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/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 /**
 530  * splice_from_pipe_next - wait for some data to splice from
 531  * @pipe:       pipe to splice from
 532  * @sd:         information about the splice operation
 533  *
 534  * Description:
 535  *    This function will wait for some data and return a positive
 536  *    value (one) if pipe buffers are available.  It will return zero
 537  *    or -errno if no more data needs to be spliced.
 538  */
 539 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 540 {
 541         /*
 542          * Check for signal early to make process killable when there are
 543          * always buffers available
 544          */
 545         if (signal_pending(current))
 546                 return -ERESTARTSYS;
 547
 548         while (pipe_empty(pipe->head, pipe->tail)) {
 549                 if (!pipe->writers)
 550                         return 0;
 551
 552                 if (sd->num_spliced)
 553                         return 0;
 554
 555                 if (sd->flags & SPLICE_F_NONBLOCK)
 556                         return -EAGAIN;
 557
 558                 if (signal_pending(current))
 559                         return -ERESTARTSYS;
 560
 561                 if (sd->need_wakeup) {
 562                         wakeup_pipe_writers(pipe);
 563                         sd->need_wakeup = false;
 564                 }
 565
 566                 pipe_wait(pipe);
 567         }
 568
 569         return 1;
 570 }
 571
 572 /**
 573  * splice_from_pipe_begin - start splicing from pipe
 574  * @sd:         information about the splice operation
 575  *
 576  * Description:
 577  *    This function should be called before a loop containing
 578  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 579  *    initialize the necessary fields of @sd.
 580  */
 581 static void splice_from_pipe_begin(struct splice_desc *sd)
 582 {
 583         sd->num_spliced = 0;
 584         sd->need_wakeup = false;
 585 }
 586
 587 /**
 588  * splice_from_pipe_end - finish splicing from pipe
 589  * @pipe:       pipe to splice from
 590  * @sd:         information about the splice operation
 591  *
 592  * Description:
 593  *    This function will wake up pipe writers if necessary.  It should
 594  *    be called after a loop containing splice_from_pipe_next() and
 595  *    splice_from_pipe_feed().
 596  */
 597 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 598 {
 599         if (sd->need_wakeup)
 600                 wakeup_pipe_writers(pipe);
 601 }
 602
 603 /**
 604  * __splice_from_pipe - splice data from a pipe to given actor
 605  * @pipe:       pipe to splice from
 606  * @sd:         information to @actor
 607  * @actor:      handler that splices the data
 608  *
 609  * Description:
 610  *    This function does little more than loop over the pipe and call
 611  *    @actor to do the actual moving of a single struct pipe_buffer to
 612  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 613  *    pipe_to_user.
 614  *
 615  */
 616 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 617                            splice_actor *actor)
 618 {
 619         int ret;
 620
 621         splice_from_pipe_begin(sd);
 622         do {
 623                 cond_resched();
 624                 ret = splice_from_pipe_next(pipe, sd);
 625                 if (ret > 0)
 626                         ret = splice_from_pipe_feed(pipe, sd, actor);
 627         } while (ret > 0);
 628         splice_from_pipe_end(pipe, sd);
 629
 630         return sd->num_spliced ? sd->num_spliced : ret;
 631 }
 632 EXPORT_SYMBOL(__splice_from_pipe);
 633
 634 /**
 635  * splice_from_pipe - splice data from a pipe to a file
 636  * @pipe:       pipe to splice from
 637  * @out:        file to splice to
 638  * @ppos:       position in @out
 639  * @len:        how many bytes to splice
 640  * @flags:      splice modifier flags
 641  * @actor:      handler that splices the data
 642  *
 643  * Description:
 644  *    See __splice_from_pipe. This function locks the pipe inode,
 645  *    otherwise it's identical to __splice_from_pipe().
 646  *
 647  */
 648 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 649                          loff_t *ppos, size_t len, unsigned int flags,
 650                          splice_actor *actor)
 651 {
 652         ssize_t ret;
 653         struct splice_desc sd = {
 654                 .total_len = len,
 655                 .flags = flags,
 656                 .pos = *ppos,
 657                 .u.file = out,
 658         };
 659
 660         pipe_lock(pipe);
 661         ret = __splice_from_pipe(pipe, &sd, actor);
 662         pipe_unlock(pipe);
 663
 664         return ret;
 665 }
 666
 667 /**
 668  * iter_file_splice_write - splice data from a pipe to a file
 669  * @pipe:       pipe info
 670  * @out:        file to write to
 671  * @ppos:       position in @out
 672  * @len:        number of bytes to splice
 673  * @flags:      splice modifier flags
 674  *
 675  * Description:
 676  *    Will either move or copy pages (determined by @flags options) from
 677  *    the given pipe inode to the given file.
 678  *    This one is ->write_iter-based.
