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/net.h>
  37 #include <linux/socket.h>
  38 #include <linux/sched/signal.h>
  39
  40 #include "internal.h"
  41
  42 /*
  43  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  44  * indicate they support non-blocking reads or writes, we must clear it
  45  * here if set to avoid blocking other users of this pipe if splice is
  46  * being done on it.
  47  */
  48 static noinline void noinline pipe_clear_nowait(struct file *file)
  49 {
  50         fmode_t fmode = READ_ONCE(file->f_mode);
  51
  52         do {
  53                 if (!(fmode & FMODE_NOWAIT))
  54                         break;
  55         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  56 }
  57
  58 /*
  59  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  60  * a vm helper function, it's already simplified quite a bit by the
  61  * addition of remove_mapping(). If success is returned, the caller may
  62  * attempt to reuse this page for another destination.
  63  */
  64 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  65                 struct pipe_buffer *buf)
  66 {
  67         struct folio *folio = page_folio(buf->page);
  68         struct address_space *mapping;
  69
  70         folio_lock(folio);
  71
  72         mapping = folio_mapping(folio);
  73         if (mapping) {
  74                 WARN_ON(!folio_test_uptodate(folio));
  75
  76                 /*
  77                  * At least for ext2 with nobh option, we need to wait on
  78                  * writeback completing on this folio, since we'll remove it
  79                  * from the pagecache.  Otherwise truncate wont wait on the
  80                  * folio, allowing the disk blocks to be reused by someone else
  81                  * before we actually wrote our data to them. fs corruption
  82                  * ensues.
  83                  */
  84                 folio_wait_writeback(folio);
  85
  86                 if (!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 folio *folio = page_folio(buf->page);
 123         int err;
 124
 125         if (!folio_test_uptodate(folio)) {
 126                 folio_lock(folio);
 127
 128                 /*
 129                  * Folio got truncated/unhashed. This will cause a 0-byte
 130                  * splice, if this is the first page.
 131                  */
 132                 if (!folio->mapping) {
 133                         err = -ENODATA;
 134                         goto error;
 135                 }
 136
 137                 /*
 138                  * Uh oh, read-error from disk.
 139                  */
 140                 if (!folio_test_uptodate(folio)) {
 141                         err = -EIO;
 142                         goto error;
 143                 }
 144
 145                 /* Folio is ok after all, we are done */
 146                 folio_unlock(folio);
 147         }
 148
 149         return 0;
 150 error:
 151         folio_unlock(folio);
 152         return err;
 153 }
 154
 155 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 156         .confirm        = page_cache_pipe_buf_confirm,
 157         .release        = page_cache_pipe_buf_release,
 158         .try_steal      = page_cache_pipe_buf_try_steal,
 159         .get            = generic_pipe_buf_get,
 160 };
 161
 162 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 163                 struct pipe_buffer *buf)
 164 {
 165         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 166                 return false;
 167
 168         buf->flags |= PIPE_BUF_FLAG_LRU;
 169         return generic_pipe_buf_try_steal(pipe, buf);
 170 }
 171
 172 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 173         .release        = page_cache_pipe_buf_release,
 174         .try_steal      = user_page_pipe_buf_try_steal,
 175         .get            = generic_pipe_buf_get,
 176 };
 177
 178 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 179 {
 180         smp_mb();
 181         if (waitqueue_active(&pipe->rd_wait))
 182                 wake_up_interruptible(&pipe->rd_wait);
 183         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 184 }
 185
 186 /**
 187  * splice_to_pipe - fill passed data into a pipe
 188  * @pipe:       pipe to fill
 189  * @spd:        data to fill
 190  *
 191  * Description:
 192  *    @spd contains a map of pages and len/offset tuples, along with
 193  *    the struct pipe_buf_operations associated with these pages. This
 194  *    function will link that data to the pipe.
 195  *
 196  */
 197 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 198                        struct splice_pipe_desc *spd)
 199 {
 200         unsigned int spd_pages = spd->nr_pages;
 201         unsigned int tail = pipe->tail;
 202         unsigned int head = pipe->head;
 203         unsigned int mask = pipe->ring_size - 1;
 204         int ret = 0, page_nr = 0;
 205
 206         if (!spd_pages)
 207                 return 0;
 208
 209         if (unlikely(!pipe->readers)) {
 210                 send_sig(SIGPIPE, current, 0);
 211                 ret = -EPIPE;
 212                 goto out;
 213         }
 214
 215         while (!pipe_full(head, tail, pipe->max_usage)) {
 216                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 217
 218                 buf->page = spd->pages[page_nr];
 219                 buf->offset = spd->partial[page_nr].offset;
 220                 buf->len = spd->partial[page_nr].len;
 221                 buf->private = spd->partial[page_nr].private;
 222                 buf->ops = spd->ops;
 223                 buf->flags = 0;
 224
 225                 head++;
 226                 pipe->head = head;
 227                 page_nr++;
 228                 ret += buf->len;
 229
 230                 if (!--spd->nr_pages)
 231                         break;
 232         }
 233
 234         if (!ret)
 235                 ret = -EAGAIN;
 236
 237 out:
 238         while (page_nr < spd_pages)
 239                 spd->spd_release(spd, page_nr++);
 240
 241         return ret;
 242 }
 243 EXPORT_SYMBOL_GPL(splice_to_pipe);
 244
 245 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 246 {
 247         unsigned int head = pipe->head;
 248         unsigned int tail = pipe->tail;
 249         unsigned int mask = pipe->ring_size - 1;
 250         int ret;
 251
 252         if (unlikely(!pipe->readers)) {
 253                 send_sig(SIGPIPE, current, 0);
 254                 ret = -EPIPE;
 255         } else if (pipe_full(head, tail, pipe->max_usage)) {
 256                 ret = -EAGAIN;
 257         } else {
 258                 pipe->bufs[head & mask] = *buf;
 259                 pipe->head = head + 1;
 260                 return buf->len;
 261         }
 262         pipe_buf_release(pipe, buf);
 263         return ret;
 264 }
 265 EXPORT_SYMBOL(add_to_pipe);
 266
 267 /*
 268  * Check if we need to grow the arrays holding pages and partial page
 269  * descriptions.
 270  */
 271 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 272 {
 273         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 274
 275         spd->nr_pages_max = max_usage;
 276         if (max_usage <= PIPE_DEF_BUFFERS)
 277                 return 0;
 278
 279         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 280         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 281                                      GFP_KERNEL);
 282
 283         if (spd->pages && spd->partial)
 284                 return 0;
 285
 286         kfree(spd->pages);
 287         kfree(spd->partial);
 288         return -ENOMEM;
 289 }
 290
 291 void splice_shrink_spd(struct splice_pipe_desc *spd)
 292 {
 293         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 294                 return;
 295
 296         kfree(spd->pages);
 297         kfree(spd->partial);
 298 }
 299
 300 /**
 301  * copy_splice_read -  Copy data from a file and splice the copy into a pipe
 302  * @in: The file to read from
 303  * @ppos: Pointer to the file position to read from
 304  * @pipe: The pipe to splice into
 305  * @len: The amount to splice
 306  * @flags: The SPLICE_F_* flags
 307  *
 308  * This function allocates a bunch of pages sufficient to hold the requested
 309  * amount of data (but limited by the remaining pipe capacity), passes it to
 310  * the file's ->read_iter() to read into and then splices the used pages into
 311  * the pipe.
