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