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