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 <[email protected]>
  17  * Copyright (C) 2005-2006 Linus Torvalds <[email protected]>
  18  * Copyright (C) 2006 Ingo Molnar <[email protected]>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/fsnotify.h>
  34 #include <linux/security.h>
  35 #include <linux/gfp.h>
  36 #include <linux/net.h>
  37 #include <linux/socket.h>
  38 #include <linux/sched/signal.h>
  39
  40 #include "internal.h"
  41
  42 /*
  43  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  44  * indicate they support non-blocking reads or writes, we must clear it
  45  * here if set to avoid blocking other users of this pipe if splice is
  46  * being done on it.
  47  */
  48 static noinline void noinline pipe_clear_nowait(struct file *file)
  49 {
  50         fmode_t fmode = READ_ONCE(file->f_mode);
  51
  52         do {
  53                 if (!(fmode & FMODE_NOWAIT))
  54                         break;
  55         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  56 }
  57
  58 /*
  59  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  60  * a vm helper function, it's already simplified quite a bit by the
  61  * addition of remove_mapping(). If success is returned, the caller may
  62  * attempt to reuse this page for another destination.
  63  */
  64 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  65                 struct pipe_buffer *buf)
  66 {
  67         struct folio *folio = page_folio(buf->page);
  68         struct address_space *mapping;
  69
  70         folio_lock(folio);
  71
  72         mapping = folio_mapping(folio);
  73         if (mapping) {
  74                 WARN_ON(!folio_test_uptodate(folio));
  75
  76                 /*
  77                  * At least for ext2 with nobh option, we need to wait on
  78                  * writeback completing on this folio, since we'll remove it
  79                  * from the pagecache.  Otherwise truncate wont wait on the
  80                  * folio, allowing the disk blocks to be reused by someone else
  81                  * before we actually wrote our data to them. fs corruption
  82                  * ensues.
  83                  */
  84                 folio_wait_writeback(folio);
  85
  86                 if (!filemap_release_folio(folio, GFP_KERNEL))
  87                         goto out_unlock;
  88
  89                 /*
  90                  * If we succeeded in removing the mapping, set LRU flag
  91                  * and return good.
  92                  */
  93                 if (remove_mapping(mapping, folio)) {
  94                         buf->flags |= PIPE_BUF_FLAG_LRU;
  95                         return true;
  96                 }
  97         }
  98
  99         /*
 100          * Raced with truncate or failed to remove folio from current
 101          * address space, unlock and return failure.
 102          */
 103 out_unlock:
 104         folio_unlock(folio);
 105         return false;
 106 }
 107
 108 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 109                                         struct pipe_buffer *buf)
 110 {
 111         put_page(buf->page);
 112         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 113 }
 114
 115 /*
 116  * Check whether the contents of buf is OK to access. Since the content
 117  * is a page cache page, IO may be in flight.
 118  */
 119 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 120                                        struct pipe_buffer *buf)
 121 {
 122         struct folio *folio = page_folio(buf->page);
 123         int err;
 124
 125         if (!folio_test_uptodate(folio)) {
 126                 folio_lock(folio);
 127
 128                 /*
 129                  * Folio got truncated/unhashed. This will cause a 0-byte
 130                  * splice, if this is the first page.
 131                  */
 132                 if (!folio->mapping) {
 133                         err = -ENODATA;
 134                         goto error;
 135                 }
 136
 137                 /*
 138                  * Uh oh, read-error from disk.
 139                  */
 140                 if (!folio_test_uptodate(folio)) {
 141                         err = -EIO;
 142                         goto error;
 143                 }
 144
 145                 /* Folio is ok after all, we are done */
 146                 folio_unlock(folio);
 147         }
 148
 149         return 0;
 150 error:
 151         folio_unlock(folio);
 152         return err;
 153 }
 154
 155 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 156         .confirm        = page_cache_pipe_buf_confirm,
 157         .release        = page_cache_pipe_buf_release,
 158         .try_steal      = page_cache_pipe_buf_try_steal,
 159         .get            = generic_pipe_buf_get,
 160 };
 161
 162 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 163                 struct pipe_buffer *buf)
 164 {
 165         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 166                 return false;
 167
 168         buf->flags |= PIPE_BUF_FLAG_LRU;
 169         return generic_pipe_buf_try_steal(pipe, buf);
 170 }
 171
 172 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 173         .release        = page_cache_pipe_buf_release,
 174         .try_steal      = user_page_pipe_buf_try_steal,
 175         .get            = generic_pipe_buf_get,
 176 };
 177
 178 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 179 {
 180         smp_mb();
 181         if (waitqueue_active(&pipe->rd_wait))
 182                 wake_up_interruptible(&pipe->rd_wait);
 183         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 184 }
 185
 186 /**
 187  * splice_to_pipe - fill passed data into a pipe
 188  * @pipe:       pipe to fill
 189  * @spd:        data to fill
 190  *
 191  * Description:
 192  *    @spd contains a map of pages and len/offset tuples, along with
 193  *    the struct pipe_buf_operations associated with these pages. This
 194  *    function will link that data to the pipe.
