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 (folio_has_private(folio) &&
  87                     !filemap_release_folio(folio, GFP_KERNEL))
  88                         goto out_unlock;
  89
  90                 /*
  91                  * If we succeeded in removing the mapping, set LRU flag
  92                  * and return good.
  93                  */
  94                 if (remove_mapping(mapping, folio)) {
  95                         buf->flags |= PIPE_BUF_FLAG_LRU;
  96                         return true;
  97                 }
  98         }
  99
 100         /*
 101          * Raced with truncate or failed to remove folio from current
 102          * address space, unlock and return failure.
 103          */
 104 out_unlock:
 105         folio_unlock(folio);
 106         return false;
 107 }
 108
 109 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 110                                         struct pipe_buffer *buf)
 111 {
 112         put_page(buf->page);
 113         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 114 }
 115
 116 /*
 117  * Check whether the contents of buf is OK to access. Since the content
 118  * is a page cache page, IO may be in flight.
 119  */
 120 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 121                                        struct pipe_buffer *buf)
 122 {
 123         struct folio *folio = page_folio(buf->page);
 124         int err;
 125
 126         if (!folio_test_uptodate(folio)) {
 127                 folio_lock(folio);
 128
 129                 /*
 130                  * Folio got truncated/unhashed. This will cause a 0-byte
 131                  * splice, if this is the first page.
 132                  */
 133                 if (!folio->mapping) {
 134                         err = -ENODATA;
 135                         goto error;
 136                 }
 137
 138                 /*
 139                  * Uh oh, read-error from disk.
 140                  */
 141                 if (!folio_test_uptodate(folio)) {
 142                         err = -EIO;
 143                         goto error;
 144                 }
 145
 146                 /* Folio is ok after all, we are done */
 147                 folio_unlock(folio);
 148         }
 149
 150         return 0;
 151 error:
 152         folio_unlock(folio);
 153         return err;
 154 }
 155
 156 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 157         .confirm        = page_cache_pipe_buf_confirm,
 158         .release        = page_cache_pipe_buf_release,
 159         .try_steal      = page_cache_pipe_buf_try_steal,
 160         .get            = generic_pipe_buf_get,
 161 };
 162
 163 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 164                 struct pipe_buffer *buf)
 165 {
 166         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 167                 return false;
 168
 169         buf->flags |= PIPE_BUF_FLAG_LRU;
 170         return generic_pipe_buf_try_steal(pipe, buf);
 171 }
 172
 173 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 174         .release        = page_cache_pipe_buf_release,
 175         .try_steal      = user_page_pipe_buf_try_steal,
 176         .get            = generic_pipe_buf_get,
 177 };
 178
 179 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 180 {
 181         smp_mb();
 182         if (waitqueue_active(&pipe->rd_wait))
 183                 wake_up_interruptible(&pipe->rd_wait);
 184         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 185 }
 186
 187 /**
 188  * splice_to_pipe - fill passed data into a pipe
 189  * @pipe:       pipe to fill
 190  * @spd:        data to fill
 191  *
 192  * Description:
 193  *    @spd contains a map of pages and len/offset tuples, along with
 194  *    the struct pipe_buf_operations associated with these pages. This
 195  *    function will link that data to the pipe.
 196  *
 197  */
 198 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 199                        struct splice_pipe_desc *spd)
 200 {
 201         unsigned int spd_pages = spd->nr_pages;
 202         unsigned int tail = pipe->tail;
 203         unsigned int head = pipe->head;
 204         unsigned int mask = pipe->ring_size - 1;
 205         int ret = 0, page_nr = 0;
 206
 207         if (!spd_pages)
 208                 return 0;
 209
 210         if (unlikely(!pipe->readers)) {
 211                 send_sig(SIGPIPE, current, 0);
 212                 ret = -EPIPE;
 213                 goto out;
 214         }
 215
 216         while (!pipe_full(head, tail, pipe->max_usage)) {
 217                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 218
 219                 buf->page = spd->pages[page_nr];
 220                 buf->offset = spd->partial[page_nr].offset;
 221                 buf->len = spd->partial[page_nr].len;
 222                 buf->private = spd->partial[page_nr].private;
 223                 buf->ops = spd->ops;
 224                 buf->flags = 0;
 225
 226                 head++;
 227                 pipe->head = head;
 228                 page_nr++;
 229                 ret += buf->len;
 230
 231                 if (!--spd->nr_pages)
 232                         break;
 233         }
 234
 235         if (!ret)
 236                 ret = -EAGAIN;
 237
 238 out:
 239         while (page_nr < spd_pages)
 240                 spd->spd_release(spd, page_nr++);
 241
 242         return ret;
 243 }
 244 EXPORT_SYMBOL_GPL(splice_to_pipe);
 245
 246 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 247 {
 248         unsigned int head = pipe->head;
 249         unsigned int tail = pipe->tail;
 250         unsigned int mask = pipe->ring_size - 1;
 251         int ret;
 252
 253         if (unlikely(!pipe->readers)) {
 254                 send_sig(SIGPIPE, current, 0);
 255                 ret = -EPIPE;
 256         } else if (pipe_full(head, tail, pipe->max_usage)) {
 257                 ret = -EAGAIN;
 258         } else {
 259                 pipe->bufs[head & mask] = *buf;
 260                 pipe->head = head + 1;
 261                 return buf->len;
 262         }
 263         pipe_buf_release(pipe, buf);
 264         return ret;
 265 }
 266 EXPORT_SYMBOL(add_to_pipe);
 267
 268 /*
 269  * Check if we need to grow the arrays holding pages and partial page
 270  * descriptions.
 271  */
 272 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 273 {
 274         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 275
 276         spd->nr_pages_max = max_usage;
 277         if (max_usage <= PIPE_DEF_BUFFERS)
 278                 return 0;
 279
 280         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 281         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 282                                      GFP_KERNEL);
 283
 284         if (spd->pages && spd->partial)
 285                 return 0;
 286
 287         kfree(spd->pages);
 288         kfree(spd->partial);
 289         return -ENOMEM;
 290 }
 291
 292 void splice_shrink_spd(struct splice_pipe_desc *spd)
 293 {
 294         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 295                 return;
 296
 297         kfree(spd->pages);
 298         kfree(spd->partial);
 299 }
 300
 301 /**
 302  * copy_splice_read -  Copy data from a file and splice the copy into a pipe
 303  * @in: The file to read from
 304  * @ppos: Pointer to the file position to read from
 305  * @pipe: The pipe to splice into
 306  * @len: The amount to splice
 307  * @flags: The SPLICE_F_* flags
 308  *
 309  * This function allocates a bunch of pages sufficient to hold the requested
 310  * amount of data (but limited by the remaining pipe capacity), passes it to
 311  * the file's ->read_iter() to read into and then splices the used pages into
 312  * the pipe.
 313  *
 314  * Return: On success, the number of bytes read will be returned and *@ppos
 315  * will be updated if appropriate; 0 will be returned if there is no more data
 316  * to be read; -EAGAIN will be returned if the pipe had no space, and some
 317  * other negative error code will be returned on error.  A short read may occur
 318  * if the pipe has insufficient space, we reach the end of the data or we hit a
 319  * hole.
