EAGAIN;
1790                         break;
1791                 }
1792                 if (signal_pending(current)) {
1793                         ret = -ERESTARTSYS;
1794                         break;
1795                 }
1796                 pipe->waiting_writers++;
1797                 pipe_wait(pipe);
1798                 pipe->waiting_writers--;
1799         }
1800
1801         pipe_unlock(pipe);
1802         return ret;
1803 }
1804
1805 /*
1806  * Splice contents of ipipe to opipe.
1807  */
1808 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1809                                struct pipe_inode_info *opipe,
1810                                size_t len, unsigned int flags)
1811 {
1812         struct pipe_buffer *ibuf, *obuf;
1813         int ret = 0, nbuf;
1814         bool input_wakeup = false;
1815
1816
1817 retry:
1818         ret = ipipe_prep(ipipe, flags);
1819         if (ret)
1820                 return ret;
1821
1822         ret = opipe_prep(opipe, flags);
1823         if (ret)
1824                 return ret;
1825
1826         /*
1827          * Potential ABBA deadlock, work around it by ordering lock
1828          * grabbing by pipe info address. Otherwise two different processes
1829          * could deadlock (one doing tee from A -> B, the other from B -> A).
1830          */
1831         pipe_double_lock(ipipe, opipe);
1832
1833         do {
1834                 if (!opipe->readers) {
1835                         send_sig(SIGPIPE, current, 0);
1836                         if (!ret)
1837                                 ret = -EPIPE;
1838                         break;
1839                 }
1840
1841                 if (!ipipe->nrbufs && !ipipe->writers)
1842                         break;
1843
1844                 /*
1845                  * Cannot make any progress, because either the input
1846                  * pipe is empty or the output pipe is full.
1847                  */
1848                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1849                         /* Already processed some buffers, break */
1850                         if (ret)
1851                                 break;
1852
1853                         if (flags & SPLICE_F_NONBLOCK) {
1854                                 ret = -EAGAIN;
1855                                 break;
1856                         }
1857
1858                         /*
1859                          * We raced with another reader/writer and haven't
1860                          * managed to process any buffers.  A zero return
1861                          * value means EOF, so retry instead.
1862                          */
1863                         pipe_unlock(ipipe);
1864                         pipe_unlock(opipe);
1865                         goto retry;
1866                 }
1867
1868                 ibuf = ipipe->bufs + ipipe->curbuf;
1869                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1870                 obuf = opipe->bufs + nbuf;
1871
1872                 if (len >= ibuf->len) {
1873                         /*
1874                          * Simply move the whole buffer from ipipe to opipe
1875                          */
1876                         *obuf = *ibuf;
1877                         ibuf->ops = NULL;
1878                         opipe->nrbufs++;
1879                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1880                         ipipe->nrbufs--;
1881                         input_wakeup = true;
1882                 } else {
1883                         /*
1884                          * Get a reference to this pipe buffer,
1885                          * so we can copy the contents over.
1886                          */
1887                         ibuf->ops->get(ipipe, ibuf);
1888                         *obuf = *ibuf;
1889
1890                         /*
1891                          * Don't inherit the gift flag, we need to
1892                          * prevent multiple steals of this page.
1893                          */
1894                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1895
1896                         obuf->len = len;
1897                         opipe->nrbufs++;
1898                         ibuf->offset += obuf->len;
1899                         ibuf->len -= obuf->len;
1900                 }
1901                 ret += obuf->len;
1902                 len -= obuf->len;
1903         } while (len);
1904
1905         pipe_unlock(ipipe);
1906         pipe_unlock(opipe);
1907
1908         /*
1909          * If we put data in the output pipe, wakeup any potential readers.
1910          */
1911         if (ret > 0)
1912                 wakeup_pipe_readers(opipe);
1913
1914         if (input_wakeup)
1915                 wakeup_pipe_writers(ipipe);
1916
1917         return ret;
1918 }
1919
1920 /*
1921  * Link contents of ipipe to opipe.