 312  *
 313  * Return: On success, the number of bytes read will be returned and *@ppos
 314  * will be updated if appropriate; 0 will be returned if there is no more data
 315  * to be read; -EAGAIN will be returned if the pipe had no space, and some
 316  * other negative error code will be returned on error.  A short read may occur
 317  * if the pipe has insufficient space, we reach the end of the data or we hit a
 318  * hole.
 319  */
 320 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 321                          struct pipe_inode_info *pipe,
 322                          size_t len, unsigned int flags)
 323 {
 324         struct iov_iter to;
 325         struct bio_vec *bv;
 326         struct kiocb kiocb;
 327         struct page **pages;
 328         ssize_t ret;
 329         size_t used, npages, chunk, remain, keep = 0;
 330         int i;
 331
 332         /* Work out how much data we can actually add into the pipe */
 333         used = pipe_occupancy(pipe->head, pipe->tail);
 334         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 335         len = min_t(size_t, len, npages * PAGE_SIZE);
 336         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 337
 338         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 339                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 340         if (!bv)
 341                 return -ENOMEM;
 342
 343         pages = (struct page **)(bv + npages);
 344         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 345         if (!npages) {
 346                 kfree(bv);
 347                 return -ENOMEM;
 348         }
 349
 350         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 351
 352         for (i = 0; i < npages; i++) {
 353                 chunk = min_t(size_t, PAGE_SIZE, remain);
 354                 bv[i].bv_page = pages[i];
 355                 bv[i].bv_offset = 0;
 356                 bv[i].bv_len = chunk;
 357                 remain -= chunk;
 358         }
 359
 360         /* Do the I/O */
 361         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 362         init_sync_kiocb(&kiocb, in);
 363         kiocb.ki_pos = *ppos;
 364         ret = call_read_iter(in, &kiocb, &to);
 365
 366         if (ret > 0) {
 367                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 368                 *ppos = kiocb.ki_pos;
 369         }
 370
 371         /*
 372          * Callers of ->splice_read() expect -EAGAIN on "can't put anything in
 373          * there", rather than -EFAULT.
 374          */
 375         if (ret == -EFAULT)
 376                 ret = -EAGAIN;
 377
 378         /* Free any pages that didn't get touched at all. */
 379         if (keep < npages)
 380                 release_pages(pages + keep, npages - keep);
 381
 382         /* Push the remaining pages into the pipe. */
 383         remain = ret;
 384         for (i = 0; i < keep; i++) {
 385                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 386
 387                 chunk = min_t(size_t, remain, PAGE_SIZE);
 388                 *buf = (struct pipe_buffer) {
 389                         .ops    = &default_pipe_buf_ops,
 390                         .page   = bv[i].bv_page,
 391                         .offset = 0,
 392                         .len    = chunk,
 393                 };
 394                 pipe->head++;
 395                 remain -= chunk;
 396         }
 397
 398         kfree(bv);
 399         return ret;
 400 }
 401 EXPORT_SYMBOL(copy_splice_read);
 402
 403 const struct pipe_buf_operations default_pipe_buf_ops = {
 404         .release        = generic_pipe_buf_release,
 405         .try_steal      = generic_pipe_buf_try_steal,
 406         .get            = generic_pipe_buf_get,
 407 };
 408
 409 /* Pipe buffer operations for a socket and similar. */
 410 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 411         .release        = generic_pipe_buf_release,
 412         .get            = generic_pipe_buf_get,
 413 };
 414 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 415
 416 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 417 {
 418         smp_mb();
 419         if (waitqueue_active(&pipe->wr_wait))
 420                 wake_up_interruptible(&pipe->wr_wait);
 421         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 422 }
 423
 424 /**
 425  * splice_from_pipe_feed - feed available data from a pipe to a file
 426  * @pipe:       pipe to splice from
 427  * @sd:         information to @actor
 428  * @actor:      handler that splices the data
 429  *
 430  * Description:
 431  *    This function loops over the pipe and calls @actor to do the
 432  *    actual moving of a single struct pipe_buffer to the desired
 433  *    destination.  It returns when there's no more buffers left in
 434  *    the pipe or if the requested number of bytes (@sd->total_len)
 435  *    have been copied.  It returns a positive number (one) if the
 436  *    pipe needs to be filled with more data, zero if the required
 437  *    number of bytes have been copied and -errno on error.
 438  *
 439  *    This, together with splice_from_pipe_{begin,end,next}, may be
 440  *    used to implement the functionality of __splice_from_pipe() when
 441  *    locking is required around copying the pipe buffers to the
 442  *    destination.
 443  */
 444 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 445                           splice_actor *actor)
 446 {
 447         unsigned int head = pipe->head;
 448         unsigned int tail = pipe->tail;
 449         unsigned int mask = pipe->ring_size - 1;
 450         int ret;
 451
 452         while (!pipe_empty(head, tail)) {
 453                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 454
 455                 sd->len = buf->len;
 456                 if (sd->len > sd->total_len)
 457                         sd->len = sd->total_len;
 458
 459                 ret = pipe_buf_confirm(pipe, buf);
 460                 if (unlikely(ret)) {
 461                         if (ret == -ENODATA)
 462                                 ret = 0;
 463                         return ret;
 464                 }
 465
 466                 ret = actor(pipe, buf, sd);
 467                 if (ret <= 0)
 468                         return ret;
 469
 470                 buf->offset += ret;
 471                 buf->len -= ret;
 472
 473                 sd->num_spliced += ret;
 474                 sd->len -= ret;
 475                 sd->pos += ret;
 476                 sd->total_len -= ret;
 477
 478                 if (!buf->len) {
 479                         pipe_buf_release(pipe, buf);
 480                         tail++;
 481                         pipe->tail = tail;
 482                         if (pipe->files)
 483                                 sd->need_wakeup = true;
 484                 }
 485
 486                 if (!sd->total_len)
 487                         return 0;
 488         }
 489
 490         return 1;
 491 }
 492
 493 /* We know we have a pipe buffer, but maybe it's empty? */
 494 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 495 {
 496         unsigned int tail = pipe->tail;
 497         unsigned int mask = pipe->ring_size - 1;
 498         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 499
 500         if (unlikely(!buf->len)) {
 501                 pipe_buf_release(pipe, buf);
 502                 pipe->tail = tail+1;
 503                 return true;
 504         }
 505
 506         return false;
 507 }
 508
 509 /**
 510  * splice_from_pipe_next - wait for some data to splice from
 511  * @pipe:       pipe to splice from
 512  * @sd:         information about the splice operation
 513  *
 514  * Description:
 515  *    This function will wait for some data and return a positive
 516  *    value (one) if pipe buffers are available.  It will return zero
 517  *    or -errno if no more data needs to be spliced.