 320  */
 321 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 322                          struct pipe_inode_info *pipe,
 323                          size_t len, unsigned int flags)
 324 {
 325         struct iov_iter to;
 326         struct bio_vec *bv;
 327         struct kiocb kiocb;
 328         struct page **pages;
 329         ssize_t ret;
 330         size_t used, npages, chunk, remain, keep = 0;
 331         int i;
 332
 333         /* Work out how much data we can actually add into the pipe */
 334         used = pipe_occupancy(pipe->head, pipe->tail);
 335         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 336         len = min_t(size_t, len, npages * PAGE_SIZE);
 337         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 338
 339         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 340                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 341         if (!bv)
 342                 return -ENOMEM;
 343
 344         pages = (struct page **)(bv + npages);
 345         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 346         if (!npages) {
 347                 kfree(bv);
 348                 return -ENOMEM;
 349         }
 350
 351         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 352
 353         for (i = 0; i < npages; i++) {
 354                 chunk = min_t(size_t, PAGE_SIZE, remain);
 355                 bv[i].bv_page = pages[i];
 356                 bv[i].bv_offset = 0;
 357                 bv[i].bv_len = chunk;
 358                 remain -= chunk;
 359         }
 360
 361         /* Do the I/O */
 362         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 363         init_sync_kiocb(&kiocb, in);
 364         kiocb.ki_pos = *ppos;
 365         ret = call_read_iter(in, &kiocb, &to);
 366
 367         if (ret > 0) {
 368                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 369                 *ppos = kiocb.ki_pos;
 370         }
 371
 372         /*
 373          * Callers of ->splice_read() expect -EAGAIN on "can't put anything in
 374          * there", rather than -EFAULT.
 375          */
 376         if (ret == -EFAULT)
 377                 ret = -EAGAIN;
 378
 379         /* Free any pages that didn't get touched at all. */
 380         if (keep < npages)
 381                 release_pages(pages + keep, npages - keep);
 382
 383         /* Push the remaining pages into the pipe. */
 384         remain = ret;
 385         for (i = 0; i < keep; i++) {
 386                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 387
 388                 chunk = min_t(size_t, remain, PAGE_SIZE);
 389                 *buf = (struct pipe_buffer) {
 390                         .ops    = &default_pipe_buf_ops,
 391                         .page   = bv[i].bv_page,
 392                         .offset = 0,
 393                         .len    = chunk,
 394                 };
 395                 pipe->head++;
 396                 remain -= chunk;
 397         }
 398
 399         kfree(bv);
 400         return ret;
 401 }
 402 EXPORT_SYMBOL(copy_splice_read);
 403
 404 const struct pipe_buf_operations default_pipe_buf_ops = {
 405         .release        = generic_pipe_buf_release,
 406         .try_steal      = generic_pipe_buf_try_steal,
 407         .get            = generic_pipe_buf_get,
 408 };
 409
 410 /* Pipe buffer operations for a socket and similar. */
 411 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 412         .release        = generic_pipe_buf_release,
 413         .get            = generic_pipe_buf_get,
 414 };
 415 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 416
 417 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 418 {
 419         smp_mb();
 420         if (waitqueue_active(&pipe->wr_wait))
 421                 wake_up_interruptible(&pipe->wr_wait);
 422         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 423 }
 424
 425 /**
 426  * splice_from_pipe_feed - feed available data from a pipe to a file
 427  * @pipe:       pipe to splice from
 428  * @sd:         information to @actor
 429  * @actor:      handler that splices the data
 430  *
 431  * Description:
 432  *    This function loops over the pipe and calls @actor to do the
 433  *    actual moving of a single struct pipe_buffer to the desired
 434  *    destination.  It returns when there's no more buffers left in
 435  *    the pipe or if the requested number of bytes (@sd->total_len)
 436  *    have been copied.  It returns a positive number (one) if the
 437  *    pipe needs to be filled with more data, zero if the required
 438  *    number of bytes have been copied and -errno on error.
 439  *
 440  *    This, together with splice_from_pipe_{begin,end,next}, may be
 441  *    used to implement the functionality of __splice_from_pipe() when
 442  *    locking is required around copying the pipe buffers to the
 443  *    destination.
 444  */
 445 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 446                           splice_actor *actor)
 447 {
 448         unsigned int head = pipe->head;
 449         unsigned int tail = pipe->tail;
 450         unsigned int mask = pipe->ring_size - 1;
 451         int ret;
 452
 453         while (!pipe_empty(head, tail)) {
 454                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 455
 456                 sd->len = buf->len;
 457                 if (sd->len > sd->total_len)
 458                         sd->len = sd->total_len;
 459
 460                 ret = pipe_buf_confirm(pipe, buf);
 461                 if (unlikely(ret)) {
 462                         if (ret == -ENODATA)
 463                                 ret = 0;
 464                         return ret;
 465                 }
 466
 467                 ret = actor(pipe, buf, sd);
 468                 if (ret <= 0)
 469                         return ret;
 470
 471                 buf->offset += ret;
 472                 buf->len -= ret;
 473
 474                 sd->num_spliced += ret;
 475                 sd->len -= ret;
 476                 sd->pos += ret;
 477                 sd->total_len -= ret;
 478
 479                 if (!buf->len) {
 480                         pipe_buf_release(pipe, buf);
 481                         tail++;
 482                         pipe->tail = tail;
 483                         if (pipe->files)
 484                                 sd->need_wakeup = true;
 485                 }
 486
 487                 if (!sd->total_len)
 488                         return 0;
 489         }
 490
 491         return 1;
 492 }
 493
 494 /* We know we have a pipe buffer, but maybe it's empty? */
 495 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 496 {
 497         unsigned int tail = pipe->tail;
 498         unsigned int mask = pipe->ring_size - 1;
 499         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 500
 501         if (unlikely(!buf->len)) {
 502                 pipe_buf_release(pipe, buf);
 503                 pipe->tail = tail+1;
 504                 return true;
 505         }
 506
 507         return false;
 508 }
 509
 510 /**
 511  * splice_from_pipe_next - wait for some data to splice from
 512  * @pipe:       pipe to splice from
 513  * @sd:         information about the splice operation
 514  *
 515  * Description:
 516  *    This function will wait for some data and return a positive
 517  *    value (one) if pipe buffers are available.  It will return zero
 518  *    or -errno if no more data needs to be spliced.
 519  */
 520 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 521 {
 522         /*
 523          * Check for signal early to make process killable when there are
 524          * always buffers available
 525          */
 526         if (signal_pending(current))
 527                 return -ERESTARTSYS;
 528
 529 repeat:
 530         while (pipe_empty(pipe->head, pipe->tail)) {
 531                 if (!pipe->writers)
 532                         return 0;
 533
 534                 if (sd->num_spliced)
 535                         return 0;
 536
 537                 if (sd->flags & SPLICE_F_NONBLOCK)
 538                         return -EAGAIN;
 539
 540                 if (signal_pending(current))
 541                         return -ERESTARTSYS;
 542
 543                 if (sd->need_wakeup) {
 544                         wakeup_pipe_writers(pipe);
 545                         sd->need_wakeup = false;
 546                 }
 547
 548                 pipe_wait_readable(pipe);
 549         }
 550
 551         if (eat_empty_buffer(pipe))
 552                 goto repeat;
 553
 554         return 1;
 555 }
 556
 557 /**
 558  * splice_from_pipe_begin - start splicing from pipe
 559  * @sd:         information about the splice operation
 560  *
 561  * Description:
 562  *    This function should be called before a loop containing
 563  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 564  *    initialize the necessary fields of @sd.
