fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.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) ? MSG_MORE : 0;
 695         if (sd->len < sd->total_len)
 696                 more |= MSG_SENDPAGE_NOTLAST;
 697         return file->f_op->sendpage(file, buf->page, buf->offset,
 698                                     sd->len, &pos, more);
 699 }
 700
 701 /*
 702  * This is a little more tricky than the file -> pipe splicing. There are
 703  * basically three cases:
 704  *
 705  *      - Destination page already exists in the address space and there
 706  *        are users of it. For that case we have no other option that
 707  *        copying the data. Tough luck.
 708  *      - Destination page already exists in the address space, but there
 709  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 710  *        through to last case.
 711  *      - Destination page does not exist, we can add the pipe page to
 712  *        the page cache and avoid the copy.
 713  *
 714  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 715  * sd->flags), we attempt to migrate pages from the pipe to the output
 716  * file address space page cache. This is possible if no one else has
 717  * the pipe page referenced outside of the pipe and page cache. If
 718  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 719  * a new page in the output file page cache and fill/dirty that.
 720  */
 721 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 722                  struct splice_desc *sd)
 723 {
 724         struct file *file = sd->u.file;
 725         struct address_space *mapping = file->f_mapping;
 726         unsigned int offset, this_len;
 727         struct page *page;
 728         void *fsdata;
 729         int ret;
 730
 731         offset = sd->pos & ~PAGE_CACHE_MASK;
 732
 733         this_len = sd->len;
 734         if (this_len + offset > PAGE_CACHE_SIZE)
 735                 this_len = PAGE_CACHE_SIZE - offset;
 736
 737         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 738                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 739         if (unlikely(ret))
 740                 goto out;
 741
 742         if (buf->page != page) {
 743                 char *src = buf->ops->map(pipe, buf, 1);
 744                 char *dst = kmap_atomic(page);
 745
 746                 memcpy(dst + offset, src + buf->offset, this_len);
 747                 flush_dcache_page(page);
 748                 kunmap_atomic(dst);
 749                 buf->ops->unmap(pipe, buf, src);
 750         }
 751         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 752                                 page, fsdata);
 753 out:
 754         return ret;
 755 }
 756 EXPORT_SYMBOL(pipe_to_file);
 757
 758 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 759 {
 760         smp_mb();
 761         if (waitqueue_active(&pipe->wait))
 762                 wake_up_interruptible(&pipe->wait);
 763         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 764 }
 765
 766 /**
 767  * splice_from_pipe_feed - feed available data from a pipe to a file
 768  * @pipe:       pipe to splice from
 769  * @sd:         information to @actor
 770  * @actor:      handler that splices the data
 771  *
 772  * Description:
 773  *    This function loops over the pipe and calls @actor to do the
 774  *    actual moving of a single struct pipe_buffer to the desired
 775  *    destination.  It returns when there's no more buffers left in
 776  *    the pipe or if the requested number of bytes (@sd->total_len)
 777  *    have been copied.  It returns a positive number (one) if the
 778  *    pipe needs to be filled with more data, zero if the required
 779  *    number of bytes have been copied and -errno on error.
 780  *
 781  *    This, together with splice_from_pipe_{begin,end,next}, may be
 782  *    used to implement the functionality of __splice_from_pipe() when
 783  *    locking is required around copying the pipe buffers to the
 784  *    destination.
 785  */
 786 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 787                           splice_actor *actor)
 788 {
 789         int ret;
 790
 791         while (pipe->nrbufs) {
 792                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 793                 const struct pipe_buf_operations *ops = buf->ops;
 794
 795                 sd->len = buf->len;
 796                 if (sd->len > sd->total_len)
 797                         sd->len = sd->total_len;
 798
 799                 ret = buf->ops->confirm(pipe, buf);
 800                 if (unlikely(ret)) {
 801                         if (ret == -ENODATA)
 802                                 ret = 0;
 803                         return ret;
 804                 }
 805
 806                 ret = actor(pipe, buf, sd);
 807                 if (ret <= 0)
 808                         return ret;
 809
 810                 buf->offset += ret;
 811                 buf->len -= ret;
 812
 813                 sd->num_spliced += ret;
 814                 sd->len -= ret;
 815                 sd->pos += ret;
 816                 sd->total_len -= ret;
 817
 818                 if (!buf->len) {
 819                         buf->ops = NULL;
 820                         ops->release(pipe, buf);
 821                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 822                         pipe->nrbufs--;
 823                         if (pipe->inode)
 824                                 sd->need_wakeup = true;
 825                 }
 826
 827                 if (!sd->total_len)
 828                         return 0;
 829         }
 830
 831         return 1;
 832 }
 833 EXPORT_SYMBOL(splice_from_pipe_feed);
 834
 835 /**
 836  * splice_from_pipe_next - wait for some data to splice from
 837  * @pipe:       pipe to splice from
 838  * @sd:         information about the splice operation
 839  *
 840  * Description:
 841  *    This function will wait for some data and return a positive
 842  *    value (one) if pipe buffers are available.  It will return zero
 843  *    or -errno if no more data needs to be spliced.