 518  */
 519 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 520 {
 521         /*
 522          * Check for signal early to make process killable when there are
 523          * always buffers available
 524          */
 525         if (signal_pending(current))
 526                 return -ERESTARTSYS;
 527
 528 repeat:
 529         while (pipe_empty(pipe->head, pipe->tail)) {
 530                 if (!pipe->writers)
 531                         return 0;
 532
 533                 if (sd->num_spliced)
 534                         return 0;
 535
 536                 if (sd->flags & SPLICE_F_NONBLOCK)
 537                         return -EAGAIN;
 538
 539                 if (signal_pending(current))
 540                         return -ERESTARTSYS;
 541
 542                 if (sd->need_wakeup) {
 543                         wakeup_pipe_writers(pipe);
 544                         sd->need_wakeup = false;
 545                 }
 546
 547                 pipe_wait_readable(pipe);
 548         }
 549
 550         if (eat_empty_buffer(pipe))
 551                 goto repeat;
 552
 553         return 1;
 554 }
 555
 556 /**
 557  * splice_from_pipe_begin - start splicing from pipe
 558  * @sd:         information about the splice operation
 559  *
 560  * Description:
 561  *    This function should be called before a loop containing
 562  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 563  *    initialize the necessary fields of @sd.
 564  */
 565 static void splice_from_pipe_begin(struct splice_desc *sd)
 566 {
 567         sd->num_spliced = 0;
 568         sd->need_wakeup = false;
 569 }
 570
 571 /**
 572  * splice_from_pipe_end - finish splicing from pipe
 573  * @pipe:       pipe to splice from
 574  * @sd:         information about the splice operation
 575  *
 576  * Description:
 577  *    This function will wake up pipe writers if necessary.  It should
 578  *    be called after a loop containing splice_from_pipe_next() and
 579  *    splice_from_pipe_feed().
 580  */
 581 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 582 {
 583         if (sd->need_wakeup)
 584                 wakeup_pipe_writers(pipe);
 585 }
 586
 587 /**
 588  * __splice_from_pipe - splice data from a pipe to given actor
 589  * @pipe:       pipe to splice from
 590  * @sd:         information to @actor
 591  * @actor:      handler that splices the data
 592  *
 593  * Description:
 594  *    This function does little more than loop over the pipe and call
 595  *    @actor to do the actual moving of a single struct pipe_buffer to
 596  *    the desired destination. See pipe_to_file, pipe_to_sendmsg, or
 597  *    pipe_to_user.
 598  *
 599  */
 600 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 601                            splice_actor *actor)
 602 {
 603         int ret;
 604
 605         splice_from_pipe_begin(sd);
 606         do {
 607                 cond_resched();
 608                 ret = splice_from_pipe_next(pipe, sd);
 609                 if (ret > 0)
 610                         ret = splice_from_pipe_feed(pipe, sd, actor);
 611         } while (ret > 0);
 612         splice_from_pipe_end(pipe, sd);
 613
 614         return sd->num_spliced ? sd->num_spliced : ret;
 615 }
 616 EXPORT_SYMBOL(__splice_from_pipe);
 617
 618 /**
 619  * splice_from_pipe - splice data from a pipe to a file
 620  * @pipe:       pipe to splice from
 621  * @out:        file to splice to
 622  * @ppos:       position in @out
 623  * @len:        how many bytes to splice
 624  * @flags:      splice modifier flags
 625  * @actor:      handler that splices the data
 626  *
 627  * Description:
 628  *    See __splice_from_pipe. This function locks the pipe inode,
 629  *    otherwise it's identical to __splice_from_pipe().
 630  *
 631  */
 632 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 633                          loff_t *ppos, size_t len, unsigned int flags,
 634                          splice_actor *actor)
 635 {
 636         ssize_t ret;
 637         struct splice_desc sd = {
 638                 .total_len = len,
 639                 .flags = flags,
 640                 .pos = *ppos,
 641                 .u.file = out,
 642         };
 643
 644         pipe_lock(pipe);
 645         ret = __splice_from_pipe(pipe, &sd, actor);
 646         pipe_unlock(pipe);
 647
 648         return ret;
 649 }
 650
 651 /**
 652  * iter_file_splice_write - splice data from a pipe to a file
 653  * @pipe:       pipe info
 654  * @out:        file to write to
 655  * @ppos:       position in @out
 656  * @len:        number of bytes to splice
 657  * @flags:      splice modifier flags
 658  *
 659  * Description:
 660  *    Will either move or copy pages (determined by @flags options) from
 661  *    the given pipe inode to the given file.
 662  *    This one is ->write_iter-based.
 663  *
 664  */
 665 ssize_t
 666 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 667                           loff_t *ppos, size_t len, unsigned int flags)
 668 {
 669         struct splice_desc sd = {
 670                 .total_len = len,
 671                 .flags = flags,
 672                 .pos = *ppos,
 673                 .u.file = out,
 674         };
 675         int nbufs = pipe->max_usage;
 676         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 677                                         GFP_KERNEL);
 678         ssize_t ret;
 679
 680         if (unlikely(!array))
 681                 return -ENOMEM;
 682
 683         pipe_lock(pipe);
 684
 685         splice_from_pipe_begin(&sd);
 686         while (sd.total_len) {
 687                 struct iov_iter from;
 688                 unsigned int head, tail, mask;
 689                 size_t left;
 690                 int n;
 691
 692                 ret = splice_from_pipe_next(pipe, &sd);
 693                 if (ret <= 0)
 694                         break;
 695
 696                 if (unlikely(nbufs < pipe->max_usage)) {
 697                         kfree(array);
 698                         nbufs = pipe->max_usage;
 699                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 700                                         GFP_KERNEL);
 701                         if (!array) {
 702                                 ret = -ENOMEM;
 703                                 break;
 704                         }
 705                 }
 706
 707                 head = pipe->head;
 708                 tail = pipe->tail;
 709                 mask = pipe->ring_size - 1;
 710
 711                 /* build the vector */
 712                 left = sd.total_len;
 713                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 714                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 715                         size_t this_len = buf->len;
 716
 717                         /* zero-length bvecs are not supported, skip them */
 718                         if (!this_len)
 719                                 continue;
 720                         this_len = min(this_len, left);
 721
 722                         ret = pipe_buf_confirm(pipe, buf);
 723                         if (unlikely(ret)) {
 724                                 if (ret == -ENODATA)
 725                                         ret = 0;
 726                                 goto done;
 727                         }
 728
 729                         bvec_set_page(&array[n], buf->page, this_len,
 730                                       buf->offset);
 731                         left -= this_len;
 732                         n++;
 733                 }
 734
 735                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 736                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 737                 if (ret <= 0)
 738                         break;
 739
 740                 sd.num_spliced += ret;
 741                 sd.total_len -= ret;
 742                 *ppos = sd.pos;
 743
 744                 /* dismiss the fully eaten buffers, adjust the partial one */
 745                 tail = pipe->tail;
 746                 while (ret) {
 747                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 748                         if (ret >= buf->len) {
 749                                 ret -= buf->len;
 750                                 buf->len = 0;
 751                                 pipe_buf_release(pipe, buf);
 752                                 tail++;
 753                                 pipe->tail = tail;
 754                                 if (pipe->files)
 755                                         sd.need_wakeup = true;
 756                         } else {
 757                                 buf->offset += ret;
 758                                 buf->len -= ret;
 759                                 ret = 0;
 760                         }
 761                 }
 762         }
 763 done:
 764         kfree(array);
 765         splice_from_pipe_end(pipe, &sd);
 766
 767         pipe_unlock(pipe);
 768
 769         if (sd.num_spliced)
 770                 ret = sd.num_spliced;
 771
 772         return ret;
 773 }
 774
 775 EXPORT_SYMBOL(iter_file_splice_write);
 776
 777 #ifdef CONFIG_NET
 778 /**
 779  * splice_to_socket - splice data from a pipe to a socket
 780  * @pipe:       pipe to splice from
 781  * @out:        socket to write to
 782  * @ppos:       position in @out
 783  * @len:        number of bytes to splice
 784  * @flags:      splice modifier flags
 785  *
 786  * Description:
 787  *    Will send @len bytes from the pipe to a network socket. No data copying
 788  *    is involved.
