fs/splice.c

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