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 #include <linux/compat.h>
  35 #include "internal.h"
  36
  37 /*
  38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  39  * a vm helper function, it's already simplified quite a bit by the
  40  * addition of remove_mapping(). If success is returned, the caller may
  41  * attempt to reuse this page for another destination.
  42  */
  43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  44                                      struct pipe_buffer *buf)
  45 {
  46         struct page *page = buf->page;
  47         struct address_space *mapping;
  48
  49         lock_page(page);
  50
  51         mapping = page_mapping(page);
  52         if (mapping) {
  53                 WARN_ON(!PageUptodate(page));
  54
  55                 /*
  56                  * At least for ext2 with nobh option, we need to wait on
  57                  * writeback completing on this page, since we'll remove it
  58                  * from the pagecache.  Otherwise truncate wont wait on the
  59                  * page, allowing the disk blocks to be reused by someone else
  60                  * before we actually wrote our data to them. fs corruption
  61                  * ensues.
  62                  */
  63                 wait_on_page_writeback(page);
  64
  65                 if (page_has_private(page) &&
  66                     !try_to_release_page(page, GFP_KERNEL))
  67                         goto out_unlock;
  68
  69                 /*
  70                  * If we succeeded in removing the mapping, set LRU flag
  71                  * and return good.
  72                  */
  73                 if (remove_mapping(mapping, page)) {
  74                         buf->flags |= PIPE_BUF_FLAG_LRU;
  75                         return 0;
  76                 }
  77         }
  78
  79         /*
  80          * Raced with truncate or failed to remove page from current
  81          * address space, unlock and return failure.
  82          */
  83 out_unlock:
  84         unlock_page(page);
  85         return 1;
  86 }
  87
  88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  89                                         struct pipe_buffer *buf)
  90 {
  91         page_cache_release(buf->page);
  92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  93 }
  94
  95 /*
  96  * Check whether the contents of buf is OK to access. Since the content
  97  * is a page cache page, IO may be in flight.
  98  */
  99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 100                                        struct pipe_buffer *buf)
 101 {
 102         struct page *page = buf->page;
 103         int err;
 104
 105         if (!PageUptodate(page)) {
 106                 lock_page(page);
 107
 108                 /*
 109                  * Page got truncated/unhashed. This will cause a 0-byte
 110                  * splice, if this is the first page.
 111                  */
 112                 if (!page->mapping) {
 113                         err = -ENODATA;
 114                         goto error;
 115                 }
 116
 117                 /*
 118                  * Uh oh, read-error from disk.
 119                  */
 120                 if (!PageUptodate(page)) {
 121                         err = -EIO;
 122                         goto error;
 123                 }
 124
 125                 /*
 126                  * Page is ok afterall, we are done.
 127                  */
 128                 unlock_page(page);
 129         }
 130
 131         return 0;
 132 error:
 133         unlock_page(page);
 134         return err;
 135 }
 136
 137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .map = generic_pipe_buf_map,
 140         .unmap = generic_pipe_buf_unmap,
 141         .confirm = page_cache_pipe_buf_confirm,
 142         .release = page_cache_pipe_buf_release,
 143         .steal = page_cache_pipe_buf_steal,
 144         .get = generic_pipe_buf_get,
 145 };
 146
 147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 148                                     struct pipe_buffer *buf)
 149 {
 150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 151                 return 1;
 152
 153         buf->flags |= PIPE_BUF_FLAG_LRU;
 154         return generic_pipe_buf_steal(pipe, buf);
 155 }
 156
 157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 158         .can_merge = 0,
 159         .map = generic_pipe_buf_map,
 160         .unmap = generic_pipe_buf_unmap,
 161         .confirm = generic_pipe_buf_confirm,
 162         .release = page_cache_pipe_buf_release,
 163         .steal = user_page_pipe_buf_steal,
 164         .get = generic_pipe_buf_get,
 165 };
 166
 167 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 168 {
 169         smp_mb();
 170         if (waitqueue_active(&pipe->wait))
 171                 wake_up_interruptible(&pipe->wait);
 172         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 173 }
 174
 175 /**
 176  * splice_to_pipe - fill passed data into a pipe
 177  * @pipe:       pipe to fill
 178  * @spd:        data to fill
 179  *
 180  * Description:
 181  *    @spd contains a map of pages and len/offset tuples, along with
 182  *    the struct pipe_buf_operations associated with these pages. This
 183  *    function will link that data to the pipe.
 184  *
 185  */
 186 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 187                        struct splice_pipe_desc *spd)
 188 {
 189         unsigned int spd_pages = spd->nr_pages;
 190         int ret, do_wakeup, page_nr;
 191
 192         ret = 0;
 193         do_wakeup = 0;
 194         page_nr = 0;
 195
 196         pipe_lock(pipe);
 197
 198         for (;;) {
 199                 if (!pipe->readers) {
 200                         send_sig(SIGPIPE, current, 0);
 201                         if (!ret)
 202                                 ret = -EPIPE;
 203                         break;
 204                 }
 205
 206                 if (pipe->nrbufs < pipe->buffers) {
 207                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 208                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 209
 210                         buf->page = spd->pages[page_nr];
 211                         buf->offset = spd->partial[page_nr].offset;
 212                         buf->len = spd->partial[page_nr].len;
 213                         buf->private = spd->partial[page_nr].private;
 214                         buf->ops = spd->ops;
 215                         if (spd->flags & SPLICE_F_GIFT)
 216                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 217
 218                         pipe->nrbufs++;
 219                         page_nr++;
 220                         ret += buf->len;
 221
 222                         if (pipe->files)
 223                                 do_wakeup = 1;
 224
 225                         if (!--spd->nr_pages)
 226                                 break;
 227                         if (pipe->nrbufs < pipe->buffers)
 228                                 continue;
 229
 230                         break;
 231                 }
 232
 233                 if (spd->flags & SPLICE_F_NONBLOCK) {
 234                         if (!ret)
 235                                 ret = -EAGAIN;
 236                         break;
 237                 }
 238
 239                 if (signal_pending(current)) {
 240                         if (!ret)
 241                                 ret = -ERESTARTSYS;
 242                         break;
 243                 }
 244
 245                 if (do_wakeup) {
 246                         smp_mb();
 247                         if (waitqueue_active(&pipe->wait))
 248                                 wake_up_interruptible_sync(&pipe->wait);
 249                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 250                         do_wakeup = 0;
 251                 }
 252
 253                 pipe->waiting_writers++;
 254                 pipe_wait(pipe);
 255                 pipe->waiting_writers--;
 256         }
 257
 258         pipe_unlock(pipe);
 259
 260         if (do_wakeup)
 261                 wakeup_pipe_readers(pipe);
 262
 263         while (page_nr < spd_pages)
 264                 spd->spd_release(spd, page_nr++);
 265
 266         return ret;
 267 }
 268
 269 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 270 {
 271         page_cache_release(spd->pages[i]);
 272 }
 273
 274 /*
 275  * Check if we need to grow the arrays holding pages and partial page
 276  * descriptions.
