fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/module.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33 #include <linux/gfp.h>
  34 #include <linux/socket.h>
  35
  36 /*
  37  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  38  * a vm helper function, it's already simplified quite a bit by the
  39  * addition of remove_mapping(). If success is returned, the caller may
  40  * attempt to reuse this page for another destination.
  41  */
  42 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  43                                      struct pipe_buffer *buf)
  44 {
  45         struct page *page = buf->page;
  46         struct address_space *mapping;
  47
  48         lock_page(page);
  49
  50         mapping = page_mapping(page);
  51         if (mapping) {
  52                 WARN_ON(!PageUptodate(page));
  53
  54                 /*
  55                  * At least for ext2 with nobh option, we need to wait on
  56                  * writeback completing on this page, since we'll remove it
  57                  * from the pagecache.  Otherwise truncate wont wait on the
  58                  * page, allowing the disk blocks to be reused by someone else
  59                  * before we actually wrote our data to them. fs corruption
  60                  * ensues.
  61                  */
  62                 wait_on_page_writeback(page);
  63
  64                 if (page_has_private(page) &&
  65                     !try_to_release_page(page, GFP_KERNEL))
  66                         goto out_unlock;
  67
  68                 /*
  69                  * If we succeeded in removing the mapping, set LRU flag
  70                  * and return good.
  71                  */
  72                 if (remove_mapping(mapping, page)) {
  73                         buf->flags |= PIPE_BUF_FLAG_LRU;
  74                         return 0;
  75                 }
  76         }
  77
  78         /*
  79          * Raced with truncate or failed to remove page from current
  80          * address space, unlock and return failure.
  81          */
  82 out_unlock:
  83         unlock_page(page);
  84         return 1;
  85 }
  86
  87 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  88                                         struct pipe_buffer *buf)
  89 {
  90         page_cache_release(buf->page);
  91         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  92 }
  93
  94 /*
  95  * Check whether the contents of buf is OK to access. Since the content
  96  * is a page cache page, IO may be in flight.
  97  */
  98 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  99                                        struct pipe_buffer *buf)
 100 {
 101         struct page *page = buf->page;
 102         int err;
 103
 104         if (!PageUptodate(page)) {
 105                 lock_page(page);
 106
 107                 /*
 108                  * Page got truncated/unhashed. This will cause a 0-byte
 109                  * splice, if this is the first page.
 110                  */
 111                 if (!page->mapping) {
 112                         err = -ENODATA;
 113                         goto error;
 114                 }
 115
 116                 /*
 117                  * Uh oh, read-error from disk.
 118                  */
 119                 if (!PageUptodate(page)) {
 120                         err = -EIO;
 121                         goto error;
 122                 }
 123
 124                 /*
 125                  * Page is ok afterall, we are done.
 126                  */
 127                 unlock_page(page);
 128         }
 129
 130         return 0;
 131 error:
 132         unlock_page(page);
 133         return err;
 134 }
 135
 136 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 137         .can_merge = 0,
 138         .map = generic_pipe_buf_map,
 139         .unmap = generic_pipe_buf_unmap,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 158         .map = generic_pipe_buf_map,
 159         .unmap = generic_pipe_buf_unmap,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 189         int ret, do_wakeup, page_nr;
 190
 191         ret = 0;
 192         do_wakeup = 0;
 193         page_nr = 0;
 194
 195         pipe_lock(pipe);
 196
 197         for (;;) {
 198                 if (!pipe->readers) {
 199                         send_sig(SIGPIPE, current, 0);
 200                         if (!ret)
 201                                 ret = -EPIPE;
 202                         break;
 203                 }
 204
 205                 if (pipe->nrbufs < pipe->buffers) {
 206                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 207                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 208
 209                         buf->page = spd->pages[page_nr];
 210                         buf->offset = spd->partial[page_nr].offset;
 211                         buf->len = spd->partial[page_nr].len;
 212                         buf->private = spd->partial[page_nr].private;
 213                         buf->ops = spd->ops;
 214                         if (spd->flags & SPLICE_F_GIFT)
 215                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 216
 217                         pipe->nrbufs++;
 218                         page_nr++;
 219                         ret += buf->len;
 220
 221                         if (pipe->inode)
 222                                 do_wakeup = 1;
 223
 224                         if (!--spd->nr_pages)
 225                                 break;
 226                         if (pipe->nrbufs < pipe->buffers)
 227                                 continue;
 228
 229                         break;
 230                 }
 231
 232                 if (spd->flags & SPLICE_F_NONBLOCK) {
 233                         if (!ret)
 234                                 ret = -EAGAIN;
 235                         break;
 236                 }
 237
 238                 if (signal_pending(current)) {
 239                         if (!ret)
 240                                 ret = -ERESTARTSYS;
 241                         break;
 242                 }
 243
 244                 if (do_wakeup) {
 245                         smp_mb();
 246                         if (waitqueue_active(&pipe->wait))
 247                                 wake_up_interruptible_sync(&pipe->wait);
 248                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 249                         do_wakeup = 0;
 250                 }
 251
 252                 pipe->waiting_writers++;
 253                 pipe_wait(pipe);
 254                 pipe->waiting_writers--;
 255         }
 256
 257         pipe_unlock(pipe);
 258
 259         if (do_wakeup)
 260                 wakeup_pipe_readers(pipe);
 261
 262         while (page_nr < spd_pages)
 263                 spd->spd_release(spd, page_nr++);
 264
 265         return ret;
 266 }
 267
 268 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 269 {
 270         page_cache_release(spd->pages[i]);
 271 }
 272
 273 /*
 274  * Check if we need to grow the arrays holding pages and partial page
 275  * descriptions.
 276  */
 277 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 278 {
 279         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 280
 281         spd->nr_pages_max = buffers;
 282         if (buffers <= PIPE_DEF_BUFFERS)
 283                 return 0;
 284
 285         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 286         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 287
 288         if (spd->pages && spd->partial)
 289                 return 0;
 290
 291         kfree(spd->pages);
 292         kfree(spd->partial);
 293         return -ENOMEM;
 294 }
 295
 296 void splice_shrink_spd(struct splice_pipe_desc *spd)
 297 {
 298         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 299                 return;
 300
 301         kfree(spd->pages);
 302         kfree(spd->partial);
 303 }
 304
 305 static int
 306 __generic_file_splice_read(struct file *in, loff_t *ppos,
 307                            struct pipe_inode_info *pipe, size_t len,
 308                            unsigned int flags)
 309 {
 310         struct address_space *mapping = in->f_mapping;
 311         unsigned int loff, nr_pages, req_pages;
 312         struct page *pages[PIPE_DEF_BUFFERS];
 313         struct partial_page partial[PIPE_DEF_BUFFERS];
 314         struct page *page;
 315         pgoff_t index, end_index;
 316         loff_t isize;
 317         int error, page_nr;
 318         struct splice_pipe_desc spd = {
 319                 .pages = pages,
 320                 .partial = partial,
 321                 .nr_pages_max = PIPE_DEF_BUFFERS,
 322                 .flags = flags,
 323                 .ops = &page_cache_pipe_buf_ops,
 324                 .spd_release = spd_release_page,
 325         };
 326
 327         if (splice_grow_spd(pipe, &spd))
 328                 return -ENOMEM;
 329
 330         index = *ppos >> PAGE_CACHE_SHIFT;
 331         loff = *ppos & ~PAGE_CACHE_MASK;
 332         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 333         nr_pages = min(req_pages, spd.nr_pages_max);
 334
 335         /*
 336          * Lookup the (hopefully) full range of pages we need.