 679  *
 680  */
 681 ssize_t
 682 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 683                           loff_t *ppos, size_t len, unsigned int flags)
 684 {
 685         struct splice_desc sd = {
 686                 .total_len = len,
 687                 .flags = flags,
 688                 .pos = *ppos,
 689                 .u.file = out,
 690         };
 691         int nbufs = pipe->max_usage;
 692         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 693                                         GFP_KERNEL);
 694         ssize_t ret;
 695
 696         if (unlikely(!array))
 697                 return -ENOMEM;
 698
 699         pipe_lock(pipe);
 700
 701         splice_from_pipe_begin(&sd);
 702         while (sd.total_len) {
 703                 struct iov_iter from;
 704                 unsigned int head, tail, mask;
 705                 size_t left;
 706                 int n;
 707
 708                 ret = splice_from_pipe_next(pipe, &sd);
 709                 if (ret <= 0)
 710                         break;
 711
 712                 if (unlikely(nbufs < pipe->max_usage)) {
 713                         kfree(array);
 714                         nbufs = pipe->max_usage;
 715                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 716                                         GFP_KERNEL);
 717                         if (!array) {
 718                                 ret = -ENOMEM;
 719                                 break;
 720                         }
 721                 }
 722
 723                 head = pipe->head;
 724                 tail = pipe->tail;
 725                 mask = pipe->ring_size - 1;
 726
 727                 /* build the vector */
 728                 left = sd.total_len;
 729                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++, n++) {
 730                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 731                         size_t this_len = buf->len;
 732
 733                         if (this_len > left)
 734                                 this_len = left;
 735
 736                         ret = pipe_buf_confirm(pipe, buf);
 737                         if (unlikely(ret)) {
 738                                 if (ret == -ENODATA)
 739                                         ret = 0;
 740                                 goto done;
 741                         }
 742
 743                         array[n].bv_page = buf->page;
 744                         array[n].bv_len = this_len;
 745                         array[n].bv_offset = buf->offset;
 746                         left -= this_len;
 747                 }
 748
 749                 iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
 750                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 751                 if (ret <= 0)
 752                         break;
 753
 754                 sd.num_spliced += ret;
 755                 sd.total_len -= ret;
 756                 *ppos = sd.pos;
 757
 758                 /* dismiss the fully eaten buffers, adjust the partial one */
 759                 tail = pipe->tail;
 760                 while (ret) {
 761                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 762                         if (ret >= buf->len) {
 763                                 ret -= buf->len;
 764                                 buf->len = 0;
 765                                 pipe_buf_release(pipe, buf);
 766                                 tail++;
 767                                 pipe->tail = tail;
 768                                 if (pipe->files)
 769                                         sd.need_wakeup = true;
 770                         } else {
 771                                 buf->offset += ret;
 772                                 buf->len -= ret;
 773                                 ret = 0;
 774                         }
 775                 }
 776         }
 777 done:
 778         kfree(array);
 779         splice_from_pipe_end(pipe, &sd);
 780
 781         pipe_unlock(pipe);
 782
 783         if (sd.num_spliced)
 784                 ret = sd.num_spliced;
 785
 786         return ret;
 787 }
 788
 789 EXPORT_SYMBOL(iter_file_splice_write);
 790
 791 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 792                           struct splice_desc *sd)
 793 {
 794         int ret;
 795         void *data;
 796         loff_t tmp = sd->pos;
 797
 798         data = kmap(buf->page);
 799         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 800         kunmap(buf->page);
 801
 802         return ret;
 803 }
 804
 805 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 806                                          struct file *out, loff_t *ppos,
 807                                          size_t len, unsigned int flags)
 808 {
 809         ssize_t ret;
 810
 811         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 812         if (ret > 0)
 813                 *ppos += ret;
 814
 815         return ret;
 816 }
 817
 818 /**
 819  * generic_splice_sendpage - splice data from a pipe to a socket
 820  * @pipe:       pipe to splice from
 821  * @out:        socket to write to
 822  * @ppos:       position in @out
 823  * @len:        number of bytes to splice
 824  * @flags:      splice modifier flags
 825  *
 826  * Description:
 827  *    Will send @len bytes from the pipe to a network socket. No data copying
 828  *    is involved.
 829  *
 830  */
 831 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 832                                 loff_t *ppos, size_t len, unsigned int flags)
 833 {
 834         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 835 }
 836
 837 EXPORT_SYMBOL(generic_splice_sendpage);
 838
 839 /*
 840  * Attempt to initiate a splice from pipe to file.
 841  */
 842 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 843                            loff_t *ppos, size_t len, unsigned int flags)
 844 {
 845         if (out->f_op->splice_write)
 846                 return out->f_op->splice_write(pipe, out, ppos, len, flags);
 847         return default_file_splice_write(pipe, out, ppos, len, flags);
 848 }
 849
 850 /*
 851  * Attempt to initiate a splice from a file to a pipe.
