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
  35 /*
  36  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  37  * a vm helper function, it's already simplified quite a bit by the
  38  * addition of remove_mapping(). If success is returned, the caller may
  39  * attempt to reuse this page for another destination.
  40  */
  41 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  42                                      struct pipe_buffer *buf)
  43 {
  44         struct page *page = buf->page;
  45         struct address_space *mapping;
  46
  47         lock_page(page);
  48
  49         mapping = page_mapping(page);
  50         if (mapping) {
  51                 WARN_ON(!PageUptodate(page));
  52
  53                 /*
  54                  * At least for ext2 with nobh option, we need to wait on
  55                  * writeback completing on this page, since we'll remove it
  56                  * from the pagecache.  Otherwise truncate wont wait on the
  57                  * page, allowing the disk blocks to be reused by someone else
  58                  * before we actually wrote our data to them. fs corruption
  59                  * ensues.
  60                  */
  61                 wait_on_page_writeback(page);
  62
  63                 if (page_has_private(page) &&
  64                     !try_to_release_page(page, GFP_KERNEL))
  65                         goto out_unlock;
  66
  67                 /*
  68                  * If we succeeded in removing the mapping, set LRU flag
  69                  * and return good.
  70                  */
  71                 if (remove_mapping(mapping, page)) {
  72                         buf->flags |= PIPE_BUF_FLAG_LRU;
  73                         return 0;
  74                 }
  75         }
  76
  77         /*
  78          * Raced with truncate or failed to remove page from current
  79          * address space, unlock and return failure.
  80          */
  81 out_unlock:
  82         unlock_page(page);
  83         return 1;
  84 }
  85
  86 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  87                                         struct pipe_buffer *buf)
  88 {
  89         page_cache_release(buf->page);
  90         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  91 }
  92
  93 /*
  94  * Check whether the contents of buf is OK to access. Since the content
  95  * is a page cache page, IO may be in flight.
  96  */
  97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  98                                        struct pipe_buffer *buf)
  99 {
 100         struct page *page = buf->page;
 101         int err;
 102
 103         if (!PageUptodate(page)) {
 104                 lock_page(page);
 105
 106                 /*
 107                  * Page got truncated/unhashed. This will cause a 0-byte
 108                  * splice, if this is the first page.
 109                  */
 110                 if (!page->mapping) {
 111                         err = -ENODATA;
 112                         goto error;
 113                 }
 114
 115                 /*
 116                  * Uh oh, read-error from disk.
 117                  */
 118                 if (!PageUptodate(page)) {
 119                         err = -EIO;
 120                         goto error;
 121                 }
 122
 123                 /*
 124                  * Page is ok afterall, we are done.
 125                  */
 126                 unlock_page(page);
 127         }
 128
 129         return 0;
 130 error:
 131         unlock_page(page);
 132         return err;
 133 }
 134
 135 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 136         .can_merge = 0,
 137         .map = generic_pipe_buf_map,
 138         .unmap = generic_pipe_buf_unmap,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .map = generic_pipe_buf_map,
 158         .unmap = generic_pipe_buf_unmap,
 159         .confirm = generic_pipe_buf_confirm,
 160         .release = page_cache_pipe_buf_release,
 161         .steal = user_page_pipe_buf_steal,
 162         .get = generic_pipe_buf_get,
 163 };
 164
 165 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 166 {
 167         smp_mb();
 168         if (waitqueue_active(&pipe->wait))
 169                 wake_up_interruptible(&pipe->wait);
 170         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 171 }
 172
 173 /**
 174  * splice_to_pipe - fill passed data into a pipe
 175  * @pipe:       pipe to fill
 176  * @spd:        data to fill
 177  *
 178  * Description:
 179  *    @spd contains a map of pages and len/offset tuples, along with
 180  *    the struct pipe_buf_operations associated with these pages. This
 181  *    function will link that data to the pipe.
 182  *
 183  */
 184 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 185                        struct splice_pipe_desc *spd)
 186 {
 187         unsigned int spd_pages = spd->nr_pages;
 188         int ret, do_wakeup, page_nr;
 189
 190         ret = 0;
 191         do_wakeup = 0;
 192         page_nr = 0;
 193
 194         pipe_lock(pipe);
 195
 196         for (;;) {
 197                 if (!pipe->readers) {
 198                         send_sig(SIGPIPE, current, 0);
 199                         if (!ret)
 200                                 ret = -EPIPE;
 201                         break;
 202                 }
 203
 204                 if (pipe->nrbufs < pipe->buffers) {
 205                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 206                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 207
 208                         buf->page = spd->pages[page_nr];
 209                         buf->offset = spd->partial[page_nr].offset;
 210                         buf->len = spd->partial[page_nr].len;
 211                         buf->private = spd->partial[page_nr].private;
 212                         buf->ops = spd->ops;
 213                         if (spd->flags & SPLICE_F_GIFT)
 214                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 215
 216                         pipe->nrbufs++;
 217                         page_nr++;
 218                         ret += buf->len;
 219
 220                         if (pipe->inode)
 221                                 do_wakeup = 1;
 222
 223                         if (!--spd->nr_pages)
 224                                 break;
 225                         if (pipe->nrbufs < pipe->buffers)
 226                                 continue;
 227
 228                         break;
 229                 }
 230
 231                 if (spd->flags & SPLICE_F_NONBLOCK) {
 232                         if (!ret)
 233                                 ret = -EAGAIN;
 234                         break;
 235                 }
 236
 237                 if (signal_pending(current)) {
 238                         if (!ret)
 239                                 ret = -ERESTARTSYS;
 240                         break;
 241                 }
 242
 243                 if (do_wakeup) {
 244                         smp_mb();
 245                         if (waitqueue_active(&pipe->wait))
 246                                 wake_up_interruptible_sync(&pipe->wait);
 247                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 248                         do_wakeup = 0;
 249                 }
 250
 251                 pipe->waiting_writers++;
 252                 pipe_wait(pipe);
 253                 pipe->waiting_writers--;
 254         }
 255
 256         pipe_unlock(pipe);
 257
 258         if (do_wakeup)
 259                 wakeup_pipe_readers(pipe);
 260
 261         while (page_nr < spd_pages)
 262                 spd->spd_release(spd, page_nr++);
 263
 264         return ret;
 265 }
 266
 267 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 268 {
 269         page_cache_release(spd->pages[i]);
 270 }
 271
 272 /*
 273  * Check if we need to grow the arrays holding pages and partial page
 274  * descriptions.
 275  */
 276 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 277 {
 278         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 279                 return 0;
 280
 281         spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
 282         spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
 283
 284         if (spd->pages && spd->partial)
 285                 return 0;
 286
 287         kfree(spd->pages);
 288         kfree(spd->partial);
 289         return -ENOMEM;
 290 }
 291
 292 void splice_shrink_spd(struct pipe_inode_info *pipe,
 293                        struct splice_pipe_desc *spd)
 294 {
 295         if (pipe->buffers <= PIPE_DEF_BUFFERS)
 296                 return;
 297
 298         kfree(spd->pages);
 299         kfree(spd->partial);
 300 }
 301
 302 static int
 303 __generic_file_splice_read(struct file *in, loff_t *ppos,
 304                            struct pipe_inode_info *pipe, size_t len,
 305                            unsigned int flags)
 306 {
 307         struct address_space *mapping = in->f_mapping;
 308         unsigned int loff, nr_pages, req_pages;
 309         struct page *pages[PIPE_DEF_BUFFERS];
 310         struct partial_page partial[PIPE_DEF_BUFFERS];
 311         struct page *page;
 312         pgoff_t index, end_index;
 313         loff_t isize;
 314         int error, page_nr;
 315         struct splice_pipe_desc spd = {
 316                 .pages = pages,
 317                 .partial = partial,
 318                 .flags = flags,
 319                 .ops = &page_cache_pipe_buf_ops,
 320                 .spd_release = spd_release_page,
 321         };
 322
 323         if (splice_grow_spd(pipe, &spd))
 324                 return -ENOMEM;
 325
 326         index = *ppos >> PAGE_CACHE_SHIFT;
 327         loff = *ppos & ~PAGE_CACHE_MASK;
 328         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 329         nr_pages = min(req_pages, pipe->buffers);
 330
 331         /*
 332          * Lookup the (hopefully) full range of pages we need.
