fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <[email protected]>
  17  * Copyright (C) 2005-2006 Linus Torvalds <[email protected]>
  18  * Copyright (C) 2006 Ingo Molnar <[email protected]>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/security.h>
  34 #include <linux/gfp.h>
  35 #include <linux/socket.h>
  36 #include <linux/compat.h>
  37 #include <linux/sched/signal.h>
  38
  39 #include "internal.h"
  40
  41 /*
  42  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  43  * a vm helper function, it's already simplified quite a bit by the
  44  * addition of remove_mapping(). If success is returned, the caller may
  45  * attempt to reuse this page for another destination.
  46  */
  47 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  48                 struct pipe_buffer *buf)
  49 {
  50         struct page *page = buf->page;
  51         struct address_space *mapping;
  52
  53         lock_page(page);
  54
  55         mapping = page_mapping(page);
  56         if (mapping) {
  57                 WARN_ON(!PageUptodate(page));
  58
  59                 /*
  60                  * At least for ext2 with nobh option, we need to wait on
  61                  * writeback completing on this page, since we'll remove it
  62                  * from the pagecache.  Otherwise truncate wont wait on the
  63                  * page, allowing the disk blocks to be reused by someone else
  64                  * before we actually wrote our data to them. fs corruption
  65                  * ensues.
  66                  */
  67                 wait_on_page_writeback(page);
  68
  69                 if (page_has_private(page) &&
  70                     !try_to_release_page(page, GFP_KERNEL))
  71                         goto out_unlock;
  72
  73                 /*
  74                  * If we succeeded in removing the mapping, set LRU flag
  75                  * and return good.
  76                  */
  77                 if (remove_mapping(mapping, page)) {
  78                         buf->flags |= PIPE_BUF_FLAG_LRU;
  79                         return true;
  80                 }
  81         }
  82
  83         /*
  84          * Raced with truncate or failed to remove page from current
  85          * address space, unlock and return failure.
  86          */
  87 out_unlock:
  88         unlock_page(page);
  89         return false;
  90 }
  91
  92 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  93                                         struct pipe_buffer *buf)
  94 {
  95         put_page(buf->page);
  96         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  97 }
  98
  99 /*
 100  * Check whether the contents of buf is OK to access. Since the content
 101  * is a page cache page, IO may be in flight.
 102  */
 103 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 104                                        struct pipe_buffer *buf)
 105 {
 106         struct page *page = buf->page;
 107         int err;
 108
 109         if (!PageUptodate(page)) {
 110                 lock_page(page);
 111
 112                 /*
 113                  * Page got truncated/unhashed. This will cause a 0-byte
 114                  * splice, if this is the first page.
 115                  */
 116                 if (!page->mapping) {
 117                         err = -ENODATA;
 118                         goto error;
 119                 }
 120
 121                 /*
 122                  * Uh oh, read-error from disk.
 123                  */
 124                 if (!PageUptodate(page)) {
 125                         err = -EIO;
 126                         goto error;
 127                 }
 128
 129                 /*
 130                  * Page is ok afterall, we are done.
 131                  */
 132                 unlock_page(page);
 133         }
 134
 135         return 0;
 136 error:
 137         unlock_page(page);
 138         return err;
 139 }
 140
 141 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 142         .confirm        = page_cache_pipe_buf_confirm,
 143         .release        = page_cache_pipe_buf_release,
 144         .try_steal      = page_cache_pipe_buf_try_steal,
 145         .get            = generic_pipe_buf_get,
 146 };
 147
 148 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 149                 struct pipe_buffer *buf)
 150 {
 151         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 152                 return false;
 153
 154         buf->flags |= PIPE_BUF_FLAG_LRU;
 155         return generic_pipe_buf_try_steal(pipe, buf);
 156 }
 157
 158 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 159         .release        = page_cache_pipe_buf_release,
 160         .try_steal      = user_page_pipe_buf_try_steal,
 161         .get            = generic_pipe_buf_get,
 162 };
 163
 164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 165 {
 166         smp_mb();
 167         if (waitqueue_active(&pipe->rd_wait))
 168                 wake_up_interruptible(&pipe->rd_wait);
 169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 170 }
 171
 172 /**
 173  * splice_to_pipe - fill passed data into a pipe
 174  * @pipe:       pipe to fill
 175  * @spd:        data to fill
 176  *
 177  * Description:
 178  *    @spd contains a map of pages and len/offset tuples, along with
 179  *    the struct pipe_buf_operations associated with these pages. This
 180  *    function will link that data to the pipe.
 181  *
 182  */
 183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 184                        struct splice_pipe_desc *spd)
 185 {
 186         unsigned int spd_pages = spd->nr_pages;
 187         unsigned int tail = pipe->tail;
 188         unsigned int head = pipe->head;
 189         unsigned int mask = pipe->ring_size - 1;
 190         int ret = 0, page_nr = 0;
 191
 192         if (!spd_pages)
 193                 return 0;
 194
 195         if (unlikely(!pipe->readers)) {
 196                 send_sig(SIGPIPE, current, 0);
 197                 ret = -EPIPE;
 198                 goto out;
 199         }
 200
 201         while (!pipe_full(head, tail, pipe->max_usage)) {
 202                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 203
 204                 buf->page = spd->pages[page_nr];
 205                 buf->offset = spd->partial[page_nr].offset;
 206                 buf->len = spd->partial[page_nr].len;
 207                 buf->private = spd->partial[page_nr].private;
 208                 buf->ops = spd->ops;
 209                 buf->flags = 0;
 210
 211                 head++;
 212                 pipe->head = head;
 213                 page_nr++;
 214                 ret += buf->len;
 215
 216                 if (!--spd->nr_pages)
 217                         break;
 218         }
 219
 220         if (!ret)
 221                 ret = -EAGAIN;
 222
 223 out:
 224         while (page_nr < spd_pages)
 225                 spd->spd_release(spd, page_nr++);
 226
 227         return ret;
 228 }
 229 EXPORT_SYMBOL_GPL(splice_to_pipe);
 230
 231 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 232 {
 233         unsigned int head = pipe->head;
 234         unsigned int tail = pipe->tail;
 235         unsigned int mask = pipe->ring_size - 1;
 236         int ret;
 237
 238         if (unlikely(!pipe->readers)) {
 239                 send_sig(SIGPIPE, current, 0);
 240                 ret = -EPIPE;
 241         } else if (pipe_full(head, tail, pipe->max_usage)) {
 242                 ret = -EAGAIN;
 243         } else {
 244                 pipe->bufs[head & mask] = *buf;
 245                 pipe->head = head + 1;
 246                 return buf->len;
 247         }
 248         pipe_buf_release(pipe, buf);
 249         return ret;
 250 }
 251 EXPORT_SYMBOL(add_to_pipe);
 252
 253 /*
 254  * Check if we need to grow the arrays holding pages and partial page
 255  * descriptions.