 852  */
 853 static long do_splice_to(struct file *in, loff_t *ppos,
 854                          struct pipe_inode_info *pipe, size_t len,
 855                          unsigned int flags)
 856 {
 857         int ret;
 858
 859         if (unlikely(!(in->f_mode & FMODE_READ)))
 860                 return -EBADF;
 861
 862         ret = rw_verify_area(READ, in, ppos, len);
 863         if (unlikely(ret < 0))
 864                 return ret;
 865
 866         if (unlikely(len > MAX_RW_COUNT))
 867                 len = MAX_RW_COUNT;
 868
 869         if (in->f_op->splice_read)
 870                 return in->f_op->splice_read(in, ppos, pipe, len, flags);
 871         return default_file_splice_read(in, ppos, pipe, len, flags);
 872 }
 873
 874 /**
 875  * splice_direct_to_actor - splices data directly between two non-pipes
 876  * @in:         file to splice from
 877  * @sd:         actor information on where to splice to
 878  * @actor:      handles the data splicing
 879  *
 880  * Description:
 881  *    This is a special case helper to splice directly between two
 882  *    points, without requiring an explicit pipe. Internally an allocated
 883  *    pipe is cached in the process, and reused during the lifetime of
 884  *    that process.
 885  *
 886  */
 887 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 888                                splice_direct_actor *actor)
 889 {
 890         struct pipe_inode_info *pipe;
 891         long ret, bytes;
 892         umode_t i_mode;
 893         size_t len;
 894         int i, flags, more;
 895
 896         /*
 897          * We require the input being a regular file, as we don't want to
 898          * randomly drop data for eg socket -> socket splicing. Use the
 899          * piped splicing for that!
 900          */
 901         i_mode = file_inode(in)->i_mode;
 902         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 903                 return -EINVAL;
 904
 905         /*
 906          * neither in nor out is a pipe, setup an internal pipe attached to
 907          * 'out' and transfer the wanted data from 'in' to 'out' through that
 908          */
 909         pipe = current->splice_pipe;
 910         if (unlikely(!pipe)) {
 911                 pipe = alloc_pipe_info();
 912                 if (!pipe)
 913                         return -ENOMEM;
 914
 915                 /*
 916                  * We don't have an immediate reader, but we'll read the stuff
 917                  * out of the pipe right after the splice_to_pipe(). So set
 918                  * PIPE_READERS appropriately.
 919                  */
 920                 pipe->readers = 1;
 921
 922                 current->splice_pipe = pipe;
 923         }
 924
 925         /*
 926          * Do the splice.
 927          */
 928         ret = 0;
 929         bytes = 0;
 930         len = sd->total_len;
 931         flags = sd->flags;
 932
 933         /*
 934          * Don't block on output, we have to drain the direct pipe.
 935          */
 936         sd->flags &= ~SPLICE_F_NONBLOCK;
 937         more = sd->flags & SPLICE_F_MORE;
 938
 939         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 940
 941         while (len) {
 942                 unsigned int p_space;
 943                 size_t read_len;
 944                 loff_t pos = sd->pos, prev_pos = pos;
 945
 946                 /* Don't try to read more the pipe has space for. */
 947                 p_space = pipe->max_usage -
 948                         pipe_occupancy(pipe->head, pipe->tail);
 949                 read_len = min_t(size_t, len, p_space << PAGE_SHIFT);
 950                 ret = do_splice_to(in, &pos, pipe, read_len, flags);
 951                 if (unlikely(ret <= 0))
 952                         goto out_release;
 953
 954                 read_len = ret;
 955                 sd->total_len = read_len;
 956
 957                 /*
 958                  * If more data is pending, set SPLICE_F_MORE
 959                  * If this is the last data and SPLICE_F_MORE was not set
 960                  * initially, clears it.