 277  */
 278 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 279 {
 280         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 281
 282         spd->nr_pages_max = buffers;
 283         if (buffers <= PIPE_DEF_BUFFERS)
 284                 return 0;
 285
 286         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 287         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 288
 289         if (spd->pages && spd->partial)
 290                 return 0;
 291
 292         kfree(spd->pages);
 293         kfree(spd->partial);
 294         return -ENOMEM;
 295 }
 296
 297 void splice_shrink_spd(struct splice_pipe_desc *spd)
 298 {
 299         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 300                 return;
 301
 302         kfree(spd->pages);
 303         kfree(spd->partial);
 304 }
 305
 306 static int
 307 __generic_file_splice_read(struct file *in, loff_t *ppos,
 308                            struct pipe_inode_info *pipe, size_t len,
 309                            unsigned int flags)
 310 {
 311         struct address_space *mapping = in->f_mapping;
 312         unsigned int loff, nr_pages, req_pages;
 313         struct page *pages[PIPE_DEF_BUFFERS];
 314         struct partial_page partial[PIPE_DEF_BUFFERS];
 315         struct page *page;
 316         pgoff_t index, end_index;
 317         loff_t isize;
 318         int error, page_nr;
 319         struct splice_pipe_desc spd = {
 320                 .pages = pages,
 321                 .partial = partial,
 322                 .nr_pages_max = PIPE_DEF_BUFFERS,
 323                 .flags = flags,
 324                 .ops = &page_cache_pipe_buf_ops,
 325                 .spd_release = spd_release_page,
 326         };
 327
 328         if (splice_grow_spd(pipe, &spd))
 329                 return -ENOMEM;
 330
 331         index = *ppos >> PAGE_CACHE_SHIFT;
 332         loff = *ppos & ~PAGE_CACHE_MASK;
 333         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 334         nr_pages = min(req_pages, spd.nr_pages_max);
 335
 336         /*
 337          * Lookup the (hopefully) full range of pages we need.
 338          */
 339         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 340         index += spd.nr_pages;
 341
 342         /*
 343          * If find_get_pages_contig() returned fewer pages than we needed,
 344          * readahead/allocate the rest and fill in the holes.
 345          */
 346         if (spd.nr_pages < nr_pages)
 347                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 348                                 index, req_pages - spd.nr_pages);
 349
 350         error = 0;
 351         while (spd.nr_pages < nr_pages) {
 352                 /*
 353                  * Page could be there, find_get_pages_contig() breaks on
 354                  * the first hole.
 355                  */
 356                 page = find_get_page(mapping, index);
 357                 if (!page) {
 358                         /*
 359                          * page didn't exist, allocate one.
 360                          */
 361                         page = page_cache_alloc_cold(mapping);
 362                         if (!page)
 363                                 break;
 364
 365                         error = add_to_page_cache_lru(page, mapping, index,
 366                                                 GFP_KERNEL);
 367                         if (unlikely(error)) {
 368                                 page_cache_release(page);
 369                                 if (error == -EEXIST)
 370                                         continue;
 371                                 break;
 372                         }
 373                         /*
 374                          * add_to_page_cache() locks the page, unlock it
 375                          * to avoid convoluting the logic below even more.
 376                          */
 377                         unlock_page(page);
 378                 }
 379
 380                 spd.pages[spd.nr_pages++] = page;
 381                 index++;
 382         }
 383
 384         /*
 385          * Now loop over the map and see if we need to start IO on any
 386          * pages, fill in the partial map, etc.
 387          */
 388         index = *ppos >> PAGE_CACHE_SHIFT;
 389         nr_pages = spd.nr_pages;
 390         spd.nr_pages = 0;
 391         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 392                 unsigned int this_len;
 393
 394                 if (!len)
 395                         break;
 396
 397                 /*
 398                  * this_len is the max we'll use from this page
 399                  */
 400                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 401                 page = spd.pages[page_nr];
 402
 403                 if (PageReadahead(page))
 404                         page_cache_async_readahead(mapping, &in->f_ra, in,
 405                                         page, index, req_pages - page_nr);
 406
 407                 /*
 408                  * If the page isn't uptodate, we may need to start io on it
 409                  */
 410                 if (!PageUptodate(page)) {
 411                         lock_page(page);
 412
 413                         /*
 414                          * Page was truncated, or invalidated by the
 415                          * filesystem.  Redo the find/create, but this time the
 416                          * page is kept locked, so there's no chance of another
 417                          * race with truncate/invalidate.
 418                          */
 419                         if (!page->mapping) {
 420                                 unlock_page(page);
 421                                 page = find_or_create_page(mapping, index,
 422                                                 mapping_gfp_mask(mapping));
 423
 424                                 if (!page) {
 425                                         error = -ENOMEM;
 426                                         break;
 427                                 }
 428                                 page_cache_release(spd.pages[page_nr]);
 429                                 spd.pages[page_nr] = page;
 430                         }
 431                         /*
 432                          * page was already under io and is now done, great
 433                          */
 434                         if (PageUptodate(page)) {
 435                                 unlock_page(page);
 436                                 goto fill_it;
 437                         }
 438
 439                         /*
 440                          * need to read in the page
 441                          */
 442                         error = mapping->a_ops->readpage(in, page);
 443                         if (unlikely(error)) {
 444                                 /*
 445                                  * We really should re-lookup the page here,
 446                                  * but it complicates things a lot. Instead
 447                                  * lets just do what we already stored, and
 448                                  * we'll get it the next time we are called.
 449                                  */
 450                                 if (error == AOP_TRUNCATED_PAGE)
 451                                         error = 0;
 452
 453                                 break;
 454                         }
 455                 }
 456 fill_it:
 457                 /*
 458                  * i_size must be checked after PageUptodate.
 459                  */
 460                 isize = i_size_read(mapping->host);
 461                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 462                 if (unlikely(!isize || index > end_index))
 463                         break;
 464
 465                 /*
 466                  * if this is the last page, see if we need to shrink
 467                  * the length and stop
 468                  */
 469                 if (end_index == index) {
 470                         unsigned int plen;
 471
 472                         /*
 473                          * max good bytes in this page
 474                          */
 475                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 476                         if (plen <= loff)
 477                                 break;
 478
 479                         /*
 480                          * force quit after adding this page
 481                          */
 482                         this_len = min(this_len, plen - loff);
 483                         len = this_len;
 484                 }
 485
 486                 spd.partial[page_nr].offset = loff;
 487                 spd.partial[page_nr].len = this_len;
 488                 len -= this_len;
 489                 loff = 0;
 490                 spd.nr_pages++;
 491                 index++;
 492         }
 493
 494         /*
 495          * Release any pages at the end, if we quit early. 'page_nr' is how far
 496          * we got, 'nr_pages' is how many pages are in the map.
 497          */
 498         while (page_nr < nr_pages)
 499                 page_cache_release(spd.pages[page_nr++]);
 500         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 501
 502         if (spd.nr_pages)
 503                 error = splice_to_pipe(pipe, &spd);
 504
 505         splice_shrink_spd(&spd);
 506         return error;
 507 }
 508
 509 /**
 510  * generic_file_splice_read - splice data from file to a pipe
 511  * @in:         file to splice from
 512  * @ppos:       position in @in
 513  * @pipe:       pipe to splice to
 514  * @len:        number of bytes to splice
 515  * @flags:      splice modifier flags
 516  *
 517  * Description:
 518  *    Will read pages from given file and fill them into a pipe. Can be
 519  *    used as long as the address_space operations for the source implements
 520  *    a readpage() hook.