 565  */
 566 static void splice_from_pipe_begin(struct splice_desc *sd)
 567 {
 568         sd->num_spliced = 0;
 569         sd->need_wakeup = false;
 570 }
 571
 572 /**
 573  * splice_from_pipe_end - finish splicing from pipe
 574  * @pipe:       pipe to splice from
 575  * @sd:         information about the splice operation
 576  *
 577  * Description:
 578  *    This function will wake up pipe writers if necessary.  It should
 579  *    be called after a loop containing splice_from_pipe_next() and
 580  *    splice_from_pipe_feed().
 581  */
 582 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 583 {
 584         if (sd->need_wakeup)
 585                 wakeup_pipe_writers(pipe);
 586 }
 587
 588 /**
 589  * __splice_from_pipe - splice data from a pipe to given actor
 590  * @pipe:       pipe to splice from
 591  * @sd:         information to @actor
 592  * @actor:      handler that splices the data
 593  *
 594  * Description:
 595  *    This function does little more than loop over the pipe and call
 596  *    @actor to do the actual moving of a single struct pipe_buffer to
 597  *    the desired destination. See pipe_to_file, pipe_to_sendmsg, or
 598  *    pipe_to_user.
 599  *
 600  */
 601 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 602                            splice_actor *actor)
 603 {
 604         int ret;
 605
 606         splice_from_pipe_begin(sd);
 607         do {
 608                 cond_resched();
 609                 ret = splice_from_pipe_next(pipe, sd);
 610                 if (ret > 0)
 611                         ret = splice_from_pipe_feed(pipe, sd, actor);
 612         } while (ret > 0);
 613         splice_from_pipe_end(pipe, sd);
 614
 615         return sd->num_spliced ? sd->num_spliced : ret;
 616 }
 617 EXPORT_SYMBOL(__splice_from_pipe);
 618
 619 /**
 620  * splice_from_pipe - splice data from a pipe to a file
 621  * @pipe:       pipe to splice from
 622  * @out:        file to splice to
 623  * @ppos:       position in @out
 624  * @len:        how many bytes to splice
 625  * @flags:      splice modifier flags
 626  * @actor:      handler that splices the data
 627  *
 628  * Description:
 629  *    See __splice_from_pipe. This function locks the pipe inode,
 630  *    otherwise it's identical to __splice_from_pipe().
 631  *
 632  */
 633 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 634                          loff_t *ppos, size_t len, unsigned int flags,
 635                          splice_actor *actor)
 636 {
 637         ssize_t ret;
 638         struct splice_desc sd = {
 639                 .total_len = len,
 640                 .flags = flags,
 641                 .pos = *ppos,
 642                 .u.file = out,
 643         };
 644
 645         pipe_lock(pipe);
 646         ret = __splice_from_pipe(pipe, &sd, actor);
 647         pipe_unlock(pipe);
 648
 649         return ret;
 650 }
 651
 652 /**
 653  * iter_file_splice_write - splice data from a pipe to a file
 654  * @pipe:       pipe info
 655  * @out:        file to write to
 656  * @ppos:       position in @out
 657  * @len:        number of bytes to splice
 658  * @flags:      splice modifier flags
 659  *
 660  * Description:
 661  *    Will either move or copy pages (determined by @flags options) from
 662  *    the given pipe inode to the given file.
 663  *    This one is ->write_iter-based.
 664  *
 665  */
 666 ssize_t
 667 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 668                           loff_t *ppos, size_t len, unsigned int flags)
 669 {
 670         struct splice_desc sd = {
 671                 .total_len = len,
 672                 .flags = flags,
 673                 .pos = *ppos,
 674                 .u.file = out,
 675         };
 676         int nbufs = pipe->max_usage;
 677         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 678                                         GFP_KERNEL);
 679         ssize_t ret;
 680
 681         if (unlikely(!array))
 682                 return -ENOMEM;
 683
 684         pipe_lock(pipe);
 685
 686         splice_from_pipe_begin(&sd);
 687         while (sd.total_len) {
 688                 struct iov_iter from;
 689                 unsigned int head, tail, mask;
 690                 size_t left;
 691                 int n;
 692
 693                 ret = splice_from_pipe_next(pipe, &sd);
 694                 if (ret <= 0)
 695                         break;
 696
 697                 if (unlikely(nbufs < pipe->max_usage)) {
 698                         kfree(array);
 699                         nbufs = pipe->max_usage;
 700                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 701                                         GFP_KERNEL);
 702                         if (!array) {
 703                                 ret = -ENOMEM;
 704                                 break;
 705                         }
 706                 }
 707
 708                 head = pipe->head;
 709                 tail = pipe->tail;
 710                 mask = pipe->ring_size - 1;
 711
 712                 /* build the vector */
 713                 left = sd.total_len;
 714                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 715                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 716                         size_t this_len = buf->len;
 717
 718                         /* zero-length bvecs are not supported, skip them */
 719                         if (!this_len)
 720                                 continue;
 721                         this_len = min(this_len, left);
 722
 723                         ret = pipe_buf_confirm(pipe, buf);
 724                         if (unlikely(ret)) {
 725                                 if (ret == -ENODATA)
 726                                         ret = 0;
 727                                 goto done;
 728                         }
 729
 730                         bvec_set_page(&array[n], buf->page, this_len,
 731                                       buf->offset);
 732                         left -= this_len;
 733                         n++;
 734                 }
 735
 736                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 737                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 738                 if (ret <= 0)
 739                         break;
 740
 741                 sd.num_spliced += ret;
 742                 sd.total_len -= ret;
 743                 *ppos = sd.pos;
 744
 745                 /* dismiss the fully eaten buffers, adjust the partial one */
 746                 tail = pipe->tail;
 747                 while (ret) {
 748                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 749                         if (ret >= buf->len) {
 750                                 ret -= buf->len;
 751                                 buf->len = 0;
 752                                 pipe_buf_release(pipe, buf);
 753                                 tail++;
 754                                 pipe->tail = tail;
 755                                 if (pipe->files)
 756                                         sd.need_wakeup = true;
 757                         } else {
 758                                 buf->offset += ret;
 759                                 buf->len -= ret;
 760                                 ret = 0;
 761                         }
 762                 }
 763         }
 764 done:
 765         kfree(array);
 766         splice_from_pipe_end(pipe, &sd);
 767
 768         pipe_unlock(pipe);
 769
 770         if (sd.num_spliced)
 771                 ret = sd.num_spliced;
 772
 773         return ret;
 774 }
 775
 776 EXPORT_SYMBOL(iter_file_splice_write);
 777
 778 #ifdef CONFIG_NET
 779 /**
 780  * splice_to_socket - splice data from a pipe to a socket
 781  * @pipe:       pipe to splice from
 782  * @out:        socket to write to
 783  * @ppos:       position in @out
 784  * @len:        number of bytes to splice
 785  * @flags:      splice modifier flags
 786  *
 787  * Description:
 788  *    Will send @len bytes from the pipe to a network socket. No data copying
 789  *    is involved.