 844  */
 845 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 846 {
 847         while (!pipe->nrbufs) {
 848                 if (!pipe->writers)
 849                         return 0;
 850
 851                 if (!pipe->waiting_writers && sd->num_spliced)
 852                         return 0;
 853
 854                 if (sd->flags & SPLICE_F_NONBLOCK)
 855                         return -EAGAIN;
 856
 857                 if (signal_pending(current))
 858                         return -ERESTARTSYS;
 859
 860                 if (sd->need_wakeup) {
 861                         wakeup_pipe_writers(pipe);
 862                         sd->need_wakeup = false;
 863                 }
 864
 865                 pipe_wait(pipe);
 866         }
 867
 868         return 1;
 869 }
 870 EXPORT_SYMBOL(splice_from_pipe_next);
 871
 872 /**
 873  * splice_from_pipe_begin - start splicing from pipe
 874  * @sd:         information about the splice operation
 875  *
 876  * Description:
 877  *    This function should be called before a loop containing
 878  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 879  *    initialize the necessary fields of @sd.
 880  */
 881 void splice_from_pipe_begin(struct splice_desc *sd)
 882 {
 883         sd->num_spliced = 0;
 884         sd->need_wakeup = false;
 885 }
 886 EXPORT_SYMBOL(splice_from_pipe_begin);
 887
 888 /**
 889  * splice_from_pipe_end - finish splicing from pipe
 890  * @pipe:       pipe to splice from
 891  * @sd:         information about the splice operation
 892  *
 893  * Description:
 894  *    This function will wake up pipe writers if necessary.  It should
 895  *    be called after a loop containing splice_from_pipe_next() and
 896  *    splice_from_pipe_feed().
 897  */
 898 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 899 {
 900         if (sd->need_wakeup)
 901                 wakeup_pipe_writers(pipe);
 902 }
 903 EXPORT_SYMBOL(splice_from_pipe_end);
 904
 905 /**
 906  * __splice_from_pipe - splice data from a pipe to given actor
 907  * @pipe:       pipe to splice from
 908  * @sd:         information to @actor
 909  * @actor:      handler that splices the data
 910  *
 911  * Description:
 912  *    This function does little more than loop over the pipe and call
 913  *    @actor to do the actual moving of a single struct pipe_buffer to
 914  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 915  *    pipe_to_user.
 916  *
 917  */
 918 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 919                            splice_actor *actor)
 920 {
 921         int ret;
 922
 923         splice_from_pipe_begin(sd);
 924         do {
 925                 ret = splice_from_pipe_next(pipe, sd);
 926                 if (ret > 0)
 927                         ret = splice_from_pipe_feed(pipe, sd, actor);
 928         } while (ret > 0);
 929         splice_from_pipe_end(pipe, sd);
 930
 931         return sd->num_spliced ? sd->num_spliced : ret;
 932 }
 933 EXPORT_SYMBOL(__splice_from_pipe);
 934
 935 /**
 936  * splice_from_pipe - splice data from a pipe to a file
 937  * @pipe:       pipe to splice from
 938  * @out:        file to splice to
 939  * @ppos:       position in @out
 940  * @len:        how many bytes to splice
 941  * @flags:      splice modifier flags
 942  * @actor:      handler that splices the data
 943  *
 944  * Description:
 945  *    See __splice_from_pipe. This function locks the pipe inode,
 946  *    otherwise it's identical to __splice_from_pipe().
 947  *
 948  */
 949 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 950                          loff_t *ppos, size_t len, unsigned int flags,
 951                          splice_actor *actor)
 952 {
 953         ssize_t ret;
 954         struct splice_desc sd = {
 955                 .total_len = len,
 956                 .flags = flags,
 957                 .pos = *ppos,
 958                 .u.file = out,
 959         };
 960
 961         pipe_lock(pipe);
 962         ret = __splice_from_pipe(pipe, &sd, actor);
 963         pipe_unlock(pipe);
 964
 965         return ret;
 966 }
 967
 968 /**
 969  * generic_file_splice_write - splice data from a pipe to a file
 970  * @pipe:       pipe info
 971  * @out:        file to write to
 972  * @ppos:       position in @out
 973  * @len:        number of bytes to splice
 974  * @flags:      splice modifier flags
 975  *
 976  * Description:
 977  *    Will either move or copy pages (determined by @flags options) from
 978  *    the given pipe inode to the given file.