 195  *
 196  */
 197 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 198                        struct splice_pipe_desc *spd)
 199 {
 200         unsigned int spd_pages = spd->nr_pages;
 201         unsigned int tail = pipe->tail;
 202         unsigned int head = pipe->head;
 203         unsigned int mask = pipe->ring_size - 1;
 204         ssize_t ret = 0;
 205         int 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 = in->f_op->read_iter(&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;
 678         ssize_t ret;
 679
 680         if (!out->f_op->write_iter)
 681                 return -EINVAL;
 682
 683         array = kcalloc(nbufs, sizeof(struct bio_vec), GFP_KERNEL);
 684         if (unlikely(!array))
 685                 return -ENOMEM;
 686
 687         pipe_lock(pipe);
 688
 689         splice_from_pipe_begin(&sd);
 690         while (sd.total_len) {
 691                 struct kiocb kiocb;
 692                 struct iov_iter from;
 693                 unsigned int head, tail, mask;
 694                 size_t left;
 695                 int n;
 696
 697                 ret = splice_from_pipe_next(pipe, &sd);
 698                 if (ret <= 0)
 699                         break;
 700
 701                 if (unlikely(nbufs < pipe->max_usage)) {
 702                         kfree(array);
 703                         nbufs = pipe->max_usage;
 704                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 705                                         GFP_KERNEL);
 706                         if (!array) {
 707                                 ret = -ENOMEM;
 708                                 break;
 709                         }
 710                 }
 711
 712                 head = pipe->head;
 713                 tail = pipe->tail;
 714                 mask = pipe->ring_size - 1;
 715
 716                 /* build the vector */
 717                 left = sd.total_len;
 718                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 719                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 720                         size_t this_len = buf->len;
 721
 722                         /* zero-length bvecs are not supported, skip them */
 723                         if (!this_len)
 724                                 continue;
 725                         this_len = min(this_len, left);
 726
 727                         ret = pipe_buf_confirm(pipe, buf);
 728                         if (unlikely(ret)) {
 729                                 if (ret == -ENODATA)
 730                                         ret = 0;
 731                                 goto done;
 732                         }
 733
 734                         bvec_set_page(&array[n], buf->page, this_len,
 735                                       buf->offset);
 736                         left -= this_len;
 737                         n++;
 738                 }
 739
 740                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 741                 init_sync_kiocb(&kiocb, out);
 742                 kiocb.ki_pos = sd.pos;
 743                 ret = out->f_op->write_iter(&kiocb, &from);
 744                 sd.pos = kiocb.ki_pos;
 745                 if (ret <= 0)
 746                         break;
 747
 748                 sd.num_spliced += ret;
 749                 sd.total_len -= ret;
 750                 *ppos = sd.pos;
 751
 752                 /* dismiss the fully eaten buffers, adjust the partial one */
 753                 tail = pipe->tail;
 754                 while (ret) {
 755                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 756                         if (ret >= buf->len) {
 757                                 ret -= buf->len;
 758                                 buf->len = 0;
 759                                 pipe_buf_release(pipe, buf);
 760                                 tail++;
 761                                 pipe->tail = tail;
 762                                 if (pipe->files)
 763                                         sd.need_wakeup = true;
 764                         } else {
 765                                 buf->offset += ret;
 766                                 buf->len -= ret;
 767                                 ret = 0;
 768                         }
 769                 }
 770         }
 771 done:
 772         kfree(array);
 773         splice_from_pipe_end(pipe, &sd);
 774
 775         pipe_unlock(pipe);
 776
 777         if (sd.num_spliced)
 778                 ret = sd.num_spliced;
 779
 780         return ret;
 781 }
 782
 783 EXPORT_SYMBOL(iter_file_splice_write);
 784
 785 #ifdef CONFIG_NET
 786 /**
 787  * splice_to_socket - splice data from a pipe to a socket
 788  * @pipe:       pipe to splice from
 789  * @out:        socket to write to
 790  * @ppos:       position in @out
 791  * @len:        number of bytes to splice
 792  * @flags:      splice modifier flags
 793  *
 794  * Description:
 795  *    Will send @len bytes from the pipe to a network socket. No data copying
 796  *    is involved.
 797  *
 798  */
 799 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 800                          loff_t *ppos, size_t len, unsigned int flags)
 801 {
 802         struct socket *sock = sock_from_file(out);
 803         struct bio_vec bvec[16];
 804         struct msghdr msg = {};
 805         ssize_t ret = 0;
 806         size_t spliced = 0;
 807         bool need_wakeup = false;
 808
 809         pipe_lock(pipe);
 810
 811         while (len > 0) {
 812                 unsigned int head, tail, mask, bc = 0;
 813                 size_t remain = len;
 814
 815                 /*
 816                  * Check for signal early to make process killable when there
 817                  * are always buffers available
 818                  */
 819                 ret = -ERESTARTSYS;
 820                 if (signal_pending(current))
 821                         break;
 822
 823                 while (pipe_empty(pipe->head, pipe->tail)) {
 824                         ret = 0;
 825                         if (!pipe->writers)
 826                                 goto out;
 827
 828                         if (spliced)
 829                                 goto out;
 830
 831                         ret = -EAGAIN;
 832                         if (flags & SPLICE_F_NONBLOCK)
 833                                 goto out;
 834
 835                         ret = -ERESTARTSYS;
 836                         if (signal_pending(current))
 837                                 goto out;
 838
 839                         if (need_wakeup) {
 840                                 wakeup_pipe_writers(pipe);
 841                                 need_wakeup = false;
 842                         }
 843
 844                         pipe_wait_readable(pipe);
 845                 }
 846
 847                 head = pipe->head;
 848                 tail = pipe->tail;
 849                 mask = pipe->ring_size - 1;
 850
 851                 while (!pipe_empty(head, tail)) {
 852                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 853                         size_t seg;
 854
 855                         if (!buf->len) {
 856                                 tail++;
 857                                 continue;
 858                         }
 859
 860                         seg = min_t(size_t, remain, buf->len);
 861
 862                         ret = pipe_buf_confirm(pipe, buf);
 863                         if (unlikely(ret)) {
 864                                 if (ret == -ENODATA)
 865                                         ret = 0;
 866                                 break;
 867                         }
 868
 869                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 870                         remain -= seg;
 871                         if (remain == 0 || bc >= ARRAY_SIZE(bvec))
 872                                 break;
 873                         tail++;
 874                 }
 875
 876                 if (!bc)
 877                         break;
 878
 879                 msg.msg_flags = MSG_SPLICE_PAGES;
 880                 if (flags & SPLICE_F_MORE)
 881                         msg.msg_flags |= MSG_MORE;
 882                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 883                         msg.msg_flags |= MSG_MORE;
 884                 if (out->f_flags & O_NONBLOCK)
 885                         msg.msg_flags |= MSG_DONTWAIT;
 886
 887                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 888                               len - remain);
 889                 ret = sock_sendmsg(sock, &msg);
 890                 if (ret <= 0)
 891                         break;
 892
 893                 spliced += ret;
 894                 len -= ret;
 895                 tail = pipe->tail;
 896                 while (ret > 0) {
 897                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 898                         size_t seg = min_t(size_t, ret, buf->len);
 899
 900                         buf->offset += seg;
 901                         buf->len -= seg;
 902                         ret -= seg;
 903
 904                         if (!buf->len) {
 905                                 pipe_buf_release(pipe, buf);
 906                                 tail++;
 907                         }
 908                 }
 909
 910                 if (tail != pipe->tail) {
 911                         pipe->tail = tail;
 912                         if (pipe->files)
 913                                 need_wakeup = true;
 914                 }
 915         }
 916
 917 out:
 918         pipe_unlock(pipe);
 919         if (need_wakeup)
 920                 wakeup_pipe_writers(pipe);
 921         return spliced ?: ret;
 922 }
 923 #endif
 924
 925 static int warn_unsupported(struct file *file, const char *op)
 926 {
 927         pr_debug_ratelimited(
 928                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 929                 op, file, current->pid, current->comm);
 930         return -EINVAL;
 931 }
 932
 933 /*
 934  * Attempt to initiate a splice from pipe to file.