 790  *
 791  */
 792 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 793                          loff_t *ppos, size_t len, unsigned int flags)
 794 {
 795         struct socket *sock = sock_from_file(out);
 796         struct bio_vec bvec[16];
 797         struct msghdr msg = {};
 798         ssize_t ret = 0;
 799         size_t spliced = 0;
 800         bool need_wakeup = false;
 801
 802         pipe_lock(pipe);
 803
 804         while (len > 0) {
 805                 unsigned int head, tail, mask, bc = 0;
 806                 size_t remain = len;
 807
 808                 /*
 809                  * Check for signal early to make process killable when there
 810                  * are always buffers available
 811                  */
 812                 ret = -ERESTARTSYS;
 813                 if (signal_pending(current))
 814                         break;
 815
 816                 while (pipe_empty(pipe->head, pipe->tail)) {
 817                         ret = 0;
 818                         if (!pipe->writers)
 819                                 goto out;
 820
 821                         if (spliced)
 822                                 goto out;
 823
 824                         ret = -EAGAIN;
 825                         if (flags & SPLICE_F_NONBLOCK)
 826                                 goto out;
 827
 828                         ret = -ERESTARTSYS;
 829                         if (signal_pending(current))
 830                                 goto out;
 831
 832                         if (need_wakeup) {
 833                                 wakeup_pipe_writers(pipe);
 834                                 need_wakeup = false;
 835                         }
 836
 837                         pipe_wait_readable(pipe);
 838                 }
 839
 840                 head = pipe->head;
 841                 tail = pipe->tail;
 842                 mask = pipe->ring_size - 1;
 843
 844                 while (!pipe_empty(head, tail)) {
 845                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 846                         size_t seg;
 847
 848                         if (!buf->len) {
 849                                 tail++;
 850                                 continue;
 851                         }
 852
 853                         seg = min_t(size_t, remain, buf->len);
 854
 855                         ret = pipe_buf_confirm(pipe, buf);
 856                         if (unlikely(ret)) {
 857                                 if (ret == -ENODATA)
 858                                         ret = 0;
 859                                 break;
 860                         }
 861
 862                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 863                         remain -= seg;
 864                         if (remain == 0 || bc >= ARRAY_SIZE(bvec))
 865                                 break;
 866                         tail++;
 867                 }
 868
 869                 if (!bc)
 870                         break;
 871
 872                 msg.msg_flags = MSG_SPLICE_PAGES;
 873                 if (flags & SPLICE_F_MORE)
 874                         msg.msg_flags |= MSG_MORE;
 875                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 876                         msg.msg_flags |= MSG_MORE;
 877                 if (out->f_flags & O_NONBLOCK)
 878                         msg.msg_flags |= MSG_DONTWAIT;
 879
 880                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 881                               len - remain);
 882                 ret = sock_sendmsg(sock, &msg);
 883                 if (ret <= 0)
 884                         break;
 885
 886                 spliced += ret;
 887                 len -= ret;
 888                 tail = pipe->tail;
 889                 while (ret > 0) {
 890                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 891                         size_t seg = min_t(size_t, ret, buf->len);
 892
 893                         buf->offset += seg;
 894                         buf->len -= seg;
 895                         ret -= seg;
 896
 897                         if (!buf->len) {
 898                                 pipe_buf_release(pipe, buf);
 899                                 tail++;
 900                         }
 901                 }
 902
 903                 if (tail != pipe->tail) {
 904                         pipe->tail = tail;
 905                         if (pipe->files)
 906                                 need_wakeup = true;
 907                 }
 908         }
 909
 910 out:
 911         pipe_unlock(pipe);
 912         if (need_wakeup)
 913                 wakeup_pipe_writers(pipe);
 914         return spliced ?: ret;
 915 }
 916 #endif
 917
 918 static int warn_unsupported(struct file *file, const char *op)
 919 {
 920         pr_debug_ratelimited(
 921                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 922                 op, file, current->pid, current->comm);
 923         return -EINVAL;
 924 }
 925
 926 /*
 927  * Attempt to initiate a splice from pipe to file.
 928  */
 929 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 930                            loff_t *ppos, size_t len, unsigned int flags)
 931 {
 932         if (unlikely(!out->f_op->splice_write))
 933                 return warn_unsupported(out, "write");
 934         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 935 }
 936
 937 /*
 938  * Indicate to the caller that there was a premature EOF when reading from the
 939  * source and the caller didn't indicate they would be sending more data after
 940  * this.
 941  */
 942 static void do_splice_eof(struct splice_desc *sd)
 943 {
 944         if (sd->splice_eof)
 945                 sd->splice_eof(sd);
 946 }
 947
 948 /**
 949  * vfs_splice_read - Read data from a file and splice it into a pipe
 950  * @in:         File to splice from
 951  * @ppos:       Input file offset
 952  * @pipe:       Pipe to splice to
 953  * @len:        Number of bytes to splice
 954  * @flags:      Splice modifier flags (SPLICE_F_*)
 955  *
 956  * Splice the requested amount of data from the input file to the pipe.  This
 957  * is synchronous as the caller must hold the pipe lock across the entire
 958  * operation.
 959  *
 960  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 961  * a hole and a negative error code otherwise.
 962  */
 963 long vfs_splice_read(struct file *in, loff_t *ppos,
 964                      struct pipe_inode_info *pipe, size_t len,
 965                      unsigned int flags)
 966 {
 967         unsigned int p_space;
 968         int ret;
 969
 970         if (unlikely(!(in->f_mode & FMODE_READ)))
 971                 return -EBADF;
 972         if (!len)
 973                 return 0;
 974
 975         /* Don't try to read more the pipe has space for. */
 976         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 977         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 978
 979         ret = rw_verify_area(READ, in, ppos, len);
 980         if (unlikely(ret < 0))
 981                 return ret;
 982
 983         if (unlikely(len > MAX_RW_COUNT))
 984                 len = MAX_RW_COUNT;
 985
 986         if (unlikely(!in->f_op->splice_read))
 987                 return warn_unsupported(in, "read");
 988         /*
 989          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 990          * buffer, copy into it and splice that into the pipe.
 991          */
 992         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 993                 return copy_splice_read(in, ppos, pipe, len, flags);
 994         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 995 }
 996 EXPORT_SYMBOL_GPL(vfs_splice_read);
 997
 998 /**
 999  * splice_direct_to_actor - splices data directly between two non-pipes
1000  * @in:         file to splice from
1001  * @sd:         actor information on where to splice to
1002  * @actor:      handles the data splicing
1003  *
1004  * Description:
1005  *    This is a special case helper to splice directly between two
1006  *    points, without requiring an explicit pipe. Internally an allocated
1007  *    pipe is cached in the process, and reused during the lifetime of
1008  *    that process.