 337          */
 338         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 339         index += spd.nr_pages;
 340
 341         /*
 342          * If find_get_pages_contig() returned fewer pages than we needed,
 343          * readahead/allocate the rest and fill in the holes.
 344          */
 345         if (spd.nr_pages < nr_pages)
 346                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 347                                 index, req_pages - spd.nr_pages);
 348
 349         error = 0;
 350         while (spd.nr_pages < nr_pages) {
 351                 /*
 352                  * Page could be there, find_get_pages_contig() breaks on
 353                  * the first hole.
 354                  */
 355                 page = find_get_page(mapping, index);
 356                 if (!page) {
 357                         /*
 358                          * page didn't exist, allocate one.
 359                          */
 360                         page = page_cache_alloc_cold(mapping);
 361                         if (!page)
 362                                 break;
 363
 364                         error = add_to_page_cache_lru(page, mapping, index,
 365                                                 GFP_KERNEL);
 366                         if (unlikely(error)) {
 367                                 page_cache_release(page);
 368                                 if (error == -EEXIST)
 369                                         continue;
 370                                 break;
 371                         }
 372                         /*
 373                          * add_to_page_cache() locks the page, unlock it
 374                          * to avoid convoluting the logic below even more.
 375                          */
 376                         unlock_page(page);
 377                 }
 378
 379                 spd.pages[spd.nr_pages++] = page;
 380                 index++;
 381         }
 382
 383         /*
 384          * Now loop over the map and see if we need to start IO on any
 385          * pages, fill in the partial map, etc.
 386          */
 387         index = *ppos >> PAGE_CACHE_SHIFT;
 388         nr_pages = spd.nr_pages;
 389         spd.nr_pages = 0;
 390         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 391                 unsigned int this_len;
 392
 393                 if (!len)
 394                         break;
 395
 396                 /*
 397                  * this_len is the max we'll use from this page
 398                  */
 399                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 400                 page = spd.pages[page_nr];
 401
 402                 if (PageReadahead(page))
 403                         page_cache_async_readahead(mapping, &in->f_ra, in,
 404                                         page, index, req_pages - page_nr);
 405
 406                 /*
 407                  * If the page isn't uptodate, we may need to start io on it
 408                  */
 409                 if (!PageUptodate(page)) {
 410                         lock_page(page);
 411
 412                         /*
 413                          * Page was truncated, or invalidated by the
 414                          * filesystem.  Redo the find/create, but this time the
 415                          * page is kept locked, so there's no chance of another
 416                          * race with truncate/invalidate.
 417                          */
 418                         if (!page->mapping) {
 419                                 unlock_page(page);
 420                                 page = find_or_create_page(mapping, index,
 421                                                 mapping_gfp_mask(mapping));
 422
 423                                 if (!page) {
 424                                         error = -ENOMEM;
 425                                         break;
 426                                 }
 427                                 page_cache_release(spd.pages[page_nr]);
 428                                 spd.pages[page_nr] = page;
 429                         }
 430                         /*
 431                          * page was already under io and is now done, great
 432                          */
 433                         if (PageUptodate(page)) {
 434                                 unlock_page(page);
 435                                 goto fill_it;
 436                         }
 437
 438                         /*
 439                          * need to read in the page
 440                          */
 441                         error = mapping->a_ops->readpage(in, page);
 442                         if (unlikely(error)) {
 443                                 /*
 444                                  * We really should re-lookup the page here,
 445                                  * but it complicates things a lot. Instead
 446                                  * lets just do what we already stored, and
 447                                  * we'll get it the next time we are called.
 448                                  */
 449                                 if (error == AOP_TRUNCATED_PAGE)
 450                                         error = 0;
 451
 452                                 break;
 453                         }
 454                 }
 455 fill_it:
 456                 /*
 457                  * i_size must be checked after PageUptodate.
 458                  */
 459                 isize = i_size_read(mapping->host);
 460                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 461                 if (unlikely(!isize || index > end_index))
 462                         break;
 463
 464                 /*
 465                  * if this is the last page, see if we need to shrink
 466                  * the length and stop
 467                  */
 468                 if (end_index == index) {
 469                         unsigned int plen;
 470
 471                         /*
 472                          * max good bytes in this page
 473                          */
 474                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 475                         if (plen <= loff)
 476                                 break;
 477
 478                         /*
 479                          * force quit after adding this page
 480                          */
 481                         this_len = min(this_len, plen - loff);
 482                         len = this_len;
 483                 }
 484
 485                 spd.partial[page_nr].offset = loff;
 486                 spd.partial[page_nr].len = this_len;
 487                 len -= this_len;
 488                 loff = 0;
 489                 spd.nr_pages++;
 490                 index++;
 491         }
 492
 493         /*
 494          * Release any pages at the end, if we quit early. 'page_nr' is how far
 495          * we got, 'nr_pages' is how many pages are in the map.
 496          */
 497         while (page_nr < nr_pages)
 498                 page_cache_release(spd.pages[page_nr++]);
 499         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 500
 501         if (spd.nr_pages)
 502                 error = splice_to_pipe(pipe, &spd);
 503
 504         splice_shrink_spd(&spd);
 505         return error;
 506 }
 507
 508 /**
 509  * generic_file_splice_read - splice data from file to a pipe
 510  * @in:         file to splice from
 511  * @ppos:       position in @in
 512  * @pipe:       pipe to splice to
 513  * @len:        number of bytes to splice
 514  * @flags:      splice modifier flags
 515  *
 516  * Description:
 517  *    Will read pages from given file and fill them into a pipe. Can be
 518  *    used as long as the address_space operations for the source implements
 519  *    a readpage() hook.