 979  *
 980  */
 981 ssize_t
 982 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 983                           loff_t *ppos, size_t len, unsigned int flags)
 984 {
 985         struct address_space *mapping = out->f_mapping;
 986         struct inode *inode = mapping->host;
 987         struct splice_desc sd = {
 988                 .total_len = len,
 989                 .flags = flags,
 990                 .pos = *ppos,
 991                 .u.file = out,
 992         };
 993         ssize_t ret;
 994
 995         pipe_lock(pipe);
 996
 997         splice_from_pipe_begin(&sd);
 998         do {
 999                 ret = splice_from_pipe_next(pipe, &sd);
1000                 if (ret <= 0)
1001                         break;
1002
1003                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1004                 ret = file_remove_suid(out);
1005                 if (!ret) {
1006                         file_update_time(out);
1007                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1008                 }
1009                 mutex_unlock(&inode->i_mutex);
1010         } while (ret > 0);
1011         splice_from_pipe_end(pipe, &sd);
1012
1013         pipe_unlock(pipe);
1014
1015         if (sd.num_spliced)
1016                 ret = sd.num_spliced;
1017
1018         if (ret > 0) {
1019                 unsigned long nr_pages;
1020                 int err;
1021
1022                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1023
1024                 err = generic_write_sync(out, *ppos, ret);
1025                 if (err)
1026                         ret = err;
1027                 else
1028                         *ppos += ret;
1029                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1030         }
1031
1032         return ret;
1033 }
1034
1035 EXPORT_SYMBOL(generic_file_splice_write);
1036
1037 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1038                           struct splice_desc *sd)
1039 {
1040         int ret;
1041         void *data;
1042
1043         data = buf->ops->map(pipe, buf, 0);
1044         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1045         buf->ops->unmap(pipe, buf, data);
1046
1047         return ret;
1048 }
1049
1050 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1051                                          struct file *out, loff_t *ppos,
1052                                          size_t len, unsigned int flags)
1053 {
1054         ssize_t ret;
1055
1056         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1057         if (ret > 0)
1058                 *ppos += ret;
1059
1060         return ret;
1061 }
1062
1063 /**
1064  * generic_splice_sendpage - splice data from a pipe to a socket
1065  * @pipe:       pipe to splice from
1066  * @out:        socket to write to
1067  * @ppos:       position in @out
1068  * @len:        number of bytes to splice
1069  * @flags:      splice modifier flags
1070  *
1071  * Description:
1072  *    Will send @len bytes from the pipe to a network socket. No data copying
1073  *    is involved.
1074  *
1075  */
1076 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1077                                 loff_t *ppos, size_t len, unsigned int flags)
1078 {
1079         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1080 }
1081
1082 EXPORT_SYMBOL(generic_splice_sendpage);
1083
1084 /*
1085  * Attempt to initiate a splice from pipe to file.
1086  */
1087 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1088                            loff_t *ppos, size_t len, unsigned int flags)
1089 {
1090         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1091                                 loff_t *, size_t, unsigned int);
1092         int ret;
1093
1094         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1095                 return -EBADF;
1096
1097         if (unlikely(out->f_flags & O_APPEND))
1098                 return -EINVAL;
1099
1100         ret = rw_verify_area(WRITE, out, ppos, len);
1101         if (unlikely(ret < 0))
1102                 return ret;
1103
1104         if (out->f_op && out->f_op->splice_write)
1105                 splice_write = out->f_op->splice_write;
1106         else
1107                 splice_write = default_file_splice_write;
1108
1109         return splice_write(pipe, out, ppos, len, flags);
1110 }
1111
1112 /*
1113  * Attempt to initiate a splice from a file to a pipe.
1114  */
1115 static long do_splice_to(struct file *in, loff_t *ppos,
1116                          struct pipe_inode_info *pipe, size_t len,
1117                          unsigned int flags)
1118 {
1119         ssize_t (*splice_read)(struct file *, loff_t *,
1120                                struct pipe_inode_info *, size_t, unsigned int);
1121         int ret;
1122
1123         if (unlikely(!(in->f_mode & FMODE_READ)))
1124                 return -EBADF;
1125
1126         ret = rw_verify_area(READ, in, ppos, len);
1127         if (unlikely(ret < 0))
1128                 return ret;
1129
1130         if (in->f_op && in->f_op->splice_read)
1131                 splice_read = in->f_op->splice_read;
1132         else
1133                 splice_read = default_file_splice_read;
1134
1135         return splice_read(in, ppos, pipe, len, flags);
1136 }
1137
1138 /**
1139  * splice_direct_to_actor - splices data directly between two non-pipes
1140  * @in:         file to splice from
1141  * @sd:         actor information on where to splice to
1142  * @actor:      handles the data splicing
1143  *
1144  * Description:
1145  *    This is a special case helper to splice directly between two
1146  *    points, without requiring an explicit pipe. Internally an allocated
1147  *    pipe is cached in the process, and reused during the lifetime of
1148  *    that process.
1149  *
1150  */
1151 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1152                                splice_direct_actor *actor)
1153 {
1154         struct pipe_inode_info *pipe;
1155         long ret, bytes;
1156         umode_t i_mode;
1157         size_t len;
1158         int i, flags;
1159
1160         /*
1161          * We require the input being a regular file, as we don't want to
1162          * randomly drop data for eg socket -> socket splicing. Use the
1163          * piped splicing for that!
1164          */
1165         i_mode = in->f_path.dentry->d_inode->i_mode;
1166         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1167                 return -EINVAL;
1168
1169         /*
1170          * neither in nor out is a pipe, setup an internal pipe attached to
1171          * 'out' and transfer the wanted data from 'in' to 'out' through that
1172          */
1173         pipe = current->splice_pipe;
1174         if (unlikely(!pipe)) {
1175                 pipe = alloc_pipe_info(NULL);
1176                 if (!pipe)
1177                         return -ENOMEM;
1178
1179                 /*
1180                  * We don't have an immediate reader, but we'll read the stuff
1181                  * out of the pipe right after the splice_to_pipe(). So set
1182                  * PIPE_READERS appropriately.