 961                  */
 962                 if (read_len < len)
 963                         sd->flags |= SPLICE_F_MORE;
 964                 else if (!more)
 965                         sd->flags &= ~SPLICE_F_MORE;
 966                 /*
 967                  * NOTE: nonblocking mode only applies to the input. We
 968                  * must not do the output in nonblocking mode as then we
 969                  * could get stuck data in the internal pipe:
 970                  */
 971                 ret = actor(pipe, sd);
 972                 if (unlikely(ret <= 0)) {
 973                         sd->pos = prev_pos;
 974                         goto out_release;
 975                 }
 976
 977                 bytes += ret;
 978                 len -= ret;
 979                 sd->pos = pos;
 980
 981                 if (ret < read_len) {
 982                         sd->pos = prev_pos + ret;
 983                         goto out_release;
 984                 }
 985         }
 986
 987 done:
 988         pipe->tail = pipe->head = 0;
 989         file_accessed(in);
 990         return bytes;
 991
 992 out_release:
 993         /*
 994          * If we did an incomplete transfer we must release
 995          * the pipe buffers in question:
 996          */
 997         for (i = 0; i < pipe->ring_size; i++) {
 998                 struct pipe_buffer *buf = &pipe->bufs[i];
 999
1000                 if (buf->ops)
1001                         pipe_buf_release(pipe, buf);
1002         }
1003
1004         if (!bytes)
1005                 bytes = ret;
1006
1007         goto done;
1008 }
1009 EXPORT_SYMBOL(splice_direct_to_actor);
1010
1011 static int direct_splice_actor(struct pipe_inode_info *pipe,
1012                                struct splice_desc *sd)
1013 {
1014         struct file *file = sd->u.file;
1015
1016         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1017                               sd->flags);
1018 }
1019
1020 /**
1021  * do_splice_direct - splices data directly between two files
1022  * @in:         file to splice from
1023  * @ppos:       input file offset
1024  * @out:        file to splice to
1025  * @opos:       output file offset
1026  * @len:        number of bytes to splice
1027  * @flags:      splice modifier flags
1028  *
1029  * Description:
1030  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1031  *    doing it in the application would incur an extra system call
1032  *    (splice in + splice out, as compared to just sendfile()). So this helper
1033  *    can splice directly through a process-private pipe.
1034  *
1035  */
1036 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1037                       loff_t *opos, size_t len, unsigned int flags)
1038 {
1039         struct splice_desc sd = {
1040                 .len            = len,
1041                 .total_len      = len,
1042                 .flags          = flags,
1043                 .pos            = *ppos,
1044                 .u.file         = out,
1045                 .opos           = opos,
1046         };
1047         long ret;
1048
1049         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1050                 return -EBADF;
1051
1052         if (unlikely(out->f_flags & O_APPEND))
1053                 return -EINVAL;
1054
1055         ret = rw_verify_area(WRITE, out, opos, len);
1056         if (unlikely(ret < 0))
1057                 return ret;
1058
1059         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1060         if (ret > 0)
1061                 *ppos = sd.pos;
1062
1063         return ret;
1064 }
1065 EXPORT_SYMBOL(do_splice_direct);
1066
1067 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1068 {
1069         for (;;) {
1070                 if (unlikely(!pipe->readers)) {
1071                         send_sig(SIGPIPE, current, 0);
1072                         return -EPIPE;
1073                 }
1074                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1075                         return 0;
1076                 if (flags & SPLICE_F_NONBLOCK)
1077                         return -EAGAIN;
1078                 if (signal_pending(current))
1079                         return -ERESTARTSYS;
1080                 pipe_wait(pipe);
1081         }
1082 }
1083
1084 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1085                                struct pipe_inode_info *opipe,
1086                                size_t len, unsigned int flags);
1087
1088 /*
1089  * Determine where to splice to/from.
1090  */
1091 long do_splice(struct file *in, loff_t __user *off_in,
1092                 struct file *out, loff_t __user *off_out,
1093                 size_t len, unsigned int flags)
1094 {
1095         struct pipe_inode_info *ipipe;
1096         struct pipe_inode_info *opipe;
1097         loff_t offset;
1098         long ret;
1099
1100         if (unlikely(!(in->f_mode & FMODE_READ) ||
1101                      !(out->f_mode & FMODE_WRITE)))
1102                 return -EBADF;
1103
1104         ipipe = get_pipe_info(in, true);
1105         opipe = get_pipe_info(out, true);
1106
1107         if (ipipe && opipe) {
1108                 if (off_in || off_out)
1109                         return -ESPIPE;
1110
1111                 /* Splicing to self would be fun, but... */
1112                 if (ipipe == opipe)
1113                         return -EINVAL;
1114
1115                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1116                         flags |= SPLICE_F_NONBLOCK;
1117
1118                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1119         }
1120
1121         if (ipipe) {
1122                 if (off_in)
1123                         return -ESPIPE;
1124                 if (off_out) {
1125                         if (!(out->f_mode & FMODE_PWRITE))
1126                                 return -EINVAL;
1127                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1128                                 return -EFAULT;
1129                 } else {
1130                         offset = out->f_pos;
1131                 }
1132
1133                 if (unlikely(out->f_flags & O_APPEND))
1134                         return -EINVAL;
1135
1136                 ret = rw_verify_area(WRITE, out, &offset, len);
1137                 if (unlikely(ret < 0))
1138                         return ret;
1139
1140                 if (in->f_flags & O_NONBLOCK)
1141                         flags |= SPLICE_F_NONBLOCK;
1142
1143                 file_start_write(out);
1144                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1145                 file_end_write(out);
1146
1147                 if (!off_out)
1148                         out->f_pos = offset;