 333          */
 334         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 335         index += spd.nr_pages;
 336
 337         /*
 338          * If find_get_pages_contig() returned fewer pages than we needed,
 339          * readahead/allocate the rest and fill in the holes.
 340          */
 341         if (spd.nr_pages < nr_pages)
 342                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 343                                 index, req_pages - spd.nr_pages);
 344
 345         error = 0;
 346         while (spd.nr_pages < nr_pages) {
 347                 /*
 348                  * Page could be there, find_get_pages_contig() breaks on
 349                  * the first hole.
 350                  */
 351                 page = find_get_page(mapping, index);
 352                 if (!page) {
 353                         /*
 354                          * page didn't exist, allocate one.
 355                          */
 356                         page = page_cache_alloc_cold(mapping);
 357                         if (!page)
 358                                 break;
 359
 360                         error = add_to_page_cache_lru(page, mapping, index,
 361                                                 GFP_KERNEL);
 362                         if (unlikely(error)) {
 363                                 page_cache_release(page);
 364                                 if (error == -EEXIST)
 365                                         continue;
 366                                 break;
 367                         }
 368                         /*
 369                          * add_to_page_cache() locks the page, unlock it
 370                          * to avoid convoluting the logic below even more.
 371                          */
 372                         unlock_page(page);
 373                 }
 374
 375                 spd.pages[spd.nr_pages++] = page;
 376                 index++;
 377         }
 378
 379         /*
 380          * Now loop over the map and see if we need to start IO on any
 381          * pages, fill in the partial map, etc.
 382          */
 383         index = *ppos >> PAGE_CACHE_SHIFT;
 384         nr_pages = spd.nr_pages;
 385         spd.nr_pages = 0;
 386         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 387                 unsigned int this_len;
 388
 389                 if (!len)
 390                         break;
 391
 392                 /*
 393                  * this_len is the max we'll use from this page
 394                  */
 395                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 396                 page = spd.pages[page_nr];
 397
 398                 if (PageReadahead(page))
 399                         page_cache_async_readahead(mapping, &in->f_ra, in,
 400                                         page, index, req_pages - page_nr);
 401
 402                 /*
 403                  * If the page isn't uptodate, we may need to start io on it
 404                  */
 405                 if (!PageUptodate(page)) {
 406                         lock_page(page);
 407
 408                         /*
 409                          * Page was truncated, or invalidated by the
 410                          * filesystem.  Redo the find/create, but this time the
 411                          * page is kept locked, so there's no chance of another
 412                          * race with truncate/invalidate.
 413                          */
 414                         if (!page->mapping) {
 415                                 unlock_page(page);
 416                                 page = find_or_create_page(mapping, index,
 417                                                 mapping_gfp_mask(mapping));
 418
 419                                 if (!page) {
 420                                         error = -ENOMEM;
 421                                         break;
 422                                 }
 423                                 page_cache_release(spd.pages[page_nr]);
 424                                 spd.pages[page_nr] = page;
 425                         }
 426                         /*
 427                          * page was already under io and is now done, great
 428                          */
 429                         if (PageUptodate(page)) {
 430                                 unlock_page(page);
 431                                 goto fill_it;
 432                         }
 433
 434                         /*
 435                          * need to read in the page
 436                          */
 437                         error = mapping->a_ops->readpage(in, page);
 438                         if (unlikely(error)) {
 439                                 /*
 440                                  * We really should re-lookup the page here,
 441                                  * but it complicates things a lot. Instead
 442                                  * lets just do what we already stored, and
 443                                  * we'll get it the next time we are called.
 444                                  */
 445                                 if (error == AOP_TRUNCATED_PAGE)
 446                                         error = 0;
 447
 448                                 break;
 449                         }
 450                 }
 451 fill_it:
 452                 /*
 453                  * i_size must be checked after PageUptodate.
 454                  */
 455                 isize = i_size_read(mapping->host);
 456                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 457                 if (unlikely(!isize || index > end_index))
 458                         break;
 459
 460                 /*
 461                  * if this is the last page, see if we need to shrink
 462                  * the length and stop
 463                  */
 464                 if (end_index == index) {
 465                         unsigned int plen;
 466
 467                         /*
 468                          * max good bytes in this page
 469                          */
 470                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 471                         if (plen <= loff)
 472                                 break;
 473
 474                         /*
 475                          * force quit after adding this page
 476                          */
 477                         this_len = min(this_len, plen - loff);
 478                         len = this_len;
 479                 }
 480
 481                 spd.partial[page_nr].offset = loff;
 482                 spd.partial[page_nr].len = this_len;
 483                 len -= this_len;
 484                 loff = 0;
 485                 spd.nr_pages++;
 486                 index++;
 487         }
 488
 489         /*
 490          * Release any pages at the end, if we quit early. 'page_nr' is how far
 491          * we got, 'nr_pages' is how many pages are in the map.
 492          */
 493         while (page_nr < nr_pages)
 494                 page_cache_release(spd.pages[page_nr++]);
 495         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 496
 497         if (spd.nr_pages)
 498                 error = splice_to_pipe(pipe, &spd);
 499
 500         splice_shrink_spd(pipe, &spd);
 501         return error;
 502 }
 503
 504 /**
 505  * generic_file_splice_read - splice data from file to a pipe
 506  * @in:         file to splice from
 507  * @ppos:       position in @in
 508  * @pipe:       pipe to splice to
 509  * @len:        number of bytes to splice
 510  * @flags:      splice modifier flags
 511  *
 512  * Description:
 513  *    Will read pages from given file and fill them into a pipe. Can be
 514  *    used as long as the address_space operations for the source implements
 515  *    a readpage() hook.