1009  *
1010  */
1011 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1012                                splice_direct_actor *actor)
1013 {
1014         struct pipe_inode_info *pipe;
1015         long ret, bytes;
1016         size_t len;
1017         int i, flags, more;
1018
1019         /*
1020          * We require the input to be seekable, as we don't want to randomly
1021          * drop data for eg socket -> socket splicing. Use the piped splicing
1022          * for that!
1023          */
1024         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1025                 return -EINVAL;
1026
1027         /*
1028          * neither in nor out is a pipe, setup an internal pipe attached to
1029          * 'out' and transfer the wanted data from 'in' to 'out' through that
1030          */
1031         pipe = current->splice_pipe;
1032         if (unlikely(!pipe)) {
1033                 pipe = alloc_pipe_info();
1034                 if (!pipe)
1035                         return -ENOMEM;
1036
1037                 /*
1038                  * We don't have an immediate reader, but we'll read the stuff
1039                  * out of the pipe right after the splice_to_pipe(). So set
1040                  * PIPE_READERS appropriately.
1041                  */
1042                 pipe->readers = 1;
1043
1044                 current->splice_pipe = pipe;
1045         }
1046
1047         /*
1048          * Do the splice.
1049          */
1050         bytes = 0;
1051         len = sd->total_len;
1052
1053         /* Don't block on output, we have to drain the direct pipe. */
1054         flags = sd->flags;
1055         sd->flags &= ~SPLICE_F_NONBLOCK;
1056
1057         /*
1058          * We signal MORE until we've read sufficient data to fulfill the
1059          * request and we keep signalling it if the caller set it.
1060          */
1061         more = sd->flags & SPLICE_F_MORE;
1062         sd->flags |= SPLICE_F_MORE;
1063
1064         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1065
1066         while (len) {
1067                 size_t read_len;
1068                 loff_t pos = sd->pos, prev_pos = pos;
1069
1070                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1071                 if (unlikely(ret <= 0))
1072                         goto read_failure;
1073
1074                 read_len = ret;
1075                 sd->total_len = read_len;
1076
1077                 /*
1078                  * If we now have sufficient data to fulfill the request then
1079                  * we clear SPLICE_F_MORE if it was not set initially.
1080                  */
1081                 if (read_len >= len && !more)
1082                         sd->flags &= ~SPLICE_F_MORE;
1083
1084                 /*
1085                  * NOTE: nonblocking mode only applies to the input. We
1086                  * must not do the output in nonblocking mode as then we
1087                  * could get stuck data in the internal pipe:
1088                  */
1089                 ret = actor(pipe, sd);
1090                 if (unlikely(ret <= 0)) {
1091                         sd->pos = prev_pos;
1092                         goto out_release;
1093                 }
1094
1095                 bytes += ret;
1096                 len -= ret;
1097                 sd->pos = pos;
1098
1099                 if (ret < read_len) {
1100                         sd->pos = prev_pos + ret;
1101                         goto out_release;
1102                 }
1103         }
1104
1105 done:
1106         pipe->tail = pipe->head = 0;
1107         file_accessed(in);
1108         return bytes;
1109
1110 read_failure:
1111         /*
1112          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1113          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1114          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1115          * least 1 byte *then* we will also do the ->splice_eof() call.
1116          */
1117         if (ret == 0 && !more && len > 0 && bytes)
1118                 do_splice_eof(sd);
1119 out_release:
1120         /*
1121          * If we did an incomplete transfer we must release
1122          * the pipe buffers in question:
1123          */
1124         for (i = 0; i < pipe->ring_size; i++) {
1125                 struct pipe_buffer *buf = &pipe->bufs[i];
1126
1127                 if (buf->ops)
1128                         pipe_buf_release(pipe, buf);
1129         }
1130
1131         if (!bytes)
1132                 bytes = ret;
1133
1134         goto done;
1135 }
1136 EXPORT_SYMBOL(splice_direct_to_actor);
1137
1138 static int direct_splice_actor(struct pipe_inode_info *pipe,
1139                                struct splice_desc *sd)
1140 {
1141         struct file *file = sd->u.file;
1142
1143         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1144                               sd->flags);
1145 }
1146
1147 static void direct_file_splice_eof(struct splice_desc *sd)
1148 {
1149         struct file *file = sd->u.file;
1150
1151         if (file->f_op->splice_eof)
1152                 file->f_op->splice_eof(file);
1153 }
1154
1155 /**
1156  * do_splice_direct - splices data directly between two files
1157  * @in:         file to splice from
1158  * @ppos:       input file offset
1159  * @out:        file to splice to
1160  * @opos:       output file offset
1161  * @len:        number of bytes to splice
1162  * @flags:      splice modifier flags
1163  *
1164  * Description:
1165  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1166  *    doing it in the application would incur an extra system call
1167  *    (splice in + splice out, as compared to just sendfile()). So this helper
1168  *    can splice directly through a process-private pipe.
1169  *
1170  */
1171 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1172                       loff_t *opos, size_t len, unsigned int flags)
1173 {
1174         struct splice_desc sd = {
1175                 .len            = len,
1176                 .total_len      = len,
1177                 .flags          = flags,
1178                 .pos            = *ppos,
1179                 .u.file         = out,
1180                 .splice_eof     = direct_file_splice_eof,
1181                 .opos           = opos,
1182         };
1183         long ret;
1184
1185         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1186                 return -EBADF;
1187
1188         if (unlikely(out->f_flags & O_APPEND))
1189                 return -EINVAL;
1190
1191         ret = rw_verify_area(WRITE, out, opos, len);
1192         if (unlikely(ret < 0))
1193                 return ret;
1194
1195         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1196         if (ret > 0)
1197                 *ppos = sd.pos;
1198
1199         return ret;
1200 }
1201 EXPORT_SYMBOL(do_splice_direct);
1202
1203 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1204 {
1205         for (;;) {
1206                 if (unlikely(!pipe->readers)) {
1207                         send_sig(SIGPIPE, current, 0);
1208                         return -EPIPE;
1209                 }
1210                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1211                         return 0;
1212                 if (flags & SPLICE_F_NONBLOCK)
1213                         return -EAGAIN;
1214                 if (signal_pending(current))
1215                         return -ERESTARTSYS;
1216                 pipe_wait_writable(pipe);
1217         }
1218 }
1219
1220 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1221                                struct pipe_inode_info *opipe,
1222                                size_t len, unsigned int flags);
1223
1224 long splice_file_to_pipe(struct file *in,
1225                          struct pipe_inode_info *opipe,
1226                          loff_t *offset,
1227                          size_t len, unsigned int flags)
1228 {
1229         long ret;
1230
1231         pipe_lock(opipe);
1232         ret = wait_for_space(opipe, flags);
1233         if (!ret)
1234                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1235         pipe_unlock(opipe);
1236         if (ret > 0)
1237                 wakeup_pipe_readers(opipe);
1238         return ret;
1239 }
1240
1241 /*
1242  * Determine where to splice to/from.