 935  */
 936 static ssize_t do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 937                               loff_t *ppos, size_t len, unsigned int flags)
 938 {
 939         if (unlikely(!out->f_op->splice_write))
 940                 return warn_unsupported(out, "write");
 941         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 942 }
 943
 944 /*
 945  * Indicate to the caller that there was a premature EOF when reading from the
 946  * source and the caller didn't indicate they would be sending more data after
 947  * this.
 948  */
 949 static void do_splice_eof(struct splice_desc *sd)
 950 {
 951         if (sd->splice_eof)
 952                 sd->splice_eof(sd);
 953 }
 954
 955 /*
 956  * Callers already called rw_verify_area() on the entire range.
 957  * No need to call it for sub ranges.
 958  */
 959 static ssize_t do_splice_read(struct file *in, loff_t *ppos,
 960                               struct pipe_inode_info *pipe, size_t len,
 961                               unsigned int flags)
 962 {
 963         unsigned int p_space;
 964
 965         if (unlikely(!(in->f_mode & FMODE_READ)))
 966                 return -EBADF;
 967         if (!len)
 968                 return 0;
 969
 970         /* Don't try to read more the pipe has space for. */
 971         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 972         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 973
 974         if (unlikely(len > MAX_RW_COUNT))
 975                 len = MAX_RW_COUNT;
 976
 977         if (unlikely(!in->f_op->splice_read))
 978                 return warn_unsupported(in, "read");
 979         /*
 980          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 981          * buffer, copy into it and splice that into the pipe.
 982          */
 983         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 984                 return copy_splice_read(in, ppos, pipe, len, flags);
 985         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 986 }
 987
 988 /**
 989  * vfs_splice_read - Read data from a file and splice it into a pipe
 990  * @in:         File to splice from
 991  * @ppos:       Input file offset
 992  * @pipe:       Pipe to splice to
 993  * @len:        Number of bytes to splice
 994  * @flags:      Splice modifier flags (SPLICE_F_*)
 995  *
 996  * Splice the requested amount of data from the input file to the pipe.  This
 997  * is synchronous as the caller must hold the pipe lock across the entire
 998  * operation.
 999  *
1000  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
1001  * a hole and a negative error code otherwise.
1002  */
1003 ssize_t vfs_splice_read(struct file *in, loff_t *ppos,
1004                         struct pipe_inode_info *pipe, size_t len,
1005                         unsigned int flags)
1006 {
1007         ssize_t ret;
1008
1009         ret = rw_verify_area(READ, in, ppos, len);
1010         if (unlikely(ret < 0))
1011                 return ret;
1012
1013         return do_splice_read(in, ppos, pipe, len, flags);
1014 }
1015 EXPORT_SYMBOL_GPL(vfs_splice_read);
1016
1017 /**
1018  * splice_direct_to_actor - splices data directly between two non-pipes
1019  * @in:         file to splice from
1020  * @sd:         actor information on where to splice to
1021  * @actor:      handles the data splicing
1022  *
1023  * Description:
1024  *    This is a special case helper to splice directly between two
1025  *    points, without requiring an explicit pipe. Internally an allocated
1026  *    pipe is cached in the process, and reused during the lifetime of
1027  *    that process.
1028  *
1029  */
1030 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1031                                splice_direct_actor *actor)
1032 {
1033         struct pipe_inode_info *pipe;
1034         ssize_t ret, bytes;
1035         size_t len;
1036         int i, flags, more;
1037
1038         /*
1039          * We require the input to be seekable, as we don't want to randomly
1040          * drop data for eg socket -> socket splicing. Use the piped splicing
1041          * for that!
1042          */
1043         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1044                 return -EINVAL;
1045
1046         /*
1047          * neither in nor out is a pipe, setup an internal pipe attached to
1048          * 'out' and transfer the wanted data from 'in' to 'out' through that
1049          */
1050         pipe = current->splice_pipe;
1051         if (unlikely(!pipe)) {
1052                 pipe = alloc_pipe_info();
1053                 if (!pipe)
1054                         return -ENOMEM;
1055
1056                 /*
1057                  * We don't have an immediate reader, but we'll read the stuff
1058                  * out of the pipe right after the splice_to_pipe(). So set
1059                  * PIPE_READERS appropriately.
1060                  */
1061                 pipe->readers = 1;
1062
1063                 current->splice_pipe = pipe;
1064         }
1065
1066         /*
1067          * Do the splice.