 520  *
 521  */
 522 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 523                                  struct pipe_inode_info *pipe, size_t len,
 524                                  unsigned int flags)
 525 {
 526         loff_t isize, left;
 527         int ret;
 528
 529         isize = i_size_read(in->f_mapping->host);
 530         if (unlikely(*ppos >= isize))
 531                 return 0;
 532
 533         left = isize - *ppos;
 534         if (unlikely(left < len))
 535                 len = left;
 536
 537         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 538         if (ret > 0) {
 539                 *ppos += ret;
 540                 file_accessed(in);
 541         }
 542
 543         return ret;
 544 }
 545 EXPORT_SYMBOL(generic_file_splice_read);
 546
 547 static const struct pipe_buf_operations default_pipe_buf_ops = {
 548         .can_merge = 0,
 549         .map = generic_pipe_buf_map,
 550         .unmap = generic_pipe_buf_unmap,
 551         .confirm = generic_pipe_buf_confirm,
 552         .release = generic_pipe_buf_release,
 553         .steal = generic_pipe_buf_steal,
 554         .get = generic_pipe_buf_get,
 555 };
 556
 557 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 558                                     struct pipe_buffer *buf)
 559 {
 560         return 1;
 561 }
 562
 563 /* Pipe buffer operations for a socket and similar. */
 564 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 565         .can_merge = 0,
 566         .map = generic_pipe_buf_map,
 567         .unmap = generic_pipe_buf_unmap,
 568         .confirm = generic_pipe_buf_confirm,
 569         .release = generic_pipe_buf_release,
 570         .steal = generic_pipe_buf_nosteal,
 571         .get = generic_pipe_buf_get,
 572 };
 573 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 574
 575 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 576                             unsigned long vlen, loff_t offset)
 577 {
 578         mm_segment_t old_fs;
 579         loff_t pos = offset;
 580         ssize_t res;
 581
 582         old_fs = get_fs();
 583         set_fs(get_ds());
 584         /* The cast to a user pointer is valid due to the set_fs() */
 585         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 586         set_fs(old_fs);
 587
 588         return res;
 589 }
 590
 591 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 592                             loff_t pos)
 593 {
 594         mm_segment_t old_fs;
 595         ssize_t res;
 596
 597         old_fs = get_fs();
 598         set_fs(get_ds());
 599         /* The cast to a user pointer is valid due to the set_fs() */
 600         res = vfs_write(file, (const char __user *)buf, count, &pos);
 601         set_fs(old_fs);
 602
 603         return res;
 604 }
 605
 606 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 607                                  struct pipe_inode_info *pipe, size_t len,
 608                                  unsigned int flags)
 609 {
 610         unsigned int nr_pages;
 611         unsigned int nr_freed;
 612         size_t offset;
 613         struct page *pages[PIPE_DEF_BUFFERS];
 614         struct partial_page partial[PIPE_DEF_BUFFERS];
 615         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 616         ssize_t res;
 617         size_t this_len;
 618         int error;
 619         int i;
 620         struct splice_pipe_desc spd = {
 621                 .pages = pages,
 622                 .partial = partial,
 623                 .nr_pages_max = PIPE_DEF_BUFFERS,
 624                 .flags = flags,
 625                 .ops = &default_pipe_buf_ops,
 626                 .spd_release = spd_release_page,
 627         };
 628
 629         if (splice_grow_spd(pipe, &spd))
 630                 return -ENOMEM;
 631
 632         res = -ENOMEM;
 633         vec = __vec;
 634         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 635                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 636                 if (!vec)
 637                         goto shrink_ret;
 638         }
 639
 640         offset = *ppos & ~PAGE_CACHE_MASK;
 641         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 642
 643         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 644                 struct page *page;
 645
 646                 page = alloc_page(GFP_USER);
 647                 error = -ENOMEM;
 648                 if (!page)
 649                         goto err;
 650
 651                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 652                 vec[i].iov_base = (void __user *) page_address(page);
 653                 vec[i].iov_len = this_len;
 654                 spd.pages[i] = page;
 655                 spd.nr_pages++;
 656                 len -= this_len;
 657                 offset = 0;
 658         }
 659
 660         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 661         if (res < 0) {
 662                 error = res;
 663                 goto err;
 664         }
 665
 666         error = 0;
 667         if (!res)
 668                 goto err;
 669
 670         nr_freed = 0;
 671         for (i = 0; i < spd.nr_pages; i++) {
 672                 this_len = min_t(size_t, vec[i].iov_len, res);
 673                 spd.partial[i].offset = 0;
 674                 spd.partial[i].len = this_len;
 675                 if (!this_len) {
 676                         __free_page(spd.pages[i]);
 677                         spd.pages[i] = NULL;
 678                         nr_freed++;
 679                 }
 680                 res -= this_len;
 681         }
 682         spd.nr_pages -= nr_freed;
 683
 684         res = splice_to_pipe(pipe, &spd);
 685         if (res > 0)
 686                 *ppos += res;
 687
 688 shrink_ret:
 689         if (vec != __vec)
 690                 kfree(vec);
 691         splice_shrink_spd(&spd);
 692         return res;
 693
 694 err:
 695         for (i = 0; i < spd.nr_pages; i++)
 696                 __free_page(spd.pages[i]);
 697
 698         res = error;
 699         goto shrink_ret;
 700 }
 701 EXPORT_SYMBOL(default_file_splice_read);
 702
 703 /*
 704  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 705  * using sendpage(). Return the number of bytes sent.
 706  */
 707 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 708                             struct pipe_buffer *buf, struct splice_desc *sd)
 709 {
 710         struct file *file = sd->u.file;
 711         loff_t pos = sd->pos;
 712         int more;
 713
 714         if (!likely(file->f_op && file->f_op->sendpage))
 715                 return -EINVAL;
 716
 717         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 718
 719         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 720                 more |= MSG_SENDPAGE_NOTLAST;
 721
 722         return file->f_op->sendpage(file, buf->page, buf->offset,
 723                                     sd->len, &pos, more);
 724 }
 725
 726 /*
 727  * This is a little more tricky than the file -> pipe splicing. There are
 728  * basically three cases:
 729  *
 730  *      - Destination page already exists in the address space and there
 731  *        are users of it. For that case we have no other option that
 732  *        copying the data. Tough luck.
 733  *      - Destination page already exists in the address space, but there
 734  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 735  *        through to last case.
 736  *      - Destination page does not exist, we can add the pipe page to
 737  *        the page cache and avoid the copy.
 738  *
 739  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 740  * sd->flags), we attempt to migrate pages from the pipe to the output
 741  * file address space page cache. This is possible if no one else has
 742  * the pipe page referenced outside of the pipe and page cache. If
 743  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 744  * a new page in the output file page cache and fill/dirty that.
 745  */
 746 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 747                  struct splice_desc *sd)
 748 {
 749         struct file *file = sd->u.file;
 750         struct address_space *mapping = file->f_mapping;
 751         unsigned int offset, this_len;
 752         struct page *page;
 753         void *fsdata;
 754         int ret;
 755
 756         offset = sd->pos & ~PAGE_CACHE_MASK;
 757
 758         this_len = sd->len;
 759         if (this_len + offset > PAGE_CACHE_SIZE)
 760                 this_len = PAGE_CACHE_SIZE - offset;
 761
 762         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 763                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 764         if (unlikely(ret))
 765                 goto out;
 766
 767         if (buf->page != page) {
 768                 /*
 769                  * Careful, ->map() uses KM_USER0!
 770                  */
 771                 char *src = buf->ops->map(pipe, buf, 1);
 772                 char *dst = kmap_atomic(page, KM_USER1);
 773
 774                 memcpy(dst + offset, src + buf->offset, this_len);
 775                 flush_dcache_page(page);
 776                 kunmap_atomic(dst, KM_USER1);
 777                 buf->ops->unmap(pipe, buf, src);
 778         }
 779         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 780                                 page, fsdata);
 781 out:
 782         return ret;
 783 }
 784 EXPORT_SYMBOL(pipe_to_file);
 785
 786 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 787 {
 788         smp_mb();
 789         if (waitqueue_active(&pipe->wait))
 790                 wake_up_interruptible(&pipe->wait);
 791         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 792 }
 793
 794 /**
 795  * splice_from_pipe_feed - feed available data from a pipe to a file
 796  * @pipe:       pipe to splice from
 797  * @sd:         information to @actor
 798  * @actor:      handler that splices the data
 799  *
 800  * Description:
 801  *    This function loops over the pipe and calls @actor to do the
 802  *    actual moving of a single struct pipe_buffer to the desired
 803  *    destination.  It returns when there's no more buffers left in
 804  *    the pipe or if the requested number of bytes (@sd->total_len)
 805  *    have been copied.  It returns a positive number (one) if the
 806  *    pipe needs to be filled with more data, zero if the required
 807  *    number of bytes have been copied and -errno on error.
 808  *
 809  *    This, together with splice_from_pipe_{begin,end,next}, may be
 810  *    used to implement the functionality of __splice_from_pipe() when
 811  *    locking is required around copying the pipe buffers to the
 812  *    destination.
