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
 878                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 879                               len - remain);
 880                 ret = sock_sendmsg(sock, &msg);
 881                 if (ret <= 0)
 882                         break;
 883
 884                 spliced += ret;
 885                 len -= ret;
 886                 tail = pipe->tail;
 887                 while (ret > 0) {
 888                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 889                         size_t seg = min_t(size_t, ret, buf->len);
 890
 891                         buf->offset += seg;
 892                         buf->len -= seg;
 893                         ret -= seg;
 894
 895                         if (!buf->len) {
 896                                 pipe_buf_release(pipe, buf);
 897                                 tail++;
 898                         }
 899                 }
 900
 901                 if (tail != pipe->tail) {
 902                         pipe->tail = tail;
 903                         if (pipe->files)
 904                                 need_wakeup = true;
 905                 }
 906         }
 907
 908 out:
 909         pipe_unlock(pipe);
 910         if (need_wakeup)
 911                 wakeup_pipe_writers(pipe);
 912         return spliced ?: ret;
 913 }
 914 #endif
 915
 916 static int warn_unsupported(struct file *file, const char *op)
 917 {
 918         pr_debug_ratelimited(
 919                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 920                 op, file, current->pid, current->comm);
 921         return -EINVAL;
 922 }
 923
 924 /*
 925  * Attempt to initiate a splice from pipe to file.
 926  */
 927 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 928                            loff_t *ppos, size_t len, unsigned int flags)
 929 {
 930         if (unlikely(!out->f_op->splice_write))
 931                 return warn_unsupported(out, "write");
 932         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 933 }
 934
 935 /*
 936  * Indicate to the caller that there was a premature EOF when reading from the
 937  * source and the caller didn't indicate they would be sending more data after
 938  * this.
 939  */
 940 static void do_splice_eof(struct splice_desc *sd)
 941 {
 942         if (sd->splice_eof)
 943                 sd->splice_eof(sd);
 944 }
 945
 946 /**
 947  * vfs_splice_read - Read data from a file and splice it into a pipe
 948  * @in:         File to splice from
 949  * @ppos:       Input file offset
 950  * @pipe:       Pipe to splice to
 951  * @len:        Number of bytes to splice
 952  * @flags:      Splice modifier flags (SPLICE_F_*)
 953  *
 954  * Splice the requested amount of data from the input file to the pipe.  This
 955  * is synchronous as the caller must hold the pipe lock across the entire
 956  * operation.
 957  *
 958  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 959  * a hole and a negative error code otherwise.
 960  */
 961 long vfs_splice_read(struct file *in, loff_t *ppos,
 962                      struct pipe_inode_info *pipe, size_t len,
 963                      unsigned int flags)
 964 {
 965         unsigned int p_space;
 966         int ret;
 967
 968         if (unlikely(!(in->f_mode & FMODE_READ)))
 969                 return -EBADF;
 970         if (!len)
 971                 return 0;
 972
 973         /* Don't try to read more the pipe has space for. */
 974         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 975         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 976
 977         ret = rw_verify_area(READ, in, ppos, len);
 978         if (unlikely(ret < 0))
 979                 return ret;
 980
 981         if (unlikely(len > MAX_RW_COUNT))
 982                 len = MAX_RW_COUNT;
 983
 984         if (unlikely(!in->f_op->splice_read))
 985                 return warn_unsupported(in, "read");
 986         /*
 987          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 988          * buffer, copy into it and splice that into the pipe.
 989          */
 990         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 991                 return copy_splice_read(in, ppos, pipe, len, flags);
 992         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 993 }
 994 EXPORT_SYMBOL_GPL(vfs_splice_read);
 995
 996 /**
 997  * splice_direct_to_actor - splices data directly between two non-pipes
 998  * @in:         file to splice from
 999  * @sd:         actor information on where to splice to
1000  * @actor:      handles the data splicing
1001  *
1002  * Description:
1003  *    This is a special case helper to splice directly between two
1004  *    points, without requiring an explicit pipe. Internally an allocated
1005  *    pipe is cached in the process, and reused during the lifetime of
1006  *    that process.