1068          */
1069         bytes = 0;
1070         len = sd->total_len;
1071
1072         /* Don't block on output, we have to drain the direct pipe. */
1073         flags = sd->flags;
1074         sd->flags &= ~SPLICE_F_NONBLOCK;
1075
1076         /*
1077          * We signal MORE until we've read sufficient data to fulfill the
1078          * request and we keep signalling it if the caller set it.
1079          */
1080         more = sd->flags & SPLICE_F_MORE;
1081         sd->flags |= SPLICE_F_MORE;
1082
1083         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1084
1085         while (len) {
1086                 size_t read_len;
1087                 loff_t pos = sd->pos, prev_pos = pos;
1088
1089                 ret = do_splice_read(in, &pos, pipe, len, flags);
1090                 if (unlikely(ret <= 0))
1091                         goto read_failure;
1092
1093                 read_len = ret;
1094                 sd->total_len = read_len;
1095
1096                 /*
1097                  * If we now have sufficient data to fulfill the request then
1098                  * we clear SPLICE_F_MORE if it was not set initially.
1099                  */
1100                 if (read_len >= len && !more)
1101                         sd->flags &= ~SPLICE_F_MORE;
1102
1103                 /*
1104                  * NOTE: nonblocking mode only applies to the input. We
1105                  * must not do the output in nonblocking mode as then we
1106                  * could get stuck data in the internal pipe:
1107                  */
1108                 ret = actor(pipe, sd);
1109                 if (unlikely(ret <= 0)) {
1110                         sd->pos = prev_pos;
1111                         goto out_release;
1112                 }
1113
1114                 bytes += ret;
1115                 len -= ret;
1116                 sd->pos = pos;
1117
1118                 if (ret < read_len) {
1119                         sd->pos = prev_pos + ret;
1120                         goto out_release;
1121                 }
1122         }
1123
1124 done:
1125         pipe->tail = pipe->head = 0;
1126         file_accessed(in);
1127         return bytes;
1128
1129 read_failure:
1130         /*
1131          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1132          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1133          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1134          * least 1 byte *then* we will also do the ->splice_eof() call.
1135          */
1136         if (ret == 0 && !more && len > 0 && bytes)
1137                 do_splice_eof(sd);
1138 out_release:
1139         /*
1140          * If we did an incomplete transfer we must release
1141          * the pipe buffers in question:
1142          */
1143         for (i = 0; i < pipe->ring_size; i++) {
1144                 struct pipe_buffer *buf = &pipe->bufs[i];
1145
1146                 if (buf->ops)
1147                         pipe_buf_release(pipe, buf);
1148         }
1149
1150         if (!bytes)
1151                 bytes = ret;
1152
1153         goto done;
1154 }
1155 EXPORT_SYMBOL(splice_direct_to_actor);
1156
1157 static int direct_splice_actor(struct pipe_inode_info *pipe,
1158                                struct splice_desc *sd)
1159 {
1160         struct file *file = sd->u.file;
1161         long ret;
1162
1163         file_start_write(file);
1164         ret = do_splice_from(pipe, file, sd->opos, sd->total_len, sd->flags);
1165         file_end_write(file);
1166         return ret;
1167 }
1168
1169 static int splice_file_range_actor(struct pipe_inode_info *pipe,
1170                                         struct splice_desc *sd)
1171 {
1172         struct file *file = sd->u.file;
1173
1174         return do_splice_from(pipe, file, sd->opos, sd->total_len, sd->flags);
1175 }
1176
1177 static void direct_file_splice_eof(struct splice_desc *sd)
1178 {
1179         struct file *file = sd->u.file;
1180
1181         if (file->f_op->splice_eof)
1182                 file->f_op->splice_eof(file);
1183 }
1184
1185 static ssize_t do_splice_direct_actor(struct file *in, loff_t *ppos,
1186                                       struct file *out, loff_t *opos,
1187                                       size_t len, unsigned int flags,
1188                                       splice_direct_actor *actor)
1189 {
1190         struct splice_desc sd = {
1191                 .len            = len,
1192                 .total_len      = len,
1193                 .flags          = flags,
1194                 .pos            = *ppos,
1195                 .u.file         = out,
1196                 .splice_eof     = direct_file_splice_eof,
1197                 .opos           = opos,
1198         };
1199         ssize_t ret;
1200
1201         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1202                 return -EBADF;
1203
1204         if (unlikely(out->f_flags & O_APPEND))
1205                 return -EINVAL;
1206
1207         ret = splice_direct_to_actor(in, &sd, actor);
1208         if (ret > 0)
1209                 *ppos = sd.pos;
1210
1211         return ret;
1212 }
1213 /**
1214  * do_splice_direct - splices data directly between two files
1215  * @in:         file to splice from
1216  * @ppos:       input file offset
1217  * @out:        file to splice to
1218  * @opos:       output file offset
1219  * @len:        number of bytes to splice
1220  * @flags:      splice modifier flags
1221  *
1222  * Description:
1223  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1224  *    doing it in the application would incur an extra system call
1225  *    (splice in + splice out, as compared to just sendfile()). So this helper
1226  *    can splice directly through a process-private pipe.
1227  *
1228  * Callers already called rw_verify_area() on the entire range.
1229  */
1230 ssize_t do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1231                          loff_t *opos, size_t len, unsigned int flags)
1232 {
1233         return do_splice_direct_actor(in, ppos, out, opos, len, flags,
1234                                       direct_splice_actor);
1235 }
1236 EXPORT_SYMBOL(do_splice_direct);
1237
1238 /**
1239  * splice_file_range - splices data between two files for copy_file_range()
1240  * @in:         file to splice from
1241  * @ppos:       input file offset
1242  * @out:        file to splice to
1243  * @opos:       output file offset
1244  * @len:        number of bytes to splice
1245  *
1246  * Description:
1247  *    For use by ->copy_file_range() methods.