 256  */
 257 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 258 {
 259         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 260
 261         spd->nr_pages_max = max_usage;
 262         if (max_usage <= PIPE_DEF_BUFFERS)
 263                 return 0;
 264
 265         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 266         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 267                                      GFP_KERNEL);
 268
 269         if (spd->pages && spd->partial)
 270                 return 0;
 271
 272         kfree(spd->pages);
 273         kfree(spd->partial);
 274         return -ENOMEM;
 275 }
 276
 277 void splice_shrink_spd(struct splice_pipe_desc *spd)
 278 {
 279         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 280                 return;
 281
 282         kfree(spd->pages);
 283         kfree(spd->partial);
 284 }
 285
 286 /**
 287  * generic_file_splice_read - splice data from file to a pipe
 288  * @in:         file to splice from
 289  * @ppos:       position in @in
 290  * @pipe:       pipe to splice to
 291  * @len:        number of bytes to splice
 292  * @flags:      splice modifier flags
 293  *
 294  * Description:
 295  *    Will read pages from given file and fill them into a pipe. Can be
 296  *    used as long as it has more or less sane ->read_iter().
 297  *
 298  */
 299 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 300                                  struct pipe_inode_info *pipe, size_t len,
 301                                  unsigned int flags)
 302 {
 303         struct iov_iter to;
 304         struct kiocb kiocb;
 305         unsigned int i_head;
 306         int ret;
 307
 308         iov_iter_pipe(&to, READ, pipe, len);
 309         i_head = to.head;
 310         init_sync_kiocb(&kiocb, in);
 311         kiocb.ki_pos = *ppos;
 312         ret = call_read_iter(in, &kiocb, &to);
 313         if (ret > 0) {
 314                 *ppos = kiocb.ki_pos;
 315                 file_accessed(in);
 316         } else if (ret < 0) {
 317                 to.head = i_head;
 318                 to.iov_offset = 0;
 319                 iov_iter_advance(&to, 0); /* to free what was emitted */
 320                 /*
 321                  * callers of ->splice_read() expect -EAGAIN on
 322                  * "can't put anything in there", rather than -EFAULT.
 323                  */
 324                 if (ret == -EFAULT)
 325                         ret = -EAGAIN;
 326         }
 327
 328         return ret;
 329 }
 330 EXPORT_SYMBOL(generic_file_splice_read);
 331
 332 const struct pipe_buf_operations default_pipe_buf_ops = {
 333         .release        = generic_pipe_buf_release,
 334         .try_steal      = generic_pipe_buf_try_steal,
 335         .get            = generic_pipe_buf_get,
 336 };
 337
 338 /* Pipe buffer operations for a socket and similar. */
 339 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 340         .release        = generic_pipe_buf_release,
 341         .get            = generic_pipe_buf_get,
 342 };
 343 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 344
 345 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 346                             unsigned long vlen, loff_t offset)
 347 {
 348         mm_segment_t old_fs;
 349         loff_t pos = offset;
 350         ssize_t res;
 351
 352         old_fs = get_fs();
 353         set_fs(KERNEL_DS);
 354         /* The cast to a user pointer is valid due to the set_fs() */
 355         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 356         set_fs(old_fs);
 357
 358         return res;
 359 }
 360
 361 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 362                                  struct pipe_inode_info *pipe, size_t len,
 363                                  unsigned int flags)
 364 {
 365         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 366         struct iov_iter to;
 367         struct page **pages;
 368         unsigned int nr_pages;
 369         unsigned int mask;
 370         size_t offset, base, copied = 0;
 371         ssize_t res;
 372         int i;
 373
 374         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 375                 return -EAGAIN;
 376
 377         /*
 378          * Try to keep page boundaries matching to source pagecache ones -
 379          * it probably won't be much help, but...
 380          */
 381         offset = *ppos & ~PAGE_MASK;
 382
 383         iov_iter_pipe(&to, READ, pipe, len + offset);
 384
 385         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 386         if (res <= 0)
 387                 return -ENOMEM;
 388
 389         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 390
 391         vec = __vec;
 392         if (nr_pages > PIPE_DEF_BUFFERS) {
 393                 vec = kmalloc_array(nr_pages, sizeof(struct kvec), GFP_KERNEL);
 394                 if (unlikely(!vec)) {
 395                         res = -ENOMEM;
 396                         goto out;
 397                 }
 398         }
 399
 400         mask = pipe->ring_size - 1;
 401         pipe->bufs[to.head & mask].offset = offset;
 402         pipe->bufs[to.head & mask].len -= offset;
 403
 404         for (i = 0; i < nr_pages; i++) {
 405                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 406                 vec[i].iov_base = page_address(pages[i]) + offset;
 407                 vec[i].iov_len = this_len;
 408                 len -= this_len;
 409                 offset = 0;
 410         }
 411
 412         res = kernel_readv(in, vec, nr_pages, *ppos);
 413         if (res > 0) {
 414                 copied = res;
 415                 *ppos += res;
 416         }
 417
 418         if (vec != __vec)
 419                 kfree(vec);
 420 out:
 421         for (i = 0; i < nr_pages; i++)
 422                 put_page(pages[i]);
 423         kvfree(pages);
 424         iov_iter_advance(&to, copied);  /* truncates and discards */
 425         return res;
 426 }
 427
 428 /*
 429  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 430  * using sendpage(). Return the number of bytes sent.
 431  */
 432 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 433                             struct pipe_buffer *buf, struct splice_desc *sd)
 434 {
 435         struct file *file = sd->u.file;
 436         loff_t pos = sd->pos;
 437         int more;
 438
 439         if (!likely(file->f_op->sendpage))
 440                 return -EINVAL;
 441
 442         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 443
 444         if (sd->len < sd->total_len &&
 445             pipe_occupancy(pipe->head, pipe->tail) > 1)
 446                 more |= MSG_SENDPAGE_NOTLAST;
 447
 448         return file->f_op->sendpage(file, buf->page, buf->offset,
 449                                     sd->len, &pos, more);
 450 }
 451
 452 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 453 {
 454         smp_mb();
 455         if (waitqueue_active(&pipe->wr_wait))
 456                 wake_up_interruptible(&pipe->wr_wait);
 457         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 458 }
 459
 460 /**
 461  * splice_from_pipe_feed - feed available data from a pipe to a file
 462  * @pipe:       pipe to splice from
 463  * @sd:         information to @actor
 464  * @actor:      handler that splices the data
 465  *
 466  * Description:
 467  *    This function loops over the pipe and calls @actor to do the
 468  *    actual moving of a single struct pipe_buffer to the desired
 469  *    destination.  It returns when there's no more buffers left in
 470  *    the pipe or if the requested number of bytes (@sd->total_len)
 471  *    have been copied.  It returns a positive number (one) if the
 472  *    pipe needs to be filled with more data, zero if the required
 473  *    number of bytes have been copied and -errno on error.