1007  *
1008  */
1009 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1010                                splice_direct_actor *actor)
1011 {
1012         struct pipe_inode_info *pipe;
1013         long ret, bytes;
1014         size_t len;
1015         int i, flags, more;
1016
1017         /*
1018          * We require the input to be seekable, as we don't want to randomly
1019          * drop data for eg socket -> socket splicing. Use the piped splicing
1020          * for that!
1021          */
1022         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1023                 return -EINVAL;
1024
1025         /*
1026          * neither in nor out is a pipe, setup an internal pipe attached to
1027          * 'out' and transfer the wanted data from 'in' to 'out' through that
1028          */
1029         pipe = current->splice_pipe;
1030         if (unlikely(!pipe)) {
1031                 pipe = alloc_pipe_info();
1032                 if (!pipe)
1033                         return -ENOMEM;
1034
1035                 /*
1036                  * We don't have an immediate reader, but we'll read the stuff
1037                  * out of the pipe right after the splice_to_pipe(). So set
1038                  * PIPE_READERS appropriately.
1039                  */
1040                 pipe->readers = 1;
1041
1042                 current->splice_pipe = pipe;
1043         }
1044
1045         /*
1046          * Do the splice.
1047          */
1048         bytes = 0;
1049         len = sd->total_len;
1050
1051         /* Don't block on output, we have to drain the direct pipe. */
1052         flags = sd->flags;
1053         sd->flags &= ~SPLICE_F_NONBLOCK;
1054
1055         /*
1056          * We signal MORE until we've read sufficient data to fulfill the
1057          * request and we keep signalling it if the caller set it.
1058          */
1059         more = sd->flags & SPLICE_F_MORE;
1060         sd->flags |= SPLICE_F_MORE;
1061
1062         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1063
1064         while (len) {
1065                 size_t read_len;
1066                 loff_t pos = sd->pos, prev_pos = pos;
1067
1068                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1069                 if (unlikely(ret <= 0))
1070                         goto read_failure;
1071
1072                 read_len = ret;
1073                 sd->total_len = read_len;
1074
1075                 /*
1076                  * If we now have sufficient data to fulfill the request then
1077                  * we clear SPLICE_F_MORE if it was not set initially.
1078                  */
1079                 if (read_len >= len && !more)
1080                         sd->flags &= ~SPLICE_F_MORE;
1081
1082                 /*
1083                  * NOTE: nonblocking mode only applies to the input. We
1084                  * must not do the output in nonblocking mode as then we
1085                  * could get stuck data in the internal pipe:
1086                  */
1087                 ret = actor(pipe, sd);
1088                 if (unlikely(ret <= 0)) {
1089                         sd->pos = prev_pos;
1090                         goto out_release;
1091                 }
1092
1093                 bytes += ret;
1094                 len -= ret;
1095                 sd->pos = pos;
1096
1097                 if (ret < read_len) {
1098                         sd->pos = prev_pos + ret;
1099                         goto out_release;
1100                 }
1101         }
1102
1103 done:
1104         pipe->tail = pipe->head = 0;
1105         file_accessed(in);
1106         return bytes;
1107
1108 read_failure:
1109         /*
1110          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1111          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1112          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1113          * least 1 byte *then* we will also do the ->splice_eof() call.
1114          */
1115         if (ret == 0 && !more && len > 0 && bytes)
1116                 do_splice_eof(sd);
1117 out_release:
1118         /*
1119          * If we did an incomplete transfer we must release
1120          * the pipe buffers in question:
1121          */
1122         for (i = 0; i < pipe->ring_size; i++) {
1123                 struct pipe_buffer *buf = &pipe->bufs[i];
1124
1125                 if (buf->ops)
1126                         pipe_buf_release(pipe, buf);
1127         }
1128
1129         if (!bytes)
1130                 bytes = ret;
1131
1132         goto done;
1133 }
1134 EXPORT_SYMBOL(splice_direct_to_actor);
1135
1136 static int direct_splice_actor(struct pipe_inode_info *pipe,
1137                                struct splice_desc *sd)
1138 {
1139         struct file *file = sd->u.file;
1140
1141         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1142                               sd->flags);
1143 }
1144
1145 static void direct_file_splice_eof(struct splice_desc *sd)
1146 {
1147         struct file *file = sd->u.file;
1148
1149         if (file->f_op->splice_eof)
1150                 file->f_op->splice_eof(file);
1151 }
1152
1153 /**
1154  * do_splice_direct - splices data directly between two files
1155  * @in:         file to splice from
1156  * @ppos:       input file offset
1157  * @out:        file to splice to
1158  * @opos:       output file offset
1159  * @len:        number of bytes to splice
1160  * @flags:      splice modifier flags
1161  *
1162  * Description:
1163  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1164  *    doing it in the application would incur an extra system call
1165  *    (splice in + splice out, as compared to just sendfile()). So this helper
1166  *    can splice directly through a process-private pipe.