1149                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1150                         ret = -EFAULT;
1151
1152                 return ret;
1153         }
1154
1155         if (opipe) {
1156                 if (off_out)
1157                         return -ESPIPE;
1158                 if (off_in) {
1159                         if (!(in->f_mode & FMODE_PREAD))
1160                                 return -EINVAL;
1161                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1162                                 return -EFAULT;
1163                 } else {
1164                         offset = in->f_pos;
1165                 }
1166
1167                 if (out->f_flags & O_NONBLOCK)
1168                         flags |= SPLICE_F_NONBLOCK;
1169
1170                 pipe_lock(opipe);
1171                 ret = wait_for_space(opipe, flags);
1172                 if (!ret) {
1173                         unsigned int p_space;
1174
1175                         /* Don't try to read more the pipe has space for. */
1176                         p_space = opipe->max_usage - pipe_occupancy(opipe->head, opipe->tail);
1177                         len = min_t(size_t, len, p_space << PAGE_SHIFT);
1178
1179                         ret = do_splice_to(in, &offset, opipe, len, flags);
1180                 }
1181                 pipe_unlock(opipe);
1182                 if (ret > 0)
1183                         wakeup_pipe_readers(opipe);
1184                 if (!off_in)
1185                         in->f_pos = offset;
1186                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1187                         ret = -EFAULT;
1188
1189                 return ret;
1190         }
1191
1192         return -EINVAL;
1193 }
1194
1195 static int iter_to_pipe(struct iov_iter *from,
1196                         struct pipe_inode_info *pipe,
1197                         unsigned flags)
1198 {
1199         struct pipe_buffer buf = {
1200                 .ops = &user_page_pipe_buf_ops,
1201                 .flags = flags
1202         };
1203         size_t total = 0;
1204         int ret = 0;
1205         bool failed = false;
1206
1207         while (iov_iter_count(from) && !failed) {
1208                 struct page *pages[16];
1209                 ssize_t copied;
1210                 size_t start;
1211                 int n;
1212
1213                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1214                 if (copied <= 0) {
1215                         ret = copied;
1216                         break;
1217                 }
1218
1219                 for (n = 0; copied; n++, start = 0) {
1220                         int size = min_t(int, copied, PAGE_SIZE - start);
1221                         if (!failed) {
1222                                 buf.page = pages[n];
1223                                 buf.offset = start;
1224                                 buf.len = size;
1225                                 ret = add_to_pipe(pipe, &buf);
1226                                 if (unlikely(ret < 0)) {
1227                                         failed = true;
1228                                 } else {
1229                                         iov_iter_advance(from, ret);
1230                                         total += ret;
1231                                 }
1232                         } else {
1233                                 put_page(pages[n]);
1234                         }
1235                         copied -= size;
1236                 }
1237         }
1238         return total ? total : ret;
1239 }
1240
1241 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1242                         struct splice_desc *sd)
1243 {
1244         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1245         return n == sd->len ? n : -EFAULT;
1246 }
1247
1248 /*
1249  * For lack of a better implementation, implement vmsplice() to userspace
1250  * as a simple copy of the pipes pages to the user iov.
1251  */
1252 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1253                              unsigned int flags)
1254 {
1255         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1256         struct splice_desc sd = {
1257                 .total_len = iov_iter_count(iter),
1258                 .flags = flags,
1259                 .u.data = iter
1260         };
1261         long ret = 0;
1262
1263         if (!pipe)
1264                 return -EBADF;
1265
1266         if (sd.total_len) {
1267                 pipe_lock(pipe);
1268                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1269                 pipe_unlock(pipe);
1270         }
1271
1272         return ret;
1273 }
1274
1275 /*
1276  * vmsplice splices a user address range into a pipe. It can be thought of
1277  * as splice-from-memory, where the regular splice is splice-from-file (or
1278  * to file). In both cases the output is a pipe, naturally.
1279  */
1280 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1281                              unsigned int flags)
1282 {
1283         struct pipe_inode_info *pipe;
1284         long ret = 0;
1285         unsigned buf_flag = 0;
1286
1287         if (flags & SPLICE_F_GIFT)
1288                 buf_flag = PIPE_BUF_FLAG_GIFT;
1289
1290         pipe = get_pipe_info(file, true);
1291         if (!pipe)
1292                 return -EBADF;
1293
1294         pipe_lock(pipe);
1295         ret = wait_for_space(pipe, flags);
1296         if (!ret)
1297                 ret = iter_to_pipe(iter, pipe, buf_flag);
1298         pipe_unlock(pipe);
1299         if (ret > 0)
1300                 wakeup_pipe_readers(pipe);
1301         return ret;
1302 }
1303
1304 static int vmsplice_type(struct fd f, int *type)
1305 {
1306         if (!f.file)
1307                 return -EBADF;
1308         if (f.file->f_mode & FMODE_WRITE) {
1309                 *type = WRITE;
1310         } else if (f.file->f_mode & FMODE_READ) {
1311                 *type = READ;
1312         } else {
1313                 fdput(f);
1314                 return -EBADF;
1315         }
1316         return 0;
1317 }
1318
1319 /*
1320  * Note that vmsplice only really supports true splicing _from_ user memory
1321  * to a pipe, not the other way around. Splicing from user memory is a simple
1322  * operation that can be supported without any funky alignment restrictions
1323  * or nasty vm tricks. We simply map in the user memory and fill them into
1324  * a pipe. The reverse isn't quite as easy, though. There are two possible
1325  * solutions for that:
1326  *
1327  *      - memcpy() the data internally, at which point we might as well just
1328  *        do a regular read() on the buffer anyway.