 789  *
 790  */
 791 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 792                          loff_t *ppos, size_t len, unsigned int flags)
 793 {
 794         struct socket *sock = sock_from_file(out);
 795         struct bio_vec bvec[16];
 796         struct msghdr msg = {};
 797         ssize_t ret = 0;
 798         size_t spliced = 0;
 799         bool need_wakeup = false;
 800
 801         pipe_lock(pipe);
 802
 803         while (len > 0) {
 804                 unsigned int head, tail, mask, bc = 0;
 805                 size_t remain = len;
 806
 807                 /*
 808                  * Check for signal early to make process killable when there
 809                  * are always buffers available
 810                  */
 811                 ret = -ERESTARTSYS;
 812                 if (signal_pending(current))
 813                         break;
 814
 815                 while (pipe_empty(pipe->head, pipe->tail)) {
 816                         ret = 0;
 817                         if (!pipe->writers)
 818                                 goto out;
 819
 820                         if (spliced)
 821                                 goto out;
 822
 823                         ret = -EAGAIN;
 824                         if (flags & SPLICE_F_NONBLOCK)
 825                                 goto out;
 826
 827                         ret = -ERESTARTSYS;
 828                         if (signal_pending(current))
 829                                 goto out;
 830
 831                         if (need_wakeup) {
 832                                 wakeup_pipe_writers(pipe);
 833                                 need_wakeup = false;
 834                         }
 835
 836                         pipe_wait_readable(pipe);
 837                 }
 838
 839                 head = pipe->head;
 840                 tail = pipe->tail;
 841                 mask = pipe->ring_size - 1;
 842
 843                 while (!pipe_empty(head, tail)) {
 844                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 845                         size_t seg;
 846
 847                         if (!buf->len) {
 848                                 tail++;
 849                                 continue;
 850                         }
 851
 852                         seg = min_t(size_t, remain, buf->len);
 853
 854                         ret = pipe_buf_confirm(pipe, buf);
 855                         if (unlikely(ret)) {
 856                                 if (ret == -ENODATA)
 857                                         ret = 0;
 858                                 break;
 859                         }
 860
 861                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 862                         remain -= seg;
 863                         if (remain == 0 || bc >= ARRAY_SIZE(bvec))
 864                                 break;
 865                         tail++;
 866                 }
 867
 868                 if (!bc)
 869                         break;
 870
 871                 msg.msg_flags = MSG_SPLICE_PAGES;
 872                 if (flags & SPLICE_F_MORE)
 873                         msg.msg_flags |= MSG_MORE;
 874                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 875                         msg.msg_flags |= MSG_MORE;
 876                 if (out->f_flags & O_NONBLOCK)
 877                         msg.msg_flags |= MSG_DONTWAIT;
 878
 879                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 880                               len - remain);
 881                 ret = sock_sendmsg(sock, &msg);
 882                 if (ret <= 0)
 883                         break;
 884
 885                 spliced += ret;
 886                 len -= ret;
 887                 tail = pipe->tail;
 888                 while (ret > 0) {
 889                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 890                         size_t seg = min_t(size_t, ret, buf->len);
 891
 892                         buf->offset += seg;
 893                         buf->len -= seg;
 894                         ret -= seg;
 895
 896                         if (!buf->len) {
 897                                 pipe_buf_release(pipe, buf);
 898                                 tail++;
 899                         }
 900                 }
 901
 902                 if (tail != pipe->tail) {
 903                         pipe->tail = tail;
 904                         if (pipe->files)
 905                                 need_wakeup = true;
 906                 }
 907         }
 908
 909 out:
 910         pipe_unlock(pipe);
 911         if (need_wakeup)
 912                 wakeup_pipe_writers(pipe);
 913         return spliced ?: ret;
 914 }
 915 #endif
 916
 917 static int warn_unsupported(struct file *file, const char *op)
 918 {
 919         pr_debug_ratelimited(
 920                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 921                 op, file, current->pid, current->comm);
 922         return -EINVAL;
 923 }
 924
 925 /*
 926  * Attempt to initiate a splice from pipe to file.
 927  */
 928 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 929                            loff_t *ppos, size_t len, unsigned int flags)
 930 {
 931         if (unlikely(!out->f_op->splice_write))
 932                 return warn_unsupported(out, "write");
 933         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 934 }
 935
 936 /*
 937  * Indicate to the caller that there was a premature EOF when reading from the
 938  * source and the caller didn't indicate they would be sending more data after
 939  * this.
 940  */
 941 static void do_splice_eof(struct splice_desc *sd)
 942 {
 943         if (sd->splice_eof)
 944                 sd->splice_eof(sd);
 945 }
 946
 947 /**
 948  * vfs_splice_read - Read data from a file and splice it into a pipe
 949  * @in:         File to splice from
 950  * @ppos:       Input file offset
 951  * @pipe:       Pipe to splice to
 952  * @len:        Number of bytes to splice
 953  * @flags:      Splice modifier flags (SPLICE_F_*)
 954  *
 955  * Splice the requested amount of data from the input file to the pipe.  This
 956  * is synchronous as the caller must hold the pipe lock across the entire
 957  * operation.
 958  *
 959  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 960  * a hole and a negative error code otherwise.
 961  */
 962 long vfs_splice_read(struct file *in, loff_t *ppos,
 963                      struct pipe_inode_info *pipe, size_t len,
 964                      unsigned int flags)
 965 {
 966         unsigned int p_space;
 967         int ret;
 968
 969         if (unlikely(!(in->f_mode & FMODE_READ)))
 970                 return -EBADF;
 971         if (!len)
 972                 return 0;
 973
 974         /* Don't try to read more the pipe has space for. */
 975         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 976         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 977
 978         ret = rw_verify_area(READ, in, ppos, len);
 979         if (unlikely(ret < 0))
 980                 return ret;
 981
 982         if (unlikely(len > MAX_RW_COUNT))
 983                 len = MAX_RW_COUNT;
 984
 985         if (unlikely(!in->f_op->splice_read))
 986                 return warn_unsupported(in, "read");
 987         /*
 988          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 989          * buffer, copy into it and splice that into the pipe.