 521  *
 522  */
 523 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 524                                  struct pipe_inode_info *pipe, size_t len,
 525                                  unsigned int flags)
 526 {
 527         loff_t isize, left;
 528         int ret;
 529
 530         isize = i_size_read(in->f_mapping->host);
 531         if (unlikely(*ppos >= isize))
 532                 return 0;
 533
 534         left = isize - *ppos;
 535         if (unlikely(left < len))
 536                 len = left;
 537
 538         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 539         if (ret > 0) {
 540                 *ppos += ret;
 541                 file_accessed(in);
 542         }
 543
 544         return ret;
 545 }
 546 EXPORT_SYMBOL(generic_file_splice_read);
 547
 548 static const struct pipe_buf_operations default_pipe_buf_ops = {
 549         .can_merge = 0,
 550         .map = generic_pipe_buf_map,
 551         .unmap = generic_pipe_buf_unmap,
 552         .confirm = generic_pipe_buf_confirm,
 553         .release = generic_pipe_buf_release,
 554         .steal = generic_pipe_buf_steal,
 555         .get = generic_pipe_buf_get,
 556 };
 557
 558 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 559                             unsigned long vlen, loff_t offset)
 560 {
 561         mm_segment_t old_fs;
 562         loff_t pos = offset;
 563         ssize_t res;
 564
 565         old_fs = get_fs();
 566         set_fs(get_ds());
 567         /* The cast to a user pointer is valid due to the set_fs() */
 568         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 569         set_fs(old_fs);
 570
 571         return res;
 572 }
 573
 574 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 575                             loff_t pos)
 576 {
 577         mm_segment_t old_fs;
 578         ssize_t res;
 579
 580         old_fs = get_fs();
 581         set_fs(get_ds());
 582         /* The cast to a user pointer is valid due to the set_fs() */
 583         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 584         set_fs(old_fs);
 585
 586         return res;
 587 }
 588 EXPORT_SYMBOL(kernel_write);
 589
 590 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 591                                  struct pipe_inode_info *pipe, size_t len,
 592                                  unsigned int flags)
 593 {
 594         unsigned int nr_pages;
 595         unsigned int nr_freed;
 596         size_t offset;
 597         struct page *pages[PIPE_DEF_BUFFERS];
 598         struct partial_page partial[PIPE_DEF_BUFFERS];
 599         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 600         ssize_t res;
 601         size_t this_len;
 602         int error;
 603         int i;
 604         struct splice_pipe_desc spd = {
 605                 .pages = pages,
 606                 .partial = partial,
 607                 .nr_pages_max = PIPE_DEF_BUFFERS,
 608                 .flags = flags,
 609                 .ops = &default_pipe_buf_ops,
 610                 .spd_release = spd_release_page,
 611         };
 612
 613         if (splice_grow_spd(pipe, &spd))
 614                 return -ENOMEM;
 615
 616         res = -ENOMEM;
 617         vec = __vec;
 618         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 619                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 620                 if (!vec)
 621                         goto shrink_ret;
 622         }
 623
 624         offset = *ppos & ~PAGE_CACHE_MASK;
 625         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 626
 627         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 628                 struct page *page;
 629
 630                 page = alloc_page(GFP_USER);
 631                 error = -ENOMEM;
 632                 if (!page)
 633                         goto err;
 634
 635                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 636                 vec[i].iov_base = (void __user *) page_address(page);
 637                 vec[i].iov_len = this_len;
 638                 spd.pages[i] = page;
 639                 spd.nr_pages++;
 640                 len -= this_len;
 641                 offset = 0;
 642         }
 643
 644         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 645         if (res < 0) {
 646                 error = res;
 647                 goto err;
 648         }
 649
 650         error = 0;
 651         if (!res)
 652                 goto err;
 653
 654         nr_freed = 0;
 655         for (i = 0; i < spd.nr_pages; i++) {
 656                 this_len = min_t(size_t, vec[i].iov_len, res);
 657                 spd.partial[i].offset = 0;
 658                 spd.partial[i].len = this_len;
 659                 if (!this_len) {
 660                         __free_page(spd.pages[i]);
 661                         spd.pages[i] = NULL;
 662                         nr_freed++;
 663                 }
 664                 res -= this_len;
 665         }
 666         spd.nr_pages -= nr_freed;
 667
 668         res = splice_to_pipe(pipe, &spd);
 669         if (res > 0)
 670                 *ppos += res;
 671
 672 shrink_ret:
 673         if (vec != __vec)
 674                 kfree(vec);
 675         splice_shrink_spd(&spd);
 676         return res;
 677
 678 err:
 679         for (i = 0; i < spd.nr_pages; i++)
 680                 __free_page(spd.pages[i]);
 681
 682         res = error;
 683         goto shrink_ret;
 684 }
 685 EXPORT_SYMBOL(default_file_splice_read);
 686
 687 /*
 688  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 689  * using sendpage(). Return the number of bytes sent.
 690  */
 691 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 692                             struct pipe_buffer *buf, struct splice_desc *sd)
 693 {
 694         struct file *file = sd->u.file;
 695         loff_t pos = sd->pos;
 696         int more;
 697
 698         if (!likely(file->f_op->sendpage))
 699                 return -EINVAL;
 700
 701         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 702
 703         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 704                 more |= MSG_SENDPAGE_NOTLAST;
 705
 706         return file->f_op->sendpage(file, buf->page, buf->offset,
 707                                     sd->len, &pos, more);
 708 }
 709
 710 /*
 711  * This is a little more tricky than the file -> pipe splicing. There are
 712  * basically three cases:
 713  *
 714  *      - Destination page already exists in the address space and there
 715  *        are users of it. For that case we have no other option that
 716  *        copying the data. Tough luck.
 717  *      - Destination page already exists in the address space, but there
 718  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 719  *        through to last case.
 720  *      - Destination page does not exist, we can add the pipe page to
 721  *        the page cache and avoid the copy.
 722  *
 723  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 724  * sd->flags), we attempt to migrate pages from the pipe to the output
 725  * file address space page cache. This is possible if no one else has
 726  * the pipe page referenced outside of the pipe and page cache. If
 727  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 728  * a new page in the output file page cache and fill/dirty that.
 729  */
 730 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 731                  struct splice_desc *sd)
 732 {
 733         struct file *file = sd->u.file;
 734         struct address_space *mapping = file->f_mapping;
 735         unsigned int offset, this_len;
 736         struct page *page;
 737         void *fsdata;
 738         int ret;
 739
 740         offset = sd->pos & ~PAGE_CACHE_MASK;
 741
 742         this_len = sd->len;
 743         if (this_len + offset > PAGE_CACHE_SIZE)
 744                 this_len = PAGE_CACHE_SIZE - offset;
 745
 746         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 747                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 748         if (unlikely(ret))
 749                 goto out;
 750
 751         if (buf->page != page) {
 752                 char *src = buf->ops->map(pipe, buf, 1);
 753                 char *dst = kmap_atomic(page);
 754
 755                 memcpy(dst + offset, src + buf->offset, this_len);
 756                 flush_dcache_page(page);
 757                 kunmap_atomic(dst);
 758                 buf->ops->unmap(pipe, buf, src);
 759         }
 760         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 761                                 page, fsdata);
 762 out:
 763         return ret;
 764 }
 765 EXPORT_SYMBOL(pipe_to_file);
 766
 767 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 768 {
 769         smp_mb();
 770         if (waitqueue_active(&pipe->wait))
 771                 wake_up_interruptible(&pipe->wait);
 772         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 773 }
 774
 775 /**
 776  * splice_from_pipe_feed - feed available data from a pipe to a file
 777  * @pipe:       pipe to splice from
 778  * @sd:         information to @actor
 779  * @actor:      handler that splices the data
 780  *
 781  * Description:
 782  *    This function loops over the pipe and calls @actor to do the
 783  *    actual moving of a single struct pipe_buffer to the desired
 784  *    destination.  It returns when there's no more buffers left in
 785  *    the pipe or if the requested number of bytes (@sd->total_len)
 786  *    have been copied.  It returns a positive number (one) if the
 787  *    pipe needs to be filled with more data, zero if the required
 788  *    number of bytes have been copied and -errno on error.
 789  *
 790  *    This, together with splice_from_pipe_{begin,end,next}, may be
 791  *    used to implement the functionality of __splice_from_pipe() when
 792  *    locking is required around copying the pipe buffers to the
 793  *    destination.
