fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  17  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/fsnotify.h>
  34 #include <linux/security.h>
  35 #include <linux/gfp.h>
  36 #include <linux/socket.h>
  37 #include <linux/sched/signal.h>
  38
  39 #include "internal.h"
  40
  41 /*
  42  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  43  * indicate they support non-blocking reads or writes, we must clear it
  44  * here if set to avoid blocking other users of this pipe if splice is
  45  * being done on it.
  46  */
  47 static noinline void noinline pipe_clear_nowait(struct file *file)
  48 {
  49         fmode_t fmode = READ_ONCE(file->f_mode);
  50
  51         do {
  52                 if (!(fmode & FMODE_NOWAIT))
  53                         break;
  54         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  55 }
  56
  57 /*
  58  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  59  * a vm helper function, it's already simplified quite a bit by the
  60  * addition of remove_mapping(). If success is returned, the caller may
  61  * attempt to reuse this page for another destination.
  62  */
  63 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  64                 struct pipe_buffer *buf)
  65 {
  66         struct folio *folio = page_folio(buf->page);
  67         struct address_space *mapping;
  68
  69         folio_lock(folio);
  70
  71         mapping = folio_mapping(folio);
  72         if (mapping) {
  73                 WARN_ON(!folio_test_uptodate(folio));
  74
  75                 /*
  76                  * At least for ext2 with nobh option, we need to wait on
  77                  * writeback completing on this folio, since we'll remove it
  78                  * from the pagecache.  Otherwise truncate wont wait on the
  79                  * folio, allowing the disk blocks to be reused by someone else
  80                  * before we actually wrote our data to them. fs corruption
  81                  * ensues.
  82                  */
  83                 folio_wait_writeback(folio);
  84
  85                 if (folio_has_private(folio) &&
  86                     !filemap_release_folio(folio, GFP_KERNEL))
  87                         goto out_unlock;
  88
  89                 /*
  90                  * If we succeeded in removing the mapping, set LRU flag
  91                  * and return good.
  92                  */
  93                 if (remove_mapping(mapping, folio)) {
  94                         buf->flags |= PIPE_BUF_FLAG_LRU;
  95                         return true;
  96                 }
  97         }
  98
  99         /*
 100          * Raced with truncate or failed to remove folio from current
 101          * address space, unlock and return failure.
 102          */
 103 out_unlock:
 104         folio_unlock(folio);
 105         return false;
 106 }
 107
 108 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 109                                         struct pipe_buffer *buf)
 110 {
 111         put_page(buf->page);
 112         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 113 }
 114
 115 /*
 116  * Check whether the contents of buf is OK to access. Since the content
 117  * is a page cache page, IO may be in flight.
 118  */
 119 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 120                                        struct pipe_buffer *buf)
 121 {
 122         struct page *page = buf->page;
 123         int err;
 124
 125         if (!PageUptodate(page)) {
 126                 lock_page(page);
 127
 128                 /*
 129                  * Page got truncated/unhashed. This will cause a 0-byte
 130                  * splice, if this is the first page.
 131                  */
 132                 if (!page->mapping) {
 133                         err = -ENODATA;
 134                         goto error;
 135                 }
 136
 137                 /*
 138                  * Uh oh, read-error from disk.
 139                  */
 140                 if (!PageUptodate(page)) {
 141                         err = -EIO;
 142                         goto error;
 143                 }
 144
 145                 /*
 146                  * Page is ok afterall, we are done.
 147                  */
 148                 unlock_page(page);
 149         }
 150
 151         return 0;
 152 error:
 153         unlock_page(page);
 154         return err;
 155 }
 156
 157 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 158         .confirm        = page_cache_pipe_buf_confirm,
 159         .release        = page_cache_pipe_buf_release,
 160         .try_steal      = page_cache_pipe_buf_try_steal,
 161         .get            = generic_pipe_buf_get,
 162 };
 163
 164 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 165                 struct pipe_buffer *buf)
 166 {
 167         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 168                 return false;
 169
 170         buf->flags |= PIPE_BUF_FLAG_LRU;
 171         return generic_pipe_buf_try_steal(pipe, buf);
 172 }
 173
 174 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 175         .release        = page_cache_pipe_buf_release,
 176         .try_steal      = user_page_pipe_buf_try_steal,
 177         .get            = generic_pipe_buf_get,
 178 };
 179
 180 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 181 {
 182         smp_mb();
 183         if (waitqueue_active(&pipe->rd_wait))
 184                 wake_up_interruptible(&pipe->rd_wait);
 185         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 186 }
 187
 188 /**
 189  * splice_to_pipe - fill passed data into a pipe
 190  * @pipe:       pipe to fill
 191  * @spd:        data to fill
 192  *
 193  * Description:
 194  *    @spd contains a map of pages and len/offset tuples, along with
 195  *    the struct pipe_buf_operations associated with these pages. This
 196  *    function will link that data to the pipe.
 197  *
 198  */
 199 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 200                        struct splice_pipe_desc *spd)
 201 {
 202         unsigned int spd_pages = spd->nr_pages;
 203         unsigned int tail = pipe->tail;
 204         unsigned int head = pipe->head;
 205         unsigned int mask = pipe->ring_size - 1;
 206         int ret = 0, page_nr = 0;
 207
 208         if (!spd_pages)
 209                 return 0;
 210
 211         if (unlikely(!pipe->readers)) {
 212                 send_sig(SIGPIPE, current, 0);
 213                 ret = -EPIPE;
 214                 goto out;
 215         }
 216
 217         while (!pipe_full(head, tail, pipe->max_usage)) {
 218                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 219
 220                 buf->page = spd->pages[page_nr];
 221                 buf->offset = spd->partial[page_nr].offset;
 222                 buf->len = spd->partial[page_nr].len;
 223                 buf->private = spd->partial[page_nr].private;
 224                 buf->ops = spd->ops;
 225                 buf->flags = 0;
 226
 227                 head++;
 228                 pipe->head = head;
 229                 page_nr++;
 230                 ret += buf->len;
 231
 232                 if (!--spd->nr_pages)
 233                         break;
 234         }
 235
 236         if (!ret)
 237                 ret = -EAGAIN;
 238
 239 out:
 240         while (page_nr < spd_pages)
 241                 spd->spd_release(spd, page_nr++);
 242
 243         return ret;
 244 }
 245 EXPORT_SYMBOL_GPL(splice_to_pipe);
 246
 247 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 248 {
 249         unsigned int head = pipe->head;
 250         unsigned int tail = pipe->tail;
 251         unsigned int mask = pipe->ring_size - 1;
 252         int ret;
 253
 254         if (unlikely(!pipe->readers)) {
 255                 send_sig(SIGPIPE, current, 0);
 256                 ret = -EPIPE;
 257         } else if (pipe_full(head, tail, pipe->max_usage)) {
 258                 ret = -EAGAIN;
 259         } else {
 260                 pipe->bufs[head & mask] = *buf;
 261                 pipe->head = head + 1;
 262                 return buf->len;
 263         }
 264         pipe_buf_release(pipe, buf);
 265         return ret;
 266 }
 267 EXPORT_SYMBOL(add_to_pipe);
 268
 269 /*
 270  * Check if we need to grow the arrays holding pages and partial page
 271  * descriptions.