 516  *
 517  */
 518 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 519                                  struct pipe_inode_info *pipe, size_t len,
 520                                  unsigned int flags)
 521 {
 522         loff_t isize, left;
 523         int ret;
 524
 525         isize = i_size_read(in->f_mapping->host);
 526         if (unlikely(*ppos >= isize))
 527                 return 0;
 528
 529         left = isize - *ppos;
 530         if (unlikely(left < len))
 531                 len = left;
 532
 533         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 534         if (ret > 0) {
 535                 *ppos += ret;
 536                 file_accessed(in);
 537         }
 538
 539         return ret;
 540 }
 541 EXPORT_SYMBOL(generic_file_splice_read);
 542
 543 static const struct pipe_buf_operations default_pipe_buf_ops = {
 544         .can_merge = 0,
 545         .map = generic_pipe_buf_map,
 546         .unmap = generic_pipe_buf_unmap,
 547         .confirm = generic_pipe_buf_confirm,
 548         .release = generic_pipe_buf_release,
 549         .steal = generic_pipe_buf_steal,
 550         .get = generic_pipe_buf_get,
 551 };
 552
 553 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 554                             unsigned long vlen, loff_t offset)
 555 {
 556         mm_segment_t old_fs;
 557         loff_t pos = offset;
 558         ssize_t res;
 559
 560         old_fs = get_fs();
 561         set_fs(get_ds());
 562         /* The cast to a user pointer is valid due to the set_fs() */
 563         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 564         set_fs(old_fs);
 565
 566         return res;
 567 }
 568
 569 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 570                             loff_t pos)
 571 {
 572         mm_segment_t old_fs;
 573         ssize_t res;
 574
 575         old_fs = get_fs();
 576         set_fs(get_ds());
 577         /* The cast to a user pointer is valid due to the set_fs() */
 578         res = vfs_write(file, (const char __user *)buf, count, &pos);
 579         set_fs(old_fs);
 580
 581         return res;
 582 }
 583
 584 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 585                                  struct pipe_inode_info *pipe, size_t len,
 586                                  unsigned int flags)
 587 {
 588         unsigned int nr_pages;
 589         unsigned int nr_freed;
 590         size_t offset;
 591         struct page *pages[PIPE_DEF_BUFFERS];
 592         struct partial_page partial[PIPE_DEF_BUFFERS];
 593         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 594         ssize_t res;
 595         size_t this_len;
 596         int error;
 597         int i;
 598         struct splice_pipe_desc spd = {
 599                 .pages = pages,
 600                 .partial = partial,
 601                 .flags = flags,
 602                 .ops = &default_pipe_buf_ops,
 603                 .spd_release = spd_release_page,
 604         };
 605
 606         if (splice_grow_spd(pipe, &spd))
 607                 return -ENOMEM;
 608
 609         res = -ENOMEM;
 610         vec = __vec;
 611         if (pipe->buffers > PIPE_DEF_BUFFERS) {
 612                 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
 613                 if (!vec)
 614                         goto shrink_ret;
 615         }
 616
 617         offset = *ppos & ~PAGE_CACHE_MASK;
 618         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 619
 620         for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
 621                 struct page *page;
 622
 623                 page = alloc_page(GFP_USER);
 624                 error = -ENOMEM;
 625                 if (!page)
 626                         goto err;
 627
 628                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 629                 vec[i].iov_base = (void __user *) page_address(page);
 630                 vec[i].iov_len = this_len;
 631                 spd.pages[i] = page;
 632                 spd.nr_pages++;
 633                 len -= this_len;
 634                 offset = 0;
 635         }
 636
 637         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 638         if (res < 0) {
 639                 error = res;
 640                 goto err;
 641         }
 642
 643         error = 0;
 644         if (!res)
 645                 goto err;
 646
 647         nr_freed = 0;
 648         for (i = 0; i < spd.nr_pages; i++) {
 649                 this_len = min_t(size_t, vec[i].iov_len, res);
 650                 spd.partial[i].offset = 0;
 651                 spd.partial[i].len = this_len;
 652                 if (!this_len) {
 653                         __free_page(spd.pages[i]);
 654                         spd.pages[i] = NULL;
 655                         nr_freed++;
 656                 }
 657                 res -= this_len;
 658         }
 659         spd.nr_pages -= nr_freed;
 660
 661         res = splice_to_pipe(pipe, &spd);
 662         if (res > 0)
 663                 *ppos += res;
 664
 665 shrink_ret:
 666         if (vec != __vec)
 667                 kfree(vec);
 668         splice_shrink_spd(pipe, &spd);
 669         return res;
 670
 671 err:
 672         for (i = 0; i < spd.nr_pages; i++)
 673                 __free_page(spd.pages[i]);
 674
 675         res = error;
 676         goto shrink_ret;
 677 }
 678 EXPORT_SYMBOL(default_file_splice_read);
 679
 680 /*
 681  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 682  * using sendpage(). Return the number of bytes sent.
 683  */
 684 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 685                             struct pipe_buffer *buf, struct splice_desc *sd)
 686 {
 687         struct file *file = sd->u.file;
 688         loff_t pos = sd->pos;
 689         int more;
 690
 691         if (!likely(file->f_op && file->f_op->sendpage))
 692                 return -EINVAL;
 693
 694         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
 695         return file->f_op->sendpage(file, buf->page, buf->offset,
 696                                     sd->len, &pos, more);
 697 }
 698
 699 /*
 700  * This is a little more tricky than the file -> pipe splicing. There are
 701  * basically three cases:
 702  *
 703  *      - Destination page already exists in the address space and there
 704  *        are users of it. For that case we have no other option that
 705  *        copying the data. Tough luck.
 706  *      - Destination page already exists in the address space, but there
 707  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 708  *        through to last case.
 709  *      - Destination page does not exist, we can add the pipe page to
 710  *        the page cache and avoid the copy.
 711  *
 712  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 713  * sd->flags), we attempt to migrate pages from the pipe to the output
 714  * file address space page cache. This is possible if no one else has
 715  * the pipe page referenced outside of the pipe and page cache. If
 716  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 717  * a new page in the output file page cache and fill/dirty that.
 718  */
 719 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 720                  struct splice_desc *sd)
 721 {
 722         struct file *file = sd->u.file;
 723         struct address_space *mapping = file->f_mapping;
 724         unsigned int offset, this_len;
 725         struct page *page;
 726         void *fsdata;
 727         int ret;
 728
 729         offset = sd->pos & ~PAGE_CACHE_MASK;
 730
 731         this_len = sd->len;
 732         if (this_len + offset > PAGE_CACHE_SIZE)
 733                 this_len = PAGE_CACHE_SIZE - offset;
 734
 735         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 736                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 737         if (unlikely(ret))
 738                 goto out;
 739
 740         if (buf->page != page) {
 741                 /*
 742                  * Careful, ->map() uses KM_USER0!
 743                  */
 744                 char *src = buf->ops->map(pipe, buf, 1);
 745                 char *dst = kmap_atomic(page, KM_USER1);
 746
 747                 memcpy(dst + offset, src + buf->offset, this_len);
 748                 flush_dcache_page(page);
 749                 kunmap_atomic(dst, KM_USER1);
 750                 buf->ops->unmap(pipe, buf, src);
 751         }
 752         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 753                                 page, fsdata);
 754 out:
 755         return ret;
 756 }
 757 EXPORT_SYMBOL(pipe_to_file);
 758
 759 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 760 {
 761         smp_mb();
 762         if (waitqueue_active(&pipe->wait))
 763                 wake_up_interruptible(&pipe->wait);
 764         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 765 }
 766
 767 /**
 768  * splice_from_pipe_feed - feed available data from a pipe to a file
 769  * @pipe:       pipe to splice from
 770  * @sd:         information to @actor
 771  * @actor:      handler that splices the data
 772  *
 773  * Description:
 774  *    This function loops over the pipe and calls @actor to do the
 775  *    actual moving of a single struct pipe_buffer to the desired
 776  *    destination.  It returns when there's no more buffers left in
 777  *    the pipe or if the requested number of bytes (@sd->total_len)
 778  *    have been copied.  It returns a positive number (one) if the
 779  *    pipe needs to be filled with more data, zero if the required
 780  *    number of bytes have been copied and -errno on error.
 781  *
 782  *    This, together with splice_from_pipe_{begin,end,next}, may be
 783  *    used to implement the functionality of __splice_from_pipe() when
 784  *    locking is required around copying the pipe buffers to the
 785  *    destination.