 990          */
 991         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 992                 return copy_splice_read(in, ppos, pipe, len, flags);
 993         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 994 }
 995 EXPORT_SYMBOL_GPL(vfs_splice_read);
 996
 997 /**
 998  * splice_direct_to_actor - splices data directly between two non-pipes
 999  * @in:         file to splice from
1000  * @sd:         actor information on where to splice to
1001  * @actor:      handles the data splicing
1002  *
1003  * Description:
1004  *    This is a special case helper to splice directly between two
1005  *    points, without requiring an explicit pipe. Internally an allocated
1006  *    pipe is cached in the process, and reused during the lifetime of
1007  *    that process.
1008  *
1009  */
1010 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1011                                splice_direct_actor *actor)
1012 {
1013         struct pipe_inode_info *pipe;
1014         long ret, bytes;
1015         size_t len;
1016         int i, flags, more;
1017
1018         /*
1019          * We require the input to be seekable, as we don't want to randomly
1020          * drop data for eg socket -> socket splicing. Use the piped splicing
1021          * for that!
1022          */
1023         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1024                 return -EINVAL;
1025
1026         /*
1027          * neither in nor out is a pipe, setup an internal pipe attached to
1028          * 'out' and transfer the wanted data from 'in' to 'out' through that
1029          */
1030         pipe = current->splice_pipe;
1031         if (unlikely(!pipe)) {
1032                 pipe = alloc_pipe_info();
1033                 if (!pipe)
1034                         return -ENOMEM;
1035
1036                 /*
1037                  * We don't have an immediate reader, but we'll read the stuff
1038                  * out of the pipe right after the splice_to_pipe(). So set
1039                  * PIPE_READERS appropriately.
1040                  */
1041                 pipe->readers = 1;
1042
1043                 current->splice_pipe = pipe;
1044         }
1045
1046         /*
1047          * Do the splice.
1048          */
1049         bytes = 0;
1050         len = sd->total_len;
1051
1052         /* Don't block on output, we have to drain the direct pipe. */
1053         flags = sd->flags;
1054         sd->flags &= ~SPLICE_F_NONBLOCK;
1055
1056         /*
1057          * We signal MORE until we've read sufficient data to fulfill the
1058          * request and we keep signalling it if the caller set it.
1059          */
1060         more = sd->flags & SPLICE_F_MORE;
1061         sd->flags |= SPLICE_F_MORE;
1062
1063         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1064
1065         while (len) {
1066                 size_t read_len;
1067                 loff_t pos = sd->pos, prev_pos = pos;
1068
1069                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1070                 if (unlikely(ret <= 0))
1071                         goto read_failure;
1072
1073                 read_len = ret;
1074                 sd->total_len = read_len;
1075
1076                 /*
1077                  * If we now have sufficient data to fulfill the request then
1078                  * we clear SPLICE_F_MORE if it was not set initially.
1079                  */
1080                 if (read_len >= len && !more)
1081                         sd->flags &= ~SPLICE_F_MORE;
1082
1083                 /*
1084                  * NOTE: nonblocking mode only applies to the input. We
1085                  * must not do the output in nonblocking mode as then we
1086                  * could get stuck data in the internal pipe:
1087                  */
1088                 ret = actor(pipe, sd);
1089                 if (unlikely(ret <= 0)) {
1090                         sd->pos = prev_pos;
1091                         goto out_release;
1092                 }
1093
1094                 bytes += ret;
1095                 len -= ret;
1096                 sd->pos = pos;
1097
1098                 if (ret < read_len) {
1099                         sd->pos = prev_pos + ret;
1100                         goto out_release;
1101                 }
1102         }
1103
1104 done:
1105         pipe->tail = pipe->head = 0;
1106         file_accessed(in);
1107         return bytes;
1108
1109 read_failure:
1110         /*
1111          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1112          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1113          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1114          * least 1 byte *then* we will also do the ->splice_eof() call.
1115          */
1116         if (ret == 0 && !more && len > 0 && bytes)
1117                 do_splice_eof(sd);
1118 out_release:
1119         /*
1120          * If we did an incomplete transfer we must release
1121          * the pipe buffers in question:
1122          */
1123         for (i = 0; i < pipe->ring_size; i++) {
1124                 struct pipe_buffer *buf = &pipe->bufs[i];
1125
1126                 if (buf->ops)
1127                         pipe_buf_release(pipe, buf);
1128         }
1129
1130         if (!bytes)
1131                 bytes = ret;
1132
1133         goto done;
1134 }
1135 EXPORT_SYMBOL(splice_direct_to_actor);
1136
1137 static int direct_splice_actor(struct pipe_inode_info *pipe,
1138                                struct splice_desc *sd)
1139 {
1140         struct file *file = sd->u.file;
1141
1142         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1143                               sd->flags);
1144 }
1145
1146 static void direct_file_splice_eof(struct splice_desc *sd)
1147 {
1148         struct file *file = sd->u.file;
1149
1150         if (file->f_op->splice_eof)
1151                 file->f_op->splice_eof(file);
1152 }
1153
1154 /**
1155  * do_splice_direct - splices data directly between two files
1156  * @in:         file to splice from
1157  * @ppos:       input file offset
1158  * @out:        file to splice to
1159  * @opos:       output file offset
1160  * @len:        number of bytes to splice
1161  * @flags:      splice modifier flags
1162  *
1163  * Description:
1164  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1165  *    doing it in the application would incur an extra system call
1166  *    (splice in + splice out, as compared to just sendfile()). So this helper
1167  *    can splice directly through a process-private pipe.
1168  *
1169  */
1170 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1171                       loff_t *opos, size_t len, unsigned int flags)
1172 {
1173         struct splice_desc sd = {
1174                 .len            = len,
1175                 .total_len      = len,
1176                 .flags          = flags,
1177                 .pos            = *ppos,
1178                 .u.file         = out,
1179                 .splice_eof     = direct_file_splice_eof,
1180                 .opos           = opos,
1181         };
1182         long ret;
1183
1184         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1185                 return -EBADF;
1186
1187         if (unlikely(out->f_flags & O_APPEND))
1188                 return -EINVAL;
1189
1190         ret = rw_verify_area(WRITE, out, opos, len);
1191         if (unlikely(ret < 0))
1192                 return ret;
1193
1194         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1195         if (ret > 0)
1196                 *ppos = sd.pos;
1197
1198         return ret;
1199 }
1200 EXPORT_SYMBOL(do_splice_direct);
1201
1202 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1203 {
1204         for (;;) {
1205                 if (unlikely(!pipe->readers)) {
1206                         send_sig(SIGPIPE, current, 0);
1207                         return -EPIPE;
1208                 }
1209                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1210                         return 0;
1211                 if (flags & SPLICE_F_NONBLOCK)
1212                         return -EAGAIN;
1213                 if (signal_pending(current))
1214                         return -ERESTARTSYS;
1215                 pipe_wait_writable(pipe);
1216         }
1217 }
1218
1219 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1220                                struct pipe_inode_info *opipe,
1221                                size_t len, unsigned int flags);
1222
1223 long splice_file_to_pipe(struct file *in,
1224                          struct pipe_inode_info *opipe,
1225                          loff_t *offset,
1226                          size_t len, unsigned int flags)
1227 {
1228         long ret;
1229
1230         pipe_lock(opipe);
1231         ret = wait_for_space(opipe, flags);
1232         if (!ret)
1233                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1234         pipe_unlock(opipe);
1235         if (ret > 0)
1236                 wakeup_pipe_readers(opipe);
1237         return ret;
1238 }
1239
1240 /*
1241  * Determine where to splice to/from.