1248  *    Like do_splice_direct(), but vfs_copy_file_range() already holds
1249  *    start_file_write() on @out file.
1250  *
1251  * Callers already called rw_verify_area() on the entire range.
1252  */
1253 ssize_t splice_file_range(struct file *in, loff_t *ppos, struct file *out,
1254                           loff_t *opos, size_t len)
1255 {
1256         lockdep_assert(file_write_started(out));
1257
1258         return do_splice_direct_actor(in, ppos, out, opos,
1259                                       min_t(size_t, len, MAX_RW_COUNT),
1260                                       0, splice_file_range_actor);
1261 }
1262 EXPORT_SYMBOL(splice_file_range);
1263
1264 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1265 {
1266         for (;;) {
1267                 if (unlikely(!pipe->readers)) {
1268                         send_sig(SIGPIPE, current, 0);
1269                         return -EPIPE;
1270                 }
1271                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1272                         return 0;
1273                 if (flags & SPLICE_F_NONBLOCK)
1274                         return -EAGAIN;
1275                 if (signal_pending(current))
1276                         return -ERESTARTSYS;
1277                 pipe_wait_writable(pipe);
1278         }
1279 }
1280
1281 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1282                                struct pipe_inode_info *opipe,
1283                                size_t len, unsigned int flags);
1284
1285 ssize_t splice_file_to_pipe(struct file *in,
1286                             struct pipe_inode_info *opipe,
1287                             loff_t *offset,
1288                             size_t len, unsigned int flags)
1289 {
1290         ssize_t ret;
1291
1292         pipe_lock(opipe);
1293         ret = wait_for_space(opipe, flags);
1294         if (!ret)
1295                 ret = do_splice_read(in, offset, opipe, len, flags);
1296         pipe_unlock(opipe);
1297         if (ret > 0)
1298                 wakeup_pipe_readers(opipe);
1299         return ret;
1300 }
1301
1302 /*
1303  * Determine where to splice to/from.
1304  */
1305 ssize_t do_splice(struct file *in, loff_t *off_in, struct file *out,
1306                   loff_t *off_out, size_t len, unsigned int flags)
1307 {
1308         struct pipe_inode_info *ipipe;
1309         struct pipe_inode_info *opipe;
1310         loff_t offset;
1311         ssize_t ret;
1312
1313         if (unlikely(!(in->f_mode & FMODE_READ) ||
1314                      !(out->f_mode & FMODE_WRITE)))
1315                 return -EBADF;
1316
1317         ipipe = get_pipe_info(in, true);
1318         opipe = get_pipe_info(out, true);
1319
1320         if (ipipe && opipe) {
1321                 if (off_in || off_out)
1322                         return -ESPIPE;
1323
1324                 /* Splicing to self would be fun, but... */
1325                 if (ipipe == opipe)
1326                         return -EINVAL;
1327
1328                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1329                         flags |= SPLICE_F_NONBLOCK;
1330
1331                 ret = splice_pipe_to_pipe(ipipe, opipe, len, flags);
1332         } else if (ipipe) {
1333                 if (off_in)
1334                         return -ESPIPE;
1335                 if (off_out) {
1336                         if (!(out->f_mode & FMODE_PWRITE))
1337                                 return -EINVAL;
1338                         offset = *off_out;
1339                 } else {
1340                         offset = out->f_pos;
1341                 }
1342
1343                 if (unlikely(out->f_flags & O_APPEND))
1344                         return -EINVAL;
1345
1346                 ret = rw_verify_area(WRITE, out, &offset, len);
1347                 if (unlikely(ret < 0))
1348                         return ret;
1349
1350                 if (in->f_flags & O_NONBLOCK)
1351                         flags |= SPLICE_F_NONBLOCK;
1352
1353                 file_start_write(out);
1354                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1355                 file_end_write(out);
1356
1357                 if (!off_out)
1358                         out->f_pos = offset;
1359                 else
1360                         *off_out = offset;
1361         } else if (opipe) {
1362                 if (off_out)
1363                         return -ESPIPE;
1364                 if (off_in) {
1365                         if (!(in->f_mode & FMODE_PREAD))
1366                                 return -EINVAL;
1367                         offset = *off_in;
1368                 } else {
1369                         offset = in->f_pos;
1370                 }
1371
1372                 ret = rw_verify_area(READ, in, &offset, len);
1373                 if (unlikely(ret < 0))
1374                         return ret;
1375
1376                 if (out->f_flags & O_NONBLOCK)
1377                         flags |= SPLICE_F_NONBLOCK;
1378
1379                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1380
1381                 if (!off_in)
1382                         in->f_pos = offset;
1383                 else
1384                         *off_in = offset;
1385         } else {
1386                 ret = -EINVAL;
1387         }
1388
1389         if (ret > 0) {
1390                 /*
1391                  * Generate modify out before access in:
1392                  * do_splice_from() may've already sent modify out,
1393                  * and this ensures the events get merged.
