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