1329  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1330  *        has restriction limitations on both ends of the pipe).
1331  *
1332  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1333  *
1334  */
1335 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1336 {
1337         if (unlikely(flags & ~SPLICE_F_ALL))
1338                 return -EINVAL;
1339
1340         if (!iov_iter_count(iter))
1341                 return 0;
1342
1343         if (iov_iter_rw(iter) == WRITE)
1344                 return vmsplice_to_pipe(f, iter, flags);
1345         else
1346                 return vmsplice_to_user(f, iter, flags);
1347 }
1348
1349 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1350                 unsigned long, nr_segs, unsigned int, flags)
1351 {
1352         struct iovec iovstack[UIO_FASTIOV];
1353         struct iovec *iov = iovstack;
1354         struct iov_iter iter;
1355         ssize_t error;
1356         struct fd f;
1357         int type;
1358
1359         f = fdget(fd);
1360         error = vmsplice_type(f, &type);
1361         if (error)
1362                 return error;
1363
1364         error = import_iovec(type, uiov, nr_segs,
1365                              ARRAY_SIZE(iovstack), &iov, &iter);
1366         if (error >= 0) {
1367                 error = do_vmsplice(f.file, &iter, flags);
1368                 kfree(iov);
1369         }
1370         fdput(f);
1371         return error;
1372 }
1373
1374 #ifdef CONFIG_COMPAT
1375 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1376                     unsigned int, nr_segs, unsigned int, flags)
1377 {
1378         struct iovec iovstack[UIO_FASTIOV];
1379         struct iovec *iov = iovstack;
1380         struct iov_iter iter;
1381         ssize_t error;
1382         struct fd f;
1383         int type;
1384
1385         f = fdget(fd);
1386         error = vmsplice_type(f, &type);
1387         if (error)
1388                 return error;
1389
1390         error = compat_import_iovec(type, iov32, nr_segs,
1391                              ARRAY_SIZE(iovstack), &iov, &iter);
1392         if (error >= 0) {
1393                 error = do_vmsplice(f.file, &iter, flags);
1394                 kfree(iov);
1395         }
1396         fdput(f);
1397         return error;
1398 }
1399 #endif
1400
1401 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1402                 int, fd_out, loff_t __user *, off_out,
1403                 size_t, len, unsigned int, flags)
1404 {
1405         struct fd in, out;
1406         long error;
1407
1408         if (unlikely(!len))
1409                 return 0;
1410
1411         if (unlikely(flags & ~SPLICE_F_ALL))
1412                 return -EINVAL;
1413
1414         error = -EBADF;
1415         in = fdget(fd_in);
1416         if (in.file) {
1417                 out = fdget(fd_out);
1418                 if (out.file) {
1419                         error = do_splice(in.file, off_in, out.file, off_out,
1420                                           len, flags);
1421                         fdput(out);
1422                 }
1423                 fdput(in);
1424         }
1425         return error;
1426 }
1427
1428 /*
1429  * Make sure there's data to read. Wait for input if we can, otherwise
1430  * return an appropriate error.
1431  */
1432 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1433 {
1434         int ret;
1435
1436         /*
1437          * Check the pipe occupancy without the inode lock first. This function
1438          * is speculative anyways, so missing one is ok.
1439          */
1440         if (!pipe_empty(pipe->head, pipe->tail))
1441                 return 0;
1442
1443         ret = 0;
1444         pipe_lock(pipe);
1445
1446         while (pipe_empty(pipe->head, pipe->tail)) {
1447                 if (signal_pending(current)) {
1448                         ret = -ERESTARTSYS;
1449                         break;
1450                 }
1451                 if (!pipe->writers)
1452                         break;
1453                 if (flags & SPLICE_F_NONBLOCK) {
1454                         ret = -EAGAIN;
1455                         break;
1456                 }
1457                 pipe_wait(pipe);
1458         }
1459
1460         pipe_unlock(pipe);
1461         return ret;
1462 }
1463
1464 /*
1465  * Make sure there's writeable room. Wait for room if we can, otherwise
1466  * return an appropriate error.
1467  */
1468 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1469 {
1470         int ret;
1471
1472         /*
1473          * Check pipe occupancy without the inode lock first. This function
1474          * is speculative anyways, so missing one is ok.