1394                  */
1395                 fsnotify_modify(out);
1396                 fsnotify_access(in);
1397         }
1398
1399         return ret;
1400 }
1401
1402 static ssize_t __do_splice(struct file *in, loff_t __user *off_in,
1403                            struct file *out, loff_t __user *off_out,
1404                            size_t len, unsigned int flags)
1405 {
1406         struct pipe_inode_info *ipipe;
1407         struct pipe_inode_info *opipe;
1408         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1409         ssize_t ret;
1410
1411         ipipe = get_pipe_info(in, true);
1412         opipe = get_pipe_info(out, true);
1413
1414         if (ipipe) {
1415                 if (off_in)
1416                         return -ESPIPE;
1417                 pipe_clear_nowait(in);
1418         }
1419         if (opipe) {
1420                 if (off_out)
1421                         return -ESPIPE;
1422                 pipe_clear_nowait(out);
1423         }
1424
1425         if (off_out) {
1426                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1427                         return -EFAULT;
1428                 __off_out = &offset;
1429         }
1430         if (off_in) {
1431                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1432                         return -EFAULT;
1433                 __off_in = &offset;
1434         }
1435
1436         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1437         if (ret < 0)
1438                 return ret;
1439
1440         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1441                 return -EFAULT;
1442         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1443                 return -EFAULT;
1444
1445         return ret;
1446 }
1447
1448 static ssize_t iter_to_pipe(struct iov_iter *from,
1449                             struct pipe_inode_info *pipe,
1450                             unsigned int flags)
1451 {
1452         struct pipe_buffer buf = {
1453                 .ops = &user_page_pipe_buf_ops,
1454                 .flags = flags
1455         };
1456         size_t total = 0;
1457         ssize_t ret = 0;
1458
1459         while (iov_iter_count(from)) {
1460                 struct page *pages[16];
1461                 ssize_t left;
1462                 size_t start;
1463                 int i, n;
1464
1465                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1466                 if (left <= 0) {
1467                         ret = left;
1468                         break;
1469                 }
1470
1471                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1472                 for (i = 0; i < n; i++) {
1473                         int size = min_t(int, left, PAGE_SIZE - start);
1474
1475                         buf.page = pages[i];
1476                         buf.offset = start;
1477                         buf.len = size;
1478                         ret = add_to_pipe(pipe, &buf);
1479                         if (unlikely(ret < 0)) {
1480                                 iov_iter_revert(from, left);
1481                                 // this one got dropped by add_to_pipe()
1482                                 while (++i < n)
1483                                         put_page(pages[i]);
1484                                 goto out;
1485                         }
1486                         total += ret;
1487                         left -= size;
1488                         start = 0;
1489                 }
1490         }
1491 out:
1492         return total ? total : ret;
1493 }
1494
1495 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1496                         struct splice_desc *sd)
1497 {
1498         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1499         return n == sd->len ? n : -EFAULT;
1500 }
1501
1502 /*
1503  * For lack of a better implementation, implement vmsplice() to userspace
1504  * as a simple copy of the pipes pages to the user iov.
1505  */
1506 static ssize_t vmsplice_to_user(struct file *file, struct iov_iter *iter,
1507                                 unsigned int flags)
1508 {
1509         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1510         struct splice_desc sd = {
1511                 .total_len = iov_iter_count(iter),
1512                 .flags = flags,
1513                 .u.data = iter
1514         };
1515         ssize_t ret = 0;
1516
1517         if (!pipe)
1518                 return -EBADF;
1519
1520         pipe_clear_nowait(file);
1521
1522         if (sd.total_len) {
1523                 pipe_lock(pipe);
1524                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1525                 pipe_unlock(pipe);
1526         }
1527
1528         if (ret > 0)
1529                 fsnotify_access(file);
1530
1531         return ret;
1532 }
1533
1534 /*
1535  * vmsplice splices a user address range into a pipe. It can be thought of
1536  * as splice-from-memory, where the regular splice is splice-from-file (or
1537  * to file). In both cases the output is a pipe, naturally.
1538  */
1539 static ssize_t vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1540                                 unsigned int flags)
1541 {
1542         struct pipe_inode_info *pipe;
1543         ssize_t ret = 0;
1544         unsigned buf_flag = 0;
1545
1546         if (flags & SPLICE_F_GIFT)
1547                 buf_flag = PIPE_BUF_FLAG_GIFT;
1548
1549         pipe = get_pipe_info(file, true);
1550         if (!pipe)
1551                 return -EBADF;
1552
1553         pipe_clear_nowait(file);
1554
1555         pipe_lock(pipe);
1556         ret = wait_for_space(pipe, flags);
1557         if (!ret)
1558                 ret = iter_to_pipe(iter, pipe, buf_flag);
1559         pipe_unlock(pipe);
1560         if (ret > 0) {
1561                 wakeup_pipe_readers(pipe);
1562                 fsnotify_modify(file);
1563         }
1564         return ret;
1565 }
1566
1567 /*
1568  * Note that vmsplice only really supports true splicing _from_ user memory
1569  * to a pipe, not the other way around. Splicing from user memory is a simple
1570  * operation that can be supported without any funky alignment restrictions
1571  * or nasty vm tricks. We simply map in the user memory and fill them into
1572  * a pipe. The reverse isn't quite as easy, though. There are two possible
1573  * solutions for that:
1574  *
1575  *      - memcpy() the data internally, at which point we might as well just
1576  *        do a regular read() on the buffer anyway.
1577  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1578  *        has restriction limitations on both ends of the pipe).
1579  *
1580  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1581  *
1582  */
1583 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1584                 unsigned long, nr_segs, unsigned int, flags)
1585 {
1586         struct iovec iovstack[UIO_FASTIOV];
1587         struct iovec *iov = iovstack;
1588         struct iov_iter iter;
1589         ssize_t error;
1590         int type;
1591
1592         if (unlikely(flags & ~SPLICE_F_ALL))
1593                 return -EINVAL;
1594
1595         CLASS(fd, f)(fd);
1596         if (fd_empty(f))
1597                 return -EBADF;
1598         if (fd_file(f)->f_mode & FMODE_WRITE)
1599                 type = ITER_SOURCE;
1600         else if (fd_file(f)->f_mode & FMODE_READ)
1601                 type = ITER_DEST;
1602         else
1603                 return -EBADF;
1604
1605         error = import_iovec(type, uiov, nr_segs,
1606                              ARRAY_SIZE(iovstack), &iov, &iter);
1607         if (error < 0)
1608                 return error;
1609
1610         if (!iov_iter_count(&iter))
1611                 error = 0;
1612         else if (type == ITER_SOURCE)
1613                 error = vmsplice_to_pipe(fd_file(f), &iter, flags);
1614         else
1615                 error = vmsplice_to_user(fd_file(f), &iter, flags);
1616
1617         kfree(iov);
1618         return error;
1619 }
1620
1621 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1622                 int, fd_out, loff_t __user *, off_out,
1623                 size_t, len, unsigned int, flags)
1624 {
1625         if (unlikely(!len))
1626                 return 0;
1627
1628         if (unlikely(flags & ~SPLICE_F_ALL))
1629                 return -EINVAL;
1630
1631         CLASS(fd, in)(fd_in);
1632         if (fd_empty(in))
1633                 return -EBADF;
1634
1635         CLASS(fd, out)(fd_out);
1636         if (fd_empty(out))
1637                 return -EBADF;
1638
1639         return __do_splice(fd_file(in), off_in, fd_file(out), off_out,
1640                                             len, flags);
1641 }
1642
1643 /*
1644  * Make sure there's data to read. Wait for input if we can, otherwise
1645  * return an appropriate error.