 794  */
 795 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 796                           splice_actor *actor)
 797 {
 798         int ret;
 799
 800         while (pipe->nrbufs) {
 801                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 802                 const struct pipe_buf_operations *ops = buf->ops;
 803
 804                 sd->len = buf->len;
 805                 if (sd->len > sd->total_len)
 806                         sd->len = sd->total_len;
 807
 808                 ret = buf->ops->confirm(pipe, buf);
 809                 if (unlikely(ret)) {
 810                         if (ret == -ENODATA)
 811                                 ret = 0;
 812                         return ret;
 813                 }
 814
 815                 ret = actor(pipe, buf, sd);
 816                 if (ret <= 0)
 817                         return ret;
 818
 819                 buf->offset += ret;
 820                 buf->len -= ret;
 821
 822                 sd->num_spliced += ret;
 823                 sd->len -= ret;
 824                 sd->pos += ret;
 825                 sd->total_len -= ret;
 826
 827                 if (!buf->len) {
 828                         buf->ops = NULL;
 829                         ops->release(pipe, buf);
 830                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 831                         pipe->nrbufs--;
 832                         if (pipe->files)
 833                                 sd->need_wakeup = true;
 834                 }
 835
 836                 if (!sd->total_len)
 837                         return 0;
 838         }
 839
 840         return 1;
 841 }
 842 EXPORT_SYMBOL(splice_from_pipe_feed);
 843
 844 /**
 845  * splice_from_pipe_next - wait for some data to splice from
 846  * @pipe:       pipe to splice from
 847  * @sd:         information about the splice operation
 848  *
 849  * Description:
 850  *    This function will wait for some data and return a positive
 851  *    value (one) if pipe buffers are available.  It will return zero
 852  *    or -errno if no more data needs to be spliced.
 853  */
 854 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 855 {
 856         while (!pipe->nrbufs) {
 857                 if (!pipe->writers)
 858                         return 0;
 859
 860                 if (!pipe->waiting_writers && sd->num_spliced)
 861                         return 0;
 862
 863                 if (sd->flags & SPLICE_F_NONBLOCK)
 864                         return -EAGAIN;
 865
 866                 if (signal_pending(current))
 867                         return -ERESTARTSYS;
 868
 869                 if (sd->need_wakeup) {
 870                         wakeup_pipe_writers(pipe);
 871                         sd->need_wakeup = false;
 872                 }
 873
 874                 pipe_wait(pipe);
 875         }
 876
 877         return 1;
 878 }
 879 EXPORT_SYMBOL(splice_from_pipe_next);
 880
 881 /**
 882  * splice_from_pipe_begin - start splicing from pipe
 883  * @sd:         information about the splice operation
 884  *
 885  * Description:
 886  *    This function should be called before a loop containing
 887  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 888  *    initialize the necessary fields of @sd.
 889  */
 890 void splice_from_pipe_begin(struct splice_desc *sd)
 891 {
 892         sd->num_spliced = 0;
 893         sd->need_wakeup = false;
 894 }
 895 EXPORT_SYMBOL(splice_from_pipe_begin);
 896
 897 /**
 898  * splice_from_pipe_end - finish splicing from pipe
 899  * @pipe:       pipe to splice from
 900  * @sd:         information about the splice operation
 901  *
 902  * Description:
 903  *    This function will wake up pipe writers if necessary.  It should
 904  *    be called after a loop containing splice_from_pipe_next() and
 905  *    splice_from_pipe_feed().
 906  */
 907 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 908 {
 909         if (sd->need_wakeup)
 910                 wakeup_pipe_writers(pipe);
 911 }
 912 EXPORT_SYMBOL(splice_from_pipe_end);
 913
 914 /**
 915  * __splice_from_pipe - splice data from a pipe to given actor
 916  * @pipe:       pipe to splice from
 917  * @sd:         information to @actor
 918  * @actor:      handler that splices the data
 919  *
 920  * Description:
 921  *    This function does little more than loop over the pipe and call
 922  *    @actor to do the actual moving of a single struct pipe_buffer to
 923  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 924  *    pipe_to_user.
 925  *
 926  */
 927 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 928                            splice_actor *actor)
 929 {
 930         int ret;
 931
 932         splice_from_pipe_begin(sd);
 933         do {
 934                 ret = splice_from_pipe_next(pipe, sd);
 935                 if (ret > 0)
 936                         ret = splice_from_pipe_feed(pipe, sd, actor);
 937         } while (ret > 0);
 938         splice_from_pipe_end(pipe, sd);
 939
 940         return sd->num_spliced ? sd->num_spliced : ret;
 941 }
 942 EXPORT_SYMBOL(__splice_from_pipe);
 943
 944 /**
 945  * splice_from_pipe - splice data from a pipe to a file
 946  * @pipe:       pipe to splice from
 947  * @out:        file to splice to
 948  * @ppos:       position in @out
 949  * @len:        how many bytes to splice
 950  * @flags:      splice modifier flags
 951  * @actor:      handler that splices the data
 952  *
 953  * Description:
 954  *    See __splice_from_pipe. This function locks the pipe inode,
 955  *    otherwise it's identical to __splice_from_pipe().
 956  *
 957  */
 958 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 959                          loff_t *ppos, size_t len, unsigned int flags,
 960                          splice_actor *actor)
 961 {
 962         ssize_t ret;
 963         struct splice_desc sd = {
 964                 .total_len = len,
 965                 .flags = flags,
 966                 .pos = *ppos,
 967                 .u.file = out,
 968         };
 969
 970         pipe_lock(pipe);
 971         ret = __splice_from_pipe(pipe, &sd, actor);
 972         pipe_unlock(pipe);
 973
 974         return ret;
 975 }
 976
 977 /**
 978  * generic_file_splice_write - splice data from a pipe to a file
 979  * @pipe:       pipe info
 980  * @out:        file to write to
 981  * @ppos:       position in @out
 982  * @len:        number of bytes to splice
 983  * @flags:      splice modifier flags
 984  *
 985  * Description:
 986  *    Will either move or copy pages (determined by @flags options) from
 987  *    the given pipe inode to the given file.
 988  *
 989  */
 990 ssize_t
 991 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 992                           loff_t *ppos, size_t len, unsigned int flags)
 993 {
 994         struct address_space *mapping = out->f_mapping;
 995         struct inode *inode = mapping->host;
 996         struct splice_desc sd = {
 997                 .total_len = len,
 998                 .flags = flags,
 999                 .pos = *ppos,
1000                 .u.file = out,
1001         };
1002         ssize_t ret;
1003
1004         pipe_lock(pipe);
1005
1006         splice_from_pipe_begin(&sd);
1007         do {
1008                 ret = splice_from_pipe_next(pipe, &sd);
1009                 if (ret <= 0)
1010                         break;
1011
1012                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1013                 ret = file_remove_suid(out);
1014                 if (!ret) {
1015                         ret = file_update_time(out);
1016                         if (!ret)
1017                                 ret = splice_from_pipe_feed(pipe, &sd,
1018                                                             pipe_to_file);
1019                 }
1020                 mutex_unlock(&inode->i_mutex);
1021         } while (ret > 0);
1022         splice_from_pipe_end(pipe, &sd);
1023
1024         pipe_unlock(pipe);
1025
1026         if (sd.num_spliced)
1027                 ret = sd.num_spliced;
1028
1029         if (ret > 0) {
1030                 int err;
1031
1032                 err = generic_write_sync(out, *ppos, ret);
1033                 if (err)
1034                         ret = err;
1035                 else
1036                         *ppos += ret;
1037                 balance_dirty_pages_ratelimited(mapping);
1038         }
1039
1040         return ret;
1041 }
1042
1043 EXPORT_SYMBOL(generic_file_splice_write);
1044
1045 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1046                           struct splice_desc *sd)
1047 {
1048         int ret;
1049         void *data;
1050         loff_t tmp = sd->pos;
1051
1052         data = buf->ops->map(pipe, buf, 0);
1053         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1054         buf->ops->unmap(pipe, buf, data);
1055
1056         return ret;
1057 }
1058
1059 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1060                                          struct file *out, loff_t *ppos,
1061                                          size_t len, unsigned int flags)
1062 {
1063         ssize_t ret;
1064
1065         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1066         if (ret > 0)
1067                 *ppos += ret;
1068
1069         return ret;
1070 }
1071
1072 /**
1073  * generic_splice_sendpage - splice data from a pipe to a socket
1074  * @pipe:       pipe to splice from
1075  * @out:        socket to write to
1076  * @ppos:       position in @out
1077  * @len:        number of bytes to splice
1078  * @flags:      splice modifier flags
1079  *
1080  * Description:
1081  *    Will send @len bytes from the pipe to a network socket. No data copying
1082  *    is involved.