 272  */
 273 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 274 {
 275         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 276
 277         spd->nr_pages_max = max_usage;
 278         if (max_usage <= PIPE_DEF_BUFFERS)
 279                 return 0;
 280
 281         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 282         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 283                                      GFP_KERNEL);
 284
 285         if (spd->pages && spd->partial)
 286                 return 0;
 287
 288         kfree(spd->pages);
 289         kfree(spd->partial);
 290         return -ENOMEM;
 291 }
 292
 293 void splice_shrink_spd(struct splice_pipe_desc *spd)
 294 {
 295         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 296                 return;
 297
 298         kfree(spd->pages);
 299         kfree(spd->partial);
 300 }
 301
 302 /**
 303  * copy_splice_read -  Copy data from a file and splice the copy into a pipe
 304  * @in: The file to read from
 305  * @ppos: Pointer to the file position to read from
 306  * @pipe: The pipe to splice into
 307  * @len: The amount to splice
 308  * @flags: The SPLICE_F_* flags
 309  *
 310  * This function allocates a bunch of pages sufficient to hold the requested
 311  * amount of data (but limited by the remaining pipe capacity), passes it to
 312  * the file's ->read_iter() to read into and then splices the used pages into
 313  * the pipe.
 314  *
 315  * Return: On success, the number of bytes read will be returned and *@ppos
 316  * will be updated if appropriate; 0 will be returned if there is no more data
 317  * to be read; -EAGAIN will be returned if the pipe had no space, and some
 318  * other negative error code will be returned on error.  A short read may occur
 319  * if the pipe has insufficient space, we reach the end of the data or we hit a
 320  * hole.
 321  */
 322 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 323                          struct pipe_inode_info *pipe,
 324                          size_t len, unsigned int flags)
 325 {
 326         struct iov_iter to;
 327         struct bio_vec *bv;
 328         struct kiocb kiocb;
 329         struct page **pages;
 330         ssize_t ret;
 331         size_t used, npages, chunk, remain, keep = 0;
 332         int i;
 333
 334         /* Work out how much data we can actually add into the pipe */
 335         used = pipe_occupancy(pipe->head, pipe->tail);
 336         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 337         len = min_t(size_t, len, npages * PAGE_SIZE);
 338         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 339
 340         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 341                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 342         if (!bv)
 343                 return -ENOMEM;
 344
 345         pages = (struct page **)(bv + npages);
 346         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 347         if (!npages) {
 348                 kfree(bv);
 349                 return -ENOMEM;
 350         }
 351
 352         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 353
 354         for (i = 0; i < npages; i++) {
 355                 chunk = min_t(size_t, PAGE_SIZE, remain);
 356                 bv[i].bv_page = pages[i];
 357                 bv[i].bv_offset = 0;
 358                 bv[i].bv_len = chunk;
 359                 remain -= chunk;
 360         }
 361
 362         /* Do the I/O */
 363         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 364         init_sync_kiocb(&kiocb, in);
 365         kiocb.ki_pos = *ppos;
 366         ret = call_read_iter(in, &kiocb, &to);
 367
 368         if (ret > 0) {
 369                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 370                 *ppos = kiocb.ki_pos;
 371                 file_accessed(in);
 372         } else if (ret < 0) {
 373                 /*
 374                  * callers of ->splice_read() expect -EAGAIN on
 375                  * "can't put anything in there", rather than -EFAULT.
 376                  */
 377                 if (ret == -EFAULT)
 378                         ret = -EAGAIN;
 379         }
 380
 381         /* Free any pages that didn't get touched at all. */
 382         if (keep < npages)
 383                 release_pages(pages + keep, npages - keep);
 384
 385         /* Push the remaining pages into the pipe. */
 386         remain = ret;
 387         for (i = 0; i < keep; i++) {
 388                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 389
 390                 chunk = min_t(size_t, remain, PAGE_SIZE);
 391                 *buf = (struct pipe_buffer) {
 392                         .ops    = &default_pipe_buf_ops,
 393                         .page   = bv[i].bv_page,
 394                         .offset = 0,
 395                         .len    = chunk,
 396                 };
 397                 pipe->head++;
 398                 remain -= chunk;
 399         }
 400
 401         kfree(bv);
 402         return ret;
 403 }
 404 EXPORT_SYMBOL(copy_splice_read);
 405
 406 const struct pipe_buf_operations default_pipe_buf_ops = {
 407         .release        = generic_pipe_buf_release,
 408         .try_steal      = generic_pipe_buf_try_steal,
 409         .get            = generic_pipe_buf_get,
 410 };
 411
 412 /* Pipe buffer operations for a socket and similar. */
 413 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 414         .release        = generic_pipe_buf_release,
 415         .get            = generic_pipe_buf_get,
 416 };
 417 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 418
 419 /*
 420  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 421  * using sendpage(). Return the number of bytes sent.
 422  */
 423 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 424                             struct pipe_buffer *buf, struct splice_desc *sd)
 425 {
 426         struct file *file = sd->u.file;
 427         loff_t pos = sd->pos;
 428         int more;
 429
 430         if (!likely(file->f_op->sendpage))
 431                 return -EINVAL;
 432
 433         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 434
 435         if (sd->len < sd->total_len &&
 436             pipe_occupancy(pipe->head, pipe->tail) > 1)
 437                 more |= MSG_SENDPAGE_NOTLAST;
 438
 439         return file->f_op->sendpage(file, buf->page, buf->offset,
 440                                     sd->len, &pos, more);
 441 }
 442
 443 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 444 {
 445         smp_mb();
 446         if (waitqueue_active(&pipe->wr_wait))
 447                 wake_up_interruptible(&pipe->wr_wait);
 448         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 449 }
 450
 451 /**
 452  * splice_from_pipe_feed - feed available data from a pipe to a file
 453  * @pipe:       pipe to splice from
 454  * @sd:         information to @actor
 455  * @actor:      handler that splices the data
 456  *
 457  * Description:
 458  *    This function loops over the pipe and calls @actor to do the
 459  *    actual moving of a single struct pipe_buffer to the desired
 460  *    destination.  It returns when there's no more buffers left in
 461  *    the pipe or if the requested number of bytes (@sd->total_len)
 462  *    have been copied.  It returns a positive number (one) if the
 463  *    pipe needs to be filled with more data, zero if the required
 464  *    number of bytes have been copied and -errno on error.