 813  */
 814 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 815                           splice_actor *actor)
 816 {
 817         int ret;
 818
 819         while (pipe->nrbufs) {
 820                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 821                 const struct pipe_buf_operations *ops = buf->ops;
 822
 823                 sd->len = buf->len;
 824                 if (sd->len > sd->total_len)
 825                         sd->len = sd->total_len;
 826
 827                 ret = buf->ops->confirm(pipe, buf);
 828                 if (unlikely(ret)) {
 829                         if (ret == -ENODATA)
 830                                 ret = 0;
 831                         return ret;
 832                 }
 833
 834                 ret = actor(pipe, buf, sd);
 835                 if (ret <= 0)
 836                         return ret;
 837
 838                 buf->offset += ret;
 839                 buf->len -= ret;
 840
 841                 sd->num_spliced += ret;
 842                 sd->len -= ret;
 843                 sd->pos += ret;
 844                 sd->total_len -= ret;
 845
 846                 if (!buf->len) {
 847                         buf->ops = NULL;
 848                         ops->release(pipe, buf);
 849                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 850                         pipe->nrbufs--;
 851                         if (pipe->inode)
 852                                 sd->need_wakeup = true;
 853                 }
 854
 855                 if (!sd->total_len)
 856                         return 0;
 857         }
 858
 859         return 1;
 860 }
 861 EXPORT_SYMBOL(splice_from_pipe_feed);
 862
 863 /**
 864  * splice_from_pipe_next - wait for some data to splice from
 865  * @pipe:       pipe to splice from
 866  * @sd:         information about the splice operation
 867  *
 868  * Description:
 869  *    This function will wait for some data and return a positive
 870  *    value (one) if pipe buffers are available.  It will return zero
 871  *    or -errno if no more data needs to be spliced.
 872  */
 873 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 874 {
 875         while (!pipe->nrbufs) {
 876                 if (!pipe->writers)
 877                         return 0;
 878
 879                 if (!pipe->waiting_writers && sd->num_spliced)
 880                         return 0;
 881
 882                 if (sd->flags & SPLICE_F_NONBLOCK)
 883                         return -EAGAIN;
 884
 885                 if (signal_pending(current))
 886                         return -ERESTARTSYS;
 887
 888                 if (sd->need_wakeup) {
 889                         wakeup_pipe_writers(pipe);
 890                         sd->need_wakeup = false;
 891                 }
 892
 893                 pipe_wait(pipe);
 894         }
 895
 896         return 1;
 897 }
 898 EXPORT_SYMBOL(splice_from_pipe_next);
 899
 900 /**
 901  * splice_from_pipe_begin - start splicing from pipe
 902  * @sd:         information about the splice operation
 903  *
 904  * Description:
 905  *    This function should be called before a loop containing
 906  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 907  *    initialize the necessary fields of @sd.
 908  */
 909 void splice_from_pipe_begin(struct splice_desc *sd)
 910 {
 911         sd->num_spliced = 0;
 912         sd->need_wakeup = false;
 913 }
 914 EXPORT_SYMBOL(splice_from_pipe_begin);
 915
 916 /**
 917  * splice_from_pipe_end - finish splicing from pipe
 918  * @pipe:       pipe to splice from
 919  * @sd:         information about the splice operation
 920  *
 921  * Description:
 922  *    This function will wake up pipe writers if necessary.  It should
 923  *    be called after a loop containing splice_from_pipe_next() and
 924  *    splice_from_pipe_feed().
 925  */
 926 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 927 {
 928         if (sd->need_wakeup)
 929                 wakeup_pipe_writers(pipe);
 930 }
 931 EXPORT_SYMBOL(splice_from_pipe_end);
 932
 933 /**
 934  * __splice_from_pipe - splice data from a pipe to given actor
 935  * @pipe:       pipe to splice from
 936  * @sd:         information to @actor
 937  * @actor:      handler that splices the data
 938  *
 939  * Description:
 940  *    This function does little more than loop over the pipe and call
 941  *    @actor to do the actual moving of a single struct pipe_buffer to
 942  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 943  *    pipe_to_user.
 944  *
 945  */
 946 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 947                            splice_actor *actor)
 948 {
 949         int ret;
 950
 951         splice_from_pipe_begin(sd);
 952         do {
 953                 ret = splice_from_pipe_next(pipe, sd);
 954                 if (ret > 0)
 955                         ret = splice_from_pipe_feed(pipe, sd, actor);
 956         } while (ret > 0);
 957         splice_from_pipe_end(pipe, sd);
 958
 959         return sd->num_spliced ? sd->num_spliced : ret;
 960 }
 961 EXPORT_SYMBOL(__splice_from_pipe);
 962
 963 /**
 964  * splice_from_pipe - splice data from a pipe to a file
 965  * @pipe:       pipe to splice from
 966  * @out:        file to splice to
 967  * @ppos:       position in @out
 968  * @len:        how many bytes to splice
 969  * @flags:      splice modifier flags
 970  * @actor:      handler that splices the data
 971  *
 972  * Description:
 973  *    See __splice_from_pipe. This function locks the pipe inode,
 974  *    otherwise it's identical to __splice_from_pipe().
 975  *
 976  */
 977 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 978                          loff_t *ppos, size_t len, unsigned int flags,
 979                          splice_actor *actor)
 980 {
 981         ssize_t ret;
 982         struct splice_desc sd = {
 983                 .total_len = len,
 984                 .flags = flags,
 985                 .pos = *ppos,
 986                 .u.file = out,
 987         };
 988
 989         pipe_lock(pipe);
 990         ret = __splice_from_pipe(pipe, &sd, actor);
 991         pipe_unlock(pipe);
 992
 993         return ret;
 994 }
 995
 996 /**
 997  * generic_file_splice_write - splice data from a pipe to a file
 998  * @pipe:       pipe info
 999  * @out:        file to write to
1000  * @ppos:       position in @out
1001  * @len:        number of bytes to splice
1002  * @flags:      splice modifier flags
1003  *
1004  * Description:
1005  *    Will either move or copy pages (determined by @flags options) from
1006  *    the given pipe inode to the given file.
1007  *
1008  */
1009 ssize_t
1010 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1011                           loff_t *ppos, size_t len, unsigned int flags)
1012 {
1013         struct address_space *mapping = out->f_mapping;
1014         struct inode *inode = mapping->host;
1015         struct splice_desc sd = {
1016                 .flags = flags,
1017                 .u.file = out,
1018         };
1019         ssize_t ret;
1020
1021         ret = generic_write_checks(out, ppos, &len, S_ISBLK(inode->i_mode));
1022         if (ret)
1023                 return ret;
1024         sd.total_len = len;
1025         sd.pos = *ppos;
1026
1027         pipe_lock(pipe);
1028
1029         splice_from_pipe_begin(&sd);
1030         do {
1031                 ret = splice_from_pipe_next(pipe, &sd);
1032                 if (ret <= 0)
1033                         break;
1034
1035                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1036                 ret = file_remove_suid(out);
1037                 if (!ret) {
1038                         file_update_time(out);
1039                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1040                 }
1041                 mutex_unlock(&inode->i_mutex);
1042         } while (ret > 0);
1043         splice_from_pipe_end(pipe, &sd);
1044
1045         pipe_unlock(pipe);
1046
1047         if (sd.num_spliced)
1048                 ret = sd.num_spliced;
1049
1050         if (ret > 0) {
1051                 unsigned long nr_pages;
1052                 int err;
1053
1054                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1055
1056                 err = generic_write_sync(out, *ppos, ret);
1057                 if (err)
1058                         ret = err;
1059                 else
1060                         *ppos += ret;
1061                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1062         }
1063
1064         return ret;
1065 }
1066
1067 EXPORT_SYMBOL(generic_file_splice_write);
1068
1069 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1070                           struct splice_desc *sd)
1071 {
1072         int ret;
1073         void *data;
1074
1075         data = buf->ops->map(pipe, buf, 0);
1076         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1077         buf->ops->unmap(pipe, buf, data);
1078
1079         return ret;
1080 }
1081
1082 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1083                                          struct file *out, loff_t *ppos,
1084                                          size_t len, unsigned int flags)
1085 {
1086         ssize_t ret;
1087
1088         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1089         if (ret > 0)
1090                 *ppos += ret;
1091
1092         return ret;
1093 }
1094
1095 /**
1096  * generic_splice_sendpage - splice data from a pipe to a socket
1097  * @pipe:       pipe to splice from
1098  * @out:        socket to write to
1099  * @ppos:       position in @out
1100  * @len:        number of bytes to splice
1101  * @flags:      splice modifier flags
1102  *
1103  * Description:
1104  *    Will send @len bytes from the pipe to a network socket. No data copying
1105  *    is involved.