1167  *
1168  */
1169 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1170                       loff_t *opos, size_t len, unsigned int flags)
1171 {
1172         struct splice_desc sd = {
1173                 .len            = len,
1174                 .total_len      = len,
1175                 .flags          = flags,
1176                 .pos            = *ppos,
1177                 .u.file         = out,
1178                 .splice_eof     = direct_file_splice_eof,
1179                 .opos           = opos,
1180         };
1181         long ret;
1182
1183         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1184                 return -EBADF;
1185
1186         if (unlikely(out->f_flags & O_APPEND))
1187                 return -EINVAL;
1188
1189         ret = rw_verify_area(WRITE, out, opos, len);
1190         if (unlikely(ret < 0))
1191                 return ret;
1192
1193         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1194         if (ret > 0)
1195                 *ppos = sd.pos;
1196
1197         return ret;
1198 }
1199 EXPORT_SYMBOL(do_splice_direct);
1200
1201 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1202 {
1203         for (;;) {
1204                 if (unlikely(!pipe->readers)) {
1205                         send_sig(SIGPIPE, current, 0);
1206                         return -EPIPE;
1207                 }
1208                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1209                         return 0;
1210                 if (flags & SPLICE_F_NONBLOCK)
1211                         return -EAGAIN;
1212                 if (signal_pending(current))
1213                         return -ERESTARTSYS;
1214                 pipe_wait_writable(pipe);
1215         }
1216 }
1217
1218 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1219                                struct pipe_inode_info *opipe,
1220                                size_t len, unsigned int flags);
1221
1222 long splice_file_to_pipe(struct file *in,
1223                          struct pipe_inode_info *opipe,
1224                          loff_t *offset,
1225                          size_t len, unsigned int flags)
1226 {
1227         long ret;
1228
1229         pipe_lock(opipe);
1230         ret = wait_for_space(opipe, flags);
1231         if (!ret)
1232                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1233         pipe_unlock(opipe);
1234         if (ret > 0)
1235                 wakeup_pipe_readers(opipe);
1236         return ret;
1237 }
1238
1239 /*
1240  * Determine where to splice to/from.