 786  */
 787 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 788                           splice_actor *actor)
 789 {
 790         int ret;
 791
 792         while (pipe->nrbufs) {
 793                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 794                 const struct pipe_buf_operations *ops = buf->ops;
 795
 796                 sd->len = buf->len;
 797                 if (sd->len > sd->total_len)
 798                         sd->len = sd->total_len;
 799
 800                 ret = buf->ops->confirm(pipe, buf);
 801                 if (unlikely(ret)) {
 802                         if (ret == -ENODATA)
 803                                 ret = 0;
 804                         return ret;
 805                 }
 806
 807                 ret = actor(pipe, buf, sd);
 808                 if (ret <= 0)
 809                         return ret;
 810
 811                 buf->offset += ret;
 812                 buf->len -= ret;
 813
 814                 sd->num_spliced += ret;
 815                 sd->len -= ret;
 816                 sd->pos += ret;
 817                 sd->total_len -= ret;
 818
 819                 if (!buf->len) {
 820                         buf->ops = NULL;
 821                         ops->release(pipe, buf);
 822                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 823                         pipe->nrbufs--;
 824                         if (pipe->inode)
 825                                 sd->need_wakeup = true;
 826                 }
 827
 828                 if (!sd->total_len)
 829                         return 0;
 830         }
 831
 832         return 1;
 833 }
 834 EXPORT_SYMBOL(splice_from_pipe_feed);
 835
 836 /**
 837  * splice_from_pipe_next - wait for some data to splice from
 838  * @pipe:       pipe to splice from
 839  * @sd:         information about the splice operation
 840  *
 841  * Description:
 842  *    This function will wait for some data and return a positive
 843  *    value (one) if pipe buffers are available.  It will return zero
 844  *    or -errno if no more data needs to be spliced.
 845  */
 846 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 847 {
 848         while (!pipe->nrbufs) {
 849                 if (!pipe->writers)
 850                         return 0;
 851
 852                 if (!pipe->waiting_writers && sd->num_spliced)
 853                         return 0;
 854
 855                 if (sd->flags & SPLICE_F_NONBLOCK)
 856                         return -EAGAIN;
 857
 858                 if (signal_pending(current))
 859                         return -ERESTARTSYS;
 860
 861                 if (sd->need_wakeup) {
 862                         wakeup_pipe_writers(pipe);
 863                         sd->need_wakeup = false;
 864                 }
 865
 866                 pipe_wait(pipe);
 867         }
 868
 869         return 1;
 870 }
 871 EXPORT_SYMBOL(splice_from_pipe_next);
 872
 873 /**
 874  * splice_from_pipe_begin - start splicing from pipe
 875  * @sd:         information about the splice operation
 876  *
 877  * Description:
 878  *    This function should be called before a loop containing
 879  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 880  *    initialize the necessary fields of @sd.
 881  */
 882 void splice_from_pipe_begin(struct splice_desc *sd)
 883 {
 884         sd->num_spliced = 0;
 885         sd->need_wakeup = false;
 886 }
 887 EXPORT_SYMBOL(splice_from_pipe_begin);
 888
 889 /**
 890  * splice_from_pipe_end - finish splicing from pipe
 891  * @pipe:       pipe to splice from
 892  * @sd:         information about the splice operation
 893  *
 894  * Description:
 895  *    This function will wake up pipe writers if necessary.  It should
 896  *    be called after a loop containing splice_from_pipe_next() and
 897  *    splice_from_pipe_feed().
 898  */
 899 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 900 {
 901         if (sd->need_wakeup)
 902                 wakeup_pipe_writers(pipe);
 903 }
 904 EXPORT_SYMBOL(splice_from_pipe_end);
 905
 906 /**
 907  * __splice_from_pipe - splice data from a pipe to given actor
 908  * @pipe:       pipe to splice from
 909  * @sd:         information to @actor
 910  * @actor:      handler that splices the data
 911  *
 912  * Description:
 913  *    This function does little more than loop over the pipe and call
 914  *    @actor to do the actual moving of a single struct pipe_buffer to
 915  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 916  *    pipe_to_user.
 917  *
 918  */
 919 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 920                            splice_actor *actor)
 921 {
 922         int ret;
 923
 924         splice_from_pipe_begin(sd);
 925         do {
 926                 ret = splice_from_pipe_next(pipe, sd);
 927                 if (ret > 0)
 928                         ret = splice_from_pipe_feed(pipe, sd, actor);
 929         } while (ret > 0);
 930         splice_from_pipe_end(pipe, sd);
 931
 932         return sd->num_spliced ? sd->num_spliced : ret;
 933 }
 934 EXPORT_SYMBOL(__splice_from_pipe);
 935
 936 /**
 937  * splice_from_pipe - splice data from a pipe to a file
 938  * @pipe:       pipe to splice from
 939  * @out:        file to splice to
 940  * @ppos:       position in @out
 941  * @len:        how many bytes to splice
 942  * @flags:      splice modifier flags
 943  * @actor:      handler that splices the data
 944  *
 945  * Description:
 946  *    See __splice_from_pipe. This function locks the pipe inode,
 947  *    otherwise it's identical to __splice_from_pipe().
 948  *
 949  */
 950 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 951                          loff_t *ppos, size_t len, unsigned int flags,
 952                          splice_actor *actor)
 953 {
 954         ssize_t ret;
 955         struct splice_desc sd = {
 956                 .total_len = len,
 957                 .flags = flags,
 958                 .pos = *ppos,
 959                 .u.file = out,
 960         };
 961
 962         pipe_lock(pipe);
 963         ret = __splice_from_pipe(pipe, &sd, actor);
 964         pipe_unlock(pipe);
 965
 966         return ret;
 967 }
 968
 969 /**
 970  * generic_file_splice_write - splice data from a pipe to a file
 971  * @pipe:       pipe info
 972  * @out:        file to write to
 973  * @ppos:       position in @out
 974  * @len:        number of bytes to splice
 975  * @flags:      splice modifier flags
 976  *
 977  * Description:
 978  *    Will either move or copy pages (determined by @flags options) from
 979  *    the given pipe inode to the given file.
 980  *
 981  */
 982 ssize_t
 983 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 984                           loff_t *ppos, size_t len, unsigned int flags)
 985 {
 986         struct address_space *mapping = out->f_mapping;
 987         struct inode *inode = mapping->host;
 988         struct splice_desc sd = {
 989                 .total_len = len,
 990                 .flags = flags,
 991                 .pos = *ppos,
 992                 .u.file = out,
 993         };
 994         ssize_t ret;
 995
 996         pipe_lock(pipe);
 997
 998         splice_from_pipe_begin(&sd);
 999         do {
1000                 ret = splice_from_pipe_next(pipe, &sd);
1001                 if (ret <= 0)
1002                         break;
1003
1004                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1005                 ret = file_remove_suid(out);
1006                 if (!ret) {
1007                         file_update_time(out);
1008                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1009                 }
1010                 mutex_unlock(&inode->i_mutex);
1011         } while (ret > 0);
1012         splice_from_pipe_end(pipe, &sd);
1013
1014         pipe_unlock(pipe);
1015
1016         if (sd.num_spliced)
1017                 ret = sd.num_spliced;
1018
1019         if (ret > 0) {
1020                 unsigned long nr_pages;
1021                 int err;
1022
1023                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1024
1025                 err = generic_write_sync(out, *ppos, ret);
1026                 if (err)
1027                         ret = err;
1028                 else
1029                         *ppos += ret;
1030                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1031         }
1032
1033         return ret;
1034 }
1035
1036 EXPORT_SYMBOL(generic_file_splice_write);
1037
1038 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1039                           struct splice_desc *sd)
1040 {
1041         int ret;
1042         void *data;
1043
1044         data = buf->ops->map(pipe, buf, 0);
1045         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1046         buf->ops->unmap(pipe, buf, data);
1047
1048         return ret;
1049 }
1050
1051 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1052                                          struct file *out, loff_t *ppos,
1053                                          size_t len, unsigned int flags)
1054 {
1055         ssize_t ret;
1056
1057         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1058         if (ret > 0)
1059                 *ppos += ret;
1060
1061         return ret;
1062 }
1063
1064 /**
1065  * generic_splice_sendpage - splice data from a pipe to a socket
1066  * @pipe:       pipe to splice from
1067  * @out:        socket to write to
1068  * @ppos:       position in @out
1069  * @len:        number of bytes to splice
1070  * @flags:      splice modifier flags
1071  *
1072  * Description:
1073  *    Will send @len bytes from the pipe to a network socket. No data copying
1074  *    is involved.
