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