1922  */
1923 static int link_pipe(struct pipe_inode_info *ipipe,
1924                      struct pipe_inode_info *opipe,
1925                      size_t len, unsigned int flags)
1926 {
1927         struct pipe_buffer *ibuf, *obuf;
1928         int ret = 0, i = 0, nbuf;
1929
1930         /*
1931          * Potential ABBA deadlock, work around it by ordering lock
1932          * grabbing by pipe info address. Otherwise two different processes
1933          * could deadlock (one doing tee from A -> B, the other from B -> A).
1934          */
1935         pipe_double_lock(ipipe, opipe);
1936
1937         do {
1938                 if (!opipe->readers) {
1939                         send_sig(SIGPIPE, current, 0);
1940                         if (!ret)
1941                                 ret = -EPIPE;
1942                         break;
1943                 }
1944
1945                 /*
1946                  * If we have iterated all input buffers or ran out of
1947                  * output room, break.
1948                  */
1949                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1950                         break;
1951
1952                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1953                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1954
1955                 /*
1956                  * Get a reference to this pipe buffer,
1957                  * so we can copy the contents over.
1958                  */
1959                 ibuf->ops->get(ipipe, ibuf);
1960
1961                 obuf = opipe->bufs + nbuf;
1962                 *obuf = *ibuf;
1963
1964                 /*
1965                  * Don't inherit the gift flag, we need to
1966                  * prevent multiple steals of this page.
1967                  */
1968                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1969
1970                 if (obuf->len > len)
1971                         obuf->len = len;
1972
1973                 opipe->nrbufs++;
1974                 ret += obuf->len;
1975                 len -= obuf->len;
1976                 i++;
1977         } while (len);
1978
1979         /*
1980          * return EAGAIN if we have the potential of some data in the
1981          * future, otherwise just return 0
1982          */
1983         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1984                 ret = -EAGAIN;
1985
1986         pipe_unlock(ipipe);
1987         pipe_unlock(opipe);
1988
1989         /*
1990          * If we put data in the output pipe, wakeup any potential readers.
1991          */
1992         if (ret > 0)
1993                 wakeup_pipe_readers(opipe);
1994
1995         return ret;
1996 }
1997
1998 /*
1999  * This is a tee(1) implementation that works on pipes. It doesn't copy
2000  * any data, it simply references the 'in' pages on the 'out' pipe.
2001  * The 'flags' used are the SPLICE_F_* variants, currently the only
2002  * applicable one is SPLICE_F_NONBLOCK.
2003  */
2004 static long do_tee(struct file *in, struct file *out, size_t len,
2005                    unsigned int flags)
2006 {
2007         struct pipe_inode_info *ipipe = get_pipe_info(in);
2008         struct pipe_inode_info *opipe = get_pipe_info(out);
2009         int ret = -EINVAL;
2010
2011         /*
2012          * Duplicate the contents of ipipe to opipe without actually
2013          * copying the data.
2014          */
2015         if (ipipe && opipe && ipipe != opipe) {
2016                 /*
2017                  * Keep going, unless we encounter an error. The ipipe/opipe
2018                  * ordering doesn't really matter.
2019                  */
2020                 ret = ipipe_prep(ipipe, flags);
2021                 if (!ret) {
2022                         ret = opipe_prep(opipe, flags);
2023                         if (!ret)
2024                                 ret = link_pipe(ipipe, opipe, len, flags);
2025                 }
2026         }
2027
2028         return ret;
2029 }
2030
2031 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2032 {
2033         struct file *in;
2034         int error, fput_in;
2035
2036         if (unlikely(!len))
2037                 return 0;
2038
2039         error = -EBADF;
2040         in = fget_light(fdin, &fput_in);
2041         if (in) {
2042                 if (in->f_mode & FMODE_READ) {
2043                         int fput_out;
2044                         struct file *out = fget_light(fdout, &fput_out);
2045
2046                         if (out) {
2047                                 if (out->f_mode & FMODE_WRITE)
2048                                         error = do_tee(in, out, len, flags);
2049                                 fput_light(out, fput_out);
2050                         }
2051                 }
2052                 fput_light(in, fput_in);
2053         }
2054
2055         return error;
2056 }