1242  */
1243 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1244                loff_t *off_out, size_t len, unsigned int flags)
1245 {
1246         struct pipe_inode_info *ipipe;
1247         struct pipe_inode_info *opipe;
1248         loff_t offset;
1249         long ret;
1250
1251         if (unlikely(!(in->f_mode & FMODE_READ) ||
1252                      !(out->f_mode & FMODE_WRITE)))
1253                 return -EBADF;
1254
1255         ipipe = get_pipe_info(in, true);
1256         opipe = get_pipe_info(out, true);
1257
1258         if (ipipe && opipe) {
1259                 if (off_in || off_out)
1260                         return -ESPIPE;
1261
1262                 /* Splicing to self would be fun, but... */
1263                 if (ipipe == opipe)
1264                         return -EINVAL;
1265
1266                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1267                         flags |= SPLICE_F_NONBLOCK;
1268
1269                 ret = splice_pipe_to_pipe(ipipe, opipe, len, flags);
1270         } else if (ipipe) {
1271                 if (off_in)
1272                         return -ESPIPE;
1273                 if (off_out) {
1274                         if (!(out->f_mode & FMODE_PWRITE))
1275                                 return -EINVAL;
1276                         offset = *off_out;
1277                 } else {
1278                         offset = out->f_pos;
1279                 }
1280
1281                 if (unlikely(out->f_flags & O_APPEND))
1282                         return -EINVAL;
1283
1284                 ret = rw_verify_area(WRITE, out, &offset, len);
1285                 if (unlikely(ret < 0))
1286                         return ret;
1287
1288                 if (in->f_flags & O_NONBLOCK)
1289                         flags |= SPLICE_F_NONBLOCK;
1290
1291                 file_start_write(out);
1292                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1293                 file_end_write(out);
1294
1295                 if (!off_out)
1296                         out->f_pos = offset;
1297                 else
1298                         *off_out = offset;
1299         } else if (opipe) {
1300                 if (off_out)
1301                         return -ESPIPE;
1302                 if (off_in) {
1303                         if (!(in->f_mode & FMODE_PREAD))
1304                                 return -EINVAL;
1305                         offset = *off_in;
1306                 } else {
1307                         offset = in->f_pos;
1308                 }
1309
1310                 if (out->f_flags & O_NONBLOCK)
1311                         flags |= SPLICE_F_NONBLOCK;
1312
1313                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1314
1315                 if (!off_in)
1316                         in->f_pos = offset;
1317                 else
1318                         *off_in = offset;
1319         } else {
1320                 ret = -EINVAL;
1321         }
1322
1323         if (ret > 0) {
1324                 /*
1325                  * Generate modify out before access in:
1326                  * do_splice_from() may've already sent modify out,
1327                  * and this ensures the events get merged.
1328                  */
1329                 fsnotify_modify(out);
1330                 fsnotify_access(in);
1331         }
1332
1333         return ret;
1334 }
1335
1336 static long __do_splice(struct file *in, loff_t __user *off_in,
1337                         struct file *out, loff_t __user *off_out,
1338                         size_t len, unsigned int flags)
1339 {
1340         struct pipe_inode_info *ipipe;
1341         struct pipe_inode_info *opipe;
1342         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1343         long ret;
1344
1345         ipipe = get_pipe_info(in, true);
1346         opipe = get_pipe_info(out, true);
1347
1348         if (ipipe) {
1349                 if (off_in)
1350                         return -ESPIPE;
1351                 pipe_clear_nowait(in);
1352         }
1353         if (opipe) {
1354                 if (off_out)
1355                         return -ESPIPE;
1356                 pipe_clear_nowait(out);
1357         }
1358
1359         if (off_out) {
1360                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1361                         return -EFAULT;
1362                 __off_out = &offset;
1363         }
1364         if (off_in) {
1365                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1366                         return -EFAULT;
1367                 __off_in = &offset;
1368         }
1369
1370         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1371         if (ret < 0)
1372                 return ret;
1373
1374         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1375                 return -EFAULT;
1376         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1377                 return -EFAULT;
1378
1379         return ret;
1380 }
1381
1382 static int iter_to_pipe(struct iov_iter *from,
1383                         struct pipe_inode_info *pipe,
1384                         unsigned flags)
1385 {
1386         struct pipe_buffer buf = {
1387                 .ops = &user_page_pipe_buf_ops,
1388                 .flags = flags
1389         };
1390         size_t total = 0;
1391         int ret = 0;
1392
1393         while (iov_iter_count(from)) {
1394                 struct page *pages[16];
1395                 ssize_t left;
1396                 size_t start;
1397                 int i, n;
1398
1399                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1400                 if (left <= 0) {
1401                         ret = left;
1402                         break;
1403                 }
1404
1405                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1406                 for (i = 0; i < n; i++) {
1407                         int size = min_t(int, left, PAGE_SIZE - start);
1408
1409                         buf.page = pages[i];
1410                         buf.offset = start;
1411                         buf.len = size;
1412                         ret = add_to_pipe(pipe, &buf);
1413                         if (unlikely(ret < 0)) {
1414                                 iov_iter_revert(from, left);
1415                                 // this one got dropped by add_to_pipe()
1416                                 while (++i < n)
1417                                         put_page(pages[i]);
1418                                 goto out;
1419                         }
1420                         total += ret;
1421                         left -= size;
1422                         start = 0;
1423                 }
1424         }
1425 out:
1426         return total ? total : ret;
1427 }
1428
1429 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1430                         struct splice_desc *sd)
1431 {
1432         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1433         return n == sd->len ? n : -EFAULT;
1434 }
1435
1436 /*
1437  * For lack of a better implementation, implement vmsplice() to userspace
1438  * as a simple copy of the pipes pages to the user iov.
1439  */
1440 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1441                              unsigned int flags)
1442 {
1443         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1444         struct splice_desc sd = {
1445                 .total_len = iov_iter_count(iter),
1446                 .flags = flags,
1447                 .u.data = iter
1448         };
1449         long ret = 0;
1450
1451         if (!pipe)
1452                 return -EBADF;
1453
1454         pipe_clear_nowait(file);
1455
1456         if (sd.total_len) {
1457                 pipe_lock(pipe);
1458                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1459                 pipe_unlock(pipe);
1460         }
1461
1462         if (ret > 0)
1463                 fsnotify_access(file);
1464
1465         return ret;
1466 }
1467
1468 /*
1469  * vmsplice splices a user address range into a pipe. It can be thought of
1470  * as splice-from-memory, where the regular splice is splice-from-file (or
1471  * to file). In both cases the output is a pipe, naturally.