1075  *
1076  */
1077 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1078                                 loff_t *ppos, size_t len, unsigned int flags)
1079 {
1080         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1081 }
1082
1083 EXPORT_SYMBOL(generic_splice_sendpage);
1084
1085 /*
1086  * Attempt to initiate a splice from pipe to file.
1087  */
1088 long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1089                     loff_t *ppos, size_t len, unsigned int flags)
1090 {
1091         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1092                                 loff_t *, size_t, unsigned int);
1093         int ret;
1094
1095         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1096                 return -EBADF;
1097
1098         if (unlikely(out->f_flags & O_APPEND))
1099                 return -EINVAL;
1100
1101         ret = rw_verify_area(WRITE, out, ppos, len);
1102         if (unlikely(ret < 0))
1103                 return ret;
1104
1105         if (out->f_op && out->f_op->splice_write)
1106                 splice_write = out->f_op->splice_write;
1107         else
1108                 splice_write = default_file_splice_write;
1109
1110         return splice_write(pipe, out, ppos, len, flags);
1111 }
1112 EXPORT_SYMBOL(do_splice_from);
1113
1114 /*
1115  * Attempt to initiate a splice from a file to a pipe.
1116  */
1117 long do_splice_to(struct file *in, loff_t *ppos,
1118                   struct pipe_inode_info *pipe, size_t len,
1119                   unsigned int flags)
1120 {
1121         ssize_t (*splice_read)(struct file *, loff_t *,
1122                                struct pipe_inode_info *, size_t, unsigned int);
1123         int ret;
1124
1125         if (unlikely(!(in->f_mode & FMODE_READ)))
1126                 return -EBADF;
1127
1128         ret = rw_verify_area(READ, in, ppos, len);
1129         if (unlikely(ret < 0))
1130                 return ret;
1131
1132         if (in->f_op && in->f_op->splice_read)
1133                 splice_read = in->f_op->splice_read;
1134         else
1135                 splice_read = default_file_splice_read;
1136
1137         return splice_read(in, ppos, pipe, len, flags);
1138 }
1139 EXPORT_SYMBOL(do_splice_to);
1140
1141 /**
1142  * splice_direct_to_actor - splices data directly between two non-pipes
1143  * @in:         file to splice from
1144  * @sd:         actor information on where to splice to
1145  * @actor:      handles the data splicing
1146  *
1147  * Description:
1148  *    This is a special case helper to splice directly between two
1149  *    points, without requiring an explicit pipe. Internally an allocated
1150  *    pipe is cached in the process, and reused during the lifetime of
1151  *    that process.
1152  *
1153  */
1154 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1155                                splice_direct_actor *actor)
1156 {
1157         struct pipe_inode_info *pipe;
1158         long ret, bytes;
1159         umode_t i_mode;
1160         size_t len;
1161         int i, flags;
1162
1163         /*
1164          * We require the input being a regular file, as we don't want to
1165          * randomly drop data for eg socket -> socket splicing. Use the
1166          * piped splicing for that!
1167          */
1168         i_mode = in->f_path.dentry->d_inode->i_mode;
1169         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1170                 return -EINVAL;
1171
1172         /*
1173          * neither in nor out is a pipe, setup an internal pipe attached to
1174          * 'out' and transfer the wanted data from 'in' to 'out' through that
1175          */
1176         pipe = current->splice_pipe;
1177         if (unlikely(!pipe)) {
1178                 pipe = alloc_pipe_info(NULL);
1179                 if (!pipe)
1180                         return -ENOMEM;
1181
1182                 /*
1183                  * We don't have an immediate reader, but we'll read the stuff
1184                  * out of the pipe right after the splice_to_pipe(). So set
1185                  * PIPE_READERS appropriately.
1186                  */
1187                 pipe->readers = 1;
1188
1189                 current->splice_pipe = pipe;
1190         }
1191
1192         /*
1193          * Do the splice.
1194          */
1195         ret = 0;
1196         bytes = 0;
1197         len = sd->total_len;
1198         flags = sd->flags;
1199
1200         /*
1201          * Don't block on output, we have to drain the direct pipe.
1202          */
1203         sd->flags &= ~SPLICE_F_NONBLOCK;
1204
1205         while (len) {
1206                 size_t read_len;
1207                 loff_t pos = sd->pos, prev_pos = pos;
1208
1209                 ret = do_splice_to(in, &pos, pipe, len, flags);
1210                 if (unlikely(ret <= 0))
1211                         goto out_release;
1212
1213                 read_len = ret;
1214                 sd->total_len = read_len;
1215
1216                 /*
1217                  * NOTE: nonblocking mode only applies to the input. We
1218                  * must not do the output in nonblocking mode as then we
1219                  * could get stuck data in the internal pipe:
1220                  */
1221                 ret = actor(pipe, sd);
1222                 if (unlikely(ret <= 0)) {
1223                         sd->pos = prev_pos;
1224                         goto out_release;
1225                 }
1226
1227                 bytes += ret;
1228                 len -= ret;
1229                 sd->pos = pos;
1230
1231                 if (ret < read_len) {
1232                         sd->pos = prev_pos + ret;
1233                         goto out_release;
1234                 }
1235         }
1236
1237 done:
1238         pipe->nrbufs = pipe->curbuf = 0;
1239         file_accessed(in);
1240         return bytes;
1241
1242 out_release:
1243         /*
1244          * If we did an incomplete transfer we must release
1245          * the pipe buffers in question:
1246          */
1247         for (i = 0; i < pipe->buffers; i++) {
1248                 struct pipe_buffer *buf = pipe->bufs + i;
1249
1250                 if (buf->ops) {
1251                         buf->ops->release(pipe, buf);
1252                         buf->ops = NULL;
1253                 }
1254         }
1255
1256         if (!bytes)
1257                 bytes = ret;
1258
1259         goto done;
1260 }
1261 EXPORT_SYMBOL(splice_direct_to_actor);
1262
1263 static int direct_splice_actor(struct pipe_inode_info *pipe,
1264                                struct splice_desc *sd)
1265 {
1266         struct file *file = sd->u.file;
1267
1268         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1269                               sd->flags);
1270 }
1271
1272 /**
1273  * do_splice_direct - splices data directly between two files
1274  * @in:         file to splice from
1275  * @ppos:       input file offset
1276  * @out:        file to splice to
1277  * @len:        number of bytes to splice
1278  * @flags:      splice modifier flags
1279  *
1280  * Description:
1281  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1282  *    doing it in the application would incur an extra system call
1283  *    (splice in + splice out, as compared to just sendfile()). So this helper
1284  *    can splice directly through a process-private pipe.