1183                  */
1184                 pipe->readers = 1;
1185
1186                 current->splice_pipe = pipe;
1187         }
1188
1189         /*
1190          * Do the splice.
1191          */
1192         ret = 0;
1193         bytes = 0;
1194         len = sd->total_len;
1195         flags = sd->flags;
1196
1197         /*
1198          * Don't block on output, we have to drain the direct pipe.
1199          */
1200         sd->flags &= ~SPLICE_F_NONBLOCK;
1201
1202         while (len) {
1203                 size_t read_len;
1204                 loff_t pos = sd->pos, prev_pos = pos;
1205
1206                 ret = do_splice_to(in, &pos, pipe, len, flags);
1207                 if (unlikely(ret <= 0))
1208                         goto out_release;
1209
1210                 read_len = ret;
1211                 sd->total_len = read_len;
1212
1213                 /*
1214                  * NOTE: nonblocking mode only applies to the input. We
1215                  * must not do the output in nonblocking mode as then we
1216                  * could get stuck data in the internal pipe:
1217                  */
1218                 ret = actor(pipe, sd);
1219                 if (unlikely(ret <= 0)) {
1220                         sd->pos = prev_pos;
1221                         goto out_release;
1222                 }
1223
1224                 bytes += ret;
1225                 len -= ret;
1226                 sd->pos = pos;
1227
1228                 if (ret < read_len) {
1229                         sd->pos = prev_pos + ret;
1230                         goto out_release;
1231                 }
1232         }
1233
1234 done:
1235         pipe->nrbufs = pipe->curbuf = 0;
1236         file_accessed(in);
1237         return bytes;
1238
1239 out_release:
1240         /*
1241          * If we did an incomplete transfer we must release
1242          * the pipe buffers in question:
1243          */
1244         for (i = 0; i < pipe->buffers; i++) {
1245                 struct pipe_buffer *buf = pipe->bufs + i;
1246
1247                 if (buf->ops) {
1248                         buf->ops->release(pipe, buf);
1249                         buf->ops = NULL;
1250                 }
1251         }
1252
1253         if (!bytes)
1254                 bytes = ret;
1255
1256         goto done;
1257 }
1258 EXPORT_SYMBOL(splice_direct_to_actor);
1259
1260 static int direct_splice_actor(struct pipe_inode_info *pipe,
1261                                struct splice_desc *sd)
1262 {
1263         struct file *file = sd->u.file;
1264
1265         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1266                               sd->flags);
1267 }
1268
1269 /**
1270  * do_splice_direct - splices data directly between two files
1271  * @in:         file to splice from
1272  * @ppos:       input file offset
1273  * @out:        file to splice to
1274  * @len:        number of bytes to splice
1275  * @flags:      splice modifier flags
1276  *
1277  * Description:
1278  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1279  *    doing it in the application would incur an extra system call
1280  *    (splice in + splice out, as compared to just sendfile()). So this helper
1281  *    can splice directly through a process-private pipe.
1282  *
1283  */
1284 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1285                       size_t len, unsigned int flags)
1286 {
1287         struct splice_desc sd = {
1288                 .len            = len,
1289                 .total_len      = len,
1290                 .flags          = flags,
1291                 .pos            = *ppos,
1292                 .u.file         = out,
1293         };
1294         long ret;
1295
1296         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1297         if (ret > 0)
1298                 *ppos = sd.pos;
1299
1300         return ret;
1301 }
1302
1303 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1304                                struct pipe_inode_info *opipe,
1305                                size_t len, unsigned int flags);
1306
1307 /*
1308  * Determine where to splice to/from.
1309  */
1310 static long do_splice(struct file *in, loff_t __user *off_in,
1311                       struct file *out, loff_t __user *off_out,
1312                       size_t len, unsigned int flags)
1313 {
1314         struct pipe_inode_info *ipipe;
1315         struct pipe_inode_info *opipe;
1316         loff_t offset, *off;
1317         long ret;
1318
1319         ipipe = get_pipe_info(in);
1320         opipe = get_pipe_info(out);
1321
1322         if (ipipe && opipe) {
1323                 if (off_in || off_out)
1324                         return -ESPIPE;
1325
1326                 if (!(in->f_mode & FMODE_READ))
1327                         return -EBADF;
1328
1329                 if (!(out->f_mode & FMODE_WRITE))
1330                         return -EBADF;
1331
1332                 /* Splicing to self would be fun, but... */
1333                 if (ipipe == opipe)
1334                         return -EINVAL;
1335
1336                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1337         }
1338
1339         if (ipipe) {
1340                 if (off_in)
1341                         return -ESPIPE;
1342                 if (off_out) {
1343                         if (!(out->f_mode & FMODE_PWRITE))
1344                                 return -EINVAL;
1345                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1346                                 return -EFAULT;
1347                         off = &offset;
1348                 } else
1349                         off = &out->f_pos;
1350
1351                 ret = do_splice_from(ipipe, out, off, len, flags);
1352
1353                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1354                         ret = -EFAULT;
1355
1356                 return ret;
1357         }
1358
1359         if (opipe) {
1360                 if (off_out)
1361                         return -ESPIPE;
1362                 if (off_in) {
1363                         if (!(in->f_mode & FMODE_PREAD))
1364                                 return -EINVAL;
1365                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1366                                 return -EFAULT;
1367                         off = &offset;
1368                 } else
1369                         off = &in->f_pos;
1370
1371                 ret = do_splice_to(in, off, opipe, len, flags);
1372
1373                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1374                         ret = -EFAULT;
1375
1376                 return ret;
1377         }
1378
1379         return -EINVAL;
1380 }
1381
1382 /*
1383  * Map an iov into an array of pages and offset/length tupples. With the
1384  * partial_page structure, we can map several non-contiguous ranges into
1385  * our ones pages[] map instead of splitting that operation into pieces.