 465  *
 466  *    This, together with splice_from_pipe_{begin,end,next}, may be
 467  *    used to implement the functionality of __splice_from_pipe() when
 468  *    locking is required around copying the pipe buffers to the
 469  *    destination.
 470  */
 471 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 472                           splice_actor *actor)
 473 {
 474         unsigned int head = pipe->head;
 475         unsigned int tail = pipe->tail;
 476         unsigned int mask = pipe->ring_size - 1;
 477         int ret;
 478
 479         while (!pipe_empty(head, tail)) {
 480                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 481
 482                 sd->len = buf->len;
 483                 if (sd->len > sd->total_len)
 484                         sd->len = sd->total_len;
 485
 486                 ret = pipe_buf_confirm(pipe, buf);
 487                 if (unlikely(ret)) {
 488                         if (ret == -ENODATA)
 489                                 ret = 0;
 490                         return ret;
 491                 }
 492
 493                 ret = actor(pipe, buf, sd);
 494                 if (ret <= 0)
 495                         return ret;
 496
 497                 buf->offset += ret;
 498                 buf->len -= ret;
 499
 500                 sd->num_spliced += ret;
 501                 sd->len -= ret;
 502                 sd->pos += ret;
 503                 sd->total_len -= ret;
 504
 505                 if (!buf->len) {
 506                         pipe_buf_release(pipe, buf);
 507                         tail++;
 508                         pipe->tail = tail;
 509                         if (pipe->files)
 510                                 sd->need_wakeup = true;
 511                 }
 512
 513                 if (!sd->total_len)
 514                         return 0;
 515         }
 516
 517         return 1;
 518 }
 519
 520 /* We know we have a pipe buffer, but maybe it's empty? */
 521 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 522 {
 523         unsigned int tail = pipe->tail;
 524         unsigned int mask = pipe->ring_size - 1;
 525         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 526
 527         if (unlikely(!buf->len)) {
 528                 pipe_buf_release(pipe, buf);
 529                 pipe->tail = tail+1;
 530                 return true;
 531         }
 532
 533         return false;
 534 }
 535
 536 /**
 537  * splice_from_pipe_next - wait for some data to splice from
 538  * @pipe:       pipe to splice from
 539  * @sd:         information about the splice operation
 540  *
 541  * Description:
 542  *    This function will wait for some data and return a positive
 543  *    value (one) if pipe buffers are available.  It will return zero
 544  *    or -errno if no more data needs to be spliced.
 545  */
 546 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 547 {
 548         /*
 549          * Check for signal early to make process killable when there are
 550          * always buffers available
 551          */
 552         if (signal_pending(current))
 553                 return -ERESTARTSYS;
 554
 555 repeat:
 556         while (pipe_empty(pipe->head, pipe->tail)) {
 557                 if (!pipe->writers)
 558                         return 0;
 559
 560                 if (sd->num_spliced)
 561                         return 0;
 562
 563                 if (sd->flags & SPLICE_F_NONBLOCK)
 564                         return -EAGAIN;
 565
 566                 if (signal_pending(current))
 567                         return -ERESTARTSYS;
 568
 569                 if (sd->need_wakeup) {
 570                         wakeup_pipe_writers(pipe);
 571                         sd->need_wakeup = false;
 572                 }
 573
 574                 pipe_wait_readable(pipe);
 575         }
 576
 577         if (eat_empty_buffer(pipe))
 578                 goto repeat;
 579
 580         return 1;
 581 }
 582
 583 /**
 584  * splice_from_pipe_begin - start splicing from pipe
 585  * @sd:         information about the splice operation
 586  *
 587  * Description:
 588  *    This function should be called before a loop containing
 589  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 590  *    initialize the necessary fields of @sd.
 591  */
 592 static void splice_from_pipe_begin(struct splice_desc *sd)
 593 {
 594         sd->num_spliced = 0;
 595         sd->need_wakeup = false;
 596 }
 597
 598 /**
 599  * splice_from_pipe_end - finish splicing from pipe
 600  * @pipe:       pipe to splice from
 601  * @sd:         information about the splice operation
 602  *
 603  * Description:
 604  *    This function will wake up pipe writers if necessary.  It should
 605  *    be called after a loop containing splice_from_pipe_next() and
 606  *    splice_from_pipe_feed().
 607  */
 608 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 609 {
 610         if (sd->need_wakeup)
 611                 wakeup_pipe_writers(pipe);
 612 }
 613
 614 /**
 615  * __splice_from_pipe - splice data from a pipe to given actor
 616  * @pipe:       pipe to splice from
 617  * @sd:         information to @actor
 618  * @actor:      handler that splices the data
 619  *
 620  * Description:
 621  *    This function does little more than loop over the pipe and call
 622  *    @actor to do the actual moving of a single struct pipe_buffer to
 623  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 624  *    pipe_to_user.
 625  *
 626  */
 627 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 628                            splice_actor *actor)
 629 {
 630         int ret;
 631
 632         splice_from_pipe_begin(sd);
 633         do {
 634                 cond_resched();
 635                 ret = splice_from_pipe_next(pipe, sd);
 636                 if (ret > 0)
 637                         ret = splice_from_pipe_feed(pipe, sd, actor);
 638         } while (ret > 0);
 639         splice_from_pipe_end(pipe, sd);
 640
 641         return sd->num_spliced ? sd->num_spliced : ret;
 642 }
 643 EXPORT_SYMBOL(__splice_from_pipe);
 644
 645 /**
 646  * splice_from_pipe - splice data from a pipe to a file
 647  * @pipe:       pipe to splice from
 648  * @out:        file to splice to
 649  * @ppos:       position in @out
 650  * @len:        how many bytes to splice
 651  * @flags:      splice modifier flags
 652  * @actor:      handler that splices the data
 653  *
 654  * Description:
 655  *    See __splice_from_pipe. This function locks the pipe inode,
 656  *    otherwise it's identical to __splice_from_pipe().
 657  *
 658  */
 659 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 660                          loff_t *ppos, size_t len, unsigned int flags,
 661                          splice_actor *actor)
 662 {
 663         ssize_t ret;
 664         struct splice_desc sd = {
 665                 .total_len = len,
 666                 .flags = flags,
 667                 .pos = *ppos,
 668                 .u.file = out,
 669         };
 670
 671         pipe_lock(pipe);
 672         ret = __splice_from_pipe(pipe, &sd, actor);
 673         pipe_unlock(pipe);
 674
 675         return ret;
 676 }
 677
 678 /**
 679  * iter_file_splice_write - splice data from a pipe to a file
 680  * @pipe:       pipe info
 681  * @out:        file to write to
 682  * @ppos:       position in @out
 683  * @len:        number of bytes to splice
 684  * @flags:      splice modifier flags
 685  *
 686  * Description:
 687  *    Will either move or copy pages (determined by @flags options) from
 688  *    the given pipe inode to the given file.