1083  *
1084  */
1085 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1086                                 loff_t *ppos, size_t len, unsigned int flags)
1087 {
1088         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1089 }
1090
1091 EXPORT_SYMBOL(generic_splice_sendpage);
1092
1093 /*
1094  * Attempt to initiate a splice from pipe to file.
1095  */
1096 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1097                            loff_t *ppos, size_t len, unsigned int flags)
1098 {
1099         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1100                                 loff_t *, size_t, unsigned int);
1101
1102         if (out->f_op->splice_write)
1103                 splice_write = out->f_op->splice_write;
1104         else
1105                 splice_write = default_file_splice_write;
1106
1107         return splice_write(pipe, out, ppos, len, flags);
1108 }
1109
1110 /*
1111  * Attempt to initiate a splice from a file to a pipe.
1112  */
1113 static long do_splice_to(struct file *in, loff_t *ppos,
1114                          struct pipe_inode_info *pipe, size_t len,
1115                          unsigned int flags)
1116 {
1117         ssize_t (*splice_read)(struct file *, loff_t *,
1118                                struct pipe_inode_info *, size_t, unsigned int);
1119         int ret;
1120
1121         if (unlikely(!(in->f_mode & FMODE_READ)))
1122                 return -EBADF;
1123
1124         ret = rw_verify_area(READ, in, ppos, len);
1125         if (unlikely(ret < 0))
1126                 return ret;
1127
1128         if (in->f_op->splice_read)
1129                 splice_read = in->f_op->splice_read;
1130         else
1131                 splice_read = default_file_splice_read;
1132
1133         return splice_read(in, ppos, pipe, len, flags);
1134 }
1135
1136 /**
1137  * splice_direct_to_actor - splices data directly between two non-pipes
1138  * @in:         file to splice from
1139  * @sd:         actor information on where to splice to
1140  * @actor:      handles the data splicing
1141  *
1142  * Description:
1143  *    This is a special case helper to splice directly between two
1144  *    points, without requiring an explicit pipe. Internally an allocated
1145  *    pipe is cached in the process, and reused during the lifetime of
1146  *    that process.
1147  *
1148  */
1149 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1150                                splice_direct_actor *actor)
1151 {
1152         struct pipe_inode_info *pipe;
1153         long ret, bytes;
1154         umode_t i_mode;
1155         size_t len;
1156         int i, flags;
1157
1158         /*
1159          * We require the input being a regular file, as we don't want to
1160          * randomly drop data for eg socket -> socket splicing. Use the
1161          * piped splicing for that!
1162          */
1163         i_mode = file_inode(in)->i_mode;
1164         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1165                 return -EINVAL;
1166
1167         /*
1168          * neither in nor out is a pipe, setup an internal pipe attached to
1169          * 'out' and transfer the wanted data from 'in' to 'out' through that
1170          */
1171         pipe = current->splice_pipe;
1172         if (unlikely(!pipe)) {
1173                 pipe = alloc_pipe_info();
1174                 if (!pipe)
1175                         return -ENOMEM;
1176
1177                 /*
1178                  * We don't have an immediate reader, but we'll read the stuff
1179                  * out of the pipe right after the splice_to_pipe(). So set
1180                  * PIPE_READERS appropriately.
1181                  */
1182                 pipe->readers = 1;
1183
1184                 current->splice_pipe = pipe;
1185         }
1186
1187         /*
1188          * Do the splice.
1189          */
1190         ret = 0;
1191         bytes = 0;
1192         len = sd->total_len;
1193         flags = sd->flags;
1194
1195         /*
1196          * Don't block on output, we have to drain the direct pipe.
1197          */
1198         sd->flags &= ~SPLICE_F_NONBLOCK;
1199
1200         while (len) {
1201                 size_t read_len;
1202                 loff_t pos = sd->pos, prev_pos = pos;
1203
1204                 ret = do_splice_to(in, &pos, pipe, len, flags);
1205                 if (unlikely(ret <= 0))
1206                         goto out_release;
1207
1208                 read_len = ret;
1209                 sd->total_len = read_len;
1210
1211                 /*
1212                  * NOTE: nonblocking mode only applies to the input. We
1213                  * must not do the output in nonblocking mode as then we
1214                  * could get stuck data in the internal pipe:
1215                  */
1216                 ret = actor(pipe, sd);
1217                 if (unlikely(ret <= 0)) {
1218                         sd->pos = prev_pos;
1219                         goto out_release;
1220                 }
1221
1222                 bytes += ret;
1223                 len -= ret;
1224                 sd->pos = pos;
1225
1226                 if (ret < read_len) {
1227                         sd->pos = prev_pos + ret;
1228                         goto out_release;
1229                 }
1230         }
1231
1232 done:
1233         pipe->nrbufs = pipe->curbuf = 0;
1234         file_accessed(in);
1235         return bytes;
1236
1237 out_release:
1238         /*
1239          * If we did an incomplete transfer we must release
1240          * the pipe buffers in question:
1241          */
1242         for (i = 0; i < pipe->buffers; i++) {
1243                 struct pipe_buffer *buf = pipe->bufs + i;
1244
1245                 if (buf->ops) {
1246                         buf->ops->release(pipe, buf);
1247                         buf->ops = NULL;
1248                 }
1249         }
1250
1251         if (!bytes)
1252                 bytes = ret;
1253
1254         goto done;
1255 }
1256 EXPORT_SYMBOL(splice_direct_to_actor);
1257
1258 static int direct_splice_actor(struct pipe_inode_info *pipe,
1259                                struct splice_desc *sd)
1260 {
1261         struct file *file = sd->u.file;
1262
1263         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1264                               sd->flags);
1265 }
1266
1267 /**
1268  * do_splice_direct - splices data directly between two files
1269  * @in:         file to splice from
1270  * @ppos:       input file offset
1271  * @out:        file to splice to
1272  * @opos:       output file offset
1273  * @len:        number of bytes to splice
1274  * @flags:      splice modifier flags
1275  *
1276  * Description:
1277  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1278  *    doing it in the application would incur an extra system call
1279  *    (splice in + splice out, as compared to just sendfile()). So this helper
1280  *    can splice directly through a process-private pipe.