 474  *
 475  *    This, together with splice_from_pipe_{begin,end,next}, may be
 476  *    used to implement the functionality of __splice_from_pipe() when
 477  *    locking is required around copying the pipe buffers to the
 478  *    destination.
 479  */
 480 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 481                           splice_actor *actor)
 482 {
 483         unsigned int head = pipe->head;
 484         unsigned int tail = pipe->tail;
 485         unsigned int mask = pipe->ring_size - 1;
 486         int ret;
 487
 488         while (!pipe_empty(head, tail)) {
 489                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 490
 491                 sd->len = buf->len;
 492                 if (sd->len > sd->total_len)
 493                         sd->len = sd->total_len;
 494
 495                 ret = pipe_buf_confirm(pipe, buf);
 496                 if (unlikely(ret)) {
 497                         if (ret == -ENODATA)
 498                                 ret = 0;
 499                         return ret;
 500                 }
 501
 502                 ret = actor(pipe, buf, sd);
 503                 if (ret <= 0)
 504                         return ret;
 505
 506                 buf->offset += ret;
 507                 buf->len -= ret;
 508
 509                 sd->num_spliced += ret;
 510                 sd->len -= ret;
 511                 sd->pos += ret;
 512                 sd->total_len -= ret;
 513
 514                 if (!buf->len) {
 515                         pipe_buf_release(pipe, buf);
 516                         tail++;
 517                         pipe->tail = tail;
 518                         if (pipe->files)
 519                                 sd->need_wakeup = true;
 520                 }
 521
 522                 if (!sd->total_len)
 523                         return 0;
 524         }
 525
 526         return 1;
 527 }
 528
 529 /* We know we have a pipe buffer, but maybe it's empty? */
 530 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 531 {
 532         unsigned int tail = pipe->tail;
 533         unsigned int mask = pipe->ring_size - 1;
 534         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 535
 536         if (unlikely(!buf->len)) {
 537                 pipe_buf_release(pipe, buf);
 538                 pipe->tail = tail+1;
 539                 return true;
 540         }
 541
 542         return false;
 543 }
 544
 545 /**
 546  * splice_from_pipe_next - wait for some data to splice from
 547  * @pipe:       pipe to splice from
 548  * @sd:         information about the splice operation
 549  *
 550  * Description:
 551  *    This function will wait for some data and return a positive
 552  *    value (one) if pipe buffers are available.  It will return zero
 553  *    or -errno if no more data needs to be spliced.
 554  */
 555 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 556 {
 557         /*
 558          * Check for signal early to make process killable when there are
 559          * always buffers available
 560          */
 561         if (signal_pending(current))
 562                 return -ERESTARTSYS;
 563
 564 repeat:
 565         while (pipe_empty(pipe->head, pipe->tail)) {
 566                 if (!pipe->writers)
 567                         return 0;
 568
 569                 if (sd->num_spliced)
 570                         return 0;
 571
 572                 if (sd->flags & SPLICE_F_NONBLOCK)
 573                         return -EAGAIN;
 574
 575                 if (signal_pending(current))
 576                         return -ERESTARTSYS;
 577
 578                 if (sd->need_wakeup) {
 579                         wakeup_pipe_writers(pipe);
 580                         sd->need_wakeup = false;
 581                 }
 582
 583                 pipe_wait_readable(pipe);
 584         }
 585
 586         if (eat_empty_buffer(pipe))
 587                 goto repeat;
 588
 589         return 1;
 590 }
 591
 592 /**
 593  * splice_from_pipe_begin - start splicing from pipe
 594  * @sd:         information about the splice operation
 595  *
 596  * Description:
 597  *    This function should be called before a loop containing
 598  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 599  *    initialize the necessary fields of @sd.
 600  */
 601 static void splice_from_pipe_begin(struct splice_desc *sd)
 602 {
 603         sd->num_spliced = 0;
 604         sd->need_wakeup = false;
 605 }
 606
 607 /**
 608  * splice_from_pipe_end - finish splicing from pipe
 609  * @pipe:       pipe to splice from
 610  * @sd:         information about the splice operation
 611  *
 612  * Description:
 613  *    This function will wake up pipe writers if necessary.  It should
 614  *    be called after a loop containing splice_from_pipe_next() and
 615  *    splice_from_pipe_feed().
 616  */
 617 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 618 {
 619         if (sd->need_wakeup)
 620                 wakeup_pipe_writers(pipe);
 621 }
 622
 623 /**
 624  * __splice_from_pipe - splice data from a pipe to given actor
 625  * @pipe:       pipe to splice from
 626  * @sd:         information to @actor
 627  * @actor:      handler that splices the data
 628  *
 629  * Description:
 630  *    This function does little more than loop over the pipe and call
 631  *    @actor to do the actual moving of a single struct pipe_buffer to
 632  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 633  *    pipe_to_user.
 634  *
 635  */
 636 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 637                            splice_actor *actor)
 638 {
 639         int ret;
 640
 641         splice_from_pipe_begin(sd);
 642         do {
 643                 cond_resched();
 644                 ret = splice_from_pipe_next(pipe, sd);
 645                 if (ret > 0)
 646                         ret = splice_from_pipe_feed(pipe, sd, actor);
 647         } while (ret > 0);
 648         splice_from_pipe_end(pipe, sd);
 649
 650         return sd->num_spliced ? sd->num_spliced : ret;
 651 }
 652 EXPORT_SYMBOL(__splice_from_pipe);
 653
 654 /**
 655  * splice_from_pipe - splice data from a pipe to a file
 656  * @pipe:       pipe to splice from
 657  * @out:        file to splice to
 658  * @ppos:       position in @out
 659  * @len:        how many bytes to splice
 660  * @flags:      splice modifier flags
 661  * @actor:      handler that splices the data
 662  *
 663  * Description:
 664  *    See __splice_from_pipe. This function locks the pipe inode,
 665  *    otherwise it's identical to __splice_from_pipe().
 666  *
 667  */
 668 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 669                          loff_t *ppos, size_t len, unsigned int flags,
 670                          splice_actor *actor)
 671 {
 672         ssize_t ret;
 673         struct splice_desc sd = {
 674                 .total_len = len,
 675                 .flags = flags,
 676                 .pos = *ppos,
 677                 .u.file = out,
 678         };
 679
 680         pipe_lock(pipe);
 681         ret = __splice_from_pipe(pipe, &sd, actor);
 682         pipe_unlock(pipe);
 683
 684         return ret;
 685 }
 686
 687 /**
 688  * iter_file_splice_write - splice data from a pipe to a file
 689  * @pipe:       pipe info
 690  * @out:        file to write to
 691  * @ppos:       position in @out
 692  * @len:        number of bytes to splice
 693  * @flags:      splice modifier flags
 694  *
 695  * Description:
 696  *    Will either move or copy pages (determined by @flags options) from
 697  *    the given pipe inode to the given file.