1241  */
1242 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1243                loff_t *off_out, size_t len, unsigned int flags)
1244 {
1245         struct pipe_inode_info *ipipe;
1246         struct pipe_inode_info *opipe;
1247         loff_t offset;
1248         long ret;
1249
1250         if (unlikely(!(in->f_mode & FMODE_READ) ||
1251                      !(out->f_mode & FMODE_WRITE)))
1252                 return -EBADF;
1253
1254         ipipe = get_pipe_info(in, true);
1255         opipe = get_pipe_info(out, true);
1256
1257         if (ipipe && opipe) {
1258                 if (off_in || off_out)
1259                         return -ESPIPE;
1260
1261                 /* Splicing to self would be fun, but... */
1262                 if (ipipe == opipe)
1263                         return -EINVAL;
1264
1265                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1266                         flags |= SPLICE_F_NONBLOCK;
1267
1268                 ret = splice_pipe_to_pipe(ipipe, opipe, len, flags);
1269         } else if (ipipe) {
1270                 if (off_in)
1271                         return -ESPIPE;
1272                 if (off_out) {
1273                         if (!(out->f_mode & FMODE_PWRITE))
1274                                 return -EINVAL;
1275                         offset = *off_out;
1276                 } else {
1277                         offset = out->f_pos;
1278                 }
1279
1280                 if (unlikely(out->f_flags & O_APPEND))
1281                         return -EINVAL;
1282
1283                 ret = rw_verify_area(WRITE, out, &offset, len);
1284                 if (unlikely(ret < 0))
1285                         return ret;
1286
1287                 if (in->f_flags & O_NONBLOCK)
1288                         flags |= SPLICE_F_NONBLOCK;
1289
1290                 file_start_write(out);
1291                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1292                 file_end_write(out);
1293
1294                 if (!off_out)
1295                         out->f_pos = offset;
1296                 else
1297                         *off_out = offset;
1298         } else if (opipe) {
1299                 if (off_out)
1300                         return -ESPIPE;
1301                 if (off_in) {
1302                         if (!(in->f_mode & FMODE_PREAD))
1303                                 return -EINVAL;
1304                         offset = *off_in;
1305                 } else {
1306                         offset = in->f_pos;
1307                 }
1308
1309                 if (out->f_flags & O_NONBLOCK)
1310                         flags |= SPLICE_F_NONBLOCK;
1311
1312                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1313
1314                 if (!off_in)
1315                         in->f_pos = offset;
1316                 else
1317                         *off_in = offset;
1318         } else {
1319                 ret = -EINVAL;
1320         }
1321
1322         if (ret > 0) {
1323                 /*
1324                  * Generate modify out before access in:
1325                  * do_splice_from() may've already sent modify out,
1326                  * and this ensures the events get merged.
1327                  */
1328                 fsnotify_modify(out);
1329                 fsnotify_access(in);
1330         }
1331
1332         return ret;
1333 }
1334
1335 static long __do_splice(struct file *in, loff_t __user *off_in,
1336                         struct file *out, loff_t __user *off_out,
1337                         size_t len, unsigned int flags)
1338 {
1339         struct pipe_inode_info *ipipe;
1340         struct pipe_inode_info *opipe;
1341         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1342         long ret;
1343
1344         ipipe = get_pipe_info(in, true);
1345         opipe = get_pipe_info(out, true);
1346
1347         if (ipipe) {
1348                 if (off_in)
1349                         return -ESPIPE;
1350                 pipe_clear_nowait(in);
1351         }
1352         if (opipe) {
1353                 if (off_out)
1354                         return -ESPIPE;
1355                 pipe_clear_nowait(out);
1356         }
1357
1358         if (off_out) {
1359                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1360                         return -EFAULT;
1361                 __off_out = &offset;
1362         }
1363         if (off_in) {
1364                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1365                         return -EFAULT;
1366                 __off_in = &offset;
1367         }
1368
1369         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1370         if (ret < 0)
1371                 return ret;
1372
1373         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1374                 return -EFAULT;
1375         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1376                 return -EFAULT;
1377
1378         return ret;
1379 }
1380
1381 static int iter_to_pipe(struct iov_iter *from,
1382                         struct pipe_inode_info *pipe,
1383                         unsigned flags)
1384 {
1385         struct pipe_buffer buf = {
1386                 .ops = &user_page_pipe_buf_ops,
1387                 .flags = flags
1388         };
1389         size_t total = 0;
1390         int ret = 0;
1391
1392         while (iov_iter_count(from)) {
1393                 struct page *pages[16];
1394                 ssize_t left;
1395                 size_t start;
1396                 int i, n;
1397
1398                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1399                 if (left <= 0) {
1400                         ret = left;
1401                         break;
1402                 }
1403
1404                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1405                 for (i = 0; i < n; i++) {
1406                         int size = min_t(int, left, PAGE_SIZE - start);
1407
1408                         buf.page = pages[i];
1409                         buf.offset = start;
1410                         buf.len = size;
1411                         ret = add_to_pipe(pipe, &buf);
1412                         if (unlikely(ret < 0)) {
1413                                 iov_iter_revert(from, left);
1414                                 // this one got dropped by add_to_pipe()
1415                                 while (++i < n)
1416                                         put_page(pages[i]);
1417                                 goto out;
1418                         }
1419                         total += ret;
1420                         left -= size;
1421                         start = 0;
1422                 }
1423         }
1424 out:
1425         return total ? total : ret;
1426 }
1427
1428 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1429                         struct splice_desc *sd)
1430 {
1431         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1432         return n == sd->len ? n : -EFAULT;
1433 }
1434
1435 /*
1436  * For lack of a better implementation, implement vmsplice() to userspace
1437  * as a simple copy of the pipes pages to the user iov.