1285  *
1286  */
1287 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1288                       size_t len, unsigned int flags)
1289 {
1290         struct splice_desc sd = {
1291                 .len            = len,
1292                 .total_len      = len,
1293                 .flags          = flags,
1294                 .pos            = *ppos,
1295                 .u.file         = out,
1296         };
1297         long ret;
1298
1299         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1300         if (ret > 0)
1301                 *ppos = sd.pos;
1302
1303         return ret;
1304 }
1305
1306 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1307                                struct pipe_inode_info *opipe,
1308                                size_t len, unsigned int flags);
1309
1310 /*
1311  * Determine where to splice to/from.
1312  */
1313 static long do_splice(struct file *in, loff_t __user *off_in,
1314                       struct file *out, loff_t __user *off_out,
1315                       size_t len, unsigned int flags)
1316 {
1317         struct pipe_inode_info *ipipe;
1318         struct pipe_inode_info *opipe;
1319         loff_t offset, *off;
1320         long ret;
1321
1322         ipipe = get_pipe_info(in);
1323         opipe = get_pipe_info(out);
1324
1325         if (ipipe && opipe) {
1326                 if (off_in || off_out)
1327                         return -ESPIPE;
1328
1329                 if (!(in->f_mode & FMODE_READ))
1330                         return -EBADF;
1331
1332                 if (!(out->f_mode & FMODE_WRITE))
1333                         return -EBADF;
1334
1335                 /* Splicing to self would be fun, but... */
1336                 if (ipipe == opipe)
1337                         return -EINVAL;
1338
1339                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1340         }
1341
1342         if (ipipe) {
1343                 if (off_in)
1344                         return -ESPIPE;
1345                 if (off_out) {
1346                         if (!(out->f_mode & FMODE_PWRITE))
1347                                 return -EINVAL;
1348                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1349                                 return -EFAULT;
1350                         off = &offset;
1351                 } else
1352                         off = &out->f_pos;
1353
1354                 ret = do_splice_from(ipipe, out, off, len, flags);
1355
1356                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1357                         ret = -EFAULT;
1358
1359                 return ret;
1360         }
1361
1362         if (opipe) {
1363                 if (off_out)
1364                         return -ESPIPE;
1365                 if (off_in) {
1366                         if (!(in->f_mode & FMODE_PREAD))
1367                                 return -EINVAL;
1368                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1369                                 return -EFAULT;
1370                         off = &offset;
1371                 } else
1372                         off = &in->f_pos;
1373
1374                 ret = do_splice_to(in, off, opipe, len, flags);
1375
1376                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1377                         ret = -EFAULT;
1378
1379                 return ret;
1380         }
1381
1382         return -EINVAL;
1383 }
1384
1385 /*
1386  * Map an iov into an array of pages and offset/length tupples. With the
1387  * partial_page structure, we can map several non-contiguous ranges into
1388  * our ones pages[] map instead of splitting that operation into pieces.
1389  * Could easily be exported as a generic helper for other users, in which
1390  * case one would probably want to add a 'max_nr_pages' parameter as well.
1391  */
1392 static int get_iovec_page_array(const struct iovec __user *iov,
1393                                 unsigned int nr_vecs, struct page **pages,
1394                                 struct partial_page *partial, int aligned,
1395                                 unsigned int pipe_buffers)
1396 {
1397         int buffers = 0, error = 0;
1398
1399         while (nr_vecs) {
1400                 unsigned long off, npages;
1401                 struct iovec entry;
1402                 void __user *base;
1403                 size_t len;
1404                 int i;
1405
1406                 error = -EFAULT;
1407                 if (copy_from_user(&entry, iov, sizeof(entry)))
1408                         break;
1409
1410                 base = entry.iov_base;
1411                 len = entry.iov_len;
1412
1413                 /*
1414                  * Sanity check this iovec. 0 read succeeds.
1415                  */
1416                 error = 0;
1417                 if (unlikely(!len))
1418                         break;
1419                 error = -EFAULT;
1420                 if (!access_ok(VERIFY_READ, base, len))
1421                         break;
1422
1423                 /*
1424                  * Get this base offset and number of pages, then map
1425                  * in the user pages.
1426                  */
1427                 off = (unsigned long) base & ~PAGE_MASK;
1428
1429                 /*
1430                  * If asked for alignment, the offset must be zero and the
1431                  * length a multiple of the PAGE_SIZE.
1432                  */
1433                 error = -EINVAL;
1434                 if (aligned && (off || len & ~PAGE_MASK))
1435                         break;
1436
1437                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1438                 if (npages > pipe_buffers - buffers)
1439                         npages = pipe_buffers - buffers;
1440
1441                 error = get_user_pages_fast((unsigned long)base, npages,
1442                                         0, &pages[buffers]);
1443
1444                 if (unlikely(error <= 0))
1445                         break;
1446
1447                 /*
1448                  * Fill this contiguous range into the partial page map.
1449                  */
1450                 for (i = 0; i < error; i++) {
1451                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1452
1453                         partial[buffers].offset = off;
1454                         partial[buffers].len = plen;
1455
1456                         off = 0;
1457                         len -= plen;
1458                         buffers++;
1459                 }
1460
1461                 /*
1462                  * We didn't complete this iov, stop here since it probably
1463                  * means we have to move some of this into a pipe to
1464                  * be able to continue.
1465                  */
1466                 if (len)
1467                         break;
1468
1469                 /*
1470                  * Don't continue if we mapped fewer pages than we asked for,
1471                  * or if we mapped the max number of pages that we have
1472                  * room for.
1473                  */
1474                 if (error < npages || buffers == pipe_buffers)
1475                         break;
1476
1477                 nr_vecs--;
1478                 iov++;
1479         }
1480
1481         if (buffers)
1482                 return buffers;
1483
1484         return error;
1485 }
1486
1487 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1488                         struct splice_desc *sd)
1489 {
1490         char *src;
1491         int ret;
1492
1493         /*
1494          * See if we can use the atomic maps, by prefaulting in the
1495          * pages and doing an atomic copy
1496          */
1497         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1498                 src = buf->ops->map(pipe, buf, 1);
1499                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1500                                                         sd->len);
1501                 buf->ops->unmap(pipe, buf, src);
1502                 if (!ret) {
1503                         ret = sd->len;
1504                         goto out;
1505                 }
1506         }
1507
1508         /*
1509          * No dice, use slow non-atomic map and copy
1510          */
1511         src = buf->ops->map(pipe, buf, 0);
1512
1513         ret = sd->len;
1514         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1515                 ret = -EFAULT;
1516
1517         buf->ops->unmap(pipe, buf, src);
1518 out:
1519         if (ret > 0)
1520                 sd->u.userptr += ret;
1521         return ret;
1522 }
1523
1524 /*
1525  * For lack of a better implementation, implement vmsplice() to userspace
1526  * as a simple copy of the pipes pages to the user iov.
1527  */
1528 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1529                              unsigned long nr_segs, unsigned int flags)
1530 {
1531         struct pipe_inode_info *pipe;
1532         struct splice_desc sd;
1533         ssize_t size;
1534         int error;
1535         long ret;
1536
1537         pipe = get_pipe_info(file);
1538         if (!pipe)
1539                 return -EBADF;
1540
1541         pipe_lock(pipe);
1542
1543         error = ret = 0;
1544         while (nr_segs) {
1545                 void __user *base;
1546                 size_t len;
1547
1548                 /*
1549                  * Get user address base and length for this iovec.