 698  *    This one is ->write_iter-based.
 699  *
 700  */
 701 ssize_t
 702 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 703                           loff_t *ppos, size_t len, unsigned int flags)
 704 {
 705         struct splice_desc sd = {
 706                 .total_len = len,
 707                 .flags = flags,
 708                 .pos = *ppos,
 709                 .u.file = out,
 710         };
 711         int nbufs = pipe->max_usage;
 712         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 713                                         GFP_KERNEL);
 714         ssize_t ret;
 715
 716         if (unlikely(!array))
 717                 return -ENOMEM;
 718
 719         pipe_lock(pipe);
 720
 721         splice_from_pipe_begin(&sd);
 722         while (sd.total_len) {
 723                 struct iov_iter from;
 724                 unsigned int head, tail, mask;
 725                 size_t left;
 726                 int n;
 727
 728                 ret = splice_from_pipe_next(pipe, &sd);
 729                 if (ret <= 0)
 730                         break;
 731
 732                 if (unlikely(nbufs < pipe->max_usage)) {
 733                         kfree(array);
 734                         nbufs = pipe->max_usage;
 735                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 736                                         GFP_KERNEL);
 737                         if (!array) {
 738                                 ret = -ENOMEM;
 739                                 break;
 740                         }
 741                 }
 742
 743                 head = pipe->head;
 744                 tail = pipe->tail;
 745                 mask = pipe->ring_size - 1;
 746
 747                 /* build the vector */
 748                 left = sd.total_len;
 749                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++, n++) {
 750                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 751                         size_t this_len = buf->len;
 752
 753                         if (this_len > left)
 754                                 this_len = left;
 755
 756                         ret = pipe_buf_confirm(pipe, buf);
 757                         if (unlikely(ret)) {
 758                                 if (ret == -ENODATA)
 759                                         ret = 0;
 760                                 goto done;
 761                         }
 762
 763                         array[n].bv_page = buf->page;
 764                         array[n].bv_len = this_len;
 765                         array[n].bv_offset = buf->offset;
 766                         left -= this_len;
 767                 }
 768
 769                 iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
 770                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 771                 if (ret <= 0)
 772                         break;
 773
 774                 sd.num_spliced += ret;
 775                 sd.total_len -= ret;
 776                 *ppos = sd.pos;
 777
 778                 /* dismiss the fully eaten buffers, adjust the partial one */
 779                 tail = pipe->tail;
 780                 while (ret) {
 781                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 782                         if (ret >= buf->len) {
 783                                 ret -= buf->len;
 784                                 buf->len = 0;
 785                                 pipe_buf_release(pipe, buf);
 786                                 tail++;
 787                                 pipe->tail = tail;
 788                                 if (pipe->files)
 789                                         sd.need_wakeup = true;
 790                         } else {
 791                                 buf->offset += ret;
 792                                 buf->len -= ret;
 793                                 ret = 0;
 794                         }
 795                 }
 796         }
 797 done:
 798         kfree(array);
 799         splice_from_pipe_end(pipe, &sd);
 800
 801         pipe_unlock(pipe);
 802
 803         if (sd.num_spliced)
 804                 ret = sd.num_spliced;
 805
 806         return ret;
 807 }
 808
 809 EXPORT_SYMBOL(iter_file_splice_write);
 810
 811 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 812                           struct splice_desc *sd)
 813 {
 814         int ret;
 815         void *data;
 816         loff_t tmp = sd->pos;
 817
 818         data = kmap(buf->page);
 819         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 820         kunmap(buf->page);
 821
 822         return ret;
 823 }
 824
 825 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 826                                          struct file *out, loff_t *ppos,
 827                                          size_t len, unsigned int flags)
 828 {
 829         ssize_t ret;
 830
 831         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 832         if (ret > 0)
 833                 *ppos += ret;
 834
 835         return ret;
 836 }
 837
 838 /**
 839  * generic_splice_sendpage - splice data from a pipe to a socket
 840  * @pipe:       pipe to splice from
 841  * @out:        socket to write to
 842  * @ppos:       position in @out
 843  * @len:        number of bytes to splice
 844  * @flags:      splice modifier flags
 845  *
 846  * Description:
 847  *    Will send @len bytes from the pipe to a network socket. No data copying
 848  *    is involved.
 849  *
 850  */
 851 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 852                                 loff_t *ppos, size_t len, unsigned int flags)
 853 {
 854         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 855 }
 856
 857 EXPORT_SYMBOL(generic_splice_sendpage);
 858
 859 /*
 860  * Attempt to initiate a splice from pipe to file.
 861  */
 862 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 863                            loff_t *ppos, size_t len, unsigned int flags)
 864 {
 865         if (out->f_op->splice_write)
 866                 return out->f_op->splice_write(pipe, out, ppos, len, flags);
 867         return default_file_splice_write(pipe, out, ppos, len, flags);
 868 }
 869
 870 /*
 871  * Attempt to initiate a splice from a file to a pipe.
 872  */
 873 static long do_splice_to(struct file *in, loff_t *ppos,
 874                          struct pipe_inode_info *pipe, size_t len,
 875                          unsigned int flags)
 876 {
 877         int ret;
 878
 879         if (unlikely(!(in->f_mode & FMODE_READ)))
 880                 return -EBADF;
 881
 882         ret = rw_verify_area(READ, in, ppos, len);
 883         if (unlikely(ret < 0))
 884                 return ret;
 885
 886         if (unlikely(len > MAX_RW_COUNT))
 887                 len = MAX_RW_COUNT;
 888
 889         if (in->f_op->splice_read)
 890                 return in->f_op->splice_read(in, ppos, pipe, len, flags);
 891         return default_file_splice_read(in, ppos, pipe, len, flags);
 892 }
 893
 894 /**
 895  * splice_direct_to_actor - splices data directly between two non-pipes
 896  * @in:         file to splice from
 897  * @sd:         actor information on where to splice to
 898  * @actor:      handles the data splicing
 899  *
 900  * Description:
 901  *    This is a special case helper to splice directly between two
 902  *    points, without requiring an explicit pipe. Internally an allocated
 903  *    pipe is cached in the process, and reused during the lifetime of
 904  *    that process.
 905  *
 906  */
 907 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 908                                splice_direct_actor *actor)
 909 {
 910         struct pipe_inode_info *pipe;
 911         long ret, bytes;
 912         umode_t i_mode;
 913         size_t len;
 914         int i, flags, more;
 915
 916         /*
 917          * We require the input being a regular file, as we don't want to
 918          * randomly drop data for eg socket -> socket splicing. Use the
 919          * piped splicing for that!