 689  *    This one is ->write_iter-based.
 690  *
 691  */
 692 ssize_t
 693 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 694                           loff_t *ppos, size_t len, unsigned int flags)
 695 {
 696         struct splice_desc sd = {
 697                 .total_len = len,
 698                 .flags = flags,
 699                 .pos = *ppos,
 700                 .u.file = out,
 701         };
 702         int nbufs = pipe->max_usage;
 703         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 704                                         GFP_KERNEL);
 705         ssize_t ret;
 706
 707         if (unlikely(!array))
 708                 return -ENOMEM;
 709
 710         pipe_lock(pipe);
 711
 712         splice_from_pipe_begin(&sd);
 713         while (sd.total_len) {
 714                 struct iov_iter from;
 715                 unsigned int head, tail, mask;
 716                 size_t left;
 717                 int n;
 718
 719                 ret = splice_from_pipe_next(pipe, &sd);
 720                 if (ret <= 0)
 721                         break;
 722
 723                 if (unlikely(nbufs < pipe->max_usage)) {
 724                         kfree(array);
 725                         nbufs = pipe->max_usage;
 726                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 727                                         GFP_KERNEL);
 728                         if (!array) {
 729                                 ret = -ENOMEM;
 730                                 break;
 731                         }
 732                 }
 733
 734                 head = pipe->head;
 735                 tail = pipe->tail;
 736                 mask = pipe->ring_size - 1;
 737
 738                 /* build the vector */
 739                 left = sd.total_len;
 740                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 741                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 742                         size_t this_len = buf->len;
 743
 744                         /* zero-length bvecs are not supported, skip them */
 745                         if (!this_len)
 746                                 continue;
 747                         this_len = min(this_len, left);
 748
 749                         ret = pipe_buf_confirm(pipe, buf);
 750                         if (unlikely(ret)) {
 751                                 if (ret == -ENODATA)
 752                                         ret = 0;
 753                                 goto done;
 754                         }
 755
 756                         bvec_set_page(&array[n], buf->page, this_len,
 757                                       buf->offset);
 758                         left -= this_len;
 759                         n++;
 760                 }
 761
 762                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 763                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 764                 if (ret <= 0)
 765                         break;
 766
 767                 sd.num_spliced += ret;
 768                 sd.total_len -= ret;
 769                 *ppos = sd.pos;
 770
 771                 /* dismiss the fully eaten buffers, adjust the partial one */
 772                 tail = pipe->tail;
 773                 while (ret) {
 774                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 775                         if (ret >= buf->len) {
 776                                 ret -= buf->len;
 777                                 buf->len = 0;
 778                                 pipe_buf_release(pipe, buf);
 779                                 tail++;
 780                                 pipe->tail = tail;
 781                                 if (pipe->files)
 782                                         sd.need_wakeup = true;
 783                         } else {
 784                                 buf->offset += ret;
 785                                 buf->len -= ret;
 786                                 ret = 0;
 787                         }
 788                 }
 789         }
 790 done:
 791         kfree(array);
 792         splice_from_pipe_end(pipe, &sd);
 793
 794         pipe_unlock(pipe);
 795
 796         if (sd.num_spliced)
 797                 ret = sd.num_spliced;
 798
 799         return ret;
 800 }
 801
 802 EXPORT_SYMBOL(iter_file_splice_write);
 803
 804 /**
 805  * generic_splice_sendpage - splice data from a pipe to a socket
 806  * @pipe:       pipe to splice from
 807  * @out:        socket to write to
 808  * @ppos:       position in @out
 809  * @len:        number of bytes to splice
 810  * @flags:      splice modifier flags
 811  *
 812  * Description:
 813  *    Will send @len bytes from the pipe to a network socket. No data copying
 814  *    is involved.
 815  *
 816  */
 817 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 818                                 loff_t *ppos, size_t len, unsigned int flags)
 819 {
 820         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 821 }
 822
 823 EXPORT_SYMBOL(generic_splice_sendpage);
 824
 825 static int warn_unsupported(struct file *file, const char *op)
 826 {
 827         pr_debug_ratelimited(
 828                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 829                 op, file, current->pid, current->comm);
 830         return -EINVAL;
 831 }
 832
 833 /*
 834  * Attempt to initiate a splice from pipe to file.
 835  */
 836 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 837                            loff_t *ppos, size_t len, unsigned int flags)
 838 {
 839         if (unlikely(!out->f_op->splice_write))
 840                 return warn_unsupported(out, "write");
 841         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 842 }
 843
 844 /**
 845  * vfs_splice_read - Read data from a file and splice it into a pipe
 846  * @in:         File to splice from
 847  * @ppos:       Input file offset
 848  * @pipe:       Pipe to splice to
 849  * @len:        Number of bytes to splice
 850  * @flags:      Splice modifier flags (SPLICE_F_*)
 851  *
 852  * Splice the requested amount of data from the input file to the pipe.  This
 853  * is synchronous as the caller must hold the pipe lock across the entire
 854  * operation.
 855  *
 856  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 857  * a hole and a negative error code otherwise.
 858  */
 859 long vfs_splice_read(struct file *in, loff_t *ppos,
 860                      struct pipe_inode_info *pipe, size_t len,
 861                      unsigned int flags)
 862 {
 863         unsigned int p_space;
 864         int ret;
 865
 866         if (unlikely(!(in->f_mode & FMODE_READ)))
 867                 return -EBADF;
 868         if (!len)
 869                 return 0;
 870
 871         /* Don't try to read more the pipe has space for. */
 872         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 873         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 874
 875         ret = rw_verify_area(READ, in, ppos, len);
 876         if (unlikely(ret < 0))
 877                 return ret;
 878
 879         if (unlikely(len > MAX_RW_COUNT))
 880                 len = MAX_RW_COUNT;
 881
 882         if (unlikely(!in->f_op->splice_read))
 883                 return warn_unsupported(in, "read");
 884         /*
 885          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 886          * buffer, copy into it and splice that into the pipe.
 887          */
 888         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 889                 return copy_splice_read(in, ppos, pipe, len, flags);
 890         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 891 }
 892 EXPORT_SYMBOL_GPL(vfs_splice_read);
 893
 894 /**
 895  * splice_direct_to_actor - splices data directly between two non-pipes
 896  * @in:         file to splice from
 897  * @sd:         actor information on where to splice to
 898  * @actor:      handles the data splicing
 899  *
 900  * Description:
 901  *    This is a special case helper to splice directly between two
 902  *    points, without requiring an explicit pipe. Internally an allocated
 903  *    pipe is cached in the process, and reused during the lifetime of
 904  *    that process.