1475          */
1476         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1477                 return 0;
1478
1479         ret = 0;
1480         pipe_lock(pipe);
1481
1482         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1483                 if (!pipe->readers) {
1484                         send_sig(SIGPIPE, current, 0);
1485                         ret = -EPIPE;
1486                         break;
1487                 }
1488                 if (flags & SPLICE_F_NONBLOCK) {
1489                         ret = -EAGAIN;
1490                         break;
1491                 }
1492                 if (signal_pending(current)) {
1493                         ret = -ERESTARTSYS;
1494                         break;
1495                 }
1496                 pipe_wait(pipe);
1497         }
1498
1499         pipe_unlock(pipe);
1500         return ret;
1501 }
1502
1503 /*
1504  * Splice contents of ipipe to opipe.
1505  */
1506 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1507                                struct pipe_inode_info *opipe,
1508                                size_t len, unsigned int flags)
1509 {
1510         struct pipe_buffer *ibuf, *obuf;
1511         unsigned int i_head, o_head;
1512         unsigned int i_tail, o_tail;
1513         unsigned int i_mask, o_mask;
1514         int ret = 0;
1515         bool input_wakeup = false;
1516
1517
1518 retry:
1519         ret = ipipe_prep(ipipe, flags);
1520         if (ret)
1521                 return ret;
1522
1523         ret = opipe_prep(opipe, flags);
1524         if (ret)
1525                 return ret;
1526
1527         /*
1528          * Potential ABBA deadlock, work around it by ordering lock
1529          * grabbing by pipe info address. Otherwise two different processes
1530          * could deadlock (one doing tee from A -> B, the other from B -> A).
1531          */
1532         pipe_double_lock(ipipe, opipe);
1533
1534         i_tail = ipipe->tail;
1535         i_mask = ipipe->ring_size - 1;
1536         o_head = opipe->head;
1537         o_mask = opipe->ring_size - 1;
1538
1539         do {
1540                 size_t o_len;
1541
1542                 if (!opipe->readers) {
1543                         send_sig(SIGPIPE, current, 0);
1544                         if (!ret)
1545                                 ret = -EPIPE;
1546                         break;
1547                 }
1548
1549                 i_head = ipipe->head;
1550                 o_tail = opipe->tail;
1551
1552                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1553                         break;
1554
1555                 /*
1556                  * Cannot make any progress, because either the input
1557                  * pipe is empty or the output pipe is full.
1558                  */
1559                 if (pipe_empty(i_head, i_tail) ||
1560                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1561                         /* Already processed some buffers, break */
1562                         if (ret)
1563                                 break;
1564
1565                         if (flags & SPLICE_F_NONBLOCK) {
1566                                 ret = -EAGAIN;
1567                                 break;
1568                         }
1569
1570                         /*
1571                          * We raced with another reader/writer and haven't
1572                          * managed to process any buffers.  A zero return
1573                          * value means EOF, so retry instead.
1574                          */
1575                         pipe_unlock(ipipe);
1576                         pipe_unlock(opipe);
1577                         goto retry;
1578                 }
1579
1580                 ibuf = &ipipe->bufs[i_tail & i_mask];
1581                 obuf = &opipe->bufs[o_head & o_mask];
1582
1583                 if (len >= ibuf->len) {
1584                         /*
1585                          * Simply move the whole buffer from ipipe to opipe
1586                          */
1587                         *obuf = *ibuf;
1588                         ibuf->ops = NULL;
1589                         i_tail++;
1590                         ipipe->tail = i_tail;
1591                         input_wakeup = true;
1592                         o_len = obuf->len;
1593                         o_head++;
1594                         opipe->head = o_head;
1595                 } else {
1596                         /*
1597                          * Get a reference to this pipe buffer,
1598                          * so we can copy the contents over.
1599                          */
1600                         if (!pipe_buf_get(ipipe, ibuf)) {
1601                                 if (ret == 0)
1602                                         ret = -EFAULT;
1603                                 break;
1604                         }
1605                         *obuf = *ibuf;
1606
1607                         /*
1608                          * Don't inherit the gift and merge flags, we need to
1609                          * prevent multiple steals of this page.
1610                          */
1611                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1612                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1613
1614                         obuf->len = len;
1615                         ibuf->offset += len;
1616                         ibuf->len -= len;
1617                         o_len = len;
1618                         o_head++;
1619                         opipe->head = o_head;
1620                 }
1621                 ret += o_len;
1622                 len -= o_len;
1623         } while (len);
1624
1625         pipe_unlock(ipipe);
1626         pipe_unlock(opipe);
1627
1628         /*
1629          * If we put data in the output pipe, wakeup any potential readers.