 920          */
 921         i_mode = file_inode(in)->i_mode;
 922         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 923                 return -EINVAL;
 924
 925         /*
 926          * neither in nor out is a pipe, setup an internal pipe attached to
 927          * 'out' and transfer the wanted data from 'in' to 'out' through that
 928          */
 929         pipe = current->splice_pipe;
 930         if (unlikely(!pipe)) {
 931                 pipe = alloc_pipe_info();
 932                 if (!pipe)
 933                         return -ENOMEM;
 934
 935                 /*
 936                  * We don't have an immediate reader, but we'll read the stuff
 937                  * out of the pipe right after the splice_to_pipe(). So set
 938                  * PIPE_READERS appropriately.
 939                  */
 940                 pipe->readers = 1;
 941
 942                 current->splice_pipe = pipe;
 943         }
 944
 945         /*
 946          * Do the splice.
 947          */
 948         ret = 0;
 949         bytes = 0;
 950         len = sd->total_len;
 951         flags = sd->flags;
 952
 953         /*
 954          * Don't block on output, we have to drain the direct pipe.
 955          */
 956         sd->flags &= ~SPLICE_F_NONBLOCK;
 957         more = sd->flags & SPLICE_F_MORE;
 958
 959         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 960
 961         while (len) {
 962                 unsigned int p_space;
 963                 size_t read_len;
 964                 loff_t pos = sd->pos, prev_pos = pos;
 965
 966                 /* Don't try to read more the pipe has space for. */
 967                 p_space = pipe->max_usage -
 968                         pipe_occupancy(pipe->head, pipe->tail);
 969                 read_len = min_t(size_t, len, p_space << PAGE_SHIFT);
 970                 ret = do_splice_to(in, &pos, pipe, read_len, flags);
 971                 if (unlikely(ret <= 0))
 972                         goto out_release;
 973
 974                 read_len = ret;
 975                 sd->total_len = read_len;
 976
 977                 /*
 978                  * If more data is pending, set SPLICE_F_MORE
 979                  * If this is the last data and SPLICE_F_MORE was not set
 980                  * initially, clears it.
 981                  */
 982                 if (read_len < len)
 983                         sd->flags |= SPLICE_F_MORE;
 984                 else if (!more)
 985                         sd->flags &= ~SPLICE_F_MORE;
 986                 /*
 987                  * NOTE: nonblocking mode only applies to the input. We
 988                  * must not do the output in nonblocking mode as then we
 989                  * could get stuck data in the internal pipe:
 990                  */
 991                 ret = actor(pipe, sd);
 992                 if (unlikely(ret <= 0)) {
 993                         sd->pos = prev_pos;
 994                         goto out_release;
 995                 }
 996
 997                 bytes += ret;
 998                 len -= ret;
 999                 sd->pos = pos;
1000
1001                 if (ret < read_len) {
1002                         sd->pos = prev_pos + ret;
1003                         goto out_release;
1004                 }
1005         }
1006
1007 done:
1008         pipe->tail = pipe->head = 0;
1009         file_accessed(in);
1010         return bytes;
1011
1012 out_release:
1013         /*
1014          * If we did an incomplete transfer we must release
1015          * the pipe buffers in question:
1016          */
1017         for (i = 0; i < pipe->ring_size; i++) {
1018                 struct pipe_buffer *buf = &pipe->bufs[i];
1019
1020                 if (buf->ops)
1021                         pipe_buf_release(pipe, buf);
1022         }
1023
1024         if (!bytes)
1025                 bytes = ret;
1026
1027         goto done;
1028 }
1029 EXPORT_SYMBOL(splice_direct_to_actor);
1030
1031 static int direct_splice_actor(struct pipe_inode_info *pipe,
1032                                struct splice_desc *sd)
1033 {
1034         struct file *file = sd->u.file;
1035
1036         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1037                               sd->flags);
1038 }
1039
1040 /**
1041  * do_splice_direct - splices data directly between two files
1042  * @in:         file to splice from
1043  * @ppos:       input file offset
1044  * @out:        file to splice to
1045  * @opos:       output file offset
1046  * @len:        number of bytes to splice
1047  * @flags:      splice modifier flags
1048  *
1049  * Description:
1050  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1051  *    doing it in the application would incur an extra system call
1052  *    (splice in + splice out, as compared to just sendfile()). So this helper
1053  *    can splice directly through a process-private pipe.
1054  *
1055  */
1056 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1057                       loff_t *opos, size_t len, unsigned int flags)
1058 {
1059         struct splice_desc sd = {
1060                 .len            = len,
1061                 .total_len      = len,
1062                 .flags          = flags,
1063                 .pos            = *ppos,
1064                 .u.file         = out,
1065                 .opos           = opos,
1066         };
1067         long ret;
1068
1069         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1070                 return -EBADF;
1071
1072         if (unlikely(out->f_flags & O_APPEND))
1073                 return -EINVAL;
1074
1075         ret = rw_verify_area(WRITE, out, opos, len);
1076         if (unlikely(ret < 0))
1077                 return ret;
1078
1079         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1080         if (ret > 0)
1081                 *ppos = sd.pos;
1082
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(do_splice_direct);
1086
1087 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1088 {
1089         for (;;) {
1090                 if (unlikely(!pipe->readers)) {
1091                         send_sig(SIGPIPE, current, 0);
1092                         return -EPIPE;
1093                 }
1094                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1095                         return 0;
1096                 if (flags & SPLICE_F_NONBLOCK)
1097                         return -EAGAIN;
1098                 if (signal_pending(current))
1099                         return -ERESTARTSYS;
1100                 pipe_wait_writable(pipe);
1101         }
1102 }
1103
1104 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1105                                struct pipe_inode_info *opipe,
1106                                size_t len, unsigned int flags);
1107
1108 /*
1109  * Determine where to splice to/from.