 905  *
 906  */
 907 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 908                                splice_direct_actor *actor)
 909 {
 910         struct pipe_inode_info *pipe;
 911         long ret, bytes;
 912         size_t len;
 913         int i, flags, more;
 914
 915         /*
 916          * We require the input to be seekable, as we don't want to randomly
 917          * drop data for eg socket -> socket splicing. Use the piped splicing
 918          * for that!
 919          */
 920         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
 921                 return -EINVAL;
 922
 923         /*
 924          * neither in nor out is a pipe, setup an internal pipe attached to
 925          * 'out' and transfer the wanted data from 'in' to 'out' through that
 926          */
 927         pipe = current->splice_pipe;
 928         if (unlikely(!pipe)) {
 929                 pipe = alloc_pipe_info();
 930                 if (!pipe)
 931                         return -ENOMEM;
 932
 933                 /*
 934                  * We don't have an immediate reader, but we'll read the stuff
 935                  * out of the pipe right after the splice_to_pipe(). So set
 936                  * PIPE_READERS appropriately.
 937                  */
 938                 pipe->readers = 1;
 939
 940                 current->splice_pipe = pipe;
 941         }
 942
 943         /*
 944          * Do the splice.
 945          */
 946         bytes = 0;
 947         len = sd->total_len;
 948         flags = sd->flags;
 949
 950         /*
 951          * Don't block on output, we have to drain the direct pipe.
 952          */
 953         sd->flags &= ~SPLICE_F_NONBLOCK;
 954         more = sd->flags & SPLICE_F_MORE;
 955
 956         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 957
 958         while (len) {
 959                 size_t read_len;
 960                 loff_t pos = sd->pos, prev_pos = pos;
 961
 962                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
 963                 if (unlikely(ret <= 0))
 964                         goto out_release;
 965
 966                 read_len = ret;
 967                 sd->total_len = read_len;
 968
 969                 /*
 970                  * If more data is pending, set SPLICE_F_MORE
 971                  * If this is the last data and SPLICE_F_MORE was not set
 972                  * initially, clears it.
 973                  */
 974                 if (read_len < len)
 975                         sd->flags |= SPLICE_F_MORE;
 976                 else if (!more)
 977                         sd->flags &= ~SPLICE_F_MORE;
 978                 /*
 979                  * NOTE: nonblocking mode only applies to the input. We
 980                  * must not do the output in nonblocking mode as then we
 981                  * could get stuck data in the internal pipe:
 982                  */
 983                 ret = actor(pipe, sd);
 984                 if (unlikely(ret <= 0)) {
 985                         sd->pos = prev_pos;
 986                         goto out_release;
 987                 }
 988
 989                 bytes += ret;
 990                 len -= ret;
 991                 sd->pos = pos;
 992
 993                 if (ret < read_len) {
 994                         sd->pos = prev_pos + ret;
 995                         goto out_release;
 996                 }
 997         }
 998
 999 done:
1000         pipe->tail = pipe->head = 0;
1001         file_accessed(in);
1002         return bytes;
1003
1004 out_release:
1005         /*
1006          * If we did an incomplete transfer we must release
1007          * the pipe buffers in question:
1008          */
1009         for (i = 0; i < pipe->ring_size; i++) {
1010                 struct pipe_buffer *buf = &pipe->bufs[i];
1011
1012                 if (buf->ops)
1013                         pipe_buf_release(pipe, buf);
1014         }
1015
1016         if (!bytes)
1017                 bytes = ret;
1018
1019         goto done;
1020 }
1021 EXPORT_SYMBOL(splice_direct_to_actor);
1022
1023 static int direct_splice_actor(struct pipe_inode_info *pipe,
1024                                struct splice_desc *sd)
1025 {
1026         struct file *file = sd->u.file;
1027
1028         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1029                               sd->flags);
1030 }
1031
1032 /**
1033  * do_splice_direct - splices data directly between two files
1034  * @in:         file to splice from
1035  * @ppos:       input file offset
1036  * @out:        file to splice to
1037  * @opos:       output file offset
1038  * @len:        number of bytes to splice
1039  * @flags:      splice modifier flags
1040  *
1041  * Description:
1042  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1043  *    doing it in the application would incur an extra system call
1044  *    (splice in + splice out, as compared to just sendfile()). So this helper
1045  *    can splice directly through a process-private pipe.
1046  *
1047  */
1048 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1049                       loff_t *opos, size_t len, unsigned int flags)
1050 {
1051         struct splice_desc sd = {
1052                 .len            = len,
1053                 .total_len      = len,
1054                 .flags          = flags,
1055                 .pos            = *ppos,
1056                 .u.file         = out,
1057                 .opos           = opos,
1058         };
1059         long ret;
1060
1061         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1062                 return -EBADF;
1063
1064         if (unlikely(out->f_flags & O_APPEND))
1065                 return -EINVAL;
1066
1067         ret = rw_verify_area(WRITE, out, opos, len);
1068         if (unlikely(ret < 0))
1069                 return ret;
1070
1071         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1072         if (ret > 0)
1073                 *ppos = sd.pos;
1074
1075         return ret;
1076 }
1077 EXPORT_SYMBOL(do_splice_direct);
1078
1079 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1080 {
1081         for (;;) {
1082                 if (unlikely(!pipe->readers)) {
1083                         send_sig(SIGPIPE, current, 0);
1084                         return -EPIPE;
1085                 }
1086                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1087                         return 0;
1088                 if (flags & SPLICE_F_NONBLOCK)
1089                         return -EAGAIN;
1090                 if (signal_pending(current))
1091                         return -ERESTARTSYS;
1092                 pipe_wait_writable(pipe);
1093         }
1094 }
1095
1096 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1097                                struct pipe_inode_info *opipe,
1098                                size_t len, unsigned int flags);
1099
1100 long splice_file_to_pipe(struct file *in,
1101                          struct pipe_inode_info *opipe,
1102                          loff_t *offset,
1103                          size_t len, unsigned int flags)
1104 {
1105         long ret;
1106
1107         pipe_lock(opipe);
1108         ret = wait_for_space(opipe, flags);
1109         if (!ret)
1110                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1111         pipe_unlock(opipe);
1112         if (ret > 0)
1113                 wakeup_pipe_readers(opipe);
1114         return ret;
1115 }
1116
1117 /*
1118  * Determine where to splice to/from.