1106  *
1107  */
1108 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1109                                 loff_t *ppos, size_t len, unsigned int flags)
1110 {
1111         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1112 }
1113
1114 EXPORT_SYMBOL(generic_splice_sendpage);
1115
1116 /*
1117  * Attempt to initiate a splice from pipe to file.
1118  */
1119 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1120                            loff_t *ppos, size_t len, unsigned int flags)
1121 {
1122         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1123                                 loff_t *, size_t, unsigned int);
1124         int ret;
1125
1126         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1127                 return -EBADF;
1128
1129         if (unlikely(out->f_flags & O_APPEND))
1130                 return -EINVAL;
1131
1132         ret = rw_verify_area(WRITE, out, ppos, len);
1133         if (unlikely(ret < 0))
1134                 return ret;
1135
1136         if (out->f_op && out->f_op->splice_write)
1137                 splice_write = out->f_op->splice_write;
1138         else
1139                 splice_write = default_file_splice_write;
1140
1141         return splice_write(pipe, out, ppos, len, flags);
1142 }
1143
1144 /*
1145  * Attempt to initiate a splice from a file to a pipe.
1146  */
1147 static long do_splice_to(struct file *in, loff_t *ppos,
1148                          struct pipe_inode_info *pipe, size_t len,
1149                          unsigned int flags)
1150 {
1151         ssize_t (*splice_read)(struct file *, loff_t *,
1152                                struct pipe_inode_info *, size_t, unsigned int);
1153         int ret;
1154
1155         if (unlikely(!(in->f_mode & FMODE_READ)))
1156                 return -EBADF;
1157
1158         ret = rw_verify_area(READ, in, ppos, len);
1159         if (unlikely(ret < 0))
1160                 return ret;
1161
1162         if (in->f_op && in->f_op->splice_read)
1163                 splice_read = in->f_op->splice_read;
1164         else
1165                 splice_read = default_file_splice_read;
1166
1167         return splice_read(in, ppos, pipe, len, flags);
1168 }
1169
1170 /**
1171  * splice_direct_to_actor - splices data directly between two non-pipes
1172  * @in:         file to splice from
1173  * @sd:         actor information on where to splice to
1174  * @actor:      handles the data splicing
1175  *
1176  * Description:
1177  *    This is a special case helper to splice directly between two
1178  *    points, without requiring an explicit pipe. Internally an allocated
1179  *    pipe is cached in the process, and reused during the lifetime of
1180  *    that process.
1181  *
1182  */
1183 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1184                                splice_direct_actor *actor)
1185 {
1186         struct pipe_inode_info *pipe;
1187         long ret, bytes;
1188         umode_t i_mode;
1189         size_t len;
1190         int i, flags, more;
1191
1192         /*
1193          * We require the input being a regular file, as we don't want to
1194          * randomly drop data for eg socket -> socket splicing. Use the
1195          * piped splicing for that!
1196          */
1197         i_mode = in->f_path.dentry->d_inode->i_mode;
1198         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1199                 return -EINVAL;
1200
1201         /*
1202          * neither in nor out is a pipe, setup an internal pipe attached to
1203          * 'out' and transfer the wanted data from 'in' to 'out' through that
1204          */
1205         pipe = current->splice_pipe;
1206         if (unlikely(!pipe)) {
1207                 pipe = alloc_pipe_info(NULL);
1208                 if (!pipe)
1209                         return -ENOMEM;
1210
1211                 /*
1212                  * We don't have an immediate reader, but we'll read the stuff
1213                  * out of the pipe right after the splice_to_pipe(). So set
1214                  * PIPE_READERS appropriately.
1215                  */
1216                 pipe->readers = 1;
1217
1218                 current->splice_pipe = pipe;
1219         }
1220
1221         /*
1222          * Do the splice.
1223          */
1224         ret = 0;
1225         bytes = 0;
1226         len = sd->total_len;
1227         flags = sd->flags;
1228
1229         /*
1230          * Don't block on output, we have to drain the direct pipe.
1231          */
1232         sd->flags &= ~SPLICE_F_NONBLOCK;
1233         more = sd->flags & SPLICE_F_MORE;
1234
1235         while (len) {
1236                 size_t read_len;
1237                 loff_t pos = sd->pos, prev_pos = pos;
1238
1239                 ret = do_splice_to(in, &pos, pipe, len, flags);
1240                 if (unlikely(ret <= 0))
1241                         goto out_release;
1242
1243                 read_len = ret;
1244                 sd->total_len = read_len;
1245
1246                 /*
1247                  * If more data is pending, set SPLICE_F_MORE
1248                  * If this is the last data and SPLICE_F_MORE was not set
1249                  * initially, clears it.
1250                  */
1251                 if (read_len < len)
1252                         sd->flags |= SPLICE_F_MORE;
1253                 else if (!more)
1254                         sd->flags &= ~SPLICE_F_MORE;
1255                 /*
1256                  * NOTE: nonblocking mode only applies to the input. We
1257                  * must not do the output in nonblocking mode as then we
1258                  * could get stuck data in the internal pipe:
1259                  */
1260                 ret = actor(pipe, sd);
1261                 if (unlikely(ret <= 0)) {
1262                         sd->pos = prev_pos;
1263                         goto out_release;
1264                 }
1265
1266                 bytes += ret;
1267                 len -= ret;
1268                 sd->pos = pos;
1269
1270                 if (ret < read_len) {
1271                         sd->pos = prev_pos + ret;
1272                         goto out_release;
1273                 }
1274         }
1275
1276 done:
1277         pipe->nrbufs = pipe->curbuf = 0;
1278         file_accessed(in);
1279         return bytes;
1280
1281 out_release:
1282         /*
1283          * If we did an incomplete transfer we must release
1284          * the pipe buffers in question:
1285          */
1286         for (i = 0; i < pipe->buffers; i++) {
1287                 struct pipe_buffer *buf = pipe->bufs + i;
1288
1289                 if (buf->ops) {
1290                         buf->ops->release(pipe, buf);
1291                         buf->ops = NULL;
1292                 }
1293         }
1294
1295         if (!bytes)
1296                 bytes = ret;
1297
1298         goto done;
1299 }
1300 EXPORT_SYMBOL(splice_direct_to_actor);
1301
1302 static int direct_splice_actor(struct pipe_inode_info *pipe,
1303                                struct splice_desc *sd)
1304 {
1305         struct file *file = sd->u.file;
1306
1307         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1308                               sd->flags);
1309 }
1310
1311 /**
1312  * do_splice_direct - splices data directly between two files
1313  * @in:         file to splice from
1314  * @ppos:       input file offset
1315  * @out:        file to splice to
1316  * @len:        number of bytes to splice
1317  * @flags:      splice modifier flags
1318  *
1319  * Description:
1320  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1321  *    doing it in the application would incur an extra system call
1322  *    (splice in + splice out, as compared to just sendfile()). So this helper
1323  *    can splice directly through a process-private pipe.