1472  */
1473 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1474                              unsigned int flags)
1475 {
1476         struct pipe_inode_info *pipe;
1477         long ret = 0;
1478         unsigned buf_flag = 0;
1479
1480         if (flags & SPLICE_F_GIFT)
1481                 buf_flag = PIPE_BUF_FLAG_GIFT;
1482
1483         pipe = get_pipe_info(file, true);
1484         if (!pipe)
1485                 return -EBADF;
1486
1487         pipe_clear_nowait(file);
1488
1489         pipe_lock(pipe);
1490         ret = wait_for_space(pipe, flags);
1491         if (!ret)
1492                 ret = iter_to_pipe(iter, pipe, buf_flag);
1493         pipe_unlock(pipe);
1494         if (ret > 0) {
1495                 wakeup_pipe_readers(pipe);
1496                 fsnotify_modify(file);
1497         }
1498         return ret;
1499 }
1500
1501 static int vmsplice_type(struct fd f, int *type)
1502 {
1503         if (!f.file)
1504                 return -EBADF;
1505         if (f.file->f_mode & FMODE_WRITE) {
1506                 *type = ITER_SOURCE;
1507         } else if (f.file->f_mode & FMODE_READ) {
1508                 *type = ITER_DEST;
1509         } else {
1510                 fdput(f);
1511                 return -EBADF;
1512         }
1513         return 0;
1514 }
1515
1516 /*
1517  * Note that vmsplice only really supports true splicing _from_ user memory
1518  * to a pipe, not the other way around. Splicing from user memory is a simple
1519  * operation that can be supported without any funky alignment restrictions
1520  * or nasty vm tricks. We simply map in the user memory and fill them into
1521  * a pipe. The reverse isn't quite as easy, though. There are two possible
1522  * solutions for that:
1523  *
1524  *      - memcpy() the data internally, at which point we might as well just
1525  *        do a regular read() on the buffer anyway.
1526  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1527  *        has restriction limitations on both ends of the pipe).
1528  *
1529  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1530  *
1531  */
1532 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1533                 unsigned long, nr_segs, unsigned int, flags)
1534 {
1535         struct iovec iovstack[UIO_FASTIOV];
1536         struct iovec *iov = iovstack;
1537         struct iov_iter iter;
1538         ssize_t error;
1539         struct fd f;
1540         int type;
1541
1542         if (unlikely(flags & ~SPLICE_F_ALL))
1543                 return -EINVAL;
1544
1545         f = fdget(fd);
1546         error = vmsplice_type(f, &type);
1547         if (error)
1548                 return error;
1549
1550         error = import_iovec(type, uiov, nr_segs,
1551                              ARRAY_SIZE(iovstack), &iov, &iter);
1552         if (error < 0)
1553                 goto out_fdput;
1554
1555         if (!iov_iter_count(&iter))
1556                 error = 0;
1557         else if (type == ITER_SOURCE)
1558                 error = vmsplice_to_pipe(f.file, &iter, flags);
1559         else
1560                 error = vmsplice_to_user(f.file, &iter, flags);
1561
1562         kfree(iov);
1563 out_fdput:
1564         fdput(f);
1565         return error;
1566 }
1567
1568 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1569                 int, fd_out, loff_t __user *, off_out,
1570                 size_t, len, unsigned int, flags)
1571 {
1572         struct fd in, out;
1573         long error;
1574
1575         if (unlikely(!len))
1576                 return 0;
1577
1578         if (unlikely(flags & ~SPLICE_F_ALL))
1579                 return -EINVAL;
1580
1581         error = -EBADF;
1582         in = fdget(fd_in);
1583         if (in.file) {
1584                 out = fdget(fd_out);
1585                 if (out.file) {
1586                         error = __do_splice(in.file, off_in, out.file, off_out,
1587                                                 len, flags);
1588                         fdput(out);
1589                 }
1590                 fdput(in);
1591         }
1592         return error;
1593 }
1594
1595 /*
1596  * Make sure there's data to read. Wait for input if we can, otherwise
1597  * return an appropriate error.
1598  */
1599 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1600 {
1601         int ret;
1602
1603         /*
1604          * Check the pipe occupancy without the inode lock first. This function
1605          * is speculative anyways, so missing one is ok.
1606          */
1607         if (!pipe_empty(pipe->head, pipe->tail))
1608                 return 0;
1609
1610         ret = 0;
1611         pipe_lock(pipe);
1612
1613         while (pipe_empty(pipe->head, pipe->tail)) {
1614                 if (signal_pending(current)) {
1615                         ret = -ERESTARTSYS;
1616                         break;
1617                 }
1618                 if (!pipe->writers)
1619                         break;
1620                 if (flags & SPLICE_F_NONBLOCK) {
1621                         ret = -EAGAIN;
1622                         break;
1623                 }
1624                 pipe_wait_readable(pipe);
1625         }
1626
1627         pipe_unlock(pipe);
1628         return ret;
1629 }
1630
1631 /*
1632  * Make sure there's writeable room. Wait for room if we can, otherwise
1633  * return an appropriate error.
1634  */
1635 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1636 {
1637         int ret;
1638
1639         /*
1640          * Check pipe occupancy without the inode lock first. This function
1641          * is speculative anyways, so missing one is ok.
1642          */
1643         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1644                 return 0;
1645
1646         ret = 0;
1647         pipe_lock(pipe);
1648
1649         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1650                 if (!pipe->readers) {
1651                         send_sig(SIGPIPE, current, 0);
1652                         ret = -EPIPE;
1653                         break;
1654                 }
1655                 if (flags & SPLICE_F_NONBLOCK) {
1656                         ret = -EAGAIN;
1657                         break;
1658                 }
1659                 if (signal_pending(current)) {
1660                         ret = -ERESTARTSYS;
1661                         break;
1662                 }
1663                 pipe_wait_writable(pipe);
1664         }
1665
1666         pipe_unlock(pipe);
1667         return ret;
1668 }
1669
1670 /*
1671  * Splice contents of ipipe to opipe.
1672  */
1673 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1674                                struct pipe_inode_info *opipe,
1675                                size_t len, unsigned int flags)
1676 {
1677         struct pipe_buffer *ibuf, *obuf;
1678         unsigned int i_head, o_head;
1679         unsigned int i_tail, o_tail;
1680         unsigned int i_mask, o_mask;
1681         int ret = 0;
1682         bool input_wakeup = false;
1683
1684
1685 retry:
1686         ret = ipipe_prep(ipipe, flags);
1687         if (ret)
1688                 return ret;
1689
1690         ret = opipe_prep(opipe, flags);
1691         if (ret)
1692                 return ret;
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                 size_t o_len;
1708
1709                 if (!opipe->readers) {
1710                         send_sig(SIGPIPE, current, 0);
1711                         if (!ret)
1712                                 ret = -EPIPE;
1713                         break;
1714                 }
1715
1716                 i_head = ipipe->head;
1717                 o_tail = opipe->tail;
1718
1719                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1720                         break;
1721
1722                 /*
1723                  * Cannot make any progress, because either the input
1724                  * pipe is empty or the output pipe is full.