1110  */
1111 long do_splice(struct file *in, loff_t __user *off_in,
1112                 struct file *out, loff_t __user *off_out,
1113                 size_t len, unsigned int flags)
1114 {
1115         struct pipe_inode_info *ipipe;
1116         struct pipe_inode_info *opipe;
1117         loff_t offset;
1118         long ret;
1119
1120         if (unlikely(!(in->f_mode & FMODE_READ) ||
1121                      !(out->f_mode & FMODE_WRITE)))
1122                 return -EBADF;
1123
1124         ipipe = get_pipe_info(in, true);
1125         opipe = get_pipe_info(out, true);
1126
1127         if (ipipe && opipe) {
1128                 if (off_in || off_out)
1129                         return -ESPIPE;
1130
1131                 /* Splicing to self would be fun, but... */
1132                 if (ipipe == opipe)
1133                         return -EINVAL;
1134
1135                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1136                         flags |= SPLICE_F_NONBLOCK;
1137
1138                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1139         }
1140
1141         if (ipipe) {
1142                 if (off_in)
1143                         return -ESPIPE;
1144                 if (off_out) {
1145                         if (!(out->f_mode & FMODE_PWRITE))
1146                                 return -EINVAL;
1147                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1148                                 return -EFAULT;
1149                 } else {
1150                         offset = out->f_pos;
1151                 }
1152
1153                 if (unlikely(out->f_flags & O_APPEND))
1154                         return -EINVAL;
1155
1156                 ret = rw_verify_area(WRITE, out, &offset, len);
1157                 if (unlikely(ret < 0))
1158                         return ret;
1159
1160                 if (in->f_flags & O_NONBLOCK)
1161                         flags |= SPLICE_F_NONBLOCK;
1162
1163                 file_start_write(out);
1164                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1165                 file_end_write(out);
1166
1167                 if (!off_out)
1168                         out->f_pos = offset;
1169                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1170                         ret = -EFAULT;
1171
1172                 return ret;
1173         }
1174
1175         if (opipe) {
1176                 if (off_out)
1177                         return -ESPIPE;
1178                 if (off_in) {
1179                         if (!(in->f_mode & FMODE_PREAD))
1180                                 return -EINVAL;
1181                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1182                                 return -EFAULT;
1183                 } else {
1184                         offset = in->f_pos;
1185                 }
1186
1187                 if (out->f_flags & O_NONBLOCK)
1188                         flags |= SPLICE_F_NONBLOCK;
1189
1190                 pipe_lock(opipe);
1191                 ret = wait_for_space(opipe, flags);
1192                 if (!ret) {
1193                         unsigned int p_space;
1194
1195                         /* Don't try to read more the pipe has space for. */
1196                         p_space = opipe->max_usage - pipe_occupancy(opipe->head, opipe->tail);
1197                         len = min_t(size_t, len, p_space << PAGE_SHIFT);
1198
1199                         ret = do_splice_to(in, &offset, opipe, len, flags);
1200                 }
1201                 pipe_unlock(opipe);
1202                 if (ret > 0)
1203                         wakeup_pipe_readers(opipe);
1204                 if (!off_in)
1205                         in->f_pos = offset;
1206                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1207                         ret = -EFAULT;
1208
1209                 return ret;
1210         }
1211
1212         return -EINVAL;
1213 }
1214
1215 static int iter_to_pipe(struct iov_iter *from,
1216                         struct pipe_inode_info *pipe,
1217                         unsigned flags)
1218 {
1219         struct pipe_buffer buf = {
1220                 .ops = &user_page_pipe_buf_ops,
1221                 .flags = flags
1222         };
1223         size_t total = 0;
1224         int ret = 0;
1225         bool failed = false;
1226
1227         while (iov_iter_count(from) && !failed) {
1228                 struct page *pages[16];
1229                 ssize_t copied;
1230                 size_t start;
1231                 int n;
1232
1233                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1234                 if (copied <= 0) {
1235                         ret = copied;
1236                         break;
1237                 }
1238
1239                 for (n = 0; copied; n++, start = 0) {
1240                         int size = min_t(int, copied, PAGE_SIZE - start);
1241                         if (!failed) {
1242                                 buf.page = pages[n];
1243                                 buf.offset = start;
1244                                 buf.len = size;
1245                                 ret = add_to_pipe(pipe, &buf);
1246                                 if (unlikely(ret < 0)) {
1247                                         failed = true;
1248                                 } else {
1249                                         iov_iter_advance(from, ret);
1250                                         total += ret;
1251                                 }
1252                         } else {
1253                                 put_page(pages[n]);
1254                         }
1255                         copied -= size;
1256                 }
1257         }
1258         return total ? total : ret;
1259 }
1260
1261 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1262                         struct splice_desc *sd)
1263 {
1264         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1265         return n == sd->len ? n : -EFAULT;
1266 }
1267
1268 /*
1269  * For lack of a better implementation, implement vmsplice() to userspace
1270  * as a simple copy of the pipes pages to the user iov.
1271  */
1272 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1273                              unsigned int flags)
1274 {
1275         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1276         struct splice_desc sd = {
1277                 .total_len = iov_iter_count(iter),
1278                 .flags = flags,
1279                 .u.data = iter
1280         };
1281         long ret = 0;
1282
1283         if (!pipe)
1284                 return -EBADF;
1285
1286         if (sd.total_len) {
1287                 pipe_lock(pipe);
1288                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1289                 pipe_unlock(pipe);
1290         }
1291
1292         return ret;
1293 }
1294
1295 /*
1296  * vmsplice splices a user address range into a pipe. It can be thought of
1297  * as splice-from-memory, where the regular splice is splice-from-file (or
1298  * to file). In both cases the output is a pipe, naturally.
1299  */
1300 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1301                              unsigned int flags)
1302 {
1303         struct pipe_inode_info *pipe;
1304         long ret = 0;
1305         unsigned buf_flag = 0;
1306
1307         if (flags & SPLICE_F_GIFT)
1308                 buf_flag = PIPE_BUF_FLAG_GIFT;
1309
1310         pipe = get_pipe_info(file, true);
1311         if (!pipe)
1312                 return -EBADF;
1313
1314         pipe_lock(pipe);
1315         ret = wait_for_space(pipe, flags);
1316         if (!ret)
1317                 ret = iter_to_pipe(iter, pipe, buf_flag);
1318         pipe_unlock(pipe);
1319         if (ret > 0)
1320                 wakeup_pipe_readers(pipe);
1321         return ret;
1322 }
1323
1324 static int vmsplice_type(struct fd f, int *type)
1325 {
1326         if (!f.file)
1327                 return -EBADF;
1328         if (f.file->f_mode & FMODE_WRITE) {
1329                 *type = WRITE;
1330         } else if (f.file->f_mode & FMODE_READ) {
1331                 *type = READ;
1332         } else {
1333                 fdput(f);
1334                 return -EBADF;
1335         }
1336         return 0;
1337 }
1338
1339 /*
1340  * Note that vmsplice only really supports true splicing _from_ user memory
1341  * to a pipe, not the other way around. Splicing from user memory is a simple
1342  * operation that can be supported without any funky alignment restrictions
1343  * or nasty vm tricks. We simply map in the user memory and fill them into
1344  * a pipe. The reverse isn't quite as easy, though. There are two possible
1345  * solutions for that:
1346  *
1347  *      - memcpy() the data internally, at which point we might as well just
1348  *        do a regular read() on the buffer anyway.
1349  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1350  *        has restriction limitations on both ends of the pipe).