1324  *
1325  */
1326 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1327                       size_t len, unsigned int flags)
1328 {
1329         struct splice_desc sd = {
1330                 .len            = len,
1331                 .total_len      = len,
1332                 .flags          = flags,
1333                 .pos            = *ppos,
1334                 .u.file         = out,
1335         };
1336         long ret;
1337
1338         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1339         if (ret > 0)
1340                 *ppos = sd.pos;
1341
1342         return ret;
1343 }
1344
1345 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1346                                struct pipe_inode_info *opipe,
1347                                size_t len, unsigned int flags);
1348
1349 /*
1350  * Determine where to splice to/from.
1351  */
1352 static long do_splice(struct file *in, loff_t __user *off_in,
1353                       struct file *out, loff_t __user *off_out,
1354                       size_t len, unsigned int flags)
1355 {
1356         struct pipe_inode_info *ipipe;
1357         struct pipe_inode_info *opipe;
1358         loff_t offset, *off;
1359         long ret;
1360
1361         ipipe = get_pipe_info(in);
1362         opipe = get_pipe_info(out);
1363
1364         if (ipipe && opipe) {
1365                 if (off_in || off_out)
1366                         return -ESPIPE;
1367
1368                 if (!(in->f_mode & FMODE_READ))
1369                         return -EBADF;
1370
1371                 if (!(out->f_mode & FMODE_WRITE))
1372                         return -EBADF;
1373
1374                 /* Splicing to self would be fun, but... */
1375                 if (ipipe == opipe)
1376                         return -EINVAL;
1377
1378                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1379         }
1380
1381         if (ipipe) {
1382                 if (off_in)
1383                         return -ESPIPE;
1384                 if (off_out) {
1385                         if (!(out->f_mode & FMODE_PWRITE))
1386                                 return -EINVAL;
1387                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1388                                 return -EFAULT;
1389                         off = &offset;
1390                 } else
1391                         off = &out->f_pos;
1392
1393                 ret = do_splice_from(ipipe, out, off, len, flags);
1394
1395                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1396                         ret = -EFAULT;
1397
1398                 return ret;
1399         }
1400
1401         if (opipe) {
1402                 if (off_out)
1403                         return -ESPIPE;
1404                 if (off_in) {
1405                         if (!(in->f_mode & FMODE_PREAD))
1406                                 return -EINVAL;
1407                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1408                                 return -EFAULT;
1409                         off = &offset;
1410                 } else
1411                         off = &in->f_pos;
1412
1413                 ret = do_splice_to(in, off, opipe, len, flags);
1414
1415                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1416                         ret = -EFAULT;
1417
1418                 return ret;
1419         }
1420
1421         return -EINVAL;
1422 }
1423
1424 /*
1425  * Map an iov into an array of pages and offset/length tupples. With the
1426  * partial_page structure, we can map several non-contiguous ranges into
1427  * our ones pages[] map instead of splitting that operation into pieces.
1428  * Could easily be exported as a generic helper for other users, in which
1429  * case one would probably want to add a 'max_nr_pages' parameter as well.
1430  */
1431 static int get_iovec_page_array(const struct iovec __user *iov,
1432                                 unsigned int nr_vecs, struct page **pages,
1433                                 struct partial_page *partial, int aligned,
1434                                 unsigned int pipe_buffers)
1435 {
1436         int buffers = 0, error = 0;
1437
1438         while (nr_vecs) {
1439                 unsigned long off, npages;
1440                 struct iovec entry;
1441                 void __user *base;
1442                 size_t len;
1443                 int i;
1444
1445                 error = -EFAULT;
1446                 if (copy_from_user(&entry, iov, sizeof(entry)))
1447                         break;
1448
1449                 base = entry.iov_base;
1450                 len = entry.iov_len;
1451
1452                 /*
1453                  * Sanity check this iovec. 0 read succeeds.
1454                  */
1455                 error = 0;
1456                 if (unlikely(!len))
1457                         break;
1458                 error = -EFAULT;
1459                 if (!access_ok(VERIFY_READ, base, len))
1460                         break;
1461
1462                 /*
1463                  * Get this base offset and number of pages, then map
1464                  * in the user pages.
1465                  */
1466                 off = (unsigned long) base & ~PAGE_MASK;
1467
1468                 /*
1469                  * If asked for alignment, the offset must be zero and the
1470                  * length a multiple of the PAGE_SIZE.
1471                  */
1472                 error = -EINVAL;
1473                 if (aligned && (off || len & ~PAGE_MASK))
1474                         break;
1475
1476                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1477                 if (npages > pipe_buffers - buffers)
1478                         npages = pipe_buffers - buffers;
1479
1480                 error = get_user_pages_fast((unsigned long)base, npages,
1481                                         0, &pages[buffers]);
1482
1483                 if (unlikely(error <= 0))
1484                         break;
1485
1486                 /*
1487                  * Fill this contiguous range into the partial page map.
1488                  */
1489                 for (i = 0; i < error; i++) {
1490                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1491
1492                         partial[buffers].offset = off;
1493                         partial[buffers].len = plen;
1494
1495                         off = 0;
1496                         len -= plen;
1497                         buffers++;
1498                 }
1499
1500                 /*
1501                  * We didn't complete this iov, stop here since it probably
1502                  * means we have to move some of this into a pipe to
1503                  * be able to continue.
1504                  */
1505                 if (len)
1506                         break;
1507
1508                 /*
1509                  * Don't continue if we mapped fewer pages than we asked for,
1510                  * or if we mapped the max number of pages that we have
1511                  * room for.
1512                  */
1513                 if (error < npages || buffers == pipe_buffers)
1514                         break;
1515
1516                 nr_vecs--;
1517                 iov++;
1518         }
1519
1520         if (buffers)
1521                 return buffers;
1522
1523         return error;
1524 }
1525
1526 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1527                         struct splice_desc *sd)
1528 {
1529         char *src;
1530         int ret;
1531
1532         /*
1533          * See if we can use the atomic maps, by prefaulting in the
1534          * pages and doing an atomic copy
1535          */
1536         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1537                 src = buf->ops->map(pipe, buf, 1);
1538                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1539                                                         sd->len);
1540                 buf->ops->unmap(pipe, buf, src);
1541                 if (!ret) {
1542                         ret = sd->len;
1543                         goto out;
1544                 }
1545         }
1546
1547         /*
1548          * No dice, use slow non-atomic map and copy
1549          */
1550         src = buf->ops->map(pipe, buf, 0);
1551
1552         ret = sd->len;
1553         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1554                 ret = -EFAULT;
1555
1556         buf->ops->unmap(pipe, buf, src);
1557 out:
1558         if (ret > 0)
1559                 sd->u.userptr += ret;
1560         return ret;
1561 }
1562
1563 /*
1564  * For lack of a better implementation, implement vmsplice() to userspace
1565  * as a simple copy of the pipes pages to the user iov.
