fs/splice.c

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