1630          */
1631         if (ret > 0)
1632                 wakeup_pipe_readers(opipe);
1633
1634         if (input_wakeup)
1635                 wakeup_pipe_writers(ipipe);
1636
1637         return ret;
1638 }
1639
1640 /*
1641  * Link contents of ipipe to opipe.
1642  */
1643 static int link_pipe(struct pipe_inode_info *ipipe,
1644                      struct pipe_inode_info *opipe,
1645                      size_t len, unsigned int flags)
1646 {
1647         struct pipe_buffer *ibuf, *obuf;
1648         unsigned int i_head, o_head;
1649         unsigned int i_tail, o_tail;
1650         unsigned int i_mask, o_mask;
1651         int ret = 0;
1652
1653         /*
1654          * Potential ABBA deadlock, work around it by ordering lock
1655          * grabbing by pipe info address. Otherwise two different processes
1656          * could deadlock (one doing tee from A -> B, the other from B -> A).
1657          */
1658         pipe_double_lock(ipipe, opipe);
1659
1660         i_tail = ipipe->tail;
1661         i_mask = ipipe->ring_size - 1;
1662         o_head = opipe->head;
1663         o_mask = opipe->ring_size - 1;
1664
1665         do {
1666                 if (!opipe->readers) {
1667                         send_sig(SIGPIPE, current, 0);
1668                         if (!ret)
1669                                 ret = -EPIPE;
1670                         break;
1671                 }
1672
1673                 i_head = ipipe->head;
1674                 o_tail = opipe->tail;
1675
1676                 /*
1677                  * If we have iterated all input buffers or run out of
1678                  * output room, break.
1679                  */
1680                 if (pipe_empty(i_head, i_tail) ||
1681                     pipe_full(o_head, o_tail, opipe->max_usage))
1682                         break;
1683
1684                 ibuf = &ipipe->bufs[i_tail & i_mask];
1685                 obuf = &opipe->bufs[o_head & o_mask];
1686
1687                 /*
1688                  * Get a reference to this pipe buffer,
1689                  * so we can copy the contents over.
1690                  */
1691                 if (!pipe_buf_get(ipipe, ibuf)) {
1692                         if (ret == 0)
1693                                 ret = -EFAULT;
1694                         break;
1695                 }
1696
1697                 *obuf = *ibuf;
1698
1699                 /*
1700                  * Don't inherit the gift and merge flag, we need to prevent
1701                  * multiple steals of this page.
1702                  */
1703                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1704                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1705
1706                 if (obuf->len > len)
1707                         obuf->len = len;
1708                 ret += obuf->len;
1709                 len -= obuf->len;
1710
1711                 o_head++;
1712                 opipe->head = o_head;
1713                 i_tail++;
1714         } while (len);
1715
1716         pipe_unlock(ipipe);
1717         pipe_unlock(opipe);
1718
1719         /*
1720          * If we put data in the output pipe, wakeup any potential readers.
1721          */
1722         if (ret > 0)
1723                 wakeup_pipe_readers(opipe);
1724
1725         return ret;
1726 }
1727
1728 /*
1729  * This is a tee(1) implementation that works on pipes. It doesn't copy
1730  * any data, it simply references the 'in' pages on the 'out' pipe.
1731  * The 'flags' used are the SPLICE_F_* variants, currently the only
1732  * applicable one is SPLICE_F_NONBLOCK.
1733  */
1734 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1735 {
1736         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1737         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1738         int ret = -EINVAL;
1739
1740         if (unlikely(!(in->f_mode & FMODE_READ) ||
1741                      !(out->f_mode & FMODE_WRITE)))
1742                 return -EBADF;
1743
1744         /*
1745          * Duplicate the contents of ipipe to opipe without actually
1746          * copying the data.
1747          */
1748         if (ipipe && opipe && ipipe != opipe) {
1749                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1750                         flags |= SPLICE_F_NONBLOCK;
1751
1752                 /*
1753                  * Keep going, unless we encounter an error. The ipipe/opipe
1754                  * ordering doesn't really matter.
1755                  */
1756                 ret = ipipe_prep(ipipe, flags);
1757                 if (!ret) {
1758                         ret = opipe_prep(opipe, flags);
1759                         if (!ret)
1760                                 ret = link_pipe(ipipe, opipe, len, flags);
1761                 }
1762         }
1763
1764         return ret;
1765 }
1766
1767 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1768 {
1769         struct fd in, out;
1770         int error;
1771
1772         if (unlikely(flags & ~SPLICE_F_ALL))
1773                 return -EINVAL;
1774
1775         if (unlikely(!len))
1776                 return 0;
1777
1778         error = -EBADF;
1779         in = fdget(fdin);
1780         if (in.file) {
1781                 out = fdget(fdout);
1782                 if (out.file) {
1783                         error = do_tee(in.file, out.file, len, flags);
1784                         fdput(out);
1785                 }
1786                 fdput(in);
1787         }
1788
1789         return error;
1790 }