1243  */
1244 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1245                loff_t *off_out, size_t len, unsigned int flags)
1246 {
1247         struct pipe_inode_info *ipipe;
1248         struct pipe_inode_info *opipe;
1249         loff_t offset;
1250         long ret;
1251
1252         if (unlikely(!(in->f_mode & FMODE_READ) ||
1253                      !(out->f_mode & FMODE_WRITE)))
1254                 return -EBADF;
1255
1256         ipipe = get_pipe_info(in, true);
1257         opipe = get_pipe_info(out, true);
1258
1259         if (ipipe && opipe) {
1260                 if (off_in || off_out)
1261                         return -ESPIPE;
1262
1263                 /* Splicing to self would be fun, but... */
1264                 if (ipipe == opipe)
1265                         return -EINVAL;
1266
1267                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1268                         flags |= SPLICE_F_NONBLOCK;
1269
1270                 ret = splice_pipe_to_pipe(ipipe, opipe, len, flags);
1271         } else if (ipipe) {
1272                 if (off_in)
1273                         return -ESPIPE;
1274                 if (off_out) {
1275                         if (!(out->f_mode & FMODE_PWRITE))
1276                                 return -EINVAL;
1277                         offset = *off_out;
1278                 } else {
1279                         offset = out->f_pos;
1280                 }
1281
1282                 if (unlikely(out->f_flags & O_APPEND))
1283                         return -EINVAL;
1284
1285                 ret = rw_verify_area(WRITE, out, &offset, len);
1286                 if (unlikely(ret < 0))
1287                         return ret;
1288
1289                 if (in->f_flags & O_NONBLOCK)
1290                         flags |= SPLICE_F_NONBLOCK;
1291
1292                 file_start_write(out);
1293                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1294                 file_end_write(out);
1295
1296                 if (!off_out)
1297                         out->f_pos = offset;
1298                 else
1299                         *off_out = offset;
1300         } else if (opipe) {
1301                 if (off_out)
1302                         return -ESPIPE;
1303                 if (off_in) {
1304                         if (!(in->f_mode & FMODE_PREAD))
1305                                 return -EINVAL;
1306                         offset = *off_in;
1307                 } else {
1308                         offset = in->f_pos;
1309                 }
1310
1311                 if (out->f_flags & O_NONBLOCK)
1312                         flags |= SPLICE_F_NONBLOCK;
1313
1314                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1315
1316                 if (!off_in)
1317                         in->f_pos = offset;
1318                 else
1319                         *off_in = offset;
1320         } else {
1321                 ret = -EINVAL;
1322         }
1323
1324         if (ret > 0) {
1325                 /*
1326                  * Generate modify out before access in:
1327                  * do_splice_from() may've already sent modify out,
1328                  * and this ensures the events get merged.
1329                  */
1330                 fsnotify_modify(out);
1331                 fsnotify_access(in);
1332         }
1333
1334         return ret;
1335 }
1336
1337 static long __do_splice(struct file *in, loff_t __user *off_in,
1338                         struct file *out, loff_t __user *off_out,
1339                         size_t len, unsigned int flags)
1340 {
1341         struct pipe_inode_info *ipipe;
1342         struct pipe_inode_info *opipe;
1343         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1344         long ret;
1345
1346         ipipe = get_pipe_info(in, true);
1347         opipe = get_pipe_info(out, true);
1348
1349         if (ipipe) {
1350                 if (off_in)
1351                         return -ESPIPE;
1352                 pipe_clear_nowait(in);
1353         }
1354         if (opipe) {
1355                 if (off_out)
1356                         return -ESPIPE;
1357                 pipe_clear_nowait(out);
1358         }
1359
1360         if (off_out) {
1361                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1362                         return -EFAULT;
1363                 __off_out = &offset;
1364         }
1365         if (off_in) {
1366                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1367                         return -EFAULT;
1368                 __off_in = &offset;
1369         }
1370
1371         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1372         if (ret < 0)
1373                 return ret;
1374
1375         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1376                 return -EFAULT;
1377         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1378                 return -EFAULT;
1379
1380         return ret;
1381 }
1382
1383 static int iter_to_pipe(struct iov_iter *from,
1384                         struct pipe_inode_info *pipe,
1385                         unsigned flags)
1386 {
1387         struct pipe_buffer buf = {
1388                 .ops = &user_page_pipe_buf_ops,
1389                 .flags = flags
1390         };
1391         size_t total = 0;
1392         int ret = 0;
1393
1394         while (iov_iter_count(from)) {
1395                 struct page *pages[16];
1396                 ssize_t left;
1397                 size_t start;
1398                 int i, n;
1399
1400                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1401                 if (left <= 0) {
1402                         ret = left;
1403                         break;
1404                 }
1405
1406                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1407                 for (i = 0; i < n; i++) {
1408                         int size = min_t(int, left, PAGE_SIZE - start);
1409
1410                         buf.page = pages[i];
1411                         buf.offset = start;
1412                         buf.len = size;
1413                         ret = add_to_pipe(pipe, &buf);
1414                         if (unlikely(ret < 0)) {
1415                                 iov_iter_revert(from, left);
1416                                 // this one got dropped by add_to_pipe()
1417                                 while (++i < n)
1418                                         put_page(pages[i]);
1419                                 goto out;
1420                         }
1421                         total += ret;
1422                         left -= size;
1423                         start = 0;
1424                 }
1425         }
1426 out:
1427         return total ? total : ret;
1428 }
1429
1430 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1431                         struct splice_desc *sd)
1432 {
1433         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1434         return n == sd->len ? n : -EFAULT;
1435 }
1436
1437 /*
1438  * For lack of a better implementation, implement vmsplice() to userspace
1439  * as a simple copy of the pipes pages to the user iov.
1440  */
1441 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1442                              unsigned int flags)
1443 {
1444         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1445         struct splice_desc sd = {
1446                 .total_len = iov_iter_count(iter),
1447                 .flags = flags,
1448                 .u.data = iter
1449         };
1450         long ret = 0;
1451
1452         if (!pipe)
1453                 return -EBADF;
1454
1455         pipe_clear_nowait(file);
1456
1457         if (sd.total_len) {
1458                 pipe_lock(pipe);
1459                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1460                 pipe_unlock(pipe);
1461         }
1462
1463         if (ret > 0)
1464                 fsnotify_access(file);
1465
1466         return ret;
1467 }
1468
1469 /*
1470  * vmsplice splices a user address range into a pipe. It can be thought of
1471  * as splice-from-memory, where the regular splice is splice-from-file (or
1472  * to file). In both cases the output is a pipe, naturally.