1646  */
1647 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1648 {
1649         int ret;
1650
1651         /*
1652          * Check the pipe occupancy without the inode lock first. This function
1653          * is speculative anyways, so missing one is ok.
1654          */
1655         if (!pipe_empty(pipe->head, pipe->tail))
1656                 return 0;
1657
1658         ret = 0;
1659         pipe_lock(pipe);
1660
1661         while (pipe_empty(pipe->head, pipe->tail)) {
1662                 if (signal_pending(current)) {
1663                         ret = -ERESTARTSYS;
1664                         break;
1665                 }
1666                 if (!pipe->writers)
1667                         break;
1668                 if (flags & SPLICE_F_NONBLOCK) {
1669                         ret = -EAGAIN;
1670                         break;
1671                 }
1672                 pipe_wait_readable(pipe);
1673         }
1674
1675         pipe_unlock(pipe);
1676         return ret;
1677 }
1678
1679 /*
1680  * Make sure there's writeable room. Wait for room if we can, otherwise
1681  * return an appropriate error.
1682  */
1683 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1684 {
1685         int ret;
1686
1687         /*
1688          * Check pipe occupancy without the inode lock first. This function
1689          * is speculative anyways, so missing one is ok.
1690          */
1691         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1692                 return 0;
1693
1694         ret = 0;
1695         pipe_lock(pipe);
1696
1697         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1698                 if (!pipe->readers) {
1699                         send_sig(SIGPIPE, current, 0);
1700                         ret = -EPIPE;
1701                         break;
1702                 }
1703                 if (flags & SPLICE_F_NONBLOCK) {
1704                         ret = -EAGAIN;
1705                         break;
1706                 }
1707                 if (signal_pending(current)) {
1708                         ret = -ERESTARTSYS;
1709                         break;
1710                 }
1711                 pipe_wait_writable(pipe);
1712         }
1713
1714         pipe_unlock(pipe);
1715         return ret;
1716 }
1717
1718 /*
1719  * Splice contents of ipipe to opipe.
1720  */
1721 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1722                                struct pipe_inode_info *opipe,
1723                                size_t len, unsigned int flags)
1724 {
1725         struct pipe_buffer *ibuf, *obuf;
1726         unsigned int i_head, o_head;
1727         unsigned int i_tail, o_tail;
1728         unsigned int i_mask, o_mask;
1729         int ret = 0;
1730         bool input_wakeup = false;
1731
1732
1733 retry:
1734         ret = ipipe_prep(ipipe, flags);
1735         if (ret)
1736                 return ret;
1737
1738         ret = opipe_prep(opipe, flags);
1739         if (ret)
1740                 return ret;
1741
1742         /*
1743          * Potential ABBA deadlock, work around it by ordering lock
1744          * grabbing by pipe info address. Otherwise two different processes
1745          * could deadlock (one doing tee from A -> B, the other from B -> A).
1746          */
1747         pipe_double_lock(ipipe, opipe);
1748
1749         i_tail = ipipe->tail;
1750         i_mask = ipipe->ring_size - 1;
1751         o_head = opipe->head;
1752         o_mask = opipe->ring_size - 1;
1753
1754         do {
1755                 size_t o_len;
1756
1757                 if (!opipe->readers) {
1758                         send_sig(SIGPIPE, current, 0);
1759                         if (!ret)
1760                                 ret = -EPIPE;
1761                         break;
1762                 }
1763
1764                 i_head = ipipe->head;
1765                 o_tail = opipe->tail;
1766
1767                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1768                         break;
1769
1770                 /*
1771                  * Cannot make any progress, because either the input
1772                  * pipe is empty or the output pipe is full.
1773                  */
1774                 if (pipe_empty(i_head, i_tail) ||
1775                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1776                         /* Already processed some buffers, break */
1777                         if (ret)
1778                                 break;
1779
1780                         if (flags & SPLICE_F_NONBLOCK) {
1781                                 ret = -EAGAIN;
1782                                 break;
1783                         }
1784
1785                         /*
1786                          * We raced with another reader/writer and haven't
1787                          * managed to process any buffers.  A zero return
1788                          * value means EOF, so retry instead.
1789                          */
1790                         pipe_unlock(ipipe);
1791                         pipe_unlock(opipe);
1792                         goto retry;
1793                 }
1794
1795                 ibuf = &ipipe->bufs[i_tail & i_mask];
1796                 obuf = &opipe->bufs[o_head & o_mask];
1797
1798                 if (len >= ibuf->len) {
1799                         /*
1800                          * Simply move the whole buffer from ipipe to opipe
1801                          */
1802                         *obuf = *ibuf;
1803                         ibuf->ops = NULL;
1804                         i_tail++;
1805                         ipipe->tail = i_tail;
1806                         input_wakeup = true;
1807                         o_len = obuf->len;
1808                         o_head++;
1809                         opipe->head = o_head;
1810                 } else {
1811                         /*
1812                          * Get a reference to this pipe buffer,
1813                          * so we can copy the contents over.