1119  */
1120 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1121                loff_t *off_out, size_t len, unsigned int flags)
1122 {
1123         struct pipe_inode_info *ipipe;
1124         struct pipe_inode_info *opipe;
1125         loff_t offset;
1126         long ret;
1127
1128         if (unlikely(!(in->f_mode & FMODE_READ) ||
1129                      !(out->f_mode & FMODE_WRITE)))
1130                 return -EBADF;
1131
1132         ipipe = get_pipe_info(in, true);
1133         opipe = get_pipe_info(out, true);
1134
1135         if (ipipe && opipe) {
1136                 if (off_in || off_out)
1137                         return -ESPIPE;
1138
1139                 /* Splicing to self would be fun, but... */
1140                 if (ipipe == opipe)
1141                         return -EINVAL;
1142
1143                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1144                         flags |= SPLICE_F_NONBLOCK;
1145
1146                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1147         }
1148
1149         if (ipipe) {
1150                 if (off_in)
1151                         return -ESPIPE;
1152                 if (off_out) {
1153                         if (!(out->f_mode & FMODE_PWRITE))
1154                                 return -EINVAL;
1155                         offset = *off_out;
1156                 } else {
1157                         offset = out->f_pos;
1158                 }
1159
1160                 if (unlikely(out->f_flags & O_APPEND))
1161                         return -EINVAL;
1162
1163                 ret = rw_verify_area(WRITE, out, &offset, len);
1164                 if (unlikely(ret < 0))
1165                         return ret;
1166
1167                 if (in->f_flags & O_NONBLOCK)
1168                         flags |= SPLICE_F_NONBLOCK;
1169
1170                 file_start_write(out);
1171                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1172                 file_end_write(out);
1173
1174                 if (ret > 0)
1175                         fsnotify_modify(out);
1176
1177                 if (!off_out)
1178                         out->f_pos = offset;
1179                 else
1180                         *off_out = offset;
1181
1182                 return ret;
1183         }
1184
1185         if (opipe) {
1186                 if (off_out)
1187                         return -ESPIPE;
1188                 if (off_in) {
1189                         if (!(in->f_mode & FMODE_PREAD))
1190                                 return -EINVAL;
1191                         offset = *off_in;
1192                 } else {
1193                         offset = in->f_pos;
1194                 }
1195
1196                 if (out->f_flags & O_NONBLOCK)
1197                         flags |= SPLICE_F_NONBLOCK;
1198
1199                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1200
1201                 if (ret > 0)
1202                         fsnotify_access(in);
1203
1204                 if (!off_in)
1205                         in->f_pos = offset;
1206                 else
1207                         *off_in = offset;
1208
1209                 return ret;
1210         }
1211
1212         return -EINVAL;
1213 }
1214
1215 static long __do_splice(struct file *in, loff_t __user *off_in,
1216                         struct file *out, loff_t __user *off_out,
1217                         size_t len, unsigned int flags)
1218 {
1219         struct pipe_inode_info *ipipe;
1220         struct pipe_inode_info *opipe;
1221         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1222         long ret;
1223
1224         ipipe = get_pipe_info(in, true);
1225         opipe = get_pipe_info(out, true);
1226
1227         if (ipipe) {
1228                 if (off_in)
1229                         return -ESPIPE;
1230                 pipe_clear_nowait(in);
1231         }
1232         if (opipe) {
1233                 if (off_out)
1234                         return -ESPIPE;
1235                 pipe_clear_nowait(out);
1236         }
1237
1238         if (off_out) {
1239                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1240                         return -EFAULT;
1241                 __off_out = &offset;
1242         }
1243         if (off_in) {
1244                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1245                         return -EFAULT;
1246                 __off_in = &offset;
1247         }
1248
1249         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1250         if (ret < 0)
1251                 return ret;
1252
1253         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1254                 return -EFAULT;
1255         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1256                 return -EFAULT;
1257
1258         return ret;
1259 }
1260
1261 static int iter_to_pipe(struct iov_iter *from,
1262                         struct pipe_inode_info *pipe,
1263                         unsigned flags)
1264 {
1265         struct pipe_buffer buf = {
1266                 .ops = &user_page_pipe_buf_ops,
1267                 .flags = flags
1268         };
1269         size_t total = 0;
1270         int ret = 0;
1271
1272         while (iov_iter_count(from)) {
1273                 struct page *pages[16];
1274                 ssize_t left;
1275                 size_t start;
1276                 int i, n;
1277
1278                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1279                 if (left <= 0) {
1280                         ret = left;
1281                         break;
1282                 }
1283
1284                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1285                 for (i = 0; i < n; i++) {
1286                         int size = min_t(int, left, PAGE_SIZE - start);
1287
1288                         buf.page = pages[i];
1289                         buf.offset = start;
1290                         buf.len = size;
1291                         ret = add_to_pipe(pipe, &buf);
1292                         if (unlikely(ret < 0)) {
1293                                 iov_iter_revert(from, left);
1294                                 // this one got dropped by add_to_pipe()
1295                                 while (++i < n)
1296                                         put_page(pages[i]);
1297                                 goto out;
1298                         }
1299                         total += ret;
1300                         left -= size;
1301                         start = 0;
1302                 }
1303         }
1304 out:
1305         return total ? total : ret;
1306 }
1307
1308 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1309                         struct splice_desc *sd)
1310 {
1311         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1312         return n == sd->len ? n : -EFAULT;
1313 }
1314
1315 /*
1316  * For lack of a better implementation, implement vmsplice() to userspace
1317  * as a simple copy of the pipes pages to the user iov.
1318  */
1319 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1320                              unsigned int flags)
1321 {
1322         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1323         struct splice_desc sd = {
1324                 .total_len = iov_iter_count(iter),
1325                 .flags = flags,
1326                 .u.data = iter
1327         };
1328         long ret = 0;
1329
1330         if (!pipe)
1331                 return -EBADF;
1332
1333         pipe_clear_nowait(file);
1334
1335         if (sd.total_len) {
1336                 pipe_lock(pipe);
1337                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1338                 pipe_unlock(pipe);
1339         }
1340
1341         return ret;
1342 }
1343
1344 /*
1345  * vmsplice splices a user address range into a pipe. It can be thought of
1346  * as splice-from-memory, where the regular splice is splice-from-file (or
1347  * to file). In both cases the output is a pipe, naturally.
1348  */
1349 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1350                              unsigned int flags)
1351 {
1352         struct pipe_inode_info *pipe;
1353         long ret = 0;
1354         unsigned buf_flag = 0;
1355
1356         if (flags & SPLICE_F_GIFT)
1357                 buf_flag = PIPE_BUF_FLAG_GIFT;
1358
1359         pipe = get_pipe_info(file, true);
1360         if (!pipe)
1361                 return -EBADF;
1362
1363         pipe_clear_nowait(file);
1364
1365         pipe_lock(pipe);
1366         ret = wait_for_space(pipe, flags);
1367         if (!ret)
1368                 ret = iter_to_pipe(iter, pipe, buf_flag);
1369         pipe_unlock(pipe);
1370         if (ret > 0)
1371                 wakeup_pipe_readers(pipe);
1372         return ret;
1373 }
1374
1375 static int vmsplice_type(struct fd f, int *type)
1376 {
1377         if (!f.file)
1378                 return -EBADF;
1379         if (f.file->f_mode & FMODE_WRITE) {
1380                 *type = ITER_SOURCE;
1381         } else if (f.file->f_mode & FMODE_READ) {
1382                 *type = ITER_DEST;
1383         } else {
1384                 fdput(f);
1385                 return -EBADF;
1386         }
1387         return 0;
1388 }
1389
1390 /*
1391  * Note that vmsplice only really supports true splicing _from_ user memory
1392  * to a pipe, not the other way around. Splicing from user memory is a simple
1393  * operation that can be supported without any funky alignment restrictions
1394  * or nasty vm tricks. We simply map in the user memory and fill them into
1395  * a pipe. The reverse isn't quite as easy, though. There are two possible
1396  * solutions for that:
1397  *
1398  *      - memcpy() the data internally, at which point we might as well just
1399  *        do a regular read() on the buffer anyway.