1438  */
1439 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1440                              unsigned int flags)
1441 {
1442         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1443         struct splice_desc sd = {
1444                 .total_len = iov_iter_count(iter),
1445                 .flags = flags,
1446                 .u.data = iter
1447         };
1448         long ret = 0;
1449
1450         if (!pipe)
1451                 return -EBADF;
1452
1453         pipe_clear_nowait(file);
1454
1455         if (sd.total_len) {
1456                 pipe_lock(pipe);
1457                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1458                 pipe_unlock(pipe);
1459         }
1460
1461         if (ret > 0)
1462                 fsnotify_access(file);
1463
1464         return ret;
1465 }
1466
1467 /*
1468  * vmsplice splices a user address range into a pipe. It can be thought of
1469  * as splice-from-memory, where the regular splice is splice-from-file (or
1470  * to file). In both cases the output is a pipe, naturally.
1471  */
1472 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1473                              unsigned int flags)
1474 {
1475         struct pipe_inode_info *pipe;
1476         long ret = 0;
1477         unsigned buf_flag = 0;
1478
1479         if (flags & SPLICE_F_GIFT)
1480                 buf_flag = PIPE_BUF_FLAG_GIFT;
1481
1482         pipe = get_pipe_info(file, true);
1483         if (!pipe)
1484                 return -EBADF;
1485
1486         pipe_clear_nowait(file);
1487
1488         pipe_lock(pipe);
1489         ret = wait_for_space(pipe, flags);
1490         if (!ret)
1491                 ret = iter_to_pipe(iter, pipe, buf_flag);
1492         pipe_unlock(pipe);
1493         if (ret > 0) {
1494                 wakeup_pipe_readers(pipe);
1495                 fsnotify_modify(file);
1496         }
1497         return ret;
1498 }
1499
1500 static int vmsplice_type(struct fd f, int *type)
1501 {
1502         if (!f.file)
1503                 return -EBADF;
1504         if (f.file->f_mode & FMODE_WRITE) {
1505                 *type = ITER_SOURCE;
1506         } else if (f.file->f_mode & FMODE_READ) {
1507                 *type = ITER_DEST;
1508         } else {
1509                 fdput(f);
1510                 return -EBADF;
1511         }
1512         return 0;
1513 }
1514
1515 /*
1516  * Note that vmsplice only really supports true splicing _from_ user memory
1517  * to a pipe, not the other way around. Splicing from user memory is a simple
1518  * operation that can be supported without any funky alignment restrictions
1519  * or nasty vm tricks. We simply map in the user memory and fill them into
1520  * a pipe. The reverse isn't quite as easy, though. There are two possible
1521  * solutions for that:
1522  *
1523  *      - memcpy() the data internally, at which point we might as well just
1524  *        do a regular read() on the buffer anyway.
1525  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1526  *        has restriction limitations on both ends of the pipe).