1281  *
1282  */
1283 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1284                       loff_t *opos, size_t len, unsigned int flags)
1285 {
1286         struct splice_desc sd = {
1287                 .len            = len,
1288                 .total_len      = len,
1289                 .flags          = flags,
1290                 .pos            = *ppos,
1291                 .u.file         = out,
1292                 .opos           = opos,
1293         };
1294         long ret;
1295
1296         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1297                 return -EBADF;
1298
1299         if (unlikely(out->f_flags & O_APPEND))
1300                 return -EINVAL;
1301
1302         ret = rw_verify_area(WRITE, out, opos, len);
1303         if (unlikely(ret < 0))
1304                 return ret;
1305
1306         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1307         if (ret > 0)
1308                 *ppos = sd.pos;
1309
1310         return ret;
1311 }
1312
1313 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1314                                struct pipe_inode_info *opipe,
1315                                size_t len, unsigned int flags);
1316
1317 /*
1318  * Determine where to splice to/from.
1319  */
1320 static long do_splice(struct file *in, loff_t __user *off_in,
1321                       struct file *out, loff_t __user *off_out,
1322                       size_t len, unsigned int flags)
1323 {
1324         struct pipe_inode_info *ipipe;
1325         struct pipe_inode_info *opipe;
1326         loff_t offset;
1327         long ret;
1328
1329         ipipe = get_pipe_info(in);
1330         opipe = get_pipe_info(out);
1331
1332         if (ipipe && opipe) {
1333                 if (off_in || off_out)
1334                         return -ESPIPE;
1335
1336                 if (!(in->f_mode & FMODE_READ))
1337                         return -EBADF;
1338
1339                 if (!(out->f_mode & FMODE_WRITE))
1340                         return -EBADF;
1341
1342                 /* Splicing to self would be fun, but... */
1343                 if (ipipe == opipe)
1344                         return -EINVAL;
1345
1346                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1347         }
1348
1349         if (ipipe) {
1350                 if (off_in)
1351                         return -ESPIPE;
1352                 if (off_out) {
1353                         if (!(out->f_mode & FMODE_PWRITE))
1354                                 return -EINVAL;
1355                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1356                                 return -EFAULT;
1357                 } else {
1358                         offset = out->f_pos;
1359                 }
1360
1361                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1362                         return -EBADF;
1363
1364                 if (unlikely(out->f_flags & O_APPEND))
1365                         return -EINVAL;
1366
1367                 ret = rw_verify_area(WRITE, out, &offset, len);
1368                 if (unlikely(ret < 0))
1369                         return ret;
1370
1371                 file_start_write(out);
1372                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1373                 file_end_write(out);
1374
1375                 if (!off_out)
1376                         out->f_pos = offset;
1377                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1378                         ret = -EFAULT;
1379
1380                 return ret;
1381         }
1382
1383         if (opipe) {
1384                 if (off_out)
1385                         return -ESPIPE;
1386                 if (off_in) {
1387                         if (!(in->f_mode & FMODE_PREAD))
1388                                 return -EINVAL;
1389                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1390                                 return -EFAULT;
1391                 } else {
1392                         offset = in->f_pos;
1393                 }
1394
1395                 ret = do_splice_to(in, &offset, opipe, len, flags);
1396
1397                 if (!off_in)
1398                         in->f_pos = offset;
1399                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1400                         ret = -EFAULT;
1401
1402                 return ret;
1403         }
1404
1405         return -EINVAL;
1406 }
1407
1408 /*
1409  * Map an iov into an array of pages and offset/length tupples. With the
1410  * partial_page structure, we can map several non-contiguous ranges into
1411  * our ones pages[] map instead of splitting that operation into pieces.
1412  * Could easily be exported as a generic helper for other users, in which
1413  * case one would probably want to add a 'max_nr_pages' parameter as well.
1414  */
1415 static int get_iovec_page_array(const struct iovec __user *iov,
1416                                 unsigned int nr_vecs, struct page **pages,
1417                                 struct partial_page *partial, bool aligned,
1418                                 unsigned int pipe_buffers)
1419 {
1420         int buffers = 0, error = 0;
1421
1422         while (nr_vecs) {
1423                 unsigned long off, npages;
1424                 struct iovec entry;
1425                 void __user *base;
1426                 size_t len;
1427                 int i;
1428
1429                 error = -EFAULT;
1430                 if (copy_from_user(&entry, iov, sizeof(entry)))
1431                         break;
1432
1433                 base = entry.iov_base;
1434                 len = entry.iov_len;
1435
1436                 /*
1437                  * Sanity check this iovec. 0 read succeeds.
1438                  */
1439                 error = 0;
1440                 if (unlikely(!len))
1441                         break;
1442                 error = -EFAULT;
1443                 if (!access_ok(VERIFY_READ, base, len))
1444                         break;
1445
1446                 /*
1447                  * Get this base offset and number of pages, then map
1448                  * in the user pages.
1449                  */
1450                 off = (unsigned long) base & ~PAGE_MASK;
1451
1452                 /*
1453                  * If asked for alignment, the offset must be zero and the
1454                  * length a multiple of the PAGE_SIZE.
1455                  */
1456                 error = -EINVAL;
1457                 if (aligned && (off || len & ~PAGE_MASK))
1458                         break;
1459
1460                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1461                 if (npages > pipe_buffers - buffers)
1462                         npages = pipe_buffers - buffers;
1463
1464                 error = get_user_pages_fast((unsigned long)base, npages,
1465                                         0, &pages[buffers]);
1466
1467                 if (unlikely(error <= 0))
1468                         break;
1469
1470                 /*
1471                  * Fill this contiguous range into the partial page map.
1472                  */
1473                 for (i = 0; i < error; i++) {
1474                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1475
1476                         partial[buffers].offset = off;
1477                         partial[buffers].len = plen;
1478
1479                         off = 0;
1480                         len -= plen;
1481                         buffers++;
1482                 }
1483
1484                 /*
1485                  * We didn't complete this iov, stop here since it probably
1486                  * means we have to move some of this into a pipe to
1487                  * be able to continue.
1488                  */
1489                 if (len)
1490                         break;
1491
1492                 /*
1493                  * Don't continue if we mapped fewer pages than we asked for,
1494                  * or if we mapped the max number of pages that we have
1495                  * room for.
1496                  */
1497                 if (error < npages || buffers == pipe_buffers)
1498                         break;
1499
1500                 nr_vecs--;
1501                 iov++;
1502         }
1503
1504         if (buffers)
1505                 return buffers;
1506
1507         return error;
1508 }
1509
1510 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1511                         struct splice_desc *sd)
1512 {
1513         char *src;
1514         int ret;
1515
1516         /*
1517          * See if we can use the atomic maps, by prefaulting in the
1518          * pages and doing an atomic copy
1519          */
1520         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1521                 src = buf->ops->map(pipe, buf, 1);
1522                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1523                                                         sd->len);
1524                 buf->ops->unmap(pipe, buf, src);
1525                 if (!ret) {
1526                         ret = sd->len;
1527                         goto out;
1528                 }
1529         }
1530
1531         /*
1532          * No dice, use slow non-atomic map and copy
1533          */
1534         src = buf->ops->map(pipe, buf, 0);
1535
1536         ret = sd->len;
1537         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1538                 ret = -EFAULT;
1539
1540         buf->ops->unmap(pipe, buf, src);
1541 out:
1542         if (ret > 0)
1543                 sd->u.userptr += ret;
1544         return ret;
1545 }
1546
1547 /*
1548  * For lack of a better implementation, implement vmsplice() to userspace
1549  * as a simple copy of the pipes pages to the user iov.
1550  */
1551 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1552                              unsigned long nr_segs, unsigned int flags)
1553 {
1554         struct pipe_inode_info *pipe;
1555         struct splice_desc sd;
1556         ssize_t size;
1557         int error;
1558         long ret;
1559
1560         pipe = get_pipe_info(file);
1561         if (!pipe)
1562                 return -EBADF;
1563
1564         pipe_lock(pipe);
1565
1566         error = ret = 0;
1567         while (nr_segs) {
1568                 void __user *base;
1569                 size_t len;
1570
1571                 /*
1572                  * Get user address base and length for this iovec.