1400  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1401  *        has restriction limitations on both ends of the pipe).
1402  *
1403  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1404  *
1405  */
1406 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1407                 unsigned long, nr_segs, unsigned int, flags)
1408 {
1409         struct iovec iovstack[UIO_FASTIOV];
1410         struct iovec *iov = iovstack;
1411         struct iov_iter iter;
1412         ssize_t error;
1413         struct fd f;
1414         int type;
1415
1416         if (unlikely(flags & ~SPLICE_F_ALL))
1417                 return -EINVAL;
1418
1419         f = fdget(fd);
1420         error = vmsplice_type(f, &type);
1421         if (error)
1422                 return error;
1423
1424         error = import_iovec(type, uiov, nr_segs,
1425                              ARRAY_SIZE(iovstack), &iov, &iter);
1426         if (error < 0)
1427                 goto out_fdput;
1428
1429         if (!iov_iter_count(&iter))
1430                 error = 0;
1431         else if (type == ITER_SOURCE)
1432                 error = vmsplice_to_pipe(f.file, &iter, flags);
1433         else
1434                 error = vmsplice_to_user(f.file, &iter, flags);
1435
1436         kfree(iov);
1437 out_fdput:
1438         fdput(f);
1439         return error;
1440 }
1441
1442 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1443                 int, fd_out, loff_t __user *, off_out,
1444                 size_t, len, unsigned int, flags)
1445 {
1446         struct fd in, out;
1447         long error;
1448
1449         if (unlikely(!len))
1450                 return 0;
1451
1452         if (unlikely(flags & ~SPLICE_F_ALL))
1453                 return -EINVAL;
1454
1455         error = -EBADF;
1456         in = fdget(fd_in);
1457         if (in.file) {
1458                 out = fdget(fd_out);
1459                 if (out.file) {
1460                         error = __do_splice(in.file, off_in, out.file, off_out,
1461                                                 len, flags);
1462                         fdput(out);
1463                 }
1464                 fdput(in);
1465         }
1466         return error;
1467 }
1468
1469 /*
1470  * Make sure there's data to read. Wait for input if we can, otherwise
1471  * return an appropriate error.
1472  */
1473 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1474 {
1475         int ret;
1476
1477         /*
1478          * Check the pipe occupancy without the inode lock first. This function
1479          * is speculative anyways, so missing one is ok.
1480          */
1481         if (!pipe_empty(pipe->head, pipe->tail))
1482                 return 0;
1483
1484         ret = 0;
1485         pipe_lock(pipe);
1486
1487         while (pipe_empty(pipe->head, pipe->tail)) {
1488                 if (signal_pending(current)) {
1489                         ret = -ERESTARTSYS;
1490                         break;
1491                 }
1492                 if (!pipe->writers)
1493                         break;
1494                 if (flags & SPLICE_F_NONBLOCK) {
1495                         ret = -EAGAIN;
1496                         break;
1497                 }
1498                 pipe_wait_readable(pipe);
1499         }
1500
1501         pipe_unlock(pipe);
1502         return ret;
1503 }
1504
1505 /*
1506  * Make sure there's writeable room. Wait for room if we can, otherwise
1507  * return an appropriate error.
1508  */
1509 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1510 {
1511         int ret;
1512
1513         /*
1514          * Check pipe occupancy without the inode lock first. This function
1515          * is speculative anyways, so missing one is ok.
1516          */
1517         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1518                 return 0;
1519
1520         ret = 0;
1521         pipe_lock(pipe);
1522
1523         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1524                 if (!pipe->readers) {
1525                         send_sig(SIGPIPE, current, 0);
1526                         ret = -EPIPE;
1527                         break;
1528                 }
1529                 if (flags & SPLICE_F_NONBLOCK) {
1530                         ret = -EAGAIN;
1531                         break;
1532                 }
1533                 if (signal_pending(current)) {
1534                         ret = -ERESTARTSYS;
1535                         break;
1536                 }
1537                 pipe_wait_writable(pipe);
1538         }
1539
1540         pipe_unlock(pipe);
1541         return ret;
1542 }
1543
1544 /*
1545  * Splice contents of ipipe to opipe.
1546  */
1547 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1548                                struct pipe_inode_info *opipe,
1549                                size_t len, unsigned int flags)
1550 {
1551         struct pipe_buffer *ibuf, *obuf;
1552         unsigned int i_head, o_head;
1553         unsigned int i_tail, o_tail;
1554         unsigned int i_mask, o_mask;
1555         int ret = 0;
1556         bool input_wakeup = false;
1557
1558
1559 retry:
1560         ret = ipipe_prep(ipipe, flags);
1561         if (ret)
1562                 return ret;
1563
1564         ret = opipe_prep(opipe, flags);
1565         if (ret)
1566                 return ret;
1567
1568         /*
1569          * Potential ABBA deadlock, work around it by ordering lock
1570          * grabbing by pipe info address. Otherwise two different processes
1571          * could deadlock (one doing tee from A -> B, the other from B -> A).
1572          */
1573         pipe_double_lock(ipipe, opipe);
1574
1575         i_tail = ipipe->tail;
1576         i_mask = ipipe->ring_size - 1;
1577         o_head = opipe->head;
1578         o_mask = opipe->ring_size - 1;
1579
1580         do {
1581                 size_t o_len;
1582
1583                 if (!opipe->readers) {
1584                         send_sig(SIGPIPE, current, 0);
1585                         if (!ret)
1586                                 ret = -EPIPE;
1587                         break;
1588                 }
1589
1590                 i_head = ipipe->head;
1591                 o_tail = opipe->tail;
1592
1593                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1594                         break;
1595
1596                 /*
1597                  * Cannot make any progress, because either the input
1598                  * pipe is empty or the output pipe is full.
1599                  */
1600                 if (pipe_empty(i_head, i_tail) ||
1601                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1602                         /* Already processed some buffers, break */
1603                         if (ret)
1604                                 break;
1605
1606                         if (flags & SPLICE_F_NONBLOCK) {
1607                                 ret = -EAGAIN;
1608                                 break;
1609                         }
1610
1611                         /*
1612                          * We raced with another reader/writer and haven't
1613                          * managed to process any buffers.  A zero return
1614                          * value means EOF, so retry instead.