1814                          */
1815                         if (!pipe_buf_get(ipipe, ibuf)) {
1816                                 if (ret == 0)
1817                                         ret = -EFAULT;
1818                                 break;
1819                         }
1820                         *obuf = *ibuf;
1821
1822                         /*
1823                          * Don't inherit the gift and merge flags, we need to
1824                          * prevent multiple steals of this page.
1825                          */
1826                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1827                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1828
1829                         obuf->len = len;
1830                         ibuf->offset += len;
1831                         ibuf->len -= len;
1832                         o_len = len;
1833                         o_head++;
1834                         opipe->head = o_head;
1835                 }
1836                 ret += o_len;
1837                 len -= o_len;
1838         } while (len);
1839
1840         pipe_unlock(ipipe);
1841         pipe_unlock(opipe);
1842
1843         /*
1844          * If we put data in the output pipe, wakeup any potential readers.
1845          */
1846         if (ret > 0)
1847                 wakeup_pipe_readers(opipe);
1848
1849         if (input_wakeup)
1850                 wakeup_pipe_writers(ipipe);
1851
1852         return ret;
1853 }
1854
1855 /*
1856  * Link contents of ipipe to opipe.
1857  */
1858 static ssize_t link_pipe(struct pipe_inode_info *ipipe,
1859                          struct pipe_inode_info *opipe,
1860                          size_t len, unsigned int flags)
1861 {
1862         struct pipe_buffer *ibuf, *obuf;
1863         unsigned int i_head, o_head;
1864         unsigned int i_tail, o_tail;
1865         unsigned int i_mask, o_mask;
1866         ssize_t ret = 0;
1867
1868         /*
1869          * Potential ABBA deadlock, work around it by ordering lock
1870          * grabbing by pipe info address. Otherwise two different processes
1871          * could deadlock (one doing tee from A -> B, the other from B -> A).
1872          */
1873         pipe_double_lock(ipipe, opipe);
1874
1875         i_tail = ipipe->tail;
1876         i_mask = ipipe->ring_size - 1;
1877         o_head = opipe->head;
1878         o_mask = opipe->ring_size - 1;
1879
1880         do {
1881                 if (!opipe->readers) {
1882                         send_sig(SIGPIPE, current, 0);
1883                         if (!ret)
1884                                 ret = -EPIPE;
1885                         break;
1886                 }
1887
1888                 i_head = ipipe->head;
1889                 o_tail = opipe->tail;
1890
1891                 /*
1892                  * If we have iterated all input buffers or run out of
1893                  * output room, break.
1894                  */
1895                 if (pipe_empty(i_head, i_tail) ||
1896                     pipe_full(o_head, o_tail, opipe->max_usage))
1897                         break;
1898
1899                 ibuf = &ipipe->bufs[i_tail & i_mask];
1900                 obuf = &opipe->bufs[o_head & o_mask];
1901
1902                 /*
1903                  * Get a reference to this pipe buffer,
1904                  * so we can copy the contents over.
1905                  */
1906                 if (!pipe_buf_get(ipipe, ibuf)) {
1907                         if (ret == 0)
1908                                 ret = -EFAULT;
1909                         break;
1910                 }
1911
1912                 *obuf = *ibuf;
1913
1914                 /*
1915                  * Don't inherit the gift and merge flag, we need to prevent
1916                  * multiple steals of this page.
1917                  */
1918                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1919                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1920
1921                 if (obuf->len > len)
1922                         obuf->len = len;
1923                 ret += obuf->len;
1924                 len -= obuf->len;
1925
1926                 o_head++;
1927                 opipe->head = o_head;
1928                 i_tail++;
1929         } while (len);
1930
1931         pipe_unlock(ipipe);
1932         pipe_unlock(opipe);
1933
1934         /*
1935          * If we put data in the output pipe, wakeup any potential readers.
1936          */
1937         if (ret > 0)
1938                 wakeup_pipe_readers(opipe);
1939
1940         return ret;
1941 }
1942
1943 /*
1944  * This is a tee(1) implementation that works on pipes. It doesn't copy
1945  * any data, it simply references the 'in' pages on the 'out' pipe.
1946  * The 'flags' used are the SPLICE_F_* variants, currently the only
1947  * applicable one is SPLICE_F_NONBLOCK.
1948  */
1949 ssize_t do_tee(struct file *in, struct file *out, size_t len,
1950                unsigned int flags)
1951 {
1952         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1953         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1954         ssize_t ret = -EINVAL;
1955
1956         if (unlikely(!(in->f_mode & FMODE_READ) ||
1957                      !(out->f_mode & FMODE_WRITE)))
1958                 return -EBADF;
1959
1960         /*
1961          * Duplicate the contents of ipipe to opipe without actually
1962          * copying the data.
1963          */
1964         if (ipipe && opipe && ipipe != opipe) {
1965                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1966                         flags |= SPLICE_F_NONBLOCK;
1967
1968                 /*
1969                  * Keep going, unless we encounter an error. The ipipe/opipe
1970                  * ordering doesn't really matter.
1971                  */
1972                 ret = ipipe_prep(ipipe, flags);
1973                 if (!ret) {
1974                         ret = opipe_prep(opipe, flags);
1975                         if (!ret)
1976                                 ret = link_pipe(ipipe, opipe, len, flags);
1977                 }
1978         }
1979
1980         if (ret > 0) {
1981                 fsnotify_access(in);
1982                 fsnotify_modify(out);
1983         }
1984
1985         return ret;
1986 }
1987
1988 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1989 {
1990         if (unlikely(flags & ~SPLICE_F_ALL))
1991                 return -EINVAL;
1992
1993         if (unlikely(!len))
1994                 return 0;
1995
1996         CLASS(fd, in)(fdin);
1997         if (fd_empty(in))
1998                 return -EBADF;
1999
2000         CLASS(fd, out)(fdout);
2001         if (fd_empty(out))
2002                 return -EBADF;
2003
2004         return do_tee(fd_file(in), fd_file(out), len, flags);
2005 }