1527  *
1528  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1529  *
1530  */
1531 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1532                 unsigned long, nr_segs, unsigned int, flags)
1533 {
1534         struct iovec iovstack[UIO_FASTIOV];
1535         struct iovec *iov = iovstack;
1536         struct iov_iter iter;
1537         ssize_t error;
1538         struct fd f;
1539         int type;
1540
1541         if (unlikely(flags & ~SPLICE_F_ALL))
1542                 return -EINVAL;
1543
1544         f = fdget(fd);
1545         error = vmsplice_type(f, &type);
1546         if (error)
1547                 return error;
1548
1549         error = import_iovec(type, uiov, nr_segs,
1550                              ARRAY_SIZE(iovstack), &iov, &iter);
1551         if (error < 0)
1552                 goto out_fdput;
1553
1554         if (!iov_iter_count(&iter))
1555                 error = 0;
1556         else if (type == ITER_SOURCE)
1557                 error = vmsplice_to_pipe(f.file, &iter, flags);
1558         else
1559                 error = vmsplice_to_user(f.file, &iter, flags);
1560
1561         kfree(iov);
1562 out_fdput:
1563         fdput(f);
1564         return error;
1565 }
1566
1567 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1568                 int, fd_out, loff_t __user *, off_out,
1569                 size_t, len, unsigned int, flags)
1570 {
1571         struct fd in, out;
1572         long error;
1573
1574         if (unlikely(!len))
1575                 return 0;
1576
1577         if (unlikely(flags & ~SPLICE_F_ALL))
1578                 return -EINVAL;
1579
1580         error = -EBADF;
1581         in = fdget(fd_in);
1582         if (in.file) {
1583                 out = fdget(fd_out);
1584                 if (out.file) {
1585                         error = __do_splice(in.file, off_in, out.file, off_out,
1586                                                 len, flags);
1587                         fdput(out);
1588                 }
1589                 fdput(in);
1590         }
1591         return error;
1592 }
1593
1594 /*
1595  * Make sure there's data to read. Wait for input if we can, otherwise
1596  * return an appropriate error.
1597  */
1598 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1599 {
1600         int ret;
1601
1602         /*
1603          * Check the pipe occupancy without the inode lock first. This function
1604          * is speculative anyways, so missing one is ok.
1605          */
1606         if (!pipe_empty(pipe->head, pipe->tail))
1607                 return 0;
1608
1609         ret = 0;
1610         pipe_lock(pipe);
1611
1612         while (pipe_empty(pipe->head, pipe->tail)) {
1613                 if (signal_pending(current)) {
1614                         ret = -ERESTARTSYS;
1615                         break;
1616                 }
1617                 if (!pipe->writers)
1618                         break;
1619                 if (flags & SPLICE_F_NONBLOCK) {
1620                         ret = -EAGAIN;
1621                         break;
1622                 }
1623                 pipe_wait_readable(pipe);
1624         }
1625
1626         pipe_unlock(pipe);
1627         return ret;
1628 }
1629
1630 /*
1631  * Make sure there's writeable room. Wait for room if we can, otherwise
1632  * return an appropriate error.
1633  */
1634 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1635 {
1636         int ret;
1637
1638         /*
1639          * Check pipe occupancy without the inode lock first. This function
1640          * is speculative anyways, so missing one is ok.
1641          */
1642         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1643                 return 0;
1644
1645         ret = 0;
1646         pipe_lock(pipe);
1647
1648         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1649                 if (!pipe->readers) {
1650                         send_sig(SIGPIPE, current, 0);
1651                         ret = -EPIPE;
1652                         break;
1653                 }
1654                 if (flags & SPLICE_F_NONBLOCK) {
1655                         ret = -EAGAIN;
1656                         break;
1657                 }
1658                 if (signal_pending(current)) {
1659                         ret = -ERESTARTSYS;
1660                         break;
1661                 }
1662                 pipe_wait_writable(pipe);
1663         }
1664
1665         pipe_unlock(pipe);
1666         return ret;
1667 }
1668
1669 /*
1670  * Splice contents of ipipe to opipe.
1671  */
1672 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1673                                struct pipe_inode_info *opipe,
1674                                size_t len, unsigned int flags)
1675 {
1676         struct pipe_buffer *ibuf, *obuf;
1677         unsigned int i_head, o_head;
1678         unsigned int i_tail, o_tail;
1679         unsigned int i_mask, o_mask;
1680         int ret = 0;
1681         bool input_wakeup = false;
1682
1683
1684 retry:
1685         ret = ipipe_prep(ipipe, flags);
1686         if (ret)
1687                 return ret;
1688
1689         ret = opipe_prep(opipe, flags);
1690         if (ret)
1691                 return ret;
1692
1693         /*
1694          * Potential ABBA deadlock, work around it by ordering lock
1695          * grabbing by pipe info address. Otherwise two different processes
1696          * could deadlock (one doing tee from A -> B, the other from B -> A).