1473  */
1474 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1475                              unsigned int flags)
1476 {
1477         struct pipe_inode_info *pipe;
1478         long ret = 0;
1479         unsigned buf_flag = 0;
1480
1481         if (flags & SPLICE_F_GIFT)
1482                 buf_flag = PIPE_BUF_FLAG_GIFT;
1483
1484         pipe = get_pipe_info(file, true);
1485         if (!pipe)
1486                 return -EBADF;
1487
1488         pipe_clear_nowait(file);
1489
1490         pipe_lock(pipe);
1491         ret = wait_for_space(pipe, flags);
1492         if (!ret)
1493                 ret = iter_to_pipe(iter, pipe, buf_flag);
1494         pipe_unlock(pipe);
1495         if (ret > 0) {
1496                 wakeup_pipe_readers(pipe);
1497                 fsnotify_modify(file);
1498         }
1499         return ret;
1500 }
1501
1502 static int vmsplice_type(struct fd f, int *type)
1503 {
1504         if (!f.file)
1505                 return -EBADF;
1506         if (f.file->f_mode & FMODE_WRITE) {
1507                 *type = ITER_SOURCE;
1508         } else if (f.file->f_mode & FMODE_READ) {
1509                 *type = ITER_DEST;
1510         } else {
1511                 fdput(f);
1512                 return -EBADF;
1513         }
1514         return 0;
1515 }
1516
1517 /*
1518  * Note that vmsplice only really supports true splicing _from_ user memory
1519  * to a pipe, not the other way around. Splicing from user memory is a simple
1520  * operation that can be supported without any funky alignment restrictions
1521  * or nasty vm tricks. We simply map in the user memory and fill them into
1522  * a pipe. The reverse isn't quite as easy, though. There are two possible
1523  * solutions for that:
1524  *
1525  *      - memcpy() the data internally, at which point we might as well just
1526  *        do a regular read() on the buffer anyway.
1527  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1528  *        has restriction limitations on both ends of the pipe).
1529  *
1530  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1531  *
1532  */
1533 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1534                 unsigned long, nr_segs, unsigned int, flags)
1535 {
1536         struct iovec iovstack[UIO_FASTIOV];
1537         struct iovec *iov = iovstack;
1538         struct iov_iter iter;
1539         ssize_t error;
1540         struct fd f;
1541         int type;
1542
1543         if (unlikely(flags & ~SPLICE_F_ALL))
1544                 return -EINVAL;
1545
1546         f = fdget(fd);
1547         error = vmsplice_type(f, &type);
1548         if (error)
1549                 return error;
1550
1551         error = import_iovec(type, uiov, nr_segs,
1552                              ARRAY_SIZE(iovstack), &iov, &iter);
1553         if (error < 0)
1554                 goto out_fdput;
1555
1556         if (!iov_iter_count(&iter))
1557                 error = 0;
1558         else if (type == ITER_SOURCE)
1559                 error = vmsplice_to_pipe(f.file, &iter, flags);
1560         else
1561                 error = vmsplice_to_user(f.file, &iter, flags);
1562
1563         kfree(iov);
1564 out_fdput:
1565         fdput(f);
1566         return error;
1567 }
1568
1569 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1570                 int, fd_out, loff_t __user *, off_out,
1571                 size_t, len, unsigned int, flags)
1572 {
1573         struct fd in, out;
1574         long error;
1575
1576         if (unlikely(!len))
1577                 return 0;
1578
1579         if (unlikely(flags & ~SPLICE_F_ALL))
1580                 return -EINVAL;
1581
1582         error = -EBADF;
1583         in = fdget(fd_in);
1584         if (in.file) {
1585                 out = fdget(fd_out);
1586                 if (out.file) {
1587                         error = __do_splice(in.file, off_in, out.file, off_out,
1588                                                 len, flags);
1589                         fdput(out);
1590                 }
1591                 fdput(in);
1592         }
1593         return error;
1594 }
1595
1596 /*
1597  * Make sure there's data to read. Wait for input if we can, otherwise
1598  * return an appropriate error.
1599  */
1600 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1601 {
1602         int ret;
1603
1604         /*
1605          * Check the pipe occupancy without the inode lock first. This function
1606          * is speculative anyways, so missing one is ok.
1607          */
1608         if (!pipe_empty(pipe->head, pipe->tail))
1609                 return 0;
1610
1611         ret = 0;
1612         pipe_lock(pipe);
1613
1614         while (pipe_empty(pipe->head, pipe->tail)) {
1615                 if (signal_pending(current)) {
1616                         ret = -ERESTARTSYS;
1617                         break;
1618                 }
1619                 if (!pipe->writers)
1620                         break;
1621                 if (flags & SPLICE_F_NONBLOCK) {
1622                         ret = -EAGAIN;
1623                         break;
1624                 }
1625                 pipe_wait_readable(pipe);
1626         }
1627
1628         pipe_unlock(pipe);
1629         return ret;
1630 }
1631
1632 /*
1633  * Make sure there's writeable room. Wait for room if we can, otherwise
1634  * return an appropriate error.
1635  */
1636 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1637 {
1638         int ret;
1639
1640         /*
1641          * Check pipe occupancy without the inode lock first. This function
1642          * is speculative anyways, so missing one is ok.
1643          */
1644         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1645                 return 0;
1646
1647         ret = 0;
1648         pipe_lock(pipe);
1649
1650         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1651                 if (!pipe->readers) {
1652                         send_sig(SIGPIPE, current, 0);
1653                         ret = -EPIPE;
1654                         break;
1655                 }
1656                 if (flags & SPLICE_F_NONBLOCK) {
1657                         ret = -EAGAIN;
1658                         break;
1659                 }
1660                 if (signal_pending(current)) {
1661                         ret = -ERESTARTSYS;
1662                         break;
1663                 }
1664                 pipe_wait_writable(pipe);
1665         }
1666
1667         pipe_unlock(pipe);
1668         return ret;
1669 }
1670
1671 /*
1672  * Splice contents of ipipe to opipe.
1673  */
1674 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1675                                struct pipe_inode_info *opipe,
1676                                size_t len, unsigned int flags)
1677 {
1678         struct pipe_buffer *ibuf, *obuf;
1679         unsigned int i_head, o_head;
1680         unsigned int i_tail, o_tail;
1681         unsigned int i_mask, o_mask;
1682         int ret = 0;
1683         bool input_wakeup = false;
1684
1685
1686 retry:
1687         ret = ipipe_prep(ipipe, flags);
1688         if (ret)
1689                 return ret;
1690
1691         ret = opipe_prep(opipe, flags);
1692         if (ret)
1693                 return ret;
1694
1695         /*
1696          * Potential ABBA deadlock, work around it by ordering lock
1697          * grabbing by pipe info address. Otherwise two different processes
1698          * could deadlock (one doing tee from A -> B, the other from B -> A).
1699          */
1700         pipe_double_lock(ipipe, opipe);
1701
1702         i_tail = ipipe->tail;
1703         i_mask = ipipe->ring_size - 1;
1704         o_head = opipe->head;
1705         o_mask = opipe->ring_size - 1;
1706
1707         do {
1708                 size_t o_len;
1709
1710                 if (!opipe->readers) {
1711                         send_sig(SIGPIPE, current, 0);
1712                         if (!ret)
1713                                 ret = -EPIPE;
1714                         break;
1715                 }
1716
1717                 i_head = ipipe->head;
1718                 o_tail = opipe->tail;
1719
1720                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1721                         break;
1722
1723                 /*
1724                  * Cannot make any progress, because either the input
1725                  * pipe is empty or the output pipe is full.