1566  */
1567 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1568                              unsigned long nr_segs, unsigned int flags)
1569 {
1570         struct pipe_inode_info *pipe;
1571         struct splice_desc sd;
1572         ssize_t size;
1573         int error;
1574         long ret;
1575
1576         pipe = get_pipe_info(file);
1577         if (!pipe)
1578                 return -EBADF;
1579
1580         pipe_lock(pipe);
1581
1582         error = ret = 0;
1583         while (nr_segs) {
1584                 void __user *base;
1585                 size_t len;
1586
1587                 /*
1588                  * Get user address base and length for this iovec.
1589                  */
1590                 error = get_user(base, &iov->iov_base);
1591                 if (unlikely(error))
1592                         break;
1593                 error = get_user(len, &iov->iov_len);
1594                 if (unlikely(error))
1595                         break;
1596
1597                 /*
1598                  * Sanity check this iovec. 0 read succeeds.
1599                  */
1600                 if (unlikely(!len))
1601                         break;
1602                 if (unlikely(!base)) {
1603                         error = -EFAULT;
1604                         break;
1605                 }
1606
1607                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1608                         error = -EFAULT;
1609                         break;
1610                 }
1611
1612                 sd.len = 0;
1613                 sd.total_len = len;
1614                 sd.flags = flags;
1615                 sd.u.userptr = base;
1616                 sd.pos = 0;
1617
1618                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1619                 if (size < 0) {
1620                         if (!ret)
1621                                 ret = size;
1622
1623                         break;
1624                 }
1625
1626                 ret += size;
1627
1628                 if (size < len)
1629                         break;
1630
1631                 nr_segs--;
1632                 iov++;
1633         }
1634
1635         pipe_unlock(pipe);
1636
1637         if (!ret)
1638                 ret = error;
1639
1640         return ret;
1641 }
1642
1643 /*
1644  * vmsplice splices a user address range into a pipe. It can be thought of
1645  * as splice-from-memory, where the regular splice is splice-from-file (or
1646  * to file). In both cases the output is a pipe, naturally.
1647  */
1648 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1649                              unsigned long nr_segs, unsigned int flags)
1650 {
1651         struct pipe_inode_info *pipe;
1652         struct page *pages[PIPE_DEF_BUFFERS];
1653         struct partial_page partial[PIPE_DEF_BUFFERS];
1654         struct splice_pipe_desc spd = {
1655                 .pages = pages,
1656                 .partial = partial,
1657                 .nr_pages_max = PIPE_DEF_BUFFERS,
1658                 .flags = flags,
1659                 .ops = &user_page_pipe_buf_ops,
1660                 .spd_release = spd_release_page,
1661         };
1662         long ret;
1663
1664         pipe = get_pipe_info(file);
1665         if (!pipe)
1666                 return -EBADF;
1667
1668         if (splice_grow_spd(pipe, &spd))
1669                 return -ENOMEM;
1670
1671         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1672                                             spd.partial, flags & SPLICE_F_GIFT,
1673                                             spd.nr_pages_max);
1674         if (spd.nr_pages <= 0)
1675                 ret = spd.nr_pages;
1676         else
1677                 ret = splice_to_pipe(pipe, &spd);
1678
1679         splice_shrink_spd(&spd);
1680         return ret;
1681 }
1682
1683 /*
1684  * Note that vmsplice only really supports true splicing _from_ user memory
1685  * to a pipe, not the other way around. Splicing from user memory is a simple
1686  * operation that can be supported without any funky alignment restrictions
1687  * or nasty vm tricks. We simply map in the user memory and fill them into
1688  * a pipe. The reverse isn't quite as easy, though. There are two possible
1689  * solutions for that:
1690  *
1691  *      - memcpy() the data internally, at which point we might as well just
1692  *        do a regular read() on the buffer anyway.
1693  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1694  *        has restriction limitations on both ends of the pipe).
1695  *
1696  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1697  *
1698  */
1699 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1700                 unsigned long, nr_segs, unsigned int, flags)
1701 {
1702         struct file *file;
1703         long error;
1704         int fput;
1705
1706         if (unlikely(nr_segs > UIO_MAXIOV))
1707                 return -EINVAL;
1708         else if (unlikely(!nr_segs))
1709                 return 0;
1710
1711         error = -EBADF;
1712         file = fget_light(fd, &fput);
1713         if (file) {
1714                 if (file->f_mode & FMODE_WRITE)
1715                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1716                 else if (file->f_mode & FMODE_READ)
1717                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1718
1719                 fput_light(file, fput);
1720         }
1721
1722         return error;
1723 }
1724
1725 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1726                 int, fd_out, loff_t __user *, off_out,
1727                 size_t, len, unsigned int, flags)
1728 {
1729         long error;
1730         struct file *in, *out;
1731         int fput_in, fput_out;
1732
1733         if (unlikely(!len))
1734                 return 0;
1735
1736         error = -EBADF;
1737         in = fget_light(fd_in, &fput_in);
1738         if (in) {
1739                 if (in->f_mode & FMODE_READ) {
1740                         out = fget_light(fd_out, &fput_out);
1741                         if (out) {
1742                                 if (out->f_mode & FMODE_WRITE)
1743                                         error = do_splice(in, off_in,
1744                                                           out, off_out,
1745                                                           len, flags);
1746                                 fput_light(out, fput_out);
1747                         }
1748                 }
1749
1750                 fput_light(in, fput_in);
1751         }
1752
1753         return error;
1754 }
1755
1756 /*
1757  * Make sure there's data to read. Wait for input if we can, otherwise
1758  * return an appropriate error.
1759  */
1760 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1761 {
1762         int ret;
1763
1764         /*
1765          * Check ->nrbufs without the inode lock first. This function
1766          * is speculative anyways, so missing one is ok.
1767          */
1768         if (pipe->nrbufs)
1769                 return 0;
1770
1771         ret = 0;
1772         pipe_lock(pipe);
1773
1774         while (!pipe->nrbufs) {
1775                 if (signal_pending(current)) {
1776                         ret = -ERESTARTSYS;
1777                         break;
1778                 }
1779                 if (!pipe->writers)
1780                         break;
1781                 if (!pipe->waiting_writers) {
1782                         if (flags & SPLICE_F_NONBLOCK) {
1783                                 ret = -EAGAIN;
1784                                 break;
1785                         }
1786                 }
1787                 pipe_wait(pipe);
1788         }
1789
1790         pipe_unlock(pipe);
1791         return ret;
1792 }
1793
1794 /*
1795  * Make sure there's writeable room. Wait for room if we can, otherwise
1796  * return an appropriate error.
1797  */
1798 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1799 {
1800         int ret;
1801
1802         /*
1803          * Check ->nrbufs without the inode lock first. This function
1804          * is speculative anyways, so missing one is ok.
1805          */
1806         if (pipe->nrbufs < pipe->buffers)
1807                 return 0;
1808
1809         ret = 0;
1810         pipe_lock(pipe);
1811
1812         while (pipe->nrbufs >= pipe->buffers) {
1813                 if (!pipe->readers) {
1814                         send_sig(SIGPIPE, current, 0);
1815                         ret = -EPIPE;
1816                         break;
1817                 }
1818                 if (flags & SPLICE_F_NONBLOCK) {
1819                         ret = -EAGAIN;
1820                         break;
1821                 }
1822                 if (signal_pending(current)) {
1823                         ret = -ERESTARTSYS;
1824                         break;
1825                 }
1826                 pipe->waiting_writers++;
1827                 pipe_wait(pipe);
1828                 pipe->waiting_writers--;
1829         }
1830
1831         pipe_unlock(pipe);
1832         return ret;
1833 }
1834
1835 /*
1836  * Splice contents of ipipe to opipe.