1697          */
1698         pipe_double_lock(ipipe, opipe);
1699
1700         i_tail = ipipe->tail;
1701         i_mask = ipipe->ring_size - 1;
1702         o_head = opipe->head;
1703         o_mask = opipe->ring_size - 1;
1704
1705         do {
1706                 size_t o_len;
1707
1708                 if (!opipe->readers) {
1709                         send_sig(SIGPIPE, current, 0);
1710                         if (!ret)
1711                                 ret = -EPIPE;
1712                         break;
1713                 }
1714
1715                 i_head = ipipe->head;
1716                 o_tail = opipe->tail;
1717
1718                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1719                         break;
1720
1721                 /*
1722                  * Cannot make any progress, because either the input
1723                  * pipe is empty or the output pipe is full.
1724                  */
1725                 if (pipe_empty(i_head, i_tail) ||
1726                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1727                         /* Already processed some buffers, break */
1728                         if (ret)
1729                                 break;
1730
1731                         if (flags & SPLICE_F_NONBLOCK) {
1732                                 ret = -EAGAIN;
1733                                 break;
1734                         }
1735
1736                         /*
1737                          * We raced with another reader/writer and haven't
1738                          * managed to process any buffers.  A zero return
1739                          * value means EOF, so retry instead.
1740                          */
1741                         pipe_unlock(ipipe);
1742                         pipe_unlock(opipe);
1743                         goto retry;
1744                 }
1745
1746                 ibuf = &ipipe->bufs[i_tail & i_mask];
1747                 obuf = &opipe->bufs[o_head & o_mask];
1748
1749                 if (len >= ibuf->len) {
1750                         /*
1751                          * Simply move the whole buffer from ipipe to opipe
1752                          */
1753                         *obuf = *ibuf;
1754                         ibuf->ops = NULL;
1755                         i_tail++;
1756                         ipipe->tail = i_tail;
1757                         input_wakeup = true;
1758                         o_len = obuf->len;
1759                         o_head++;
1760                         opipe->head = o_head;
1761                 } else {
1762                         /*
1763                          * Get a reference to this pipe buffer,
1764                          * so we can copy the contents over.
1765                          */
1766                         if (!pipe_buf_get(ipipe, ibuf)) {
1767                                 if (ret == 0)
1768                                         ret = -EFAULT;
1769                                 break;
1770                         }
1771                         *obuf = *ibuf;
1772
1773                         /*
1774                          * Don't inherit the gift and merge flags, we need to
1775                          * prevent multiple steals of this page.
1776                          */
1777                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1778                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1779
1780                         obuf->len = len;
1781                         ibuf->offset += len;
1782                         ibuf->len -= len;
1783                         o_len = len;
1784                         o_head++;
1785                         opipe->head = o_head;
1786                 }
1787                 ret += o_len;
1788                 len -= o_len;
1789         } while (len);
1790
1791         pipe_unlock(ipipe);
1792         pipe_unlock(opipe);
1793
1794         /*
1795          * If we put data in the output pipe, wakeup any potential readers.
1796          */
1797         if (ret > 0)
1798                 wakeup_pipe_readers(opipe);
1799
1800         if (input_wakeup)
1801                 wakeup_pipe_writers(ipipe);
1802
1803         return ret;
1804 }
1805
1806 /*
1807  * Link contents of ipipe to opipe.
1808  */
1809 static int link_pipe(struct pipe_inode_info *ipipe,
1810                      struct pipe_inode_info *opipe,
1811                      size_t len, unsigned int flags)
1812 {
1813         struct pipe_buffer *ibuf, *obuf;
1814         unsigned int i_head, o_head;
1815         unsigned int i_tail, o_tail;
1816         unsigned int i_mask, o_mask;
1817         int ret = 0;
1818
1819         /*
1820          * Potential ABBA deadlock, work around it by ordering lock
1821          * grabbing by pipe info address. Otherwise two different processes
1822          * could deadlock (one doing tee from A -> B, the other from B -> A).