1615                          */
1616                         pipe_unlock(ipipe);
1617                         pipe_unlock(opipe);
1618                         goto retry;
1619                 }
1620
1621                 ibuf = &ipipe->bufs[i_tail & i_mask];
1622                 obuf = &opipe->bufs[o_head & o_mask];
1623
1624                 if (len >= ibuf->len) {
1625                         /*
1626                          * Simply move the whole buffer from ipipe to opipe
1627                          */
1628                         *obuf = *ibuf;
1629                         ibuf->ops = NULL;
1630                         i_tail++;
1631                         ipipe->tail = i_tail;
1632                         input_wakeup = true;
1633                         o_len = obuf->len;
1634                         o_head++;
1635                         opipe->head = o_head;
1636                 } else {
1637                         /*
1638                          * Get a reference to this pipe buffer,
1639                          * so we can copy the contents over.
1640                          */
1641                         if (!pipe_buf_get(ipipe, ibuf)) {
1642                                 if (ret == 0)
1643                                         ret = -EFAULT;
1644                                 break;
1645                         }
1646                         *obuf = *ibuf;
1647
1648                         /*
1649                          * Don't inherit the gift and merge flags, we need to
1650                          * prevent multiple steals of this page.
1651                          */
1652                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1653                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1654
1655                         obuf->len = len;
1656                         ibuf->offset += len;
1657                         ibuf->len -= len;
1658                         o_len = len;
1659                         o_head++;
1660                         opipe->head = o_head;
1661                 }
1662                 ret += o_len;
1663                 len -= o_len;
1664         } while (len);
1665
1666         pipe_unlock(ipipe);
1667         pipe_unlock(opipe);
1668
1669         /*
1670          * If we put data in the output pipe, wakeup any potential readers.
1671          */
1672         if (ret > 0)
1673                 wakeup_pipe_readers(opipe);
1674
1675         if (input_wakeup)
1676                 wakeup_pipe_writers(ipipe);
1677
1678         return ret;
1679 }
1680
1681 /*
1682  * Link contents of ipipe to opipe.
1683  */
1684 static int link_pipe(struct pipe_inode_info *ipipe,
1685                      struct pipe_inode_info *opipe,
1686                      size_t len, unsigned int flags)
1687 {
1688         struct pipe_buffer *ibuf, *obuf;
1689         unsigned int i_head, o_head;
1690         unsigned int i_tail, o_tail;
1691         unsigned int i_mask, o_mask;
1692         int ret = 0;
1693
1694         /*
1695          * Potential ABBA deadlock, work around it by ordering lock
1696          * grabbing by pipe info address. Otherwise two different processes
1697          * could deadlock (one doing tee from A -> B, the other from B -> A).
1698          */
1699         pipe_double_lock(ipipe, opipe);
1700
1701         i_tail = ipipe->tail;
1702         i_mask = ipipe->ring_size - 1;
1703         o_head = opipe->head;
1704         o_mask = opipe->ring_size - 1;
1705
1706         do {
1707                 if (!opipe->readers) {
1708                         send_sig(SIGPIPE, current, 0);
1709                         if (!ret)
1710                                 ret = -EPIPE;
1711                         break;
1712                 }
1713
1714                 i_head = ipipe->head;
1715                 o_tail = opipe->tail;
1716
1717                 /*
1718                  * If we have iterated all input buffers or run out of
1719                  * output room, break.
1720                  */
1721                 if (pipe_empty(i_head, i_tail) ||
1722                     pipe_full(o_head, o_tail, opipe->max_usage))
1723                         break;
1724
1725                 ibuf = &ipipe->bufs[i_tail & i_mask];
1726                 obuf = &opipe->bufs[o_head & o_mask];
1727
1728                 /*
1729                  * Get a reference to this pipe buffer,
1730                  * so we can copy the contents over.
1731                  */
1732                 if (!pipe_buf_get(ipipe, ibuf)) {
1733                         if (ret == 0)
1734                                 ret = -EFAULT;
1735                         break;
1736                 }
1737
1738                 *obuf = *ibuf;
1739
1740                 /*
1741                  * Don't inherit the gift and merge flag, we need to prevent
1742                  * multiple steals of this page.
1743                  */
1744                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1745                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1746
1747                 if (obuf->len > len)
1748                         obuf->len = len;
1749                 ret += obuf->len;
1750                 len -= obuf->len;
1751
1752                 o_head++;
1753                 opipe->head = o_head;
1754                 i_tail++;
1755         } while (len);
1756
1757         pipe_unlock(ipipe);
1758         pipe_unlock(opipe);
1759
1760         /*
1761          * If we put data in the output pipe, wakeup any potential readers.
1762          */
1763         if (ret > 0)
1764                 wakeup_pipe_readers(opipe);
1765
1766         return ret;
1767 }
1768
1769 /*
1770  * This is a tee(1) implementation that works on pipes. It doesn't copy
1771  * any data, it simply references the 'in' pages on the 'out' pipe.
1772  * The 'flags' used are the SPLICE_F_* variants, currently the only
1773  * applicable one is SPLICE_F_NONBLOCK.
1774  */
1775 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1776 {
1777         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1778         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1779         int ret = -EINVAL;
1780
1781         if (unlikely(!(in->f_mode & FMODE_READ) ||
1782                      !(out->f_mode & FMODE_WRITE)))
1783                 return -EBADF;
1784
1785         /*
1786          * Duplicate the contents of ipipe to opipe without actually
1787          * copying the data.
1788          */
1789         if (ipipe && opipe && ipipe != opipe) {
1790                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1791                         flags |= SPLICE_F_NONBLOCK;
1792
1793                 /*
1794                  * Keep going, unless we encounter an error. The ipipe/opipe
1795                  * ordering doesn't really matter.
1796                  */
1797                 ret = ipipe_prep(ipipe, flags);
1798                 if (!ret) {
1799                         ret = opipe_prep(opipe, flags);
1800                         if (!ret)
1801                                 ret = link_pipe(ipipe, opipe, len, flags);
1802                 }
1803         }
1804
1805         return ret;
1806 }
1807
1808 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1809 {
1810         struct fd in, out;
1811         int error;
1812
1813         if (unlikely(flags & ~SPLICE_F_ALL))
1814                 return -EINVAL;
1815
1816         if (unlikely(!len))
1817                 return 0;
1818
1819         error = -EBADF;
1820         in = fdget(fdin);
1821         if (in.file) {
1822                 out = fdget(fdout);
1823                 if (out.file) {
1824                         error = do_tee(in.file, out.file, len, flags);
1825                         fdput(out);
1826                 }
1827                 fdput(in);
1828         }
1829
1830         return error;
1831 }