1837  */
1838 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1839                                struct pipe_inode_info *opipe,
1840                                size_t len, unsigned int flags)
1841 {
1842         struct pipe_buffer *ibuf, *obuf;
1843         int ret = 0, nbuf;
1844         bool input_wakeup = false;
1845
1846
1847 retry:
1848         ret = ipipe_prep(ipipe, flags);
1849         if (ret)
1850                 return ret;
1851
1852         ret = opipe_prep(opipe, flags);
1853         if (ret)
1854                 return ret;
1855
1856         /*
1857          * Potential ABBA deadlock, work around it by ordering lock
1858          * grabbing by pipe info address. Otherwise two different processes
1859          * could deadlock (one doing tee from A -> B, the other from B -> A).
1860          */
1861         pipe_double_lock(ipipe, opipe);
1862
1863         do {
1864                 if (!opipe->readers) {
1865                         send_sig(SIGPIPE, current, 0);
1866                         if (!ret)
1867                                 ret = -EPIPE;
1868                         break;
1869                 }
1870
1871                 if (!ipipe->nrbufs && !ipipe->writers)
1872                         break;
1873
1874                 /*
1875                  * Cannot make any progress, because either the input
1876                  * pipe is empty or the output pipe is full.
1877                  */
1878                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1879                         /* Already processed some buffers, break */
1880                         if (ret)
1881                                 break;
1882
1883                         if (flags & SPLICE_F_NONBLOCK) {
1884                                 ret = -EAGAIN;
1885                                 break;
1886                         }
1887
1888                         /*
1889                          * We raced with another reader/writer and haven't
1890                          * managed to process any buffers.  A zero return
1891                          * value means EOF, so retry instead.
1892                          */
1893                         pipe_unlock(ipipe);
1894                         pipe_unlock(opipe);
1895                         goto retry;
1896                 }
1897
1898                 ibuf = ipipe->bufs + ipipe->curbuf;
1899                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1900                 obuf = opipe->bufs + nbuf;
1901
1902                 if (len >= ibuf->len) {
1903                         /*
1904                          * Simply move the whole buffer from ipipe to opipe
1905                          */
1906                         *obuf = *ibuf;
1907                         ibuf->ops = NULL;
1908                         opipe->nrbufs++;
1909                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1910                         ipipe->nrbufs--;
1911                         input_wakeup = true;
1912                 } else {
1913                         /*
1914                          * Get a reference to this pipe buffer,
1915                          * so we can copy the contents over.
1916                          */
1917                         ibuf->ops->get(ipipe, ibuf);
1918                         *obuf = *ibuf;
1919
1920                         /*
1921                          * Don't inherit the gift flag, we need to
1922                          * prevent multiple steals of this page.
1923                          */
1924                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1925
1926                         obuf->len = len;
1927                         opipe->nrbufs++;
1928                         ibuf->offset += obuf->len;
1929                         ibuf->len -= obuf->len;
1930                 }
1931                 ret += obuf->len;
1932                 len -= obuf->len;
1933         } while (len);
1934
1935         pipe_unlock(ipipe);
1936         pipe_unlock(opipe);
1937
1938         /*
1939          * If we put data in the output pipe, wakeup any potential readers.
1940          */
1941         if (ret > 0)
1942                 wakeup_pipe_readers(opipe);
1943
1944         if (input_wakeup)
1945                 wakeup_pipe_writers(ipipe);
1946
1947         return ret;
1948 }
1949
1950 /*
1951  * Link contents of ipipe to opipe.
1952  */
1953 static int link_pipe(struct pipe_inode_info *ipipe,
1954                      struct pipe_inode_info *opipe,
1955                      size_t len, unsigned int flags)
1956 {
1957         struct pipe_buffer *ibuf, *obuf;
1958         int ret = 0, i = 0, nbuf;
1959
1960         /*
1961          * Potential ABBA deadlock, work around it by ordering lock
1962          * grabbing by pipe info address. Otherwise two different processes
1963          * could deadlock (one doing tee from A -> B, the other from B -> A).
1964          */
1965         pipe_double_lock(ipipe, opipe);
1966
1967         do {
1968                 if (!opipe->readers) {
1969                         send_sig(SIGPIPE, current, 0);
1970                         if (!ret)
1971                                 ret = -EPIPE;
1972                         break;
1973                 }
1974
1975                 /*
1976                  * If we have iterated all input buffers or ran out of
1977                  * output room, break.
1978                  */
1979                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1980                         break;
1981
1982                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1983                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1984
1985                 /*
1986                  * Get a reference to this pipe buffer,
1987                  * so we can copy the contents over.
1988                  */
1989                 ibuf->ops->get(ipipe, ibuf);
1990
1991                 obuf = opipe->bufs + nbuf;
1992                 *obuf = *ibuf;
1993
1994                 /*
1995                  * Don't inherit the gift flag, we need to
1996                  * prevent multiple steals of this page.
1997                  */
1998                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1999
2000                 if (obuf->len > len)
2001                         obuf->len = len;
2002
2003                 opipe->nrbufs++;
2004                 ret += obuf->len;
2005                 len -= obuf->len;
2006                 i++;
2007         } while (len);
2008
2009         /*
2010          * return EAGAIN if we have the potential of some data in the
2011          * future, otherwise just return 0
2012          */
2013         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2014                 ret = -EAGAIN;
2015
2016         pipe_unlock(ipipe);
2017         pipe_unlock(opipe);
2018
2019         /*
2020          * If we put data in the output pipe, wakeup any potential readers.
2021          */
2022         if (ret > 0)
2023                 wakeup_pipe_readers(opipe);
2024
2025         return ret;
2026 }
2027
2028 /*
2029  * This is a tee(1) implementation that works on pipes. It doesn't copy
2030  * any data, it simply references the 'in' pages on the 'out' pipe.
2031  * The 'flags' used are the SPLICE_F_* variants, currently the only
2032  * applicable one is SPLICE_F_NONBLOCK.
2033  */
2034 static long do_tee(struct file *in, struct file *out, size_t len,
2035                    unsigned int flags)
2036 {
2037         struct pipe_inode_info *ipipe = get_pipe_info(in);
2038         struct pipe_inode_info *opipe = get_pipe_info(out);
2039         int ret = -EINVAL;
2040
2041         /*
2042          * Duplicate the contents of ipipe to opipe without actually
2043          * copying the data.
2044          */
2045         if (ipipe && opipe && ipipe != opipe) {
2046                 /*
2047                  * Keep going, unless we encounter an error. The ipipe/opipe
2048                  * ordering doesn't really matter.
2049                  */
2050                 ret = ipipe_prep(ipipe, flags);
2051                 if (!ret) {
2052                         ret = opipe_prep(opipe, flags);
2053                         if (!ret)
2054                                 ret = link_pipe(ipipe, opipe, len, flags);
2055                 }
2056         }
2057
2058         return ret;
2059 }
2060
2061 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2062 {
2063         struct file *in;
2064         int error, fput_in;
2065
2066         if (unlikely(!len))
2067                 return 0;
2068
2069         error = -EBADF;
2070         in = fget_light(fdin, &fput_in);
2071         if (in) {
2072                 if (in->f_mode & FMODE_READ) {
2073                         int fput_out;
2074                         struct file *out = fget_light(fdout, &fput_out);
2075
2076                         if (out) {
2077                                 if (out->f_mode & FMODE_WRITE)
2078                                         error = do_tee(in, out, len, flags);
2079                                 fput_light(out, fput_out);
2080                         }
2081                 }
2082                 fput_light(in, fput_in);
2083         }
2084
2085         return error;
2086 }