1823          */
1824         pipe_double_lock(ipipe, opipe);
1825
1826         i_tail = ipipe->tail;
1827         i_mask = ipipe->ring_size - 1;
1828         o_head = opipe->head;
1829         o_mask = opipe->ring_size - 1;
1830
1831         do {
1832                 if (!opipe->readers) {
1833                         send_sig(SIGPIPE, current, 0);
1834                         if (!ret)
1835                                 ret = -EPIPE;
1836                         break;
1837                 }
1838
1839                 i_head = ipipe->head;
1840                 o_tail = opipe->tail;
1841
1842                 /*
1843                  * If we have iterated all input buffers or run out of
1844                  * output room, break.
1845                  */
1846                 if (pipe_empty(i_head, i_tail) ||
1847                     pipe_full(o_head, o_tail, opipe->max_usage))
1848                         break;
1849
1850                 ibuf = &ipipe->bufs[i_tail & i_mask];
1851                 obuf = &opipe->bufs[o_head & o_mask];
1852
1853                 /*
1854                  * Get a reference to this pipe buffer,
1855                  * so we can copy the contents over.
1856                  */
1857                 if (!pipe_buf_get(ipipe, ibuf)) {
1858                         if (ret == 0)
1859                                 ret = -EFAULT;
1860                         break;
1861                 }
1862
1863                 *obuf = *ibuf;
1864
1865                 /*
1866                  * Don't inherit the gift and merge flag, we need to prevent
1867                  * multiple steals of this page.
1868                  */
1869                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1870                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1871
1872                 if (obuf->len > len)
1873                         obuf->len = len;
1874                 ret += obuf->len;
1875                 len -= obuf->len;
1876
1877                 o_head++;
1878                 opipe->head = o_head;
1879                 i_tail++;
1880         } while (len);
1881
1882         pipe_unlock(ipipe);
1883         pipe_unlock(opipe);
1884
1885         /*
1886          * If we put data in the output pipe, wakeup any potential readers.
1887          */
1888         if (ret > 0)
1889                 wakeup_pipe_readers(opipe);
1890
1891         return ret;
1892 }
1893
1894 /*
1895  * This is a tee(1) implementation that works on pipes. It doesn't copy
1896  * any data, it simply references the 'in' pages on the 'out' pipe.
1897  * The 'flags' used are the SPLICE_F_* variants, currently the only
1898  * applicable one is SPLICE_F_NONBLOCK.
1899  */
1900 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1901 {
1902         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1903         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1904         int ret = -EINVAL;
1905
1906         if (unlikely(!(in->f_mode & FMODE_READ) ||
1907                      !(out->f_mode & FMODE_WRITE)))
1908                 return -EBADF;
1909
1910         /*
1911          * Duplicate the contents of ipipe to opipe without actually
1912          * copying the data.
1913          */
1914         if (ipipe && opipe && ipipe != opipe) {
1915                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1916                         flags |= SPLICE_F_NONBLOCK;
1917
1918                 /*
1919                  * Keep going, unless we encounter an error. The ipipe/opipe
1920                  * ordering doesn't really matter.
1921                  */
1922                 ret = ipipe_prep(ipipe, flags);
1923                 if (!ret) {
1924                         ret = opipe_prep(opipe, flags);
1925                         if (!ret)
1926                                 ret = link_pipe(ipipe, opipe, len, flags);
1927                 }
1928         }
1929
1930         if (ret > 0) {
1931                 fsnotify_access(in);
1932                 fsnotify_modify(out);
1933         }
1934
1935         return ret;
1936 }
1937
1938 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1939 {
1940         struct fd in, out;
1941         int error;
1942
1943         if (unlikely(flags & ~SPLICE_F_ALL))
1944                 return -EINVAL;
1945
1946         if (unlikely(!len))
1947                 return 0;
1948
1949         error = -EBADF;
1950         in = fdget(fdin);
1951         if (in.file) {
1952                 out = fdget(fdout);
1953                 if (out.file) {
1954                         error = do_tee(in.file, out.file, len, flags);
1955                         fdput(out);
1956                 }
1957                 fdput(in);
1958         }
1959
1960         return error;
1961 }