1573                  */
1574                 error = get_user(base, &iov->iov_base);
1575                 if (unlikely(error))
1576                         break;
1577                 error = get_user(len, &iov->iov_len);
1578                 if (unlikely(error))
1579                         break;
1580
1581                 /*
1582                  * Sanity check this iovec. 0 read succeeds.
1583                  */
1584                 if (unlikely(!len))
1585                         break;
1586                 if (unlikely(!base)) {
1587                         error = -EFAULT;
1588                         break;
1589                 }
1590
1591                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1592                         error = -EFAULT;
1593                         break;
1594                 }
1595
1596                 sd.len = 0;
1597                 sd.total_len = len;
1598                 sd.flags = flags;
1599                 sd.u.userptr = base;
1600                 sd.pos = 0;
1601
1602                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1603                 if (size < 0) {
1604                         if (!ret)
1605                                 ret = size;
1606
1607                         break;
1608                 }
1609
1610                 ret += size;
1611
1612                 if (size < len)
1613                         break;
1614
1615                 nr_segs--;
1616                 iov++;
1617         }
1618
1619         pipe_unlock(pipe);
1620
1621         if (!ret)
1622                 ret = error;
1623
1624         return ret;
1625 }
1626
1627 /*
1628  * vmsplice splices a user address range into a pipe. It can be thought of
1629  * as splice-from-memory, where the regular splice is splice-from-file (or
1630  * to file). In both cases the output is a pipe, naturally.
1631  */
1632 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1633                              unsigned long nr_segs, unsigned int flags)
1634 {
1635         struct pipe_inode_info *pipe;
1636         struct page *pages[PIPE_DEF_BUFFERS];
1637         struct partial_page partial[PIPE_DEF_BUFFERS];
1638         struct splice_pipe_desc spd = {
1639                 .pages = pages,
1640                 .partial = partial,
1641                 .nr_pages_max = PIPE_DEF_BUFFERS,
1642                 .flags = flags,
1643                 .ops = &user_page_pipe_buf_ops,
1644                 .spd_release = spd_release_page,
1645         };
1646         long ret;
1647
1648         pipe = get_pipe_info(file);
1649         if (!pipe)
1650                 return -EBADF;
1651
1652         if (splice_grow_spd(pipe, &spd))
1653                 return -ENOMEM;
1654
1655         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1656                                             spd.partial, false,
1657                                             spd.nr_pages_max);
1658         if (spd.nr_pages <= 0)
1659                 ret = spd.nr_pages;
1660         else
1661                 ret = splice_to_pipe(pipe, &spd);
1662
1663         splice_shrink_spd(&spd);
1664         return ret;
1665 }
1666
1667 /*
1668  * Note that vmsplice only really supports true splicing _from_ user memory
1669  * to a pipe, not the other way around. Splicing from user memory is a simple
1670  * operation that can be supported without any funky alignment restrictions
1671  * or nasty vm tricks. We simply map in the user memory and fill them into
1672  * a pipe. The reverse isn't quite as easy, though. There are two possible
1673  * solutions for that:
1674  *
1675  *      - memcpy() the data internally, at which point we might as well just
1676  *        do a regular read() on the buffer anyway.
1677  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1678  *        has restriction limitations on both ends of the pipe).
1679  *
1680  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1681  *
1682  */
1683 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1684                 unsigned long, nr_segs, unsigned int, flags)
1685 {
1686         struct fd f;
1687         long error;
1688
1689         if (unlikely(nr_segs > UIO_MAXIOV))
1690                 return -EINVAL;
1691         else if (unlikely(!nr_segs))
1692                 return 0;
1693
1694         error = -EBADF;
1695         f = fdget(fd);
1696         if (f.file) {
1697                 if (f.file->f_mode & FMODE_WRITE)
1698                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1699                 else if (f.file->f_mode & FMODE_READ)
1700                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1701
1702                 fdput(f);
1703         }
1704
1705         return error;
1706 }
1707
1708 #ifdef CONFIG_COMPAT
1709 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1710                     unsigned int, nr_segs, unsigned int, flags)
1711 {
1712         unsigned i;
1713         struct iovec __user *iov;
1714         if (nr_segs > UIO_MAXIOV)
1715                 return -EINVAL;
1716         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1717         for (i = 0; i < nr_segs; i++) {
1718                 struct compat_iovec v;
1719                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1720                     get_user(v.iov_len, &iov32[i].iov_len) ||
1721                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1722                     put_user(v.iov_len, &iov[i].iov_len))
1723                         return -EFAULT;
1724         }
1725         return sys_vmsplice(fd, iov, nr_segs, flags);
1726 }
1727 #endif
1728
1729 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1730                 int, fd_out, loff_t __user *, off_out,
1731                 size_t, len, unsigned int, flags)
1732 {
1733         struct fd in, out;
1734         long error;
1735
1736         if (unlikely(!len))
1737                 return 0;
1738
1739         error = -EBADF;
1740         in = fdget(fd_in);
1741         if (in.file) {
1742                 if (in.file->f_mode & FMODE_READ) {
1743                         out = fdget(fd_out);
1744                         if (out.file) {
1745                                 if (out.file->f_mode & FMODE_WRITE)
1746                                         error = do_splice(in.file, off_in,
1747                                                           out.file, off_out,
1748                                                           len, flags);
1749                                 fdput(out);
1750                         }
1751                 }
1752                 fdput(in);
1753         }
1754         return error;
1755 }
1756
1757 /*
1758  * Make sure there's data to read. Wait for input if we can, otherwise
1759  * return an appropriate error.
1760  */
1761 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1762 {
1763         int ret;
1764
1765         /*
1766          * Check ->nrbufs without the inode lock first. This function
1767          * is speculative anyways, so missing one is ok.
1768          */
1769         if (pipe->nrbufs)
1770                 return 0;
1771
1772         ret = 0;
1773         pipe_lock(pipe);
1774
1775         while (!pipe->nrbufs) {
1776                 if (signal_pending(current)) {
1777                         ret = -ERESTARTSYS;
1778                         break;
1779                 }
1780                 if (!pipe->writers)
1781                         break;
1782                 if (!pipe->waiting_writers) {
1783                         if (flags & SPLICE_F_NONBLOCK) {
1784                                 ret = -EAGAIN;
1785                                 break;
1786                         }
1787                 }
1788                 pipe_wait(pipe);
1789         }
1790
1791         pipe_unlock(pipe);
1792         return ret;
1793 }
1794
1795 /*
1796  * Make sure there's writeable room. Wait for room if we can, otherwise
1797  * return an appropriate error.
1798  */
1799 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1800 {
1801         int ret;
1802
1803         /*
1804          * Check ->nrbufs without the inode lock first. This function
1805          * is speculative anyways, so missing one is ok.
1806          */
1807         if (pipe->nrbufs < pipe->buffers)
1808                 return 0;
1809
1810         ret = 0;
1811         pipe_lock(pipe);
1812
1813         while (pipe->nrbufs >= pipe->buffers) {
1814                 if (!pipe->readers) {
1815                         send_sig(SIGPIPE, current, 0);
1816                         ret = -EPIPE;
1817                         break;
1818                 }
1819                 if (flags & SPLICE_F_NONBLOCK) {
1820                         ret = -EAGAIN;
1821                         break;
1822                 }
1823                 if (signal_pending(current)) {
1824                         ret = -ERESTARTSYS;
1825                         break;
1826                 }
1827                 pipe->waiting_writers++;
1828                 pipe_wait(pipe);
1829                 pipe->waiting_writers--;
1830         }
1831
1832         pipe_unlock(pipe);
1833         return ret;
1834 }
1835
1836 /*
1837  * Splice contents of ipipe to opipe.