1351  *
1352  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1353  *
1354  */
1355 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1356 {
1357         if (unlikely(flags & ~SPLICE_F_ALL))
1358                 return -EINVAL;
1359
1360         if (!iov_iter_count(iter))
1361                 return 0;
1362
1363         if (iov_iter_rw(iter) == WRITE)
1364                 return vmsplice_to_pipe(f, iter, flags);
1365         else
1366                 return vmsplice_to_user(f, iter, flags);
1367 }
1368
1369 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1370                 unsigned long, nr_segs, unsigned int, flags)
1371 {
1372         struct iovec iovstack[UIO_FASTIOV];
1373         struct iovec *iov = iovstack;
1374         struct iov_iter iter;
1375         ssize_t error;
1376         struct fd f;
1377         int type;
1378
1379         f = fdget(fd);
1380         error = vmsplice_type(f, &type);
1381         if (error)
1382                 return error;
1383
1384         error = import_iovec(type, uiov, nr_segs,
1385                              ARRAY_SIZE(iovstack), &iov, &iter);
1386         if (error >= 0) {
1387                 error = do_vmsplice(f.file, &iter, flags);
1388                 kfree(iov);
1389         }
1390         fdput(f);
1391         return error;
1392 }
1393
1394 #ifdef CONFIG_COMPAT
1395 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1396                     unsigned int, nr_segs, unsigned int, flags)
1397 {
1398         struct iovec iovstack[UIO_FASTIOV];
1399         struct iovec *iov = iovstack;
1400         struct iov_iter iter;
1401         ssize_t error;
1402         struct fd f;
1403         int type;
1404
1405         f = fdget(fd);
1406         error = vmsplice_type(f, &type);
1407         if (error)
1408                 return error;
1409
1410         error = compat_import_iovec(type, iov32, nr_segs,
1411                              ARRAY_SIZE(iovstack), &iov, &iter);
1412         if (error >= 0) {
1413                 error = do_vmsplice(f.file, &iter, flags);
1414                 kfree(iov);
1415         }
1416         fdput(f);
1417         return error;
1418 }
1419 #endif
1420
1421 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1422                 int, fd_out, loff_t __user *, off_out,
1423                 size_t, len, unsigned int, flags)
1424 {
1425         struct fd in, out;
1426         long error;
1427
1428         if (unlikely(!len))
1429                 return 0;
1430
1431         if (unlikely(flags & ~SPLICE_F_ALL))
1432                 return -EINVAL;
1433
1434         error = -EBADF;
1435         in = fdget(fd_in);
1436         if (in.file) {
1437                 out = fdget(fd_out);
1438                 if (out.file) {
1439                         error = do_splice(in.file, off_in, out.file, off_out,
1440                                           len, flags);
1441                         fdput(out);
1442                 }
1443                 fdput(in);
1444         }
1445         return error;
1446 }
1447
1448 /*
1449  * Make sure there's data to read. Wait for input if we can, otherwise
1450  * return an appropriate error.
1451  */
1452 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1453 {
1454         int ret;
1455
1456         /*
1457          * Check the pipe occupancy without the inode lock first. This function
1458          * is speculative anyways, so missing one is ok.
1459          */
1460         if (!pipe_empty(pipe->head, pipe->tail))
1461                 return 0;
1462
1463         ret = 0;
1464         pipe_lock(pipe);
1465
1466         while (pipe_empty(pipe->head, pipe->tail)) {
1467                 if (signal_pending(current)) {
1468                         ret = -ERESTARTSYS;
1469                         break;
1470                 }
1471                 if (!pipe->writers)
1472                         break;
1473                 if (flags & SPLICE_F_NONBLOCK) {
1474                         ret = -EAGAIN;
1475                         break;
1476                 }
1477                 pipe_wait_readable(pipe);
1478         }
1479
1480         pipe_unlock(pipe);
1481         return ret;
1482 }
1483
1484 /*
1485  * Make sure there's writeable room. Wait for room if we can, otherwise
1486  * return an appropriate error.
1487  */
1488 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1489 {
1490         int ret;
1491
1492         /*
1493          * Check pipe occupancy without the inode lock first. This function
1494          * is speculative anyways, so missing one is ok.
1495          */
1496         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1497                 return 0;
1498
1499         ret = 0;
1500         pipe_lock(pipe);
1501
1502         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1503                 if (!pipe->readers) {
1504                         send_sig(SIGPIPE, current, 0);
1505                         ret = -EPIPE;
1506                         break;
1507                 }
1508                 if (flags & SPLICE_F_NONBLOCK) {
1509                         ret = -EAGAIN;
1510                         break;
1511                 }
1512                 if (signal_pending(current)) {
1513                         ret = -ERESTARTSYS;
1514                         break;
1515                 }
1516                 pipe_wait_writable(pipe);
1517         }
1518
1519         pipe_unlock(pipe);
1520         return ret;
1521 }
1522
1523 /*
1524  * Splice contents of ipipe to opipe.
1525  */
1526 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1527                                struct pipe_inode_info *opipe,
1528                                size_t len, unsigned int flags)
1529 {
1530         struct pipe_buffer *ibuf, *obuf;
1531         unsigned int i_head, o_head;
1532         unsigned int i_tail, o_tail;
1533         unsigned int i_mask, o_mask;
1534         int ret = 0;
1535         bool input_wakeup = false;
1536
1537
1538 retry:
1539         ret = ipipe_prep(ipipe, flags);
1540         if (ret)
1541                 return ret;
1542
1543         ret = opipe_prep(opipe, flags);
1544         if (ret)
1545                 return ret;
1546
1547         /*
1548          * Potential ABBA deadlock, work around it by ordering lock
1549          * grabbing by pipe info address. Otherwise two different processes
1550          * could deadlock (one doing tee from A -> B, the other from B -> A).
1551          */
1552         pipe_double_lock(ipipe, opipe);
1553
1554         i_tail = ipipe->tail;
1555         i_mask = ipipe->ring_size - 1;
1556         o_head = opipe->head;
1557         o_mask = opipe->ring_size - 1;
1558
1559         do {
1560                 size_t o_len;
1561
1562                 if (!opipe->readers) {
1563                         send_sig(SIGPIPE, current, 0);
1564                         if (!ret)
1565                                 ret = -EPIPE;
1566                         break;
1567                 }
1568
1569                 i_head = ipipe->head;
1570                 o_tail = opipe->tail;
1571
1572                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1573                         break;
1574
1575                 /*
1576                  * Cannot make any progress, because either the input
1577                  * pipe is empty or the output pipe is full.
1578                  */
1579                 if (pipe_empty(i_head, i_tail) ||
1580                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1581                         /* Already processed some buffers, break */
1582                         if (ret)
1583                                 break;
1584
1585                         if (flags & SPLICE_F_NONBLOCK) {
1586                                 ret = -EAGAIN;
1587                                 break;
1588                         }
1589
1590                         /*
1591                          * We raced with another reader/writer and haven't
1592                          * managed to process any buffers.  A zero return
1593                          * value means EOF, so retry instead.