1726                  */
1727                 if (pipe_empty(i_head, i_tail) ||
1728                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1729                         /* Already processed some buffers, break */
1730                         if (ret)
1731                                 break;
1732
1733                         if (flags & SPLICE_F_NONBLOCK) {
1734                                 ret = -EAGAIN;
1735                                 break;
1736                         }
1737
1738                         /*
1739                          * We raced with another reader/writer and haven't
1740                          * managed to process any buffers.  A zero return
1741                          * value means EOF, so retry instead.
1742                          */
1743                         pipe_unlock(ipipe);
1744                         pipe_unlock(opipe);
1745                         goto retry;
1746                 }
1747
1748                 ibuf = &ipipe->bufs[i_tail & i_mask];
1749                 obuf = &opipe->bufs[o_head & o_mask];
1750
1751                 if (len >= ibuf->len) {
1752                         /*
1753                          * Simply move the whole buffer from ipipe to opipe
1754                          */
1755                         *obuf = *ibuf;
1756                         ibuf->ops = NULL;
1757                         i_tail++;
1758                         ipipe->tail = i_tail;
1759                         input_wakeup = true;
1760                         o_len = obuf->len;
1761                         o_head++;
1762                         opipe->head = o_head;
1763                 } else {
1764                         /*
1765                          * Get a reference to this pipe buffer,
1766                          * so we can copy the contents over.
1767                          */
1768                         if (!pipe_buf_get(ipipe, ibuf)) {
1769                                 if (ret == 0)
1770                                         ret = -EFAULT;
1771                                 break;
1772                         }
1773                         *obuf = *ibuf;
1774
1775                         /*
1776                          * Don't inherit the gift and merge flags, we need to
1777                          * prevent multiple steals of this page.
1778                          */
1779                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1780                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1781
1782                         obuf->len = len;
1783                         ibuf->offset += len;
1784                         ibuf->len -= len;
1785                         o_len = len;
1786                         o_head++;
1787                         opipe->head = o_head;
1788                 }
1789                 ret += o_len;
1790                 len -= o_len;
1791         } while (len);
1792
1793         pipe_unlock(ipipe);
1794         pipe_unlock(opipe);
1795
1796         /*
1797          * If we put data in the output pipe, wakeup any potential readers.
1798          */
1799         if (ret > 0)
1800                 wakeup_pipe_readers(opipe);
1801
1802         if (input_wakeup)
1803                 wakeup_pipe_writers(ipipe);
1804
1805         return ret;
1806 }
1807
1808 /*
1809  * Link contents of ipipe to opipe.
1810  */
1811 static int link_pipe(struct pipe_inode_info *ipipe,
1812                      struct pipe_inode_info *opipe,
1813                      size_t len, unsigned int flags)
1814 {
1815         struct pipe_buffer *ibuf, *obuf;
1816         unsigned int i_head, o_head;
1817         unsigned int i_tail, o_tail;
1818         unsigned int i_mask, o_mask;
1819         int ret = 0;
1820
1821         /*
1822          * Potential ABBA deadlock, work around it by ordering lock
1823          * grabbing by pipe info address. Otherwise two different processes
1824          * could deadlock (one doing tee from A -> B, the other from B -> A).
1825          */
1826         pipe_double_lock(ipipe, opipe);
1827
1828         i_tail = ipipe->tail;
1829         i_mask = ipipe->ring_size - 1;
1830         o_head = opipe->head;
1831         o_mask = opipe->ring_size - 1;
1832
1833         do {
1834                 if (!opipe->readers) {
1835                         send_sig(SIGPIPE, current, 0);
1836                         if (!ret)
1837                                 ret = -EPIPE;
1838                         break;
1839                 }
1840
1841                 i_head = ipipe->head;
1842                 o_tail = opipe->tail;
1843
1844                 /*
1845                  * If we have iterated all input buffers or run out of
1846                  * output room, break.
1847                  */
1848                 if (pipe_empty(i_head, i_tail) ||
1849                     pipe_full(o_head, o_tail, opipe->max_usage))
1850                         break;
1851
1852                 ibuf = &ipipe->bufs[i_tail & i_mask];
1853                 obuf = &opipe->bufs[o_head & o_mask];
1854
1855                 /*
1856                  * Get a reference to this pipe buffer,
1857                  * so we can copy the contents over.
1858                  */
1859                 if (!pipe_buf_get(ipipe, ibuf)) {
1860                         if (ret == 0)
1861                                 ret = -EFAULT;
1862                         break;
1863                 }
1864
1865                 *obuf = *ibuf;
1866
1867                 /*
1868                  * Don't inherit the gift and merge flag, we need to prevent
1869                  * multiple steals of this page.
1870                  */
1871                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1872                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1873
1874                 if (obuf->len > len)
1875                         obuf->len = len;
1876                 ret += obuf->len;
1877                 len -= obuf->len;
1878
1879                 o_head++;
1880                 opipe->head = o_head;
1881                 i_tail++;
1882         } while (len);
1883
1884         pipe_unlock(ipipe);
1885         pipe_unlock(opipe);
1886
1887         /*
1888          * If we put data in the output pipe, wakeup any potential readers.
1889          */
1890         if (ret > 0)
1891                 wakeup_pipe_readers(opipe);
1892
1893         return ret;
1894 }
1895
1896 /*
1897  * This is a tee(1) implementation that works on pipes. It doesn't copy
1898  * any data, it simply references the 'in' pages on the 'out' pipe.
1899  * The 'flags' used are the SPLICE_F_* variants, currently the only
1900  * applicable one is SPLICE_F_NONBLOCK.
1901  */
1902 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1903 {
1904         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1905         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1906         int ret = -EINVAL;
1907
1908         if (unlikely(!(in->f_mode & FMODE_READ) ||
1909                      !(out->f_mode & FMODE_WRITE)))
1910                 return -EBADF;
1911
1912         /*
1913          * Duplicate the contents of ipipe to opipe without actually
1914          * copying the data.
1915          */
1916         if (ipipe && opipe && ipipe != opipe) {
1917                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1918                         flags |= SPLICE_F_NONBLOCK;
1919
1920                 /*
1921                  * Keep going, unless we encounter an error. The ipipe/opipe
1922                  * ordering doesn't really matter.
1923                  */
1924                 ret = ipipe_prep(ipipe, flags);
1925                 if (!ret) {
1926                         ret = opipe_prep(opipe, flags);
1927                         if (!ret)
1928                                 ret = link_pipe(ipipe, opipe, len, flags);
1929                 }
1930         }
1931
1932         if (ret > 0) {
1933                 fsnotify_access(in);
1934                 fsnotify_modify(out);
1935         }
1936
1937         return ret;
1938 }
1939
1940 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1941 {
1942         struct fd in, out;
1943         int error;
1944
1945         if (unlikely(flags & ~SPLICE_F_ALL))
1946                 return -EINVAL;
1947
1948         if (unlikely(!len))
1949                 return 0;
1950
1951         error = -EBADF;
1952         in = fdget(fdin);
1953         if (in.file) {
1954                 out = fdget(fdout);
1955                 if (out.file) {
1956                         error = do_tee(in.file, out.file, len, flags);
1957                         fdput(out);
1958                 }
1959                 fdput(in);
1960         }
1961
1962         return error;
1963 }