1386  * Could easily be exported as a generic helper for other users, in which
1387  * case one would probably want to add a 'max_nr_pages' parameter as well.
1388  */
1389 static int get_iovec_page_array(const struct iovec __user *iov,
1390                                 unsigned int nr_vecs, struct page **pages,
1391                                 struct partial_page *partial, bool aligned,
1392                                 unsigned int pipe_buffers)
1393 {
1394         int buffers = 0, error = 0;
1395
1396         while (nr_vecs) {
1397                 unsigned long off, npages;
1398                 struct iovec entry;
1399                 void __user *base;
1400                 size_t len;
1401                 int i;
1402
1403                 error = -EFAULT;
1404                 if (copy_from_user(&entry, iov, sizeof(entry)))
1405                         break;
1406
1407                 base = entry.iov_base;
1408                 len = entry.iov_len;
1409
1410                 /*
1411                  * Sanity check this iovec. 0 read succeeds.
1412                  */
1413                 error = 0;
1414                 if (unlikely(!len))
1415                         break;
1416                 error = -EFAULT;
1417                 if (!access_ok(VERIFY_READ, base, len))
1418                         break;
1419
1420                 /*
1421                  * Get this base offset and number of pages, then map
1422                  * in the user pages.
1423                  */
1424                 off = (unsigned long) base & ~PAGE_MASK;
1425
1426                 /*
1427                  * If asked for alignment, the offset must be zero and the
1428                  * length a multiple of the PAGE_SIZE.
1429                  */
1430                 error = -EINVAL;
1431                 if (aligned && (off || len & ~PAGE_MASK))
1432                         break;
1433
1434                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1435                 if (npages > pipe_buffers - buffers)
1436                         npages = pipe_buffers - buffers;
1437
1438                 error = get_user_pages_fast((unsigned long)base, npages,
1439                                         0, &pages[buffers]);
1440
1441                 if (unlikely(error <= 0))
1442                         break;
1443
1444                 /*
1445                  * Fill this contiguous range into the partial page map.
1446                  */
1447                 for (i = 0; i < error; i++) {
1448                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1449
1450                         partial[buffers].offset = off;
1451                         partial[buffers].len = plen;
1452
1453                         off = 0;
1454                         len -= plen;
1455                         buffers++;
1456                 }
1457
1458                 /*
1459                  * We didn't complete this iov, stop here since it probably
1460                  * means we have to move some of this into a pipe to
1461                  * be able to continue.
1462                  */
1463                 if (len)
1464                         break;
1465
1466                 /*
1467                  * Don't continue if we mapped fewer pages than we asked for,
1468                  * or if we mapped the max number of pages that we have
1469                  * room for.
1470                  */
1471                 if (error < npages || buffers == pipe_buffers)
1472                         break;
1473
1474                 nr_vecs--;
1475                 iov++;
1476         }
1477
1478         if (buffers)
1479                 return buffers;
1480
1481         return error;
1482 }
1483
1484 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1485                         struct splice_desc *sd)
1486 {
1487         char *src;
1488         int ret;
1489
1490         /*
1491          * See if we can use the atomic maps, by prefaulting in the
1492          * pages and doing an atomic copy
1493          */
1494         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1495                 src = buf->ops->map(pipe, buf, 1);
1496                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1497                                                         sd->len);
1498                 buf->ops->unmap(pipe, buf, src);
1499                 if (!ret) {
1500                         ret = sd->len;
1501                         goto out;
1502                 }
1503         }
1504
1505         /*
1506          * No dice, use slow non-atomic map and copy
1507          */
1508         src = buf->ops->map(pipe, buf, 0);
1509
1510         ret = sd->len;
1511         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1512                 ret = -EFAULT;
1513
1514         buf->ops->unmap(pipe, buf, src);
1515 out:
1516         if (ret > 0)
1517                 sd->u.userptr += ret;
1518         return ret;
1519 }
1520
1521 /*
1522  * For lack of a better implementation, implement vmsplice() to userspace
1523  * as a simple copy of the pipes pages to the user iov.