1594                          */
1595                         pipe_unlock(ipipe);
1596                         pipe_unlock(opipe);
1597                         goto retry;
1598                 }
1599
1600                 ibuf = &ipipe->bufs[i_tail & i_mask];
1601                 obuf = &opipe->bufs[o_head & o_mask];
1602
1603                 if (len >= ibuf->len) {
1604                         /*
1605                          * Simply move the whole buffer from ipipe to opipe
1606                          */
1607                         *obuf = *ibuf;
1608                         ibuf->ops = NULL;
1609                         i_tail++;
1610                         ipipe->tail = i_tail;
1611                         input_wakeup = true;
1612                         o_len = obuf->len;
1613                         o_head++;
1614                         opipe->head = o_head;
1615                 } else {
1616                         /*
1617                          * Get a reference to this pipe buffer,
1618                          * so we can copy the contents over.
1619                          */
1620                         if (!pipe_buf_get(ipipe, ibuf)) {
1621                                 if (ret == 0)
1622                                         ret = -EFAULT;
1623                                 break;
1624                         }
1625                         *obuf = *ibuf;
1626
1627                         /*
1628                          * Don't inherit the gift and merge flags, we need to
1629                          * prevent multiple steals of this page.
1630                          */
1631                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1632                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1633
1634                         obuf->len = len;
1635                         ibuf->offset += len;
1636                         ibuf->len -= len;
1637                         o_len = len;
1638                         o_head++;
1639                         opipe->head = o_head;
1640                 }
1641                 ret += o_len;
1642                 len -= o_len;
1643         } while (len);
1644
1645         pipe_unlock(ipipe);
1646         pipe_unlock(opipe);
1647
1648         /*
1649          * If we put data in the output pipe, wakeup any potential readers.
1650          */
1651         if (ret > 0)
1652                 wakeup_pipe_readers(opipe);
1653
1654         if (input_wakeup)
1655                 wakeup_pipe_writers(ipipe);
1656
1657         return ret;
1658 }
1659
1660 /*
1661  * Link contents of ipipe to opipe.
1662  */
1663 static int link_pipe(struct pipe_inode_info *ipipe,
1664                      struct pipe_inode_info *opipe,
1665                      size_t len, unsigned int flags)
1666 {
1667         struct pipe_buffer *ibuf, *obuf;
1668         unsigned int i_head, o_head;
1669         unsigned int i_tail, o_tail;
1670         unsigned int i_mask, o_mask;
1671         int ret = 0;
1672
1673         /*
1674          * Potential ABBA deadlock, work around it by ordering lock
1675          * grabbing by pipe info address. Otherwise two different processes
1676          * could deadlock (one doing tee from A -> B, the other from B -> A).
1677          */
1678         pipe_double_lock(ipipe, opipe);
1679
1680         i_tail = ipipe->tail;
1681         i_mask = ipipe->ring_size - 1;
1682         o_head = opipe->head;
1683         o_mask = opipe->ring_size - 1;
1684
1685         do {
1686                 if (!opipe->readers) {
1687                         send_sig(SIGPIPE, current, 0);
1688                         if (!ret)
1689                                 ret = -EPIPE;
1690                         break;
1691                 }
1692
1693                 i_head = ipipe->head;
1694                 o_tail = opipe->tail;
1695
1696                 /*
1697                  * If we have iterated all input buffers or run out of
1698                  * output room, break.
1699                  */
1700                 if (pipe_empty(i_head, i_tail) ||
1701                     pipe_full(o_head, o_tail, opipe->max_usage))
1702                         break;
1703
1704                 ibuf = &ipipe->bufs[i_tail & i_mask];
1705                 obuf = &opipe->bufs[o_head & o_mask];
1706
1707                 /*
1708                  * Get a reference to this pipe buffer,
1709                  * so we can copy the contents over.
1710                  */
1711                 if (!pipe_buf_get(ipipe, ibuf)) {
1712                         if (ret == 0)
1713                                 ret = -EFAULT;
1714                         break;
1715                 }
1716
1717                 *obuf = *ibuf;
1718
1719                 /*
1720                  * Don't inherit the gift and merge flag, we need to prevent
1721                  * multiple steals of this page.
1722                  */
1723                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1724                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1725
1726                 if (obuf->len > len)
1727                         obuf->len = len;
1728                 ret += obuf->len;
1729                 len -= obuf->len;
1730
1731                 o_head++;
1732                 opipe->head = o_head;
1733                 i_tail++;
1734         } while (len);
1735
1736         pipe_unlock(ipipe);
1737         pipe_unlock(opipe);
1738
1739         /*
1740          * If we put data in the output pipe, wakeup any potential readers.
1741          */
1742         if (ret > 0)
1743                 wakeup_pipe_readers(opipe);
1744
1745         return ret;
1746 }
1747
1748 /*
1749  * This is a tee(1) implementation that works on pipes. It doesn't copy
1750  * any data, it simply references the 'in' pages on the 'out' pipe.
1751  * The 'flags' used are the SPLICE_F_* variants, currently the only
1752  * applicable one is SPLICE_F_NONBLOCK.
1753  */
1754 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1755 {
1756         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1757         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1758         int ret = -EINVAL;
1759
1760         if (unlikely(!(in->f_mode & FMODE_READ) ||
1761                      !(out->f_mode & FMODE_WRITE)))
1762                 return -EBADF;
1763
1764         /*
1765          * Duplicate the contents of ipipe to opipe without actually
1766          * copying the data.
1767          */
1768         if (ipipe && opipe && ipipe != opipe) {
1769                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1770                         flags |= SPLICE_F_NONBLOCK;
1771
1772                 /*
1773                  * Keep going, unless we encounter an error. The ipipe/opipe
1774                  * ordering doesn't really matter.
1775                  */
1776                 ret = ipipe_prep(ipipe, flags);
1777                 if (!ret) {
1778                         ret = opipe_prep(opipe, flags);
1779                         if (!ret)
1780                                 ret = link_pipe(ipipe, opipe, len, flags);
1781                 }
1782         }
1783
1784         return ret;
1785 }
1786
1787 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1788 {
1789         struct fd in, out;
1790         int error;
1791
1792         if (unlikely(flags & ~SPLICE_F_ALL))
1793                 return -EINVAL;
1794
1795         if (unlikely(!len))
1796                 return 0;
1797
1798         error = -EBADF;
1799         in = fdget(fdin);
1800         if (in.file) {
1801                 out = fdget(fdout);
1802                 if (out.file) {
1803                         error = do_tee(in.file, out.file, len, flags);
1804                         fdput(out);
1805                 }
1806                 fdput(in);
1807         }
1808
1809         return error;
1810 }