1550                  */
1551                 error = get_user(base, &iov->iov_base);
1552                 if (unlikely(error))
1553                         break;
1554                 error = get_user(len, &iov->iov_len);
1555                 if (unlikely(error))
1556                         break;
1557
1558                 /*
1559                  * Sanity check this iovec. 0 read succeeds.
1560                  */
1561                 if (unlikely(!len))
1562                         break;
1563                 if (unlikely(!base)) {
1564                         error = -EFAULT;
1565                         break;
1566                 }
1567
1568                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1569                         error = -EFAULT;
1570                         break;
1571                 }
1572
1573                 sd.len = 0;
1574                 sd.total_len = len;
1575                 sd.flags = flags;
1576                 sd.u.userptr = base;
1577                 sd.pos = 0;
1578
1579                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1580                 if (size < 0) {
1581                         if (!ret)
1582                                 ret = size;
1583
1584                         break;
1585                 }
1586
1587                 ret += size;
1588
1589                 if (size < len)
1590                         break;
1591
1592                 nr_segs--;
1593                 iov++;
1594         }
1595
1596         pipe_unlock(pipe);
1597
1598         if (!ret)
1599                 ret = error;
1600
1601         return ret;
1602 }
1603
1604 /*
1605  * vmsplice splices a user address range into a pipe. It can be thought of
1606  * as splice-from-memory, where the regular splice is splice-from-file (or
1607  * to file). In both cases the output is a pipe, naturally.
1608  */
1609 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1610                              unsigned long nr_segs, unsigned int flags)
1611 {
1612         struct pipe_inode_info *pipe;
1613         struct page *pages[PIPE_DEF_BUFFERS];
1614         struct partial_page partial[PIPE_DEF_BUFFERS];
1615         struct splice_pipe_desc spd = {
1616                 .pages = pages,
1617                 .partial = partial,
1618                 .flags = flags,
1619                 .ops = &user_page_pipe_buf_ops,
1620                 .spd_release = spd_release_page,
1621         };
1622         long ret;
1623
1624         pipe = get_pipe_info(file);
1625         if (!pipe)
1626                 return -EBADF;
1627
1628         if (splice_grow_spd(pipe, &spd))
1629                 return -ENOMEM;
1630
1631         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1632                                             spd.partial, flags & SPLICE_F_GIFT,
1633                                             pipe->buffers);
1634         if (spd.nr_pages <= 0)
1635                 ret = spd.nr_pages;
1636         else
1637                 ret = splice_to_pipe(pipe, &spd);
1638
1639         splice_shrink_spd(pipe, &spd);
1640         return ret;
1641 }
1642
1643 /*
1644  * Note that vmsplice only really supports true splicing _from_ user memory
1645  * to a pipe, not the other way around. Splicing from user memory is a simple
1646  * operation that can be supported without any funky alignment restrictions
1647  * or nasty vm tricks. We simply map in the user memory and fill them into
1648  * a pipe. The reverse isn't quite as easy, though. There are two possible
1649  * solutions for that:
1650  *
1651  *      - memcpy() the data internally, at which point we might as well just
1652  *        do a regular read() on the buffer anyway.
1653  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1654  *        has restriction limitations on both ends of the pipe).
1655  *
1656  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1657  *
1658  */
1659 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1660                 unsigned long, nr_segs, unsigned int, flags)
1661 {
1662         struct file *file;
1663         long error;
1664         int fput;
1665
1666         if (unlikely(nr_segs > UIO_MAXIOV))
1667                 return -EINVAL;
1668         else if (unlikely(!nr_segs))
1669                 return 0;
1670
1671         error = -EBADF;
1672         file = fget_light(fd, &fput);
1673         if (file) {
1674                 if (file->f_mode & FMODE_WRITE)
1675                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1676                 else if (file->f_mode & FMODE_READ)
1677                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1678
1679                 fput_light(file, fput);
1680         }
1681
1682         return error;
1683 }
1684
1685 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1686                 int, fd_out, loff_t __user *, off_out,
1687                 size_t, len, unsigned int, flags)
1688 {
1689         long error;
1690         struct file *in, *out;
1691         int fput_in, fput_out;
1692
1693         if (unlikely(!len))
1694                 return 0;
1695
1696         error = -EBADF;
1697         in = fget_light(fd_in, &fput_in);
1698         if (in) {
1699                 if (in->f_mode & FMODE_READ) {
1700                         out = fget_light(fd_out, &fput_out);
1701                         if (out) {
1702                                 if (out->f_mode & FMODE_WRITE)
1703                                         error = do_splice(in, off_in,
1704                                                           out, off_out,
1705                                                           len, flags);
1706                                 fput_light(out, fput_out);
1707                         }
1708                 }
1709
1710                 fput_light(in, fput_in);
1711         }
1712
1713         return error;
1714 }
1715
1716 /*
1717  * Make sure there's data to read. Wait for input if we can, otherwise
1718  * return an appropriate error.
1719  */
1720 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1721 {
1722         int ret;
1723
1724         /*
1725          * Check ->nrbufs without the inode lock first. This function
1726          * is speculative anyways, so missing one is ok.
1727          */
1728         if (pipe->nrbufs)
1729                 return 0;
1730
1731         ret = 0;
1732         pipe_lock(pipe);
1733
1734         while (!pipe->nrbufs) {
1735                 if (signal_pending(current)) {
1736                         ret = -ERESTARTSYS;
1737                         break;
1738                 }
1739                 if (!pipe->writers)
1740                         break;
1741                 if (!pipe->waiting_writers) {
1742                         if (flags & SPLICE_F_NONBLOCK) {
1743                                 ret = -EAGAIN;
1744                                 break;
1745                         }
1746                 }
1747                 pipe_wait(pipe);
1748         }
1749
1750         pipe_unlock(pipe);
1751         return ret;
1752 }
1753
1754 /*
1755  * Make sure there's writeable room. Wait for room if we can, otherwise
1756  * return an appropriate error.
1757  */
1758 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1759 {
1760         int ret;
1761
1762         /*
1763          * Check ->nrbufs without the inode lock first. This function
1764          * is speculative anyways, so missing one is ok.
1765          */
1766         if (pipe->nrbufs < pipe->buffers)
1767                 return 0;
1768
1769         ret = 0;
1770         pipe_lock(pipe);
1771
1772         while (pipe->nrbufs >= pipe->buffers) {
1773                 if (!pipe->readers) {
1774                         send_sig(SIGPIPE, current, 0);
1775                         ret = -EPIPE;
1776                         break;
1777                 }
1778                 if (flags & SPLICE_F_NONBLOCK) {
1779                         ret = -EAGAIN;
1780                         break;
1781                 }
1782                 if (signal_pending(current)) {
1783                         ret = -ERESTARTSYS;
1784                         break;
1785                 }
1786                 pipe->waiting_writers++;
1787                 pipe_wait(pipe);
1788                 pipe->waiting_writers--;
1789         }
1790
1791         pipe_unlock(pipe);
1792         return ret;
1793 }
1794
1795 /*
1796  * Splice contents of ipipe to opipe.