1524  */
1525 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1526                              unsigned long nr_segs, unsigned int flags)
1527 {
1528         struct pipe_inode_info *pipe;
1529         struct splice_desc sd;
1530         ssize_t size;
1531         int error;
1532         long ret;
1533
1534         pipe = get_pipe_info(file);
1535         if (!pipe)
1536                 return -EBADF;
1537
1538         pipe_lock(pipe);
1539
1540         error = ret = 0;
1541         while (nr_segs) {
1542                 void __user *base;
1543                 size_t len;
1544
1545                 /*
1546                  * Get user address base and length for this iovec.
1547                  */
1548                 error = get_user(base, &iov->iov_base);
1549                 if (unlikely(error))
1550                         break;
1551                 error = get_user(len, &iov->iov_len);
1552                 if (unlikely(error))
1553                         break;
1554
1555                 /*
1556                  * Sanity check this iovec. 0 read succeeds.
1557                  */
1558                 if (unlikely(!len))
1559                         break;
1560                 if (unlikely(!base)) {
1561                         error = -EFAULT;
1562                         break;
1563                 }
1564
1565                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1566                         error = -EFAULT;
1567                         break;
1568                 }
1569
1570                 sd.len = 0;
1571                 sd.total_len = len;
1572                 sd.flags = flags;
1573                 sd.u.userptr = base;
1574                 sd.pos = 0;
1575
1576                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1577                 if (size < 0) {
1578                         if (!ret)
1579                                 ret = size;
1580
1581                         break;
1582                 }
1583
1584                 ret += size;
1585
1586                 if (size < len)
1587                         break;
1588
1589                 nr_segs--;
1590                 iov++;
1591         }
1592
1593         pipe_unlock(pipe);
1594
1595         if (!ret)
1596                 ret = error;
1597
1598         return ret;
1599 }
1600
1601 /*
1602  * vmsplice splices a user address range into a pipe. It can be thought of
1603  * as splice-from-memory, where the regular splice is splice-from-file (or
1604  * to file). In both cases the output is a pipe, naturally.
1605  */
1606 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1607                              unsigned long nr_segs, unsigned int flags)
1608 {
1609         struct pipe_inode_info *pipe;
1610         struct page *pages[PIPE_DEF_BUFFERS];
1611         struct partial_page partial[PIPE_DEF_BUFFERS];
1612         struct splice_pipe_desc spd = {
1613                 .pages = pages,
1614                 .partial = partial,
1615                 .flags = flags,
1616                 .ops = &user_page_pipe_buf_ops,
1617                 .spd_release = spd_release_page,
1618         };
1619         long ret;
1620
1621         pipe = get_pipe_info(file);
1622         if (!pipe)
1623                 return -EBADF;
1624
1625         if (splice_grow_spd(pipe, &spd))
1626                 return -ENOMEM;
1627
1628         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1629                                             spd.partial, false,
1630                                             pipe->buffers);
1631         if (spd.nr_pages <= 0)
1632                 ret = spd.nr_pages;
1633         else
1634                 ret = splice_to_pipe(pipe, &spd);
1635
1636         splice_shrink_spd(pipe, &spd);
1637         return ret;
1638 }
1639
1640 /*
1641  * Note that vmsplice only really supports true splicing _from_ user memory
1642  * to a pipe, not the other way around. Splicing from user memory is a simple
1643  * operation that can be supported without any funky alignment restrictions
1644  * or nasty vm tricks. We simply map in the user memory and fill them into
1645  * a pipe. The reverse isn't quite as easy, though. There are two possible
1646  * solutions for that:
1647  *
1648  *      - memcpy() the data internally, at which point we might as well just
1649  *        do a regular read() on the buffer anyway.
1650  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1651  *        has restriction limitations on both ends of the pipe).
1652  *
1653  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1654  *
1655  */
1656 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1657                 unsigned long, nr_segs, unsigned int, flags)
1658 {
1659         struct file *file;
1660         long error;
1661         int fput;
1662
1663         if (unlikely(nr_segs > UIO_MAXIOV))
1664                 return -EINVAL;
1665         else if (unlikely(!nr_segs))
1666                 return 0;
1667
1668         error = -EBADF;
1669         file = fget_light(fd, &fput);
1670         if (file) {
1671                 if (file->f_mode & FMODE_WRITE)
1672                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1673                 else if (file->f_mode & FMODE_READ)
1674                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1675
1676                 fput_light(file, fput);
1677         }
1678
1679         return error;
1680 }
1681
1682 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1683                 int, fd_out, loff_t __user *, off_out,
1684                 size_t, len, unsigned int, flags)
1685 {
1686         long error;
1687         struct file *in, *out;
1688         int fput_in, fput_out;
1689
1690         if (unlikely(!len))
1691                 return 0;
1692
1693         error = -EBADF;
1694         in = fget_light(fd_in, &fput_in);
1695         if (in) {
1696                 if (in->f_mode & FMODE_READ) {
1697                         out = fget_light(fd_out, &fput_out);
1698                         if (out) {
1699                                 if (out->f_mode & FMODE_WRITE)
1700                                         error = do_splice(in, off_in,
1701                                                           out, off_out,
1702                                                           len, flags);
1703                                 fput_light(out, fput_out);
1704                         }
1705                 }
1706
1707                 fput_light(in, fput_in);
1708         }
1709
1710         return error;
1711 }
1712
1713 /*
1714  * Make sure there's data to read. Wait for input if we can, otherwise
1715  * return an appropriate error.