1838  */
1839 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1840                                struct pipe_inode_info *opipe,
1841                                size_t len, unsigned int flags)
1842 {
1843         struct pipe_buffer *ibuf, *obuf;
1844         int ret = 0, nbuf;
1845         bool input_wakeup = false;
1846
1847
1848 retry:
1849         ret = ipipe_prep(ipipe, flags);
1850         if (ret)
1851                 return ret;
1852
1853         ret = opipe_prep(opipe, flags);
1854         if (ret)
1855                 return ret;
1856
1857         /*
1858          * Potential ABBA deadlock, work around it by ordering lock
1859          * grabbing by pipe info address. Otherwise two different processes
1860          * could deadlock (one doing tee from A -> B, the other from B -> A).
1861          */
1862         pipe_double_lock(ipipe, opipe);
1863
1864         do {
1865                 if (!opipe->readers) {
1866                         send_sig(SIGPIPE, current, 0);
1867                         if (!ret)
1868                                 ret = -EPIPE;
1869                         break;
1870                 }
1871
1872                 if (!ipipe->nrbufs && !ipipe->writers)
1873                         break;
1874
1875                 /*
1876                  * Cannot make any progress, because either the input
1877                  * pipe is empty or the output pipe is full.
1878                  */
1879                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1880                         /* Already processed some buffers, break */
1881                         if (ret)
1882                                 break;
1883
1884                         if (flags & SPLICE_F_NONBLOCK) {
1885                                 ret = -EAGAIN;
1886                                 break;
1887                         }
1888
1889                         /*
1890                          * We raced with another reader/writer and haven't
1891                          * managed to process any buffers.  A zero return
1892                          * value means EOF, so retry instead.
1893                          */
1894                         pipe_unlock(ipipe);
1895                         pipe_unlock(opipe);
1896                         goto retry;
1897                 }
1898
1899                 ibuf = ipipe->bufs + ipipe->curbuf;
1900                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1901                 obuf = opipe->bufs + nbuf;
1902
1903                 if (len >= ibuf->len) {
1904                         /*
1905                          * Simply move the whole buffer from ipipe to opipe
1906                          */
1907                         *obuf = *ibuf;
1908                         ibuf->ops = NULL;
1909                         opipe->nrbufs++;
1910                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1911                         ipipe->nrbufs--;
1912                         input_wakeup = true;
1913                 } else {
1914                         /*
1915                          * Get a reference to this pipe buffer,
1916                          * so we can copy the contents over.
1917                          */
1918                         ibuf->ops->get(ipipe, ibuf);
1919                         *obuf = *ibuf;
1920
1921                         /*
1922                          * Don't inherit the gift flag, we need to
1923                          * prevent multiple steals of this page.
1924                          */
1925                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1926
1927                         obuf->len = len;
1928                         opipe->nrbufs++;
1929                         ibuf->offset += obuf->len;
1930                         ibuf->len -= obuf->len;
1931                 }
1932                 ret += obuf->len;
1933                 len -= obuf->len;
1934         } while (len);
1935
1936         pipe_unlock(ipipe);
1937         pipe_unlock(opipe);
1938
1939         /*
1940          * If we put data in the output pipe, wakeup any potential readers.
1941          */
1942         if (ret > 0)
1943                 wakeup_pipe_readers(opipe);
1944
1945         if (input_wakeup)
1946                 wakeup_pipe_writers(ipipe);
1947
1948         return ret;
1949 }
1950
1951 /*
1952  * Link contents of ipipe to opipe.
1953  */
1954 static int link_pipe(struct pipe_inode_info *ipipe,
1955                      struct pipe_inode_info *opipe,
1956                      size_t len, unsigned int flags)
1957 {
1958         struct pipe_buffer *ibuf, *obuf;
1959         int ret = 0, i = 0, nbuf;
1960
1961         /*
1962          * Potential ABBA deadlock, work around it by ordering lock
1963          * grabbing by pipe info address. Otherwise two different processes
1964          * could deadlock (one doing tee from A -> B, the other from B -> A).
1965          */
1966         pipe_double_lock(ipipe, opipe);
1967
1968         do {
1969                 if (!opipe->readers) {
1970                         send_sig(SIGPIPE, current, 0);
1971                         if (!ret)
1972                                 ret = -EPIPE;
1973                         break;
1974                 }
1975
1976                 /*
1977                  * If we have iterated all input buffers or ran out of
1978                  * output room, break.
1979                  */
1980                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1981                         break;
1982
1983                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1984                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1985
1986                 /*
1987                  * Get a reference to this pipe buffer,
1988                  * so we can copy the contents over.
1989                  */
1990                 ibuf->ops->get(ipipe, ibuf);
1991
1992                 obuf = opipe->bufs + nbuf;
1993                 *obuf = *ibuf;
1994
1995                 /*
1996                  * Don't inherit the gift flag, we need to
1997                  * prevent multiple steals of this page.
1998                  */
1999                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2000
2001                 if (obuf->len > len)
2002                         obuf->len = len;
2003
2004                 opipe->nrbufs++;
2005                 ret += obuf->len;
2006                 len -= obuf->len;
2007                 i++;
2008         } while (len);
2009
2010         /*
2011          * return EAGAIN if we have the potential of some data in the
2012          * future, otherwise just return 0
2013          */
2014         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2015                 ret = -EAGAIN;
2016
2017         pipe_unlock(ipipe);
2018         pipe_unlock(opipe);
2019
2020         /*
2021          * If we put data in the output pipe, wakeup any potential readers.
2022          */
2023         if (ret > 0)
2024                 wakeup_pipe_readers(opipe);
2025
2026         return ret;
2027 }
2028
2029 /*
2030  * This is a tee(1) implementation that works on pipes. It doesn't copy
2031  * any data, it simply references the 'in' pages on the 'out' pipe.
2032  * The 'flags' used are the SPLICE_F_* variants, currently the only
2033  * applicable one is SPLICE_F_NONBLOCK.
2034  */
2035 static long do_tee(struct file *in, struct file *out, size_t len,
2036                    unsigned int flags)
2037 {
2038         struct pipe_inode_info *ipipe = get_pipe_info(in);
2039         struct pipe_inode_info *opipe = get_pipe_info(out);
2040         int ret = -EINVAL;
2041
2042         /*
2043          * Duplicate the contents of ipipe to opipe without actually
2044          * copying the data.
2045          */
2046         if (ipipe && opipe && ipipe != opipe) {
2047                 /*
2048                  * Keep going, unless we encounter an error. The ipipe/opipe
2049                  * ordering doesn't really matter.
2050                  */
2051                 ret = ipipe_prep(ipipe, flags);
2052                 if (!ret) {
2053                         ret = opipe_prep(opipe, flags);
2054                         if (!ret)
2055                                 ret = link_pipe(ipipe, opipe, len, flags);
2056                 }
2057         }
2058
2059         return ret;
2060 }
2061
2062 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2063 {
2064         struct fd in;
2065         int error;
2066
2067         if (unlikely(!len))
2068                 return 0;
2069
2070         error = -EBADF;
2071         in = fdget(fdin);
2072         if (in.file) {
2073                 if (in.file->f_mode & FMODE_READ) {
2074                         struct fd out = fdget(fdout);
2075                         if (out.file) {
2076                                 if (out.file->f_mode & FMODE_WRITE)
2077                                         error = do_tee(in.file, out.file,
2078                                                         len, flags);
2079                                 fdput(out);
2080                         }
2081                 }
2082                 fdput(in);
2083         }
2084
2085         return error;
2086 }