1797  */
1798 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1799                                struct pipe_inode_info *opipe,
1800                                size_t len, unsigned int flags)
1801 {
1802         struct pipe_buffer *ibuf, *obuf;
1803         int ret = 0, nbuf;
1804         bool input_wakeup = false;
1805
1806
1807 retry:
1808         ret = ipipe_prep(ipipe, flags);
1809         if (ret)
1810                 return ret;
1811
1812         ret = opipe_prep(opipe, flags);
1813         if (ret)
1814                 return ret;
1815
1816         /*
1817          * Potential ABBA deadlock, work around it by ordering lock
1818          * grabbing by pipe info address. Otherwise two different processes
1819          * could deadlock (one doing tee from A -> B, the other from B -> A).
1820          */
1821         pipe_double_lock(ipipe, opipe);
1822
1823         do {
1824                 if (!opipe->readers) {
1825                         send_sig(SIGPIPE, current, 0);
1826                         if (!ret)
1827                                 ret = -EPIPE;
1828                         break;
1829                 }
1830
1831                 if (!ipipe->nrbufs && !ipipe->writers)
1832                         break;
1833
1834                 /*
1835                  * Cannot make any progress, because either the input
1836                  * pipe is empty or the output pipe is full.
1837                  */
1838                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1839                         /* Already processed some buffers, break */
1840                         if (ret)
1841                                 break;
1842
1843                         if (flags & SPLICE_F_NONBLOCK) {
1844                                 ret = -EAGAIN;
1845                                 break;
1846                         }
1847
1848                         /*
1849                          * We raced with another reader/writer and haven't
1850                          * managed to process any buffers.  A zero return
1851                          * value means EOF, so retry instead.
1852                          */
1853                         pipe_unlock(ipipe);
1854                         pipe_unlock(opipe);
1855                         goto retry;
1856                 }
1857
1858                 ibuf = ipipe->bufs + ipipe->curbuf;
1859                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1860                 obuf = opipe->bufs + nbuf;
1861
1862                 if (len >= ibuf->len) {
1863                         /*
1864                          * Simply move the whole buffer from ipipe to opipe
1865                          */
1866                         *obuf = *ibuf;
1867                         ibuf->ops = NULL;
1868                         opipe->nrbufs++;
1869                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1870                         ipipe->nrbufs--;
1871                         input_wakeup = true;
1872                 } else {
1873                         /*
1874                          * Get a reference to this pipe buffer,
1875                          * so we can copy the contents over.
1876                          */
1877                         ibuf->ops->get(ipipe, ibuf);
1878                         *obuf = *ibuf;
1879
1880                         /*
1881                          * Don't inherit the gift flag, we need to
1882                          * prevent multiple steals of this page.
1883                          */
1884                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1885
1886                         obuf->len = len;
1887                         opipe->nrbufs++;
1888                         ibuf->offset += obuf->len;
1889                         ibuf->len -= obuf->len;
1890                 }
1891                 ret += obuf->len;
1892                 len -= obuf->len;
1893         } while (len);
1894
1895         pipe_unlock(ipipe);
1896         pipe_unlock(opipe);
1897
1898         /*
1899          * If we put data in the output pipe, wakeup any potential readers.
1900          */
1901         if (ret > 0)
1902                 wakeup_pipe_readers(opipe);
1903
1904         if (input_wakeup)
1905                 wakeup_pipe_writers(ipipe);
1906
1907         return ret;
1908 }
1909
1910 /*
1911  * Link contents of ipipe to opipe.
1912  */
1913 static int link_pipe(struct pipe_inode_info *ipipe,
1914                      struct pipe_inode_info *opipe,
1915                      size_t len, unsigned int flags)
1916 {
1917         struct pipe_buffer *ibuf, *obuf;
1918         int ret = 0, i = 0, nbuf;
1919
1920         /*
1921          * Potential ABBA deadlock, work around it by ordering lock
1922          * grabbing by pipe info address. Otherwise two different processes
1923          * could deadlock (one doing tee from A -> B, the other from B -> A).
1924          */
1925         pipe_double_lock(ipipe, opipe);
1926
1927         do {
1928                 if (!opipe->readers) {
1929                         send_sig(SIGPIPE, current, 0);
1930                         if (!ret)
1931                                 ret = -EPIPE;
1932                         break;
1933                 }
1934
1935                 /*
1936                  * If we have iterated all input buffers or ran out of
1937                  * output room, break.
1938                  */
1939                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1940                         break;
1941
1942                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1943                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1944
1945                 /*
1946                  * Get a reference to this pipe buffer,
1947                  * so we can copy the contents over.
1948                  */
1949                 ibuf->ops->get(ipipe, ibuf);
1950
1951                 obuf = opipe->bufs + nbuf;
1952                 *obuf = *ibuf;
1953
1954                 /*
1955                  * Don't inherit the gift flag, we need to
1956                  * prevent multiple steals of this page.
1957                  */
1958                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1959
1960                 if (obuf->len > len)
1961                         obuf->len = len;
1962
1963                 opipe->nrbufs++;
1964                 ret += obuf->len;
1965                 len -= obuf->len;
1966                 i++;
1967         } while (len);
1968
1969         /*
1970          * return EAGAIN if we have the potential of some data in the
1971          * future, otherwise just return 0
1972          */
1973         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1974                 ret = -EAGAIN;
1975
1976         pipe_unlock(ipipe);
1977         pipe_unlock(opipe);
1978
1979         /*
1980          * If we put data in the output pipe, wakeup any potential readers.
1981          */
1982         if (ret > 0)
1983                 wakeup_pipe_readers(opipe);
1984
1985         return ret;
1986 }
1987
1988 /*
1989  * This is a tee(1) implementation that works on pipes. It doesn't copy
1990  * any data, it simply references the 'in' pages on the 'out' pipe.
1991  * The 'flags' used are the SPLICE_F_* variants, currently the only
1992  * applicable one is SPLICE_F_NONBLOCK.
1993  */
1994 static long do_tee(struct file *in, struct file *out, size_t len,
1995                    unsigned int flags)
1996 {
1997         struct pipe_inode_info *ipipe = get_pipe_info(in);
1998         struct pipe_inode_info *opipe = get_pipe_info(out);
1999         int ret = -EINVAL;
2000
2001         /*
2002          * Duplicate the contents of ipipe to opipe without actually
2003          * copying the data.
2004          */
2005         if (ipipe && opipe && ipipe != opipe) {
2006                 /*
2007                  * Keep going, unless we encounter an error. The ipipe/opipe
2008                  * ordering doesn't really matter.
2009                  */
2010                 ret = ipipe_prep(ipipe, flags);
2011                 if (!ret) {
2012                         ret = opipe_prep(opipe, flags);
2013                         if (!ret)
2014                                 ret = link_pipe(ipipe, opipe, len, flags);
2015                 }
2016         }
2017
2018         return ret;
2019 }
2020
2021 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2022 {
2023         struct file *in;
2024         int error, fput_in;
2025
2026         if (unlikely(!len))
2027                 return 0;
2028
2029         error = -EBADF;
2030         in = fget_light(fdin, &fput_in);
2031         if (in) {
2032                 if (in->f_mode & FMODE_READ) {
2033                         int fput_out;
2034                         struct file *out = fget_light(fdout, &fput_out);
2035
2036                         if (out) {
2037                                 if (out->f_mode & FMODE_WRITE)
2038                                         error = do_tee(in, out, len, flags);
2039                                 fput_light(out, fput_out);
2040                         }
2041                 }
2042                 fput_light(in, fput_in);
2043         }
2044
2045         return error;
2046 }