1716  */
1717 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1718 {
1719         int ret;
1720
1721         /*
1722          * Check ->nrbufs without the inode lock first. This function
1723          * is speculative anyways, so missing one is ok.
1724          */
1725         if (pipe->nrbufs)
1726                 return 0;
1727
1728         ret = 0;
1729         pipe_lock(pipe);
1730
1731         while (!pipe->nrbufs) {
1732                 if (signal_pending(current)) {
1733                         ret = -ERESTARTSYS;
1734                         break;
1735                 }
1736                 if (!pipe->writers)
1737                         break;
1738                 if (!pipe->waiting_writers) {
1739                         if (flags & SPLICE_F_NONBLOCK) {
1740                                 ret = -EAGAIN;
1741                                 break;
1742                         }
1743                 }
1744                 pipe_wait(pipe);
1745         }
1746
1747         pipe_unlock(pipe);
1748         return ret;
1749 }
1750
1751 /*
1752  * Make sure there's writeable room. Wait for room if we can, otherwise
1753  * return an appropriate error.
1754  */
1755 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1756 {
1757         int ret;
1758
1759         /*
1760          * Check ->nrbufs without the inode lock first. This function
1761          * is speculative anyways, so missing one is ok.
1762          */
1763         if (pipe->nrbufs < pipe->buffers)
1764                 return 0;
1765
1766         ret = 0;
1767         pipe_lock(pipe);
1768
1769         while (pipe->nrbufs >= pipe->buffers) {
1770                 if (!pipe->readers) {
1771                         send_sig(SIGPIPE, current, 0);
1772                         ret = -EPIPE;
1773                         break;
1774                 }
1775                 if (flags & SPLICE_F_NONBLOCK) {
1776                         ret = -EAGAIN;
1777                         break;
1778                 }
1779                 if (signal_pending(current)) {
1780                         ret = -ERESTARTSYS;
1781                         break;
1782                 }
1783                 pipe->waiting_writers++;
1784                 pipe_wait(pipe);
1785                 pipe->waiting_writers--;
1786         }
1787
1788         pipe_unlock(pipe);
1789         return ret;
1790 }
1791
1792 /*
1793  * Splice contents of ipipe to opipe.
1794  */
1795 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1796                                struct pipe_inode_info *opipe,
1797                                size_t len, unsigned int flags)
1798 {
1799         struct pipe_buffer *ibuf, *obuf;
1800         int ret = 0, nbuf;
1801         bool input_wakeup = false;
1802
1803
1804 retry:
1805         ret = ipipe_prep(ipipe, flags);
1806         if (ret)
1807                 return ret;
1808
1809         ret = opipe_prep(opipe, flags);
1810         if (ret)
1811                 return ret;
1812
1813         /*
1814          * Potential ABBA deadlock, work around it by ordering lock
1815          * grabbing by pipe info address. Otherwise two different processes
1816          * could deadlock (one doing tee from A -> B, the other from B -> A).
1817          */
1818         pipe_double_lock(ipipe, opipe);
1819
1820         do {
1821                 if (!opipe->readers) {
1822                         send_sig(SIGPIPE, current, 0);
1823                         if (!ret)
1824                                 ret = -EPIPE;
1825                         break;
1826                 }
1827
1828                 if (!ipipe->nrbufs && !ipipe->writers)
1829                         break;
1830
1831                 /*
1832                  * Cannot make any progress, because either the input
1833                  * pipe is empty or the output pipe is full.
1834                  */
1835                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1836                         /* Already processed some buffers, break */
1837                         if (ret)
1838                                 break;
1839
1840                         if (flags & SPLICE_F_NONBLOCK) {
1841                                 ret = -EAGAIN;
1842                                 break;
1843                         }
1844
1845                         /*
1846                          * We raced with another reader/writer and haven't
1847                          * managed to process any buffers.  A zero return
1848                          * value means EOF, so retry instead.
1849                          */
1850                         pipe_unlock(ipipe);
1851                         pipe_unlock(opipe);
1852                         goto retry;
1853                 }
1854
1855                 ibuf = ipipe->bufs + ipipe->curbuf;
1856                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1857                 obuf = opipe->bufs + nbuf;
1858
1859                 if (len >= ibuf->len) {
1860                         /*
1861                          * Simply move the whole buffer from ipipe to opipe
1862                          */
1863                         *obuf = *ibuf;
1864                         ibuf->ops = NULL;
1865                         opipe->nrbufs++;
1866                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1867                         ipipe->nrbufs--;
1868                         input_wakeup = true;
1869                 } else {
1870                         /*
1871                          * Get a reference to this pipe buffer,
1872                          * so we can copy the contents over.