1725                  */
1726                 if (pipe_empty(i_head, i_tail) ||
1727                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1728                         /* Already processed some buffers, break */
1729                         if (ret)
1730                                 break;
1731
1732                         if (flags & SPLICE_F_NONBLOCK) {
1733                                 ret = -EAGAIN;
1734                                 break;
1735                         }
1736
1737                         /*
1738                          * We raced with another reader/writer and haven't
1739                          * managed to process any buffers.  A zero return
1740                          * value means EOF, so retry instead.
1741                          */
1742                         pipe_unlock(ipipe);
1743                         pipe_unlock(opipe);
1744                         goto retry;
1745                 }
1746
1747                 ibuf = &ipipe->bufs[i_tail & i_mask];
1748                 obuf = &opipe->bufs[o_head & o_mask];
1749
1750                 if (len >= ibuf->len) {
1751                         /*
1752                          * Simply move the whole buffer from ipipe to opipe
1753                          */
1754                         *obuf = *ibuf;
1755                         ibuf->ops = NULL;
1756                         i_tail++;
1757                         ipipe->tail = i_tail;
1758                         input_wakeup = true;
1759                         o_len = obuf->len;
1760                         o_head++;
1761                         opipe->head = o_head;
1762                 } else {
1763                         /*
1764                          * Get a reference to this pipe buffer,
1765                          * so we can copy the contents over.
1766                          */
1767                         if (!pipe_buf_get(ipipe, ibuf)) {
1768                                 if (ret == 0)
1769                                         ret = -EFAULT;
1770                                 break;
1771                         }
1772                         *obuf = *ibuf;
1773
1774                         /*
1775                          * Don't inherit the gift and merge flags, we need to
1776                          * prevent multiple steals of this page.
1777                          */
1778                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1779                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1780
1781                         obuf->len = len;
1782                         ibuf->offset += len;
1783                         ibuf->len -= len;
1784                         o_len = len;
1785                         o_head++;
1786                         opipe->head = o_head;
1787                 }
1788                 ret += o_len;
1789                 len -= o_len;
1790         } while (len);
1791
1792         pipe_unlock(ipipe);
1793         pipe_unlock(opipe);
1794
1795         /*
1796          * If we put data in the output pipe, wakeup any potential readers.
1797          */
1798         if (ret > 0)
1799                 wakeup_pipe_readers(opipe);
1800
1801         if (input_wakeup)
1802                 wakeup_pipe_writers(ipipe);
1803
1804         return ret;
1805 }
1806
1807 /*
1808  * Link contents of ipipe to opipe.
1809  */
1810 static int link_pipe(struct pipe_inode_info *ipipe,
1811                      struct pipe_inode_info *opipe,
1812                      size_t len, unsigned int flags)
1813 {
1814         struct pipe_buffer *ibuf, *obuf;
1815         unsigned int i_head, o_head;
1816         unsigned int i_tail, o_tail;
1817         unsigned int i_mask, o_mask;
1818         int ret = 0;
1819
1820         /*
1821          * Potential ABBA deadlock, work around it by ordering lock
1822          * grabbing by pipe info address. Otherwise two different processes
1823          * could deadlock (one doing tee from A -> B, the other from B -> A).
1824          */
1825         pipe_double_lock(ipipe, opipe);
1826
1827         i_tail = ipipe->tail;
1828         i_mask = ipipe->ring_size - 1;
1829         o_head = opipe->head;
1830         o_mask = opipe->ring_size - 1;
1831
1832         do {
1833                 if (!opipe->readers) {
1834                         send_sig(SIGPIPE, current, 0);
1835                         if (!ret)
1836                                 ret = -EPIPE;
1837                         break;
1838                 }
1839
1840                 i_head = ipipe->head;
1841                 o_tail = opipe->tail;
1842
1843                 /*
1844                  * If we have iterated all input buffers or run out of
1845                  * output room, break.
1846                  */
1847                 if (pipe_empty(i_head, i_tail) ||
1848                     pipe_full(o_head, o_tail, opipe->max_usage))
1849                         break;
1850
1851                 ibuf = &ipipe->bufs[i_tail & i_mask];
1852                 obuf = &opipe->bufs[o_head & o_mask];
1853
1854                 /*
1855                  * Get a reference to this pipe buffer,
1856                  * so we can copy the contents over.
1857                  */
1858                 if (!pipe_buf_get(ipipe, ibuf)) {
1859                         if (ret == 0)
1860                                 ret = -EFAULT;
1861                         break;
1862                 }
1863
1864                 *obuf = *ibuf;
1865
1866                 /*
1867                  * Don't inherit the gift and merge flag, we need to prevent
1868                  * multiple steals of this page.
1869                  */
1870                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1871                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1872
1873                 if (obuf->len > len)
1874                         obuf->len = len;
1875                 ret += obuf->len;
1876                 len -= obuf->len;
1877
1878                 o_head++;
1879                 opipe->head = o_head;
1880                 i_tail++;
1881         } while (len);
1882
1883         pipe_unlock(ipipe);
1884         pipe_unlock(opipe);
1885
1886         /*
1887          * If we put data in the output pipe, wakeup any potential readers.
1888          */
1889         if (ret > 0)
1890                 wakeup_pipe_readers(opipe);
1891
1892         return ret;
1893 }
1894
1895 /*
1896  * This is a tee(1) implementation that works on pipes. It doesn't copy
1897  * any data, it simply references the 'in' pages on the 'out' pipe.
1898  * The 'flags' used are the SPLICE_F_* variants, currently the only
1899  * applicable one is SPLICE_F_NONBLOCK.
1900  */
1901 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1902 {
1903         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1904         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1905         int ret = -EINVAL;
1906
1907         if (unlikely(!(in->f_mode & FMODE_READ) ||
1908                      !(out->f_mode & FMODE_WRITE)))
1909                 return -EBADF;
1910
1911         /*
1912          * Duplicate the contents of ipipe to opipe without actually
1913          * copying the data.
1914          */
1915         if (ipipe && opipe && ipipe != opipe) {
1916                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1917                         flags |= SPLICE_F_NONBLOCK;
1918
1919                 /*
1920                  * Keep going, unless we encounter an error. The ipipe/opipe
1921                  * ordering doesn't really matter.
1922                  */
1923                 ret = ipipe_prep(ipipe, flags);
1924                 if (!ret) {
1925                         ret = opipe_prep(opipe, flags);
1926                         if (!ret)
1927                                 ret = link_pipe(ipipe, opipe, len, flags);
1928                 }
1929         }
1930
1931         if (ret > 0) {
1932                 fsnotify_access(in);
1933                 fsnotify_modify(out);
1934         }
1935
1936         return ret;
1937 }
1938
1939 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1940 {
1941         struct fd in, out;
1942         int error;
1943
1944         if (unlikely(flags & ~SPLICE_F_ALL))
1945                 return -EINVAL;
1946
1947         if (unlikely(!len))
1948                 return 0;
1949
1950         error = -EBADF;
1951         in = fdget(fdin);
1952         if (in.file) {
1953                 out = fdget(fdout);
1954                 if (out.file) {
1955                         error = do_tee(in.file, out.file, len, flags);
1956                         fdput(out);
1957                 }
1958                 fdput(in);
1959         }
1960
1961         return error;
1962 }