1873                          */
1874                         ibuf->ops->get(ipipe, ibuf);
1875                         *obuf = *ibuf;
1876
1877                         /*
1878                          * Don't inherit the gift flag, we need to
1879                          * prevent multiple steals of this page.
1880                          */
1881                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1882
1883                         obuf->len = len;
1884                         opipe->nrbufs++;
1885                         ibuf->offset += obuf->len;
1886                         ibuf->len -= obuf->len;
1887                 }
1888                 ret += obuf->len;
1889                 len -= obuf->len;
1890         } while (len);
1891
1892         pipe_unlock(ipipe);
1893         pipe_unlock(opipe);
1894
1895         /*
1896          * If we put data in the output pipe, wakeup any potential readers.
1897          */
1898         if (ret > 0)
1899                 wakeup_pipe_readers(opipe);
1900
1901         if (input_wakeup)
1902                 wakeup_pipe_writers(ipipe);
1903
1904         return ret;
1905 }
1906
1907 /*
1908  * Link contents of ipipe to opipe.
1909  */
1910 static int link_pipe(struct pipe_inode_info *ipipe,
1911                      struct pipe_inode_info *opipe,
1912                      size_t len, unsigned int flags)
1913 {
1914         struct pipe_buffer *ibuf, *obuf;
1915         int ret = 0, i = 0, nbuf;
1916
1917         /*
1918          * Potential ABBA deadlock, work around it by ordering lock
1919          * grabbing by pipe info address. Otherwise two different processes
1920          * could deadlock (one doing tee from A -> B, the other from B -> A).
1921          */
1922         pipe_double_lock(ipipe, opipe);
1923
1924         do {
1925                 if (!opipe->readers) {
1926                         send_sig(SIGPIPE, current, 0);
1927                         if (!ret)
1928                                 ret = -EPIPE;
1929                         break;
1930                 }
1931
1932                 /*
1933                  * If we have iterated all input buffers or ran out of
1934                  * output room, break.
1935                  */
1936                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1937                         break;
1938
1939                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1940                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1941
1942                 /*
1943                  * Get a reference to this pipe buffer,
1944                  * so we can copy the contents over.
1945                  */
1946                 ibuf->ops->get(ipipe, ibuf);
1947
1948                 obuf = opipe->bufs + nbuf;
1949                 *obuf = *ibuf;
1950
1951                 /*
1952                  * Don't inherit the gift flag, we need to
1953                  * prevent multiple steals of this page.
1954                  */
1955                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1956
1957                 if (obuf->len > len)
1958                         obuf->len = len;
1959
1960                 opipe->nrbufs++;
1961                 ret += obuf->len;
1962                 len -= obuf->len;
1963                 i++;
1964         } while (len);
1965
1966         /*
1967          * return EAGAIN if we have the potential of some data in the
1968          * future, otherwise just return 0
1969          */
1970         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1971                 ret = -EAGAIN;
1972
1973         pipe_unlock(ipipe);
1974         pipe_unlock(opipe);
1975
1976         /*
1977          * If we put data in the output pipe, wakeup any potential readers.
1978          */
1979         if (ret > 0)
1980                 wakeup_pipe_readers(opipe);
1981
1982         return ret;
1983 }
1984
1985 /*
1986  * This is a tee(1) implementation that works on pipes. It doesn't copy
1987  * any data, it simply references the 'in' pages on the 'out' pipe.
1988  * The 'flags' used are the SPLICE_F_* variants, currently the only
1989  * applicable one is SPLICE_F_NONBLOCK.
1990  */
1991 static long do_tee(struct file *in, struct file *out, size_t len,
1992                    unsigned int flags)
1993 {
1994         struct pipe_inode_info *ipipe = get_pipe_info(in);
1995         struct pipe_inode_info *opipe = get_pipe_info(out);
1996         int ret = -EINVAL;
1997
1998         /*
1999          * Duplicate the contents of ipipe to opipe without actually
2000          * copying the data.
2001          */
2002         if (ipipe && opipe && ipipe != opipe) {
2003                 /*
2004                  * Keep going, unless we encounter an error. The ipipe/opipe
2005                  * ordering doesn't really matter.
2006                  */
2007                 ret = ipipe_prep(ipipe, flags);
2008                 if (!ret) {
2009                         ret = opipe_prep(opipe, flags);
2010                         if (!ret)
2011                                 ret = link_pipe(ipipe, opipe, len, flags);
2012                 }
2013         }
2014
2015         return ret;
2016 }
2017
2018 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2019 {
2020         struct file *in;
2021         int error, fput_in;
2022
2023         if (unlikely(!len))
2024                 return 0;
2025
2026         error = -EBADF;
2027         in = fget_light(fdin, &fput_in);
2028         if (in) {
2029                 if (in->f_mode & FMODE_READ) {
2030                         int fput_out;
2031                         struct file *out = fget_light(fdout, &fput_out);
2032
2033                         if (out) {
2034                                 if (out->f_mode & FMODE_WRITE)
2035                                         error = do_tee(in, out, len, flags);
2036                                 fput_light(out, fput_out);
2037                         }
2038                 }
2039                 fput_light(in, fput_in);
2040         }
2041
2042         return error;
2043 }